Abstract: Aiming to address the dual-challenges of complex pattern interference and real-time demand in color-coated sheet surface defect detection, this study proposes an unsupervised color-coated sheet defect detection method based on cross-layer attention fuzzy. By constructing a dynamic fuzzy adjustment mechanism, the fuzzy parameters are adaptively optimized by combining the local gradient information of the image, and the edge features of the defects are retained while the background noise is suppressed. A cross-layer attentional feature enhancement network is designed, which fuses multi-scale features and adaptively weights them, and significantly improves the sensitivity to tiny defects and general defects. And an adversarial training framework is introduced, which realizes the alignment of unsupervised feature distributions, and reduces the dependence on the labeled data. The experimental results show that the method achieves 98.21% accuracy in color-coated sheet surface defect detection, the false alarm rate is reduced to 1.2%, the 4 K image processing latency is 32 ms, the recall rate of micro-defects (less than 0.2 mm) reaches 96.12%, and the recall rate of general defects (0.2 mm and above) reaches 98.32%, and the robustness reaches 95.67% under the complex texture scenario.In the drawing, spotting, abrasive defect detection tasks, the comprehensive performance is significantly better than the comparison models, in which the accuracy of drawing defect detection reaches 96.55%, speckle defect reaches 95.89%, and abrasive defect reaches 97.21%. The method solves the bottleneck of traditional algorithms in texture-defect decoupling and real-time balance through the synergistic effect of dynamic fuzzy and cross-layer attention mechanisms, and provides a feasible solution for high-precision real-time detection in industrial scenarios.