Big Data Analytics For Intelligent Manufacturing Systems A Review

Journal of Manufacturing Systems 62 (2022) 738–752
Contents lists available at ScienceDirect
Journal of Manufacturing Systems

journal homepage: www.elsevier.com/locate/jmansys
Big data analytics for intelligent manufacturing systems: A review

Junliang Wang a, b, c, Chuqiao Xu d, Jie Zhang a, b, *, Ray Zhong e
a
Institute of Artificial Intelligence, Donghua University, Shanghai, China
b
Shanghai Engineering Research Center of Industrial Big Data and Intelligent System, Donghua University, Shanghai, China
c
National Synthetic Fiber Engineering Technology Research Center, Beijing Chonglee Machinery Engineering Co., Ltd, Beijing, China
d
School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai, China
e
Department of Industrial and Manufacturing Systems Engineering, The University of Hong Kong, Hong Kong
A R T I C L E I N F O A B S T R A C T
Keywords: With the development of Internet of Things (IoT), 5 G, and cloud computing technologies, the amount of data
Big data analytics (BDA) from manufacturing systems has been increasing rapidly. With massive industrial data, achievements beyond
Intelligent manufacturing expectations have been made in the product design, manufacturing, and maintain process. Big data analytics
Artificial intelligence
(BDA) have been a core technology to empower intelligent manufacturing systems. In order to fully report BDA
Manufacturing systems
for intelligent manufacturing systems, this paper provides a comprehensive review of associated topics such as
the concept of big data, model driven and data driven methodologies. The framework, development, key tech
nologies, and applications of BDA for intelligent manufacturing systems are discussed. The challenges and op
portunities for future research are highlighted. Through this work, it is hoped to spark new ideas in the effort to
realize the BDA for intelligent manufacturing systems.
1. Introduction [6–8]. To understand the current state of the research and provide in
sights for future studies, this paper systematically analyzes current
With the arrival of a new round of industrial revolution, information research efforts, derive prominent research themes, and identify gaps in
technology has accelerated its integration with manufacturing systems, the current literature about big data analytics for intelligent
and the data owned by enterprises have become increasingly rich, with manufacturing systems (BDAIMS).
features of volume, variety, and velocity [1]. In intelligent A bibliometric analysis was conducted with published data from
manufacturing, industrial big data not only promotes enterprises to 2011 to 2020 regarding intelligent manufacturing driven by big data
accurately perceive the internal and external environment changes in gathered from the Web of Science database (as shown in Fig. 1), which
the system but also facilitates scientific analysis and decision making to shows an almost steady increase in papers on this topic. Fig. 1(a) shows
optimize the production process, reduce costs and improve operational the number of published articles about intelligent manufacturing from
efficiency. With massive data, new business modes are generated out of 2011 to 2020. From 2011–2014, the amounts of articles are small and
expectation (such as mass customization [2] and precision marketing have little change. But from 2014 to 2019, the amounts have been
[3]) to empower the social development and economic growth. As a increased sharply and the ratio keeps going up. Affected by COVID-19,
result, big industrial data is regarded as a mean of production to driving the amounts of relevant articles published in 2020 have decreased.
intelligent manufacturing. Fig. 1(b) shows the top areas related to intelligent manufacturing. As
With the development of artificial intelligence, big data analytics shown in the chart, the top five areas are the Engineering, the Computer
(BDA) have been greatly improved to effectively mine both structured Science, the Automation Control Systems, the Business Economics and
and unstructured industrial data in intelligent manufacturing, which has the Telecommunications. The Computer Science is the most related area
become a new research hotspot [4]. Continuous learning from big data for intelligent manufacturing. With the technology developing rapidly,
of manufacturing system enables the system to be self-learning, self- the Computer Science will be the key to break through the bottleneck of
optimization and self-regulation [5]. With the further development of intelligent manufacturing. Fig. 1(c) shows the productive scholars
BDA, the operation of manufacturing systems will be deeply changed publishing in this area. Fig. 1(d) lists the top universities or research
* Corresponding author at: Institute of Artificial Intelligence, Donghua University, Shanghai, China.
E-mail address: [email protected] (J. Zhang).
https://doi.org/10.1016/j.jmsy.2021.03.005
Received 30 November 2020; Received in revised form 4 March 2021; Accepted 5 March 2021
Available online 12 June 2021
0278-6125/© 2021 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
J. Wang et al. Journal of Manufacturing Systems 62 (2022) 738–752
institutes publishing in this area. The top five universities are South continuous manufacturing process, various sensing devices and
China University of Technology, Chinese Academy of Sciences, Shanghai real-time and efficient data transmission make the data possess the
Jiao Tong University, University of Hong Kong and Huazhong Univer typical characteristics of "3V" of big data of volume, variety and velocity
sity of Science Technology. This figure shows that most of them are [7]. Through further analysis of the data in manufacturing workshops, it
Chinese universities, which corresponds to the following analytics about is found to be characterized by multi-source, multi-dimension,
countries or regions. As shown in Fig. 1(e), countries or regions that are multi-noise, imbalance, and time series.
active in this field, of which China, the United States, and the United Multi-source: Along with the rise of IoT technologies and the
Kingdom are the top three. These articles are sourced from the Web of number of sensors in workshops, the manufacturing data is stored and
Science database with a focus on key concepts of intelligent captured in multi-sourced information systems [8] with a different data
manufacturing and big data. By analyzing these key technologies and structure [9]. To integrate multi-source data in the manufacturing pro
related academic movements, the way forward is clearer and the next cess, Zhu [10] proposed a big data analytics framework for smart tool
new wave is coming soon. condition monitoring (TCM), using multi-source information fusion
method to process image data, 3-d point cloud data, and frequency
2. Concepts and definitions signal data, achieve the monitor and adaptively control of machining
process in real time under varying working conditions. Zhang [11]
2.1. Concept of big data proposed a pixel level fusion method that combines raw data from
multiple sources into single resolution data, and summarizes several
With the development of IoT, intelligent manufacturing has focused trends tending to broaden the application of multi-source data fusion.
on the collection of enormous data called big data [6]. However, it is Tao [12] analyzed the dependency of data in product lifecycle man
facing great challenges when making full use of such data. The agement under big data environment, and proposed a conceptual
Fig. 1. Statistics from Web of Science database (search keywords: “intelligent manufacturing” and “big data”; Date: 25 November 2020). (a) Published articles per
year; (b) published articles by areas; (c) published articles by author; (d) published articles by affiliation; (e) published articles by country/region.
739
framework with higher flexibility, accuracy, and less computing time, time series data. In this algorithm, the random decision trees were
which can deal with multi-source data and massive data. Zhang [13] created via modified algorithm of bagging and unbiased split selection
proposed an overall architecture of multi-source lifecycle big data for based on conditional probabil probability index so as to construct
product lifecycle (BDA-PL), to make better product lifecycle manage random forests. In petroleum production, there are a large amount of
ment (PLM) and (cleaner production) CP decisions in manufacturing nonlinearity and complexity of time-series data, Sagheer [32] intro
processes. duced a deep learning approach capable to address the limitations of
Multi-dimension: In manufacturing systems, the performance pa traditional forecasting approaches and improve the accuracy of
rameters are influenced by different factors, which results in a multi- predictions.
dimension problem in prediction and control [14]. The current
methods for multi-dimension problems can be classified into numerous 2.2. Concept of model driven and data driven
types. For example, in semiconductor wafer manufacturing systems, the
cycle time of wafer products is influenced by more than one thousand There are two main paradigms for BDA: model-driven, and data-
factors, such as the processing time of each operation, the size of waiting driven modes.
queue for each machine, and the utilization of a machine [15]. In health Model-Driven is the way to start with a solid idea of how the physical
condition monitoring of the machines, to fully inspect the health con system works. Model-driven approaches are powerful because they rely
ditions, condition monitoring systems are used to collect real-time data on a deep understanding of the system or process, and can benefit from
from multiple sensors after the long-time operation [16]. Aiming at the scientifically established relationships [33]. Models can’t accommodate
multi-dimension feature of manufacturing data, Liu [17] proposed a infinite complexity and generally must be simplified. They have trouble
novel integrated process planning and control method based on intelli in constructing a well fitted model accounting for noisy data and
gent software agents and multi-dimension manufacturing features. Luo non-included variables. In real practice scenarios, building up a so
[18] proposed a cloud manufacturing (CMfg) architecture to process phisticated model integrates the physical, mechanical, electronic, data
multi-dimension data, achieving on-demand use, dynamic collabora flow, or other appropriate details of the complex systems. Furthermore,
tion, and circulation of manufacturing resources. modeling takes time. It is inherently a trial-and-error approach, rooted
Multi-noise: Industrial big data also has the characteristics of multi- in the old scientific method of theory-based hypothesis formation and
noise at the same time since the electromagnetic interference and harsh experiment-based testing. Finding a suitable model and refining it until
environment [19]. For example, in a dataset of wafer cycle time fore it produces the desired results is often a lengthy process [34].
casting, about five thousand records are containing missing value, and Data driven is another model-free way based on the correlation be
fifteen hundred records containing abnormal value, in every 8 million tween system status parameters and the target estimated by carious
records [20]. To solve the tasks with multi-noise data, Lee [21] proposed artificial intelligent models, which have high accuracy for accurate
a systematic and data-driven approach to process these missing values, optimization [35]. Find an algorithm that can spot connections and
and random sampling data for identifying semiconductor wafers defects. correlations by extracting knowledge from a database that can be ob
In order to reduce the impact of noise data, Zhao [22] developed the tained from historical measurements. Validations reveal their excellent
deep residual networks with dynamically weighted wavelet coefficients performance in computing efficiency, accuracy, and decision-making
(DRN + DWWC) to adaptively filter out these multi-noise data, which rationality. However, data driven methods strongly rely on expert
improved the accuracy of gearbox fault diagnosis. experience in designing machine learning algorithms and their perfor
Imbalance: The number of samples in the industry also has the mance is also affected by the quantity and quality of the
characteristic of imbalance. For example, the data samples with defects prior-knowledge database. Data driven methods work based on the
are much less than normal ones [23]. As indicated by the previous BDA co-distribution hypothesis. The machine learning tools that discover
studies, the class imbalance is a critical issue accounting for the bad features and train-up classifiers learn from examples, and there need to
performance, since they are biased toward the majority classes and tend be enough examples to cover the full range of expected variation and
to misclassify minority class examples [24]. In pattern recognition null cases. To train the generalized model and discover viable feature
problems, several approaches have been proposed to solve the classifi sets and decision criteria, big data is truly needed to get meaningful
cation tasks with imbalanced data, which can be classified into two results. The data driven methods training with big data are called to be
