Lightweight transformer-based generative adversarial network for acoustic anomaly detection in converter valves

Anomaly detection of converter valves via acoustic analysis has been a hot topic in the high-voltage direct current research field. However, several factors have significantly hindered the practical deployment of such acoustic anomaly detection methods, including the infrequency of valve anomalies leading to imbalanced distributions of acoustic samples, limitations in computational resources, and inherent class imbalance in acoustic signals. To address these challenges, this paper proposes a novel unsupervised anomaly detection framework, named lightweight transformer-based generative adversarial networks (LT-GAN). It introduces two lightweight modules, MobileNet V2 and D-MobileNet V2, to perform downsampling and upsampling of Mel-spectrograms derived from acoustic signals. Furthermore, it incorporates a K-ViT block to enhance global representation learning of spectral images and reduce network parameters. Experiments on real-world acoustic datasets show the superiority of the proposed LT-GAN, which achieves an AUC of 0.9806 on the ROC curve, significantly outperforming baseline methods. In the low false-positive regions [0, 0.1], [0, 0.2], and [0, 0.3], the p-AUCs reach 0.9295, 0.9122, and 0.9559, respectively. Moreover, LT-GAN exhibits exceptional lightweight characteristics, with model complexity metrics of 9.626 million parameters, 0.506 GFLOPs, and a model size of 37.48 MB. These results validate the effectiveness of the proposed approach in terms of anomaly detection performance and resource efficiency.