Anomalous sound detection for machine condition monitoring using 3D tensor representation of sound and 3D deep convolutional neural network

This study introduces a novel approach that utilizes a three-dimensional tensor representation of machine-generated audio signals, serving as a suitable input for a three-dimensional convolutional neural network. The proposed method involves calculating the reconstructed phase space of the audio signal, followed by converting the resulting three-dimensional reconstructed phase space into a three-dimensional tensor format. This technique offers superiority by capturing nonlinear dynamic features and uncovering hidden system variables, which can improve discrimination and classification, enabling accurate detection of anomalous sound patterns, with valuable information encoded in the shape of the data cloud within the tensors. Subsequently, these tensors are employed as input to a three-dimensional deep convolutional neural network, facilitating effective analysis and classification of the audio signals. To assess the effectiveness of the proposed method, we conduct a comprehensive evaluation on three benchmark datasets: MFPT, MIMII, and ToyADAMOS, employing a 5-fold cross-validation scheme. The evaluation metrics employed include Sensitivity, Specificity, Accuracy, and F1 Score to ensure a thorough examination of the method's performance across diverse datasets, encompassing different machine types and acoustic environments. The experimental results showed a high average accuracy of 97.63% on the MFPT dataset. However, in the MIMII dataset, the slider machinery achieved the highest average accuracy rate of 92.02%, while the pump machinery had the lowest average accuracy rate of 90.54%. For the ToyADAMOS dataset, an average accuracy rate of approximately 94% was obtained. These findings underscore the method's potential for accurately detecting anomalies across various machine types and acoustic environments.