Process of Building AI Models for Predicting Engine Performance and Emissions

Simulation-based product development requires efficient engine models that can not only predict engine performance and emissions accurately but also produce results at real-time computing speeds. These models are used to predict in a closed-loop system-level environment integrated with controls and other components. 0D/1D physics-based software can model engine performance with good accuracy, but these models cannot be deployed in a closed-loop environment with controls and other components because they have a significantly higher runtime—about 20 to 40 times slower than real time. Investigation was done to find a low-fidelity model or algorithm with similar performance that could replace high-fidelity physics-based models in the simulation domain and can be deployed in a system level environment. The robustness of AI models’ ability to learn from the data encouraged Cummins to try several machine learning models. A detailed approach was taken to build LSTM-based deep neural network models that achieved target model accuracy with a runtime of about one-eighth that of real time. Cummins developed a total of 26 models using MATLAB^® to predict 26 different responses of the engine which consist of temperature, pressure, and flows across multiple locations of the engine along with emissions, efficiencies, and engine brake torque. In the process of building these deep learning models, multiple techniques were evaluated such as defining an optimal deep neural network structure consisting of the type and number of layers, using different activation functions, and optimizing associated hyperparameters. In future, the process of building these models will be automated, and the use of MATLAB to train the models in parallel will be a significant advantage. As these models meet all the requirements, future work would be done to integrate these models in the pure simulation domain, followed by integration with actual hardware and control.