Cochlear transform and self-organized DarkNet based automated motor fault classification using sound signals

Abstract Motor fault detection and classification are critical tasks in industrial applications, where sound signals are commonly employed for fault diagnosis. This study aims to classify motor faults using a cochlear transform and a self-organized DarkNet-based model to achieve high detection performance. A motor fault sound dataset comprising 3727 sound signals across five categories was collected. A novel approach based on cochlear transform and a self-organized, pretrained convolutional neural network is proposed. In this approach: (i) each sound signal is converted into an image using the cochlear transform; (ii) three feature vectors are extracted using pretrained DarkNet19 and DarkNet53 architectures; (iii) the top 500 features from each vector are selected using the Chi-square (Chi2) selector; and (iv) the selected 500-dimensional feature vectors are classified using a support vector machine (SVM) with 10-fold cross-validation. This pipeline represents a self-organized deep feature engineering model for motor fault classification. The primary goal of the proposed model is to maximize classification accuracy. The model achieved accuracies of 99.87%, 99.92%, and 99.70% using the three generated feature vectors, with the highest classification accuracy of 99.92% being selected. The achieved classification accuracy of 99.92% demonstrates the effectiveness and reliability of the proposed method for motor fault classification.