Multi-Channel Time-Frequency Domain Deep CNN Approach for Machinery Fault Recognition Using Multi-Sensor Time-Series

Abstract

In the industry, machinery failure causes catastrophic accidents and destructive damage to the machines. It causes the machinery to stop and reduces production, causing financial losses to the industry. As a result, identifying machine faults at an early stage is critical. With the rapid advancement in artificial intelligence-based methods, developing automated systems that can diagnose machinery faults is necessary and challenging. This paper proposes a multi-channel time-frequency domain deep convolutional neural network (CNN)-based approach for machinery fault diagnosis using multivariate time-series data from multisensors (tachometer, microphone, underhang bearing accelerometer, and overhand bearing accelerometer). The wavelet synchro-squeezed transform (WSST) based technique is used to evaluate the time-frequency images from the multivariate time-series data. The time-frequency images are fed into the multi-channel deep CNN model for automated fault detection. The proposed multi-channel deep CNN model is multi-headed, considering the time-frequency domain information of each channel time-series data for automated fault detection. The proposed model’s performance is compared to benchmark models regarding testing accuracy, total parameters, and model size. Experiments have shown that the proposed model outperforms benchmark models regarding classification accuracy. The proposed multi-channel CNN model has obtained the accuracy and F1-score values of 99.48% and 99% for fault classification using time-frequency images of multi-sensor data. Finally, the proposed model’s performance is measured regarding inference time when deployed on edge computing devices such as the Raspberry Pi and the Nvidia Jetson AGX Xavier.