Power Transformer Audio Anomaly Detection Based on the Hybrid Time Attention Spectrogram Fusion Network

This paper proposes an audio anomaly detection method for power transformers based on hybrid time attention spectrogram fusion network. The method extracts spectral features by MelAttnConvNet module which using one-dimensional convolution and ConvNeXt block, and enhances the local information representation by using depthwise separable convolution and channel attention mechanism; meanwhile, the periodic and temporal dynamic features of the signals are efficiently captured by TimesNet module. On this basis, a preliminary classification module is designed to model the global semantics by fusing traditional CNN (Convolutional Neural Networks, CNN) feature extraction with Transformer encoder and employing learnable [CLS] token to aggregate key information. Finally, angular boundary constraints on the features improve the discriminative power of the classifier. Experiments show that the method has high accuracy rate and strong generalization ability in audio anomaly detection, demonstrating the effectiveness of multi-channel feature fusion and joint time-frequency modeling in audio classification tasks.