types: model modifying method, and data processing method [25]. Kang big data driven.
and Zhou [26] proposed a distance based weighted down-sampling
scheme for support vector machine (SVM) for imbalanced classifica 3. Framework and development
tion. On the opposite, the oversampling method synthetic data instances
to augment the minority classes to change the distribution of classes. 3.1. The framework of big data analytics for intelligent manufacturing
Zhang et al. [27] designed a weighted minority oversampling (WMO) to systems
rebalance the data distribution during fault diagnosis of rotating
machinery. The primary science paradigm of BDAIMS is data science, which was
Time series：：A lot of industrial data are collected in time series, highlighted in the book entitled the fourth paradigm: data intensive
such as the vibration signal, and temperature data. In wafer scientific discovery. In 2009, Tony et al. pointed out from the perspec
manufacturing production, the machine processing data, wafer mea tive of scientific research paradigm that data intensive science would
surements, defect distribution data, wafer electrical properties data, become the next paradigm after experimental science, theoretical deri
actual production monitoring data and machine operation status data vation and simulation [36]. The traditional scientific research para
are all time series data. These time series data are very important to digms construct sophisticated mathematical models to approximate real
provide variation tendency for improving the accuracy of prediction systems through experiment, derivation and simulation, and analyzes
[28]. To deal with the time series data, Vafeiadis [29] proposed a and optimizes the system. In complicated large-scaled dynamic systems,
self-adaptive sliding data window method to extract time series features. well experimental, mathematical, and simulation models are hard to be
To process the captured data from sensors (time-series data) for pre constructed. On the other hand, big data extracts knowledge by mining
dictive maintenance of equipment, Villalobos [30] presented a system the correlation between data, which can provide stronger insight,
I4TSRS1, analyzing time-series data by combining the time-series analysis and decision-making ability.
reduction techniques with extracted features from industrial Under the paradigm of data science, the big data driven operation of
time-series. In order to predict the quality of products in textile pro manufacturing systems change to emerge a framework of "correlation +
duction, Seçkin [31] proposed an improved random forest algorithm for prediction + regulation" (shown in Fig. 2). (1) correlation analysis
740
Fig. 2. The framework of big data driven intelligent manufacturing (adjusted from [37]).
means to quantify the relationship between different factors in the The first generation of BDA aims to solve complex problems with
manufacturing systems from the perspective of data. (2) Prediction re simple algorithms. The representative work was influenza epidemics
fers to further forecast the performance indicators of manufacturing detector from google [39]. This detector estimated the probability that a
systems (e.g. cycle time and yield) by machine learning methods. (3) random physician visit in a particular region was related to an ILI in
Regulation refers to optimizing the controllable variables to improve the dicator which was equivalent to the percentage of ILI-related physician
system performance. From the view of data processing cycle, the BDAIM visits. In this detector, a simple linear regression model was developed to
works in four steps: (1) The manufacturing data is integrated and pre model the relationship between an ILI physician visit and the log-odds of
processed to provide reliable and reusable data. (2) The correlative an ILI-related search query. With large numbers of Google search
analysis is conducted to obtain the explanatory factors of performance queries, the influenza-like illness in a population was accurately tracked
indicators of manufacturing systems. With the explanatory factors, the for each of the nine surveillance regions of the United States without any
fluctuation of system performance indicators can be modeled. (3) With influenza transmission models. In the first generation of BDA, data is
the explanatory factors, the system performance indicators can be pre treated as the most important ingredient to be a success.
dicted to provide insights for decision making. Different kinds of ma
chine learning models are developed for accurate prediction. In this 2) The second generation of BDA
section, prediction is generalized to consist fault detection, classifica
tion, and other extended prediction tasks in manufacturing systems. (4) The second generation of BDA aims to solve a complex problem with
Based on the predicted value, the decision-making methods can be complicated algorithms. Deep learning models are central enablers for
implemented to improve the systems performance. Usually, the func this generation. The massive parameters (such as weights and biases) in
tion, structure, process of product can be optimized by design data the deep neural networks can be adjusted to handle high dimensional
analysis. The efficiency of manufacturing systems can be optimized problems. In this period, outstanding progress has been made in the
through planning and scheduling. The yield of products was controlled complicated high dimensional problems, such as prognostics health
and improved by process and quality control systems. The system management of equipment, surface defect recognition, neural language
robustness was guaranteed by prognostics health management in intel processing. The representative mark was the AlphaGo designed by Deep
ligent maintain. Mind [40], which introduced Monte Carlo search tree and deep learning
models into Go games. The latest version AlphaGo Zero was trained by
3.2. The development of big data analytics self-play during 4.9 million games for approximately 3 days. However,
the number of games in the life of a human player could be ten thousand
The big data analytics are enabled by the evolution of artificial in class. With the power of training and deep learning algorithms, no
telligence [38], which can be divided into three generations as shown in human professional player can beat AlphaGo Zero. In manufacturing
Fig. 3. systems, different kinds of problems can be modeled and solved by
various kinds of deep learning methods. For example, the recurrent
1) The first generation of BDA neural network has been integrated with convolutional neural networks
to tackle the fault detection problems with time series data in the
741
Fig. 3. The development of big data analytics (adjusted from [38]).
intelligent maintain of a machine. theory and develop safe, trust worthy, reliable, and extensible AI tech
In the second generation of BDA, well designed complex models and nology. Hassabis et al. [42] argued that better understanding of bio
high-quality big data are the fundamental issues to succeed in logical brains could play a vital role in building intelligent machines. To
manufacturing systems. Nevertheless, the need for generalization and gain insights from bio-science for AI has become a consensus view.
interpretability represent substantial obstacles to the development and Shimon Ullman [43], director of Weizmann AI center, considered
translation of BIDA. The first important challenge in developing realiz bio-structure as a critical success factor. He thought combining deep
able and efficient systems is access to large, specific and well-annotated learning with brain-like innate structures may guide network models
datasets. Deep learning is unique in its capability of recognizing mean toward human-like learning. In industrial practice, the BDA model is
ingful patterns from raw data without explicit directions when provided learnt and evaluated during the interaction with the dynamic environ
with sufficient examples. Current published BDA works on the co- ments. The mechanisms and development of model structure topology
distribution hypothesis, where the distribution of training and testing and behavior may be an important issue to empower the next generation
data should be identical. However, the data collected from industrial of BDA [44].
practice is changing and unpredictable since manufacturing systems are
dynamic evolutionary. The second important challenge is that the inner 4. Key technologies of big data analytics
workings and decision-making processes of BDA algorithms remain
opaque. For the security considerations, the industrial applications 4.1. Computation framework
require decision support tools to explain the rationale or support for its
decisions to enable the users to independently review the basis of their With the development of data sensing and Internet of Things tech
recommendations. To meet this requirement and gain trust from in nologies, the scale of data continues to increase to reach TB, PB, and
dustrial users, BDA should provide explanations for their outputs. even higher levels, which can no longer be completed by a single com
Overcoming these two major barriers to industrial implementation can puter. This puts forward higher requirements for the computing power
facilitate the development and adoption of BDA into industrial practice. and timeliness of the computer to process big data. Big data processing
involves the configuration of the storage system, the division of
3) The next generation of BDA computing tasks, the distribution of computing load, the data migration
between computers, and the data security when the computer or
The next generation of BDA is placed great expectations to approach network fails. The situation is much more complicated. Distributed
the level of human intelligence with the development of AI. Zhang, one computing technology, in simple terms, is that multiple machines share
of the pioneers of AI in China, regards the integration of knowledge, computing tasks, which is the main solution at present. Furthermore, in
data, algorithm, and computation as the core driving force of the next some industrial scenarios, due to the high real-time data calculation and
generation of BDA [41]. He thought the combination of symbolism and analysis requirements of a large number of processes and equipment at
connectionism is imperative to establish a new, explainable, robust AI the bottom of the production process, the edge-cloud computing
742
architecture was born. e) After all nodes have performed the above steps, MapReduce starts
the Reduce task. Each group corresponds to a Reduce task.
1) Distributed technology f) The process executing the reduce task obtains all the key-value pairs
of the specified group through the network.
Around 2004, three papers introducing distributed file system GFS g) Combine values with the same key name into a list.
(Google file system) [45], parallel computing model (MapReduce) [46], h) Perform reduce task by processing the list corresponding to each key,
and non-relational data storage system (BigTable) [47] were published and then outputting the result.
by Google. This was the first time to propose a reusable solution for the
distributed processing of big data. Inspired by the ideas of Google, (2) The second generation of distributed computing framework
Yahoo engineers Doug and Mike developed Hadoop. Based on It can be seen that through the combination of multiple MapReduce,
improving Hadoop, dozens of big data computing frameworks applied in complex calculation problems can be expressed [51]. However, the
distributed environments have been born, as shown in Fig. 4. combination process requires manual design and the process is relatively
(1) The first generation of distributed computing framework cumbersome. In addition, all computers need to be synchronized at each
The first generation of distributed computing framework represented stage, which affects execution efficiency [52]. In order to overcome the
by Hadoop initially mainly included two parts: Hadoop distributed file above problems, a directed acyclic graph (henceforth referred to as
system (HDFS) and the computing framework (MapReduce) [48]. In DAG) calculation model is proposed [53]. The core idea of DAG is to
version 2.0, the resource management and task scheduling functions are decompose the task internally into a number of sequential subtasks,
stripped from MapReduce to form YARN [49], so that other frameworks which can express various complex dependencies more flexibly.
can also run on Hadoop like MapReduce. Compared with the previous Microsoft Dryad [54], Google FlumeJava [55], and Apache Tez [56] are
distributed computing framework, Hadoop hides many tedious details, the earliest DAG models. Dryad defines several simple DAG models such
such as fault tolerance, load balancing, etc., making it easier to use [50]. as series connection, full connection, and fusion, and describes complex
MapReduce can be understood as combining a bunch of chaotic data tasks by combining these simple structures. FlumeJava and Tez form
according to certain characteristics, and then processing and obtaining DAG model by combining several MapReduces.
the final result, as shown in Fig. 5. The basic processing steps are as Another shortcoming of MapReduce is the use of disks to store in
