A self-supervised acoustic anomaly detection model with a multi-branch attention mechanism

Abstract Acoustic Anomaly Detection plays a crucial role in industrial scenarios, especially in maintaining equipment stability and early warning capabilities. However, due to the scarcity and complexity of anomalous patterns, current mainstream unsupervised methods—particularly those based on autoencoders—still face significant limitations in feature extraction and modeling capacity. To address these challenges, this paper proposes an enhanced self-supervised anomaly detection framework featuring a Hierarchically Enhanced Feature Extraction Module (HEFEM), which integrates multi-path convolution and channel attention mechanisms to better capture localized anomalous patterns in industrial audio signals. Additionally, we incorporate a standard Transformer module between the encoder and decoder, replacing the original LSTM architecture to model long-range dependencies and contextual interactions from a global perspective. Experimental results demonstrate that the proposed method exhibits superior robustness and detection accuracy compared to existing approaches, validating its effectiveness and practical potential in real-world industrial applications.