A RES-GANomaly Method for Machine Sound Anomaly Detection

Abstract

Automatic detection of mechanical failure is an essential technology in the fourth industrial revolution, which involves artificial-intelligence-based factory automation. The primary challenge in encoder-based machine sound anomaly detection lies in ensuring high-quality reconstruction of feature maps, as this directly impacts the precise definition of reconstruction error thresholds for normal and abnormal sound feature maps. This study proposes an improved deep convolutional generative adversarial network combined with the GANomaly method to introduce a unique anomaly detection model. This model leverages a residual deep convolutional generative adversarial network with an integrated attention mechanism as the generator and a multi-scale, multi-layer convolutional neural network as the discriminator to address the issue of information loss in reconstruction feature maps with increasing network depth, enhancing model generalization capabilities. The proposed approach introduces custom hyperparameters and tailored loss functions, utilizing Wasserstein distance to measure sample differences and promote model convergence. Researching the DCASE Challenge 2023 Task 2 development dataset, improvements are observed in experimental metrics such as AUC and pAUC, demonstrating the superiority of the model. We also analyze aspects such as feature map quality, parameter settings, and experimental ablation, and compare with other state-of-the-art methods to showcase the contributions of our model.

Keywords

• Anomaly detection
• attention mechanism
• generative adversarial networks
• residual connectivity
• unsupervised