follows: termediate results, which seriously affects the performance of the sys
tem, which is more obvious in situations that require iterative
a) The input file is divided into pieces according to certain standards, calculations such as machine learning. Spark, developed by the AMP
and each piece corresponds to a map task. In general, MapReduce laboratory of the University of California, Berkeley, overcomes the
and HDFS run on the same set of computers, that is, each computer is above problems [57]. Spark has improved the early DAG model, and
responsible for storage and calculation tasks at the same time, proposed a memory-based distributed storage abstract model resilient
sharding usually does not involve data replication between distributed dataset (henceforth referred to as RDD), which selectively
computers. loads and resides in the intermediate data in the memory, reducing the
b) The content in the fragment is usually parsed into key-value pairs IO overhead of disks. Compared with Hadoop, memory-based Spark
according to a predefined certain standard. operations are more than 100 times faster, and disk-based operations are
c) Perform map tasks by processing each key-value pair, and then also more than 10 times faster.
outputting key-value pairs. (3) The third generation of distributed computing framework
d) MapReduce obtains the grouping method defined by the application Machine learning and artificial intelligence are also one of the trends
and sorts the key-value pairs output by the map task according to the in big data computing. Spark and Flink launched machine learning li
grouping. Each key name forms a group by default. braries Spark ML and Flink ML, respectively. More platforms provide
machine learning on third-party big data computing frameworks, such
Fig. 4. Classification of distributed computing frameworks.
743
Fig. 5. MapReduce processing process (adjusted from [48]).
as Mahout, Oryx and the Apache incubation project SystemML, Hive implemented Structured Streaming), the graph computing framework
Mall, PredictionIO, SAMOA, MADLib. These machine learning platforms GraphX, and the machine learning library Spark ML. While Flink pro
generally support multiple computing frameworks at the same time. For vides low-latency stream computing, batch processing, relational
example, Mahout uses Spark, Flink, and H2O as engines, while SAMOA computing, graph computing, and machine learning keep running, and
uses S4, Storm, and Samza. After deep learning has caused extensive the goal is to go straight to the big data general computing platform.
discussion, some communities have explored the combination of deep Google’s Beam is trying to incorporate computing frameworks such as
learning frameworks with existing distributed computing frameworks, Spark, Flink, and Apex under its own standards, which with the inten
such as projects such as SparkNet, Caffe on Spark, and TensorFrames tion of taking a monopoly on the market.
[58].
Supporting multiple frameworks on the same platform is also one of 2) Cloud- Edge computing technology
the development trends, especially for communities with strong devel
opment capabilities. Spark takes the batch processing model as its core, With the increasing pace of the manufacturing process, the timeli
and implements the interactive analysis framework Spark SQL, the ness requirements for data capturing and analysis are getting higher and
stream computing framework Spark Streaming (and the currently higher. Industrial applications such as real-time monitoring of the
Fig. 6. The framework of “plug and play” edge computing model (adjusted from [59]).
744
production process, online quality inspection, and decision-making Therefore, the deployment of edge computing in 5 G networks has
control all have high requirements for the timeliness of computing. become an important research area, and comprehensive solutions need
Under the high timeliness requirements of industrial applications, the to be developed to meet the challenges of heterogeneity, scalability,
traditional way of data capturing by sensors, realizing data transmission flexibility. Duan et al. [60] proposed a decentralized structure for ser
through industrial Ethernet, and realizing data analysis instruction vice orchestration/federation across admin-domains in the 5 GEx and 5
issuance in the enterprise private cloud cannot meet the requirements. G-Transformer projects, as shown in Fig. 7.
Edge computing, which is optimized by edge nodes, has obvious ad
vantages in low-latency applications. It is a powerful supplement to the
cloud computing system. 4.2. Data processing technology
(1) “plug and play” edge computing
Wang et al. [59] proposed a novel “plug and play” edge computing The importance of big data is not how much data you have, but what
mode, as shown in Fig. 6, which consisted of the edge layer, the PaaS you do with it. We can perform high-performance analysis on the
layer, and the application layer. In the special edge layer, each resource massive production process data to find answers to reduce costs, shorten
unit in workshop equipped with an industry driver for better accessi time, develop new products and optimize products, and make smart
bility. The industrial driver is an agent that contains the corresponding decisions. The mainstream data processing technology can be divided
hardware and software acting as an abstraction layer between the edge into batch processing and stream processing in terms of data processing
and cloud. Benefited by the great computing power from the cloud, the mode. The characteristics of batch processing and stream processing
industrial driver contained “3C-3I” structures: communication, compu were summarized and compared [61], as shown in Table 1.
tation, control, identification, insight and interoperation. Different from
the driver in operation systems, the driver of a resource unit had func 1) Batch processing framework
tions of computing, and insight, which enabled the industrial drivers to
monitor, analysis, and forecast the status of resource units to achieve Table 1
initiative, efficient optimization with low time latency. Batch processing vs stream processing (adjusted from [61]).
(b) Integration of 5 G network and edge computing Characteristics Batch Processing Stream Processing
The 5 G network provides an environment in which edge computing Data scope Queries or processing over Queries or processing over data
can be widely deployed, and has the advantages of ultra-low commu all or most of the data in within a rolling time window, or on
nication delay and high positioning accuracy. Therefore, the combina the dataset. just the most recent data record.
tion of 5 G network and edge computing may become a typical scenario Data size Large batches of data. Individual records or micro batches
consisting of a few records.
of edge-cloud convergence. Meanwhile, the unprecedented complexity
Performance Latencies in minutes to Requires latency in the order of
of 5 G network also brings challenges to its integration with edge hours. seconds or milliseconds.
computing. This complexity comes from many aspects, such as dense Analyses Complex analytics. Simple response functions,
network functions, heterogeneous, and highly diversified applications. aggregates, and rolling metrics.
Fig. 7. Decentralized structure for service orchestration/federation across admin-domains in the 5 GEx and 5 G-Transformer projects (adjusted from [60]).
745
Batch processing mainly operates on large static data sets and returns Continuous real-time data from external systems streams through these
the results after the computation is completed. Batch task can be inter nodes and concatenates them. If the data stream is compared to water,
preted as a collection of data points that have been grouped at a fixed the input node is like the sprinkler head, which keeps pouring water out,
time interval (each piece of data is called a data point and corresponds to and the computing node is like the interface of the water pipe.
the state information of a point in time). We can also call this a window In the stream computing framework, Storm [63] is one of the most
of data. popular and widely used. This is due to Storm’s simple programming
The data sets used in bath processing mode always conform to these model and support for Java, Ruby, Python, and other development
characteristics: languages. Storm also has good performance, handling millions of
messages per second on multi-node clusters. Storm’s solution is to assign
a) Bounded: A batch data set represents a finite set of data. an ID to each message as a unique identifier and include the ID of the
b) Persistent: Data is usually always stored in some type of persistent original input message in the message. The state of each original input
storage location. message is maintained with a response center (Acker) with an initial
c) Vast: Batch operations are often the only way to process extremely value of the ID of the original input message. After the successful
large data sets. execution of each computing node, the ids of the input and output
messages are xor, and then xor the state of the corresponding original
Batch processing is well suited for computations that require access input message. Since each message is xor once respectively during
to a complete set of records. For example, when calculating totals and generation and processing, all messages are xor twice after successful
averages, the data set must be treated as a whole rather than as a execution, and the corresponding state of the original input message is 0.
collection of multiple records. These operations require the data to Therefore, when the state is 0, the content of the original input message
maintain its own state as the calculation proceeds. can be safely cleared. If the state is not 0 after the specified time interval,
Tasks that require processing large amounts of data are often best it is considered that there is something wrong with the processing of the
handled by batch operations. Whether the data set is processed directly message and needs to be re-executed.
from the persistent storage device or loaded into memory first, the batch
processing system is designed with the amount of data in mind to pro 3) "Batch + stream" processing
vide sufficient processing resources. Because batch processing is
particularly good at handling large amounts of persistent data, it is often With the further development of big data, the separate batch pro
used to analyze historical data. cessing and stream processing framework could not fully meet the cur
The processing of large amounts of data requires a lot of time, so rent needs of the enterprise, hence the hybrid mode combined batch
batch processing is not suitable for situations with high processing time processing and stream processing was born, as shown in Fig. 8.
requirements. Batch tasks are suitable for the following situations: Apache Spark [64] is the typical framework for "batch + stream"
processing. Spark is an optimization based on the Hadoop MapReduce
a) Aggregate function operations are needed to run on data over a computing model. Spark greatly improves the processing capacity of
period of time, such as the average, maximum, and minimum values data through the in-memory computing model and execution optimi
of these data points. zation (in different cases, the speed can be 10–100 times faster than MR
b) The alert function does not need to run every data point, or the state or even higher). However, Spark’s Streaming capability is provided by
represented by the data, and does not change frequently. Spark Streaming module. Spark introduced the concept of Micro-batch,
c) Reduce the frequency of collection to a large extent (downsample). which treats access data over a short period of time as a single
Only the most obvious part of the huge data is needed to explain the micro-batch. However, compared with native stream processing systems
change of the overall state. such as Storm, Spark Streaming has a relatively high latency.
d) A little extra latency won’t have a big impact on your overall Apache Flink [65] also supports stream processing and batch pro
business. cessing. Flink’s design idea is "stateful stream processing", which treats
e) Kapacitor processes data slower than it writes data when the cluster input data item by item as a real stream and batch tasks as a bounded
is running the timing database. stream. In the current field of streaming data processing framework,
Flink is unique. While Spark also provides batch and stream processing
Apache Hadoop and its MapReduce processing engine provide a capabilities, the micro-batch architecture of Spark stream processing
proven batch processing model that is best suited for handling very large makes its response time slightly longer. Flink stream-first approach
data sets with low time requirements [62]. A fully functional Hadoop achieves low latency, high throughput and true item-by-item processing,
cluster can be built with very low-cost components, making this cheap which is why Flink has received more and more attention in recent
and efficient processing technique flexible for many cases. Compatibility years.
and integration with other frameworks and engines make Hadoop the
underlying foundation for a variety of workload processing platforms
using different technologies.
2) Stream processing
In the era of big data, data is usually generated continuously and

