Machine learning in manufacturing and industry 4.0 applications

Abstract

The machine learning (ML) field has deeply impacted the manufacturing industry in the context of the Industry 4.0 paradigm. The industry 4.0 paradigm encourages the usage of smart sensors, devices, and machines, to enable smart factories that continuously collect data pertaining to production. ML techniques enable the generation of actionable intelligence by processing the collected data to increase manufacturing efficiency without significantly changing the required resources. Additionally, the ability of ML techniques to provide predictive insights has enabled discerning complex manufacturing patterns and offers a pathway for an intelligent decision support system in a variety of manufacturing tasks such as intelligent and continuous inspection, predictive maintenance, quality improvement, process optimisation, supply chain management, and task scheduling. While different ML techniques have been used in a variety of manufacturing applications in the past, many open questions and challenges remain, from Big data curation, storage, and understanding, data reasoning to enable real-time actionable intelligence to topics such as edge computing and cybersecurity aspects of smart manufacturing. Hence, this special issue is focused on bringing together a wide range of researchers to report the latest efforts in the fundamental theoretical as well as experimental aspects of ML and their applications in manufacturing and productionsystems.

KEYWORDS:

• machine learning
• industry 4.0
• manufacturing analytics
• big data in manufacturing
• cloud manufacturing
• fault detection

Introduction

Manufacturing can be defined as the fabrication or assembly of components into finished products on a large scale (Encyclopaedia Britannica 2020). It is one of the most important industries in the world’s economy wherein, it accounts for approximately 16% of the global GDP in 2019 and has generated an output of 13.9 trillion globally (Sodani 2021). One of the most critical manufacturing goals is to produce more high quality products at minimum costs. But manufacturing products can be a very expensive and complicated process for businesses that do not have the associated resources and tools to design and develop quality products (AMFG AI 2019). Over the last couple of centuries, the history of manufacturing has changed dramatically. Instead of items being produced using manual labour, industries sought machines in order to produce the items, leading to the Industrial Revolution in the 18^th century. As of 2016, the world entered the fourth Industrial Revolution coined as ‘Industry 4.0’, which promotes the computerisation of manufacturing by encompassing three technological trends: connectivity, intelligence, and flexible automation (AMFG AI 2019).

Industry 4.0 has given rise to an emerging sector in manufacturing called Smart Manufacturing that opens doors for analytics in the industry. Figure 1 represents the various components attributed within the domain of Smart Manufacturing (Sodani 2021). It is a technology-driven approach that utilises the IoT and internet-connected devices to produce goods and monitor processes. Its goal is to automate the processes involved in manufacturing to maximise efficiency, increase sustainability, supply chain management, and identify the systems barriers even before they occur by generating, optimising, and implementing enormous volumes of data. With the application of advanced analytics to industrial data, manufacturers can gain insights to optimise the productivity of individual assets as well as the total manufacturing operation prevailing artificial intelligence (AI) and machine learning (ML).

Figure 1. Global Smart Manufacturing Segmentation.

Figure 2. The multi-faceted usage of AI and machine learning in manufacturing domain.

Industry 4.0 converges information technology and operational technology to create a cyber-physical environment with digital solutions and advanced technologies including,

Table

• Industrial Internet of Things (IIoT)	• Additive manufacturing (AM)
• Digital twin	• Advanced robotics
• Cloud computing	• Augmented and virtual reality

In manufacturing, a paradigm shift is happening right now. Advances in Big data and Machine Learning (ML) is changing the traditional manufacturing era into the smart manufacturing era of Industry 4.0 (I4.0). This paradigm shift is creating new opportunities. IIoT is a subset of the Internet of Things (IoT) which refers to a network of physical devices that are digitally interconnected, facilitating the communication and exchange of data through the internet IIoT interconnects machines through various sensors, RFID tags, software, and electronics which are integrated with machines to collect real-time data. The abundance of smart sensors and the Internet of Things are the key enabler of curating and storing a remarkable amount of industrial data-rich environments related to all aspects of production. Digital twin runs an online simulation based on data received from IIoT. A digital twin is a digital representation of a real-world product, machine, process, or system that allows companies to better understand, analyise and optimise their processes through real-timesimulations.

