Knowledge-based diagnosis for automotive body assembly: Methodology and implementation

This thesis presents a knowledge-based diagnosis method developed and implemented for automotive body assembly. This method enables quick detection, localization, and isolation of faults in the assembly process based on in-line dimensional measurements. The formulation of the fault and knowledge domains for diagnosis was based on the a priori investigation of the most severe faults occurring during an 18-month product development cycle. Based on this study, the diagnostic method focused on the dominant faults related to the assembly fixtures.;The developed diagnostic method includes an automotive body assembly knowledge representation and a diagnostic reasoning mechanism. The knowledge representation is based on the functional characteristics of the product, tooling, process, and measurement, which are in turn defined as collections of hierarchical groups. The diagnostic reasoning includes four major tasks: (a) fault detection, (b) determination of the Candidate Component (part affected by failure), (c) determination of the Candidate Station (assembly station causing the failure), and (d) fault isolation.;The first three tasks are driven by the variation level of the measurement points, selecting for the most severe fault first. The fault isolation method exploits the root causes of the dimensional faults related to the tooling elements. This is a novel approach that allows effective fault isolation by integrating in-line dimensional measurements, advanced statistics and pattern recognition with product and fixture design. It presents a variation pattern model for each hypothetical fault derived, based on the CAD data for the fixture. The model for a single fault is unique, deriving analytically proved relations between fixture geometry, dimensional variation and eigenvalue-eigenvector pair. The fault is isolated by mapping the pre-determined model of the fault with the pattern of measurement data represented by eigenvalue-eigenvector pair.;A major advantage of the proposed knowledge representation and diagnostic reasoning is that no pre-solved cases are necessary to create the knowledge base. This allows prompt localization of the candidate component and station even in the absence of a particular fault history, which makes this approach a viable strategy during development stages such as launch or pre-production.;The diagnosability of the developed method is analyzed with respect to robustness to noise in a real industrial environment. It was derived from the general class of covariance matrices describing tooling faults. The presented analysis estimates that the diagnostic method gives correct results in an industrial environment with up to a 173% level of additive uncorrelated noise.;The developed method was implemented in the Expert System shell Nexpert Object produced by Neuron Data Inc. The implementation is divided into knowledge reasoning (rule network) and fact representation (object network) using Nexpert's ability to create a hybrid representation of the knowledge (rules and objects can be merged). The developed and implemented approach was evaluated using production data. The evaluation indicates that 83.3% of the faults were successfully localized.