dynamically. In many cases, data needs to be processed in a very short
period of time, and fault tolerance, congestion control and other issues
need to be considered to avoid data omission or double-counting.
Stream processing is the solution to this kind of problem. DAG
(directed acyclic graph) model is generally used in stream processing
framework. The nodes in the figure are divided into two categories. One
is the input node of data, which is responsible for interacting with the
outside world and providing data to the system. The other is the
computing node of the data, which is responsible for performing certain
processing functions such as filtering, accumulating, merging, etc. Fig. 8. "Batch + stream" processing.
746
5. Application in manufacturing systems parameters [68]. Tao [69] analyzed the isolation of data in the product
life cycle management under the big data environment, and proposed a
Driven by intelligent sensing, Internet of Things (IoT), distributed digital twin-driven model of product design, manufacturing and service.
storage computing, machine learning and other technologies, big data- Geiger [35] analyzed the reliability of components in automobile design
driven intelligent manufacturing applications have begun to emerge. by using field measurement data and user usage data to improve design
As is shown in Fig. 9. the applications in product design, planning and parameters and enhance product reliability. Tucker [70] proposed big
scheduling, quality optimization, equipment operation and maintenance data-driven product optimization method, in which the design decision
have been carried out to substantialize the intelligent manufacturing tree model was used to realize the combinatorial optimization of prod
systems. ucts. Qin [71] applied big data analyzing technology to identify key
parameters affecting diesel engine power by correlation analysis which
5.1. Big data-driven product design effectively reduced the cost of design.
Big data provides a mechanism of fine design with streamlined
Design plays an important role in manufacturing process, which design processes, promoted product innovations, and developed
determines most of a product’s performance and performance [12]. customized products by researching and understanding customer de
With the development of big data technology, product design is shifting mands, behaviors, and preferences [72]. Ireland [73] brought Internet
towards data-driven design from subjective conceptual design [66]. Big user evaluation into product design to achieve a quantitative analysis of
data-driven product design analyzes the market demand through users’ product functions and provided support for design decisions. Xu [74]
evaluation [67]. Then, it generates the design scheme through the sale, proposed a framework of data-driven product design for capturing
customer, manufacturing data, and enhances the decision-making abil product visual aesthetics UX to effectively identify the useful design
ity based on history learning. Finally achieves the closed-loop optimi concepts from consumer preferences to consumer response. Promotion
zation of self-monitoring, self-analyzing, self-learning and strategies were developed to suit for corresponding segments of different
self-adaptation [37]. Intelligent design can be divided into two aspects customers. Li [75] designed a framework of smart product. In this
according to the design goals, one is product design for quality framework, pairwise comparative algorithm was employed to cluster
improvement, and the other is humanized shape design focused on technical attributes into modules assembled into smart products. Wire
customer preferences. less sensor network was developed to monitor smart product, and
In the field of product design, a new digitized and parameterized K-means algorithm was adopted to deal with monitoring data. Trian
perspective are shaped in the era of big data. The main factors affecting gular fuzzy numbers were used to evaluate the maturity of the moni
the quality can be obtained by analyzing various types of process data toring function. Finally, the design of control function was determined
through correlation analysis, and established the mapping model of to give the optimal strategy by Q-learning algorithm to realize the closed
quality influencing factors to effectively optimize the product design loop optimization of design [76].
Fig. 9. The applications of BDA in manufacturing systems.
747
5.2. Big data-driven planning and scheduling dependence but also noises caused by transitive effects of correlations.
As a result, the analyzed correlation may be heightened to mislead the
Production planning and scheduling is one of the core issues of en key factor identification. To exactly identify the root cause factors, the
terprise operation and management [77]. In the ideal scenario, a feasible causal inference embedded BDA could be further addressed to improve
schedule can efficiently allocate production resources while coping with the robustness and accuracy of BDA.
the impact of various on-site uncertainties. However, traditional shop In the self-adaptive optimization and control of product quality, BDA
floor scheduling methods mainly focus on the optimization of workshop models of product quality are constructed to find the relationship be
performance and resource utilization in static environment, which may tween the raw collected data and product quality [91,92]. Gustavo et al.
result in deterioration or infeasibility of static schedule. Big data ana used data mining methods to predict 8 diesel quality performance pa
lytics provide a series of effective methods and tools for the optimization rameters, which significantly reduced the detection time and cost [93].
of production planning and scheduling in dynamic environment [78]. Rokach L et al. applied data mining methods to improve the quality of
BDA can improve the generation of scheme in the planning and the manufacturing process, and achieved good results in integrated
scheduling. A large amount of historical data and real-time information circuit manufacturing [94]. Nakazawa et al. designed a five-layer con
of the production process should be collected through various enterprise volutional network to achieve accurate recognition of 22 defect patterns
information systems, embedded sensors and intelligent machines, such on wafers [94]. Kiryong et al. designed a sub-fully connected layer
as the processing time of jobs, the setup time of machines, and the network for each defect mode by sharing the convolution feature
transportation time of materials [79–81]. Based on the collected data, extraction layer. Each sub-network judged whether the defect mode
big data analytics can provide comprehensive information support for existed in two categories, which effectively identified the defect mode of
scheduling decisions. For example, before making production schedules, the wafer map [95]. Current methods performed well with complete,
the distribution information of various uncertain factors can be accu clean, and balanced data. The quality prediction or defect detection
rately estimated by statistical analysis techniques. After that, robust methods with imbalanced data [24], inequivalent data [96] is still a
optimization [82,83] and intelligent optimization [84,85] algorithms concern. In addition, the model is susceptible to input noise, and it is
are used to formulate robust production plans and schedules under difficult to obtain high accuracy. For example, detecting product quality
uncertain environment. with noisy mixed- defects is a difficult issue [97]. At the same time,
In the execution process of a scheme, by extracting the current status when the process parameters change, the special detection model that
information of workshop, such as manufacturing resource status, pro relies on data distribution training cannot dynamically adapt to the data
duction progress and inventory data, deep learning and other machine change, resulting in model failure. Models in different case scenarios
learning methods [86] can be used to predict the completion time of cannot be used universally so that the scene detection cannot be copied
products, so as to evaluate the production capacity. By establishing the quickly, which influences the wide application of BDA in actual in
risk probability evaluation model of schedules, high-risk tasks can be dustries. The generalized BDA will be a next focused area of quality
identified and reallocated [87]. Finally, the complete actual operation forecasting research.
data and the execution results of dispatching rules provide useful Controlling product quality is an essential part of the development
knowledge about planning and scheduling, which can be used to and production process. Jin et al. proposed a feedforward control
establish reliable and practical heuristic scheduling rules based on method based on the picewise linear model [93]. The designed
learning strategies [88]. In addition, the production process can be engineering-driven reconstruction method of the picewise linear
optimized by analyzing the correlation of production parameters in regression tree further reduced the complexity of the model by merging
