Abstract: Real-time monitoring of conveyor belts is a crucial aspect of safety in industrial production settings. Addressing the issues of insufficient existing conveyor belt defect data and the inadequate per-formance of defect detection models, this study proposes an unsupervised belt defect detection model based on an improved Generative Adversarial Network (GAN). The model's training relies solely on easily accessible images of normal belts, effectively circumventing the challenges in obtaining and labeling defect data for belts. To tailor the GAN for belt defect detection, several enhancements were implemented: the generator was modified into an end-to-end Autoencoder (AE) structure to address misalignment between reconstructed and input images; to enhance the gener-ator's ability to reconstruct and repair unknown irregular defects and alleviate identity mapping issues, we introduced irregular simulated defects using Perlin noise into the training images, and the generator loss function was refined into a weighted sum of multiple loss functions. Furthermore, to bolster the model's capability in detecting minute defects, a pre-trained block discriminator replaced the original discriminator structure. The anomaly score is determined by the highest block anomaly score and the difference in the latent space of the discriminator between the original and recon-structed images. Experimental results indicate that the proposed model achieved an ROC-AUC of 0.999 and PR-AUC of 0.997, with a detection time of 13.51ms per image. Compared to other methods such as Structural Similarity Index Measure – AutoEncoder（SSIM-AE），Fast-Anomaly GAN (F-AnoGAN), GANomaly, and Omni-frequency Channel-selection Reconstruction GAN (OCR-GAN), our model shows a 4.5% increase in ROC-AUC and an 8% increase in PR-AUC. It satisfies the real-time detection requirement for conveyor belts moving at 3 m/s. Furthermore, the model demonstrates a more distinct distribution difference in anomaly scores between positive and negative examples, offering robustness and accuracy in practical applications.