Condition-Aware Online Cleaning and Context-Aware Imputation of Torque Signals in Sliding-Bearing Friction Tests Under Variable Operating Conditions

Torque signals recorded during sliding-bearing friction tests under variable operating conditions are non-stationary and are contaminated by load–speed transitions, electromagnetic disturbance, short mechanical shocks and communication errors, so that global anomaly detectors misclassify normal high-load samples as outliers. This paper presents SG-IF-BiLSTM, a condition-aware framework for the online segmentation, anomaly detection and repair of such signals. Candidate operating-condition boundaries are located from Savitzky–Golay derivative extrema and accepted only when the shift between the local means on both sides exceed three local standard deviations. An Isolation Forest is then trained inside each regime, and its decision limit is calibrated by the interquartile range of the local anomaly score. Detected abnormal samples are finally reconstructed by bidirectional LSTM that uses the context on both sides of the gap. The method is validated on two independent 10 min records acquired at 50 Hz on a sliding-bearing friction and wear test bench. It reaches 93.1% segmentation accuracy and reduces the false-positive rate from 38.7% to 4.6%.