Open access peer-reviewed chapter
Abstract
With the ever-growing complexity of different assets in a factory, the main focus of predictive maintenance solutions has shifted from model-based approaches to data-driven and hybrid approaches. This shift as a result highlights the importance and the inevitable impact of data, data quality, model maintenance, and model interpretability on the performance and acceptability of these predictive maintenance approaches in industry. In this chapter, the hurdles for developing effective predictive maintenance solutions for original equipment manufacturers (OEMs) and small and medium-sized enterprises (SMEs) with different levels of digitalization are introduced. Furthermore, it is discussed how to choose a suitable strategy for developing a predictive maintenance model, given the different constraints in the availability of data and the requirements of the customer.
Keywords
- predictive maintenance
- data quality
- Industry 4.0
- machine learning
- trustworthy AI
1. Introduction
In the landscape of industrial operations, the concept of maintenance has undergone a profound evolution. Traditionally, maintenance strategies were predominantly reactive or scheduled, often resulting in downtime, inefficiencies, and unexpected costs. Assets would be repaired or replaced only after failure, leading to disruptions in production and compromising overall efficiency. Furthermore, scheduled maintenance, while aiming to prevent breakdowns, often resulted in unnecessary servicing of equipment that may not have required immediate attention, leading to inefficiencies and increased costs.
However, with the advent of the digital era and advancements in technology, a paradigm shift has occurred, ushering in the era of predictive maintenance. At the heart of predictive maintenance lies the integration of data, models, and performance evaluation, offering a proactive approach to asset management and optimization. By harnessing data from various sources, such as sensors, Internet of Things (IoT) devices, and historical records, predictive maintenance enables the prediction of equipment failures before they occur, allowing for timely intervention and maintenance activities. This transition from reactive or scheduled maintenance to predictive maintenance represents a significant leap forward in operational efficiency and cost-effectiveness.
The complexity of assets within modern factories has surged exponentially, accompanied by a corresponding rise in the complexity of maintenance challenges. To address these challenges, predictive maintenance solutions have transitioned from traditional model-based approaches to more data-driven and hybrid methodologies. This shift underscores the paramount importance of data and their quality, along with the ongoing maintenance and interpretability of models, in ensuring the efficacy and acceptance of predictive maintenance solutions across industries.
In this context, the availability and quality of data play a critical role in the success of predictive maintenance initiatives. The proliferation of sensors and IoT devices has enabled the generation of vast amounts of data from industrial equipment and processes. However, challenges, such as data silos, interoperability issues, and data quality concerns, can hinder the effective utilization of these data for predictive maintenance purposes. Addressing these challenges requires robust data management strategies, including data integration, cleansing, and preprocessing, to ensure the reliability and accuracy of predictive maintenance models.
Moreover, the concept of the Asset Administration Shell (AAS) in the context of predictive maintenance deserves attention. The AAS, as defined in the Reference Architecture Model Industrie 4.0 (RAMI4.0), offers a standardized representation of assets and their associated data. This framework facilitates interoperability and data exchange across heterogeneous systems, ensuring seamless integration and communication within industrial environments. By adopting the AAS framework, organizations can enhance the transparency, accessibility, and integrity of asset-related data, thereby enabling more effective predictive maintenance strategies.
As predictive maintenance solutions evolve to meet the diverse needs of original equipment manufacturers (OEMs) and small and medium-sized enterprises (SMEs) operating at varying levels of digitalization, it becomes essential to consider the specific requirements and constraints of each context. Customer requirements, including cost considerations, operational priorities, and regulatory compliance, must be carefully evaluated when developing predictive maintenance solutions. Balancing these requirements with the constraints imposed by data availability and technological capabilities is crucial in selecting an optimal strategy for predictive maintenance model development.
Throughout this chapter, we will navigate through the complexities and nuances of predictive maintenance in industrial settings, exploring challenges, opportunities, and best practices for leveraging data, models, and performance evaluation to drive operational excellence. By the conclusion of this chapter, readers will gain a comprehensive understanding of the role of predictive maintenance in enhancing asset management and operational efficiency, thereby empowering organizations to thrive in the dynamic landscape of industrial operations.
2. From laboratory to industrial settings
For finding trends in asset signal readings (see Figure 1) which indicate a potential failure in the system, the most convenient way would be to train a prediction model with abundant samples of such sequences. For ensuring the generalization power of the prediction model for all the impending failures in the system, the gathered samples should cover different failure types. Once such data are acquired from the system, a classification or regression model can be trained to estimate and/or predict its status. Nevertheless, these samples are normally hard to attain in industry, as failures in such settings mean loss of productivity, reduced throughput, maintenance costs, and replanning costs. Thus, predictive maintenance solutions tend to be different in industry from the ones developed in perfect conditions and laboratories. In what follows, namely in Section 2.1, some of the reasons for lack of data and also annotated data are introduced.
