Xu et al. Chin. J. Mech. Eng.

(2020) 33:43
https://doi.org/10.1186/s10033-020-00459-x Chinese Journal of Mechanical
Engineering

REVIEW Open Access

Advanced Data Collection and Analysis

in Data‑Driven Manufacturing Process
Ke Xu1, Yingguang Li1* , Changqing Liu1, Xu Liu2, Xiaozhong Hao1, James Gao3 and Paul G. Maropoulos4

Abstract
The rapidly increasing demand and complexity of manufacturing process potentiates the usage of manufacturing
data with the highest priority to achieve precise analyze and control, rather than using simplified physical models and
human expertise. In the era of data-driven manufacturing, the explosion of data amount revolutionized how data is
collected and analyzed. This paper overviews the advance of technologies developed for in-process manufacturing
data collection and analysis. It can be concluded that groundbreaking sensoring technology to facilitate direct meas-
urement is one important leading trend for advanced data collection, due to the complexity and uncertainty during
indirect measurement. On the other hand, physical model-based data analysis contains inevitable simplifications and
sometimes ill-posed solutions due to the limited capacity of describing complex manufacturing process. Machine
learning, especially deep learning approach has great potential for making better decisions to automate the process
when fed with abundant data, while trending data-driven manufacturing approaches succeeded by using limited
data to achieve similar or even better decisions. And these trends can demonstrated be by analyzing some typical
applications of manufacturing process.
Keywords: Data-driven manufacturing, Intelligent manufacturing, Process monitoring, Data analysis, Machine
learning

1 Introduction as IoT-enabled manufacturing, digital twin, cloud manu-

The concept of “smart manufacturing” (also known as facturing. Manufacturing data can be captured at all
intelligent manufacturing) is experiencing an explosive stages in a product life, ranging from explicit values such
propagation. According to Tan et al. [1], smart manu- as material properties, process temperature, vibration to
facturing is the fusion of intelligent technologies and implicit ones like supply chain resource and customers’
manufacturing technologies. Smart manufacturing is an preferences. The volume of data generated by manufac-
umbrella term of manufacturing technologies and para- turing systems is growing rapidly with over 1000 EB [3]
digms aiming to automate production and transaction by collected in the year 2015, and is expected to increase
taking full advantage of advanced information technolo- 20-fold in the next ten years. Data has been playing a cru-
gies [2]. Key technologies involved in smart manufac- cial role since the fourth industrial revolution initiated in
turing include but not limited to the Internet of Things Germany [4]. Data-driven decisions, on the other hand,
(IoT), cyber-physical system (CPS), cloud computing and distinguishes modern manufacturing from traditional
big data analytics. These technologies, when integrated ones in that decisions are made based on data of facts,
with manufacturing capabilities, initiated various para- not theoretical physical models, opinions and guesses.
digms belonging to the smart manufacturing family, such So far, researchers have been investigating in the area
of data-driven manufacturing for decades and have pub-
lished a great number of review articles about the latest
*Correspondence: [email protected]
1 achievements from different aspects. It all came about
Nanjing University of Aeronautics and Astronautics, Nanjing 210016,
China when large volumes of data were generated as an out-
Full list of author information is available at the end of the article come of digital manufacturing, along with data mining

© The Author(s) 2020. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and
the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material
in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material
is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the
permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat​iveco​
mmons​.org/licen​ses/by/4.0/.
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 2 of 21

techniques developed since the 1990s [5]. Later, upon performance of the manufacturing system. This closed-
wide acceptance of data-driven methods, process diagno- loop form of manufacturing inscribed a fundamental
sis techniques were adopted to automate fault detection paradigm of data-driven manufacturing, as opposed
in industrial processes [6]. Wuest et al. [7] gave a com- to conventional model-based manufacturing. As the
prehensive review of machine learning methods utilized development of various sensoring technologies and rel-
in manufacturing tasks. Since most manufacturing data evant infrastructures, modern manufacturing systems
are labeled data, supervised learning played a dominant are equipped with a large number of sensors captur-
role in practical applications. Zhong et al. [2] surveyed ing data at an unprecedented rate [12]. New challenges
associated topics in the context of Industry 4.0, including are thus raised: First, erroneous or patchy data can dis-
Internet of Things, cloud manufacturing, cyber-physical tort results and lead to faulty decisions [11]. Maintain-
systems, etc. based on which they provided detailed anal- ing the veracity of data with respect to the concerned
ysis on how these key constitutive technologies together target is challenging, because most data captured via
can realize Industry 4.0. When data became available generic sensors cannot directly reflect the actual on-
everywhere, data fusion techniques were also devel- site situation. Secondly, transforming these data into
oped to facilitate industrial prognosis [8]. Kong et al. [9] useful knowledge and decision is also challenging, as
reviewed the latest multisensor measurement and data the volume, variety and velocity of the captured data
fusion technology in precision monitoring systems. Tao are already beyond normal capacity [13]. Inaccurate
et al. [10] summarized the state-of-art development of methods to analyze only partial information from col-
new technologies through the lifecycle of manufactur- lected data can also mislead the final decision and
ing data, including data collection, storage, processing, performance.
visualization, etc. These technologies altogether initiated To deal with these challenges, researchers have been
smart manufacturing applications, such as smart design, centralized in this area and yielded rich outcomes. The
smart maintenance and quality control. main purpose of this paper is to summarize the devel-
As an increasing number of researchers started to opment and trend of data collection and analysis in
realize the importance of manufacturing data, data the era of data-driven manufacturing by conducting
collection and analysis have been broadly studied and a thorough review of the state-of-art. The rest of this
incorporated into modern manufacturing. Neverthe- paper is organized as follows: Section 2 will elaborate
less, few of the aforementioned review articles focused the framework of data-driven manufacturing and dem-
the evolution pattern of data collection and analysis onstrate some representative applications in different
towards modern manufacturing processes. As depicted aspects. The evolution of data collection and analyt-
in Figure 1, data collector, e.g., sensor is designated ics will be separately discussed in Sections 3 and 4.
to capture useful physical values generated by manu- Section 5 will summarize and give outlook of future
facturing event. The acquired data is further analyzed trend in advanced monitoring systems in modern
and interpreted into optimal decisions to enhance the manufacturing.

Data collector Data analyzer

Manufacturing
Data Decision
event

Figure 1 Closing the loop of data-driven manufacturing

Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 3 of 21

2 Concept and Key Components of Data‑Driven layer locates on top of manufacturing layer via sensor
Manufacturing interface. Various types of sensors are integrated into the
2.1 Data‑Driven Manufacturing manufacturing system to monitor and inspect during the
The major distinction between the paradigm of modern manufacturing process. In data layer, data is collected,
and conventional manufacturing can be viewed in Fig- stored and visualized for the preparation of data pro-
ure 2. Conventional manufacturing automation can be cessing. In knowledge layer, raw data transformed into
regarded as model-based manufacturing [14]. Experts insightful features and knowledge via data processing
gain experience by making physical observations, such technologies. In decision layer, through the utilization of
as visual inspection, noise recognition from manufac- intelligence, knowledge eventually becomes decisions to
turing systems. These experiences together with human make accurate simulation, evaluation and prediction, etc.
intelligence will derive physical models using theo- to facilitate smart manufacturing.
retical, experimental and numerical methods, to bet- The major distinctions between data-driven manufac-
ter understand the mechanism behind. Although great turing and conventional manufacturing are the genera-
achievements have been made and applied in various tion, collection and utilization of data, which have been
applications, such as simulation [15] and performance regarded as the key enabler to realize smart manufactur-
evaluation [16], these model-based methods are nev- ing [17]. As can be implied from Figure 3, data eventu-
ertheless inferior with their limited effective range and ally becomes decisions to automate the manufacturing
accuracy. This is mainly because a great deal of simpli- process and enhance its performance. In this manner,
fications and assumptions are made when deriving the accuracy of the decision predominates the manufac-
physical models, while the human experts are not assured turing outcome, e.g., a false decision could potentially
to be mentally stable and impartial towards all gained jeopardize the delicate product or even the manufactur-
experiences. ing system. It is conceivable that through all these layers,
Modern manufacturing, on the other hand, is data- accuracy and fidelity of the final decision will decrease
driven [10], in the sense that data generated through according to several reasons. In data layer, data acquisi-
manufacturing activities is fully utilized to positively tion may cause accuracy loss depending on the speci-
enhance manufacturing quality and thus enrich flexibil- fication of sensors adopted. The correlation between
ity and autonomy of the system. The framework of data- the acquired data and the actual physical value actually
driven manufacturing is outlined in Figure 3 consisting of involves certain assumptions/simplifications. In knowl-
four layers. The bottom layer is known as the manufac- edge layer, the extraction of knowledge from raw data
turing layer comprising different types of manufacturing further induces error since extracted features may not
processes, through which a product is designed, manu- perfectly define the overall profile of the original data.
factured, assembled and evaluated from scratch. Data In decision layer, improper data analysis could lead to

Observaon Sensor

Experience Manufacturing Data

system
Model-based Data-driven
manufacturing manufacturing
Human intelligence Machine intelligence

Physical Smart
model decision
Figure 2 Model-based manufacturing and data-driven manufacturing
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 4 of 21

Decision layer

Simulation Evaluation Prediction

Data analysis interface

Knowledge layer

Data processing Feature extraction Knowledge extraction

Data processing interface

Data layer

Data collection Data storage Data visualization

Sensor interface
Manufacturing layer

Manufacturing process
Figure 3 Framework of data-driven manufacturing

misunderstanding of the features, thus mislead the final reviews in social media [19]. Cloud-based CAx software
decision. In total, the aggregated error can be tremen- such as AutoCAD 360 was also invented to enable real-
dous, which poses great challenge to develop advanced time monitoring and collaboration among design teams
manufacturing systems. While detailed evolutional trend that are geographically apart. In addition to the cloud-
of data-driven manufacturing process will be demon- based infrastructure, high performance cloud computing
strated in Sections 3 and 4, we first analyze some typical and big data analytics have enabled expensive computa-
applications in Section 2.2. tions such as analysis of market preference and customer
demands at a reduced cost [20].
2.2 Typical Applications in Data‑Driven Manufacturing
2.2.1 Product Design 2.2.2 Logistics and Supply Chain Management
Product design is an iterative decision-making pro- Manufacturing supply chain refers to the flow of raw
cess to seek optimal solutions to the target customer materials from distributed original suppliers to manufac-
needs. The cost of product design can go up to 75% of turing sites, and finally to places of consumption. Trace-
the entire product cost, according to Li et al. [13], which ability of the supply chain is an important feature for
was mainly due to the constant trial-and-error itera- modern manufacturing enterprise to reduce logistic cost
tions during the product design phase until the custom- and increase its production efficiency in a long run. Aim-
ers get satisfied. It was particularly difficult for customers ing at a better supply chain visibility and tracking, radio
to monitor and dominate the designing process which frequency identification (RFID) and GPS work together
truly reflected their needs until the popularization of to provide a seamless and detailed trace of the product
rapid prototyping and Internet. As an additive manu- [21]. Supply chain analytics (SCA) has been extensively
facturing technology, rapid prototyping [18] revolution- investigated to assist decision makers in identifying and
ized the way how 3D product can be fabricated quickly assessing supply chain risks, and improving supply chain
from virtual design, offering the most intuitive feedback flexibility and capability. According to a latest review in
to the designers as well as to the customers. Internet this subject [22], analytic techniques in SCA include sta-
increases direct communications between the customers tistical analysis, simulation and optimization, which take
and the company, through which customers can directly full advantage of big data to analyze the supply chain per-
post their demands, share first-hand experience and formance and to make appropriate decisions.
even participate in a customized designing process of a
product. Regarding the state of the art, a new paradigm 2.2.3 Shop Floor Monitoring
named cloud-based design (CBD) was established to let Shop floor monitoring is essential to keep track of the
design engineers conduct market research more effec- running state of each machine, make adaptive scheduling
tively and efficiently through spreading feedbacks and and maintenance. Modern manufacturing shop floors are
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 5 of 21

