Due to the increasing complexity of technical systems, efficient and effective fault diagnosis is crucial in order to reduce service costs and system downtime.
Particularly in domains where maintenance poses an extensive part of the entire operation cost, accurate identification of failure sources has a large economic impact.
Model-based diagnosis, as a subfield of Artificial Intelligence, derives explanations for observed anomalies by relying on a formal representation of the artifact under consideration. Two notions of model-based diagnosis exist; first, the consistency-based variation determining root causes through discrepancies between the predicted and actual behavior and second, the abductive form which is founded in logic-based abduction.
Although the theoretical background was established decades ago and various research prototypes have been implemented, industrial applications are sparse. The barriers separating science from practice include among others the initial modeling effort, the high computational complexity associated with the fault identification task, and the difficulties of consolidating model-based diagnosis tools and existing software.
To bridge the gap between academic research and industrial applications, this thesis proposes a methodology for applying abductive model-based diagnosis to an industrial setting.
We define a process that positions the diagnostic task in the overall operational life of a technical system and enables an easy integration into practice by splitting model generation, fault detection, and fault identification into separate modules. To reduce or even eliminate the initial modeling effort, we automatically construct suitable diagnosis models from failure assessments available in practice.
Within the troubleshooting portion of the process, we show a general technique for combining diagnosis, solution ranking, and the selection of probing points.
An essential aspect when transporting research into practical applications is to ensure the calculation of solutions is done in a reasonable time frame. Hence, we investigate computational approaches to diagnostic problem-solving which allow us to extract explanations efficiently. Thus, based on the characteristics of the models generated on top of the failure records, we explore different notions of abductive reasoning such as conflict-driven or direct procedures. Our experimental evaluations show that for simple bipartite abduction problems deriving explanations via hitting set computation is favorable. Yet, for more expressive representations we could not determine that either conflict-directed or direct methods are superior.
The practicality of the proposed framework for incorporating abductive model-based diagnosis into real-world domains is evaluated in an industrial use case on wind turbines. Turbine maintenance costs are high and idle times lead to substantial revenue loss making exact and fast fault identification a highly relevant pursuit. Therefore, we apply the theoretically formalized diagnosis process by exploiting automatically retrieved turbine health variables and expert knowledge of failures and their effects. To consolidate the fault identification engine with the existing monitoring software and current work processes, we present a user interface and interaction design as well as a workflow that should foster user acceptance of the diagnosis application. The evaluations up to this point have shown that the diagnosis approach is appropriate and permits practical usage.