The problem of sample imbalance will lead to poor generalization ability of the deep learning model algorithm, and the phenomenon of overfitting during network training, which limits the accuracy of intelligent fault diagnosis of switchgear equipment. In view of this, this paper proposes a data augmentation method for switchgear defect samples based on Wasserstein generative adversarial network with the partial discharge live detection data of the substation and the real-time switchgear partial discharge simulation experimental data. This method can improve the imbalanced distribution of data, and solve the problems such as the disappearance of gradients and model collapses in the classic generative adversarial network model, and greatly improve the stability of training. Verification through examples and comparison with traditional data augmentation methods. The results show that the data augmentation method mentioned in this paper can more effectively reduce the data imbalance, improve the performance of data-driven technology, and provide data support for subsequent fault diagnosis of switchgear equipment.
The traditional switchgear partial discharge pattern recognition method lacks a certain generalization performance and the recognition accuracy is low, which is difficult to apply in practical engineering. In view of this, this paper proposes a switchgear partial discharge pattern recognition method based on residual convolutional neural network. By adding a residual module to the network to solve the problem of deterioration after the saturation of the network leads to the saturation of accuracy, and comprehensively utilize the shallow and deep features of switchgear partial discharge data fusion learning to achieve pattern recognition. In this paper, through partial discharge simulation experiments of different insulation defect types of switchgear and on-site detection of distribution stations, a sample database of partial discharge data of switchgear is constructed and analyzed. The experimental results show that the recognition accuracy rate of the proposed method is 96.06%, which is at least 20.22% higher than the traditional recognition methods, and with the increase of the number of samples in the training set, the recognition rate is greatly improved. The comprehensive use of the feature layer fusion module and the residual module significantly improves the generalization performance of the model, and it is more suitable for practical engineering.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.