Using Feature Alignment Learning to Improve the Output Repeatability of Classification Neural Network

Deep learning is extensively applied in industrial defect detection, with deep neural networks classifying product images for quality assessment. However, disturbances in captured images due to factors such as assembly tolerances and vibrations can compromise the consistency of network outputs, affecting Attribute Reproducibility and Repeatability (AR&R) evaluations. This paper proposes a feature alignment method for neural networks to improve output repeatability, which has shown significant enhancements in classification consistency and accuracy on an inductance defect dataset.