Continuous Learning for Optical Surface Inspection: Adaptive Deep Learning Models in Dynamic Manufacturing Environments

Abstract

Deep learning models deployed for optical surface inspection in real manufacturing environments face a challenge that is largely absent from research settings: the data distribution is not stationary. New product variants are introduced, manufacturing processes are refined, defect patterns evolve, and measurement instruments are upgraded—all changes that can cause a deployed model to become progressively less accurate without periodic retraining. This phenomenon, known as model degradation or concept drift, can silently erode inspection quality, leading to increased escapes (false acceptances of defective parts) or increased false rejects (unnecessary rework). This study proposes a continuous learning framework for optical surface inspection that enables deployed models to adapt to evolving data distributions without requiring the massive retraining datasets or computational resources of full model retraining. Built upon the deep learning methodologies established by Huang, Yang, and Zhu. (2023) in 4D thermal imaging and the optical metrology innovations of Huang, Tang, Liu, and Huang (2026), the framework combines streaming anomaly detection to identify distribution shifts, an elastic weight consolidation mechanism to preserve learned knowledge while adapting to new patterns, and a memory-replay training strategy that prevents catastrophic forgetting. Evaluated on a longitudinal production dataset spanning 18 months of real manufacturing data from a precision optical component line, the proposed framework maintains inspection accuracy within 3.1% of a fully retrained model while requiring only 0.4% of the retraining computational cost and no interruption to production. The framework provides a practical pathway toward self-sustaining, long-term deployment of deep learning optical inspection systems in dynamic manufacturing environments.