equipped with smart sensors, among which RFID sen- fidelity to a whole new level. The other one is to excavate
sors are widely adopted to enable enormous data acquisi- valuable knowledge from low-value data using advanced
tion [23]. Data pre-treatment and analysis using machine machine intelligence. As mentioned earlier, data collec-
learning algorithms are then applied for shop floor tor and data analyzer are the key components to achieve
scheduling and fault prediction [24]. Real-time machine these two targets. Pertaining to the former, the acquired
availability and execution status monitoring is also an manufacturing data from advanced sensors is of unprec-
important issue to render distributed process schedul- edented fidelity and accuracy compared to the one from
ing in shop floor and cloud-based manufacturing. Wang legacy data collectors. On the other hand, data analyzers
[25] proposed a tiered system architecture with function utilizes the latest data processing and machine learning
blocks for monitoring the machine availability and execu- technologies to make better decisions than ever before.
tion status in real time, such that a closed-loop informa- In the remainder of this paper, a thorough investigation is
tion flow can be established. As the variety and volume made to review the state-of-the-art development of data
of data keeps increasing, the integrated data becomes too collection and data analysis towards data-driven manu-
intricate to handle and perceive. Cyber physical system facturing process, and to discuss the future trend of data-
(CPS) provides an ultimate solution to this issue by estab- driven smart manufacturing.
lishing a synchronized virtual shop floor to the actual one
[26]. In this way, a series of smart operations can be real- 3 Advanced Data Collection in Data‑Driven
ized, including smart interaction, smart control and man- Manufacturing Process
agement, etc. making the networked machines perform Modern manufacturing system is equipped with
more efficiently, responsively and collaboratively. advanced sensors collecting sequential data from differ-
ent physical events. These data are of low value density
2.2.4 Manufacturing Process if treated individually, but they together form great value
The importance of manufacturing process can never be for the system to keep track of the manufacturing pro-
overstated. In the last few decades, new types of manu- cess, in order to make simulations, evaluations and pre-
facturing process, such as high speed machining (HSM), dictions, etc. Therefore, high-quality data collection is a
additive manufacturing (AM) and hybrid manufacturing, desirable target in modern manufacturing by means of
have been rapidly emerged to satisfy growing demands of various types of sensors.
product. The complexity of modern manufacturing pro- As alluded earlier, the lifecycle of manufacturing data
cess has already gone beyond the level to be manually consists of data collection and data analysis. In data
observed and controlled. Monitoring the manufacturing collection stage, manufacturing data is generated and
process via dedicated monitoring system has become a collected from equipment, human operators and prod-
critical and essential target to avoid anomalies and reduce ucts. These data can be classified into structured, semi-
the maintenance cost. Towards this target, machining structured and unstructured data [32], depending on the
monitoring system became a research hotspot in recent selection of sensors and their working principles. In data
literature [27]. Machining monitoring system encom- analysis, the target is to extract informative knowledge/
passes signal acquisition, signal processing and decision decision from the raw data end to end. High dimensional
making steps, in order to identify tool conditions, chip raw data sometimes needs a prior feature extractor to
conditions, processing conditions [28] and part surface extract low dimensional representative features in either
conditions [29]. Monitoring system for additive manu- time domain or frequency domain [29], the extracted
facturing, especially for metal-based AM, has been fully features from different data sources are fused together
investigated to enhance the part quality and repeatability to make valuable decisions to control the manufacturing
in order to satisfy the stringent requirements from aero- process.
space and healthcare sectors [30]. Vision and camera- Figure 4 summarizes two typical workflow of data-
based monitoring systems are widely adopted for in-situ driven manufacturing process, which are based on direct
metrology inspection and closed-loop control of additive and indirect data measurement. In direct measurement,
manufacturing [31]. sensors are specifically designed to measure the physi-
As the development of advanced sensors and artificial cal value or its direct covariant during the process. These
intelligence, data-driven manufacturing process is also sensors are usually expensive and exclusive to certain
in its evolution. There has been two diverse trends in working environment. The captured data from direct
modern manufacturing process. The first one is to devise monitoring is of high fidelity and accuracy. For example,
smart sensors for direct measurement of those high-value the touch-trigger probes offer a direct way to precisely
data. These direct measurement approach can effectively measure the coordinates of the part by discrete physi-
bypass the tedious data processing stage and improve the cal contact points. Alternatively, indirect measurement
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 6 of 21

Direct measurement

Collector Analyzer
Data Decision
Target
Controller
Manufacturing value
Data collection Data analysis
system
Correlated Decision
value
Transfer Data
Collector
Data Analyzer
function

Indirect measurement
Figure 4 Workflow of direct and indirect measurement

offers a more cost-effective way to collect indirect but 3.1 Data Collection in Manufacturing Process
correlated value using generic sensors, such as cur- Extensive studies in manufacturing data collection have
rent sensor, accelerometer. The major difference from been conducted for various applications. From the per-
direct monitoring is that the captured physical value is spective of machining process, as depicted in Figure 5,
no longer the target one but a correlated value through people care mostly about the real time condition of the
a non-deterministic transfer function. For example, large process, the tool and the part. Applications include cut-
spindle current sometimes implies a large cutting force ting force monitoring, chatter detection, tool wear/break-
during metal cutting process, and sometimes only indi- age diagnosis, online inspection of surface roughness and
cates an accelerating spindle speed, which is hard to tell dimensional accuracy.
unless more information is provided. Building up the
exact inverse transfer function to deduce the target value 3.1.1 Cutting Force Monitoring
is impossible, which always involves simplifications and Cutting force monitoring is among the earliest achievable
assumptions, inevitably leading to accuracy loss, which capabilities in numerical controlled machining, for its
has become the major issue for indirect monitoring. high correlation with the in-process workpiece and tool
Abellan-Nebot and Subiron [33] gave a comprehen- status. Large cutting force is detrimental to the part accu-
sive review of machining monitoring systems developed racy as well as to the cutting effectiveness [35]. Initially,
so far, including sensors, signal processing and fea- the measurement of cutting force value was conducted
ture extraction. In their point of view, they argued that indirectly by using current signals of servo motors [36] or
indirect measurement was more prevalent for its cost- feed motors [37]. These methods are cheap and easy to
effectiveness and versatility. Lauro et al. [27] in their lat- implement, but with very limited upper bound in terms
est review suggested to take great care of the choice of of accuracy. Albrecht et al. [38] proposed an innovative
measurement due to implementation cost and require- indirect force measurement by integrating capacitance
ments. For tool condition monitoring, such as tool wear displacement sensor into the spindle. The sensor was
diagnosis, direct methods, e.g., optical and radioactive capable of measuring deflection of the tool and finally
sensors were deprecated due to the inaccessibility of the converted to the value of force. At certain frequency
cutting area during the cutting process [34]. However, (650 Hz), the sensor reliably measured cutting force
direct vision/camera-based systems were widely used for
monitoring in-situ metrology for additive manufactur-
ing process [31], in order to achieve a close-loop identi-
fication of material discontinuities and failure modes. As
manufacturing processes are becoming more and more Tool condition
complicated, generic sensors may not satisfy the increas- Process condition
ing demands of high accuracy because of the inevitable Part condition
simplifications and assumptions between the target and
the captured value. The following section will demon-
strate the evolutional trends for various monitoring tasks
in manufacturing process. Figure 5 A typical machining process
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 7 of 21

with around 10% error in magnitude. The major draw- and accelerometer [47] have been demonstrated as effi-
back of this indirect sensor was its limited bandwidth, cient and effective solutions for chatter recognition. They
which, even after applying a Kalman filter, can only reach however suffer from a common drawback that the ambi-
1000 Hz. These indirect methods were either of low accu- ent sound/vibration could introduce noise to the target
racy or with limited frequency bandwidth. Direct force signal. Especially for the microphone, the suppression of
sensors were developed which was equipped with sensing environmental noise is mandatory to make it truly appli-
elements to convert external force load into deformation cable. Later on, indirect methods came out focusing on
of the elastic element. Piezoelectric transducer and strain the correlated effect of chattering and utilized relevant
gage are two major branches in modern cutting force signals for chatter detection, such as using cutting force
dynamometers. Strain gauges force transducers offer signal [48], motor current signal [49], acoustic emission
high frequency response and long-term stability of defor- [50] and the fusion of multiple signals [51]. The correla-
mation under an external force. Yaldiz et al. [39] devel- tion between these signals and the chatter occurrence
oped a table dynamometer using strain gauge to measure needs to be meticulously analyzed to achieve feasible
static and dynamic milling forces. An octagonal ring was results. The accuracy of these indirect measurements
manufactured to locate the strain gauges, whose orienta- has been greatly enhanced after adopting machine learn-
tions and locations were carefully determined to maxi- ing algorithms, such as in [49], using a support vector
mize the overall sensitivity. After calibration, the final machine to recognize the chatter pattern based on servo
accuracy can reach up to 98.5% in real milling process. motor current signal can reach over 95% in terms of
Piezoelectric sensors, as compared to strain gage, are accuracy rate. Nevertheless, the frequency bandwidth of
superior for dynamic force measurement for their high these generic indirect sensors may not suffice the detec-
dynamic range and sensitivity [40], which were usually tion of chatter, especially in high-speed machining. Con-
mounted on the spindle side for dynamic force measure- sequently, direct measurements using microphone has
ment. As for the state-of-the-art, advanced measurement been revived after the reliability of sensor was improved
apparatus were developed to measure micro-cutting in monitoring milling operations. Specifically, the micro-
force in wireless manner, and with higher accuracy up to phone response inside the machine-tool chamber was
99.8% [41]. Polyvinylidene fluoride (PVDF) sensors were sensibly corrected using equalization filters to ensure
embedded in each inserts of the cutter to estimate real- adequate accuracy in chatter detection task [52]. Optical
time working condition for separate insert in a wireless measurement such as using a laser beam and an optical
manner [42]. Some recent studies tried to develop a so- position detector (OPD) to identify the vibration of the
called smart tool with built-in piezoelectric sensor array in-process tool was also regarded as a direct method
[43, 44], which could be a future trend towards smart [53]. In this study, the laser beam was reflected on the
manufacturing process. Table 1 lists the evolution of cut- rotating cutter and captured by the OPD, by which the
ting force measurement. displacement of the cutter can be recorded in real time.
The development of high-resolution vision system also
3.1.2 Machining Chatter Inspection facilitated the online measurement of chatter by analyz-
Machining chatter has always been an important issue in ing the surface texture/marks in real time [54]. Ding et al.
manufacturing process, for its complex physical mecha- [55] invented an active control system to detect and sup-
nism and negative effects leading to poor surface finish, press machining chatter. Chatter was detected by directly
tool damage, etc. [45]. Real time chatter monitoring has sensing the workpiece displacement using a displace-
also been classified into indirect and direct methods. ment sensor and then controlled via a voice coil motor.
As the outcome of chatter is usually in the form of self- In terms of offline chatter identification, chatter stability
excited vibration, direct methods using microphone [46] diagram offers a scientific reference for a proper choice
of chatter-free machining parameters [56], the genera-
tion of which relies hugely on the frequency response
Table 1 Evolution of cutting force measurement functions (FRF) at the tool tip. Accelerometers are widely
adopted for FRF measurement [57] based on standard
Year Indirect measurement Direct measurement
impact test using a hammer integrated with force sensor.
Before 2000 Current sensor [36, 37] The impact test is nominated as a direct measurement
2001–2010 Displacement sensor [38] Strain gauge [39] for FRF determination but requires tedious setting-ups
Piezoelectric sensor [40] for pose-dependent tool tip dynamics in bi-rotary milling
2011–2020 Smart tooling [43, 44] head [58]. Table 2 lists the evolution of machining chatter
Wireless cutting force sensor
[41] inspection.
Embedded PVDF [42]
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 8 of 21

Table 2 Evolution of machining chatter inspection

Year Indirect chatter inspection Direct chatter inspection

Before 2000 Acoustic emission sensor [50] Microphone [46]

Accelerometer [47]
2001–2010 Dynamometer [48] Optical sensor [53]
Sensor fusion [51]
2011–2020 Current sensor [49] Integrated microphone transducer [52]
Vision-based system [54]
Displacement sensor [55]

