基于域泛化的无监督机器设备异常声检测

Anomalous sound detection of unsupervised machines based on domain generalization

在基于声信号的机器设备异常检测中,机器运行时发出的声音可能会因机器运转状态的改变而发生变化,环境噪声也会改变场景中的声学特征,而传统的机器异常声音检测系统会因为声学特征的改变导致正常的声音被错误地标记为异常。针对这一问题,提出了基于域泛化技术的无监督机器异常声音检测方法。首先,将声信号的对数梅尔谱图输入深度学习神经网络模型,结合域混合方法和坐标注意力机制模块,提高系统域泛化能力和异常识别性能;然后,在DCASE开源数据集上进行实验,与两种基线评估系统的AUC和pAUC得分进行对比。结果表明,所提出的方法在域泛化条件下的异常检测性能得到明显提升。

In the abnormal condition detection of machine and equipment based on acoustic signals,the sound emitted during machine operation may change due to the change of machine operation status,and the environmental noise may also change the acoustic features in the scene.The traditional machine anomaly sound detection system may cause the normal sound to be incorrectly marked as anomaly due to the change of acoustic features.To address this problem,an anomalous sound detection method of unsupervised machinery based on domain generalization technology was proposed.The Log-Mel spectrogram of acoustic signal was input into a deep learning neural network model,and the domain mixing method and coordinate attention mechanism module were combined to improve the system domain generalization ability and abnormal recognition performance.Experiments were conducted on the DCASE open source dataset,and compared with the A UC and pA UC scores of two baseline evaluation systems.The results show that the anomaly detection performance of the proposed method under domain generalization conditions is significantly improved.