This paper proposes a novel image detection method for identifying small hardware defects in power transmission line systems using zero-shot learning. The method aims to improve detection accuracy, especially when limited sample data is available. It employs a generative adversarial network (GAN) to create synthetic datasets for transmission line components and defects, addressing the challenge of insufficient labeled data. Additionally, the defect recognition algorithm is optimized, pruned, and compressed for efficient deployment at the front end. Zero-shot transfer learning techniques are applied for fine-grained target recognition, facilitating the development of a defect recognition tool. Experimental results demonstrate that the proposed method outperforms traditional detection techniques, achieving higher accuracy and efficiency. The approach shows considerable promise for defect detection in power transmission systems and has the potential for broad application in other critical infrastructure domains, offering an effective solution for proactive maintenance and system reliability.