Online defect identification for hot-rolled strip steel requires the rapid recognition of high-cost defects under strict production takt-time constraints, so as to reduce losses caused by rework and missed quality inspection. However, in practical deployment, this task remains challenged by the coupled effects of class imbalance, high inter-class texture similarity, asymmetric misclassification costs, and cross-line or cross-sensor domain shifts. To address these issues, this study proposes a domain-knowledge-guided sparse generative key-focus framework, termed DefectDiverseDiffusion Net ( \({\textrm{D}}^{3}\) Net). Specifically, SparDiff, a Sparse-Aware Diffusion Generator, performs class-conditional modeling in the discrete latent space of a vector-quantized variational autoencoder (VQ-VAE) and incorporates a sparsity loss via differentiable clean-latent estimation to enhance the structural fidelity, intra-class diversity, and inter-class separability of minority-class samples. DISDFormer, a Discrepancy Enhancement with Domain Knowledge Guidance, incorporates misclassification costs and inter-class similarity into cost-sensitive weights, enabling the network to suppress high-cost, highly similar confusions and optimize decision boundaries. Evaluations on the NEU-CLS and X-SDD datasets, including cross-line and cross-sensor generalization, confirm the effectiveness of \({\textrm{D}}^{3}\) Net, which achieves robust gains on minority and confounding classes. With 32.14M parameters and 6.48 GFLOPs, \({\textrm{D}}^{3}\) Net delivers 83.3 FPS real-time inference, demonstrating strong potential for industrial online inspection, early warning of costly defects, and quality control decisions in hot-rolling production.
