Diagnostic automatisé en environnement de production à faible volume et haute complexité

Abstract

Commercial pressures having resulted in the transfer of most high-volume printed circuit board assembly operations to Asia, remaining North-American manufacturers have had little choice but to reposition themselves in the more difficult market segment of lower volume, higher complexity products, also called Low-Volume, High-Mix (LVHM). These products, commonly targeted to the industrial, medical or military sectors, integrate expensive components and their high unit costs justify substantial diagnosis and repair efforts when defects are detected in production. Although automated diagnostics is desirable, the low production volumes impose severe limits on available data and make conventional machine learning techniques impractical. In this thesis, we propose a novel approach based on knowledge modeling and case-based reasoning for automated diagnosis of printed circuit boards in an LVHM production environment. Our hybrid approach, whose effectiveness we have demonstrated using a prototype we developed, can overcome the knowledge-acquisition bottleneck even though it is targeting a data-poor environment. The proposed approach does not require a contribution from product designer or experts and is designed to operate using information available during manufacturing only. Our approach is based on using a reasoning system to accumulate experience coupled with a repository of domain and product knowledge. In order to accelerate the acquisition of experience on a new product, we generate synthetic cases using a boundary-scan level board emulator connected to the same test equipment used on the real board in production. Our test results demonstrate that these synthetic cases allow our diagnostic system to detect, locate and classify all single faults and multiple faults affecting up to three neighboring nodes with a better success rate than the reference commercial tool. Moreover, case base data, including board layout information and user feedback from previous repairs, are used to feed a recommender system tasked with performing root-cause analysis to identify faulty components and suggest repairs. A production and repair data simulator is described and integrated with our prototype system in order to verify functionality of the recommender system and evaluate its effectiveness. Results are promising and allow us to consider using the proposed system in a commercial environment.

Date	16 Sept 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Claude Thibeault (Supervisor)