Novel preprocessing of multimodal condition monitoring data for classifying induction motor faults using deep learning methods

Induction motors are widely used in manufacturing industries failures in them could be fatal and costly. Hence their health condition must be adequately monitored because defects grow over time, and the earlier the faults are detected, the less their severity and risks. This study contributes to the preprocessing of multimodal condition data, which will be used as input in the induction motor fault classification process utilizing CNN’s Deep Learning (DL) capabilities. The proposed method presents a holistic and reliable multimodal fault classification approach that combines mixed inputs to identify induction motor bearing faults through the use of mixed channel inputs in particular vibration signals and thermal images. The preprocessing includes signal-to-image encoding of vibration signals, namely, Continuous Wavelet Transform (CWT) and Gramian Angular Field (GAF), which are used for image fusion to be presented as inputs to two Convolutional Neural Network (CNN) architectures known as ResNet and SqueezeNet trained through transfer learning algorithm. The results showed that the proposed approach of fused image performed well when there was a reduction in thermal image quality, which makes it more reliable and accurate compared with using thermal image on its own as a single channel input. Further, the results indicated that the vibration images are slightly better classified in terms of accuracy using GAF compared to CWT by almost 1.72% using ResNet-18 and 1.29% using SqueezeNet. The proposed methodology enhanced the overall fault classification accuracy by 14.81% using ResNet-18 when compared to using thermal images only. It was also noted that in the best-case scenario, the inner fault type classification was improved using the proposed method by 31% for ResNet-18 and 14% for SqueezeNet