historical data and the influence on production performance. leaf nodes under the constraints of control accuracy requirements. The
effectiveness of the proposed method was verified in a multi-stage wafer
5.3. Big data-driven quality management manufacturing process. Borja et al. used Reinforcement Learning (RL)
analysis solutions to derive the required feedback controller for the ball
Big data-driven product quality management realizes product screw feed drive to provide positioning accuracy. The proposed method
traceability and optimization based on product manufacturing process performed better than PID controller in computational experiments
data and quality inspection data [89]. The main factors affecting prod [98]. Liu et al. regarded the feature tensors extracted by CNN as inter
uct quality most, such as raw material performance parameters, equip dependent feature sequences in the process of molten pool state recog
ment status parameters, process parameters, and workshop nition, and used LSTM to mine the relationship between the feature
environmental parameters can be identified by correlation analysis. sequences to realize the effective integration of redundant features
Meanwhile, a mapping model between quality influencing factors and extracted by CNN [99]. Liu et al. designed a coarse-grained regulariza
quality performance can be established to effectively predict product tion method in view of the small sample problem faced in the process of
quality. Intelligent optimization algorithms are further used to adap molten pool state recognition based on deep learning, considering the
tively adjust control parameters that affect product quality in real time characteristics of the convolution operation, which effectively pre
to achieve adaptive control and optimization of product quality. vented the model from overfitting [100]. Chien et al. used a neural
Identifying the key influencing parameters affecting the product network model to fit the relationship between WIP parameters, job size,
quality from mass manufacturing parameters plays an important role as the length of the waiting queue before the bottleneck station, the length
the input of the product quality control model. Chien et al. screened out of the waiting queue on the process route and the construction period to
12 highly correlated WAT factors through expert experience as model form a schedule control strategy [100]. Chen et al. used the
input, and designed an analysis method based on modified Partial Least Gauss-Newton regression model to measure the influence of the average
Square (mPLS) to screen key parameters [90]. This method required the number of layers, production capacity, movement, WIP and other pa
expert experience to select key parameters, which was difficult for rameters on the construction period, and realized the construction
quality analysts without experience to quickly master this skill. Qin et al. period control on this basis [101]. In the actual product manufacturing
used bench test data in the quality control of diesel engines to perform process, product quality often needs to be measured by multiple quality
correlation analysis on potential parameters that affect power consis indicators, each of which is related to each other and involves multiple
tency, and effectively improved power consistency [71]. These methods processes. The cooperative control with different variable controlled
analyze the correlation between candidate factors and the quality in variables in the integration multi-process with design, manufacturing
dicators to reveal the root cause of fluctuation. However, the correlation and assembly is still an open issue.
evaluated by current BDA is different from the true situation under some
conditions, since the observed correlation contains not only true
748
5.4. Big data-driven prognostics health management 6.2. Cognitive security of big data analytics
Big data-driven prognostics health management (PHM) reveals the The security of big data analytics in manufacturing systems is
time-series changes in system fault characteristics through real-time another major concern in the application. Although great achievements
monitoring of manufacturing process data, equipment performance have been made, the overall performance of AI for BDA is sometimes
parameters and other time-series data. The PHM data is fully analyzed to surprising. The Google brain team in California observed an interesting
proactively identify potential anomalies in the system operation process phenomenon named antagonistic aggression, which describes that the
in advance, diagnose the root cause of anomalies, and forecast the current deep learning model is easy to be confused in an image classi
remaining life of a system [102]. fication task by added a small fluorescent sticker in the corner of the
The analyzed PHM data consists of acceleration signal, acoustic image [113]. Researchers at Kyushu University in Japan found that by
emission signals, optical signals, electrical data, temperature data, etc. changing only one pixel on an image, a neural network could be tricked
In the PHM of a rotating machinery, Sinha et al. analyzed the vibration during an image recognition task. Experiments on cifar-10 and Image
signals collected by an acceleration sensor [103]. For complicated ma Net data sets achieved a deception success rate of 68.36 % and 41.22 %,
chinery, Lei et al. acquired and integrated several acceleration sensors to respectively. This technique is therefore known as "One Pixel Attack".
monitor the health condition of planetary gearboxes [104]. In the PHM These cases essentially reflect that learning from massive data to
of high-speed moving machinery, Zhang et al. extended real-time improve the accuracy through black-box models is insufficient [114]. In
nearfield acoustic holography (RT-NAH) to reconstruct the instanta manufacturing systems, the BDA is a part of the human cyber physical
neous surface normal velocity of a vibrating structure from the systems (HCPS), which requires high reliability and security. For BDA, it
time-dependent pressure measured in the near field, which provides a is essential to develop the cognitive security which ensures AI models
non-contact and real-time method to measure and visualize the transient operating properly from the attack of fake content, interference signals
vibration of the structure [105]. Acoustic emission is another signal and other adversarial issues.
widely used in the PHM of slew bearing [106], cutting tool, and struc In this way the interpretability of machine learning models should be
ture health monitoring applications. From the perspective of data, most further concerned [115]. At present, deep neural networks obtain high
of the current research using data within a single media. The discrimination power at the cost of low interpretability of their
cross-media BDA extracting information from different medias will be black-box representations [116]. In industrial scenarios, the trans
an important issue in the PHM. parency, and interpretability are required for security issues. The visu
Before the development of deep learning for BDA, the PHM methods alization of black-box models is a practical approach to understand the
make decisions based on pre-extracted features. There are several crit working process of deep learning methods. Another approach is based
ical and commonly used features, including time-frequency domain- on statistical tools [117]. The sensitive analysis of different parameters
based feature, spectral cliff-based feature, higher-order statistical and the role of the subcomponent can be investigated by knowledge
variable-based feature, and random resonance-based feature. The BDA distillation and network pruning [118].
merges two steps together, since deep learning methods can automati
cally extract features from raw time-domain signals [107]. Zheng et al. 6.3. Industrial big data governance
developed a transfer locality preserving projection based intelligent
fault identification method, which embeds the data to a subspace Governance of big data handles data integrity, quality, provenance,
through preserving a priori distribution structure properties and trains a retention, processing, and analysis in full data lifecycle [119]. The
classifier to identify the condition of target machine by the historical governance of industrial big data considers the issues of security and
data [108]. To train the PHM model adaptively, Wen et al. developed a privacy [120]. In the BDA, data from all sections are combined to form
novel learning rate scheduler based on reinforcement learning for con an integrated environment for system optimization [121]. In the ma
volutional neural network in fault classification, which can schedule the jority of manufacturing organizations, raw data is not allowed to be
learning rate efficiently and automatically [109]. Lei’s group designed a transmitted to a remote commercial cloud center directly, since private
feature-based transfer neural network to learn transferable features from raw data can easily be copied during storage and transmission. The
