Server monitoring system using an improved Faster RCNN approach

The Data Center contains lots of servers whose indicator LEDs can provide the fault information which is important for information security. In order to monitor the working status of the server in real time, a novel server recognition scheme combined with deep learning and recognition computing was proposed. In this method, the state-of-the-art Faster RCNN framework was improved by appropriate anchors selection, hard negative mining and non-maximum suppression. Morphological operations were used to strengthen the robustness of the traditional LEDs detection algorithms. For Resnet model, our system achieved a frame rate of 14 fps and object accuracy of 96% on a NVIDIA Titan X. The proposed scheme obtained excellent detection performance in real conditions, making it much more accurate and efficient to monitor the fault information of the servers.