Conditional vector quantized variational autoencoders for machine sound anomaly detection under domain shift

In industrial production, monitoring abnormal machine sounds is crucial for early fault detection, reducing accidents, and improving efficiency. However, domain shift-caused by variations in operating states and environments-poses significant challenges to model generalization. To address this, this paper proposes an unsupervised anomaly detection method that integrates a Vector Quantized Conditional Variational Autoencoder (VQ-CVAE) with the autoregressive model PixelSNAIL. A lightweight audio-text contrastive pre-trained model extracts textual features describing equipment physical attributes (e.g., type, parameters, operating time) and employs them as conditional inputs to the VQ-CVAE, thereby constructing fused audio-text multimodal representations. Anomaly scoring is jointly performed in time-frequency and latent spaces. In the time-frequency space, the VQ-CVAE codebook mitigates excessive feature shifts, while an attention- and gate-enhanced residual module strengthens feature extraction and noise robustness. In the latent space, PixelSNAIL predicts the discrete code matrix produced by the VQ-VAE, and the prediction loss is used to further refine anomaly detection results. Experiments on the DCASE2024 Task 2 dataset demonstrate superior performance across multiple machine types, confirming the method's effectiveness and robustness for machine sound anomaly detection in complex industrial scenarios.