An Efficient Anomalous Sound Detection by Robust Processing and Reformation of Objective

Boosted by the Internet of Things (IoT) and neural network (NN), an accurate anomalous sound detection (ASD) system is critical in Industry 4.0 for enabling proactive maintenance. Previous works have enhanced ASD’s accuracy by incorporating multi-domain features through advanced NN techniques. However, this often results in significant computational and memory overhead, which contradicts the IoT scenario. Therefore, we propose a lightweight ASD system by inspecting the core issues. First, to mitigate the noise effect of IoT, we propose dynamic signal processing to stabilize the feature extractor. Second, to adapt to the varying IoT environment, we apply structural learning to enhance the NN’s generalization. Additionally, to overcome ASD’s data scarcity, we improve the training strategy to strengthen NN’s interpretation capability. Finally, compared to the state-of-the-art method on the DCASE 2020 dataset, our approach shows 1.34% accuracy improvement, 16.9% computation reduction, and 43.5% storage reduction.