An Improved Prototype Learning Method for Domain Generalized Industrial Machine Anomaly Sound Detection

Authors:

Abstract
In recent years, deep learning methods have made significant progress in the field of industrial machine anomaly sound detection. However, the model’s generalization ability remains significantly constrained in practical applications, such as cross-machine and cross-operating conditions, due to domain shift issues arising from differences in acoustic feature distributions. To address this problem, this paper proposes a domain-generalization-based industrial machine anomaly sound detection method. The method is built upon a prototype learning framework and integrates both time-domain and frequency-domain features to construct multi-source acoustic representations. Furthermore, a convolutional block attention module (CBAM) is introduced during the feature extraction stage to enhance the model's ability to focus on key time-frequency regions. In addition, a feature enhancement strategy specific to the target domain is incorporated to effectively alleviate the distribution discrepancies between domains, thus improving the model's robustness in unknown domains. Experimental results on the DCASE2022 Task 2 development dataset show that the proposed method achieves a harmonic mean score of 80.35%, representing a 3.92% improvement over the state-of-the-art unsupervised method MSL. Further validation on the gearbox dataset collected from a PT700 rotor fault test bench demonstrates that the method maintains good detection stability and generalization ability in real industrial scenarios. These results indicate that the proposed framework effectively addresses the domain generalization challenges in machine anomaly sound detection tasks and holds promising engineering application potential.