Safety risk point identification and localization in power equipment based on electromagnetic signals and a multiscale attention convolutional neural network

INTRODUCTION: Safety risk points in power equipment often produce weak and non-stationary electromagnetic responses. These responses are difficult to identify when partial discharge, contact anomalies, insulation deterioration, and shielding defects occur under complex operating conditions. OBJECTIVES: To improve risk recognition and spatial localization, a multichannel electromagnetic signal identification method is proposed. METHODS: Sliding windows are used to construct signal samples. Outlier correction, normalization, and wavelet threshold denoising are then performed to suppress background interference while preserving transient pulse structures. Time-domain statistics, frequency-domain energy characteristics, and short-time Fourier transform (STFT) spectra are extracted to describe waveform fluctuation, band energy migration, and time-frequency coupling. A multiscale convolutional neural network (CNN) with an attention mechanism is designed to enhance key frequency bands and sensitive measurement channels. Spatial correlation weights are further introduced to estimate the location of risk points. RESULTS: Experimental results show that the proposed method achieves a recognition accuracy of 95.28%, an F1-score of 94.58%, and an area under the receiver operating characteristic curve (AUC) of 0.981. The mean localization error is 8.46 cm. CONCLUSION: Compared with conventional methods, the proposed method improves the discrimination of weak electromagnetic disturbances and provides a practical solution for safety risk identification in power equipment.