This thesis contributes to the application of Deep Learning (DL) in Remaining useful life (RUL) prediction of industrial equipment, addressing significant challenges in this field. Our research is driven by the need to develop DL architectures that mitigate performance degradation under various operating conditions, to improve model interpretability, and to address data scarcity by leveraging external (un)labeled data. We structured our work into two principal parts. In the first part, we explore architectures capable of handling data variability resulting from different operating conditions, without manual feature engineering. This led us to propose an MLP-LSTM-MLP architecture. By employing an MLP at the first stage, we were able to normalize this variability, thus improving performances under such settings. Furthermore, To enhance interpretability, we proposed to replaced the first-stage MLP stage with a Gated mixture of experts (GMoE) system, enabling interpretable decomposition based on operating conditions. The second part of the thesis addresses the issue of data scarcity, a widely recognized challenge in the Prognostics and health management (PHM) field. Through the introduction of adapters, i.e. task-specific layers that address the challenge of handling multiple input/output structures, we proposed an auxiliary training approach that leverages external labeled data, presenting a method that surpasses traditional techniques found in the literature. Moreover, to utilize external unlabeled data in auxiliary training, We proposed a meta-learning approach to automatically derive auxiliary objectives from these data by pseudo-labeling them in an end-task aware manner. The goal of this part was to leverage broader spectrum of available data to improve RUL prediction performances. In reflecting upon our work, we acknowledge the limitations of the proposed approaches and suggest both immediate and long-term directions for future research. These include tackling the challenges of processing long sequence data, further improving model interpretability, addressing data scarcity with more advanced training methodologies, and exploring the potential of federated learning and large language models in industrial settings.

• Emmanuel Ramasso – Associate professor HDR – Institut FEMTO-ST
• Céline Hudelot – Professor – Centralesupelec University of Paris-Saclay

• Bernardetta Addis – Professor – Université de Lorraine
• Birgit Vogel-Heuser – Professor – Technische Universität München
• Raphaël Couturier – Professor – Université de Franche-Comté

• Christophe Cerisara – Researcher HDR (CR) – Université de Lorraine, LORIA
• Alexandre Voisin – Associate professor HDR – Université de Lorraine, CRAN