Data-driven diagnosis of permanent magnet synchronous motor demagnetization faults using machine learning and statistical process control
Permanent magnet synchronous motors (PMSMs) are widely utilized in electric vehicle and industrial drive applications due to their exceptional efficiency and power density. Nevertheless, the irreversible demagnetization of permanent magnets poses a significant degradation mechanism that negatively impacts torque capability, efficiency, and long-term reliability. A physics-guided hybrid diagnostic architecture has been developed for assessing demagnetization in PMSM. This architecture integrates statistical process control (SPC), Isolation Forest for anomaly validation, and K-Nearest Neighbors (KNNs) for severity estimation, thereby forming a unified sequential monitoring framework. Instead of depending solely on an isolated algorithm, this proposed methodology establishes a structured, hierarchical diagnostic pipeline. This pipeline systematically combines initial statistical screening with subsequent unsupervised anomaly confirmation, ultimately leading to a quantitative degradation assessment, all designed for continuous, real-time monitoring of motor health. Integrated multi-sensor data, encompassing temperature, magnetic flux density, stator currents, and rotor speed, forms the basis for facilitating non-invasive, real-time health monitoring. The analytical process involves three distinct methods, which are executed in a strict sequential pipeline. Initially, SPC is applied to continuously monitor magnetic flux density using both Shewhart and Exponentially Weighted Moving Average control charts. This step identifies observations that exceed the three-sigma control limits, classifying them as statistically deviant. These flagged observations, along with all incoming multi-feature vectors, are subsequently transferred to the Isolation Forest algorithm. This algorithm assigns an anomaly score to each sample by measuring the mean path length necessary to isolate it within an ensemble of randomized trees. Samples with scores above a predefined contamination threshold are then designated as anomalies. Finally, only these labeled anomalous samples are directed to the KNN regression model. The KNN model retrieves the k most similar historical degradation records in the feature space, utilizing Euclidean distance, and then calculates the predicted demagnetization percentage as the average of their respective target values. Validation of the approach was performed on an experimental 1 kW PMSM setup exposed to varying thermal and electrical loading in order to simulate sensor-based degradation markers associated with the early stages of demagnetization-related degradation. The proposed framework was validated using experimentally acquired multi-sensor degradation indicators rather than direct measurements of irreversible permanent magnet remanence. The ability to consistently identify the degradation pattern and agree with the physics-guided demagnetization indicators computed using multi-sensor data is evident from this result. Severity estimation is based on experimentally obtained degradation indicators compared with other indicators, not on the actual magnet degradation levels themselves. The proposed framework provides a scalable, interpretable, and cost-effective solution for predictive maintenance and real-time fault diagnosis of PMSMs.