data collected in a laboratory to achieve higher diagnosis accuracy for Blockchain providing cybersecurity [122] may be a novel method for
real-case machines [110]. Current BDA provides an end-to-end archi industrial big data governance. The architecture, mechanisms, tech
tecture to train PHM models with massive data directly. In industrial nologies, and applications for big data governance keeping the privacy
practice, the features, and potential abnormal signals, are changing with and security during big data sharing and integration is another open
the dynamic operation environment. The BDA can discover knowledge issue in the BDA.
about PHM with a self-learning mechanism to overcome the unpre
dictable scenarios. In the future, the end-to-end architecture may be 7. Conclusion
improved with interpretable transformable quantizable learning tech
nologies to substantialize a hybrid model. As increasing attention is given to intelligent manufacturing systems,
BDA is becoming a core technology to provide forecasting and decision
6. Challenges for future research making in manufacturing systems. BDA is a key future perspective in
both research and industrial community, as it provides added value to
6.1. Knowledge embedded industrial big data analytics various products and systems by applying cutting-edge technologies to
traditional products in manufacturing. Key concepts, frameworks, key
The BDA was considered as one of the most important technologies, technologies, applications are discussed in this paper. Open issues for
due to its capacity to explore large and varied datasets to uncover hidden future research are highlighted after a systematic review. We hope this
patterns and knowledge as well as other useful information [111]. The paper helps to promote the theoretical and technical research and en
discovered patterns and knowledge can help manufacturing systems to gineering application of big data-driven intelligent manufacturing sys
make more-informed decisions, and to achieve the whole lifecycle tems [123].
optimization and more sustainable production [112]. The integration of
current knowledge and the discovered patterns and knowledge from big Declaration of Competing Interest
data will be an interesting issue. From another perspective, the reuse
discovered patterns and knowledge from big data into BDA process will The authors declare that they have no known competing financial
improve the performance of BDA methods. interests or personal relationships that could have appeared to influence
749
the work reported in this paper. [24] Wang J, Yang Z, Zhang J, Zhang Q, Chien W-TK. AdaBalGAN: an improved
generative adversarial network with imbalanced learning for wafer defective
pattern recognition. IEEE Trans Semicond Manuf 2019;32(3). https://doi.org/
Acknowledgements 10.1109/TSM.2019.2925361.
[25] Sun Y, Kamel MS, Wong AKC, Wang Y. Cost-sensitive boosting for classification of
This work was supported in part by National Natural Science Foun imbalanced data. Pattern Recognit 2007;40(12). https://doi.org/10.1016/j.
patcog.2007.04.009.
dation of China under Grant 51905091, and in part by Shanghai Sailing [26] Kang Q, Shi L, Zhou MC, Wang XS, Di Wu Q, Wei Z. A distance-based weighted
Program under Grant 19YF1401500, Shanghai Science and Technology undersampling scheme for support vector machines and its application to
Planning Project under Grant 20DZ2251400. The authors want to thank imbalanced classification. IEEE Trans Neural Networks Learn. Syst 2018;29(9).
https://doi.org/10.1109/TNNLS.2017.2755595.
Mr. Peng Zheng, Mr. Jie Ren, Mr. Shuxuan Zhao, Mr. Pengjie Gao, and [27] Zhang Y, Li X, Gao L, Wang L, Wen L. Imbalanced data fault diagnosis of rotating
Ms. Zixun Zhu for their help in literature collection. machinery using synthetic oversampling and feature learning. J Manuf Syst 2018;
48(August 2017):34–50. https://doi.org/10.1016/j.jmsy.2018.04.005.
[28] Fernandez V. Wavelet- and SVM-based forecasts: an analysis of the U.S. Metal and
References materials manufacturing industry. Resour Policy 2007;32(1–2):80–9. https://doi.
org/10.1016/j.resourpol.2007.06.002.
[1] Gao RX, Wang L, Helu M, Teti R. Big data analytics for smart factories of the [29] Vafeiadis T, Krinidis S, Ziogou C, Ioannidis D, Voutetakis S, Tzovaras D. Robust
future. CIRP Ann 2020;69(2):668–92. https://doi.org/10.1016/j. malfunction diagnosis in process industry time series. IEEE International
cirp.2020.05.002. Conference on Industrial Informatics (INDIN) 2016. https://doi.org/10.1109/
[2] Leng J, et al. A loosely-coupled deep reinforcement learning approach for order INDIN.2016.7819143.
acceptance decision of mass-individualized printed circuit board manufacturing [30] Villalobos K, Diez B, Illarramendi A, Go i A, Blanco JM. I4TSRs: a system to assist
in industry 4.0. J Clean Prod 2021;280. https://doi.org/10.1016/j. a data engineer in time-series dimensionality reduction in industry 4.0 scenarios.
jclepro.2020.124405. 2018. https://doi.org/10.1145/3269206.3269213.
[3] Wang G, Gunasekaran A, Ngai EWT, Papadopoulos T. Big data analytics in [31] Seçkin M, Seçkin AÇ, Coşkun A. Production fault simulation and forecasting from
logistics and supply chain management: certain investigations for research and time series data with machine learning in glove textile industry. J Eng Fiber Fabr
applications. Int J Prod Econ 2016;176:98–110. https://doi.org/10.1016/j. 2019;14. https://doi.org/10.1177/1558925019883462.
ijpe.2016.03.014. [32] Sagheer A, Kotb M. Time series forecasting of petroleum production using deep
[4] Zhang Z, et al. Pathologist-level interpretable whole-slide cancer diagnosis with LSTM recurrent networks. Neurocomputing 2019;323. https://doi.org/10.1016/
deep learning. Nat Mach Intell 2019;1(May). https://doi.org/10.1038/s42256- j.neucom.2018.09.082.
019-0052-1. [33] Benosman M. Model-based vs data-driven adaptive control: an overview. Int J
[5] Parisi GI, Kemker R, Part JL, Kanan C, Wermter S. Continual lifelong learning Adapt Control Signal Process 2018;32(5):753–76. https://doi.org/10.1002/
with neural networks : a review. Neural Netw 2019;113:54–71. https://doi.org/ acs.2862.
10.1016/j.neunet.2019.01.012. [34] Wang Q, Li F, Tang Y, Xu Y. Integrating model-driven and data-driven methods
[6] Kusiak A. Smart manufacturing must embrace big data. Nature 2017;544(7648). for power system frequency stability assessment and control. IEEE Trans Power
https://doi.org/10.1038/544023a. Syst 2019;34(6):4557–68. https://doi.org/10.1109/TPWRS.2019.2919522.
[7] Zhong RY, Newman ST, Huang GQ, Lan S. Big Data for supply chain management [35] Geiger C, Sarakakis G. Data driven design for reliability. Proceedings - Annual
in the service and manufacturing sectors: challenges, opportunities, and future Reliability and Maintainability Symposium, 2016-April; 2016. https://doi.org/
perspectives. Comput Ind Eng 2016;101. https://doi.org/10.1016/j. 10.1109/RAMS.2016.7448023.
cie.2016.07.013. [36] Hey T, Tansley S, Tolle K. The fourth paradigm: data-intensive scientific
[8] Dutton W, et al. Clouds, big data, and smart assets: ten tech-enabled business discovery. Microsoft Research. 2009.
trends to watch. McKinsey Q 2010;(4). [37] Zhang J, Wang J, Lyu Y, Bao J. Big data driven intelligent manufacturing.
[9] Yager RR. A framework for multi-source data fusion. Inf Sci (Ny) 2004;163(1–3). Zhongguo Jixie Gongcheng/China Mech Eng 2019;30(2). https://doi.org/
https://doi.org/10.1016/j.ins.2003.03.018. 10.3969/j.issn.1004-132X.2019.02.001.
[10] Zhu K, Li G, Zhang Y. Big data oriented smart tool condition monitoring system. [38] Wang L. From intelligence science to intelligent manufacturing. Engineering
IEEE Trans Ind Inform 2019;16(6):1. https://doi.org/10.1109/tii.2019.2957107. 2019;5(4):615–8. https://doi.org/10.1016/j.eng.2019.04.011.
[11] Zhang J. Multi-source remote sensing data fusion: status and trends. Int J Image [39] Ginsberg J, Mohebbi MH, Patel RS, Brammer L, Smolinski MS, Brilliant L.
Data Fusion 2010;1(1). https://doi.org/10.1080/19479830903561035. Detecting influenza epidemics using search engine query data. Nature 2009;457
[12] Tao F, Qi Q, Liu A, Kusiak A. Data-driven smart manufacturing. Int J Ind Manuf (7232):1012–4. https://doi.org/10.1038/nature07634.
Syst Eng 2018;48:157–69. https://doi.org/10.1016/j.jmsy.2018.01.006. [40] Silver D, et al. Article mastering the game of Go without human knowledge. Nat
[13] Zhang Y, Ren S, Liu Y, Si S. A big data analytics architecture for cleaner Publ Gr 2017;550(7676):354–9. https://doi.org/10.1038/nature24270.
manufacturing and maintenance processes of complex products. J Clean Prod [41] Zhang B, Zhu J, Su H. Toward the third generation of artificial intelligence. Sci
2017;142. https://doi.org/10.1016/j.jclepro.2016.07.123. Sin Informationis 2020;50(9):1281–302. https://doi.org/10.1360/SSI-2020-
[14] Wang J, Zheng P, Zhang J. Big data analytics for cycle time related feature 0204.
selection in the semiconductor wafer fabrication system. Comput Ind Eng 2020; [42] Hassabis D, Kumaran D, Summerfield C, Botvinick M. Neuroscience-inspired
143(February):106362. https://doi.org/10.1016/j.cie.2020.106362. artificial intelligence. Neuron 2017;95(July (2)):245–58. https://doi.org/
