Human Reliability and Interaction with Intelligent systems

Foundational services that are essential to the security, development, and welfare of a society are defined as Critical infrastructures. These are often of considerable age and were not designed to cope with modern-day usage. In addition, there exist new extraordinary risks, such as climate change and terrorist attacks that are not predictable on the basis of historical data. This is confounded by the increased interconnectivity of assets, with the function of one dependent on constant input from others, this provides mechanisms for one event acutely damaging one assets performance to cascade to others.
Although the human intervention remains the ultimate resource to mitigate the consequences of extreme events, Humans also remain the weakest link of such critical infrastructure. In particular, the interaction of human an autonomous and intelligent systems remains largely unknown.

This project will develop models to assess human performance after the occurrence of a disruptive event of interconnected critical infrastructures. The existing Human Reliability Analysis methods are focused on the human actions needed to maintain or activate the protection and mitigation measures of a system. It means that they are focused on the probability of a human making an error that initiates an accident event. However, if the protection measures in place fail to contain the evolution of the disruptive event, new human actions are needed.
Existing research on human performance on this phase usually covers the human escaping behaviour, to define better escape routes, but not the human performance for taking the necessary actions to recover the system.
Currently qualitative reliability methods do not provide the human error probability, but only its identification and possible solutions to prevent or mitigate human errors [2]. Although some safety regulators do accept qualitative analysis on human errors, human error probabilities are required by probabilistic safety assessment. Quantitative human reliability methods such as THERP, SPAR-H, HEART, CREAM and ATHEANA are often affected by imprecision, leading to under-estimated or over-estimated probabilities. This uncertainty may be one of the causes that is preventing industries from adopting risk assessments that account for human errors.
The present research proposes to develop an innovative approach to construct and calibrate a model for calculating the human error probability using data extracted from existing major accident investigation reports [1]. This approach has the potential to provide data that depict contexts and scenarios not fully achieved by simulator, near-misses and expert elicitation data [3,4]. Since the number of such reports are usually very small, additional information will be gained by accessing simulated data (e.g. from simulators and learning from similarities). The methodology allows to minimise the expert judgement in the definition of human error probability, assess the uncertainty and variability of human errors under different scenarios.