Surveillance et diagnostic des défauts des machines industrielles par un réseau de capteurs sans fil.

Abstract

This thesis proposes BearNet, a deep transfer learning framework for bearing fault diagnosis in resource-constrained WSN environments. BearNet adapts the pre-trained YAMNet modeloriginally designed for audio recognitionto classify bearing defects from vibration signals. Raw acceleration data from low-power nodes are transformed into Mel-spectrograms, serving as compact and informative inputs. Only selected YAMNet layers are fine-tuned, while additional layers enhance sensitivity to bearing features, reducing computational and memory demands without sacrificing accuracy. The method is validated on the CWRU and MFPT datasets. Results show that BearNet improves accuracy by up to 3.1% over standard YAMNet approaches and surpasses other state-of-the-art models by 0.02--5.26%. It also scales from binary detection to multi-class and severity-level classification, remaining robust across operating conditions. By combining vibration sensing, spectrogram representation, and selective transfer learning, BearNet delivers a practical, energy-efficient solution for real-time bearing fault diagnosis in WSN-based monitoring systems. The findings demonstrate the feasibility of embedding high-performance AI into wireless infrastructures and open avenues for future research on adaptive, on-node fault detection for modern industrial environments