Predictive maintenance (PdM) is increasingly being used to improve system reliability, reduce downtime, and optimize operating costs in complex industrial environments through anomaly detection. As industrial systems become more complex, with components interacting with one another and monitored by ever-expanding sensor networks, anomaly detection faces new methodological challenges. In particular, there is a gap in the literature regarding anomaly detection using traditional machine learning (ML) methods based on a Single Label Classification (SLC) approach, which is unsuitable when multiple failures can occur simultaneously.
Traditional anomaly detection approaches based on ML methods using an SLC formulation assume that each observation belongs to a single anomaly category. This assumption does not hold when multiple failures occur simultaneously on the same equipment, resulting in a loss of information. We begin by illustrating this limitation through a concrete example showing that using an SLC approach instead of a multi-label formulation degrades the classification of simultaneous anomalies. To address this issue, anomaly detection is reformulated as a Multi-Label Classification (MLC) problem, enabling the detection of multiple failures within a single time step. Following this reformulation, the issue of selecting multi-label classifiers suited to the context of Predictive Maintenance (PdM) must be examined. The general conclusion from the literature is that the performance of classifiers depends heavily on the context and that while there are many detection methods, none is universally dominant. Although in-depth comparative studies exist in the field of MLC, datasets from complex systems are absent from these comparisons, even though they exhibit specific characteristics such as multivariate time-series structures, significant label imbalance, and complex interactions between components.
To address this gap, a structured and reproducible evaluation protocol is proposed to evaluate eight state-of-the-art ML methods in an MLC framework across three public industrial datasets. The results confirm that the performance of classifiers depends heavily on the characteristics of the dataset and that no single method consistently outperforms others across all scenarios. However, conducting this type of comparison for each new industrial context is costly in terms of time, computational resources, and data requirements, which limits its feasibility for real-world industrial deployment. The thesis also addresses this issue through dimensionality reduction and variable selection methods. PM systems generate large multivariate datasets in which only a subset of the variables is actually relevant for anomaly detection. A self-adaptive evolutionary strategy is proposed to perform wrapper-style variable selection and obtain a subset containing only the most informative variables.
Experimental results show that reducing the variable space improves computational efficiency and, in most cases, anomaly prediction performance. A comparison of the proposed method with state-of-the-art metaheuristic approaches on three PdM datasets and five anomaly detection methods confirms its competitiveness in terms of predictive performance and the optimal size of the selected feature subset. Finally, feature space reduction not only optimizes the resources required for detection but also reduces the number of signals that technicians must analyze during root cause investigations. This simplification aims to strengthen confidence in the system. Finally, the thesis addresses the practical aspects of deploying detection methods in complex systems. Among these, there is a lack of trust and transparency among maintenance technicians and decision-makers.
An analysis of the interpretability and robustness of anomaly detection methods is provided to directly contribute to their deployment in real-world conditions. First, the proposed variable selection strategy is evaluated on a public PdM dataset using SHAP-based explanations to verify the consistency between the selected variables and those identified as the most important by tools dedicated to interpretability. The results show that the selected variables are consistent with those identified as the most important by these tools. Furthermore, visual analysis of the most influential variables reveals that anomalies are associated with a limited and well-defined subset of sensor signals, which helps improve transparency and strengthen confidence in automatic detection systems.
At the same time, it is necessary to conduct a study of the robustness of multi-label classifiers in controlled sensor degradation scenarios, including different types of degradation and various levels of severity. The results show that the performance of classifiers can deteriorate significantly, particularly in the presence of gradual drifts. To mitigate this effect, introducing sensor degradation during training appears to be a relevant strategy.
Finally, this thesis proposes a structured framework for multi-label anomaly detection in complex systems, covering classification methods, feature space optimization, interpretability, and robustness. For each of these aspects, the thesis adopts an industrial perspective and takes into account the differences between controlled laboratory experimental conditions and the additional constraints of real-world complex environments.
