Safety Predictive Model with Machine Learning and Its Application in DART Analysis in Utility

Abstract

Workplace safety is critical for utilities and field operations, where complex work orders pose significant risks to employees and operations. Safety Predictive Model (SPM) is a data-driven solution that proactively identifies high-risk work orders and supports mitigation strategies. SPM uses enterprise data such as material group codes, work order types, location attributes, seasonal factors, and circuit details. Injury and Serious Injury/Fatality (SIF) records are linked to work orders to strengthen model accuracy. Historical data informs modelling, while new data validates performance. Over 90 variables are engineered for machine learning, with predictive strength assessed via Information Value. Algorithms tested include logistic regression, decision trees, SVM, and gradient boosting, with ensemble methods selected using ROC AUC and KS statistics. A composite risk score flags top deciles as high risk, applying district-specific thresholds. Beyond assessing upcoming work orders, SPM reveals key risk drivers, enabling utilities to anticipate and mitigate safety challenges effectively.

Keywords: Workplace Safety, Machine Learning, Artificial Intelligence, DART, Injury And Serious Injury/fatality (SIF), OSHA

DOI: 10.54941/ahfe1007695