Research On Inspection System And Recognition Algorithm For Data Center

With the development of digitalization,data centers and communication rooms with a large number of server equipment have become an important part of the construction of digital information systems,and are increasingly important in the process of promoting digital construction.Currently,most domestic data rooms mainly rely on manual on-site visual methods to complete inspections.However,as the service scope of data centers expands,the difficulty of manual inspections is gradually increasing.To reduce costs and increase efficiency,mobile robot-based data center inspection systems have emerged.At present,the image recognition accuracy of the inspection system is affected to varying degrees in two stages.In the instrument positioning stage,traditional computer vision mainly uses methods such as template matching and Hough transform.Such methods have problems such as environmental noise,single template,and weak generalization ability.In the dial content recognition stage,recognition methods based on deep learning face difficulties due to the loss of target feature pixels caused by continuous downsampling.On the basis of clarifying the respective advantages and disadvantages of traditional computer vision algorithms and deep learning algorithms,this article combines the technical advantages of the two to apply and improve the algorithms according to the actual recognition targets and their requirements for recognition accuracy,in order to achieve the optimization of recognition accuracy.The main research contents of this article are summarized as follows:By analyzing the functional indicators corresponding to the inspection environment and functional requirements of the data center,an intelligent data center inspection system framework is built and the core components of the system are selected.At the same time,the mechanical structure of the robot body equipment is designed and a prototype is prepared to provide hardware support for subsequent image acquisition and algorithm application.Instrument recognition in the data center inspection process mainly focuses on the positioning and recognition of pointer-type and digital display-type instruments.In the instrument positioning stage,the YOLOv7-tiny model is introduced,and the mainstream Faster R-CNN model is improved.A data set for data center inspection is constructed through actual sampling.Based on this,the improved Faster R-CNN model accuracy is analyzed.Improving the pooling strategy and adopting a deeper Res Net152 residual network is advantageous for feature extraction,significantly improving positioning accuracy.The performance of the YOLO series model in the instrument positioning experiment is compared based on actual data,including m AP,P,Recall,and parameter quantity indexes.The strong learning ability of the YOLOv7-tiny model in instrument positioning is verified,providing a new method for instrument positioning.Through comprehensive evaluation,it can be concluded that the improved Faster R-CNN model can be selected when there is sufficient computing power and higher accuracy is needed.The YOLOv7-tiny model is more suitable when accuracy and algorithm deployment costs need to be balanced.In the digit recognition stage,for digital display-type instruments,character segmentation based on image features is carried out first.Then,a digital recognition model based on support vector machines is introduced,and the optimal parameters of the classifier are obtained through the gray wolf optimization algorithm.Finally,the GWO-SVM model proposed in this paper is verified through the digital tube data set,with a recognition accuracy of up to 98.1%.For pointer-type instrument readings,the affine transformation is applied to the instrument image through its rectangular features,and the application of Hough transform in pointer detection is analyzed.The reading is obtained by calculating the angle,which has been proven to be efficient and accurate through experiments.A recognition approach based on computer vision is proposed for identifying indicator light images that cannot be accurately recognized by deep learning models.Firstly,the image is denoised by grayscale and smoothing processing.Secondly,a detection algorithm based on adaptive Canny edge detection and Hough circle transform is used to identify the indicator light.Finally,the image is transformed into the HSV color space to effectively obtain the color feature information of the indicator light.Through comparative experiments,it is verified that the proposed recognition approach has superiority in identifying indicator lights,with a recognition rate of up to 93.9%,providing a new solution for small target detection and recognition tasks.Finally,the described inspection system has been successfully deployed in a certain rail transit signal company’s machine room,efficiently and accurately completing the tasks of instrument detection and digital display recognition,making beneficial attempts to achieve unmanned operation of machine rooms.