Circuit breaker sound source localization method based on cross-channel attention fusion

Identifying the active circuit breaker in densely packed distribution cabinets is challenging, as acoustic signatures from neighboring breakers often overlap. At the same time, microphone-array-based localization remains expensive to deploy. To address these limitations, we propose a stereo-enhanced cross-channel attention framework that localizes the source using only a fixed dual-microphone recorder. The input is represented as a six-channel time-frequency feature map that combines left and right log-Mel spectrograms, interaural level difference (ILD), interaural phase difference (IPD) encoded as cosine and sine components, and a GCC-PHAT peak map, enabling a clear separation between content and spatial cues. On top of this representation, we introduce Stereo-Enhanced Cross-Channel Attention (SECCA), in which the content branch provides Query features while the spatial branch provides Key and Value features for cross-channel guidance. The module injects directional information into content modeling through lightweight global pooling and gating. The resulting features are fed into a ConvNeXt-T backbone with hierarchical dual-head supervision for coarse region classification and fine breaker classification, with only the fine head retained during inference. File-level predictions are obtained by averaging clip-level posteriors within each recording. On a 10-kV switchgear dataset comprising 900 recordings across nine breaker positions, two operation types, and five load-current levels, ConvNeXt-T + SECCA achieves a file-level Macro-F1 of 84.79 %±2.26 % (+4.10 points) and a file-level localization accuracy of 84.90 %, with a peak of 88.00 %. Attention visualization and LR-swap consistency analysis further show that SECCA learns frequency-selective and spatially informative responses, rather than acting as a simple parameter increase. These results indicate that the proposed method provides a non-contact, low-cost solution for on-site cabinet inspection and online maintenance. The reported performance should be interpreted under a controlled fixed-recorder geometry.