Integrating explainable artificial intelligence with vision-based surface defect detection systems for realizing zero-defect manufacturing

The successful implementation of Zero-Defect Manufacturing (ZDM) necessitates the development of robust, reliable, and accurate inspection systems for detecting defective components, minimising waste, and achieving sustainability. Vision-based systems employing Convolutional Neural Networks (CNNs) have shown excellent potential for 100% in-line surface inspection, supporting ZDM goals. However, the complex black-box architecture, coupled with a lack of explainability in system predictions, hinders the realisation of ZDM goals. Vision system developers struggle to select an appropriate network architecture for a given application, while end-users lack confidence in the trustworthiness of inspection outcomes. The present work proposes an assistive system utilising Explainable Artificial Intelligence (XAI) to enable the broader adoption of CNN-powered inspection. The XAI-based assistant provides classification scores, heatmaps, and bounding boxes as additional indices alongside system predictions, supporting informed decision-making by models. The additional information provided by integrating an explainable interface with the vision system increases end-user trust. A case study on in-line tapered roller inspection provides decision support for model developers and end-users using the assistant, enhancing broader acceptance of CNN-powered surface defect detection systems. The outcomes highlight the potential of an XAI-integrated vision-based system in realising ZDM goals, paving the way toward a zero-defect and zero-waste future.