Existing knowledge distillation methods for anomaly detection face architectural constraints and limited generalization. This paper proposes Denoising Heterogeneous Knowledge Distillation (DHKD), featuring a heterogeneous Teacher−Student (T−S) framework where the teacher employs flow-based probabilistic density estimation while the student uses feedforward architecture. This design intentionally amplifies representation discrepancies, preventing complete capability replication. DHKD reduces pre-trained model dependency through generative architecture. Gaussian noise injection at feature level simulates anomalous perturbations, enhancing normal sample reconstruction. A Dual-Domain Reconstruction Module (DDRM) filters anomalies and strengthens normal data discrimination. T−S representation affinity loss integrates local features with global semantic context. By mining latent patterns in anomalous data, the method automatically discovers complex anomaly representations, improving implicit pattern recognition. Experimental results demonstrate state-of-the-art performance on MVTec AD, achieving an Area Under the Receiver Operating Characteristic curve (AUROC) of 99.3% for detection and 98.6% for localization, with superior generalization across real-world datasets.