Simpler Machine Learning Methods Outperform Deep Learning in Motor Fault Detection

In motor condition monitoring, deep learning techniques have been considered by using 2-dimensional plots as datasets instead of time-series signals. For example, a Convolutional Neural Network (CNN) can be trained using recurrence or frequency-occurrence plots. While previous studies demonstrated promising results using CNNs, the lack of discernible differences in the plots rendered the model's inner workings seem like a black box. This study applies ten traditional machine learning (ML) techniques and compares them with recent deep learning (DL) techniques used in motor fault diagnosis using the same dataset. The synthetically prepared motor current signal dataset with 3,750 samples has five classes - healthy and four faulty motor conditions, under five loading conditions - 0, 25, 50, 75, and 100%. After similar training and testing phases, the light gradient-boosting machine (LightGBM) showed the best overall classification accuracy of 93.20 %, significantly outperforming by at least 10.4% the three CNN-based models, which obtained performances that range from 74.80 to 82.80%. LightGBM also has the best average performances in other metrics, such as F1 score, precision, and recall. Five out of ten ML models performed better than these three CNN-based models. Given the excellent performance of traditional ML models such as LightGBM, care and consideration has to be taken in the use of deep learning architectures especially since they are more computationally expensive and memory-intensive because there seems to be no guarantee that they will perform better than traditional models, especially on simpler problems, such as the motor fault classification using current signals that we have presented in this paper.