Hybrid SVM–Random Forest Approach for Fault Diagnosis, Prognostics and Web-Based Predictive Maintenance Decision Support in Induction Motors

Abstract

Unplanned motor failure remains one of the costliest disruptions in continuous-process industries, and neither breakdown repair nor fixed-interval servicing makes use of the condition information a running motor already generates through its temperature, vibration and load signatures. This work builds a sensor-driven predictive-maintenance pipeline for three-phase induction motors that converts six raw channels — stator, rotor and winding temperature, bearing temperature, vibration and load — into three interpretable outputs: a Health Index, a Remaining Useful Life (RUL) estimate, and a Failure-Risk probability. A radial-basis-function Support Vector Machine assigns each operating snapshot to one of five condition classes, while two independently trained Random Forest regressors refine RUL and failure-risk estimates derived from a deterministic, weighted Health-Index formula. A four-condition rule engine converts these outputs into a specific maintenance recommendation, and the complete pipeline is exposed through a role-based Flask dashboard with live gauges, trend charts and downloadable reports. On a held-out 200-record test partition, the classifier reached 76.0% accuracy, performing well on the majority Critical-Failure and Healthy classes but failing entirely on the underrepresented Overload class — a class-imbalance effect examined in detail alongside a circularity caveat that affects the regression evaluation, since both regression targets are themselves formulas of the same input features. The paper closes by outlining the field-validation, class-balancing and temporal-modelling work needed before deployment.