Vibration-Based Health-State Monitoring and Time-to-Critical-Threshold Prediction of a Motor-Generator System Using ML/DL Models

Rotating electrical machines are critical components in industrial systems, and variations in vibration behavior can provide valuable information about their operational health condition. This study proposes a vibration-based framework for health-state monitoring and time-to-critical-threshold prediction in a laboratory motor–generator system under variable loading conditions. A wound-rotor induction motor mechanically coupled to a direct-current shunt generator was instrumented with an ADcmXL3021 triaxial MEMS vibration sensor. Measurements were collected under a reference condition and three generator-side resistive load conditions. The three acceleration components were combined into a resultant vibration magnitude and segmented into fixed-size windows of 1024 samples. After preprocessing, 38,799 feature vectors were generated from approximately 181.25 s of vibration measurements. Seven statistical time-domain features were extracted and fused using Principal Component Analysis (PCA) to construct a Health Indicator (HI). Baseline-derived warning and critical thresholds were defined to characterize Normal, Warning, and Critical vibration states. Several Machine Learning and Deep Learning models, including Random Forest, Support Vector Regression, LSTM, GRU, BiLSTM, CNN1D, and xLSTM-inspired models, were evaluated for HI prediction and time-to-critical-threshold estimation.

The PCA analysis showed that the first principal component explained 49.1% of the total variance, while the first three components captured 88.9%. Support Vector Regression achieved the best HI prediction performance, with an MAE of 0.009, an RMSE of 0.012, and an R² of 0.955. CNN1D provided the most accurate time-to-critical-threshold prediction, achieving an MAE of 1.821 s, an RMSE of 3.415 s, and an R² of 0.967. The proposed framework provides an interpretable solution for vibration-based condition monitoring under controlled operating variations. The predicted remaining time corresponds to the duration required for the HI to reach the predefined critical threshold and does not represent physical run-to-failure Remaining Useful Life estimation. The results highlight the potential of the approach for future edge-based monitoring and Digital Twin applications.