Figure 1.
Data acquisition from different assets in a smart factory.
Additionally, all decisions made in an industrial plant have to show their potential in having the return of investment as fast as possible. Therefore, for such use cases, it is vital to offer solutions that first of all do not cause substantial costs for sensor installation and asset verification (see Section 2.2) and can be also verified, given the operator’s domain knowledge about the system (see Section 2.3).
Furthermore, given the constant changes in the production line, the deployed predictive maintenance solution must be aware of and robust towards these frequent changes impacting the read data. In Section 2.4, issues causing the aforementioned changes are discussed in detail.
In what follows, the briefly introduced concepts will be explained for a better comparison between predictive maintenance solutions developed for the industry and the ones developed in the controlled settings of a lab.
2.1 Lack of failure samples
In numerous industrial predictive maintenance settings, no instances of failures from the targeted asset are available, which could be due to the following:
The targeted asset is new and thus no annotated data have been previously gathered from it.
The industry owner did not store any historical data from the production line and/or did not annotate the data.
Gathering failure sample from the asset cannot be financially justified as the costs of developing the predictive maintenance solution would be higher than preventive maintenance.
Incidents leading to a failure in the asset can be dangerous for operators and/or users of the asset such as commercial planes.
In what follows, different strategies for dealing with lack of annotated for predictive maintenance are presented.
2.1.1 Predictive maintenance for new assets
There are several possible cases where a predictive maintenance solution has to be developed or adapted for a new asset. Occasionally, given the wear and tear during production, despite the normal usage of the asset, or the release of new series of the aforementioned asset, the target asset needs to be changed to increase productivity during production.
The former is easier to deal with as principally, the new replacement for the asset, has the same build and logic. In a simplified case, the similarity in the build and the logic of the asset must result in the same signal readings from the system as the physical characteristics of the asset have not changed and the behavior of the asset, controlled by the logic unit of the asset, is also as it previously was. Nonetheless, for more complex cases, no two instances of the same assets behave exactly the same, as assets after production go through calibration steps to ensure their required performance. Apart from the build and logic of the asset, there are numerous factors that potentially impact the gathered data from the asset, which are shown in Figure 2, but are not relevant for the current case and will be discussed further in Section 2.4.
Figure 2.
Internal and external factors impacting the data gathered from a production asset.
On a separate note, it is also the case for many SMEs that they assign new tasks for different assets, as either they are not being used for production or priorities require dedicating more production assets for a specific customer. In this case, the build of the asset has not changed but its logic has changed, given the new production instructions. Regardless of the sources of changes in the read data from an asset, in case that annotated historical data from the previous asset are available, it is possible to adapt the available predictive maintenance solution to mitigate the changes in the read data from the new asset. In fact, this issue is a well-studied area of research in machine learning called domain adaptation. In short, domain adaptation seeks to learn a model from a provided source of annotated data, which can later be generalized to a target domain by minimizing the difference between the source and target distributions (see Figure 3) or by relying on features that are source independent [1, 2, 3, 4].
Figure 3.
Reducing distribution gap between the source and target domains by domain adaptation.
On the other hand, while developing a predictive maintenance solution for a new production line or a new asset, from which no former historical data are available, there are two possible ways to deal with lack of annotated data:
Using physical model of the system for data generation.
Using anomaly detection.
If the complexity of the targeted asset does not prevent creating a physical model, capable of recreating its behavior, the aforementioned model can be used to create annotated data [5, 6]. Nevertheless, one important note here is that, it must be feasible to create different scenarios in the simulation model which represent different failure types in the physical system. Otherwise, the data generated using the physical model can be used as a basis for the anomaly detection solution introduced later (Figure 4).
Figure 4.
Simulation model data generation for predictive maintenance.
Anomaly detection can be used as a basis for any new predictive maintenance solution, as a main building block of condition monitoring of the asset. Concisely, anomaly detection aims to flag parts of the signal readings from the asset that deviate from a previously provided pattern [7]. Moreover, anomaly detection as one of the most important use cases of unsupervised learning does not require annotated data and thus can significantly facilitate creating a predictive maintenance solution for a new asset or new production settings. One of the requirements for this method is a reference behavior from the system. As soon as this reference is available, different machine learning algorithms, such as autoencoder, one-class support vector machine, isolation forest, and different clustering algorithms, can be used to detect datapoints that are deviating from the provided reference behavior. Nonetheless, although this method may appear favorable, solutions relying on anomaly detection are mostly heavily dependent on their hyperparameters and metrics. Such hyperparameters help determine if a given datapoint is
In addition, as data readings from industry are not perfect, usually it is not easy to find the perfect decision boundary separating the datapoints representing the normal working condition of the asset from the erroneous conditions and failures, given the signal readings. This exacerbates further when no samples of asset failure are available as there can be infinitely many decision boundaries that divide the available data space, given the datapoints representing the healthy status of the asset. Furthermore, decision boundary in different use cases has to weigh the importance of finding signal readings representing a failure against falsely flagging a normal working condition datapoint as erroneous (see Figure 5).
Figure 5.
Impact of decision boundary on false positive and false negative rates.
The aforementioned importance can be simply be translated into an objective function during the model training. Model improvement after deployment is also possible when more information about the importance of different classes of data is evident. Some values, which can help interpret the results of the prediction, are as follows [9]. Please note that the negative class represents scenarios where the target asset does not have any problem and is functioning as expected. The positive class, on the other hand, represents faulty asset states.
True positive rate (TPR):
True negative rate (TNR):
False positive rate (FPR):
False negative rate (FNR):
Accuracy:
Balanced accuracy:
Given the highly unbalanced nature of data available in different predictive maintenance use cases (see Section 2.4), it is suggested to use accuracy measurement criteria such as balanced accuracy over a simple accuracy value to have a better overview on the performance of the deployed prediction model for the asset. In fact, relying on only one class of data in the available asset readings is not recommended as the ratio and importance of positive and negative classes can change in time, leading to numerous issues caused by biased sample selection and model training. By monitoring the overall performance of the system,
2.1.2 Lack of physical model of the system
Production assets frequently used in OEMs and SMEs are constantly growing more complex for increased efficiency and performance, making them harder to model. Therefore, in the recent studies addressing the lack of annotated data from an asset, the focus has shifted from gathering data from a simulation model for model training to anomaly detection. With the help of anomaly detection, it is possible to constantly compare the asset readings with a previously provided normal behavior and flag parts of the readings as anomaly which deviate from the expected trend [11]. The aforementioned deviation can also be translated into a health index for the given asset. In fact, the health index can be interpreted as a probability value which indicates how probable it is that the target asset in its current condition is defective. The calculated health indices can then be used to calculate the degradation of the asset in time. In essence, what follows are the steps recommended to implement an anomaly detection-based predictive maintenance solution (Figure 6):
Gathering asset readings representing normal working condition of an arbitrary asset.
Training an anomaly detection model, given the acquired normal working condition samples.
Deploying the anomaly detection model and calculating the deviation from the provided reference asset behavior.
Converting deviation values to health indices.
Predicting the future health index values for the asset for calculating its remaining useful life.
Planning maintenance.
Figure 6.
Anomaly detection-based predictive maintenance.
Such an approach reduces the impact of the absolute value of the asset readings and puts more emphasis on the deviation from the expected values. Furthermore, in a rare case of a failure in the asset, the acquired datapoint can be used to update the thresholds and the different hyperparameters of the deployed model.
In the upcoming sections, more complex implementation and model evaluation topics for predictive maintenance solutions are introduced.
2.2 Efficient predictive maintenance: reducing sensors
There are several factors impacting the deployment costs of a predictive maintenance solution (see Figure 7). One of the main factors preventing SMEs or OEMs for integrating predictive maintenance in their ecosystem is related to the costs and effort associated with sensor installation. Especially, for companies that own fleets of production assets, installing additional sensors on each of these units can be time-consuming and expensive. The issues with sensor installation are further exasperating for pharmaceutical companies or producers of safety-critical system as the production assets have to be certified again and also reevaluated, given the adjustments made in the asset during sensor installation. This process can reduce the productivity particularly at the beginning of deploying the solution. Apart from the effort needed for sensor installation, as the information acquired from sensors is the very foundation of decisions made by the predictive maintenance solution, it is of utmost importance to ensure the quality of the data fed to it. Noisy, incomplete, and high latency in sensor reading are some of the common issues deteriorating the data quality and consequently the effectiveness of the developed predictive maintenance solution [12]. As a result, to prevent issues with data readings from the sensors, they need to be maintained as well, which could potentially lead to a never-ending loop of maintenance.
Figure 7.
Predictive maintenance costs and profit.
