- A mining company relies on high-payload-capacity trucks to haul extracted material for extraction of ore and minerals. An unscheduled downtime incident, due to an engine component failure, can result in a productivity loss of no less than $200,000. While scheduled maintenance goes a long way to reduce these events, the client wanted to take advantage of data generated by the engine systems to predict faults of specific key components, in order to replace them before they failed.
- We started by integrating, transforming, and aggregating terabytes of high-frequency (one-second interval) data produced by the monitoring systems of the trucks. A key factor in building reliable predictive models has to do with the so-called “feature engineering”, or data transformations, and this sometimes has a larger impact on model performance than the selection of the algorithm itself.
- For predictive modeling we explored various traditional methods, such as logistic regression and decision trees, as well as newer methods, such as XGBoost. In most cases, XGBoost had the best performance, and in conjunction with methods that shed light on input variable importance, we were able to make predictions and understand the key factors behind those predictions.
- Engine component failures are rare, but their rarity makes predicting events very challenging for models. The risk of producing too many false positive alerts erodes confidence in the models and diminishes or dismisses their use.
- Fortunately, newer methodologies are getting much better at rare event analysis. Depending on the engine component, we were able to predict faults or failures upwards of 80% of the time, with a minimal number of false positives. We were able to do this in a short-term horizon (up to three days in advance) but found that some of the best warning signals developed as far as 28 days out.
- With an early warning system in place, the company is able to maintain high availability and productivity of their assets and avoid the financial cost of down time.