Machine Learning-Based Condition Monitoring of Electromechanical Systems Using Sensor Fusion

The reliability of electromechanical systems operating under dynamic industrial conditions is increasingly challenged by nonlinear behaviour, load variability, and complex fault interactions. Conventional condition monitoring approaches often rely on isolated sensor measurements and static threshold-based diagnostics, limiting their ability to detect early-stage and compound faults. This study proposes a unified machine learning-driven condition monitoring framework that integrates multi-modal sensor fusion with adaptive feature learning for robust health assessment of electromechanical systems. The proposed architecture acquires synchronized data streams from heterogeneous sensors—including vibration, stator current, temperature, torque, and acoustic signals—and performs dynamic feature extraction using time–frequency domain transformations and statistical descriptors. A hierarchical sensor fusion strategy is introduced, combining adaptive feature-weight optimization with ensemble learning to enhance fault separability under varying operational regimes. Unlike traditional single-model pipelines, the framework incorporates adaptive model calibration to maintain diagnostic performance under distributional shifts. Experimental evaluation under variable load and speed conditions demonstrates improved fault classification accuracy, enhanced early anomaly sensitivity, and stable performance in noisy environments. The system supports scalable deployment in embedded and edge-computing environments, enabling real-time inference with reduced computational overhead. The proposed approach provides a generalized and resilient condition monitoring solution that enhances predictive reliability, minimizes unplanned downtime, and supports intelligent maintenance strategies in next-generation cyber-physical industrial systems.