Details
| Original language | English |
|---|---|
| Article number | 105528 |
| Journal | Safety Science |
| Volume | 146 |
| Early online date | 27 Oct 2021 |
| Publication status | Published - Feb 2022 |
Abstract
The capability of learning from accidents from different industrial sectors could prevent similar accidents to happen. With this aim, the Multi-attribute Technological Accidents Dataset (MATA-D) has been created, using a classification focused on the relation between human errors and their influencing factors (e.g., cognitive functions, organisational and technological factors). The process of collecting new data for this dataset should be constant, not only to decrease epistemic uncertainty in human reliability data but also to reflect changes in human behaviour due to evolving technology and organisational arrangements. However, reading an accident report is a time-consuming process, which delays the learning process. For this reason, this research proposes an automated approach to train the computer on a predefined classification scheme (taxonomy), which will be called the virtual human factors classifier. The virtual classifier should support human experts to analyse accident reports for organisational, technological, and individual factors that may trigger human errors. The proposed approach is based on classifying text according to previously labelled accident reports by human experts. Two case studies are used to demonstrate how data from different sectors can be used to train the machine, providing an efficient cross-discipline knowledge transfer. The accuracy of the results is promising and comparable to the classifications provided by human experts. The proposed work demonstrated to the industry the feasibility of the use of artificial intelligence to collect data and support risk and reliability assessments, and recommendations based on the study findings are suggested for investigation agencies.
Keywords
- Accident report data, Automated text classification, CREAM, Human factors taxonomy, Human reliability data
ASJC Scopus subject areas
- Engineering(all)
- Safety, Risk, Reliability and Quality
- Social Sciences(all)
- Safety Research
- Medicine(all)
- Public Health, Environmental and Occupational Health
Sustainable Development Goals
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
In: Safety Science, Vol. 146, 105528, 02.2022.
Research output: Contribution to journal › Article › Research › peer review
}
TY - JOUR
T1 - Identification of human errors and influencing factors
T2 - A machine learning approach
AU - Morais, Caroline
AU - Yung, Ka Lai
AU - Johnson, Karl
AU - Moura, Raphael
AU - Beer, Michael
AU - Patelli, Edoardo
N1 - Funding Information: Caroline Morais gratefully acknowledges the Brazilian Oil & Gas regulator ANP (Agencia Nacional do Petroleo, Gas Natural e Biocombustiveis) for the support for her research. Edoardo Patelli was partially supported by the EPSRC grant EP/R020558/2 Resilience Modelling Framework for Improved Nuclear Safety (NuRes). The authors also acknowledge Mrs. Raneesha for helping to implement the virtual classifier in the Cossan website, and Jack Tully, a final year graduation student for helping to test the code.
PY - 2022/2
Y1 - 2022/2
N2 - The capability of learning from accidents from different industrial sectors could prevent similar accidents to happen. With this aim, the Multi-attribute Technological Accidents Dataset (MATA-D) has been created, using a classification focused on the relation between human errors and their influencing factors (e.g., cognitive functions, organisational and technological factors). The process of collecting new data for this dataset should be constant, not only to decrease epistemic uncertainty in human reliability data but also to reflect changes in human behaviour due to evolving technology and organisational arrangements. However, reading an accident report is a time-consuming process, which delays the learning process. For this reason, this research proposes an automated approach to train the computer on a predefined classification scheme (taxonomy), which will be called the virtual human factors classifier. The virtual classifier should support human experts to analyse accident reports for organisational, technological, and individual factors that may trigger human errors. The proposed approach is based on classifying text according to previously labelled accident reports by human experts. Two case studies are used to demonstrate how data from different sectors can be used to train the machine, providing an efficient cross-discipline knowledge transfer. The accuracy of the results is promising and comparable to the classifications provided by human experts. The proposed work demonstrated to the industry the feasibility of the use of artificial intelligence to collect data and support risk and reliability assessments, and recommendations based on the study findings are suggested for investigation agencies.
AB - The capability of learning from accidents from different industrial sectors could prevent similar accidents to happen. With this aim, the Multi-attribute Technological Accidents Dataset (MATA-D) has been created, using a classification focused on the relation between human errors and their influencing factors (e.g., cognitive functions, organisational and technological factors). The process of collecting new data for this dataset should be constant, not only to decrease epistemic uncertainty in human reliability data but also to reflect changes in human behaviour due to evolving technology and organisational arrangements. However, reading an accident report is a time-consuming process, which delays the learning process. For this reason, this research proposes an automated approach to train the computer on a predefined classification scheme (taxonomy), which will be called the virtual human factors classifier. The virtual classifier should support human experts to analyse accident reports for organisational, technological, and individual factors that may trigger human errors. The proposed approach is based on classifying text according to previously labelled accident reports by human experts. Two case studies are used to demonstrate how data from different sectors can be used to train the machine, providing an efficient cross-discipline knowledge transfer. The accuracy of the results is promising and comparable to the classifications provided by human experts. The proposed work demonstrated to the industry the feasibility of the use of artificial intelligence to collect data and support risk and reliability assessments, and recommendations based on the study findings are suggested for investigation agencies.
KW - Accident report data
KW - Automated text classification
KW - CREAM
KW - Human factors taxonomy
KW - Human reliability data
UR - http://www.scopus.com/inward/record.url?scp=85117809268&partnerID=8YFLogxK
U2 - 10.1016/j.ssci.2021.105528
DO - 10.1016/j.ssci.2021.105528
M3 - Article
AN - SCOPUS:85117809268
VL - 146
JO - Safety Science
JF - Safety Science
SN - 0925-7535
M1 - 105528
ER -