Abstract
The prediction of equipment failures typically falls along two paths: statistical analysis of past failures, or using sensor data to monitor the condition of equipment in the field. This study instead takes a novel approach by implementing machine learning on the data within maintenance systems (which are typically used to monitor and track maintenance activities/workflows). This implementation serves to map the correlations between maintenance data and failures, to predict failures by only monitoring an organization's maintenance system.
This involved identifying maintenance system variables or attributes that normally represent specific steps in the maintenance process, but could potentially provide indirect information on the reliability of assets or the risk of future failure (e.g. work orders created/completed, work order types, original equipment manufacturer recalls/upgrades, inventory rotation rates, procedure changes, turnaround events, etc.). Machine learning is then used to determine correlations between failures and system variables (which results in an optimized model). The resulting model can then be implemented on current data to predict future failures.
Although largely designed around workflow management, the underlying metadata from an organization's maintenance system can indeed be transformed to provide valuable insights and quantify risk. Machine Learning is able to predict failures with a meaningful degree of accuracy. Even if the transparency of model logic is insufficient to take specific actions to mitigate failures, it is enough to dynamically assign a level of risk to individual assets, and even predict risk for future time periods. When interpreting the optimized machine learning model with methods such as feature impact charts, it is also possible to identify the strongest correlations between failures and different Computerized Maintenance Management System (CMMS) derived variables (variables that were designed and transformed from actual historical data for the purpose of model building). Insights gained from this additional analysis, such as the true impact that preventive maintenance delays have on risk, offer great potential for understanding the behavior of equipment and maintenance organizations.