Performance management of safety instrumented systems (SIS) is essential for reducing major accident risk in oil and gas facilities, and has traditionally been governed by standards such as IEC 61508 and 61511. These standards prescribe how performance shall be managed for conventional systems, but provide little concrete guidance on how machine learning (ML) can be incorporated. As facilities move toward unmanned operation, where manual testing and inspection are infeasible, this lack of standardized rules for ML-based SIS becomes as challenging as the technical issues of data imbalance and sample selection bias. Without established rules, certification bodies have no framework for evaluating whether an ML-based SIS can be considered sufficiently reliable. This paper proposes an ML-based SIS advisor designed to improve operator awareness and decision support under varying operating conditions. To overcome data limitations, we introduce a simulation-based generative oversampling method (SIM-GEN) that produces abnormal-operation data from dynamic process models, thereby extending training coverage and addressing imbalance and support mismatch. Engineering priors are incorporated through HAZOP-based pseudo-labels to stabilize the training process. To address this lack of standardized measures, we further propose how the functional safety performance measure can be derived from the confusion matrix. These serve as dataset-level surrogates of the IEC 61508 failure rate categories. While these measures are presently static proxies, they can be extended in actual operation by integrating diagnostic functions and time-based monitoring, enabling convergence toward actual performance measures as defined in functional safety standards. Importantly, in the pre-deployment phase, these indices offer certification bodies a practical and reproducible basis for auditing ML model performance, supporting design verification and approval prior to operation. Experiments show that SIM-GEN improves robustness under imbalance and bias and outperforms conventional resampling techniques. The dynamic process simulator K-SPICE was used only to generate abnormal-operation scenarios during data preparation; the SIM-GEN method itself is independent of the particular simulation platform. • Formulates abnormal condition detection task in unmanned facilities as a machine learning classification problem, explicitly addressing class imbalance and sample selection bias, and linking functional safety practice with machine learning approaches. • Proposes a simulation-based generative oversampling method (SIM-GEN) that leverages dynamic process simulation to generate realistic abnormal operation data, thereby mitigating severe imbalance and sample selection bias in safety-critical training datasets. • Introduces functional safety–oriented performance indices ( η D , η S ), which, although not yet standardized, provide concrete numerical evidence beyond IEC 61508-2 guidelines and can serve certification bodies and practitioners as transparent benchmarks.
Lee et al. (Wed,) studied this question.