One effective way for reducing the number of sensors is by inspecting the impact of different sensors on the predictions made by the predictive maintenance solution. Thereafter, sensor readings can be chosen if their roll in the predictions can be physically justified, given the domain knowledge from the asset. In addition, such important readings must also have a significant contribution to the accuracy of the prediction model (see Figure 8), so that their selection can also be validated from a deployment cost point of view. This approach is thoroughly examined in the next subsection.
Figure 8.
Cost-effective predictive maintenance.
2.3 Interpretable predictive maintenance: verifying prediction models with domain knowledge
An important characteristic of data-driven predictive maintenance solutions ensuring their acceptance in industry is their interpretability. Model interpretability can help verify the trained prediction model and have an overview of how it uses different sources of information from the asset [13]. In fact, by using different techniques introduced in interpretable AI, it is possible to infer the role of different asset readings on the decisions made by the prediction model and in case of a wrong prediction, the detected error can be elucidated, given this information. Provided that the deployed predictive maintenance solution complies with the requirements of trustworthy AI [14], it is also possible to examine which datapoints were used to train the model that led to the wrong prediction of the model and prevent similar incidents in the future. Under the circumstances that some domain knowledge is available from the maintenance crew of the SME or OEM, it is also possible to visualize the feature importance of the prediction model and have the experts verify the behavior of the model. In addition, by inspecting the importance of different asset readings, it is feasible to inspect if there are any specific readings that dictate the output of the prediction model and whether or not these sensor readings can be relied on under different working conditions. For ensuring the acceptable performance of the predictive maintenance solution, it is recommended to have a prediction model that generates its output, given a wider range of sensor readings with no dominant peak in importance values (see Figure 9).
Figure 9.
Role of model interpretability on the evaluation of a predictive maintenance solution.
2.4 Maintaining predictive maintenance models in industry
One of the main goals of Industry 4.0/5.0 is to adapt different assets of a production line to meet the requirements of the end user. This adaptability in production often leads to changes in the read asset data [15, 16]. Therefore, for a successful and impactful deployment of a predictive maintenance solution it is crucial to develop measures for maintaining the prediction model [17] and protecting it from different types of data distribution shifts (see Figure 10). Some of the most evident distribution shifts or potential reasons for them in different predictive maintenance implementations are the following:
Imbalanced data: prior to deploying a new predictive maintenance solution, the availability of a previously deployed preventive maintenance solution can impact the distribution of datapoints attained from the asset from different working conditions. If the deployed preventive maintenance solution had a low performance or was nonexistent, numerous instances of system failure could have been gathered. On the other hand, if the allocated preventive maintenance solution hindered gathering samples that represent different failure types in the targeted asset, then such failure samples would be scarce. The ratio of datapoints representing normal working condition and erroneous states of the system will ultimately change when an effective predictive maintenance solution is deployed. The unbalance in the data should always be accounted for in different parts of the data pipeline and model training by either upweighting some of the gathered samples from the production asset or by under/oversampling the instances of some working conditions. In fact, imbalanced data reflect sample selection bias with a known bias value that is determined by the label of the class gathered from the asset [18].
Covariate shift: given the possible changes in the working condition and also the external factors from the production shop floor impacting an asset, these changes could potentially have not been reflected in the data used to train a predictive maintenance solution. Therefore, the model is not qualified to make predictions for such new production settings. If the aforementioned changes are not tracked for evaluating the predictions of the trained model, they could be misleading as the trained model is working under different assumptions [19].
Source component shift: in many industrial settings, different production assets are reassigned for different tasks than they were initially used for. The new production recipe in this case will result in data with different characteristics. For such occurrences, the previously trained model is again obsolete and needs to be adapted to the new production circumstances. As soon as different models for different production recipes are available, the next step would be to first identify which task the read asset data are indicating and then use the correct prediction model to evaluate the production asset status [20].
Figure 10.
Data imbalance and distribution shifts in predictive maintenance.
As it can be seen, there are numerous internal and external factors impacting the gathered data from an asset. This constant change in the data influences the performance of the data-driven predictive maintenance solution. Thus, it is inevitable to have an administration unit for tracking the changes in the deployed models and also to log the changes in the data for future references. One potential solution for fulfilling this task is the AAS [21], which is introduced in the next section.
3. AAS as a potential performance booster for predictive maintenance
The landscape of predictive maintenance is continuously evolving, driven by advancements in technology and the increasing complexity of industrial operations. As organizations strive to optimize asset management and minimize downtime, the concept of the AAS emerges as a potential game-changer in enhancing the performance of predictive maintenance solutions. In this section, we delve into the role of AAS as a facilitator of interoperability, data exchange, and performance enhancement in the context of predictive maintenance.
