Deployment of Safety Predictive Analytics to Prevent Workplace Incidents and Promote Event Reduction: A Machine Learning Approach Towards a Data-Driven Safety System

With the ongoing era of digitization and the ability to bring together huge data sets, there is enormous potential for many industries and business functions to take advantage of their “big data” sets by performing advanced analytics, allowing managers to make business decisions. When applied correctly, predictive analytics allow organizations to gain deep insight into their business and deploy their scarce resources in an optimal way (Predictive Solutions, 2012). Thus, improving safety in workplaces is no exception to using predictive analysis to predict and prevent events in real-time. Increasing process safety entails lowering the likelihood that hazardous environments may have negative effects. Unfortunately, it is impossible to measure the probability factor directly. The absence of an accident does not provide helpful information because Target Zero cannot be accomplished until the hazard is eliminated. On the other hand, numerous cause elements that could have been better monitored and controlled are discovered after a significant accident occurs. In this paper, we advocate for a proactive, dynamic system by embedding predictive analysis into two process stages. At the design stage, we suggest a predictive modeling process: 1. Define goals 2. Collect and preprocess data 3. Address missing data through data imputation 4. Build candidate predictive models 5. Train, test and cross-validate candidate models 6. Select “the best” models based on multiple evaluation metrics 7. At runtime, consume the predictions from selected predictive models using the end-user interface. Depending on the data input and the time it takes for learning algorithms to learn, we claim the system can predict future potential input variables for the system in a single model and proactively adjust to meet intermediate and longterm goals and safety needs. A subset of data supports the proposal through in-sample validation. The validation through simulation was done based on actual data extracted. Several questions and requirements were answered based on the feasibility of using our approach to lead the proactive adaptation at runtime for safety monitoring systems.