3.1.3 Tool Condition Monitoring recent studies, among which, Ramirez-Nunez et al. [73]
Tool condition monitoring (TCM) is vital to keep track of came up with a smart sensor consisting of an infrared
the remaining useful life (RUL) of the tool. Late replace- camera and a temperature sensor, which facilitates the
ment of dull or broken tool may decrease the accuracy in-process tool breakage inspection even with the exist-
and quality of the final part and cause machine break- ence of coolant fluid. The tool condition is well estimated
down [34]. Though with much effort spent in the past by processing the infrared thermography. Dai and Zhu
[59], direct inspection of the in-process tool condition [74] in their recent study proposed an integrated vision
was developed in the first place which includes the usage system for micro-milling TCM. The system was designed
of proximity sensors, radioactive sensors and vision sen- with a telecentric lens, light source and a 3-DOF motion
sors. Proximity sensors, such as ultrasonic sensor [60] platform to achieve uniform image quality and high auto-
estimate the differential of distance between cutting edge mation. As the availability of powerful image processing
and workpiece, whose accuracy is highly affected by the algorithms, direct TCM using smart sensors and inte-
thermal expansion and cutting force induced deflection. grated systems is believed to have a promising future.
Radioactive sensors [61] detect the amount of residue Table 3 lists the evolution of tool condition monitoring.
radioactive materials implanted on the flank face of the
cutting tool in order to estimate the wear percentage, 3.1.4 Part Condition Monitoring
which was regarded detrimental and thus limited for lab The condition of in-process part (a.k.a. workpiece) needs
usage. Vision-based tool condition monitoring, especially to be monitored to take timely adjustment of the pro-
using structured light [62] was also patented long ago, cess, in order to yield high-quality part. Surface finish
but required an ideal condition of lighting and cutting and dimensional accuracy are the two dominant factors
environment to achieve acceptable accuracy. of the workpiece condition to determine the final qual-
Deficiencies of these early direct TCM methods lead ity of product. Especially for the surface finish metrology,
to the prosperity of indirect TCM methods, which uti- which has been overwhelmingly concerned as a direct
lized correlated signals, such as cutting force [63], acous- indicator to the capability of modern manufacturing sys-
tic emission [64], vibration [65], current [66–68] and tem. Conventional surface inspection methods [75] are
surface roughness [69]. These representative indirect usually conducted subsequently to the manufacturing
methods advanced the development of signal processing process. These post-processing based methods can usu-
and sensor fusion techniques to enhance the prediction ally achieve higher accuracy using dedicated instruments,
accuracy. Though many review articles highly voted for such as the stylus profilometer [76], but are inactive to
the indirect TCM methods [29, 34] as the future trend take responsive actions to prevent further accuracy loss
due to the increasing accuracy, the major drawback is
still prominent in that these methods are case-sensitive
and requires fine-tuning and calibration to achieve high Table 3 Evolution of tool condition monitoring
authenticity. Direct TCM methods, especially for the Year Indirect monitoring Direct monitoring
vision-based, can bypass this issue by directly inspect-
ing the geometric change of the tool. Two-dimensional Before 2000 Accelerometer [65] Radioactive sensor [61]
Proximity sensor [60]
[70] and three-dimensional vision systems [71, 72] were Structured light [62]
developed for direct TCM and achieved sub-pixel accu- 2001–2010 Current sensor [66–68] 3D metrology [72]
racy using advanced image processing techniques. These 2011–2020 AE sensor [64] 2D vision [70]
vision-based TCM systems all require a pause between Surface roughness inspector 3D vision [71]
two sequential operations to capture a steady image of [69] Infrared thermography [73]
Dynamometer [63] Integrated vision system [74]
the tool. This inconvenience has been fully addressed in
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 9 of 21

during the process. Therefore, quality control based However, for realistic complex parts, the distribution of
on in-situ monitoring offers a more practical solution. residual stress can be elusive especially when the work-
However, it is intractable to incorporate sophisticated piece profile is constantly changing during the process. In
roughness scanners into the harsh operating environ- light of this concern, direct measurement would be a bet-
ment with metal chips, lubricants and vibrations. Con- ter choice.
sequently, indirect methods for the inspection of in-situ To directly measure the deflection, the on-machine
surface roughness took over the mainstream in the past measurement system using a touch-trigger stylus was
[77], which contain the usage of accelerometer [78], adopted to inspect the workpiece deformation, and adap-
dynamometer [79], acoustic sensor [80], ultrasonic sen- tively change the subsequent tool path for compensation
sor [81], etc. Prevalent shortcoming of these indirect [89]. The utilization of inspecting stylus was a prevalent
methods is lower achievable accuracy due to the nature choice for online measurement, it yet required the sus-
of uncertainty. Currently, vision-based surface rough- pension of the manufacturing process, which prolongs
ness evaluation system [82] has been developed for effi- the overall processing time and is technically incapable of
cient and accurate in-situ surface inspection. The essence real-time monitoring. More advanced instruments were
behind was the usage of graph theory-based image developed recently to address these issues. Luo et al. [90]
analysis to achieve real-time identification of surface devised a thin film PVDF sensor attached to the non-
roughness distribution without interrupting the machin- machining side of the thin-walled part to monitor the
ing process. For metal additive manufacturing process deflection and vibration caused by machining force. The
where quality matters, vision-based systems are also the change of output voltage faithfully reflected the high-fre-
primary choice for in-situ metrology monitoring [31]. quency deflection of workpiece during different machin-
It is also suggested to apply hybrid instrumentation as a ing stages. Real-time surface normal measurement for
future direction to overcome the compromise between maintaining high accuracy of thickness is indispensable
spatial resolution and the field of view, in which low reso- in machining freeform thin-walled part. Yuan et al. [91]
lution sensor detects the whole area while high resolu- established an online surface normal measurement using
tion sensor focuses on the area of interest. four eddy current displacement sensors installed in the
Real-time dimensional accuracy monitoring is vital to frontend of the spindle, achieving a remarkable reduction
render in-process quality control. Dimensional accuracy in displacement errors (from 12% to 1%) after compen-
is prone to be violated for parts consisting of thin-wall sation. A more intractable case of deformation is caused
features, due to either the deflection by external cutting by the release of residual stress during the removal of
force or the internal release of residual stress. As the in- raw material, such deformation remains obscure as long
process part is usually securely mounted by fixture and as the workpiece is securely fixed on the machine table.
hard to access by exotic instruments, integration of fix- Indirect prediction model of the residual stress distribu-
tures with sensing technology will be a potential direc- tion [92] is too complicated to be accurate, due to a large
tion according to the state-of-art review [83]. In terms of set of remaining uncertainties. In light of this issue, Li
the deflection caused by external force, Azouzi and Guil- et al. [93] inaugurated a novel responsive fixture appa-
lot [84] predicted the workpiece dimensional deviation ratus for direct inspection of in-process deformation of
in turning process via cutting feed, depth of cut and cut- large aerospace parts. This smart fixture automatically
ting force signal. Cutting force and vibration signal were opens up to release the deformation once the built-in
fused together for the prediction of deviation in turning stress sensor reaches its threshold. In this way, adaptive
a slender part [85]. For thin-walled part, such as a blisk, adjustment of the process can be made as long as the
a common method to identify its in-situ deflection is by final shape is still enveloped by the remaining workpiece.
simulation using cutting force value and modeling tech- Inspired by this idea, Hohring and Wiederkehr [94] fol-
niques [86], which is not only time-consuming but also lowed up with a similar intelligent fixture for the purpose
uncertain in terms of accuracy. As for the deflection error of mitigation of chatter and compensation of work-
caused by the release of residual stress, it was particu- piece distortion to achieve high performance machin-
larly tricky to predict such error since each piece of raw ing. Table 4 lists the evolution of workpiece condition
stock has its own stress pattern. Instead, people strived monitoring.
to characterize the residual stress field distribution via
nondestructive methods, such as using ultrasonic devices 3.2 Discussion and Future Trend
[87] and X-ray diffraction [88]. The key to these nonde- Data collection and analysis are two essential stages
structive methods is to formulate the stress gradient with in data-driven manufacturing process. Depending on
respect to the center frequencies, which can achieve the correlation of captured and target value, manu-
plausible accuracy in workpiece with simple geometries. facturing data can be collected via direct and indirect
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 10 of 21

Table 4 Evolution of workpiece condition monitoring In the first paradigm of data analysis, a physical model
Year Indirect accuracy Direct accuracy monitoring
describing the mechanism is developed by the expert.
monitoring Once the input data and information is imported into
the physical model as prior knowledge, a mathematical
Before 2000 Ultrasonic sensor [81]
AE sensor [80]
solver is established to find the optimal solution, i.e., the
Dynamometer [84] decision. For example, finite element analysis is a typi-
2001–2010 Dynamometer [79] cal data analysis which employs linear solver to solve a
Accelerometer [78, 85] partial differential equation, e.g., deformation of the
2011–2020 Simulation [86] Vision based system [82, 31] part. Obviously, two simplifications are involved in this
Ultrasonic sensor [87] Touch-trigger stylus [89]
X-ray diffraction [88] Thin film sensor [90]
pipeline. First, the physical models developed by human
Responsive fixture [93, 94] experts are usually based on certain assumption and
simplification which deviates from actual scenario. Sec-
ondly, solving the physical model with limited input data
is sometimes ill-posed, which can possibly lead to faulty
measurement. Although indirect measurement offer
results. However, it should not be denied that when data
more possibilities and larger scalability in diversified
is scarce and expensive to acquire, this paradigm effec-
applications and are more cost-effective, they usually
tively offers a plausible way to interpret the process.
require the establishment of a physical transfer function
The second paradigm of data analysis utilizes machine
to indicate the correlation between the measured and
learning techniques to train a shallow encoders which
target value, which inevitably induce error as long as such
consist exponentially greater number of unknown
correlation contains physical uncertainty. Consequently,
parameters than the physical model. Through sufficient
the accuracy of indirect measurement is undermined
training stage using paired feature-result set, the trained
so long as the correlation is not rigorously and math-
model is capable of producing sensible answers on new
ematically identified. This gap encouraged more studies
input features. Due to the high generalization ability of
to improve the accuracy by developing various sensor
machine learning models, it successfully bypasses the
fusion and data analysis methods [95]. On a different
model simplification encountered in the previous para-
perspective, direct measurement using dedicated sen-
digm. Nevertheless, the ability for a shallow encoder to
sors can achieve high fidelity and accuracy. Although it
directly process high-dimensional raw data is still limited,
seems to be contradictive to the big data scenario where
it thus requires careful feature engineering and consider-
obtained data is usually trivial and individually inaccu-
able domain knowledge to reduce the input dimension.
rate, the design of exclusive sensor is still one important
As the density and dimension of manufacturing raw
trend in the manufacturing field to facilitate accurate
data is experiencing a rapid growth, the key factor to the
process monitoring, and thus to make precise decision
final accuracy is how the data is processed in the first
and control. This pictures one possible future of intelli-
place. Motivated by this need, the third paradigm using
gent manufacturing.
deep learning can potentially eradicate the error-prone
handicraft of feature extraction, which instead is achieved
4 Advanced Data Analysis in Data‑Driven automatically using a general learning procedure. In this
Manufacturing Process way, feature extraction induced error can be reduced to
Manufacturing process is decisive to the whole product a great extent. It is expected that this paradigm will give
life cycle. As elaborated in the previous section, vari- the best performance on data analysis as long as the deep
ous sensors are being devised and integrated onto the model is fed with sufficient data.
machine to enable in-process monitoring by captur- In the following sections, we will first provide a com-
ing target or correlated values. Nevertheless, the data prehensive review on existing methods for feature extrac-
acquired by these sensors, no matter directly or indi- tion from manufacturing raw data, given that feature
rectly, only gives partial view of the manufacturing pro- extraction is an essential stage for the first two para-
cess. These data still needs post analysis to be converted digms. Some typical manufacturing applications using
into perceptible knowledge and decisions. Making deci- data analysis will then be elaborated according to the
sions from data rather than human knowledge has above three paradigms.
become the dominant trend in data-driven manufactur-
ing. In data analysis, we believe there have been at least 4.1 Pre‑processing of Manufacturing Data
three paradigms so far, as depicted in Figure 6, which Manufacturing raw data can be regarded as a sequential
also illustrates the evolution of modern manufacturing of digital bits if not further processed. Data processing is
process. an essential stage especially for indirect data to convert
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 11 of 21

