TOWARDS IMPROVING THE PERFORMANCE AND RESILIENCY OF POWERTRAIN SYSTEMS

The development of powertrain systems has played a pivotal role in advancing transportation technologies, shaping economic growth and societal progress. This thesis explores innovative methodologies aimed at enhancing powertrain efficiency and resilience through dynamic system modeling and Machine Learning (ML) applications. Three research projects are presented: (1) improving turbocharger turbine mapping using machine learning, (2) fault analysis and mitigation of transmission squawk noise, and (3) fault detection and diagnosis of two-stroke diesel engines. The findings contribute to improved predictive capabilities, reduced system inefficiencies, and enhanced fault detection strategies, pushing the boundaries of intelligent powertrain optimization

The first project focuses on improving the turbine mapping process Machine Learning (ML). The current method, polynomial regression used by the manufacturer, Cummins Turbo Technologies (CTT) is presented first. In an attempt to improve the results over the current method, various machine learning and deep learning models, including Gaussian Process Regression (GPR), Support Vector Machines (SVM), and several neural network architectures such as Levenberg-Marquardt (LMANN), Bayesian Regularization (BRANN), and Gaussian Back-Propagation (GBPNN) Artificial Neural Networks are investigated. These models are comparatively analysed, demonstrating a significant improvement in turbine performance prediction accuracy, offering a more efficient and cost-effective approach to turbine mapping. The Bayesian Regularization Neural Network (BRANN) demonstrated a 30% reduction in mean squared error (MSE) compared to traditional polynomial regression methods, significantly improving predictive capabilities and reducing reliance on extensive physical testing.

The second project focuses on fault analysis and mitigation of transmission squawk noise, a prevalent issue in automatic transmission systems. By employing a non-linear multi-degree-of-freedom (MDOF) spring-mass-damper model and a multi-disk clutch model, this study identifies the root causes of squawk noise during low-speed up-shifts in a 9-speed Automatic Transmission. Through simulations and experimental validations, effective mitigation strategies are proposed to enhance the overall transmission performance and reduce undesirable noise.

The third project addresses fault detection and diagnosis in two-stroke (2S) diesel engines using Machine Learning. A data-driven fault detection framework is implemented using neural networks trained on synthetic fault data generated on an 2S Diesel Engine model. The classifier achieved a fault detection accuracy of 88.8%, with an area under the curve (AUC) exceeding 0.95 for intake manifold pressure sensorfaults and 0.92 for () engine speed sensor faults, demonstrating the robustness of machine learning in early fault diagnosis.The results highlight the model's efficacy in accurately identifying and diagnosing faults, thereby improving the engine's reliability and operational efficiency.

Collectively, the three projects presented in this thesis align under the common umbrella of enhancing powertrain resiliency and performance. Each project, though distinct in methodology and scope, addresses a critical subsystem within modern automotive powertrains: turbine efficiency mapping, noise mitigation in transmissions, and fault detection in diesel engines. Together, they reflect a systems-level approach to powertrain improvement. Learnings from ML-driven turbine modeling inform predictive diagnostics in diesel engines, while squawk noise detection and squawk mitigation reinforce the need for real-time adaptive responses seen in neural-network-based fault detection. These cross-learnings emphasize the importance of integrating data-driven models, system-level dynamics, mechanisms to design future-ready, robust, and efficient powertrain systems.