Cloud computing uses internet connections to store, access, and process data and is a key enabler of the digital twin paradigm. Additionally, current generation manufacturing is supported by advanced technologies such as advanced robotics, augmented, and virtual reality and additive manufacturing. ML techniques, a subfield of Artificial Intelligence (AI), has the potential to become the main driver in uncovering fine-grained intricate production patterns in smart manufacturing paradigm and offering timely decision support in a wide range of manufacturing and production applications, to name a few, predictive maintenance, process optimisation, task scheduling, quality improvement, supply chain, and sustainability so on.

Smart use of technology and AI can help grow and drive businesses. The biggest companies around the world are utilising AI and ML in manufacturing and investing huge amounts of money in its development (Chuprina 2020). According to Mckinsey 40% of all the potential value that can be created by analytics today, comes from AI and ML techniques, wherein ML accounts for between $3.5 trillion to $5.8 trillion in annual value. Data-driven techniques automate learning from data, detect prevalent underlying patterns and make informed decisions. Industries have been successful in the application of machine learning mainly into three aspects of the business:

Table

• Operations

• Productions

• Post-Productions

Additionally, data analytics in the manufacturing industry can be used to amplify growth in the following domains (but not limited to)

1. Improve assembly-line efficiency using data analytics.

2. Improved Customer Experience, including personalisation and finding individualised value propositions.

3. Inventory management by

   1. Real-time insights and visibility into inventory along the supply lines

   2. Delivery route optimizations

4. Minimise loss associated with delayed, damaged, or lost goods in transport and for providing real-time asset management include real-time alerts

5. Reduce errors and corrections during product development and improve the product’s quality and packaging with,

   1. Analytics-backed simulations

   2. Product modelling

6. Predictive maintenance helps increase assets’ lifetime by,

   1. Asset Management.

   2. Improving asset availability.

   3. Detection of faults and defects.

   4. Prevention of unplanned downtimes.

7. Increase visibility into the supply chain with actionable data location-based IoT services.

The manufacturing environment offers enormous opportunities to leverage artificial intelligence techniques for better decision-making (Sharma, Zhang, and Rai 2021). A key enabler of AI in manufacturing environments is IIoT, which facilitated the harnessing of big data. Big data, in turn, needs AI to make decisions in real-time. The core techniques of AI like deep learning, computer vision, reinforcement learning have already been utilised in manufacturing scenarios. Computer vision has been leveraged for structural health monitoring (Yang et al. 2020). The vision-based techniques offered high spatial resolution even with inexpensive sensors. Ontology can be used to store and organise the knowledge of manufacturing systems (Ali et al. 2019). AI can be used to study the brain computer interface of manufacturing operators (Shankar and Rai 2014). This can enhance the study of industrial safety and ergonomics. The data generated by manufacturing processes, like additive manufacturing, can be analyised for real-time monitoring, control, and defect minimisation (Khadilkar, Wang, and Rai 2019). The manufacturing equipment can be diagnosed and prognosed using machine learning (Yang and Rai 2019). Augmented reality devices have been used for maintenance applications (Young and Rai 2021). They offered additional flexibility to reconfigure the maintenance procedure on the go.

While different ML techniques have been used in a variety of manufacturing applications in the past, many open questions and challenges remain, from Big data curation, storage, and understanding, data reasoning to enable real-time actionable intelligence to advanced topics such as edge computing, fog computing, cybersecurity aspects of smart manufacturing. Hence, this special issue is focused on bringing together a wide range of researchers to report the latest efforts in the fundamental theoretical as well as experimental aspects of ML and their applications in manufacturing and production systems.

Summary of papers in this special issue

A summary of papers that address several important issues at the intersection of manufacturing and ML domain is presented next. Different categories to which the papers belong and the importance of these categories in machine learning in the manufacturing context are elaborated next.

Computer vision-based inspection and monitoring

One of machine learning’s most high-impact application areas in the manufacturing domain is computer vision-based part inspection and process monitoring. Utilising cost-effective sensors such as RGB cameras wrapped with ML-based algorithms can enable high throughput part inspection. Computer vision (images and video) based approaches that are integrated with ML can enable monitoring of a product throughout the entire production process. Additionally, a computer vision-based approach can also enable high-quality continuous process monitoring.

