Acoustic signals emitted by electrical power equipment, such as transformers, generators, and circuit breakers, encode critical operational states. These signals enable non-contact, real-time condition monitoring, which is essential for predictive maintenance and grid reliability. However, traditional diagnostic methods struggle to detect multi-scale fault patterns, lack efficient cross-feature integration, and demonstrate limited generalization with scarce labeled acoustic data from power systems. To overcome these limitations, this paper proposes DBAN-Net, a hybrid dual-branch network for acoustic anomaly detection in power equipment. The model synergistically combines a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (BiLSTM), enhanced with an attention-based fusion mechanism. Mel-frequency cepstral coefficients (MFCCs) and their delta features are extracted to represent acoustic characteristics. The CNN branch captures localized frequency patterns, while the BiLSTM branch models long-term temporal dependencies associated with evolving fault conditions. An attention mechanism dynamically fuses these two feature streams to enhance discriminative patterns and suppress ambient noise. Evaluated on the pump subset of the MIMII dataset, DBAN-Net achieves a high fault detection accuracy of 98.26% and consistently surpasses benchmark methods in AUC, F1-score, and Accuracy. The results validate the efficacy of the dual-branch attention architecture for reliable acoustic-based fault diagnosis in electrical power assets.
Chen et al. (Sat,) studied this question.