Chi-Squared Detection of Physical Accident Signatures in Nuclear Power Plant Data, with Synthetic Adversarial Stress Test

Nuclear-Renewable Hybrid Energy Systems (N-RHES) integrate small modular reactors with intermittent renewable generation to deliver flexible, zero-carbon grid balancing. However, their interconnected cyber-physical architecture remains susceptible to adversarial manipulation of digital instrumentation and control pathways. Advanced consequence-driven methodologies, such as Consequence-driven Cyber-Informed Engineering (CCE), prioritize engineering out high-consequence cyber risks rather than relying solely on post hoc detection. This paper presents a benchmark of lightweight detectors motivated by the CCE analytical workflow, instantiated as an anomaly detection pipeline, evaluated on the Nuclear Power Plant Accident Data (NPPAD) dataset. We compare a classical chi-squared (Mahalanobis distance) detector against an Isolation Forest detector, alongside per-feature z-score and PCA reconstruction baselines. All detectors are evaluated at a matched 1% target false alarm rate (FAR) on a random split of the 302 normal samples, yielding an achieved FAR of 0.98%. Chi-squared achieves an AUC of 0.9989 (95% CI [0.9975, 0.9997]) and a 96.61% sample-level detection rate, significantly outperforming Isolation Forest (AUC 0.9762, 81.79% detection rate; DeLong p<0.001). We do not claim to exceed published deep learning baselines, because their evaluation protocols differ substantially. Per-accident analysis shows all 17 accident types are detected at the scenario level, 15 of 17 exceed 95% sample-level detection, and 14 of 17 achieve 100% sample-level detection, with load rejection (70.6%) and rod insertion (92.5%) as the exceptions. To bridge the gap between physical accident detection and the adversarial threat model, we introduce a synthetic false-data-injection stress test. Under bounded sensor drift at the stealthy attack band (0.5σ), chi-squared detection remains at 1.5% ± 0.7%, indistinguishable from its 0.98% false alarm rate. This demonstrates that the 96.61% physical-accident detection rate does not transfer to adversarial manipulation, and that the detector ranking changes under adversarial stress. We interpret this as a dataset-specific finding, NPPAD presents large, physically distinct anomaly signatures well-suited to distance-based detection. This work demonstrates methodological transferability and explicitly characterizes the fidelity gap between commodity-platform analysis and hardware-in-the-loop validation.