Fault diagnosis method of soft-switching inverter based on deep learning

Aiming at the complexity of open-circuit fault diagnosis of soft-switching inverters, a dual-branch fault diagnosis method based on one-dimensional convolutional neural network (1-D CNN) combined with improved deep residual shrinkage network (IDRSN) and two-dimensional convolutional neural network (2-D CNN) with attention mechanism (IDRSCNN-ACNN) is proposed. This method combines the advantages of 1-D CNN that can extract the primitive features from the original time series with the advantages of 2-D CNN that can extract high-dimensional features from images. It can mine more effective spatial features for inverter fault diagnosis. In addition, the method integrates an improved deep shrinkage residual network (IDRSN) and coordinate attention (CA) mechanism to improve performance. Firstly, the input current data on the power side of the soft-switching inverter is expanded by sliding window overlapping sampling, and the extended data is converted into time-dependent Markov images by using Markov transition field. Secondly, a parallel dual-branch IDRSCNN-ACNN fault diagnosis model is designed. Then, the original time series data and Markov image are used as the input of 1-D branch and 2-D branch respectively to train the model. Finally, the Softmax classifier is employed for precise fault classification. Experimental results show the method’s efficacy in classifying mixed-noise data across 79 fault types. IDRSCNN-ACNN has better fault diagnosis performance through ablation experiment and comparison with some traditional fault diagnosis models.