Research On Industrial Model Service Elastic Scheduling Platform Based On Prediction In Edge Computing Environment

With the rapid development of the industrial Internet,information technology is widely used in all aspects of industrial production.Enterprises have accumulated a large amount of data in production and operation.These data provide a data basis for factory scientific analysis and optimized production.Professional data analysts Use data-driven modeling technology to establish many model applications in parameter prediction and performance optimization.However,when deploying the developed industrial model in the production environment,it faces many problems and limitations.First,industrial data has the characteristics of real-time and privacy,so it needs to be collected and processed near the data source.A lot of models to be generated in the actual production environment,repeated development of similar models is inefficient and wastes of human resources,and requires automatic model construction.At the same time,during the period of providing services,the model lacks professional management,self-improvement and evolution,and cannot provide long-term effective value.In addition,a large number of model services are deployed at the edge computing gateway,and resources are limited.The traditional deployment method is to statically allocate according to the peak of service traffic,and cannot effectively use limited resources.Therefore,the model has many practical problems to be solved when providing services for actual production.In view of the difficulties encountered in the actual deployment of model services,this paper proposes an elastic scheduling platform for industrial model services based on prediction,which solves the problem of lack of life cycle management and update of model services,and can dynamically allocate model service resources according to the amount of requests.The main research work of this article is as follows:Through research on big data,elastic scaling and other technologies,an elastic platform to support the operation of model services is built,and all functional modules are researched and implemented through container technology packaging to improve the efficiency of service deployment.First,Kafka and Spark technologies are used to build real-time processing architectures,to solve the problem of real-time collection and calculation of industrial data.The processed data is stored in the database to provide a data basis for the update of the model service.Once the traditional model is built and deployed,there is a lack of dedicated personnel to manage it.Over time,the accuracy of the model will begin to decline.In response to this problem,the platform supports the life cycle management of the model service.The model can be automatically trained and updated.Design Model training can use historical data and new data to retrain the model file,the model can be seamlessly switched through the container rolling update,and the task scheduling function can be used to run the components in the model update process to achieve automatic model update.Due to the complexity of industrial data,more and more model services will be deployed in the platform.In order to meet the potential load demand,resources are usually configured according to the ability to access the peak,which results in waste of resources during off-peak periods and resource utilization is greatly reduced.Reduced,the model services that the platform can deploy will also be reduced.Because of limited resources on the industrial site,a flexible forecasting-based scheduling architecture is proposed to manage the number of model service containers,which mainly includes a monitoring module,a resource scheduling module,and a model service scaling group.In addition to collecting host and container resource usage indicators,the monitoring module also adds the collection of service application request volume and reponse time that traditional monitoring systems lack.The resource scheduling module uses a random forest algorithm to build a prediction model on the historical data of the request volume,predicts the request volume after multiple steps,allocates resources to the model service scaling group in advance,considers the startup time of the container,and reduces the shortcomings of responsive scaling resource allocation lag.A downsizing count value is designed to avoid scaling jitter.Combined with the function of Rancher container management platform,the platform proposed in this paper is realized and applied to a chemical rectification tower plant to provide parameter prediction and fault diagnosis services.Through simulation experiments,it is verified that the model service can automatically update the model after the new data is generated to maintain the long-term service quality of the service.The comparison with the responsive elastic strategy and the test under real load also verify that the platform's elastic strategy has usability and scalability.The results show that while improving the resource utilization rate of the platform,it can also maintain the service quality of the model service in both prediction accuracy and response time,which is of great significance to help companies achieve scientific decision-making,optimize production,and improve economic efficiency.