Ensemble Learning for Fault Condition Prediction and Health Status Monitoring in Military Ground Vehicles

Mun, S., Hwang, J., Bian, L., Falls, T. C., & Bond, W. G. (2023). Ensemble Learning for Fault Condition Prediction and Health Status Monitoring in Military Ground Vehicles. IISE Annual Conference & Expo 2023. New Orleans, LA.

Unexpected technical issues such as engine and transmission failure represent critical Reliability, Availability, and Maintainability (RAM) issues for military ground vehicles. It is essential to keep the vehicles in healthy condition and to provide predictive maintenance that enables efficient diagnosis and repair of vehicle failures, reduces associated operation/sustainment costs and vehicle downtime, and supports predictive logistics for critical components. This paper develops a predictive model for early fault detection using a machine learning (ML) algorithm trained by a real vehicle data set. The ML model is built on various operational time series sensor data and fault codes collected via a Digital Source Collector and a CAN bus device for several vehicles over extensive time intervals. The approach proposed here is an ensemble learning of a multivariate Long Short-Term Memory (LSTM) neural network for operational data forecasting based on select channel data such as coolant temperature, engine oil temperature/pressure, and battery voltage with respect to the driving status. The LSTM model can use recorded parameters to classify vehicle or component reliability and health. For further improvement of prediction accuracy and generalization of the model across various vehicle types, multiple independent LSTM models are trained over training/validation datasets from the randomly subsampled period of a single vehicle or those from the same families of vehicles. Outputs from the individual networks are then linearly combined to produce the output of the ensemble network. The analysis shows better prediction accuracy than the single LSTM approach, providing promising early fault detection performance.