A Machine Anomalous Sound Detection Method Based on Deep Residual Generative Adversarial Network

To address the domain shift issue arising from acoustic discrepancies between training and testing data in machine anomalous sound detection, this study proposes a novel method based on a Deep Residual Generative Adversarial Network (DRGAN). The proposed method employs an encoder-decoder architecture to process source and target domain samples separately, incorporating residual connections and a Squeeze-and-Excitation (SE) attention mechanism within the encoder to effectively mitigate information loss and gradient vanishing problems. Furthermore, adversarial training is jointly optimized with the Maximum Mean Discrepancy (MMD) criterion to align feature distributions across domains, thereby enhancing the model's capability to extract domain-invariant representations. Finally, anomaly detection is conducted based on the reconstruction error. Experimental results on the MIMII DUE machine sound dataset demonstrate that the proposed method significantly outperforms the baseline Autoencoder (AE), achieving average improvements of 5.24% in AUC and 0.85% in pAUC under domain shift conditions. Compared to Ganomaly and Transformer-based approaches, the proposed method further demonstrates superior detection performance, validating its effectiveness and robustness in cross-domain anomaly detection scenarios.