<p>Ensuring the stability and safety of cloud service centers stands as a paramount task in management. However, due to the prolonged operation and recurring maintenance and repairs, safety hazards are easily introduced. Meanwhile, research on foreign object detection in cloud server centers remains a relatively unexplored area. In this study, we designed a comprehensive strategy to address the foreign object detection task in cloud server centers. Firstly, we utilized patrol robots to collect data while proposing a fuzzy learning annotation strategy to accommodate the diversity of foreign objects found in real-world scenarios. Subsequently, we introduced a method for augmenting negative samples to address the problem of inadequate network fitting. By enhancing the network’s capacity to learn from negative samples, we ultimately achieved a perfect detection accuracy rate of 100% for the negative samples. Eventually, based on Faster R-CNN, our research introduced an improved foreign object detection network for reliable and real-time extraction of foreign objects from background scenes. The network presented herein innovatively employs a separable self-attention mechanism to achieve an enhanced refinement of the ConvMixer. This enables our model to learn global features, capture long-range dependencies within a sequence, focus on salient regional features, and generate high-quality regions of interest (ROI). Ultimately, the method proposed in this study achieved a mean average precision (mAP) of 98.45% on the foreign object dataset from cloud server centers. This holds practical significance in ensuring the stable operation of cloud server centers.</p>