Research On Technology Of Heterogeneous Deep Neural Network Model Scheduling Test Engine

With the continuous advancement of production technology,the competition between enterprises has become more and more fierce.If an enterprise wants to stand out,the quality of its products is especially important.Product quality is mainly guaranteed through quality inspection.Traditional manual inspection has been replaced by artificial intelligence.Greatly improve the detection efficiency and the detection quality.When artificial intelligence performs detection tasks,it is necessary to use a trained deep neural network model to detect the images of finished industrial products.Image detection is a very complex technology.It requires many different types of models to form a heterogeneous deep neural network model orchestration to achieve.People generally use use detection efficiency and detection accuracy to evaluate the quality of image detection,and the execution framework,size,type,and GPU of the model will affect these two points.Therefore,when performing inspection tasks,we need to select models and allocate reasonable GPUs to them.At the same time,factors such as inspection time,cost,quality,and machine load must be taken into consideration.In order to perform scheduling test on the zipper defect detection model,this paper designs a heterogeneous deep neural network model scheduling test engine and conducts research on its related technologies.First,we establish the corresponding mathematical model for the scheduling problem of the zipper defect detection model.Then,we expand and improve the non-dominated sorting genetic algorithm based on reference point(NSGA-III),propose the NSGA-III algorithm based on the dominance hierarchy and evolutionary algebra(LGNSGA-III).Finally,the scheduling algorithm is applied to the scheduling test engine of the heterogeneous deep neural network model.The main content can be divided into the following points:(1)Describe the scheduling problem of the zipper defect detection model studied in this paper,and then analyze the problem from the four aspects of time,quality,cost,and load.Then,in order to reduce the complexity of establishing the mathematical model and reduce the influence of interference factors,make reasonable assumptions for the problem,and then on the basis of problem analysis,establish five evaluation indicators from the four aspects of time,quality,cost,and load,and determine the objective function and objective constraints based on this,and establish a multi-objective zipper defect detection Mathematical model of model scheduling problem.(2)In-depth study of the process and principle of the NSGA-III algorithm,and improve the algorithm.Aiming at the influence of its fixed genetic operation on the convergence rate,an adaptive crossover and mutation operator based on evolutionary algebra and dominance level is proposed;Aiming at the problem that it is easy to fall into a local optimal solution,a hierarchical selection strategy based on evolutionary algebra and dominance hierarchy is proposed;for the problem that it cannot construct a hyperplane,a new type of extreme point selection method is used.Then,combined with the actual problem of zipper defect detection model scheduling,the three-stage coding mode of detection process,model selection,and GPU selection,as well as the corresponding decoding mode and initialization mode,are designed,and proposed the LG-NSGA-III algorithm.Finally,comparative simulation experiments were carried out on the standard objective function,standard test cases and real cases.The maximum scheduling completion time of the LG-NSGA-III algorithm was reduced by 5.18% compared with the NSGA-III algorithm,and the maximum load of the machine was reduced by 3.66%.The total load is reduced by 1.04%,the cost is reduced by8.96%,and the detection quality factor is reduced by 8.11%,which proves the effectiveness of the LG-NSGA-III algorithm.(3)Designed and implemented a heterogeneous deep neural network scheduling test engine,introduced the architecture design of the engine in detail,and then tested the three main functional modules of model management,edge management,and scheduling testing,and dispatched the zipper defect detection model proposed in this paper.The algorithm is applied to actual industrial production.After the user configures the model and model layout on the web page,the scheduling test engine invokes the improved algorithm of this article to obtain the optimal scheduling plan,and then realizes the automated deployment test through Docker technology,and obtains the zipper defect detection result.