Server Monitoring System Based On Convolutional Neural Network

With the advent of the digital era and the rapid development of communication technologies,more and more data centers are used for data transmission,and the monitoring of the working status of the server is one of the focuses of the data center.Common monitoring methods mainly include manual monitoring and network monitoring,which have the problems of low efficiency and high cost of network formation.The image recognition monitoring method based on convolutional neural network(CNN)can effectively solve the above problems,which does not need complex network formation.But it also has the following disadvantages: The precision of the detection of the servers and the working status of the indicator LEDs are low.CNN is affected by class imbalance during the training process,which leads to the degradation of the model.The structure of CNN is so complex that it is difficult to apply to embedded systems.The cost of training data collection for CNN is high.In this paper,we optimize the above issues from four aspects.Firstly,we design a server monitoring system based on CNN to detect the servers and indicator LEDs,build a server data set containing about 5,000 images,and the precision of the detection of the servers is 99.9%,its frame per second(fps)on the Raspberry Pi embedded platform reaches 120,which can output the working status of the servers in real time.Secondly,we propose a loss function and parameter update algorithm to solve the class imbalance problem by automatically adjusting the weights of hard examples and easy examples,whose recognition accuracy on the VOC2007 dataset exceeds the advanced algorithms such as YOLOv2,SSD,and Faster RCNN.Then,we propose the Light-YOLO network architecture to optimize the number of convolution kernels and channels in CNN.Compared to Tiny-YOLO,the speed is increased by 51% and the model size is reduced by 88%.Finally,we design a novel multi-scale randomized edge image synthesis method and introduce a randomized edge information fusion algorithm to make the edge information of synthetic image more realistic.We construct a multi-scale discrimination network to make synthetic images have a wider global view and better detailed information by using different scales of features.This method can effectively synthesize training pictures and save the cost of data collection,and its pixel accuracy on the Cityscapes dataset is much better than pix2 pix.The image recognition monitoring method based on CNN proposed in this paper effectively improves the detection accuracy,eases the class imbalance,compresses the network model and reduces the cost of image collection.As a supplement to the manual monitoring method,it has great application prospects in the field of server monitoring.