Multi-Sensor Vibration-Based Classification of Shaft Misalignment Severity Using Interpretable Machine Learning: An Experimental Study
This work investigates the application of data driven techniques for classification of misalignment severity using vibration measurements acquired from multiple sensor locations. Experiments were conducted under four operating conditions representing healthy operation and three levels of shaft misalignment at several motor speeds and loads. A set of time-domain and frequency-domain features was extracted and used to train Random Forest classifier. Separate models were developed for two individual sensors as well as for a combined multi-sensor feature set. While overall accuracies were comparable (0.879-0.895), cross-condition validation using Leave-Condition-Out (LCO) revealed that the sensors exhibit different sensitivities to operational changes. Analysis of prediction disagreement showed that 12% of signal segments were correctly identified by only one of the sensors, particularly at the lowest operating speed (20 Hz). This suggests that multi-sensor fusion enhances reliability in scenarios where individual sensor signals are weak. To improve interpretability, SHAP-based explanation techniques were applied to the fusion model. The analysis showed that several physically meaningful features - such as spectral bandwidth, RMS value, vibration range, and form factor - play a dominant role in the classification process and exhibit clear relationships with fault severity. The results demonstrate that while sensor fusion provides marginal gains in average accuracy, it effectively reduces diagnostic “blind spots”, offering a more robust framework for automated misalignment evaluation within the studied operational range.