Cross-Domain Fault Diagnosis of Powertrain System using Sparse Representation

Although excellent progress has been made recently in powertrain fault diagnosis based on vibration signals, most of them are based on the assumption that the fault features of the training and test data are drawn from the same probability distribution. Due to the limitation of the domain shift phenomenon, the performance of the current intelligent fault diagnosis methods is significantly reduced. Even many existing transfer learning methods have the problem of low generalization ability. Inspired by sparse representation theory, a novel cross-domain fault diagnosis method based on K-means singular value decomposition (K-SVD) and long short-term memory network (LSTM) is proposed in this study. First, K-SVD can convert source domain data into a sparse dictionary and sparse coefficient. The domain-invariant features are explored in the sparse dictionary, which contains redundant features. The sparse coefficients are input into the LSTM to obtain a primary classifier. Then, the sparse coefficients of the target domain are solved by using the sparse dictionary of the source domain. It is input into the primary classifier for fine-tuning training, and the final diagnostic classification model is obtained. The proposed method establishes knowledge transfer from the source domain to the target domain by exploring domain-invariant features in the sparse domain and bridging the distribution discrepancy. It is evaluated using powertrain operating data acquired on cross-speed, cross-load and cross-sensor working conditions. It is demonstrated that the proposed method has superior performance in dealing with data imbalance and different distributions. It offers a promising approach for industrial applications on cross-domain fault diagnosis.