Edge-Side Fault Diagnosis for Machinery Health Management Based on Lightweight Model and Uncertainty Analysis

Considering scenarios in mechanical equipment involving continuous high-frequency signal acquisition and challenges in data transmission, this paper proposes an edge-based fault diagnosis framework (EdgeDiag) based on high-confidence information extraction to enhance model training and improve the efficiency of diagnostic information processing, including data storage and uploading. First, to ensure diagnostic accuracy while reducing computational burden on edge devices, a lightweight model is designed to minimize the number of model parameters. Next, a high confidence–based training strategy and information processing framework considering risk levels are developed, utilizing uncertainty analysis to evaluate the diagnostic performance of each class in the data. Finally, the effectiveness of the proposed method is validated through Raspberry Pi and fault diagnosis of critical components in mechanical systems. This approach addresses the challenge of transmitting all data from edge devices, significantly improving the efficiency of fault diagnosis.