Search

Aiming to solve the problems of multiple internal power components, high fault probability, high similarity of the fault features of different power components, difficulty of traditional fault diagnosis feature extraction and low accuracy of fault identification in high-voltage multilevel cascaded H-bridge inverters, this paper presents a fault diagnosis method based on an attention-parallel multisource data fusion residual network. First, a parallel residual neural network model is established, and the extracted multilevel three-phase voltage before filtering and the three-phase current waveform after filtering are converted into two-dimensional image data using a wavelet transform. Subsequently, a feature fusion module is integrated into the network structure to adaptively extract features at different network levels. This module locates key features using the attention mechanism. Then, it fuses useful fault information into feature images using the feature fusion mechanism, enhancing the feature representation capability of the network. Finally, the fault features extracted by the feature fusion module undergo the complete convolution operation. The final enhanced features are used as classification features and classified using a softmax layer. Experimental results demonstrate that the proposed method exhibits high fault diagnosis accuracy and adaptability. [ABSTRACT FROM AUTHOR]

Aiming at the problems of high similarity and difficulty in extracting the fault features of power-switching tubes, as well as the high complexity of fault diagnosis models, the large number of parameters, and the long fault diagnosis time of the multilevel cascaded H-bridge inverter in medium-voltage and high-voltage applications, this study proposes a fault diagnosis method based on a lightweight shuffle–SimAM network. First, the proposed method establishes a lightweight parallel ShuffleNet network model and utilizes multi-sensor data as the input of each parallel network for the initial extraction of similar fault features. Second, a feature fusion module is constructed inside the network to weight and fuse the features extracted at each level of the parallel network. Then the fused features are successively advanced to further enhance the extraction of similar fault features. Finally, to maintain the network with high diagnostic accuracy while improving the level of lightweighting, deep separable convolution and SimAM parameter-free attention mechanisms are introduced into the diagnostic network. Experimental results show that the proposed method effectively reduced the complexity of the model and the diagnosis time while maintaining a high diagnosis accuracy. [ABSTRACT FROM AUTHOR]