DSFB-Chemical-Engineering: Read-Only Residual Semiotics for Chemometrics-Augmented Fault Detection and Diagnosis in Chemical Engineering, with a Deterministic, Byte-Exact, CUDA-Accelerated Forensic Evidence Court

AbstractChemical process monitoring relies on mature chemometrics and statistical process-monitoring tools: PCA/PLS score-space statistics, squared prediction error (SPE/Q),Hotelling’s T2, dynamic and kernel PCA, independent component analysis, contributionplots, EWMA/CUSUM-style temporal detectors, and modern machine-learning anomalymodels. These methods are effective, yet many of their intermediate outputs — residuals,reconstruction errors, contribution traces, threshold excursions, lagged deviations, anddetector disagreements — are treated as transient alarms or discarded. We introduce DSFB-Chemical-Engineering, a deterministic, read-only augmentation framework that elevatesthose residual products into a structured, auditable evidence language. It ingests residualstreams without replacing the underlying estimator, controller, historian, or alarm system,then applies Drift–Slew Fusion Bootstrap (DSFB): first-order residual drift, first-order rate-of-change slew, admissibility envelopes, deterministic detector-consensus fusion, and a chemicalheuristics bank, producing typed structural episodes with deterministic replay.We evaluate on twenty public chemical-engineering / chemometric datasets (ten measuredreal-world, nine field-standard simulation benchmarks including the Tennessee Eastmanprocess, one agreement-gated stand-in) spanning plant-wide process faults, spectroscopy,e-nose drift, batch fermentation, semiconductor and steel manufacturing, and wastewatertreatment. A second crate accelerates evidence production on an NVIDIA RTX 4080 SUPERusing fixed-point arithmetic and on-GPU SHA-256, and we show that the GPU evidence rootis byte-for-byte identical to an independent CPU reference on all twenty datasets, sealingresiduals (and the usually-discarded noise floor) into a hash-linked, replayable forensic case file.Separately, we contribute a falsifiable balance-witness applicability criterion: a mass/energybalance admits evidence only when a closed, fully-metered control volume is made to appearnon-conserved (a sensor spoof/drift, or a leak crossing the metered boundary), which weverify by firing with spatial selectivity on the real BATADAL and SWaT testbeds and bycorrectly staying silent on six datasets the closure gate rejects for distinct physical reasons.The goal is explicitly not to claim superiority over established chemometrics; it is to testwhether deterministic residual semiotics can improve fault-narrative compression, detector-agreement auditing, and reproducible operator-facing diagnostics. We report results honestly,including a first-class unknown rate and the datasets on which the framework struggles. Wealso provide a deliberately exhaustive limitations section, including a one-hundred-questionhostile-reviewer attack surface, as scientific armor rather than as an afterthought.