Machine Learning and Deep Learning for Modeling and Control of Internal Combustion Engines

  • Author / Creator
    Norouzi Yengeje, Armin
  • Internal Combustion Engines (ICEs) are ubiquitous; they power a wide range of systems. The broad use of ICEs globally causes more than 20% of the total greenhouse gas emissions. In many countries, emission legislation is transitioning from certification using only traditional chassis dynomometer testing to now requiring the inclusion of Real Driving Emissions (RDE). Complying with this legislation has led to increased challenges to meet emissions levels under on-road use of the engine. The stringent legislation governing emissions and fuel economy, in combination with the complexity of the combustion process, have led to requirements for significantly more advanced engine controllers than are currently used. Reducing the emissions of diesel engines while simultaneously increasing their thermal efficiency through online control optimization and Machine Learning (ML) are the main objectives of this thesis. ML techniques offer powerful solutions that help to address the existing challenges in ICE modeling, control, and optimization. ML can also help to reduce the time, cost, and effort required for ICE calibration for both vehicular and stationary applications.

    In this thesis, a four-cylinder medium-duty Cummins diesel engine and emission measurement system including an electrochemical fast Nitrogen Oxides (NOx) sensor, Pegasor Particle Sensor (PPS-M), and MKS Fourier-Transform Infrared Spectroscopy (FTIR) are used for experimental implementation. A dSPACE MicroAutoBox II, which is a rapid prototyping system, is used for control implementation. In order to compare the proposed control method with the existing Cummins calibrated engine control unit (ECU), all the production calibration tables are imported to the MicroAutoBox. The simulation results presented in this thesis are developed using a detailed physics-based model using the GT-power\(^{\copyright}\) software. A co-simulation of GT-power\(^{\copyright}\)/Matlab\(^{\copyright}\)/Simulink is used as an Engine Simulation Model (ESM).

    The application of ML in engine control can be divided into three main categories: i) ML in emission prediction, ii) Integration of ML and Model Predictive Control (MPC), and iii) ML in the learning-based controller.

    In the first category, a correlation-based order reduction algorithm is developed to model \nox, resulting in a simple and accurate model. This algorithm utilizes Support Vector Machine (SVM) techniques to predict \nox~emission with high accuracy. In addition, a comprehensive study involving eight ML methods and five feature sets is done for Particulate Matter (PM) modeling using gray-box techniques. Then using the K-means clustering algorithm, a systematic way to select the best method for a specific application is proposed.

    In the second category, two methods of combining ML and MPC were used: ML-based modeling and ML imitation control. First, ML is used to identify a model for implementation in MPC optimization problems. Additionally, ML can be used to replace MPC, where the ML controller learns the optimal control action by imitating the behavior of the MPC. Using the ESM to provide simulation data, SVM and deep recurrent neural networks, including long-short-term memory (LSTM) layers, are used to develop engine performance and emission models. Then based on these models, MPC is designed and compared to both a linear controller and the Cummins' calibrated ECU model in ESM. Then, a deep learning scheme is deployed to imitate the behavior of the developed controllers. These imitative controllers behave similarly to the online optimization of MPC but require significantly lower computational time. The LSTM-based MPC is then implemented on the real-time system using open-source software. Compared to the stock Cummins ECU, this controller has significant emission reduction, fuel economy improvement, and thermal efficiency.

    Reinforcement Learning (RL) and Iterative Learning controller (ILC) are developed to investigate learning-based controllers. Using the ESM, a model-free off-policy algorithm, Deep Deterministic Policy Gradient (DDPG), is developed. A safety filter is added to the deep RL to avoid damage to the engine. This filter guarantees output and input constraints for both RL and ILC. The developed safe RL is then compared with ILC and LSTM-NMPC.

  • Subjects / Keywords
  • Graduation date
    Fall 2022
  • Type of Item
  • Degree
    Doctor of Philosophy
  • DOI
  • License
    This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.