A novel method for anomaly detection using graph signal embedding: Application to anomalous sound detection

Anomalous Sound Detection (ASD) is an emerging field that combines audio signal processing and anomaly detection to provide accurate detection of anomalous sound events. Conventional ASD encounters two main problems: a) the inherently low amount of anomalous samples, eg. traffic accidents compared to any other event on the road, which creates an unbalanced dataset, and b) the limited ability of standard audio features, such as MFCC, to characterise anomalies, especially in highly noisy environments, such as roads, factories, cocktail party scenes, etc. In this context, we propose a Graph Signal Processing (GSP) approach based on graph embedding and filtering before undertaking anomaly detection methods, such as One-Class SVM, Isolation Forests and Autoencoders. First, acoustic features such as MFCC or log-Energy are extracted and embedded into a graph. Second, the constructed graphs are -optionally- filtered using a graph filterbank. Third, the magnitude of the joint Fourier transform is computed for each filtered graph. Finally, the features thus extracted are used to train one of the standard semi-supervised anomaly detection method using only normal data. To evaluate the advantages of the proposed approach, we tested using different standard anomaly detection methods, such as OC-SVM, isolation forests and autoencoders. Besides, experiments were conducted on different kinds of sounds and anomalies. Each anomaly detection method is carried out with and without graph embedding. In case of graph embedding, different types of graph topologies are tested, such as Path, Ring, Fully-Connected, Sensor-Network, … , etc. Experiments demonstrate the effectiveness of the proposed Graph Embedding and Filtering technique in improving the results of anomaly detection for all the methods.