Robustness of Predictive-Regime Monitoring Under Sensor Degradation and Nonstationarity

A supervisory monitoring framework becomes credible for safety-relevant deployment not only by succeeding on clean benchmarks, but by showing predictable behavior under stress, degradation, and ambiguity. We characterize the robustness of a predictive-regime monitoring framework under nine categories of degradation organized into three phases: signal-path corruptions (colored noise, sensor saturation, missing data), nonstationarity and intermittency (slow parameter drift, bursty regime switching, mixed periodic–stochastic signals, actuator transients), and structural degradation of the monitoring infrastructure itself (reduced analysis windows, partial multi-channel failure). Rather than reporting binary pass/fail outcomes, we define a formal operating envelope that maps each degradation type and intensity to a trust classification: full trust, reduced confidence, or abstain. We measure detection fidelity (classification accuracy, false positive/negative rates, boundary index deviation) and component-level sensitivity (which risk components are most vulnerable to each stress type). Applied to canonical dynamical systems during controlled parameter sweeps, the results show that the monitoring framework maintains reliable classification under moderate white noise, low missing-data rates, and moderate sensor clipping. A counterintuitive finding is that strongly correlated (Brownian) noise preserves accuracy across all tested amplitudes, because the smooth noise structure does not disrupt model selection. Bursty intermittency is detected reliably even at short epoch lengths, while mixed periodic–stochastic signals produce negligiblefalse positive rates. Progressive channel failure produces graceful degradation in the coherence subsystem. The operating envelope provides practitioners with explicit guidance on when to trust, discount, or disregard regime assessments from the monitoring framework.