Evaluating Machine Learning Efficiency: Simpler Models Outperform Deep Learning in Motor Fault Detection
Authors:-Mrs. L. Yamuna., G. Abhisekhar., K. Prasanna Lahari., K. Vivek., S. Hemanth.
Abstract- In motor condition monitoring, deep learning techniques have been explored by utilizing two-dimensional plots as datasets instead of traditional time-series signals. For instance, Convolutional Neural Networks (CNNs) have been trained using recurrence and frequency-occurrence plots. While previous studies have shown promising results with CNNs, the indistinct differences in these plots often make the model’s decision-making process appear as a black box. This study evaluates and compares ten traditional machine learning (ML) techniques with recent deep learning (DL) approaches for motor fault diagnosis using the same dataset. The dataset consists of 3,750 synthetically generated motor current signal samples, categorized into five classes—one representing healthy conditions and four representing faulty motor conditions—each tested under five loading levels (0%, 25%, 50%, 75%, and 100%). Following similar training and testing phases, the Light Gradient Boosting Machine (LightGBM) achieved the highest classification accuracy of 93.20%, outperforming three CNN-based models by at least 10.4%, whose accuracy ranged between 74.80% and 82.80%. LightGBM also demonstrated superior performance in other key evaluation metrics, including F1 score, precision, and recall. Notably, five out of ten traditional ML models surpassed the CNN-based models. These findings emphasize the importance of carefully selecting deep learning architectures, as they are computationally expensive and memory-intensive, yet do not always guarantee improved performance over traditional ML models, especially for relatively straightforward tasks like motor fault classification using current signals.