[15] Wang J, Yang J, Zhang J, Wang X, Zhang W (Chris). Big data driven cycle time 10.1016/j.neuron.2017.06.011.
parallel prediction for production planning in wafer manufacturing. Enterp Inf [43] Ullman S. Using neuroscience to develop artificial intelligence. Science (80-)
Syst 2018;12(6):714–32. https://doi.org/10.1080/17517575.2018.1450998. 2019;363(6428):692–3. https://doi.org/10.1126/science.aau6595.
[16] Lei Y, Jia F, Lin J, Xing S, Ding SX. An intelligent fault diagnosis method using [44] Rahwan I, et al. Machine behaviour. Nature 2019;568(7753):477–86. https://doi.
unsupervised feature learning towards mechanical big data. IEEE Trans Ind org/10.1038/s41586-019-1138-y.
Electron 2016;63(5). https://doi.org/10.1109/TIE.2016.2519325. [45] Ghemawat S, Gobioff H, Leung ST. The google file system. Operating Syst Rev
[17] Liu C, Li Y, Shen W. Integrated manufacturing process planning and control based (ACM) 2003;37(5). https://doi.org/10.1145/1165389.945450.
on intelligent agents and multi-dimension features. Int J Adv Manuf Technol [46] Dean J, Ghemawat S. MapReduce: simplified data processing on large clusters.
2014;75(9–12). https://doi.org/10.1007/s00170-014-6246-0. Commun ACM 2008;51(1). https://doi.org/10.1145/1327452.1327492.
[18] Luo Y, Zhang L, Zhang K, Tao F. Research on the knowledge-based multi- [47] Chang F, et al. Bigtable: A distributed storage system for structured data. ACM
dimensional information model of manufacturing capability in CMfg. Adv Mater Trans Comput Syst 2008;26(2). https://doi.org/10.1145/1365815.1365816.
Res 2012;472–475. https://doi.org/10.4028/www.scientific.net/amr.472- [48] Abouzeid A, Bajda-Pawlikowski K, Abadi D, Silberschatz A, Rasin A. HadoopDB:
475.2592. an architectural hybrid of mapreduce and DBMS technologies for analytical
[19] Xu C, Wang J, Zhang J. Forecasting the power consumption of a rotor spinning workloads. Proc VLDB Endow. 2009;2(1). https://doi.org/10.14778/
machine by using an adaptive squeeze and excitation convolutional neural 1687627.1687731.
network with imbalanced data. J Clean Prod 2020:122864. https://doi.org/ [49] Vavilapalli VK, et al. Apache hadoop YARN: yet another resource negotiator.
10.1016/j.jclepro.2020.122864. 2013. https://doi.org/10.1145/2523616.2523633.
[20] O’Donovan P, Leahy K, Bruton K, O’Sullivan DTJ. Big data in manufacturing: a [50] Huang J, Nicol DM, Campbell RH. Denial-of-service threat to hadoop/YARN
systematic mapping study. J Big Data 2015;2(1). https://doi.org/10.1186/ clusters with multi-tenancy. 2014. https://doi.org/10.1109/BigData.
s40537-015-0028-x. Congress.2014.17.
[21] Lee DH, Yang JK, Lee CH, Kim KJ. A data-driven approach to selection of critical [51] Ranger C, Raghuraman R, Penmetsa A, Bradski G, Kozyrakis C. Evaluating
process steps in the semiconductor manufacturing process considering missing MapReduce for multi-core and multiprocessor systems. 2007. https://doi.org/
and imbalanced data. J Manuf Syst 2019;52. https://doi.org/10.1016/j. 10.1109/HPCA.2007.346181.
jmsy.2019.07.001. [52] Stuart JA, Owens JD. Multi-GPU MapReduce on GPU clusters. 2011. https://doi.
[22] Zhao M, Kang M, Tang B, Pecht M. Deep residual networks with dynamically org/10.1109/IPDPS.2011.102.
weighted wavelet coefficients for fault diagnosis of planetary gearboxes. IEEE [53] Bonifaci V, Marchetti-Spaccamela A, Stiller S, Wiese A. Feasibility analysis in the
Trans Ind Electron 2018;65(5). https://doi.org/10.1109/TIE.2017.2762639. sporadic DAG task model. 2013. https://doi.org/10.1109/ECRTS.2013.32.
[23] He H, Garcia EA. Learning from imbalanced data. IEEE Trans Knowl Data Eng [54] Isard M, Budiu M, Yu Y, Birrell A, Fetterly D. Dryad: distributed data-parallel
2009;21(9). https://doi.org/10.1109/TKDE.2008.239. programs from sequential building blocks. Oper Syst Rev 2007:59–72. https://
doi.org/10.1145/1272996.1273005.
750
[55] Chambers C, et al. FlumeJava: Easy, efficient data-parallel pipelines. ACM uncertainty. Int J Comput Integr Manuf 2020. https://doi.org/10.1080/
SIGPLAN Notices 2010;45(6). https://doi.org/10.1145/1809028.1806638. 0951192X.2020.1803506.
[56] Saha B, Shah H, Seth S, Vijayaraghavan G, Murthy A, Curino C. Apache tez: a [84] Wang B, Wang X, Xie H. Bad-scenario-set robust scheduling for a job shop to
unifying framework for modeling and building data processing applications. hedge against processing time uncertainty. Int J Prod Res 2019;57(10). https://
Proceedings of the ACM SIGMOD International Conference on Management of doi.org/10.1080/00207543.2018.1555650.
Data, 2015-May; 2015. https://doi.org/10.1145/2723372.2742790. [85] Gao D, Wang G-G, Pedrycz W. Solving fuzzy job-shop scheduling problem using
[57] Zaharia M, Chowdhury M, Franklin MJ, Shenker S, Stoica I. Spark: cluster DE algorithm improved by a selection mechanism. IEEE Trans Fuzzy Syst 2020.
computing with working sets. 2010. https://doi.org/10.1109/tfuzz.2020.3003506.
[58] Lyubimov D, Palumbo A. Apache mahout: beyond MapReduce. CreateSpace [86] Wang J, Zhang J. Big data analytics for forecasting cycle time in semiconductor
Independent Publishing Platform; 2016. wafer fabrication system. Int J Prod Res 2016;54(23):7231–44. https://doi.org/
[59] Wang J, Xu C, Zhang J, Bao J, Zhong R. A collaborative architecture of the 10.1080/00207543.2016.1174789.
industrial internet platform for manufacturing systems. Robot Comput Integr [87] Ji W, Wang L. Big data analytics based fault prediction for shop floor scheduling.
Manuf 2020;61(April 2019):101854. https://doi.org/10.1016/j. J Manuf Syst 2017;43:187–94. https://doi.org/10.1016/j.jmsy.2017.03.008.
rcim.2019.101854. [88] Jun S, Lee S, Chun H. Learning dispatching rules using random forest in flexible
[60] Duan Q, Wang S, Ansari N. Convergence of networking and cloud/edge job shop scheduling problems. Int J Prod Res 2019;57(10). https://doi.org/
computing: status, challenges, and opportunities. IEEE Netw 2020;34(6). https:// 10.1080/00207543.2019.1581954.
doi.org/10.1109/MNET.011.2000089. [89] Gao Y, GAO L, Li X. A generative adversarial network-based deep learning
[61] Shahrivari S. Beyond batch processing: towards real-time and streaming big data. method for low-quality defect image reconstruction and recognition. IEEE Trans
Computers 2014;3(4):117–29. Ind Informatics 2020. https://doi.org/10.1109/tii.2020.3008703.
[62] Inoubli W, Aridhi S, Mezni H, Maddouri M, Mephu Nguifo E. An experimental [90] Chien CF, Lee PC, Dou R, Chen YJ, Chen CC. Modeling collinear WATs for
survey on big data frameworks. Future Gener Comput Syst 2018;86. https://doi. parametric yield enhancement in semiconductor manufacturing. IEEE
org/10.1016/j.future.2018.04.032. International Conference on Automation Science and Engineering, 2017 August;
[63] Christensen R, Wang L, Li F, Yi K, Tang J, Villa N. STORM: spatio-temporal online 2017. https://doi.org/10.1109/COASE.2017.8256192.
reasoning and management of large spatio-temporal data. Proceedings of the [91] Yiping G, Xinyu L, Gao L. A deep lifelong learning method for digital-twin driven
ACM SIGMOD International Conference on Management of Data, 2015-May; defect recognition with novel classes. J Comput Inf Sci Eng 2021;21(3). https://
2015. https://doi.org/10.1145/2723372.2735373. doi.org/10.1115/1.4049960.
[64] Zaharia M, Das T, Li H, Hunter T, Shenker S, Stoica I. Discretized streams: fault- [92] Miao R, Gao Y, Ge L, Jiang Z, Zhang J. Online defect recognition of narrow
tolerant streaming computation at scale. 2013. https://doi.org/10.1145/ overlap weld based on two-stage recognition model combining continuous
2517349.2522737. wavelet transform and convolutional neural network. Comput. Ind. 2019;112:
[65] Carbone P, et al. Apache FlinkTM: Stream and Batch Processing in a Single 103115. https://doi.org/10.1016/j.compind.2019.07.005.
Engine. 2015. p. 28–38. [93] Jin R, Shi J. Reconfigured piecewise linear regression tree for multistage
[66] Kusiak A, Salustri FA. Computational intelligence in product design engineering: manufacturing process control. IIE Trans (Institute Ind Eng) 2012;44(4). https://
review and trends. IEEE Trans Syst Man Cybern Part C 2007;37(5). https://doi. doi.org/10.1080/0740817X.2011.564603.
org/10.1109/TSMCC.2007.900669. [94] Nakazawa T, Kulkarni DV. Wafer map defect pattern classification and image
[67] Qi J, Zhang Z, Jeon S, Zhou Y. Mining customer requirements from online retrieval using convolutional neural network. IEEE Trans Semicond Manuf 2018;
reviews: a product improvement perspective. Inf Manag 2016;53(8). https://doi. 31(2). https://doi.org/10.1109/TSM.2018.2795466.
org/10.1016/j.im.2016.06.002. [95] Kyeong K, Kim H. Classification of mixed-type defect patterns in wafer bin maps
[68] Kusiak A. Innovation: a data-driven approach. Int J Prod Econ 2009;122(1). using convolutional neural networks. IEEE Trans Semicond Manuf 2018;31(3).
https://doi.org/10.1016/j.ijpe.2009.06.025. https://doi.org/10.1109/TSM.2018.2841416.
[69] Tao F, Cheng J, Qi Q, Zhang M, Zhang H, Sui F. Digital twin-driven product [96] Wang J, Xu C, Dai L, Zhang J, Zhong R. An unequal deep learning approach for
design, manufacturing and service with big data. Int J Adv Manuf Technol. 2018; 3D point cloud segmentation. IEEE Trans Ind Informatics 2020;3203(c):1–11.
