Attention-Based Two-Stage Multi-Sensor Feature Fusion Method for Bearing Fault Diagnosis

Abstract

Effective bearing fault diagnosis can ensure the safe operation of rotating machinery, which is important for the stable operation of rotating machinery. Feature fusion of multi-sensor data is a feasible method to improve fault diagnosis performance. To accurately detect, localize and identify bearing faults, we propose an attention-based two-stage multi-sensor feature fusion (ATS-MSFF) method for bearing fault diagnosis. The first stage, focuses on feature extraction from sensor itself, which utilizes the Channel-Attention to enhance key features in the sensor's signal. The second stage then focuses on feature fusion between sensors, such that the output of each sensor is endowed with additional critical information provided by other sensors. Experiments conducted on a publicly available PU dataset validate the effectiveness of our approach, achieving a high classification accuracy of 99.58% and maintaining stable performance in the presence of noise interference. In addition, the design of this framework takes into account the flexibility of practical applications and can adapt to different numbers of sensor configurations, providing a new solution for the accurate diagnosis of bearing faults.