‘A data-driven method for Enhancing the image-based automatic inspection of IC wire bonding defects’ by Chen et al. (Chen, Zhang, and Wu 2020) offers a data-driven method for enabling automatic inspection of integrated circuit (IC) wire bonding defects. The outlined method in the paper comprises of three steps: (1) data pre-processing for locating and separating IC chip image patches from the raw image, (2) feature engineering that extracts geometric features from the segmented wires, and (3) machine learning algorithms (such as CNN and SVM) based classification. The authors showcase the efficacy of the developed method on a set of X-ray images collected from a semiconductor factory.

‘Soft sensor of flotation froth grade classification based on hybrid deep neural network’ by Zhang and Gao (Zhang and Gao 2021) summarises their efforts to develop a soft sensor that can process froth images of flotation tailing to classify iron ore tailings grade. The two key contributions of the paper are to curate and establish a database of froth images of flotation tailings and a comparison of the accuracy of multiple deep neural network (DNN) models for the task at hand. Based on the comparison results, the authors outline a fine-tuned hybrid DNN model that has excellent accuracy and develops software around it. The experimental results show the potential of application of DNNs in the field of iron ore froth flotation.

Fault detection

Timely and accurate diagnosis of manufacturing equipment process faults provides a strategic advantage to help manufacturing companies stay competitive by reducing machine downtimes. Machine learning algorithms within the manufacturing industry will find increased usage for enabling manufacturing system fault diagnosis as more and more customers require manufacturers to expedite the delivery of high-quality products at a low cost.

‘A deformable CNN-DLSTM based transfer learning method for fault diagnosis of rolling bearing under multiple working conditions’ by Wang et al. (Wang et al. 2020) presents a transfer learning-based approach that combines deformable convolutional neural network (CNN) and deep long short-term memory (DLSTM) rolling bearing fault diagnosis. The authors specifically focus on bearing faults under multiple working conditions where it is hard to find sufficiently large-scale labelled data. The transfer learning approach allows them to pre-train fault diagnosis model using data samples under one working condition and then transfer the model with further fine-tuning (with very few data samples) to another working condition. The developed framework, when applied to an experimental data set, shows better results than the state-of-the-art methods.

‘Applications of deep learning for fault detection in industrial cold forging’ by Glaeser et al. (Glaeser et al. 2021) showcase the usage of deep learning techniques in the industrial cold forging fault detection domain. Specifically, the authors outline two different approaches based on convolutions neural network (CNN) and decision tree (DT) to understand how machine signal impacted various faults. The outlined approaches show superior performance on vibration test data sets collected for commonly encountered faults in the industrial cold forging domain.

Cloud manufacturing

Cloud storage within manufacturing is being embraced to store and manage manufacturing production big data. Aside from the understandably increased storage space, the cloud and machine learning techniques will help reduce costs and drive profitability. Understanding how hardware – Edge vs Cloud – in deploying real-world machine learning-based manufacturing solutions at scale will be key in the near future.

‘A cloud edge-based two-level hybrid scheduling learning model in cloud manufacturing’ by Jian et al. (Jian, Ping, and Zhang 2020) presents a cloud edge-based two-level hybrid scheduling learning model to enable fast prediction of the scheduling results. The proposed model at the first level decomposes and distributes multiple scheduling tasks into several first-level sub-tasks based on First-in First-out (FIFO) principle. The second level enables even granular decomposition to atomic tasks that are then scheduled to different industrial devices (factory edge nodes). The second level utilises an innovative combination of improved bat algorithm (VSSBA) and long and short-term memory networks (LSTM) for scheduling problems. Experiments performed by the authors show that the two-level approach can improve performance in real-life applications in a cloud manufacturing setting.

Process improvement and optimisation

The prescriptive analytics power of ML techniques can augment the manufacturing employees’ effort to select the optimal set of parameters related to a given manufacturing process, hence enabling manufacturing process improvement and optimisation. The discipline at the intersection of ML and process improvement that leads to manufacturing analytics insights enabling faster mass and customised production at a rapid pace with as little waste as possible will grow by leaps and bound in the next decade.

‘An investigation of the utilisation of different data sources in manufacturing with application in Injection Moulding’ by Rønsch et al. (Rønsch, Kulahci, and Dybdahl 2021) report their work on the study that explores effective utilisation of various data sources for process improvement in injection moulding. Specifically, the authors investigate whether a good prediction accuracy can be achieved by using readily available Machine Process Data on a moulding manufacturing line comprising of 100 injection moulding machines or additional sensor signals obtained at a higher cost can provide additional beneficial information. For the specific use case that was carried out in close collaboration with an industrial partner, they conclude that available machine process data does not capture the variation in the raw material that impacts element quality.

