Using Machine Learning to Reduce Manual Incident Report Review by 60%

In a given year, a single nuclear power plant can generate 10,000 - 20,000 incident reports (IRs), documenting a wide range of possible situations where the safety of the plant may have been compromised. Each IR that is generated must be reviewed by an engineer to identify if a functional failure of high safety significance has occurred. The review process for this determination is extremely resource-intensive and comes at a large cost to the utility. Additionally, an incorrect classification of an IR could result in additional scrutiny from regulatory authorities.

Our team is developing techniques to automate this review process using machine learning incident management. Data on how employees previously classified incident reports is fed into a computer, which “learns” what constitutes a functional failure by using the text and classifications as a reference. Thus far, this automation process has been able to reduce the amount of incident reports that need manual review by 60%, significantly decreasing the time, labor and cost required to process these reports.