Fault Detection and Diagnosis of DC Motor Using Long Short-Term Memory (LSTM)

Abstract

The reliability and continuous operation of DC motors are essential in industrial and engineering applications. However, unexpected faults such as brush wear and commutator faults can lead to efficiency reduction and unplanned downtime. This paper presents a machine learning approach for fault detection and diagnosis (FDD) of DC motors using Long Short-Term Memory (LSTM). A publicly available dataset was employed to train the models. To compare the performance, two other models, K-nearest neighbor (KNN) and artificial neural network (ANN), were used. The performance of the models was assessed through accuracy, precision, recall, and F1-score, and presented using confusion matrices. The results obtained from the developed scheme were compared with those obtained when KNN and ANN were used. The results revealed that KNN, while simple and computationally efficient with an accuracy of 93.3%, was prone to misclassifications in overlapping feature spaces. ANN demonstrated improved accuracy of 94.7% by capturing non-linear relationships among features but lacked the ability to effectively exploit time-dependent characteristics of the data. In contrast, LSTM achieved the highest performance, with a validation accuracy of 97.3% and strong precision and recall across all the classes, owing to its ability to capture temporal dependencies in sequential motor data. The study concludes that LSTM significantly outperforms KNN and ANN, making it the most suitable model for DC motor fault detection and diagnosis.