Sequence Based Classification for Predictive Maintenance

PACCAR, a commercial vehicles company, has provided a dataset of sensor snapshots that were taken using communication networks during the operation of several vehicles, as well as repairs that were performed for specific failures on these vehicles. Our objective is to predict future repairs for predictive maintenance. Most approaches in the literature involve predicting repairs with only one time window. We show that in addition to the repair, time window prediction is significantly more difficult due to various inherent challenges associated with the sensor snapshot and imbalanced data. Models employed for prediction are Random Forest (industry standard), LSTM RNN, and GMM (Gaussian Mixture Models). Our findings show that while time window is intractable, repair code prediction is best at 43% recall with Random Forest.