Adaptive Incremental Prognostics for Remaining Useful Life Under Changing Operating Conditions

Remaining Useful Life (RUL) prediction is central to reliability assurance and predictive maintenance. Deep learning models capture nonlinear degradation behaviour but often require substantial computation and retraining, while existing Broad Learning System (BLS) approaches have limited adaptability to evolving degradation patterns and operating conditions. This paper proposes the Adaptive Transfer Broad Learning System (ATBLS), a lightweight framework for incremental RUL prediction under changing operating conditions. ATBLS combines structured enhancement fusion, sequential projection, PCA-guided contribution weighting, and analytical ridge regression within a single adaptation pathway. Enhancement fusion reorganizes nonlinear responses before fixed sequential projections generate complementary degradation representations for prediction and cross-dataset comparison. Projection aggregates form normalized-life descriptors in a jointly recomputed shared PCA space to estimate participant relevance and source–incremental domain contributions. These contributions weight the participating transition matrices before one closed-form ridge solution is obtained, while the feature, enhancement, fusion, and projection mappings remain fixed. Thus, newly available labeled data are incorporated without end-to-end retraining. ATBLS is evaluated on three run-to-failure benchmarks covering rolling bearings and turbofan engines, including within-domain incremental learning and explicit cross-domain operating-condition transitions. Results demonstrate accurate RUL prediction, low analytical adaptation cost, and substantial improvements in several cross-domain transitions, although the benefit of incremental target-domain information remains condition dependent.