3.1 Understanding the Asset Administration Shell
Asset Administration Shell (AAS) represents a standardized framework within the context of the RAMI4.0. At its core, the AAS provides a digital representation of physical assets and their associated data, enabling seamless integration and communication across heterogeneous systems. By encapsulating comprehensive information about an asset, including its structure, behavior, and lifecycle data, the AAS fosters interoperability and transparency within industrial environments. A detailed description of the AAS information model can be found in the specification (see Figure 11) [22]. The AAS encapsulates comprehensive information about each asset, including its structure, behavior, and lifecycle data. This information is structured according to a predefined information model that defines the attributes and relationships necessary to describe an asset comprehensively. By adhering to this standardized format, the AAS ensures interoperability and transparency, enabling stakeholders to exchange data and information effectively. One of the key aspects of the AAS is its ability to establish digital twins and virtual representations of physical assets. These digital twins serve as virtual counterparts to their physical counterparts, providing real-time insights into asset health, performance trends, and maintenance requirements. By leveraging digital twins, organizations can monitor assets remotely, identify anomalies or potential issues, and optimize maintenance schedules proactively.
Figure 11.
Three-dimensional (3D) representation of the reference RAMI4.0 based on [
3.2 Leveraging AAS for predictive maintenance
Integrating the principles of AAS into predictive maintenance initiatives holds immense potential for enhancing the efficiency and effectiveness of asset management strategies. Through the establishment of digital twins and virtual representations of physical assets, organizations can gain insights into asset health, performance trends, and maintenance requirements in real-time. By adopting the AAS framework, organizations can achieve several benefits in the context of predictive maintenance:
Data integration and aggregation : the AAS enables the aggregation of disparate data sources, including sensor data, maintenance records, and operational parameters, into a unified digital representation. These aggregated data serve as the foundation for predictive analytics and machine learning algorithms, empowering organizations to extract actionable insights and make informed decisions regarding asset maintenance and optimization.Enhanced transparency : the AAS enables a high level of transparency regarding asset status, performance metrics, and maintenance requirements. By encapsulating all relevant information within the asset’s administration shell, stakeholders gain comprehensive visibility into the asset’s condition and operational history. This transparency facilitates informed decision-making regarding maintenance scheduling, resource allocation, and asset optimization strategies.Improved data accessibility : the AAS facilitates the centralized storage and management of asset-related data, ensuring easy accessibility for relevant stakeholders across the organization. With data stored in a standardized format within the asset’s administration shell, authorized personnel can retrieve and analyze critical information efficiently. This accessibility streamlines data-driven decision-making processes, enabling timely interventions and proactive maintenance activities.Integration with predictive maintenance systems : the AAS framework can seamlessly integrate with predictive maintenance systems, providing them with access to real-time asset data and status updates. By leveraging AAS interfaces and communication protocols, predictive maintenance solutions can retrieve pertinent information directly from the asset’s administration shell. This integration enhances the accuracy and effectiveness of predictive maintenance models by incorporating up-to-date data and contextual insights.Interoperability and interconnectivity : by adhering to standardized AAS formats and interfaces, predictive maintenance solutions can seamlessly communicate with diverse industrial systems and components. This interoperability facilitates the exchange of data and information across the entire value chain, enabling enhanced collaboration between stakeholders and ensuring holistic asset management strategies.Enabling data-driven insights : with the comprehensive data stored within the asset’s administration shell, organizations can harness advanced analytics and machine learning algorithms to derive actionable insights. By analyzing historical performance data, identifying patterns, and predicting future maintenance needs, organizations can optimize asset utilization, minimize downtime, and enhance overall operational efficiency. The structured nature of data within AAS facilitates sophisticated analytics, enabling organizations to unlock valuable insights for predictive maintenance optimization.Ensuring security and integrity : the AAS incorporates robust security mechanisms to safeguard asset data and prevent unauthorized access or tampering. By implementing encryption, authentication, and access control measures, organizations can mitigate cybersecurity risks and ensure the integrity of asset-related information. This security framework is essential for maintaining trust and reliability in predictive maintenance systems, particularly in industrial environments where data confidentiality and integrity are paramount.
3.3 Case studies and success stories
Real-world case studies and success stories demonstrate the tangible benefits of leveraging AAS for predictive maintenance across various industries. These case studies highlight the implementation challenges, best practices, and measurable outcomes achieved through the integration of AAS into predictive maintenance strategies. Examples include:
Automotive industry : Volkswagen Sachsen GmbH implements AAS-based solutions to optimize production line efficiency, reduce unplanned downtime, and improve product quality. By leveraging digital twins and real-time data analytics, the automotive manufacturer achieves significant cost savings and operational enhancements across its manufacturing facilities [24].Manufacturing sector : Hitachi, a multinational semiconductor manufacturer, uses the AAS as a component of its predictive maintenance framework, enabling proactive asset management and predictive analytics. By standardizing asset data formats, optimizing data exchange processes, and implementing algorithms for predictive maintenance, the company achieves new levels of operational efficiency and asset reliability, resulting in higher and consistent product quality [25].
