The thesis topic focuses on predictive maintenance based on data collected during operation, typically of a vibratory nature, representative of the health status of the monitored system, and analyzed by experts looking for symptoms of failure (“Health and Usage Monitoring System” in aeronautics). The traditional approach is based on the use of advanced signal processing methods, with the disadvantages of requiring a high degree of user expertise and being extremely time-consuming. Another approach, based on machine learning, has recently demonstrated its potential to automate expert tasks with significantly reduced response times, thus offering the possibility of large-scale data processing while guaranteeing performance identical to, and sometimes even superior to, that of the expert. Despite numerous publications demonstrating this potential, this approach has not yet been implemented in critical industrial sectors such as aeronautics or wind power due to the difficulty of certifying 

 The AI models used in predictive maintenance and studied in this thesis are deep convolutional neural networks, known as “end-to-end” networks, which take the measured vibration signals directly as input and produce a binary classification (‘healthy’ versus “defective”) or multiple classifications (associated with different types of defects) as output. These models have demonstrated excellent performance, which remains difficult to explain. They function as “black boxes,” which implies a lack of transparency in their decisions, posing a problem in the event of a diagnostic error.

Barriers: The research project aims to answer some fundamental questions that represent the current barriers in the field:

How can we guarantee a level of confidence in diagnoses based on the results of an AI model?
• How can we take into account the knowledge and experience of a human expert in the learning process of an AI model?
These general questions can be broken down into more specific ones.
• What patterns are discovered in vibration signals by a deep convolutional neural network? The answer will make it possible to verify that these patterns are relevant from the human expert's point of view, either because they corroborate the state of knowledge or because they provide new knowledge.
• How can the size of the data be reduced to a small number of latent variables with high informational content that are potentially interpretable on the basis of physical principles? Human experts will expect to find an image of the vibrational excitations here.
•    How can digital learning and symbolic learning be combined? The extraction of logical rules will facilitate the interpretation of decisions made by an AI model and their validation by human experts.