A Transfer Learning LSTM Network-Based Severity Evaluation for Intermittent Faults of an Electrical Connector

Traditional machine learning methods work well under a general assumption that the training data and the testing data should be drawn from the same distribution. However, this assumption does not hold in many real-world applications. For two structure-similar but specification-different connectors with intermittent fault, one connector has sufficient intermittent fault data with severity labeled but another connector has no labeled intermittent fault data. In this situation, the severity evaluation model trained by the samples from the former connector will have large evaluation errors for the samples from the latter connector. For realizing the transfer of the intermittent fault severity evaluation model between two similar but different connectors, this article proposes a transfer learning long short term memory (LSTM) network-based severity evaluation method for intermittent faults of an electrical connector. Conventional transfer learning methods with maximum mean discrepancy (MMD) have two shortcomings: 1) MMD only considers the mean discrepancy between the source and target domains and 2) MMD only measures the marginal distribution discrepancy of the source and target domains but ignored the conditional distribution discrepancy of different labels. Therefore, this article proposes a modified MMD metric which not only considers the higher order statistics but also the label discrepancy information to enhance the severity evaluation accuracy. Finally, this transfer learning model is applied in the intermittent fault experiment of three electrical connectors, and the effectiveness of the proposed method is verified using comparisons with four state-of-art transfer learning methods.