3.4 Future directions and emerging trends
Looking ahead, the integration of AAS into predictive maintenance is poised to drive further innovation and transformation in industrial asset management. Emerging trends, such as edge computing, artificial intelligence, and digital twins, promise to revolutionize predictive maintenance capabilities, offering organizations new opportunities for optimization and competitive advantage.
Edge computing, for instance, brings computational power closer to the data source, enabling real-time analysis and decision-making at the network’s edge. By deploying edge computing solutions in conjunction with AAS-enabled predictive maintenance, organizations can enhance responsiveness and reduce latency, thereby improving overall operational efficiency.
Artificial intelligence (AI) plays a pivotal role in augmenting predictive maintenance capabilities by enabling advanced analytics, anomaly detection, and predictive modeling. By leveraging AI algorithms, organizations can extract valuable insights from large volumes of data generated by AAS-enabled systems, facilitating proactive maintenance interventions and optimizing asset performance.
Digital twins represent another transformative trend in predictive maintenance, offering virtual replicas of physical assets that mirror their real-world behavior and characteristics. By creating digital twins within the AAS framework, organizations can simulate asset performance, conduct scenario analysis, and optimize maintenance strategies in a risk-free virtual environment.
As organizations continue to embrace digitalization and Industry 4.0 initiatives, the integration of AAS into predictive maintenance strategies will play a pivotal role in shaping the future of asset management and maintenance practices. By keeping up with these trends and embracing technological advancements, organizations can fully exploit the capabilities of predictive maintenance solutions enabled by AAS, driving continuous improvement and innovation in industrial operations.
4. Predictive maintenance in OEM vs. SME: differences and pivoting points
Currently, the main driving forces for implementing predictive maintenance for different industrial settings are reducing the downtime, avoiding micro-stops, and increasing production. However, one of the more important side benefits of predictive maintenance that is normally neglected in the related literature is the impact of production asset state on the produced product quality [26]. As an example, the wear and tear in pneumatic actuators moving the tool center point of a production asset influences highly the precision of the production asset, which as a result deteriorates the final product. For both OEMs and SMEs, it is crucial to reduce the scrap rate during production for minimizing the raw material consumption and decreasing the production cost. Apart from production cost reduction, by maintaining the production assets in an acceptable condition, it is possible to ensure certain final product quality and prevent the need for providing replacements for instances that did not meet the customer’s needs. Furthermore, avoiding high scrap rate also reduces the energy consumption during production which boosts sustainable production.
Apart from boosting performance and sustainability in production, OEMs can use predictive maintenance as an important source of income in their business model. By offering smart production assets with functioning predictive maintenance solutions, the end user will indirectly benefit from the OEM’s expertise during production for reaching their goal with comparably less effort and resource use. In addition, as soon as the end user has gathered enough data during production, they can adapt the provided generic predictive maintenance solution from the OEM to better suit their needs and also analyze the impact of the production asset on the quality of the produced product. Once the aforementioned datasets are produced, the end user can also integrate selling data to other end users with the similar production assets to their business model (see Figure 12). More details regarding privacy and data protection are provided in the next section.
Figure 12.
Smart asset and maintenance solution sharing for increasing revenue.
5. Privacy and data protection in predictive maintenance aimed towards the end customer
With the continuous advancement of data-driven solutions in different areas of research, numerous key figures in industry have grasped the significance of proper digitalization and data acquisition for their assets. The importance, effort, and costs associated with industrial data have boosted their value up to a point that some business models in industry suggest sharing and selling industrial data as a crucial source of income, especially for OEMs. Nonetheless, one major ethical question that is raised here, is
To protect the end user of a production asset and also to enable OEMs to benefit from their efforts in producing smart production assets, there are multiple ways to indirectly share data between different users without compromising privacy. Possible solutions include but are not limited to data encryption [27] and federated learning [28]. By encrypting the data, no raw data will be shared among different users protecting different companies from their production plans to be revealed. However, the data-driven models used in these scenarios need to be capable of using the encrypted data from different sources in a way that boosts their performance.
