REMAINING USEFUL LIFE PREDICTION OF BEARING UNITS OF MOBILE MACHINERY BASED ON DEGRADATION ONSET DETECTION AND AN EVALUATION PROTOCOL

Abstract

The paper addresses the problem of building the prognostic loop of an intelligent maintenance system for bearing units of mobile machinery using run-to-failure test data. Unlike diagnostic benchmarks with artificially seeded faults, run-to-failure trajectories expose the entire life of a unit, from run-in to natural failure. They therefore allow the validation of prognostic rather than purely classification methods. Four components of the methodology are proposed and experimentally verified on twelve natural degradation trajectories of the public IMS bearing dataset: four complete failure trajectories and eight censored ones. The first component of the methodology is an integral technical condition index (health index). It is constructed as the first principal component of fifteen snapshot features normalized to the baseline period of each trajectory. The trendability of the index is higher for failed bearings than for censored ones (0.654 versus 0.493, Mann–Whitney p = 0.055). However, it is not a universal indicator, because both false trends of healthy units (up to 0.737) and a failure without any trend (0.296) are documented. For this reason the index serves as an input to onset detectors rather than as a stand-alone decision criterion. The second component is a quorum ensemble of three heterogeneous degradation onset detectors: a robust MAD threshold, an EWMA control chart and an isolation forest. Each of them is confirmed by five consecutive snapshots against the self-healing effect. The ensemble detects the onset in three of four trajectories with lead times of 72–623 hours. This is sufficient to schedule the replacement outside the seasonal peak period and, in the best case, to order spare parts without urgent logistics. The k-out-of-N quorum rule turns detector disagreement on non-monotonic trajectories into an adjustable trade-off between early warning and false alarms. The third component is a defect-type identification rule based on matched scoring of the harmonic comb of bearing characteristic frequencies in the envelope spectrum. It yields three correct diagnoses out of four versus one out of four for the dominant-peak rule. Moreover, the diagnosis becomes readable almost simultaneously with the detected onset. The fourth component is an honest object-level evaluation protocol for remaining useful life regression (leave-one-bearing-out). Its comparison with the common random snapshot split quantifies the data leakage caused by temporal autocorrelation of a trajectory as a 2.5-fold understatement of the prediction error (RMSE 21.5 versus 53.8 hours). The same object-level requirement is also formulated for the calibration of conformal prediction intervals. A principal limitation of single-channel vibration monitoring is documented separately, namely a failure with no vibration manifestation in the available channel. It is translated into architectural requirements of channel and modality redundancy. The results extend the model selection methodology and the integrated diagnostic pipeline proposed in the authors' previous works with a prognostic loop of condition-based maintenance.