The proposed collaborative project with GM R&D center, Warren, MI seeks to develop hybrid model-based/data-driven/knowledge-based prognostic framework, and the associated prediction and inference algorithms, to detect and isolate incipient component degradations in coupled systems. The focus will be on chassis health determination (meaning the health status of steering, braking and suspension components, as well as functionalities that use these components such as the StabiliTrak system) to replaceable components (e.g., broken tie rod, worn out brake pads or blown tire, malfunctioning ABS control module). The diagnostic and prognostic framework will be validated on test fleet prior to implementing in the production versions.
Intellectual Merit: Existing prognostic and diagnostic algorithms employed in automotive systems tend to be component-centric. They often fail to provide correct diagnosis due to neglect of cross-subsystem failure propagation and unreliable tests. The proposed research seeks to overcome these limitations by explicitly modeling the cross-subsystem effects via a graphical model and by developing a multi-layer probabilistic reasoning process spanning sensed data → predicted features → diagnostic trouble codes → failure modes → replaceable components and subsystems. Modeling the time evolution of coupled component states as factorial hidden Markov models and the development of computationally-efficient inference algorithms in multi-layer graphical models is a novel aspect of the proposed effort. In addition, combining model-based, data-driven and knowledge-based approaches in a unified way to solve practical diagnosis and prognosis problems in the next-generation automotive vehicles is another contribution of this work.
Broader Impact: The proposed research improves the competitiveness of American automotive industry by reducing warranty costs, and enhancing vehicle availability and customer satisfaction. The PI plans to promulgate the results of this research to the broader industrial community via short courses, tutorials, conference presentations and journal manuscripts.