94(9–12). https://doi.org/10.1007/s00170-017-0233-1. https://doi.org/10.1109/TII.2020.3044106.
[70] Tucker CS, Kim HM. Data-driven decision tree classification for product portfolio [97] Wang J, Xu C, Yang Z, Zhang J, Li X. Deformable convolutional networks for
design optimization. J Comput Inf Sci Eng 2009;9(4). https://doi.org/10.1115/ efficient mixed-type wafer defect pattern recognition. IEEE Trans Semicond
1.3243634. Manuf 2020;33(4):587–96. https://doi.org/10.1109/TSM.2020.3020985.
[71] Qin W, Zha D, Zhang J. An effective approach for causal variables analysis in [98] Fernandez-Gauna B, Ansoategui I, Etxeberria-Agiriano I, Graña M. Reinforcement
diesel engine production by using mutual information and network learning of ball screw feed drive controllers. Eng Appl Artif Intell 2014;30.
deconvolution. J Intell Manuf 2020;31(7). https://doi.org/10.1007/s10845-018- https://doi.org/10.1016/j.engappai.2014.01.015.
1397-8. [99] Liu T, Bao J, Wang J, Zhang Y. A hybrid CNN− LSTM algorithm for online defect
[72] Wu D, Liu X, Hebert S, Gentzsch W, Terpenny J. Democratizing digital design and recognition of CO₂welding. Sensors (Basel) 2018;18(12). https://doi.org/
manufacturing using high performance cloud computing: performance evaluation 10.3390/s18124369.
and benchmarking. J Manuf Syst 2017;43. https://doi.org/10.1016/j. [100] Chien CF, Hsu CY, Hsiao CW. Manufacturing intelligence to forecast and reduce
jmsy.2016.09.005. semiconductor cycle time. J Intell Manuf 2012;23(6). https://doi.org/10.1007/
[73] Ireland R, Liu A. Application of data analytics for product design: sentiment s10845-011-0572-y.
analysis of online product reviews. CIRP J Manuf Sci Technol 2018;23. https:// [101] Chen T. A systematic cycle time reduction procedure for enhancing the
doi.org/10.1016/j.cirpj.2018.06.003. competitiveness and sustainability of a semiconductor manufacturer. Sustain
[74] Xu Z, Frankwick GL, Ramirez E. Effects of big data analytics and traditional 2013;5(11). https://doi.org/10.3390/su5114637.
marketing analytics on new product success: a knowledge fusion perspective. [102] Liu Y, Hu X, Zhang W. Remaining useful life prediction based on health index
J Bus Res 2016;69(5). https://doi.org/10.1016/j.jbusres.2015.10.017. similarity. Reliab Eng Syst Saf 2019;185(February):502–10. https://doi.org/
[75] Li X, Ni Y, Ming XG, Song W, Cai W. Module-based similarity measurement for 10.1016/j.ress.2019.02.002.
commercial aircraft tooling design. Int J Prod Res 2015;53(17). https://doi.org/ [103] Sinha JK, Elbhbah K. A future possibility of vibration based condition monitoring
10.1080/00207543.2015.1047973. of rotating machines. Mech Syst Signal Process 2013;34(1–2). https://doi.org/
[76] Li X, Song W. A rough VIKOR-Based QFD for prioritizing design attributes of 10.1016/j.ymssp.2012.07.001.
product-related service. Math Probl Eng 2016;2016. https://doi.org/10.1155/ [104] Lei Y, Liu Z, Wu X, Li N, Chen W, Lin J. Health condition identification of multi-
2016/9642018. stage planetary gearboxes using a mRVM-based method. Mech Syst Signal Process
[77] Rossit DA, Tohmé F, Frutos M. Production planning and scheduling in Cyber- 2015;60. https://doi.org/10.1016/j.ymssp.2015.01.014.
Physical Production Systems: a review. Int J Comput Integr Manuf 2019;32(4–5). [105] Zhang XZ, Bi CX, Bin Zhang Y, Geng L. Real-time nearfield acoustic holography
https://doi.org/10.1080/0951192X.2019.1605199. for reconstructing the instantaneous surface normal velocity. Mech Syst Signal
[78] Ismail R, Othman Z, Bakar AA. Data mining in production planning and Process 2015;52–53(1). https://doi.org/10.1016/j.ymssp.2014.06.009.
scheduling: a review. 2009. https://doi.org/10.1109/DMO.2009.5341895. [106] Caesarendra W, Kosasih B, Tieu AK, Zhu H, Moodie CAS, Zhu Q. Acoustic
[79] Zhong RY, Xu C, Chen C, Huang GQ. Big data analytics for physical Internet-based emission-based condition monitoring methods: review and application for low
intelligent manufacturing shop floors. Int J Prod Res 2017;55(9). https://doi.org/ speed slew bearing. Mech Syst Signal Process 2016;72–73. https://doi.org/
10.1080/00207543.2015.1086037. 10.1016/j.ymssp.2015.10.020.
[80] Zhong RY, Huang GQ, Lan S, Dai QY, Zhang T, Xu C. A two-level advanced [107] Wen L, Gao L, Li X, Zeng B. Convolutional neural network with automatic
production planning and scheduling model for RFID-enabled ubiquitous learning rate scheduler for fault classification. IEEE Trans Instrum Meas 2021;70.
manufacturing. Adv Eng Informatics 2015;29(4). https://doi.org/10.1016/j. https://doi.org/10.1109/TIM.2020.3048792.
aei.2015.01.002. [108] Zheng H, Wang R, Yin J, Li Y, Lu H, Xu M. A new intelligent fault identification
[81] Wang C, Jiang P. Manifold learning based rescheduling decision mechanism for method based on transfer locality preserving projection for actual diagnosis
recessive disturbances in RFID-driven job shops. J Intell Manuf 2018;29(7). scenario of rotating machinery. Mech Syst Signal Process 2020;135. https://doi.
https://doi.org/10.1007/s10845-016-1194-1. org/10.1016/j.ymssp.2019.106344.
[82] Zhang Y, Jin X, Feng Y, Rong G. Data-driven robust optimization under correlated [109] Wen L, Li X, Gao L. A new reinforcement learning based learning rate scheduler
uncertainty: a case study of production scheduling in ethylene plant. Comput for convolutional neural network in fault classification. IEEE Trans Ind Electron
Chem Eng 2018;109. https://doi.org/10.1016/j.compchemeng.2017.10.024. 2020. https://doi.org/10.1109/TIE.2020.3044808.
[83] Zheng P, Zhang P, Wang J, Zhang J, Yang C, Jin Y. A data-driven robust [110] Yang B, Lei Y, Jia F, Xing S. An intelligent fault diagnosis approach based on
optimization method for the assembly job-shop scheduling problem under transfer learning from laboratory bearings to locomotive bearings. Mech Syst
751
Signal Process 2019;122:692–706. https://doi.org/10.1016/j. [117] David Bau AT, Zhou Bolei, Khosla Aditya, Oliva Aude. Network dissection:
ymssp.2018.12.051. quantifying interpretability of deep visual representation seminar report. 2018.
[111] Ren S, Zhang Y, Liu Y, Sakao T, Huisingh D, Almeida CMVB. A comprehensive [118] Ni CC, Lin YY, Luo F, Gao J. Community detection on networks with ricci flow. Sci
review of big data analytics throughout product lifecycle to support sustainable Rep 2019;9(1):1–12. https://doi.org/10.1038/s41598-019-46380-9.
smart manufacturing: a framework, challenges and future research directions. [119] Tallon PP. Corporate governance of big data: perspectives on value, risk, and cost.
J Clean Prod 2019;210:1343–65. https://doi.org/10.1016/j.jclepro.2018.11.025. Computer (Long Beach Calif) 2013;46(6):32–8. https://doi.org/10.1109/
[112] Xu C, Wang J, Zhang J, Li X. Anomaly detection of power consumption in yarn MC.2013.155.
spinning using transfer learning. Comput Ind Eng 2021;152. https://doi.org/ [120] Usman M, Jolfaei A, Jan MA. RaSEC: an intelligent framework for reliable and
10.1016/j.cie.2020.107015. secure multilevel edge computing in industrial environments. IEEE Trans Ind
[113] Gilmer J, et al. The relationship between High-dimensional geometry and Appl 2020;56(4):4543–51. https://doi.org/10.1109/TIA.2020.2975488.
adversarial examples. 6th Int. Conf. Learn. Represent. ICLR 2018 - Work. Track [121] Wang J, Zheng P, Lv Y, Bao J, Zhang J. Fog-IBDIS: industrial big data integration
Proc. 2018. and sharing with fog computing for manufacturing systems. Engineering 2019;5
[114] Su J, Vargas DV, Sakurai K. One pixel attack for fooling deep neural networks. (4):662–70. https://doi.org/10.1016/j.eng.2018.12.013.
IEEE Trans Evol Comput 2019;23(5):828–41. https://doi.org/10.1109/ [122] Leng J, et al. Blockchain-secured smart manufacturing in industry 4.0: a survey.
TEVC.2019.2890858. IEEE Trans Syst Man Cybern Syst 2021;51(1). https://doi.org/10.1109/
[115] Marcus G. Deep Learning: A Critical Appraisal [Online]. Available: habr.com/ TSMC.2020.3040789.
post/371179/%0A. 2018. p. 1–24. https://arxiv.org/ftp/arxiv/papers/1801/ [123] Babiceanu RF, Seker R. Big Data and virtualization for manufacturing cyber-
1801.00631.pdf. physical systems: a survey of the current status and future outlook. Comput Ind
[116] Zhang Q, Zhu S-C. Visual interpretability for deep learning: a survey [Online]. 2016;81:128–37. https://doi.org/10.1016/j.compind.2016.02.004.
Available:. 2018. http://arxiv.org/abs/1802.00614.
752

Big Data Analytics For Intelligent Manufacturing Systems A Review

Uploaded by

Copyright:

Available Formats

Big Data Analytics For Intelligent Manufacturing Systems A Review

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Big Data Analytics For Intelligent Manufacturing Systems A Review

Uploaded by

Copyright:

Available Formats

Journal of Manufacturing Systems 62 (2022) 738–752

Contents lists available at ScienceDirect

Journal of Manufacturing Systems

Big data analytics for intelligent manufacturing systems: A review

Fig. 3. The development of big data analytics (adjusted from [38]).

Fig. 4. Classification of distributed computing frameworks.

Fig. 5. MapReduce processing process (adjusted from [48]).

In the era of big data, data is usually generated continuously and

Fig. 9. The applications of BDA in manufacturing systems.

You might also like