Another method for sharing production asset, without directly sharing the data from different users, is to employ federated learning. By using this method, the trained models from different users are shared and retrained using the data from other customers (see Figure 13). In case the model accuracy increases, the newly trained model will be adapted by the initial user. This solution does not require actual sharing of the data and only the trained models will be passed around. Nevertheless, it is possible to attain the initially used dataset to train a model from many solutions using deep neural networks. Therefore, in such scenarios, measures needed to protect the data from being extracted from the trained model need to be taken to ensure maximum data privacy for the end user.
Figure 13.
Preventing direct data sharing via federated learning.
6. Conclusion
In this chapter, different internal and external factors impacting various aspects of planning, implementing, and deploying effective predictive maintenance solutions were discussed. It was shown how the availability of data, data quality, and data annotation can steer the design process. Furthermore, important concepts in trustworthy AI were introduced, which can help maintain and boost the performance of the deployed predictive maintenance solutions. It was also explained in detail how model interpretability can help reduce the required number of sensors and thus lower the deployment costs, motivating different customers to equip their production assets with sensors and monitor their efficiency. In addition, AAS as a promising standardization for different Industry 4.0/5.0 use cases in Europe was introduced. It was discussed how AAS can help track asset data in different industrial settings, and how the unified digital representation can then be used in the data pipeline and also model training. The standardized data attained from different assets can ultimately enhance the performance of the deployed data-driven solutions. Afterwards, it was shown how predictive maintenance can impact the business model of OEMs and SMEs, as they can earn additional money just by providing the foundation needed for gathering data from the production assets and later share the data with other users as a service. Lastly, possible issues with direct data sharing were pointed out and federated learning as a feasible solution for improving predictive maintenance models without data sharing was introduced. The reader of this chapter must now have a broad overview on the most demanding and complicated issues for developing and deploying predictive maintenance solution in industry. Depending on their need, the reader can further pursue different topics that are currently needed in industry, such as model lifecycle support, edge device implementation of prediction models, and local and cloud data preparation and structuring for production assets.
Acknowledgments
The authors of this book chapter would like to express their deep sense of gratitude to SMC Schweiz AG, especially Mr. Alessandro Grizzetti, for their unceasing assistance in promoting Competence Centre for Automation and Digitalisation among different industrial partners in Europe.
Abbreviations
original equipment manufacturer | |
small and medium-sized enterprise | |
Internet of Things | |
Asset Administration Shell | |
Reference Architecture Model Industrie 4.0 | |
true positive rate | |
true negative rate | |
false positive rate | |
false negative rate | |
artificial intelligence |
References
- 1.
Farahani A, Voghoei S, Rasheed K, Arabnia HR. A brief review of domain adaptation. Advances in Data Science and Information Engineering: Proceedings from ICDATA 2020 and IKE 2020. 2021:877-894 - 2.
Zhou K, Liu Z, Qiao Y, Xiang T, Loy CC. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2022; 45 (4):4396-4415 - 3.
Nejjar I, Geissmann F, Zhao M, Taal C, Fink O. Domain adaptation via alignment of operation profile for remaining useful lifetime prediction. Reliability Engineering & System Safety. 2024; 242 :109718 - 4.
Rahat M, Mashhadi PS, Nowaczyk S, Rognvaldsson T, Taheri A, Abbasi A. Domain adaptation in predicting turbocharger failures using vehicle’s sensor measurements. Phm Society European Conference. 2022; 7 :432-439 - 5.
Didona D, Quaglia F, Romano P, Torre E. Enhancing performance prediction robustness by combining analytical modeling and machine learning. In: Proceedings of the 6th ACM/SPEC International Conference on Performance Engineering. 2015. pp. 145-156. DOI: 10.1145/2668930.2688047 - 6.
Lei Y, Li N, Gontarz S, Lin J, Radkowski S, Dybala J. A model-based method for remaining useful life prediction of machinery. IEEE Transactions on Reliability. 2016; 65 (3):1314-1326 - 7.
Chandola V, Banerjee A, Kumar V. Anomaly detection: A survey. ACM Computing Surveys (CSUR). 2009; 41 (3):1-58 - 8.
Fathi K, van de Venn HW, Honegger M. Predictive maintenance: An autoencoder anomaly-based approach for a 3 dof delta robot. Sensors. 2021; 21 (21):6979 - 9.
Monaghan TF, Rahman SN, Agudelo CW, Wein AJ, Lazar JM, Everaert K, et al. Foundational statistical principles in medical research: Sensitivity, specificity, positive predictive value, and negative predictive value. Medicina. 2021; 57 (5):503 - 10.
Hurtado J, Salvati D, Semola R, Bosio M, Lomonaco V. Continual learning for predictive maintenance: Overview and challenges. Intelligent Systems with Applications. 2023; 19 :200251 - 11.
