Time Stretching and Pitch Shifting for Anomalous Sound Detection in Fan Sound Signal using MobileNetV2

Anomalous sound detection means the ability to identify between normal and anomalous sounds specifically within the context of sound produced by a fan machine. This capability is of vital importance to various industries as it enables the prediction of issues and alerts maintenance teams before significant damage occurs. This leads to cost reduction and efficient resource allocation, especially considering the wide range of applications involving fan machines. In this study, we provide an approach for the unsupervised learning-based anomalous sound detection (ASD) of fan sound using a Convolutional Neural Network (CNN). We used a MobileNetV2-based system established by the Detection and Classification of Acoustic Scenes and Events (DCASE) competition organizers as a starting point to implement this system. To make the system tailored more specifically toward fan sound signals, we have tried two distinct approaches. The first approach depended on applying transformations to the data using data augmentation after experimenting with several techniques we found that time stretch, along with pitch shift, yielded the most effective results. The second approach involved incorporating the Gaussian Mixture Model (GMM) into our proposed system to enhance the accuracy of the detection. Using an exhaustive dataset that included normal and anomalous fan sound recordings, our proposed approaches were evaluated.