Expert

Input Physical Mathematic

Decision
a data model al solver
Shallow neural network

Training

Feature engineering Machine learning

Input Feature Decision

b data
Deep neural network

Training

Deep learning
Input
c Decision
data
Figure 6 Three paradigms of data analysis based on (a) physical modeling; (b) machine learning; (c) deep learning

them into perceptible values. In general, data processing hard to distinguish in time domain. Altintas et al. [100]
methods are specifically designed to extract useful fea- analyzed the cutting force and chatter stability during
tures and can be categorized into time domain, frequency dynamic cutting process using Nyquist law in frequency
domain, time-frequency domain and statistical process- domain. The analysis of tool vibrations using fast Fourier
ing [34]. Specifically, time domain data processing refers transform (FFT) was proved an effective mean for the
to direct feature extraction of the time series data, such as prediction of surface roughness [101]. Frequency spectra
the mean, peak and root mean square (RMS) value [33]. of the AE signal was identified to evaluate the tool condi-
Song et al. [96] characterized the vibration time series tion in broaching process [102]. By analyzing the motor
using autoregressive moving average (ARMA) model and current in frequency domain, the sensorless automated
discovered a linear relationship between the AR param- condition monitoring was achieved for predictive main-
eter and the surface roughness. Campatelli and Scippa tenance of machine tool [103]. FFT was also utilized to
[97] predicted the cutting force coefficients by analyz- filter out noise from the audible energy sound to achieve
ing the time domain behavior of the cutting force signal. better monitoring performance [104].
Ertekin et al. [98] calculated the RMS value of the AEDC The FFT gives the entire frequency spectrum with the
signal which was observed as the most sensitive feature average frequency composition. Practically, the sensory
to the tool wear. The average RMS feature of the current data is dynamically changing over time. Therefore, a
signal also contributes the estimation of tool wear [99]. time-frequency data processing gives a more reasonable
Frequency domain data processing can extract more outcome by partitioning the time series data into short
intrinsic features from a cyclic data series, especially time intervals for frequency analysis [29]. Specifically,
when such data contains background noise which is wavelet analysis and short time Fourier transform (STFT)
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 12 of 21

are the two prevalent techniques to analyze cutting force by using a dedicated tribometer, which consists of cut-
[105], vibration [106], AE [107, 108], current [109] and ting and thermal simulations. However, the formulation
sound signal [51]. of all these factors is subject to certain simplifications
Statistical data processing offers a better way to indi- and assumptions, and calibrating the pending coefficients
cate short term impulses and analyze variance between of the model using limited testing data would introduce
different factors. In particular, Aouici et al. [110] utilized more statistical errors, which together make the physical
a statistical analysis of variance (ANOVA) to predict the model inaccurate and unstable towards real complicated
surface roughness during hard turning. Similar approach machining process.
was found in Ref. [111] using vibration signal. Kannatey- Instead of formulating complex and error-prone physi-
Asibu and Dornfeld found that the skew and kurtosis of cal models for tool wear mechanism, most research-
the AE signal was sensitive to the tool wear [112]. Lu and ers intended to estimate tool wear in a more statistical
Wan [113] studied the high-frequency sound signal for manner, i.e., to estimate remaining useful life by fitting
tool wear monitoring using class mean scattering criteria. historical data into an empirical model. Endeavors to
Table 5 lists the processing methods for different manu- estimate the tool life can be traced back to the early 20th
facturing signals. century when FW Taylor [116] proposed the well-known
No matter what data processing strategy is utilized, it Taylor equation, which is an empirical model with two
is a primary stage in the entire data analysis for the fol- unknowns. Ever since then, various empirical models
lowing two benefits. First, raw sensory data is usually of [117–119] and experimental studies [120, 121] were pre-
high dimension and contains stochastic noise, data pro- sented targeting at different tool-workpiece combina-
cessing can tremendously reduce the dimension and filter tions. A comprehensive list of variant tool wear empirical
out disturbance without losing much valuable informa- models for dry machining can be found in Ref. [122].
tion. On the other hand, the extracted low-dimensional Their procedures were in similar fashion: first a nonlin-
features are more comprehensible in terms of develop- ear formula describing the tool condition based on the
ing analytical algorithms to make decisions accordingly, observer’s domain expertise was established ahead of
which will be discussed in the following section. time, then factorial design of physical experiments were
conducted to calibrate the unknowns of the formula,
4.2 Data Analysis in Manufacturing Process experimental validations were eventually conducted to
4.2.1 Tool Condition Analysis prove the feasibility. Although the prediction accuracy
Analyzing in-process tool condition from limited data is reported in these works can reach as high as 95% in
an important issue through which manufacturing process their experimental setups, it is perceived that any slight
can be more precise and efficient. Tool wear is the most change of the actual cutting condition would devastate
phenomenal condition people cared. Identifying the tool the accuracy. As the demand for accuracy and the com-
wear mechanism is intractable as it involves physical and plexity of manufacturing process keep growing rapidly,
chemical process, such as abrasion, adhesion, diffusion physical and empirical models have been widely depre-
and other types of wear during cutting process. A few cated. Zhao et al. [123] argued that this was mainly due
pioneered studies strived to understand the wear mecha- to the following reasons: First, the performance of these
nism as the addition of brittle fracture, mechanical abra- models was highly dependent on the domain expertise of
sion, physicochemical mechanism and others [114]. A the observer, whose robustness was unsecured due to the
recent study [115] developed a fundamental wear model uncertainty and complexity of working conditions. Sec-
ondly, these models were unable to evolve along with the
accumulation of data, and thus insensitive to the chang-
ing conditions, which lead to limited effectiveness and
Table 5 Processing methods for different manufacturing flexibility in real cases. These two deficiencies of model-
signals based approach would introduce considerable amount of
Signal Time domain Frequency Time- Statistical error, not to mention the error from the feature extrac-
domain frequency analysis tor, which together makes the physical model-based data
domain
analysis hardly compatible with wider applications.
Cutting force [97] [100] [105] [110] The advance of volume and veracity of data makes it
Vibration [96] [101] [106] [111] possible to adopt various machine learning algorithms to
Acoustic emis- [98] [102] [107, 108] [112] predict tool condition more accurately. Prevalent choices
sion of machine learning techniques for tool condition analy-
Current [99] [103] [109] sis include support vector machine (SVM), artificial neu-
Audible sound [104] [51] [113] ral networks (ANN), Hidden Markov models (HMM)
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 13 of 21

and decision tree. With sufficient training data, trained accuracy. Another recent study [135] was to monitor
ANN model using back propagation can be compara- the tool wear level based on audio signal using CNN,
tively accurate for tool wear estimation [124]. Palanisamy which strived to eradicate the need of feature extraction
[125] compared ANN against classic regression model by using the absolute values of Fourier transformation
in terms of the capability in tool wear estimation, ANN as input. As a result, the tool wear prediction accuracy
was found to be more robust and accurate for its pow- reached to as high as 96.3%. A Convolutional Bi-direc-
erful fitting ability. As a statistical learning approach, tional Long Short-Term Memory networks (CBLSTM)
SVM is superior for non-linear classification of data by was designed in Ref. [123] to eliminate feature engineer-
mapping them into higher dimensional feature space, ing in tool health monitoring. In this network, CNN
by which discretized state of tool wear can be classified. was served as local feature extractor, while LSTM was
Support vector regression (SVR) is a variant of SVM for to address sequences of varying length data and capture
continuous regression of tool wear value. Tool breakage long-term dependencies, in that tool wear was a time-
detection [126] and tool wear estimation [127] were suc- variant sequential progress.
cessfully carried out via SVM/SVR with over 99% success The extrusive challenge for the adoption of deep learn-
rate when the design parameters of the SVM model was ing to make accurate analysis is the demand of large vol-
fine-tuned. It was also noticed in a more recent study ume of labeled data, the acquisition of which is extremely
that a hybrid estimator combining analytic fuzzy classi- costly and time-consuming for many manufacturing
fier (AFC) and SVM can reach higher accuracy in tool applications. For example, the identification of tool tip
wear estimation [128]. Other learning techniques, such dynamics for a newly inserted tool needs hundreds of
as decision tree classifier [129] and HMM [130], were impact tests at different tool postures. In this situation,
also applied in application of tool wear estimation and the utilization of historical data to facilitate the training
achieved plausible performance. It was however stated of a new case becomes a potential and appealing solution.
in Ref. [129] that the performance of decision tree clas- Chen et al. [136] proposed a transfer learning-based pre-
sifier combined with a PCA was case-sensitive. It is diction for pose-dependent tool tip dynamics in five-axis
noticed that there has always been a hidden trade-off machine, by which the number of required impact tests
issue between the complexity of learning model and the is highly reduced. Sun et al. [137] utilized deep trans-
training cost. To achieve high accuracy, a more complex fer learning to predict tool life, by taking advantage of
learning model would thus require a larger training data the learnt similar characteristic across different objects.
set and heavier computational load, otherwise overfitting A recent study on tool wear prediction based on meta-
issue would lower the performance. learning was proposed by Li et al. [138]. Meta-learning
Most of the aforementioned tool condition analysis is has the ability of learning the hidden rules behind a vari-
majorly dependent on time series data such as cutting ety of different but similar tasks/models. The adoption of
force and vibration. Shallow function approximators meta-learning in this study successfully predicts the tool
like ANN and SVM are technically incapable of deal- wear status in changing cutting conditions with enhanced
ing with such high-dimensional data and thus require a accuracy, while only a few training samples are needed
dedicated feature extractor beforehand [131], as already upon a new learning task. This meta-learning approach
elaborated in Section 4.1 and illustrated in Figure 6(b). provides a new perspective to solve manufacturing prob-
Conceivably, the quality of the extracted features directly lems where the acquisition of data samples are expensive
affect the accuracy of subsequent operations. Improper and time-consuming. Table 6 lists the evolution of tool
choice of feature extractor may fundamentally suppress condition analysis.
the eventual performance. Therefore, it would be better
if one can directly handle the raw data series and bypass 4.2.2 Process Condition Analysis
the feature extraction stage. The development of deep Process condition analysis is a typical classification task.
neural networks such as convolutional neural network In machining process, the condition can be categorized
(CNN) [132] and long short-term memory (LSTM) net-
work [133] can fully satisfy this requirement. Specifically
Table 6 Evolution of tool condition analysis
in the application of tool condition analysis, Li et al. [134]
adopted CNN to detect tool breakage by spindle current Physical model Machine learning
signal, which achieved higher accuracy (93%) than that Simulation Empirical Shallow Deep Learning
of the traditional BP neural network (around 80%). How- model learning learning from fewer
ever, time-domain feature extraction was still adopted data
in this work, CNN was thus only regarded as a tradi- [115] [116–119, [124–130] [123, 134, [136–138]
tional machine learning methods with higher achievable 122] 135]
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 14 of 21

