Answers are grounded in this paper and cite their sources.
Data‐driven fault detection and diagnosis (FDD) models are accurate on standard benchmarks, but their decision logic is rarely available in a form that engineers can check against process knowledge and safety requirements. This paper studies the Learning Probabilistic Boolean Network (LPBN), a classifier that learns a circuit of two‐input Boolean gates over named threshold predicates on process statistics and collapses, after training, into a deterministic decision specification. Four research questions are evaluated on the 21‐class Tennessee Eastman Process against black‐box, probabilistic, Bayesian‐network, interpretable, and symbolic‐regression comparators. On per‐variable episode statistics, the LPBN reaches a macro‐F1 of 0.931, and no comparator is significantly more accurate after correction for multiple comparisons; with a sparse readout, its class decisions use 7–10 terms over 22 sensors at 0.91–0.92, whereas symbolic regression reaches 0.81 with 62‐node expressions. An audit case study encodes the circuit, physical axioms, and safety requirements for satisfiability solvers; it shows that the accurate circuit ignores reactor pressure, saturates on reactor temperature, and is non‐monotone outside the normal band, none of which test metrics reveal, and it re‐verifies a corrected decision procedure without loss of accuracy. The confusion between normal operation and fault 9, shared by the comparators, is traced to the episode representation rather than to the Boolean logic or the sensors: a residual statistic of a predictor fitted to normal operation separates the two classes almost perfectly, and with residual predicates the LPBN reaches, for one training seed, a macro‐F1 of 0.985 and a fault‐9 class F1 of 1.00.
No takes yet. Share an insight, caveat, or question.
Torres et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: