An Improved Deep Transfer Learning Method for Rotating Machinery Fault Diagnosis Based on Time Frequency Diagram and Pretraining Model

Under varying working conditions, the failure prediction model for the same type of equipment often proves ineffective in its deployment and application. Aiming to address the limitations of current fault diagnosis models based on neural networks, such as limited depth, insufficient feature extraction ability, lack of adaptive ability, and poor classification effectiveness in various domains, a novel fault prediction algorithm called deep adaption residual neural network (DARN) was studied and proposed. The algorithm incorporates a pretraining model to enhance its fault diagnosis capabilities. In this method, the time frequency diagram of the original time-series signal is obtained through time frequency processing. At the same time, the residual neural network pretraining model is used as the primary network for feature extraction. In addition, several loss functions are designed to minimize the discrepancy between data categories and the loss of adaptive transfer. The ablation experiments for several hyperparameters were carried out. The proposed method not only improves the accuracy of the fault prediction model but also significantly reduces the training time. Compared with the traditional neural network fault diagnosis model, this method addresses the issues of structural instability and limited feature extraction ability. It ensures that the model maintains strong predictive capabilities across various working conditions. Finally, the method was validated on a public bearing dataset and a homemade bearing dataset.