into idling, stable and chatter state. Timely and precise machining process condition analysis. Among existing
identification of process condition is always desired to deep learning algorithms, CNN is known for its powerful
make adaptive adjustment of process control. Previous image (second order tensor) processing and classification
studies made some important progress in identifying the capability. However, most captured data from machin-
mechanism of cutting process. Budak and Altintas [139] ing process is in the form of first order tensor (time
explored the mechanism of chatter during milling pro- sequence), which is not practical to be processed via
cess and came up with a physical model to identify the CNN. Fu et al. [144] innovatively transformed measured
chatter stability induced by the dynamic milling forces. signals into plotted image and employed convolutional
According to this study, the cutter is simplified as a two neural network to achieve real-time identification of cut-
degree-of-freedom system subject to a dynamic radial ting vibration state. This work realized directly use of
force, based on which the theoretical chatter stability the original signal sequence for cutting state monitoring
lobe was derived. On the other hand, the calculation of with significant performance of over 99.5% accuracy in
dynamic cutting force is also simplified using numerical most testing cases. Deep Belief Network (DBN) has been
method. This plausible offline solution may not prac- majorly dealing with voice and speech recognition [145].
tically satisfy real machining cases [45], as it requires a The in-process vibration signal is similar to the voice. Fu
complete analysis of machine dynamics including the et al. [146] got inspired by this and came up with a DBN
spindle, tool holder, tool and the workpiece, which is not approach for cutting state monitoring. It turned out that
only intractable to precisely identify but also requires DBN can steadily achieve high performance on the raw
tedious calibration works for different process conditions. vibration signal without much data preparation.
The simplifications and unpredictable systematic bias Since data is relatively convenient to acquire during the
further reduced the accuracy in offline analysis. Although manufacturing process, most deep learning approaches
researchers carried forward this theory to adapt to more can already achieve very promising accuracy in their
complex situations, e.g., five-axis machining [140], they case studies. Still, conditions can be quite different in real
were still of limited usage since the fundamental gap was machining situation where various materials, tools and
not completely filled. parameters are combined in each individual task. Trans-
When it comes to online identification of process fer learning has been attracting more attention to deal
condition, the preferred option is to make diagnosis as with varying conditions [147] and proved to be effective
early as possible, in order to prevent workpiece damage for chatter detection with accuracy up to 95%. This new
ahead of time. Traditional estimation algorithm, such as learning technology will not only reduce the data needed
maximum likelihood [141] though achieved great suc- for training a deep model, but also increase model ver-
cessful rate, but lacked the ability for early prediction. satility to adapt to complex manufacturing process sce-
The main reason is that subtle features are prone to be nario. Table 7 lists the evolution of process condition
overlooked before they become phenomenal. Machine analysis.
learning methods have been employed in this task for
the superiority in classification, especially in those hard- 4.2.3 Part Condition Analysis
to-recognize scenario. In particular, acceleration signals The well-being of in-process part directly affects the
were analyzed based on wavelet transform and SVM, this quality of final product. Surface roughness [77] and part
combination was able to detect transition state between dimensional error [148] are the two most concerned
stable and chatter state, showing excellent performance aspects, since they respectively reflect the manufactur-
with over 95% accuracy rate [142]. In this way, chatter ing quality in microscopic and macroscopic view. For the
could be firmly suppressed in its infancy stage. Later on, formal one, physical models and experimental data based
neural network approaches were also developed for pro- regression are the two mainstream solutions people uti-
cess condition classification using vibratory signal [143]. lized to understand the surface roughness mechanism.
In addition to the feature generation which is mandatory Lin and Chang [149] established a surface topogra-
for traditional machine learning approaches, this work phy simulation model incorporating the effects of tool
introduced a feature selection strategy based on envelope
analysis to rank the features according to their entropy,
and only those high-ranking features were selected for Table 7 Evolution of process condition analysis
classification. This operation essentially reduced the Physical model Machine learning
error from irrelevant features and hence increased the
final accuracy. Analytical Shallow Deep learning Learning
model learning from fewer data
To further reduce the error induced by feature extrac-
tion, deep learning methods were also utilized for [139, 140] [142, 143] [144–146] [147]
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 15 of 21

geometry, cutting parameters and tool motions to simu- When it comes to dimensional error prediction, online
late the surface finish profile during turning operation. prediction and real time compensation has always been
Kim and Chu [150] determined the surface roughness a preferable choice. Li et al. [158] developed a soft-touch
by proposing a geometrical model to calculate the maxi- sensor which provides proximity information when the
mum height of the effective scallop. This model was par- tool is approaching the workpiece, and a neuro-fuzzy
ticularly complicated as it considered the cutter runout network for predicting machining errors. This hybrid
effect and cutter marks. Others conducted experimental learning system succeeded in precise prediction of the
studies to unveil the relationship between tool life, sur- aggregate sum of thermal error, force-induced deflection
face roughness and vibration [101]. Regression analysis error and other source error in turning process. Another
was adopted to handle the experimental data. dimensional error prediction in milling process was
The dimensional error of in-process part can be catego- achieved using ANN [159]. In this work, data set of pro-
rized into plastic deformation caused by residual stress cess parameters that can affect dimensional errors was
and elastic deformation caused by large cutting load. yielded via experiments. The large number of influencing
Finite element method (FEM) was a primary choice for parameters led to the choice of ANN, which generated
the evaluation of these two types of deformation, due to more accurate models than the previous empirical mod-
the large uncertainty of part shape and stress distribution els after training process.
during the process. The distortion of thin-walled work- Conventional machine learning approaches suffice the
piece induced by machining residual force was predicted demand for real-time prediction of surface roughness
using a modified finite element model [151]. The combi- and part deformation. A foreseeable trend in this sec-
nation of experimental results with FEM was proposed to tion would be more precise identification of part condi-
predict the shape deviation of complex geometry [152]. tions, such as the types of defect and crack, by further
Elastic deflection also induces machining error, especially exploiting the advanced vision-based sensors. Towards
for thin-walled part. Wan et al. [153] estimated the cutter this goal, traditional shallow learning approaches require
deflection using a simple cantilever beam model, and the artificially defined feature descriptors from the captured
workpiece deflection using FEM simulation. The induced raw pixels, while deep networks are able to directly pro-
error was compensated accordingly [154]. cess raw data. In particular, CNN serves as a primary
Both analytical model and FEM have to make a great choice for surface inspection task. A max-pooling CNN
deal of simplifications since accurate prediction of sur- was developed in Ref. [160] to identify steel defect with
face roughness and part deformation require tedious an error rate of 7%, which outperformed the best trained
trial-and-error process and excessive computing power. classifier using artificial feature descriptors (15%). Part
Targeting at online analysis, trade-off between model et al. [161] showed that using CNN can achieve 250 times
complexity and its performance has always been a puz- faster inspecting speed compared to manpower inspec-
zling task. In light of this issue, machine learning algo- tion, without sacrificing the accuracy. Ren et al. [162]
rithms started to take over online quality analysis with proposed a CNN based feature extractor for pixel-wise
higher performance. As for the surface roughness predic- surface inspection, which did not require large-scale
tion, although people spent great effort investigating its training data using pretrained model. The heat map
mechanism, it however varied with different processes showing distribution of defects was then generated for
and conditions. Any sophisticated physical models will the identification of seven types of defects using image
only take effect in a limited range of applications. In this processing algorithms. This work showed improved accu-
case, ANN has been widely adopted [155, 156] in both racies in both classification and segmentation tasks for all
turning and milling process. Using a small number of seven defect types. Crack identification was also realized
training samples, ANN is capable of generating accurate using a deep RBM from consumer-grade camera images
prediction values but would essentially require a good [163], which provided an alternative option in addition to
design of network structure. As compared to linear and CNN. In terms of part deformation prediction and con-
exponential regression model [155], neural networks trol, the utilization of responsive fixture made it possible
were found to be capable of better predictions for sur- to measure and accumulate online deformation data for
face roughness. Support vector regression (SVR) method different parts in different machining stages. Such data
was also utilized for the prediction of roughness. A com- potentiates the training of a mixed deep learning model,
parison of three types of SVRs and ANN was conducted as proposed by Zhao et al. [164], to predict the part
in Ref. [157], results showed that SVR can achieve pre- deformation and make process adjustments in an early
diction accuracy as high as 95%, while for ANN it was stage. As can be concluded from previous studies, most
slightly lower (91.4%) and required more computational deep learning based part condition analysis takes image
time at the same time. as raw input. It is conceivable that when the amount of
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 16 of 21

training data is limited, deep neural network, such as third-party agent. Consequently, when data is abundant,
CNN, can be easily over-fitted to jeopardize the accuracy. deep learning achieves better performance than conven-
In order to reduce data dependency, Ferguson et al. [165] tional machine learning approaches.
trained a CNN using openly-available image datasets and On a different perspective, utilizing deep learning
leveraged transfer learning to adapt the pre-trained CNN may be troublesome in manufacturing field, since the
model to the detection of defects, by using small X-ray acquisition of meaningful manufacturing data is not as
dataset. Cheng et al. [166] applied a parameter-based convenient as data from internet. Advanced machine
transfer learning in modeling shape deviations during learning technologies, such as transfer learning [136] and
additive manufacturing, as one particular example to meta learning [138], already left some successful marks
represent the future trend. Table 8 lists the evolution of in manufacturing applications where data acquisition
part condition analysis. is expensive and slow. It is foreseeable in the future that
more advanced machine learning methods dealing with
4.3 Discussion and Future Trend insufficient data will emerge and apply in manufacturing
Data analysis has been comprehensively reviewed from process.
three aspects in manufacturing process: tool condition, In the state-of-the-art development of machine learn-
process condition and part condition. The evolutions of ing techniques, new types of machine learning algo-
data analysis in all three aspects follow the same routine rithms for various tasks are being developed. Specifically,
from physical modeling to machine learning and reach- deep reinforcement learning using deep Q-network was
ing deep learning in the state-of-the-art. proposed by Google DeepMind [167], which opened
Due to the great complexity of manufacturing pro- up a new era to learn successful policies directly from
cess, the establishment of physical models would induce high-dimensional inputs and achieve human-level per-
noticeable errors. First of all, the construction of physi- formance in game play. The same group later proposed
cal model requires domain expertise which may contain a meta-reinforcement learning system inspired from the
cognitive bias to the actual mechanism. These manufac- activity of dopamine system in human brain [168], which
turing process usually comprise intricate and unstable expedited the learning process from past experience.
physical/chemical processes that are hard to precisely These new findings in reinforcement learning would
constructed, which inevitably require certain level of potentially render new possibilities for manufacturing
assumption and simplification. Mathematical solution systems to understand rules from source data and real-
based on limited observable data is sometimes ill-posed, ize true automation [169]. The lately reported domain-
making the final physical model barely accurate to deliver transform manifold learning made a huge success in
satisfactory results. The development of machine learn- noise-reduced image reconstruction from raw sensory
ing techniques inaugurated a new paradigm to analyze data [170], which could also be a promising tool in manu-
manufacturing processes, without needing to manually facturing data pre-processing stage for higher fidelity.
develop complicated but inaccurate physical models.
After some crafted feature extraction and training pro- 5 Conclusions and Outlook
cess, complex manufacturing process can be established 5.1 Conclusions
in a more unified, efficient and effective way. Through Manufacturing data collection and analysis are the key
the training process, hidden and obscure correlations enablers to realize data-driven manufacturing. As the
between the input and output can be unveiled. Neverthe- two crucial components in manufacturing monitoring
less, even the most powerful feature extractor still cannot system, they have been evolving to cater to increas-
guarantee zero discrepancy and error from the raw data, ing demands in modern manufacturing. The develop-
which directly affects the final accuracy. This dilemma ment of these two components have been thoroughly
is well resolved by deep neural networks, in which these investigated from literature, with conclusion depicted
features are automatically extracted rather than by a in Figure 7. In terms of data collection, in most manu-
facturing circumstances valuable data is measured via
Table 8 Evolution of part condition analysis sensors. Direct and indirect measurement are the two
categories in this stage. While indirect measurement
Physical model Machine learning
has been more widely adopted in recent manufacturing
Topography FEM Shallow Deep Learning applications for its cost-effectiveness and high compat-
simulation simulation learning learning from fewer
data ibility, it is still facing a considerable amount of dis-
crepancies in terms of accuracy. Although people have
[149, 150] [151–153] [155, 156, [160–164] [165, 166] made great efforts to reduce the error, it is theoretically
158, 159]
incapable to achieve high precision measurement due
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 17 of 21

