Table of Contents: :: Library Catalog

Gardado en:

Detalles Bibliográficos
Autor Principal:	Malone, Ryan
Formato:	Recurso digital
Idioma:
Publicado:	Zenodo 2026
Subjects:	variance-AR1 coupling fault detection process monitoring baseline susceptibility prediction Tennessee Eastman Process DAMADICS Secure Water Treatment industrial process control critical slowing down statistical process control monitoring variable prioritization chemical engineering reliability engineering cyber-physical systems permutation test
Acceso en liña:	https://doi.org/10.5281/zenodo.19918461
Tags:	Engadir etiqueta Sen Etiquetas, Sexa o primeiro en etiquetar este rexistro!

Table of Contents:

Current fault detection methods — PCA, PLS, and neural network approaches — detect anomalies after faults manifest in process data. While contribution plots, sensitivity analysis, and fault propagation studies provide variable-level diagnostic information, few methods offer a principled, purely baseline-driven approach for identifying which process variables will respond most strongly to faults without requiring fault data, labeled training conditions, or domain expertise. This paper introduces variance-AR1 coupling C = corr(Var_W, AR1_W) as a baseline susceptibility predictor: computed from rolling windows of normal-operation time series, it characterizes relative fault response magnitude across process variables before any fault occurs. On the Tennessee Eastman Process (TEP) benchmark — the standard chemical process fault detection dataset — baseline coupling provides a strong statistical predictor of relative fault response magnitude across all 52 process variables (Pearson r = −0.781, p = 8.9 × 10⁻¹², n = 52), confirmed by permutation test (empirical p < 0.0001, 10,000 shuffles) and robust across window sizes from 45 to 180 minutes. Coupling-prioritized variables show 7.16× larger mean coupling response under fault conditions than PCA-prioritized variables (mean delta +0.529 vs +0.074, Welch t = 3.550, p = 0.002), despite PCA explaining 99.998% of process variance — evidence that coupling captures dynamical co-movement orthogonal to variance magnitude. Multivariate regression confirms coupling accounts for 62% of unique predictive variance (ΔR² = 0.620 when removed); no individual baseline statistic (mean, standard deviation, AR1, variance, coefficient of variation) is statistically significant alone. Cross-domain validation on real industrial data confirms generalizability: DAMADICS actuator benchmark from an operational sugar factory, Cukrownia Lublin, Poland (r = −0.770, p = 2.0 × 10⁻⁶, n = 28 variables) and Secure Water Treatment (SWaT) infrastructure under documented cyberattack conditions (r = −0.649, p = 1.4 × 10⁻⁵, n = 37 variables). Six robustness analyses confirm result stability: permutation test, leave-one-fault-out (r range −0.791 to −0.767 across all 21 fault type exclusions), window size variation (80-90% top-10 variable selection overlap across adjacent windows), noise robustness (r = −0.768 at 10% additive Gaussian noise), alternative predictor comparison, and multivariate leave-one-out regression. The method is baseline-driven and parameter-light, requiring only normal operation data with no fault examples, no labeled training conditions, and no prior knowledge of variable roles. It complements existing fault detection infrastructure without replacing it, functioning as a monitoring system design tool that identifies which variables warrant closest attention before any fault occurs. Practically, focusing monitoring resources on the top-10 coupling variables (19% of all variables) captures the highest-susceptibility set across all 21 TEP fault types. All analysis code and processed results are archived on Zenodo (doi.org/10.5281/zenodo.19915719). This manuscript has been submitted to Reliability Engineering and System Safety. Author: Ryan W. Malone, Independent Researcher, Spring TX 77386, USA Email: <a class="underline underline underline-offset-2 decoration-1 decoration-current/40 hover:decoration-current focus:decoration-current" href="mailto:malonrw@gmail.com">malonrw@gmail.com</a> ORCID: 0009-0002-9583-232X

Títulos similares