Creating an incident investigation framework for a complex socio-technical system: Application of multi-label text classification and Bayesian network structure learning

The power distribution sector presents a complex socio-technical system where accidents frequently occur from various technical, human, environmental, and organizational factors, resulting in fatalities and substantial economic losses. The dynamic operational environment and complex interactions among the causal factors further complicate effective risk management and accident prevention. This research proposes a methodology to identify various risk factors and develop causal networks representing the complex relationships among these factors in power distribution incident reports. First, machine learning multi-label text classification identifies the risk factors from the incident reports. Then, the relationship among these factors is determined by integrating experts’ domain knowledge and data-driven Bayesian network structure learning approaches. Finally, the most influential causal factors and their direct/indirect effects on the incidents are identified, and proper risk control measures are recommended. The proposed methodology is applied to an incident database from a Canadian power distribution company, covering power outages, injuries, environmental issues, and near misses collected from 2013 to 2020. The results highlight that human and technical factors are the most influential and affected by organizational and environmental factors. Considering their complex interaction, implementing targeted risk management for high-risk direct/indirect causal factors could prevent further incidents and improve the companies’ overall safety.