5.2 Outlook
Thanks to the development of advanced sensoring and
data analyzing technologies, modern manufacturing
outperforms with higher efficiency, accuracy and self-
diagnosis by the extensive use of data. Direct process
monitoring combined with advanced machine learn-
ing technologies have achieved remarkable effectiveness
and will perhaps trend the development of data-driven
manufacturing. Though deep learning obtained huge
success in a variety of fields, training a deep model in
manufacturing scenario remains challenging due to the
Figure 7 Evolution of data-driven manufacturing
prolonged time and cost needed for collecting sufficient
labeled data. To overcome this crucial deficiency, there
are two suggested directions. First, it is though theoreti-
to the uncertainty and simplification of the correlation
cally impractical to train a deep model with high per-
between the target value and the measured value. On
formance using insufficient sample data, one can adopt
the other hand, direct measurement though encounter-
few-shot learning to extract common rules from exist-
ing incompatibility issue in some manufacturing cases,
ing well-trained knowledge, instead of training from
it will be ultimately adopted for its high fidelity and
scratch. Another potential direction is to combine physi-
achievable accuracy. There have already been sporadic
cal mechanism, such as Newton’s law and energy con-
developments of advanced sensors that can directly
servation, with machine learning models in order to take
measure the in-process data without violating the pro-
advantage of both, which would significantly reduce the
cess condition.
amount of training data and enhance the generalization
Data analysis is another crucial phase in data-driven
of the trained model. These are perhaps among the future
manufacturing to make diagnosis, predictions and other
shapes of data-driven smart manufacturing.
decisions based on the obtained data. Three paradigms
of data analysis, i.e., physical modeling, conventional Acknowledgements
machine learning and deep learning based data analysis Not applicable.

have been investigated in this paper. The development Authors’ Contributions

of physical modeling to describe high dimensional non- KX was responsible for the literature survey, structure arrangement and paper
linear manufacturing process is challenging and experi- writing; YL proposed the future trend of data-driven manufacturing and
determined the main scope of this review paper; CL, XL and XH provided
encing low accuracy. These physical models are usually constructive advices and case studies; JG and PGM gave professional advices
oversimplified, containing observation bias, and thus on paper writing. All authors read and approved the final manuscript.
theoretically incapable of accommodating the increas-
Funding
ing demand for accurate data analysis. Machine learning Supported by National Natural Science Foundation of China (Grant No.
based data analysis was developed to resolve this issue by 51805260), National Natural Science Foundation for Distinguished Young
using generic model such as neural networks, and train- Scholars of China (Grant No. 51925505), and National Natural Science Founda-
tion of China (Grant No. 51775278).
ing process, which can successfully overcome the model
simplification issue. As long as for a sufficient training Availability of Data and Materials
procedure, the learning model can achieve high nonlin- Not applicable.

earity to describe an arbitrarily complex process. While Competing Interests

shallow learning models can only deal with low-dimen- The authors declare that they have no competing interests.
sional data due to the limited capacity, deep learning
Author Details
based data analysis can achieve an end-to-end modeling 1
Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China.
from the raw sensory data to the final decision. The deep 2
Nanjing Tech University, Nanjing 211816, China. 3 University of Greenwich,
stacks of layers combined with dedicated training pro- London, UK. 4 Queen’s University Belfast, Belfast, UK.

cess can automatically learn to extract useful features Received: 5 November 2019 Revised: 26 April 2020 Accepted: 8 May 2020
without human intervention. Practically however, since
the amount of manufacturing data is usually limited,
advanced machine learning techniques such as trans-
fer learning and meta learning that require fewer train- References
ing samples are investigated in some recent studies to [1] J Tan, D Liu, Z Liu, et al. Research on key technical approaches for the
transition from digital manufacturing to intelligent manufacturing.
achieve better results and handle varying conditions. Strategic Study of Chinese Academy of Engineering, 2017, 19(3): 34-44.
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 18 of 21

[2] R Y Zhong, X Xu, E Klotz, et al. Intelligent manufacturing in the context [28] Y G Li, C Q Liu, X Z Hao, et al.Complex structural member digital
of industry 4.0: a review. Engineering, 2017, 3(5): 616-630. control processing realtime monitoring method based on machining
[3] S Yin, O Kaynak. Big data for modern industry: challenges and trends feature: CN, 105867305B. 2018-06-12.
[point of view]. Proceedings of the IEEE, 2015, 103(2): 143-146. [29] R Teti, K Jemielniak, G O’Donnell, et al. Advanced monitoring of machin-
[4] H Lasi, P Fettke, H-G Kemper, et al. Industry 4.0. Business & Information ing operations. CIRP Annals-Manufacturing Technology, 2010, 59(2):
Systems Engineering, 2014, 6(4): 239-242. 717-739.
[5] J Harding, M Shahbaz, A Kusiak. Data mining in manufacturing: a [30] G Tapia, A Elwany. A review on process monitoring and control in
review. In: Data mining in manufacturing. VDM Verlag Dr. Müller, 2006. metal-based additive manufacturing. Journal of Manufacturing Sci-
[6] S J Qin. Survey on data-driven industrial process monitoring and diag- ence and Engineering, 2014, 136(6): 060801.
nosis. Annual Reviews in Control, 2012, 36(2): 220-234. [31] S K Everton, M Hirsch, P Stravroulakis, et al. Review of in-situ process
[7] T Wuest, D Weimer, C Irgens, et al. Machine learning in manufacturing: monitoring and in-situ metrology for metal additive manufacturing.
advantages, challenges, and applications. Production & Manufacturing Materials & Design, 2016, 95: 431-445.
Research, 2016, 4(1): 23-45. [32] I A T Hashem, I Yaqoob, N B Anuar, et al. The rise of “big data” on cloud
[8] A Diez-Olivan, J Del Ser, D Galar, et al. Data fusion and machine learning computing: Review and open research issues. Information Systems,
for industrial prognosis: Trends and perspectives towards Industry 4.0. 2015, 47: 98-115.
Information Fusion, 2019, 50: 92-111. [33] J V Abellan-Nebot, F R Subirón. A review of machining monitoring
[9] L Kong, X Peng, Y Chen, et al. Multi-sensor measurement and data systems based on artificial intelligence process models. The Interna-
fusion technology for manufacturing process monitoring: A literature tional Journal of Advanced Manufacturing Technology, 2010, 47(1-4):
review. International Journal of Extreme Manufacturing, 2020, 2(2), https​ 237-257.
://iopsc​ience​.iop.org/artic​le/10.1088/2631-7990/ab7ae​6/meta. [34] N Ambhore, D Kamble, S Chinchanikar, et al. Tool condition monitor-
[10] F Tao, Q Qi, A Liu, et al. Data-driven smart manufacturing. Jour- ing system: A review. Materials Today: Proceedings, 2015, 2(4-5):
nal of Manufacturing Systems, 2018, https​://doi.org/10.1016/j. 3419-3428.
jmsy.2018.01.006. [35] K Xu, K Tang. Five-axis tool path and feed rate optimization based on
[11] F Tao, Q Qi. Make more digital twins. Nature Publishing Group, 2019. the cutting force–area quotient potential field. The International Jour-
[12] B Esmaeilian, S Behdad, B Wang. The evolution and future of manufac- nal of Advanced Manufacturing Technology, 2014, 75(9-12): 1661-1679.
turing: A review. Journal of Manufacturing Systems, 2016, 39: 79-100. [36] T-Y Kim, J Woo, D Shin, et al. Indirect cutting force measurement in
[13] J Li, F Tao, Y Cheng, et al. Big data in product lifecycle management. The multi-axis simultaneous NC milling processes. International Journal of
International Journal of Advanced Manufacturing Technology, 2015, 81(1- Machine Tools and Manufacture, 1999, 39(11): 1717-1731.
4): 667-684. [37] G D Kim, C N Chu. Indirect cutting force measurement considering
[14] K Ehmann, S G Kapoor, R DeVor, et al. Machining process modeling: a frictional behaviour in a machining centre using feed motor current.
review. Journal of Manufacturing Science and Engineering, Transactions of The International Journal of Advanced Manufacturing Technology, 1999,
the ASME, 1997, 119(4B): 655-663. 15(7): 478-484.
[15] D Mourtzis, M Doukas, D Bernidaki. Simulation in manufacturing: [38] A Albrecht, S S Park, Y Altintas, et al. High frequency bandwidth cutting
Review and challenges. Procedia CIRP, 2014, 25: 213-229. force measurement in milling using capacitance displacement sensors.
[16] C Leachman, C C Pegels, S K Shin. Manufacturing performance: evalua- International Journal of Machine Tools and Manufacture, 2005, 45(9):
tion and determinants. International Journal of Operations & Production 993-1008.
Management, 2005, 25(9): 851-874. [39] S Yaldız, F Ünsaçar, H Sağlam, et al. Design, development and testing of
[17] A Kusiak. Smart manufacturing must embrace big data. Nature News, a four-component milling dynamometer for the measurement of cut-
2017, 544(7648): 23. ting force and torque. Mechanical Systems and Signal Processing, 2007,
[18] B Mueller. Additive manufacturing technologies–Rapid prototyping to 21(3): 1499-1511.
direct digital manufacturing. Assembly Automation, 2012, 32(2), https​:// [40] G Byrne, G O’Donnell. An integrated force sensor solution for process
doi.org/10.1108/aa.2012.03332​baa.010. monitoring of drilling operations. CIRP Annals-Manufacturing Technol-
[19] D Wu, D W Rosen, L Wang, et al. Cloud-based design and manufactur- ogy, 2007, 56(1): 89-92.
ing: A new paradigm in digital manufacturing and design innovation. [41] X Jin, T Qin, Z Zhang, et al. Development and evaluation of a three-
Computer-Aided Design, 2015, 59: 1-14. component micro-cutting force wireless measurement apparatus
[20] D Wu, X Liu, S Hebert, et al. Democratizing digital design and manu- and method in turning-milling compound machining. The Interna-
facturing using high performance cloud computing: Performance tional Journal of Advanced Manufacturing Technology, 2017, 89(5-8):
evaluation and benchmarking. Journal of Manufacturing Systems, 2017, 1367-1378.
43: 316-326. [42] M Luo, H Luo, D Axinte, et al. A wireless instrumented milling cutter
[21] W He, E L Tan, E W Lee, et al. A solution for integrated track and trace system with embedded PVDF sensors. Mechanical Systems and Signal
in supply chain based on RFID & GPS. Emerging Technologies & Factory Processing, 2018, 110: 556-568.
Automation, 2009. ETFA 2009. IEEE Conference on, 2009: 1-6. [43] C Xiao, H Ding, K Cheng, et al. Design of an innovative smart turning
[22] G Wang, A Gunasekaran, E W Ngai, et al. Big data analytics in logistics tool with application to real-time cutting force measurement. Proceed-
and supply chain management: Certain investigations for research and ings of the Institution of Mechanical Engineers, Part B: Journal of Engineer-
applications. International Journal of Production Economics, 2016, 176: ing Manufacture, 2015, 229(3): 563-568.
98-110. [44] K Chen, C W Xiao, H Ding, et al. Intelligent tool for monitoring multiple
[23] J Hu, F L Lewis, O P Gan, et al. Discrete-event shop-floor monitoring physical quantities of machining state in real time: CN, 102601399B.
system in RFID-enabled manufacturing. IEEE Transactions on Industrial 2013-10-23.
Electronics, 2014, 61(12): 7083-7091. [45] G Quintana, J Ciurana. Chatter in machining processes: A review. Inter-
[24] W Ji, L Wang. Big data analytics based fault prediction for shop floor national Journal of Machine Tools and Manufacture, 2011, 51(5): 363-376.
scheduling. Journal of Manufacturing Systems, 2017, 43: 187-194. [46] T Delio, J Tlusty, S Smith. Use of audio signals for chatter detection and
[25] L Wang. Machine availability monitoring and machining process control. Journal of Engineering for Industry, 1992, 114(2): 146-157.
planning towards Cloud manufacturing. CIRP Journal of Manufacturing [47] X Li, Y Wong, A Nee. Tool wear and chatter detection using the coher-
Science and Technology, 2013, 6(4): 263-273. ence function of two crossed accelerations. International Journal of
[26] J Lee, B Bagheri, H-A Kao. A cyber-physical systems architecture for Machine Tools and Manufacture, 1997, 37(4): 425-435.
industry 4.0-based manufacturing systems. Manufacturing Letters, 2015, [48] M Yoon, D Chin. Cutting force monitoring in the endmilling operation
3: 18-23. for chatter detection. Proceedings of the Institution of Mechanical Engi-
[27] C Lauro, L Brandão, D Baldo, et al. Monitoring and processing signal neers, Part B: Journal of Engineering Manufacture, 2005, 219(6): 455-465.
applied in machining processes–A review. Measurement, 2014, 58: [49] H Q Liu, K M Mao, F Y Peng. On-line chatter detection using servo
73-86. motor current signal in turning. Science in China, 2011, 54(12):
3119-3129.
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 19 of 21

