End-to-end digitisation of physical assets within a value chain and their integration into a digital ecosystem is the focus of the fourth industrial revolution [1]. Ubiquitous sensors and microprocessors installed on machines, embedded systems and smart devices along with the increasing horizontal and vertical networking of value chains will result in unprecedently seen amounts of data.
The application of analytics techniques to this massively growing data flow can generate substantial benefits as valuable information can be quickly extracted and used to achieve in-depth understanding, gain insights and make discoveries for accurate, evidence-based and potentially autonomous decision making in complex industrial processes [2]. The most cited advantages that analytics can generate in the manufacturing world are related to maintenance operations. Maintenance is no longer seen as a non-value-added function that merely reduces machines availability; the current view is towards considering it as a source of added value, with the potential of driving performance improvements [3].
Different maintenance strategies are available to be implemented in manufacturing systems. The typical classification identifies four main categories (Figure 1).
Figure 1 Typical classification of maintenance strategies
Predictive maintenance is undoubtedly the one that has attracted most attention in the last few years and is seen as one of the killer apps of Industry 4.0. It combines many of the technologies that underpin the new wave of industrial digitization, such as networked sensors, big data, advanced analytics, and machine learning [4].
Although the predictive maintenance concept has been recently augmented with the power of data and advanced analytics, it is certainly not a recent practice. For instance, when a technician performs a visual inspection and selects, based on his knowledge, experience and intuition the best time to shut down a piece of equipment, he is in fact performing predictive maintenance. At a subsequent level of sophistication, operators often use instruments to perform periodic inspections and obtain more specific and objective information about the condition of the asset in question. When continuous data flow from sensors is used to monitor the state of an asset and identify anomalous behaviour, condition-based monitoring, which is a precursor of predictive maintenance, is actually being performed.
Condition-based monitoring
Condition Based Monitoring (CBM) is a type of maintenance strategy that involves using sensors to measure the status of an asset over time while it is in operation. When following this strategy, mechanical and operational conditions are monitored, periodically or in real-time (sometimes known as “real-time condition monitoring”) through sensors and process data. Alarm thresholds are specified for each of the monitored variables. When a variable value crosses its threshold, an alarm is sent, and the troublesome component is identified and scheduled for maintenance. This approach can also be augmented with the introduction of more sophisticated rules that trigger alarms based on historical data patterns and anomalous behaviours (i.e. pattern recognition); in this case, a predictive element is introduced in this strategy although it is generally not used to evaluate the residual useful life of a component (i.e., and plan a maintenance action accordingly) but to anticipate an imminent failure and schedule a repair intervention in the short term. As artificial intelligence (AI) approaches can also be used to unveil anomalous patterns and anticipate failures, the distinction between condition-based maintenance and predictive maintenance is very fine and some authors classify condition-based maintenance as a predictive maintenance approach.
The type of sensors that can be used to perform this maintenance strategy range from vibration (which are the most popular) to infrared sensors, from ultrasonic to acoustic, from oil analysis to electrical parameters monitoring.
The opportunity of adopting a condition-based maintenance strategy on manufacturing assets should be based on a variety of practical observations and analyses of machine performance data conducted by maintenance engineers. Frequency and randomness of breakdowns, need for repetitive repairs, and impact on quality yield are typical factors that should be considered in the opportunity assessment.
Predictive maintenance
Predictive maintenance is a maintenance strategy that is based on the possibility of predicting future failure points of a machine component based on relevant input variables, so that the component can be replaced just before it fails. As a result of this strategy, equipment downtime is minimised, and the component lifetime is maximised with an obvious positive impact on productivity, maintenance costs and spare parts inventory management.
Predictive maintenance involves prognosis and diagnosis processes:
- Prognosis: aims at predicting the fault caused by natural wear and tear
- Diagnosis: analyses the difference between observed and expected parameters so that the effect of working conditions under certain failures can be evaluated [5].
Typical approaches used for prognostic and diagnostic predictive maintenance are summarised in Figure 2. AI approaches and, particularly, machine learning (ML) have emerged as powerful tools for developing intelligent predictive algorithms [6]. Machine learning is an area of AI that focuses on algorithms that allow a machine to learn from data. At a further level of analytics sophistication, deep learning methods can also be used to develop predictive maintenance systems. Deep learning, a frontier area of research within machine learning, uses neural networks with many layers (hence the label “deep”) to push the boundaries of machine capabilities. Conventional machine learning methods require careful engineering and considerable domain expertise to identify relevant features (i.e. parameters/variables) to include in the fault prediction model and design a feature extractor that transforms raw data into a feature vector used by the learning system to extract patterns [7]. With deep learning, raw data can be directly fed into the learning system and the representation needed for fault classification or pattern detection is automatically identified through multiple layers of transformation.
Figure 2 Typical approaches for predictive maintenance
Applications
Predictive maintenance systems can be applied to any manufacturing asset or component subject to failures; for this reason, their application is potentially viable across all manufacturing sectors. It is evident that the digital backbone required for their effective implementation has led to a widespread adoption of these systems in industries that are typically more automated and digitised, such as the oil and gas industry, the semiconductor industry and the automotive sector.
