Physics-Informed Machine Learning for Combat Vehicle Performance Predictions

For modern vehicles governed by electronically-controlled transmissions, real-time predictions of operating characteristics have become an increasingly significant element in ensuring reliable performance. The practice of using the electronic control system for derating engine torque and power output in harsh environmental conditions or when under heavy loads protects the vehicle against potential malfunction and damage. For these systems, a network of sensors is used to continuously monitor various engine health parameters, including temperature, pressure, and load conditions to provide feedback to the electronic control system. Extrapolating this data to develop a comprehensive picture of vehicle health to identify a potentially unsafe operating state can be a complex, uncertain process, especially in cases of rapidly changing or intermittent conditions. For this reason, the Army has identified a critical need for modern data processing and predictive capabilities for combat vehicle powertrains operating in prototypical wartime environments. To address this need, Creare proposes a physics-informed machine learning (ML) approach to modeling and predicting modern combat vehicle powertrain and cooling system performance. Our approach implements physics-based dynamic simulation tools, calibrated by test data, which then serve as the ground truth for training of a deep learning neural network. The overall framework is designed for near?term implementation at the ArmyÆs Yuma Test Center (YTC) to allow more accurate characterization of vehicle performance, in particular for Full Load Cooling (FLC) tests. Relative to blackbox ML solutions, the combination of physics-based models and deep learning provides higher fidelity predictions, more efficient training of ML models, and allows a greater degree of interpretability into the inner-workings of the ML. These features are imperative to improving the ArmyÆs test capabilities and ultimately to the safety of the warfighter and successful combat operations. During Phase I, we will develop detailed physics-based dynamic models of critical vehicle systems using MATLAB Simulink or similar software, calibrate these component and system-level models using test data as available, use the physics-based model for training a deep learning neural network, and finally use this framework to demonstrate simulated data generation of the combat vehicle systems for a limited set of environmental conditions. In Phase II, we will build upon the initial work by carrying out further refinements, we will develop a packaged software suite with GUI, and formulate detailed plans for field testing with the Army.