Webinar: Diagnosis and Repair Optimization with Bayesian Networks
Context
In this webinar, we construct a Bayesian network model of a known technical system for troubleshooting purposes. As a case study, we explore a domain that many of us have presumably experienced at some point, namely an engine that won't start. More specifically, we look at this problem from the perspective of a trained technician, who is entirely familiar with all aspects of the engine in question, its functionality, and its commonly-observed defects.
While we limit this webinar to studying a highly-simplified problem domain, we should note that, in practice, Bayesia's technologies are used to manage much more complex systems, ranging from airplanes to nuclear power plants.
Diagnostic Decision Support Given Perfect Domain Knowledge?
So, why do we need a diagnosis system if our knowledge of the problem domain is practically perfect? Our objective is to devise a methodology that continuously optimizes the sequence of diagnostic steps given observations. The goal is to reach a technical diagnosis in the most direct way and at the lowest cost. Our example illustrates that even a superlative familiarity with the underlying domain does not necessarily lead to an ideal diagnostic workflow. Why not? Finding the optimal sequence of diagnostic steps is not a trivial calculation. Rather, it requires a complete representation of the failure probabilities of all components and their respective interactions, which then needs to be updated repeatedly given new evidence.
Bayesian Networks for Knowledge Representation
To make this task tractable, we encode the functionality of an engine and its various subsystems in a Bayesian network. All such knowledge comes from experts, i.e., engineers and technicians. Additionally, we assign failure probabilities to each part. These quantities can be derived from data, e.g., historical warranty claims, or may be based on technical specifications. In the absence of either source, subjective estimates can serve as a substitute. In the same way, we specify values for diagnostic and repair costs.
Once the network is complete, we use BayesiaLab's Adaptive Questionnaire to simulate various fault scenarios and their resolution. Furthermore, we publish the Adaptive Questionnaire via the BayesiaLab WebSimulator.
De-Escalating Repair Efforts
While the diagnostic efficiency gain can be significant, another potential benefit emerges from this proposed approach. Certain diagnostic tasks can potentially be delegated to untrained end-users, thus reducing cost and focusing technicians on more involved tasks.
In this context, the U.S. Army implemented a highly-successful program for combat equipment maintenance (see Bayesian Networks for Combat Equipment Diagnostics).
Worst-Case Scenarios in Critical Environments
While a car that won't start is certainly a nuisance, such a situation rarely has catastrophic consequences. If we shift to the aviation domain, however, safety and reliability become much more critical. Here, we must not only diagnose actual defects but rather gain an in-depth understanding of the faults that could potentially occur.
A Bayesian network allows us to calculate the probabilities of hypothetical chains of events that could ultimately lead to a catastrophe. This includes studying scenarios in which a minor defect prevents the detection of a much more serious issue, such as a defective fuel gauge that hides the fuel loss caused by a leak. We use the given engine model to identify and quantify such extreme conditions.