Multi-scale deep learning framework for robust bearing fault diagnosis in rotating machinery under sensor anomalies

Accurate fault detection in rotating machinery is essential for ensuring industrial safety. However, the reliability of such detection systems is often compromised by faulty sensors or inaccurate sensor data. To address this challenge, this paper proposes a fault-tolerant hybrid approach that integrates a 2D convolutional autoencoder (2D-CAE) and a 2D nested multiscale convolutional neural network (2D-NMSCNN). The 2D-CAE is trained on healthy sensor data to identify sensor anomalies, enabling the model to isolate and disregard faulty sensors. Simultaneously, the 2D-NMSCNN is designed to diagnose mechanical faults using only the data from functional sensors. This dual-module system is inherently resilient to sensor failure, maintaining high diagnostic accuracy even when up to half of the sensors are impaired. Evaluations on datasets with sensor displacements exposed to noise and bias demonstrate that the proposed method achieves robust, high-accuracy fault detection while effectively mitigating the effects of sensor noise and failure in industrial applications.