[50] I Inasaki. Application of acoustic emission sensor for monitoring [75] Y Li, P Gu. Free-form surface inspection techniques state of the art
machining processes. Ultrasonics, 1998, 36(1-5): 273-281. review. Computer-Aided Design, 2004, 36(13): 1395-1417.
[51] E Kuljanic, G Totis, M Sortino. Development of an intelligent multisen- [76] D H Lee, N G Cho. Assessment of surface profile data acquired by a
sor chatter detection system in milling. Mechanical Systems and Signal stylus profilometer. Measurement Science & Technology, 2012, 23(10):
Processing, 2009, 23(5): 1704-1718. 105601-105612(12).
[52] L Sallese, N Grossi, A Scippa, et al. Investigation and correction of actual [77] P Benardos, G-C Vosniakos. Predicting surface roughness in machining:
microphone response for chatter detection in milling operations. A review. International Journal of Machine Tools and Manufacture,
Measurement and Control, 2017, 50(2): 45-52. 2003, 43(8): 833-844.
[53] A Eppel, E T Enikov, T Insperger, et al. Feasibility study of optical [78] D R Salgado, F Alonso, I Cambero, et al. In-process surface roughness
detection of chatter vibration during milling. International Journal of prediction system using cutting vibrations in turning. The Interna-
Optomechatronics, 2010, 4(2): 195-214. tional Journal of Advanced Manufacturing Technology, 2009, 43(1-2):
[54] N Lei, M Soshi. Vision-based system for chatter identification and 40-51.
process optimization in high-speed milling. The International Journal of [79] J Z Zhang, J C Chen. The development of an in-process surface rough-
Advanced Manufacturing Technology, 2017, 89(9-12): 2757-2769. ness adaptive control system in end milling operations. The Interna-
[55] H Ding, P W Chen, H T Zhang, et al. Milling machining chatter active tional Journal of Advanced Manufacturing Technology, 2007, 31(9-10):
control system and method thereof: CN, 103252680B. 2015-05-20. 877-887.
[56] Y Altintaş, E Budak. Analytical prediction of stability lobes in milling. [80] C Beggan, M Woulfe, P Young, et al. Using acoustic emission to predict
CIRP Annals-Manufacturing Technology, 1995, 44(1): 357-362. surface quality. International Journal of Advanced Manufacturing
[57] O Özşahin, H N Özgüven, E Budak. Analysis and compensation of mass Technology, 1999, 15(10): 737-742.
loading effect of accelerometers on tool point FRF measurements for [81] S A Coker, Y C Shin. In-process control of surface roughness due to tool
chatter stability predictions. International Journal of Machine Tools and wear using a new ultrasonic system. International Journal of Machine
Manufacture, 2010, 50(6): 585-589. Tools and Manufacture, 1996, 36(3): 411-422.
[58] J Zhang, J Li, Z Xie, et al. Rapid dynamics prediction of tool point for [82] M S Tootooni, C Liu, D Roberson, et al. Online non-contact surface
bi-rotary head five-axis machine tool. Precision Engineering, 2017, 48: finish measurement in machining using graph theory-based image
203-215. analysis. Journal of Manufacturing Systems, 2016, 41: 266-276.
[59] S Kurada, C Bradley. A review of machine vision sensors for tool condi- [83] A Gameros, S Lowth, D Axinte, et al. State-of-the-art in fixture systems
tion monitoring. Computers in Industry, 1997, 34(1): 55-72. for the manufacture and assembly of rigid components: A review.
[60] N H Abu-Zahra, T H Nayfeh. Calibrated method for ultrasonic on-line International Journal of Machine Tools and Manufacture, 2017, 123:
monitoring of gradual wear during turning operations. International 1-21.
Journal of Machine Tools and Manufacture, 1997, 37(10): 1475-1484. [84] R Azouzi, M Guillot. On-line prediction of surface finish and dimensional
[61] N Cook, K Subramanian. Micro-isotope tool wear sensor. Annals of the deviation in turning using neural network based sensor fusion.
CIRP, 1978, 21(1): 67-72. International Journal of Machine Tools and Manufacture, 1997, 37(9):
[62] Y H Lee, R P Khetan, D R Sutherland. 3-D measurement of cutting tool 1201-1217.
wear: US, 5361308. 1994-11-01. [85] K Risbood, U Dixit, A Sahasrabudhe. Prediction of surface roughness
[63] M Nouri, B K Fussell, B L Ziniti, et al. Real-time tool wear monitoring in and dimensional deviation by measuring cutting forces and vibrations
milling using a cutting condition independent method. International in turning process. Journal of Materials Processing Technology, 2003,
Journal of Machine Tools and Manufacture, 2015, 89: 1-13. 132(1-3): 203-214.
[64] J-H Zhou, C K Pang, Z-W Zhong, et al. Tool wear monitoring using [86] P Kersting, D Biermann. Modeling techniques for simulating workpiece
acoustic emissions by dominant-feature identification. IEEE Transactions deflections in NC milling. Cirp Journal of Manufacturing Science & Tech-
on Instrumentation and Measurement, 2011, 60(2): 547-559. nology, 2014, 7(1): 48-54.
[65] J C Chen, W-L Chen. A tool breakage detection system using an [87] W Song, C Xu, Q Pan, et al. Nondestructive testing and characterization
accelerometer sensor. Journal of Intelligent Manufacturing, 1999, 10(2): of residual stress field using an ultrasonic method. Chinese Journal of
187-197. Mechanical Engineering, 2016, 29(2): 365-371.
[66] D Salgado, F Alonso. An approach based on current and sound signals [88] J Zhang, F Li, L Zheng, et al. Internal residual stresses in the friction stir
for in-process tool wear monitoring. International Journal of Machine weldment of 2024-T 351 al alloy determined by short wavelength X-ray
Tools and Manufacture, 2007, 47(14): 2140-2152. diffraction. Journal of Mechanical Engineering, 2013, 49(2): 28-34. (in
[67] S H Lee. Method and apparatus of detecting tool abnormality in a Chinese)
machine tool: US, 6937942B2. 2005-08-30. [89] N Huang, Q Bi, Y Wang, et al. 5-Axis adaptive flank milling of flexible
[68] B Li, H Q Liu, X Y Mao, et al. A kind of numerically-controlled machine thin-walled parts based on the on-machine measurement. International
milling process condition monitoring of tool breakage method: CN, Journal of Machine Tools and Manufacture, 2014, 84: 1-8.
103324139B. 2016-02-24. [90] M Luo, D Liu, H Luo. Real-time deflection monitoring for milling of a
[69] K V Rao, B Murthy, N M Rao. Cutting tool condition monitoring by thin-walled workpiece by using PVDF thin-film sensors with a cantile-
analyzing surface roughness, work piece vibration and volume of vered beam as a case study. Sensors, 2016, 16(9): 1470.
metal removed for AISI 1040 steel in boring. Measurement, 2013, 46(10): [91] Y Yuan, Q Bi, L Zhu, et al. Real-time normal measurement and error
4075-4084. compensation of curved aircraft surface based on on-line thickness
[70] C Zhang, J Zhang. On-line tool wear measurement for ball-end milling measurement. International Conference on Intelligent Robotics & Applica-
cutter based on machine vision. Computers in Industry, 2013, 64(6): tions, 2017.
708-719. [92] I Lazoglu, D Ulutan, B Alaca, et al. An enhanced analytical model for
[71] L Čerče, F Pušavec, J Kopač. A new approach to spatial tool wear analy- residual stress prediction in machining. CIRP Annals-Manufacturing
sis and monitoring. Strojniški vestnik-Journal of Mechanical Engineering, Technology, 2008, 57(1): 81-84.
2015, 61(9): 489-497. [93] Y Li, C Liu, X Hao, et al. Responsive fixture design using dynamic prod-
[72] T G Dawson, T R Kurfess. Quantification of tool wear using white light uct inspection and monitoring technologies for the precision machin-
interferometry and three-dimensional computational metrology. ing of large-scale aerospace parts. CIRP Annals, 2015, 64(1): 173-176.
International Journal of Machine Tools and Manufacture, 2005, 45(4-5): [94] H-C Möhring, P Wiederkehr. Intelligent fixtures for high performance
591-596. machining. Procedia CIRP, 2016, 46: 383-390.
[73] J A Ramirez-Nunez, M Trejo-Hernandez, R J Romero-Troncoso, et al. [95] D Lucke, C Constantinescu, E Westkämper. Smart factory-a step towards
Smart-sensor for tool-breakage detection in milling process under dry the next generation of manufacturing. In: Manufacturing systems and
and wet conditions based on infrared thermography. The International technologies for the new frontier, Springer, 2008: 115-118.
Journal of Advanced Manufacturing Technology, 2018: 1-13. [96] D-Y Song, N Otani, T Aoki, et al. A new approach to cutting state moni-
[74] Y Dai, K Zhu. A machine vision system for micro-milling tool condi- toring in end-mill machining. International Journal of Machine Tools and
tion monitoring. Precision Engineering, 2018, 52: 183-191. Manufacture, 2005, 45(7-8): 909-921.
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 20 of 21

