A Zero-Shot Object Detection Method for Network Operation and Maintenance Using Cross-Modal Feature Alignment

With the continuous expansion of network scale and increasing complexity, the operation and maintenance (O&M) of future internet and networking infrastructures faces significant challenges. The rapid and accurate detection and localization of critical objects (e.g., device failures and potential risk points) is essential to enhance operational efficiency and mitigate fault risks. Object detection provides significant value in network O&M scenarios including device status monitoring and anomaly behavior identification. However, traditional methods depend heavily on large amounts of high-quality annotated data, rendering them unsuitable for O&M scenarios characterized by diverse network equipment classes and high annotation costs. In this paper, we propose a zero-shot object detection method based on Cross-Modal Feature Alignment (CMFA). This method innovatively integrates multi-modal pre-trained models into network O&M systems, enabling efficient unsupervised object detection. It eliminates the need for annotated data while adapting to diverse network component categories. CMFA leverages large vision-language models to generate object detection prompts, enabling effective cross-modal feature alignment between textual descriptors and visual patterns. Furthermore, it incorporates an ensemble learning-based post-processing algorithm to model intra-class variations of network components, thereby enhancing model generalization. Experimental results demonstrate the significant advantages of the proposed method in zero-shot object detection, both on public datasets and in real-world O&M scenarios. The method has been successfully deployed in China Telecom’s network O&M infrastructure.