Artificial Neural Networks and Sensitivity Analysis to overcome performance shaping factor’s dependency in the estimation of Human Error Probabilities (HEPs)

Albati, Mohammad; Bui, Ha; Sakurahara, Tatsuya; Mohaghegh, Zahra

Permalink

https://hdl.handle.net/2142/121843 Copy

Description

Title

Artificial Neural Networks and Sensitivity Analysis to overcome performance shaping factor’s dependency in the estimation of Human Error Probabilities (HEPs)

Author(s)

• Albati, Mohammad
• Bui, Ha
• Sakurahara, Tatsuya
• Mohaghegh, Zahra

Issue Date

2023

Keyword(s)

• Human Risk Assessment (HRA)
• Artificial Neural Networks (ANN)
• Machine learning
• PSFs
• Dependency

Abstract

In probabilistic risk assessment, human errors are considered one of the main factors leading to undesirable consequences. To estimate the probability of human errors, different Human Reliability Assessment (HRA) methods have been developed that utilize the available historical/experimental data of human failures. These HRA methods use a base HEP value and different levels of Performance Shaping Factors (PSFs) with their associated numerical multipliers to estimate the final HEP. In considering the effect (the multiplier) of each PSF on the final HEP value, the effects of other PSFs are completely ignored or are not correctly modeled. Thus, dependency between the effects of the PSFs on the final HEP value is not captured. In this paper, we suggest an approach to handle such kind of dependency. The suggested approach is utilizing an Artificial Neural Network (ANN) which has input layer (with number of neurons equal to the number of PSFs plus any required additional parameters such as the type of task), a hidden layer, and an output layer with two neurons representing the failure or success of the task. The last layer utilizes SoftMax as the activation function in order to obtain a probability distribution among the two outputs of the last layer. The SoftMax value of the neuron which was used to represent failure will be considered as the probability of human error. This approach calculates the human error probability given our knowledge of the level of the performance shaping factor without using separate multipliers for each PSF. The HEP value will be a function of the joint distribution of the different PSFs. For example, in SPAR-H, the HEP value for a “high complexity” action task performed in a “high” level of stress would be equal to the base HEP multiplied by 2 (for the high stress) and 5 (for the high complexity). In our suggested approach the HEP value will be calculated as a function of the PSF levels without any multipliers. In other words, the HEP value can be represented by the equation 𝐻𝐸𝑃 = 𝑓(𝑠𝑡𝑟𝑒𝑠𝑠 𝑙𝑒𝑣𝑒𝑙, 𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝑖𝑡𝑦 𝑙𝑒𝑣𝑒𝑙).

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121843

Owning Collections