A Fault Diagnosis Method With Bitask-Based Time- and Frequency-Domain Feature Learning

Deep-learning-based methods used for fault diagnosis show remarkable performance, and these methods primarily learn features based on time or frequency domains. Generally, time and frequency domain features are important for identifying faults. They can reflect the type and severity of faults, especially, frequency domain features can reveal highly distinct patterns related to fault types. Thus, learning time and frequency domain features helps obtain comprehensive fault information and realize high accuracy for fault diagnosis. In addition, multitask learning can learn features from many related tasks simultaneously. Therefore, in this study, multitask learning is employed for learning time and frequency domain features from two tasks (learning features from the time domain and the frequency domain). A fault diagnosis method using a bitask-based time and frequency domain feature learning network (TF-FLN) is proposed. The TF-FLN learns features through pretraining, and then, the features are fed into a fault diagnosis network that is fine-tuned using labeled data. In particular, in the TF-FLN, task-specific loss functions are applied according to the characteristics of the two tasks. Moreover, the weight for each loss function is optimized automatically using a loss function optimizer. Compared with existing methods, the proposed method realizes a higher accuracy of 99.86% on the gearbox dataset and of 99.37% on the Paderborn University bearing dataset. Furthermore, experimental results demonstrate the effectiveness of multitask learning and the low computation cost of the proposed method.