Design of a virtual test cell based on GMDH-type neural network for a heavy-duty diesel engine

Abstract

The engine development process faces big challenges from new strict emission regulations in addition to the need for fuel efficiency improvements. The Software-in-the-Loop (SiL) and Hardware-in-the-Loop (HiL) environments decreases the required time during engine development, calibration, verification, and validation of the product. An accurate and easy to build dyno-engine model with real-time operational ability is required for this purpose. Artificial Neural Networks (ANN) have shown ability to model dynamic and complex systems like internal combustion engines. In this paper, the Group Method of Data Handling (GMDH) algorithm was utilized to build an ANN model of a heavy-duty diesel engine. One objective is to reduce the amount of manual labor on the results during the ANN model development process. The GMDH algorithm is a self-organizing process that will find the system laws and optimize the model structure automatically in one iteration. The GMDH model results were compared with a model developed by Levenberg-Marquardt Backpropagation (LM-BP) algorithm. The ANN models used actuator signals from an Engine Management System (EMS) to simulate the engine operation parameters. As revealed by the simulation results, the ANN models successfully predicted engine torque, fuel flow, and NOx concentration. The GMDH model as a self-organized model reduced lead time, had slightly higher transient cycle accuracy, had fewer inconsistent predictions, and demonstrated better extrapolation capability. The prediction accuracy for transient operation was improved by shifting the predicted value by calculating time delay and a decrease of 76.66% for fuel flow and 66.51% for NOX concentration in model accuracy were achieved. The GMDH dyno-engine model can be effectively applied as a virtual test cell instrument for testing, calibration, and optimization purposes.