Severity Classification of Electrical Unbalance in Variable-Speed Induction Motors Using Low-Dimensional MCSA Features: A Data-Driven Framework for Predictive Maintenance in Manufacturing

Electrical unbalance is a critical supply-related disturbance in three-phase induction motors, as it alters the electromagnetic behavior of the machine, degrades operating stability, and may adversely affect maintenance scheduling and production continuity in industrial environments. In this context, the present study proposes a compact and data-driven framework for the multi-class severity classification of electrical unbalance in variable-speed induction motors using low-dimensional Motor Current Signature Analysis (MCSA) features. A dedicated experimental dataset was established from one healthy operating condition and five unbalance severity levels (U60%, U70%, U80%, U90%, and U110%) acquired at multiple rotational speeds. Three-phase currents and voltages, neutral voltage, and rotational speed were synchronously recorded at 10 kHz over 5 s, and each acquisition was segmented into eight non-overlapping windows to generate supervised learning samples. Six physically interpretable indicators were extracted from each window in order to characterize negative-sequence effects, spectral distortion, and neutral-voltage behavior induced by electrical unbalance. The diagnostic capability of the resulting feature space was then assessed using three supervised classification models, namely a Deep Neural Network (DNN), k-Nearest Neighbors (kNN), and AdaBoost.M2. In addition to the baseline evaluation performed with the complete feature set, a systematic feature-number ablation study was carried out to quantify the effect of input dimensionality on classification performance. The results demonstrate that the proposed descriptors provide strong class separability across variable-speed operating conditions. In particular, kNN achieved perfect discrimination using only two features, whereas the DNN reached perfect performance with four features, thereby highlighting the partial redundancy of the extracted indicators. Overall, the findings confirm that accurate multi-class severity diagnosis of electrical unbalance can be achieved using a compact and physically meaningful MCSA representation, making the proposed approach attractive for real-time condition monitoring and predictive maintenance of induction-motor-driven industrial systems.