The complexity of modern electro-mechanical systems require the development of sophisticated diagnostic methods like anomaly detection capable of detecting deviations. Conventional anomaly detection approaches like signal processing and statistical modelling often struggle to effectively handle the intricacies of complex systems, particularly when dealing with multi-variate signals. In contrast, neural network-based anomaly detection methods, especially Auto-Encoders, have emerged as a compelling alternative, demonstrating remarkable performance. However, Auto-Encoders exhibit inherent opaqueness in their decision-making processes, hindering their practical implementation at scale. Addressing this opacity is essential for enhancing the interpretability and trustworthiness of anomaly detection models. In this work, we address this challenge by employing a feature selector to select features and counterfactual explanations to give a context to the model output. We tested this approach on the SKAB benchmark dataset and an industrial time-series dataset. The gradient based counterfactual explanation approach was evaluated via validity, sparsity and distance measures. Our experimental findings illustrate that our proposed counterfactual approach can offer meaningful and valuable insights into the model decision-making process, by explaining fewer signals compared to conventional approaches. These insights enhance the trustworthiness and interpretability of anomaly detection models.

Machine learning-based condition monitoring of mechanical systems, such as bearings, employs two primary approaches: unsupervised and supervised methods. Unsupervised approaches aim to characterize the healthy state of the machine and monitor deviations from this state. The advantage lies in requiring only the health condition of the component without the need for historical data until breakdown. However, the disadvantage is the lack of information regarding the root cause of any potential malfunction. On the other hand, supervised methods consider both healthy and faulty cases, aiming to maximize the difference between them through post-processing, as well as among different fault types. The advantage is the ability to analyze the specific signature of a particular fault type. Nonetheless, the disadvantage is that available data usually do not cover all possible faults that may occur.Typically, obtaining a faulty bearing involves either a time-consuming run-to-failure test or the artificial induction of faults using drills, electro-discharge pens, etc. While artificial faults offer a quicker procedure, they often fail to replicate real faults faithfully. This paper suggests using picosecond laser technology to engrave the surface of the bearing and create artificial faults. Modern laser technology allows for precise control over the dimensions of injected faults, enhancing the understanding of fault progression at various stages in the life of bearings. These measurements are crucial parameters for evaluating the robustness of diagnostic algorithms. This paper focuses on artificially damaging a ball bearing used in an independent cart systems application, which comprises a fleet of linear motors moving on the same rail. These systems have recently been proposed by different manufacturers and adopted in the field of packaging machines for their flexibility. For such systems, no prior instances of faulted bearings are available, and the size of a real fault is also unknown. Hand-made faults with drills did not produce discernible faults appreciable in post-processing of the data. Therefore, a picosecond laser with a pulse duration of 10 ps and a maximum energy per pulse of approximately 100 µJ is utilized to create a set of test bearings with increasing fault sizes on the outer race. Post-processing of the data enables the qualification of the minimum fault severity detectable in this specific application.

The introduction of cyber-physical systems with increased availability of sensor data creates a lot of research interest in prognostic algorithms for predictive maintenance. Although a lot of algorithms are successfully applied to benchmark case studies based on simulated data and experimental set-ups, deploymentin industry lags behind. From a comparison between three benchmark case studies with two real-world case studies based on prognostic metrics (monotonicity, prognosability and trendability), two main issues are observed: 1) the lack of run-to-failures and 2) low prognostic metrics due to a low signal-to-noise ratio of degradation trends, as a result of unexplained physical phenomena. To make prognostics feasible, a hybrid framework is proposed that focuses on improving system knowledge. The framework consists of a quantitative diagnostic assessments, guided by (modular) system models in which damage is induced. This quantitative damage assessment provides input for prognostics based on Bayesian filtering, enabling prognostics for assets in varying operational conditions. Implementation and validation of the framework requires investments, but modularity within the framework can accelerate development for new systems.