Carrasco J, López D, Aguilera-Martos I, García-Gil D, Markova I, Garcia-Barzana M, et al. Anomaly detection in predictive maintenance: A new evaluation framework for temporal unsupervised anomaly detection algorithms. Neurocomputing. 2021; 462 :440-452 - 12.
Renggli C, Rimanic L, Gürel NM, Karlaš B, Wu W, Zhang C. “A data quality-driven view of mlops”. 2021. arXiv preprint arXiv:2102.07750 - 13.
Steurtewagen B, Van den Poel D. Adding interpretability to predictive maintenance by machine learning on sensor data. Computers & Chemical Engineering. 2021; 152 :107381 - 14.
Li B, Qi P, Liu B, Di S, Liu J, Pei J, et al. Trustworthy ai: From principles to practices. ACM Computing Surveys. 2023; 55 (9):1-46 - 15.
Sculley D, Holt G, Golovin D, Davydov E, Phillips T, Ebner D, et al. Hidden technical debt in machine learning systems. Advances in Neural Information Processing Systems. 2015; 28 . Available from:https://proceedings.neurips.cc/paper_files/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html - 16.
Paleyes A, Urma R-G, Lawrence ND. Challenges in deploying machine learning: A survey of case studies. ACM Computing Surveys. 2022; 55 (6):1-29. Available from:https://proceedings.neurips.cc/paper_files/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html - 17.
Kreuzberger D, Kühl N, Hirschl S. Machine learning operations (mlops): Overview, definition, and architecture. IEEE Access. 2023. pp. 31866-31879 - 18.
Quiñonero-Candela J, Sugiyama M, Schwaighofer A, Lawrence ND. Dataset Shift in Machine Learning. Mit Press; 2008. Available from: https://mitpress.mit.edu/9780262545877/dataset-shift-in-machine-learning/ - 19.
Huyen C. Designing Machine Learning Systems. O’Reilly Media, Inc; 2022. Available from: https://www.oreilly.com/library/view/designing-machine-learning/9781098107956/ - 20.
Fathi K, Sadurski M, Kleinert T, van de Venn HW. Source component shift detection & classification for improved remaining useful life estimation in alarm-based predictive maintenance. In: 2023 23rd International Conference on Control, Automation and Systems (ICCAS), Yeosu, Korea. IEEE; 2023. pp. 975-980 - 21.
Cavalieri S, Salafia MG. A model for predictive maintenance based on asset administration shell. Sensors. 2020; 20 (21):6028 - 22.
Bader S, Barnstedt E, Bedenbender H, Berres B, Billmann M, Ristin M. “Details of the Asset Administration Shell-Part 1: The Exchange of Information Between Partners in the Value Chain of Industrie 4.0 (Version 3.0 rc02)”. Berlin, Germany: Federal Ministry for Economic Affairs and Climate Action (BMWK); 2022 - 23.
Spec D. 91345: Referenzarchite-kturmodell industrie 4.0 (rami4. 0). Deutsches Institut für Normung (DIN) eV. 2016 - 24.
Case IU. “A Uniform Information Model for Manufacturing”. 2024.[Online]. Available from:: https://iotusecase.com/en/use-cases/a-uniform-information-model-for-manufacturing/ - 25.
Fathi K, Stramaglia M, Ristin M, Sadurski M, Kleinert K, Schönfelder R, et al. 12th IFAC Symposium on Fault Detection, Supervision and Safety for Technical Processes: Safeprocess 2024: 4-7 June 2024: Proceedings, Ferrara, Italy. Elsevier Science Direct - IFAC PapersOnline; 2024 - 26.
Acebo AG, Ventura-Traveset A, Uriarte S d l M. Remote condition monitoring and control for maintenance activities in coordinate measurement machine: A data-driven approach for the metrology engineering industry. Predictive Maintenance in Smart Factories: Architectures, Methodologies, and Use-cases. 2021:121-147. Available from: https://link.springer.com/chapter/10.1007/978-981-16-2940-2_6#citeas - 27.
Podschwadt R, Takabi D, Hu P, Rafiei MH, Cai Z. A survey of deep learning architectures for privacy-preserving machine learning with fully homomorphic encryption. IEEE Access. 2022; 10 :117477-117500. Available from:https://link.springer.com/chapter/10.1007/978-981-16-2940-2_6#citeas - 28.
Zhang C, Xie Y, Bai H, Yu B, Li W, Gao Y. A survey on federated learning. Knowledge-Based Systems. 2021; 216 :106775