An Explainable Active Learning Approach for Enhanced Defect Detection in Manufacturing

Artificial Intelligence (AI) can significantly support manufacturing companies in their pursuit of operational excellence, by maintaining efficiency while minimizing defects. However, the complexity of AI solutions often creates a barrier to their practical application. Transparency and user-friendliness should be prioritized to ensure that the insights generated by AI can be effectively applied in real-time decision-making. To bridge this gap and foster a collaborative environment where AI and human expertise collectively drive operational excellence, this paper suggests an AI approach that targets identifying defects in production while providing understandable insights. A semi-supervised convolutional neural network (CNNs) with attention mechanisms and Layer-wise Relevance Propagation (LRP) for explainable active learning is discussed. Predictions but also feedback from human experts are used to dynamically adjust the learning focus, ensuring a continuous improvement cycle in defect detection capabilities. The proposed approach has been tested in a use case related to the manufacturing of batteries. Preliminary results demonstrate substantial improvements in prediction accuracy and operational efficiency, offering a scalable solution for industrial applications aiming at zero defects.