Date of Completion

11-23-2015

Embargo Period

11-23-2015

Major Advisor

Krishna R. Pattipati

Associate Advisor

Yaakov Bar-Shalom

Associate Advisor

Peter Willett

Field of Study

Electrical Engineering

Degree

Doctor of Philosophy

Open Access

Campus Access

Abstract

Rapid advances in electronics, control, communication and computing technologies have resulted in complex network-embedded automotive systems. A major goal for safe and reliable operation of complex systems is to develop robust analytical tools/algorithms, based on an information ensemble generated from integrated sensor-based (also referred to as data-driven), model-based and knowledge-based approaches, which enable early diagnosis and prognosis of incipient faults. Fault diagnosis is the process of fault localization (root-cause isolation) and severity estimation (identification); whereas fault prognosis is the process of early diagnosis and remaining useful life estimation.

In this dissertation, we develop fault detection, diagnosis, and prognosis algorithms to detect and isolate faults in various automotive subsystems. In the first application, a systematic model-based and data-driven diagnostic process is developed to detect and diagnose faults in a regenerative braking system in hybrid electric vehicles. The results demonstrate that highly accurate fault diagnosis is possible with state-of-the-art classification techniques.

In the second application, we develop a data-driven framework for fault diagnosis by analyzing the vehicle health data acquired from test fleet and production vehicles via dealer diagnostics. The framework features data pre-processing, data visualization, clustering, classification, and fusion techniques and is applied to real-world failure datasets. The results demonstrate various ways to achieve good accuracy in detecting/isolating faults.

In the third application, we develop incremental learning classification methodologies to classify faults based on evolving databases. The performance of adaptive (or incremental learning) classification techniques is discussed when (i) the new data has the same fault classes and same features, and (ii) when the new data has new fault classes, but with the same set of observed features. The proposed methodology is demonstrated on datasets derived from an automotive electronic throttle control subsystem.

Finally, in the fourth application, a unified prognostic framework is developed to estimate the component degradations. The framework employs Cox proportional hazards model and soft dynamic multiple fault diagnosis algorithm for inferring the degraded state trajectories of components. The capabilities of the framework are demonstrated on two different automotive subsystems. The proposed framework is modular, leading to flexible and evolvable software architecture for prognostics and health management.

COinS