For instance, in the semiconductor industry, big data is being leveraged to adopt a more predictive or even proactive approach to factory control through advanced process control (APC) solutions [8], which also include approaches to augment failure detection with failure prediction [9]. Thanks to advances in data and model integration, the landscape of predictive maintenance in semiconductor manufacturing has rapidly advanced to the point that fab-wide solutions are now available.
In the automobile industry innovative concepts, such as Electric, Connected & Autonomous vehicles and Mobility Services, allow manufacturer to offer new services using Big Data and advances in technology. Data collected through embedded sensors or even social media can be used to offer assisted/autonomous driving and safety alerts, monitor the vehicle health, extract automated insights for design and production. The use of data mining on logged vehicle data as a predictive maintenance solution has been demonstrated as a viable approach in various research studies.
Application in Machining
In the machining industry, considerable progress has been made in fault interpretation, detection, and prediction for machine centres using AI approaches. Generally, these predictive maintenance systems focus on predicting tool wear of specific components, such as gearbox and rolling element bearings. Online sensing methods can be used to monitor the tool wear status and identify anomalous behaviours or predict its remaining useful life. As a result, significant benefits in terms of machining quality and cost can be obtained. The first step to adopt a more proactive maintenance strategy would be to identify suitable sensors by proving the link between tool condition and metrics that sensors can measure. For instance, in milling operations, acoustic emissions are very effective in detecting tool breakage rather than wear since typical acoustic signal peaks are emitted when the tool breaks; on the contrary, mechanical vibration, surface roughness, cutting forces, spindle motor current have proven to correlate well with tool wear and are factors that have been concurrently considered in multi-sensor-based failure prediction models [10], [11].
E-maintenance platforms using standard communication protocols can be used by machinery manufacturers as a technological support to enhance maintenance service provision [12]. Accessibility to large datasets coming from a network of connected machines operating in different plants under different environmental conditions also allows to considerably improve the performance of predictive models and can also be used to derive optimisation opportunities for machine design and production. From a machine user’s perspective, the latest tendency is to integrate fault diagnosis and prognosis systems of specific components into comprehensive decision support tools exploiting Industry 4.0 technologies, such as Internet of Things, Internet of Services and Cyber Physical systems to enhance failure detection and optimise maintenance scheduling and implementation [5], [13].
Bibliography
[1] R. Geissbauer, V. Jesper, and S. Schrauf, “Industry 4.0: Building the digital enterprise,” PwC, 2016. [Online]. Available: https://www.pwc.com/gx/en/industries/industries-4.0/landing-page/industry-4.0-building-your-digital-enterprise-april-2016.pdf.
[2] K. Zhou, T. Liu, and L. Zhou, “Industry 4.0: Towards future industrial opportunities and challenges,” in 2015 12th International conference on fuzzy systems and knowledge discovery (FSKD), 2015, pp. 2147--2152.
[3] M. Holgado, M. Macchi, and L. Fumagalli, “Maintenance business model: a concept for driving performance improvement,” Int. J. Strateg. Eng. Asset Manag., vol. 2, no. 2, p. 159, 2015.
[4] S. Bradbury, B. Carpizo, M. Gentzel, D. Horah, and J. Thibert, “Digitally enabled reliability: Beyond predictive maintenance | McKinsey,” no. October, 2018.
[5] Z. Li, Y. Wang, and K. S. Wang, “Intelligent predictive maintenance for fault diagnosis and prognosis in machine centers: Industry 4.0 scenario,” Adv. Manuf., vol. 5, no. 4, pp. 377–387, 2017.
[6] T. P. Carvalho, F. A. A. M. N. Soares, R. Vita, R. da P. Francisco, J. P. Basto, and S. G. S. Alcalá, “A systematic literature review of machine learning methods applied to predictive maintenance,” Comput. Ind. Eng., vol. 137, no. April, p. 106024, 2019.
[7] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015.
[8] J. Moyne, J. Samantaray, and M. Armacost, “Manufacturing Advanced Process Control,” vol. 29, no. 4, pp. 283–291, 2016.
[9] H. K. Lim, Y. Kim, and M.-K. Kim, “Failure Prediction Using Sequential Pattern Mining in the Wire Bonding Process,” IEEE Trans. Semicond. Manuf., vol. 30, no. 3, pp. 285–292, 2017.
[10] D. Wu, C. Jennings, J. Terpenny, and S. Kumara, “Cloud-based machine learning for predictive analytics: Tool wear prediction in milling,” Proc. - 2016 IEEE Int. Conf. Big Data, Big Data 2016, pp. 2062–2069, 2016.
[11] P. Stavropoulos, A. Papacharalampopoulos, E. Vasiliadis, and G. Chryssolouris, “Tool wear predictability estimation in milling based on multi-sensorial data,” Int. J. Adv. Manuf. Technol., vol. 82, no. 1–4, pp. 509–521, 2016.
[12] L. Fumagalli and M. Macchi, “Integrating maintenance within the production process through a flexible E-maintenance platform,” IFAC-PapersOnLine, vol. 28, no. 3, pp. 1457–1462, 2015.
[13] K. Zhu and Y. Zhang, “A Cyber-Physical Production System Framework of Smart CNC Machining Monitoring System,” IEEE/ASME Trans. Mechatronics, vol. 23, no. 6, pp. 2579–2586, 2018.
