基于字典学习的无监督机器异常声检测

The detection of anomalous sounds for machines is a task to identify whether the sound produced by the target machine is normal or abnormal, which is very important in mechanical production. However, in real factories, actual abnormal sounds rarely occur and are highly diversified, so it is difficult to collect detailed abnormal sounds. To solve the problem of less abnormal data in the training set, this paper proposes an unsupervised detection system of anomalous sounds for machines based on a dictionary learning algorithm. The One-Class Support Vector Machine (OCSVM) commonly used in the anomaly detection algorithm is employed to find outliers, which can effectively detect unknown abnormal sounds under the condition that only normal sound samples are used as training data. In terms of audio feature selection, 16 classic traditional features (such as variance and kurticity, etc.) in the mechanical field are selected, and the audio is divided into frames to obtain more audio feature information. Compared with the baseline system in DCASE 2020 Challenge Task 2, which uses log Mel spectrum as the feature and autoencoder as the training classifier, the recognition effect of our system on some machines has been significantly improved.