Reconstruction-Based Spectrogram Augmentation for Anomalous Sound Detection System

Unsupervised anomalous sound detection(ASD) without anomalous samples is a challenging task, it has received much attention in the fields of industrial automation and medical information. In order to obtain representations that are conducive to the detection of anomalies, many studies focus on using meta-information for self-supervised representation learning. In this study, we propose an approach that includes data augmentation and contrastive learning to train the detector. This method is applied to the Log-Mel spectrogram of the waveform, converting it into an enhanced counterpart that highlights key information while reducing noise. Subsequently, we obtain temporal features through a TDNN extractor and merge them with the enhanced spectrogram. These two features are then fed into a Convolutional Neural Network(CNN) model for the ASD task. Furthermore, aiming to achieve inter-class dispersion and intra-class compactness through additional contrastive learning, we construct positive and negative samples using this AE and the CutMix method, respectively. We conduct a series of experiments, including comparative experiments with previous ASD systems, and two ablation studies for the data augmentation and contrastive learning modules, respectively. These experiments demonstrate the effectiveness of our proposed method and its superior performance compared to previous methods.