[97] G Campatelli, A Scippa. Prediction of milling cutting force coefficients [121] A Altin, M Nalbant, A Taskesen. The effects of cutting speed on tool
for Aluminum 6082-T4. Procedia CIRP, 2012, 1: 563-568. wear and tool life when machining Inconel 718 with ceramic tools.
[98] Y M Ertekin, Y Kwon, T-L B Tseng. Identification of common sensory Materials & Design, 2007, 28(9): 2518-2522.
features for the control of CNC milling operations under varying cutting [122] P Marksberry, I Jawahir. A comprehensive tool-wear/tool-life perfor-
conditions. International Journal of Machine Tools and Manufacture, mance model in the evaluation of NDM (near dry machining) for
2003, 43(9): 897-904. sustainable manufacturing. International Journal of Machine Tools and
[99] N Ghosh, Y Ravi, A Patra, et al. Estimation of tool wear during CNC mill- Manufacture, 2008, 48(7-8): 878-886.
ing using neural network-based sensor fusion. Mechanical Systems and [123] R Zhao, R Yan, J Wang, et al. Learning to monitor machine health with
Signal Processing, 2007, 21(1): 466-479. convolutional bi-directional LSTM networks. Sensors, 2017, 17(2): 273.
[100] Y Altintas, M Eynian, H Onozuka. Identification of dynamic cutting force [124] J C Chen, J C Chen. An artificial-neural-networks-based in-process tool
coefficients and chatter stability with process damping. CIRP Annals- wear prediction system in milling operations. The International Journal
Manufacturing Technology, 2008, 57(1): 371-374. of Advanced Manufacturing Technology, 2005, 25(5-6): 427-434.
[101] O Abouelatta, J Madl. Surface roughness prediction based on cutting [125] P Palanisamy, I Rajendran, S Shanmugasundaram. Prediction of tool
parameters and tool vibrations in turning operations. Journal of Materi- wear using regression and ANN models in end-milling operation. The
als Processing Technology, 2001, 118(1-3): 269-277. International Journal of Advanced Manufacturing Technology, 2008, 37(1-
[102] D A Axinte, N Gindy. Tool condition monitoring in broaching. Wear, 2): 29-41.
2003, 254(3-4): 370-382. [126] S Cho, S Asfour, A Onar, et al. Tool breakage detection using support
[103] A Verl, U Heisel, M Walther, et al. Sensorless automated condition moni- vector machine learning in a milling process. International Journal of
toring for the control of the predictive maintenance of machine tools. Machine Tools and Manufacture, 2005, 45(3): 241-249.
CIRP Annals, 2009, 58(1): 375-378. [127] D Shi, N N Gindy. Tool wear predictive model based on least squares
[104] E M Rubio, R Teti. Cutting parameters analysis for the development of support vector machines. Mechanical Systems and Signal Processing,
a milling process monitoring system based on audible energy sound. 2007, 21(4): 1799-1814.
Journal of Intelligent Manufacturing, 2009, 20(1): 43. [128] D Brezak, D Majetic, T Udiljak, et al. Tool wear estimation using an ana-
[105] P Huang, J Li, J Sun, et al. Vibration analysis in milling titanium alloy lytic fuzzy classifier and support vector machines. Journal of Intelligent
based on signal processing of cutting force. The International Journal of Manufacturing, 2012, 23(3): 797-809.
Advanced Manufacturing Technology, 2013, 64(5-8): 613-621. [129] M Elangovan, S B Devasenapati, N Sakthivel, et al. Evaluation of expert
[106] F Al-Badour, M Sunar, L Cheded. Vibration analysis of rotating machin- system for condition monitoring of a single point cutting tool using
ery using time–frequency analysis and wavelet techniques. Mechanical principle component analysis and decision tree algorithm. Expert
Systems and Signal Processing, 2011, 25(6): 2083-2101. Systems with Applications, 2011, 38(4): 4450-4459.
[107] I Marinescu, D A Axinte. A critical analysis of effectiveness of acoustic [130] R K Fish, M Ostendorf, G D Bernard, et al. Multilevel classification of mill-
emission signals to detect tool and workpiece malfunctions in milling ing tool wear with confidence estimation. IEEE Transactions on Pattern
operations. International Journal of Machine Tools and Manufacture, Analysis & Machine Intelligence, 2003, 25(1): 75-85.
2008, 48(10): 1148-1160. [131] Y LeCun, Y Bengio, G Hinton. Deep learning. Nature, 2015, 521(7553):
[108] X Li, S Dong, Z Yuan. Discrete wavelet transform for tool breakage 436.
monitoring. International Journal of Machine Tools and Manufacture, [132] S Lawrence, C L Giles, A C Tsoi, et al. Face recognition: a convolutional
1999, 39(12): 1935-1944. neural-network approach. IEEE Transactions on Neural Networks, 1997,
[109] X Li, X Guan. Time-frequency-analysis-based minor cutting edge 8(1): 98-113.
fracture detection during end milling. Mechanical Systems and Signal [133] S Hochreiter, J Schmidhuber. Long short-term memory. Neural Compu-
Processing, 2004, 18(6): 1485-1496. tation, 1997, 9(8): 1735-1780.
[110] H Aouici, M A Yallese, K Chaoui, et al. Analysis of surface roughness and [134] G Li, X Yang, D Chen, et al. Tool breakage detection using deep learning.
cutting force components in hard turning with CBN tool: Prediction 2018 IEEE International Conference on Big Data, Cloud Computing, Data
model and cutting conditions optimization. Measurement, 2012, 45(3): Science & Engineering (BCD), 2018: 37-42.
344-353. [135] A Kothuru, S P Nooka, R Liu. Audio-based tool condition monitoring
[111] Z Hessainia, A Belbah, M A Yallese, et al. On the prediction of surface in milling of the workpiece material with the hardness variation using
roughness in the hard turning based on cutting parameters and tool support vector machines and convolutional neural networks. Journal of
vibrations. Measurement, 2013, 46(5): 1671-1681. Manufacturing Science and Engineering-Transactions of the ASME, 2018,
[112] E Kannatey-Asibu Jr, D Dornfeld. A study of tool wear using statistical 140(11).
analysis of metal-cutting acoustic emission. Wear, 1982, 76(2): 247-261. [136] G Chen, Y Li, X Liu. Pose-dependent tool tip dynamics prediction using
[113] M-C Lu, B-S Wan. Study of high-frequency sound signals for tool wear transfer learning. International Journal of Machine Tools and Manufacture,
monitoring in micromilling. The International Journal of Advanced Manu- 2019, 137: 30-41.
facturing Technology, 2013, 66(9-12): 1785-1792. [137] C Sun, M Ma, Z Zhao, et al. Deep transfer learning based on sparse
[114] H Takeyama, R Murata. Basic investigation of tool wear. Journal of Engi- autoencoder for remaining useful life prediction of tool in manufactur-
neering for Industry, 1963, 85(1): 33-37. ing. IEEE Transactions on Industrial Informatics, 2018, 15(4): 2416-2425.
[115] J Rech, A Giovenco, C Courbon, et al. Toward a new tribological [138] Y Li, C Liu, J Hua, et al. A novel method for accurately monitoring and
approach to predict cutting tool wear. CIRP Annals, 2018, 67(1): 65-68. predicting tool wear under varying cutting conditions based on meta-
[116] F W Taylor. On the art of cutting metals. American Society of Mechanical learning. CIRP Annals, 2019.
Engineers, 1906. [139] E Budak, Y Altintas. Analytical prediction of chatter stability in milling—
[117] H Kishawy, S Kannan, M Balazinski. Analytical modeling of tool wear Part I: general formulation. Journal of Dynamic Systems, Measurement,
progression during turning particulate reinforced metal matrix com- and Control, 1998, 120(1): 22-30.
posites. CIRP Annals-Manufacturing Technology, 2005, 54(1): 55-58. [140] E Ozturk, E Budak. Dynamics and stability of five-axis ball-end milling.
[118] S Oraby, D Hayhurst. Tool life determination based on the measure- Journal of Manufacturing Science and Engineering, 2010, 132(2): 021003.
ment of wear and tool force ratio variation. International Journal of [141] T Choi, Y C Shin. On-line chatter detection using wavelet-based param-
Machine Tools and Manufacture, 2004, 44(12-13): 1261-1269. eter estimation. Journal of Manufacturing Science and Engineering, 2003,
[119] A S Kumar, A R Durai, T Sornakumar. The effect of tool wear on tool life 125(1): 21-28.
of alumina-based ceramic cutting tools while machining hardened [142] Z Yao, D Mei, Z Chen. On-line chatter detection and identification based
martensitic stainless steel. Journal of Materials Processing Technology, on wavelet and support vector machine. Journal of Materials Processing
2006, 173(2): 151-156. Technology, 2010, 210(5): 713-719.
[120] J Arsecularatne, L Zhang, C Montross. Wear and tool life of tungsten [143] M Lamraoui, M Barakat, M Thomas, et al. Chatter detection in milling
carbide, PCBN and PCD cutting tools. International Journal of Machine machines by neural network classification and feature selection. Journal
Tools and Manufacture, 2006, 46(5): 482-491. of Vibration and Control, 2015, 21(7): 1251-1266.
Xu et al. Chin. J. Mech. Eng. (2020) 33:43 Page 21 of 21

[144] Y Fu, Y Zhang, Y Gao, et al. Machining vibration states monitoring [158] X Li, P K Venuvinod, A Djorjevich, et al. Predicting machining errors in
based on image representation using convolutional neural networks. turning using hybrid learning. The International Journal of Advanced
Engineering Applications of Artificial Intelligence, 2017, 65: 240-251. Manufacturing Technology, 2001, 18(12): 863-872.
[145] A-r Mohamed, G E Dahl, G Hinton. Acoustic modeling using deep belief [159] Á Arnaiz-González, A Fernández-Valdivielso, A Bustillo, et al. Using artifi-
networks. IEEE Trans. Audio, Speech & Language Processing, 2012, 20(1): cial neural networks for the prediction of dimensional error on inclined
14-22. surfaces manufactured by ball-end milling. The International Journal of
[146] Y Fu, Y Zhang, H Qiao, et al. Analysis of feature extracting ability for cut- Advanced Manufacturing Technology, 2016, 83(5-8): 847-859.
ting state monitoring using deep belief networks. Procedia CIRP, 2015, [160] J Masci, U Meier, D Ciresan, et al. Steel defect classification with max-
31: 29-34. pooling convolutional neural networks. In: Neural Networks (IJCNN), The
[147] M C Yesilli, F A Khasawneh, A Otto. On transfer learning for chatter 2012 International Joint Conference on, 2012: 1-6.
detection in turning using wavelet packet transform and ensemble [161] J-K Park, B-K Kwon, J-H Park, et al. Machine learning-based imaging
empirical mode decomposition. CIRP Journal of Manufacturing Science system for surface defect inspection. International Journal of Precision
and Technology, 2019. Engineering and Manufacturing-Green Technology, 2016, 3(3): 303-310.
[148] T Yandayan, M Burdekin. In-process dimensional measurement and [162] R Ren, T Hung, K C Tan. A generic deep-learning-based approach for
control of workpiece accuracy. International Journal of Machine Tools automated surface inspection. IEEE Transactions on Cybernetics, 2018,
and Manufacture, 1997, 37(10): 1423-1439. 48(3): 929-940.
[149] S Lin, M Chang. A study on the effects of vibrations on the surface finish [163] Y Xu, S Li, D Zhang, et al. Identification framework for cracks on a steel
using a surface topography simulation model for turning. International structure surface by a restricted Boltzmann machines algorithm based
Journal of Machine Tools and Manufacture, 1998, 38(7): 763-782. on consumer‐grade camera images. Structural Control and Health
[150] B Kim, C Chu. Texture prediction of milled surfaces using texture super- Monitoring, 2018, 25(2): e2075.
position method. Computer-Aided Design, 1999, 31(8): 485-494. [164] Z Zhao, Y Li, C Liu, et al. On-line part deformation prediction based on
[151] J Wang, D Zhang, B Wu, et al. Prediction of distortion induced by deep learning. Journal of Intelligent Manufacturing, 2019: 1-14.
machining residual stresses in thin-walled components. The Interna- [165] M K Ferguson, A Ronay, Y-T T Lee, et al. Detection and segmentation of
tional Journal of Advanced Manufacturing Technology, 2018, 95(9-12): manufacturing defects with convolutional neural networks and transfer
4153-4162. learning. Smart and Sustainable Manufacturing Systems, 2018, 2.
[152] E Brinksmeier, J Sölter. Prediction of shape deviations in machining. CIRP [166] L Cheng, F Tsung, A Wang. A statistical transfer learning perspective for
Annals, 2009, 58(1): 507-510. modeling shape deviations in additive manufacturing. IEEE Robotics and
[153] M Wan, W Zhang, G Qin, et al. Strategies for error prediction and error Automation Letters, 2017, 2(4): 1988-1993.
control in peripheral milling of thin-walled workpiece. International [167] V Mnih, K Kavukcuoglu, D Silver, et al. Human-level control through
Journal of Machine Tools and Manufacture, 2008, 48(12-13): 1366-1374. deep reinforcement learning. Nature, 2015, 518(7540): 529.
[154] S Ratchev, S Liu, W Huang, et al. Milling error prediction and compensa- [168] J X Wang, Z Kurth-Nelson, D Kumaran, et al. Prefrontal cortex as a meta-
tion in machining of low-rigidity parts. International Journal of Machine reinforcement learning system. Nature Neuroscience, 2018, 21(6): 860.
Tools and Manufacture, 2004, 44(15): 1629-1641. [169] Q Xiao, C Li, Y Tang, et al. Meta-reinforcement learning of machining
[155] T Özel, Y Karpat. Predictive modeling of surface roughness and tool parameters for energy-efficient process control of flexible turning
wear in hard turning using regression and neural networks. Interna- operations. IEEE Transactions on Automation Science and Engineering,
tional Journal of Machine Tools and Manufacture, 2005, 45(4-5): 467-479. 2019.
[156] A M Zain, H Haron, S Sharif. Prediction of surface roughness in the end [170] B Zhu, J Z Liu, S F Cauley, et al. Image reconstruction by domain-trans-
milling machining using Artificial Neural Network. Expert Systems with form manifold learning. Nature, 2018, 555(7697): 487.
Applications, 2010, 37(2): 1755-1768.
[157] U Çaydaş, S Ekici. Support vector machines models for surface rough-
ness prediction in CNC turning of AISI 304 austenitic stainless steel.
Journal of Intelligent Manufacturing, 2012, 23(3): 639-650.