Improving Accuracy of Virtual Torque Measurement using Artificial Intelligence Methods

Date of Award

4-2025

Degree Name

Master of Science in Engineering

Department

Mechanical and Aerospace Engineering

First Advisor

Muralidhar K. Ghantasala, Ph.D.

Second Advisor

Richard Meyer, Ph.D.

Third Advisor

John Bair

Keywords

Artificial intelligence, long short-term memory, reciprocating torque, torque sensor, virtual torque

Access Setting

Masters Thesis-Abstract Only

Restricted to Campus until

4-1-2035

Abstract

Predicting engine torque from instantaneous engine speed has been proposed as a simple and inexpensive way to predict real-time torque of an engine. Engine torque is predicted using the n-th order Fourier series amplitude, but the accuracy of the predictions is reduced when using this method due to the effect of reciprocating torque; an inertial torque in the engine due to reciprocating masses. Accounting for this can be done using engine speed phase conditions which mathematically remove its effects. This results in mean absolute percentage errors above 10% which can be reduced through the application of artificial intelligence (AI). This leads to the development and evaluation of an artificial neural network using a long short-term memory layer and a random forest regression model. These AI architectures apply the reciprocating torque correction during model training when given engine speed phase and flywheel angular acceleration as predictors along with CAN bus data. Using AI in the torque prediction results in mean absolute percentage errors of less than 5%. The mathematical reciprocating torque correction is applied to an on-road test using a truck with a Cummins ISX12G engine. The performance of the ANN and RFR model are evaluated on a Cummins L9N engine connected to a dynamometer and the data collected from the on-road testing of the Cummins ISX12G engine.

This document is currently not available here.

Share

COinS