Comparative Analysis of LSTM, ANN, and KNN Architectures for Fault Detection and Diagnosis in Permanent Magnet DC Motors

The reliability of Direct Current (DC) motors is critical to industrial productivity, yet mechanical components such as commutators and brushes are highly susceptible to wear and failure. This paper presents a rigorous comparative analysis of machine learning-based Fault Detection and Diagnosis (FDD) frameworks for Permanent Magnet DC (PMDC) motors. Specifically, we evaluate and compare the diagnostic performance of K-Nearest Neighbors (KNNs), Artificial Neural Networks (ANNs), and Long Short-Term Memory (LSTM) networks in classifying three operational states: Normal, Brush-Wear, and Commutator-Fault. Utilizing an open-source industrial dataset, each model was optimized to distinguish between subtle fault signatures that often lead to unplanned downtime. Our experimental results demonstrate a clear performance hierarchy: the baseline KNN model achieved an accuracy of 93.3% but was vulnerable to overlapping feature spaces, whereas the ANN improved accuracy to 94.7% by capturing non-linear relationships. The proposed LSTM architecture significantly outperformed both models, achieving a superior validation accuracy of 97.4% and near-perfect precision for normal and brush-wear conditions. This superior performance is attributed to the LSTM’s specialized gating mechanisms, which effectively capture long-term temporal dependencies within motor current and vibration signals. The study concludes that temporal deep learning models offer the most robust solution for automated predictive maintenance in complex industrial environments.