Nuclear plant operators are required to understand the uncertainties associated with the deployment of prognostics toolsin order to justify their inclusion in operational decision making processes and satisfy regulatory requirements. Operationaluncertainty can cause underlying prognostics models to underperform on assets that are subject to evolving impactsof age, manufacturing tolerances, operating conditions, and operating environment effects, of which may be capturedthrough a condition monitoring (CM) system that itself may be degraded. Sources of uncertainty in the data acquisitionpipeline can impact the health of CM data used to estimate the remaining useful life (RUL) of assets. These uncertaintiescan disguise or misrepresent developing faults, where (for example) the fault identification is not achieved until it hasprogressed to an unmanageable state. This leaves little flexibility for the operator’s maintenance decisions and generallyundermines model confidence.
 One method to quantify and account for operational uncertainty is calibrated hybrid models, employing physics, knowledgeor data driven methods to improve model accuracy and robustness. Hybrid models allow known physical relations tooffset full reliance on potentially untrustworthy data, whilst reducing the need for an abundance of representative historicaldata to reliably identify the monitored asset’s underlying behavioural trends. Calibration of the model then ensuresthe model is updated and representative of the real monitored asset by accounting for differences between the physics orknowledge model and CM data. 
 In this paper, an open-source bearing knowledge informed machine learning (ML) model and CM datasets are utilizedin an illustrative bearing prognostic application. The uncertainty incurred by the decisions made at key stages in thedevelopment of the model’s data acquisition and processing pipeline are assessed and demonstrated by the resultant impacton RUL prediction performance. It was shown that design decisions could result in multiple valid pipeline designswhich generated different predicted RUL trajectories, increasing the uncertainty in the model output.

This paper presents XGBoost classifier-based methods to solve three tasks proposed by the European Prognostics and Health Management Society (PHME) 2022 conference. These tasks are based on real data from a Surface Mount Technologies line. Each of these tasks aims to improve the efficiency of the Printed Circuit Board (PCB) manufacturing process, facilitate the operator’s work and minimize the cases of manual intervention. Due to the structured nature of the problems proposed for each task, an XGBoost method based on encoding and feature engineering is proposed. The proposed methods utilise the fusion of test values and system characteristics extracted from two different testing equipment of the Surface Mount Technologies lines. This work also explores the problems of generalising prediction at the system level using information from the subsystem data. For this particular industrial case: the challenges with the changes in the number of subsystems. For Industry 4.0, the need for interpretability is very important. This is why the results of the models are analysed using Shapley values. With the proposed method, our team took the first place, capable of successfully detecting at an early stage the defective components for tasks 2 and 3.

The existing helicopter Health and Usage Management Systems (HUMS) collect and process flight operational parameters and sensors data such as vibrations to provide health monitoring of the helicopter dynamic assemblies and engines. So far, structure-related mechanical faults, such as looseness in bolted structures, have not been addressed by vibration-based condition monitoring in existing HUMS systems. Bolt loosening was identified as a potential risk to flight safety demanding periodical visual monitoring, and increased maintenance and repair expenses. Its automatic identification in helicopters by using vibration measurements is challenging due to the limited number of known events and the presence of high-energy vibrations originating in rotating parts, which shadow the low-level signals generated by the bolt loosening.
 New developed bolt loosening monitoring approach was tested on HUMS vibrations data recorded from the IAF AH-64 Apache helicopters fleet. ML-based unsupervised anomaly detection was utilized in order to address the limited number of faulty cases. The predictive power of health features was significantly improved by applying the Harmonic filtering differentiating between the high-energy vibrations generated by rotating parts compared with the low-energy structural vibrations. Different unsupervised anomaly detection techniques were examined on the dataset. The experimental results demonstrate that the developed approach enable successful bolt loosening monitoring in helicopters and can potentially be used in other health monitoring applications.

Due to the global transition to electromobility and the associated increased use of high-performance batteries, research is increasingly focused on estimating and forecasting the state of health (SOH) of lithium-ion batteries. Several data-intensive and well-performing methods for SOH forecasting have been introduced. However, these approaches are only reliable for new battery types, e.g., with a new cell chemistry, if a sufficient amount of training data is given, which is rarely the case. A promising approach is to transfer an established model of another battery type to the new battery type, using only a small amount of data of the new battery type. Such methods in machine learning are known as transfer learning. The usefulness and applicability of transfer learning and its underlying methods have been very successfully demonstrated in various fields, such as computer vision and natural language processing. Heterogeneity in battery systems, such as differences in rated capacity, cell cathode materials, as well as applied stress from use, necessitate transfer learning concepts for data-based battery SOH forecasting models. Hereby, the general electrochemical behavior of lithium-ion batteries, as a major common characteristic, supposedly provides an excellent starting point for a transfer learning approach for SOH forecasting models. In this paper, we present a transfer learning approach for SOH forecasting models using a multilayer perceptron (MLP). We apply and evaluate it on the method presented by von Bülow, Mentz, and Meisen (2021) using five battery datasets. In this regard, we investigate the optimal conditions and settings for the development of transfer learning with respect to suitable data from the target domain, as well as hyperparameters such as learning rate and frozen layers. We show that for the transfer of a SOH forecasting model to a new battery type it is more beneficial to have data of few old batteries, compared to data of many new batteries, especially in the case of superlinear degradation with knee points. Contrarily to computer vision freezing no layers is preferable in 95% of the experimental scenarios.

