In-process vision-based quality inspection system for manufacturing of electrical machine components with hidden features

In electrical machine manufacturing, inspections are currently performed at the end-of-line tests, leading to wasted time and resources. This research aims to develop a vision-based in-process quality inspection system for the manufacturing of electrical machine components with hidden features, such as thermally crimped terminations. The purpose of the system is to infer the quality of the product during and/or after the production steps.

The first research gap (Objective 1) highlights a knowledge gap in selecting an image aug-
mentation technique to enhance the performance of object detectors, as well as the rationale behind this choice. Additionally, the focus currently lies on individual use cases rather than their transferability to other domains. This study is the first to develop a framework quantifying the influence of augmentation techniques on the performance of a one-stage object detector, generalising across dissimilar datasets, for data-scarce conditions.

The second research gap (Objective 2) demonstrates that existing quality inspection systems either rely on vision-based inspection for surface defects or infrared for hidden defects, highlighting that no single sensing technique used in isolation is sufficient. Objective two presents the use of deep learning object detectors (for the localisation and classification of visible surface features) and infrared imaging with a deep learning classifier (for predicting and classifying subsurface/hidden features) for the thermally crimped terminations. It is the first to facilitate in-process, non-destructive detection of both visible and subsurface defects during thermal crimping.

The third research gap (Objective 3) identified a lack of an end-to-end in-process inspection
framework for electrical machine manufacturing. Objective three fills this gap by develop-
ing and validating an end-to-end, in-process quality inspection framework that merges a machine-learning predictor of optimal process parameters with a real-time, bimodal sensing system for thermal crimping. This novel framework provided the first end-to-end system for in-process quality assessment for thermal crimping.