Research On Unsupervised Anomalous Sound Detection

Anomaly sound detection is the task of determining the presence of anomalies by detecting the sound emitted by the target machine.With artificial intelligence-based factory automation,it is particularly important to detect anomalies in factory machines quickly and accurately.However,there are many difficulties in anomaly detection:different parts damaged may exhibit different anomaly patterns,and the probability of these anomalies occurring is so small that it is impossible to collect all the anomaly patterns.Therefore,training can only be performed with normal data,which is unsupervised anomaly detection.The fundamental problem of anomaly detection is to find out the distinguishing information between normal and anomaly and to find the appropriate distinguishing space.The main work of this paper is as follows:(1)Study of distinguishability on auxiliary supervised informationTo address the unsupervised problem of anomaly detection,this study adopts an auxiliary supervision approach.It explores how to introduce more auxiliary information,analyzes and studies the effectiveness of this auxiliary information in distinguishing anomalous data,and finally employs this auxiliary information for anomaly detection.The information introduced in this paper can be classified into two types:self-supervised information and category labeled supervised information.Self-supervised information uses the normal data itself as supervised information,and three methods are tried in this paper to extract different self-supervised information:reconstruction/prediction,generation,and comparison.Category-labeled supervised information,on the other hand,distinguishes between normal and anomaly by classifying and labeling categories of audio with similar features,so as to extract normal representations of each category by means of supervised learning.Anomaly detection is achieved based on the above-assisted supervised distinguishability study.Among the auxiliary supervised approaches,acoustic features play a decisive role in extracting the distinguishing information.In order to be better adapted to the auxiliary supervision information,several different acoustic features are listed in this paper,and the discriminative study of these acoustic features is carried out for anomaly detection combined with category labeling information.(2)Anomalous sound detection based on information fusionAnomaly sound detection has the problems of anomaly diversity and normal randomness,the way of information fusion can better solve the above problems by reflecting the distinguishability in a complementary and many-to-many way.The fusion methods in this paper are divided into two types:acoustic feature fusion and auxiliary supervised information fusion.Among the acoustic feature fusion,this paper tries to form diversity features by combining different acoustic features and pre-trained embeddings,aiming to solve the anomaly diversity problem by a suitable feature space.And the auxiliary supervised information fusion is based on multi-task learning or stage training to fuse the self-supervised information and category labeled supervised information.Finally,based on the integrated learning approach,the previously obtained multiple discriminative spaces are fused for anomaly score judgments.In summary,this paper analyzes the challenges faced by this task by conducting an in-depth study of the anomalous sound detection task under unsupervised conditions,firstly analyzing the distinguishability for the auxiliary supervised information and the input acoustic features,and finally fusing and judging the different information,and finally proposing that the performance of the improved system can achieve better results on some machines.