A possible way to increase the reliability and availability of a system is to apply an Active Fault Tolerant Control (AFTC) algorithm. This thesis aims to use this algorithm in a multiengine propulsive cluster with sensor and actuator faults. First, a Health Monitoring System (HMS) will be developed to monitor the entire propulsive cluster. The HMS will use model-based fault diagnosis techniques. Then, in case of actuator faults, the cluster will be reconfigured to minimize its effects. The reconfiguration can be made by using control allocation or modifying the control law of the engine. A simulation model of the entire cluster is under development. The model simulates the whole system, including the propellant feeding system, engines, and mechanical system. It will be used to study the effect of different faults on the system and compare different reconfiguration strategies.

In Prognostics and Health Management, there are three main approaches for implementing diagnostic and prognostic applications. These approaches are data-driven methods, physical model-based methods, and combinations of them, in the form of hybrid methods. Each of them has specific advantages but also limitations for their purposeful implementation. In the case of data-driven methods, one of the main limitations is the availability of sufficient training data that adequately cover the relevant state space. For model-based methods, on the other hand, it is often the case that the degradation process of the considered technical system is of significant complexity. In such a scenario physics-based modeling requires great effort or is not possible at all. Combinations of data-driven and model-based approaches in form of hybrid approaches offer the possibility to partially mitigate the shortcomings of the other two approaches, however, require a sufficiently detailed data-driven and physics-based model. This paper addresses the transitional field between data-driven and hybrid approaches. Despite the issues of formulating a physics-based model that provides a representation of the degradation process, basic knowledge of the considered system and of the laws governing its degradation process is usually available. Integration of such knowledge into a machine learning process is part of a research field that is either called theory-guided data science, (physics) informed machine learning, physics-based learning or physics guided machine learning. First, the state of research in Prognostics and Health Management on methods of this field is presented and existing research gaps are outlined. Then, a concept is introduced for incorporating fundamental knowledge, such as monotonicity constraints, into data-driven diagnostic and prognostic applications using approaches from theory-guided data science. A special aspect of this concept is its cross-application usability through the consideration of knowledge that repeatedly occurs in diagnostics and prognostics. This is, for example, knowledge about physically justified boundaries whose compliance makes a prediction of the data-driven model plausible in the first place.

The analysis of the behaviour of complex mechanical components in order to identify relevant patterns for health monitoring and diagnostics is a complex task. One example of this complexity is the data workflow that can be generated for analysis purposes when a new prototype of an electro-mechanical actuator (EMA) is being designed and experimented. The most accurate way of getting valuable insights during this analysis is by running tests in a controlled environment. In this regard, a test bench can be used to simulate the operating conditions experienced by the actuator during its operational lifetime in an accelerated manner.
 Depending on the number and nature of the parameters to be obtained, the prototype’s functionality under study, and the test frequencies and experiment duration, big data challenges may appear (volume, variety, and velocity). The analysis of such data and the development of adequate data-driven algorithms usually involves computations on massive amounts of data, a flexible and automated data pipeline, repeatable and reliable operations, and a collaborative approach. The use of DataOps techniques on a cloud infrastructure can provide the ideal environment for creating an end-to-end integrated architecture which meets these requirements.
 This work describes a data-driven information system developed for an electro mechanical actuator on a test bench. It uses a multivariate statistical process control (SPC) and a linear discriminant analysis (LDA) algorithm for detecting and evaluating the evolution of the actuator’s health. The information system runs an automated data pipeline on a cloud platform with signals obtained on the test bench, leveraging DataOps and machine learning techniques for a flexible and scalable data management.