A novel differentiable neural network architecture automatic search method for GIS partial discharge pattern recognition. (31st May 2022)

• Record Type: Journal Article
• Authors: Jing, Qianzhen; Yan, Jing; Wang, Yanxin; He, Ruixin; Lu, Lei
• Is Part Of: Measurement. Volume 195(2022)
• Publication Date: 2022-05-31
• Subjects: Differentiable neural architecture search -- Convolution neural network -- Partial discharge -- Pattern recognition -- Gas-insulated switchgear; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.111154
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A novel partial discharge (PD) pattern recognition method is proposed. The optimal network for PD pattern recognition can be automatically constructed. Mixed depthwise separable convolution is used to extract multiscale features of PD. Self-attention mechanism is used to further enhance the feature extraction ability. The performance of the proposed method has been verified by the PD dataset. Abstract: Convolutional neural network (CNN) has been extensively used in pattern recognition of partial discharge (PD) in gas-insulated switchgear (GIS) because of its powerful feature extraction ability. However, at this stage, manual trial and error is needed to construct the CNN. Moreover, the model is designed for specific datasets, which will cause domain bias when applied to a new dataset. Therefore, a novel differentiable neural architecture search method is proposed to automatically construct a GIS PD pattern recognition model. First, a factorized hierarchical search space is used to design the CNN architecture. Then, a discrete search space is relaxed into a continuous search space through a search strategy based on Gumbel–softmax. Experiments show that the recognition accuracy of the proposed method can reach 97.625%. Furthermore, the proposed method has strong robustness and high precision against noise and strong tolerance for unbalanced datasets.