Dynamic Integration of Probabilistic Information for Diagnostics and Decisions
The automobile has advanced from a basic means of transportation to a rolling platform that hosts safety, performance, emissions control, and entertainment systems. Due to this increase in complexity, more advanced diagnostic techniques are required. This thesis presents Multi-Modal Diagnostics (MMD), which is a probabilistic approach for diagnosing vehicles and other complex systems. MMD combines model-based diagnostics, Bayesian networks, and statistical decision analysis into a unified probabilistic framework. This thesis introduces the framework and analyzes the temporal characteristics of its components in order to understand its performance.
Multi-Modal Diagnostics is a model-based diagnostic technique and is designed to reduce costs by using both existing sensors and system models. MMD differs from other model-based diagnostic techniques in that it uses information from multiple sources and models to analyze dissimilar observable modes of a system. Operating status information and model-generated residuals are interpreted using a Bayesian network, which models the temporal characteristics of the faults and determines realtime fault probabilities. In order to solve the network efficiently, it is necessary to assume the residuals are not autocorrelated, an assumption that rarely holds. This thesis analyzes the sources and consequences of this autocorrelation and discusses methods for removing it, including whitening and downsampling. It is shown that the best method depends on the application.
The product of the combined residual generating and Bayesian network system is a continuous stream of joint fault probabilities. A method is needed to evaluate these probabilistic estimates, and existing static techniques for rating probabilistic forecasters are reviewed. It is shown that the temporal characteristics of the estimates are important. In response, this thesis introduces a new, dynamic method for scoring probabilistic estimators, which rewards estimates with desirable temporal properties.
The probabilistic treatment of diagnostics is completed by addressing how decisions can be made based on fault probabilities. Since many of the available actions are irreversible, the dynamics of these decisions must be considered. This thesis introduces a new decision technique that includes a model of these temporal relations. The method draws on decision analysis and probabilistic risk assessment theory and yields a statistically optimal on-board decision policy.