Research On Operational State Detection System Of Server Based On Thermal Images

Nowadays,with the rapid development of "big data" and cloud computing technology,it has driven a huge amount of computing power and storage demand.As the cornerstone behind the scenes,the data center has high operational and automated status diagnostics requirements.When the load of the server or the external environment changes,the server may overheat and form a local hot spot which may cause an unbalanced temperature distribution of data center.It'll result in an increasement in cooling energy consumption.Based on thermal infrared image technology,this paper proposes a server operation state diagnosis system based on image processing and pattern recognition technology.An automatic diagnosis software is designed based on the algorithm structure of the diagnosis system.The goal of this paper is to use the infrared thermal images to identify the fault state of the server effectively and provide supporting for local hotspot issues.During the experiment,the system simulates five operating states that may appear in actual situations.The infrared thermal imager is used to capture the thermal images of the server outlet as the system input data.The server status diagnosis model is built based on the pattern recognition technology.In the research process,considering the actual engineering application scenarios,the diagnostic model may face the problem of insufficient sample data of server failure and low quality of thermal images.Therefore,a system optimization method based on GAN(Generative Adversarial Networks)is proposed.Finally,the diagnostic software is designed according to the algorithm architecture of the diagnostic system.Next,the composition of the server status diagnosis system is described.Firstly,the thermal images are captured by the thermal imager at the outlet of the server under the running state;then the thermal images are subjected to standardized preprocessing;then the images enhanced by homomorphic filtering are subjected to one-dimensional maximum entropy hotspot segmentation to obtain server hotspot images;then the types of feature extraction is selected according to different classifiers.The SVM(Support Vector Machine)adopts three kinds of features extracted manually: statistical features,texture features and morphological features.CNN(Convolutional Neural Network)is improved by AlexNet network model.The convolution kernel is used to extract features automatically;finally,the performance of the diagnostic model is tested by the test set samples.Experiments show that SVM and AlexNet can achieve a diagnostic accuracy of about 90%.Due to the unbalanced number of training samples and the inconspicuous difference between the samples,the bottleneck of the system diagnostic accuracy is caused.Therefore,this paper proposes a solution for the DCGAN(Deep Convolutional Generation Adversarial Network)synthesis fault status samples to improve diagnostic accuracy by using DCGAN to extend the diagnostic model training set.Through a series of comparative experiments,an optimal combination of diagnostic systems is proposed: DCGAN + AlexNet.This optimization method can increase the diagnostic accuracy to 95.11%.Based on the structure of the diagnostic system algorithm,this paper designs and develops a server diagnostic software by Qt cross-platform development environment and OpenCV computer vision open source library.The software includes three functional sub-interfaces: data reading and storage,model training and diagnosis,and diagnostic report.It will provide scientific and reasonable technical support for data center maintenance personnel in the actual engineering application scenario.