Prediction of abnormal situations by machine learning in a predictivemaintenance context : Optimal transport theory for anomaly detection
Amina Alaoui Belghiti · HAL (Le Centre pour la Communication Scientifique Directe) · 2021
Emerging technological advances in the Internet of Things (IoT) have led to significant interference from manufacturing strategies. To this end, concepts such as "Industry 4.0", "smart manufacturing" and "digital factory" have emerged. In these contexts, predictive maintenance plays an increasingly crucial role in reducing costs and improving commercial performance because it uses heterogeneous data sources to detect abnormal behavior of equipment (diagnostics), predict modes of future failure (prognosis) and support upstream decisions.In this thesis, we present an overview of predictive maintenance architectures and we are interested in a capital pillar of these architectures, the anomaly detection as a first step of decision-making in a predictive maintenance architecture. We provide two contributions to this research question.A first method of semisupervised classification in optimal transport in two versions (parametric and nonparametric) for the detection of anomalies in time series. The experimental results of this method’s application on synthetic and real acoustic data sets prove the robustness of the metrics derived from optimal transport and further demonstrate the superiority of the performance of the method over state-of-the-art algorithms.The second contribution concerns an unsupervised anomaly detection method in multidimensional data. It identifies local outliers in a non-Euclidean topological space using optimal transport metrics.The experimental results revealed the effectiveness of the method in solving the dimensionality problem and testify to the statistically significant difference of the proposed method compared to the methods of the state-of-the art evaluated.