Awareness based internal reliability monitoring of streaming models using incremental principal component analysis

Abstract Streaming models in predictive maintenance and cyber–physical systems operate under non-stationary conditions with delayed or unavailable ground-truth labels. While existing approaches detect discrete drift events, they provide little insight into how reliably the underlying model is operating over time. This paper introduces a continuous, label-free internal reliability metric built on an Incremental Principal Component Analysis (IPCA) backbone. Reconstruction residuals are normalized using slowly updated reference statistics and temporally smoothed to produce a bounded awareness signal that reflects persistent internal strain rather than isolated residual spikes. Evaluated in a one-pass streaming setting using curated, publicly available datasets, the signal remains smooth under nominal conditions and stays close to baseline during successful online adaptation. However, when adaptation is constrained during gradual drift, the signal exhibits dataset-dependent changes that remain smooth and bounded. The observed differences are small in magnitude, indicating that the awareness signal is more suitable as a continuous internal reliability indicator than as a universal detector of frozen adaptation. Our proposed metric provides a lightweight and label-independent indicator that complements classical drift detectors by offering a continuous view of model health under evolving conditions.