Multitask learning enhanced fault diagnosis based on multimodal feature with adaptive weight fusion

Fault diagnosis has gained increasing interests to assure the security and stability of industrial systems. Simultaneously collecting multi-sensor signals is crucial for comprehensively characterizing the system's health status. Nonetheless, creating effective fusion mechanisms for multi-sensor signals to enhance diagnostic performance remains a challenge. Consequently, this study proposes an innovative fault diagnosis method enhanced by multi-task learning, incorporating multimodal feature learning and attention fusion across vibration and current signals. Initially, we introduce an attention fusion module designed to selectively extract crucial features from multimodal data obtained from vibration and current signals. Furthermore, a multitask learning module is developed to collectively optimize multiple classification tasks using fused features and two distinct unimodal features. The performance of our proposed method is evaluated on two datasets. Experimental results illustrate that our method can extract more significant features and achieve superior classification performance compared to unimodal methods which lacks attention fusion and multi-task learning modules.