Design And Implementation Of Artificial Intelligence Operation And Maintenance System For Micro-Service Based Cloud Platform

With the rapid development of cloud computing technology,cloud platforms are increasingly adopted by enterprises as the infrastructure of their information systems.With the advancement of microservice technology,more and more cloud platforms are adopting microservice architecture to improve the flexibility and scalability of platform functionality.However,the increasing number of components in microservice-based cloud platforms has led to an increase in the complexity of platform operation and maintenance.Therefore,introducing artificial intelligence operations(AIOps)for microservice-based cloud platforms has become a hotspot in the evolution of cloud platform technology and application expansion.This thesis designs and implements an intelligent operation and maintenance system for microservice-based cloud platforms.The system consists of six main functions:state data collection,service state management,anomaly detection,platform state assessment,alarm notification,and monitoring visualization.Among them,state data collection collects server and service operation status information through a status monitoring service.Service state management manages and controls various services deployed on servers.Anomaly detection includes threshold anomaly detection and model anomaly detection services.The analysis service evaluates the running state of various functions of the cloud platform based on state data and anomaly detection results.Alarm notification sends notifications to administrators through various channels by analyzing anomaly detection results.Monitoring visualization provides an interactive graphical operation and maintenance system interface for administrators.This thesis proposes an anomaly detection method StI-AD(Sequence to Image Anomaly Detection)based on a pre-trained model that migrates the multi-dimensional time series classification problem to the image field,and realizes a multi-dimensional data anomaly detection based on a pre-trained model in the visual field method.The method first proposes a method based on the Graham angle field method to convert multi-dimensional time-series data into a two-dimensional three-channel RGB image.With the help of the periodic fluctuation of the cloud platform state data on a day-by-day scale,the collected state data is converted into an RGB image on a day-by-day basis.Second,the method is based on the underlying structure of ResNet-50 with appropriate modifications,and the generated images are used for fine-tuning training of the ImageNet-based pre-trained model.Compared with methods based on pre-trained models such as Resnet-34,EfficientNet-v2,and vgg-19,this method reduces the length of the bottleneck period of retraining required for model update,and also uses polar coordinates to maintain the dependence of data on time.Through Converting time series data into polar coordinates can reflect the time domain correlation of coordinates.A series of experiments show that this method is effective in cloud platforms in different environments,and has a certain degree of performance improvement compared with other model structures mentioned above.