A Self-Evolving Transformer-Based Machine Vision Framework for Adaptive Industrial Defect Diagnosis Under Non-Stationary Environments

The proposed study suggested a machine vision system that employed a transformer to detect defects in a manufacturing environment, based on a self-evolving concept. One of the primary issues that the paradigm succeeded in resolving was the fact that traditional machine learning methods do not function in practical scenarios of production settings in which conditions are constantly being altered. The framework was also integrated into architecture and included a vision transformer with adaptive memory, distribution shift detection, meta-training, self-evolution, and quantification of uncertainties. This was effective in manufacturing large-scale testing on steel production, semiconductor production, and textile inspection, achieving 95.5% and 8.9 percentage points lower than the current methods. It was also robust in terms of the performance of the system, and had an accuracy of over 93.6% even when the environmental factors were perturbed. The framework also had the significant advantages of being able to accommodate a new set of defects (94.7% accuracy using only 10 samples per new line defect labelled in the lab) and of requiring less annotation (reducing it by 95% and saving about 95000 dollars annually). When episodic memory was employed, almost catastrophic forgetting occurred, and the elastic application of weight consolidation (3.1% average) facilitated actual lifelong learning in the system. Selves developed 95.8% and 89.4% accuracy with high confidence without the use of manual labelling, among continual updates. As demonstrated in a real-life application in steel manufacturing, the detection rate was 98.7% with reduced false alarms of 52%. Savings of 127,000 per quarter were realized and yet no downtime had to be set aside to debug the system.

1.                Baysal, E., & Bayılmış, C. (2025). Overcoming class imbalance in incremental learning using an elastic weight consolidation-assisted common encoder approach. Mathematics, 13(11), Article 1887. https://doi.org/10.3390/math13111887 2.                Bhatnagar, P., Arora, T., & Chaujar, R. (2022). Semiconductor wafer map defect classification using transfer learning. In Proceedings of the IEEE Delhi Section Conference (DELCON) (pp. 1–4). IEEE. https://doi.org/10.1109/DELCON54057.2022.9753436 3.                Borde, S. (2023). Mitigating catastrophic forgetting in continual learning-based image classification. In Proceedings of the IEEE International Conference on Service Operations and Logistics, and Informatics (SOLI) (pp. 1–16). IEEE. https://doi.org/10.1109/SOLI60636.2023.10425549 4.                Chen, Y., Chen, C. P., Han, B., & Yang, Y. (2025). Enhancement in three-dimensional depth with bionic image processing. Computers, 14(8), Article 340. https://doi.org/10.3390/computers14080340 5.                Chen, Y., et al. (2021). Surface defect detection methods for industrial products: A review. Applied Sciences, 11(16), Article 7657. https://doi.org/10.3390/APP11167657 6.                Cheng, M., Wang, H., & Long, Y. (2022). Meta-learning-based incremental few-shot object detection. IEEE Transactions on Circuits and Systems for Video Technology, 32(4), 2158–2169. https://doi.org/10.1109/TCSVT.2021.3088545 7.                Cheng, S. B., et al. (2023). Machine learning with data assimilation and uncertainty quantification for dynamical systems: A review. IEEE/CAA Journal of Automatica Sinica, 10(6), 1361–1387. https://doi.org/10.1109/JAS.2023.123537 8.                Chien, J.-C., Wu, M.-T., & Lee, J.-D. (2020). Inspection and classification of semiconductor wafer surface defects using CNN deep learning networks. Applied Sciences, 10(15), Article 5340. https://doi.org/10.3390/app10155340 9.                Contreras Ortiz, A., Santiago, R. R., Hernandez, D. E., & Lopez-Montiel, M. (2025). Multiclass evaluation of vision transformers for industrial welding defect detection. Mathematical and Computational Applications, 30(2), Article 24. https://doi.org/10.3390/mca30020024 10.             Duan, Y., et al. (2024). Learning to diagnose: Meta-learning for efficient adaptation in few-shot AIOps scenarios. Electronics, 13(11), Article 2102. https://doi.org/10.3390/electronics13112102