Machinery fault diagnosis based on a transfer learning model with local sparse structure of PMDCNN-LSTM

Abstract

Aiming at the fault diagnosis problem in practical applications where the performance of the model degrades when applied to new domains or tasks due to limited and approximate training data, a transfer learning (TL) model with a local sparse structure based on a parallel multichannel convolutional neural network (CNN) and a long-short-term memory network (LSTM) (PMDCNN-LSTM) is proposed. The model is combined with CNN, LSTM, and TL to achieve parameter sharing by training learning parameters in the source domain and transferring them to the target domain. At the same time, the model is transformed from training data to real-world applications by integrating samples from the target domain, which improves generalization while maintaining the ability to identify faults in the training set. The advantages of CNN and LSTM are combined in this model compared with the traditional TL model, which can efficiently process data containing spatio-temporal features while maintaining sensitivity to time-series variations, exhibiting higher performance and advantages in complex data processing. It is shown by the experimental results that excellent performance and generalization ability can still be maintained by the model in high-noise environments, which is considered to be of great significance for research on bearing and gear fault diagnosis.