‘Using process mining to improve productivity in make-to-stock manufacturing’ by Lorenz et al. (Lorenz et al. 2021) demonstrate a novel use case of data-driven procedure to improve productivity in make-to-stock manufacturing. Specifically, the outlined procedure utilises process mining to dynamically map and analyse manufacturing processes with high process complexity and variety in an automated manner to enable productivity improvement. The authors also empirically validate the usage of the developed procedure on a test case emanating from a leading manufacturer of sanitary products and provide tangible improvement suggestions for the manufacturers.

‘A supervised machine learning approach for the optimisation of the assembly line feeding mode selection’ by Zangaro et al. (Zangaro, Minner, and Battini 2020) outlines a supervised learning-based approach that utilises classification and regression tree (CART) algorithm for line feeding problem (LFP). The proposed approach takes as input attributes of components and manufacturing environment to create a decision tree that suggests a line feeding mode for every component. Additionally, for cases that result in an infeasible solution, they also outline a repair approach that provides feasible solutions with respectable average cost deviation. The proposed approach predicts line feeding mode with good classification accuracy.

State of the art review papers

Review papers related to machine learning applications in the manufacturing domain in this special issue bring together quantitative and qualitative components and provide new conceptual frameworks, synthesise diverse results, and give the broader research community a ‘state-of-the-art’ snapshot of essential issues related to the focus of this special issue.

‘Artificial intelligence (AI) in augmented reality (AR)-assisted manufacturing applications: a review’ by Sahu et al. (Sahu, Young, and Rai 2020) provides a comprehensive review of a very dynamic research domain at the intersection of AI and AR. The current paradigms utilised in AR for camera calibration, detection, tracking, camera position and orientation (pose) estimation, inverse rendering, procedure storage, virtual object creation, registration, and rendering are still mostly dominated by traditional non-AI approaches. The authors provide a critical appraisal of these strategies. Additionally, the authors also outline potential AI solutions for every component of the computational pipeline. Furthermore, interesting and fruitful future research work directions at the intersection of AI and AR are also listed.

‘The interpretive model of manufacturing: A theoretical framework and research agenda for machine learning in manufacturing’ by Sharma et al. (Sharma, Zhang, and Rai 2021) argues that a new framework is needed to fully comprehend the paradigm of application of machine learning in manufacturing. The authors present a comprehensive hybrid literature review approach comprising of a thematic and conceptual synthesis of the literature resulting in the interpretive model framework of manufacturing. The resultant interpretative framework is articulated as consisting of scan, store, interpret, execute, and learn as its purposive components. Additionally, the authors provide pertinent future research questions and implications for manufacturing operations, manufacturing strategy, and manufacturing.

Conclusion

Manufacturing is undertaking a definitive shift as it assimilates and is transformed by the usage of machine learning techniques. This special issue brought together a wide range of researchers to report the latest efforts in the fundamental theoretical as well as experimental aspects of ML and their applications in manufacturing and production systems to understand and contextualise this shift at the onset of this decade. The topics such as computer-vision based inspection and monitoring, fault detection, cloud manufacturing, process improvement and optimisation, and comprehensive state of the art review papers are spanned by the papers in this special issue.

Going forward, as Industry 4.0 transforms the manufacturing environments into cyber-physical environments, AI will be indispensable for data analysis and subsequent decision-making (Rai and Sahu 2020). Product design (Huang and Rai 2018), inspection (Zhang et al. 2019a), geometry processing (Zhang, Jaiswal, and Rai 2018), material informatics (Wang et al. 2021), control of dynamic systems (Zhang et al. 2021) are few more applications of ML in manufacturing. The utility of ML in manufacturing applications reflects that ML can be integrated into every phase of product lifecycle, starting from design to disposal of the product in a manufacturing environment.

It is also important to note that the ML-based decision-making in manufacturing applications can also be supplemented by the physics governing the physical phenomena. Synergizing ML with physics becomes crucial, especially in the case of manufacturing applications where data generation can be expensive and unsafe (Rai and Sahu 2020).