Adaptive fusion based on physics-constrained dictionary learning for fault diagnosis of rotating machinery

Abstract

Modern manufacturing systems rely on multiple types of sensors to monitor the manufacturing processes and machine faults to ensure product quality. The costs associated with sensor installation, maintenance, data transmission, and data storage are high. In this paper, a new data fusion method based on physics-constrained dictionary learning is proposed to improve the efficiency of data collection and the accuracy of fault diagnosis. In the proposed dictionary learning method, the measurement, basis, and classification matrices are optimized for the effective use of compressed sensing technique in process monitoring. With the optimized matrices, full-scale high-resolution signals can be reconstructed from a small number of sensor measurements. At the same time, machine states can be identified based on the sparse measurements. An adaptive weight scheme is introduced to combine low-cost sensor data so that the accuracy of fault diagnosis can be improved. The proposed method is tested with an experimental dataset of gearbox vibrations caused by gear cracks and different levels of crack severities are classified. The results show that up to 70% of data collection can be saved with the new approach while more than 95% of diagnosis accuracy is achieved. The sensitivities of the performance with respect to the number of measurements and number of sensors are also studied.