Simulation Optimization Method And Application Of Complex Electromechanical Products Based On Deep Surrogate Model

Product design based on simulation optimization can significantly reduce the expenditure of time and money caused by physical prototype test,and it has been widely used for electromechanical product development process.As there are many influencing factors to be considered for complex electromechanical products(e.g.,engineering machinery,automobiles,ships,etc)and they are coupled with each other,the simulation models are often multidisciplinary and strongly nonlinear,which makes the simulation solution time very long The simulation optimization methods based on surrogate model can reduce the number of calls to the simulation model and have gradually become indispensable technical means in complex electromechanical product development process.However,there are some shortcomings in the existing simulation optimization methods,i.e.,less attention to the large amount of simulation data generated in the past product development process,lack of reuse of the knowledge of the variation characteristics between the design variables and the performance responses implied in the methods,etc.Moreover,they mainly focus on the characterization of the variation characteristics among the low dimensional,few variables and single performance responses,so it is difficult for them to support the characterization of high-dimensional and high-order nonlinear changes among multivariable and multi-performance responses of complex products.To solve these problems,the paper proposes an efficient simulation and optimization method for complex electromechanical products based on deep agent model,which mainly focuses on the following aspects of the research work.(1)An adaptive construction method of deep surrogate model for complex electromechanical products based on a large amount of simulation data is studied.Based on the existing data,traditional deep agent modeling methods usually require artificial setting of hyperparameters of deep learning algorithms,and the training neural network parameters are used to approximate the variation characteristic relationship between the design variables of training data and the performance response.However,improper configuration of hyperparameters can easily lead to under-approximation or over-approximation of the model.To solve this problem,an adaptive construction algorithm of the deep surrogate model based on MH-TRMPS is proposed.The MH(Modified Hyperband)algorithm is used to generate a large number of hyperparameter configurations randomly,and the early approximation performance of the deep surrogate model is obtained through iterative learning and training of simulation data.Based on this,some configurations with poor performance are eliminated.Multiple frameworks and multiple initial iterative resources are designed to repeat the process,and a set of approximately optimal hyperparameters are finally obtained.Furthermore,a Trust region(TR)is constructed in the neighborhood with the optimal configuration,and Mode pursuing sampling is used to search TR to obtain more configurations that can improve the deep surrogate model’s performance.Meanwhile,the size is reduced and the position is adjusted according to the search results of MPS in order to quickly get the optimal configuration in TR,and the global test is carried out with MH.The results of multiple test cases show that the proposed algorithm can be used to adaptively construct a deep surrogate model that accurately approximates the large amounts of simulation data.(2)The transfer learning construction method of deep surrogate model based on a small amount of simulation data planning for new products is studied.In the deep surrogate model of new products obtained by using transfer learning to construct previous product variation design or adaptive design,existing methods usually have problems such as insufficient consideration of model uncertainty caused by a small amount of data and insufficient reuse of existing model features.In order to solve these problems,the optimality design criteria and multiple iteration are introduced,and combined with random Monte Carlo sampling,an Active Close-Loop Transfer Learning(ACTL)algorithm is proposed.In ACTL,the Fischer information matrix based on the design point distribution is used to represent the uncertainty of the deep surrogate model design points.And the design point chasing sampling mode which maximizes the determinant value of the matrix is established.Then the random Monte Carlo programming is used to sample the simulation data.With the aim of improving the generalization ability of the agent model,the deep surrogate model and Fischer information matrix are updated by active closed-loop transfer learning,and the control factor is introduced to improve the sampling probability of the design points with large determinant values.At the same time,the design points corresponding to the planned simulation data are made capable of covering the whole design space statistically.Experimental results show that the proposed algorithm can stably construct a deep surrogate model of the new product with strong generalization ability under the condition of a small number of samples(3)An efficient multi-objective optimization method for complex electromechanical products based on deep agent model is studied.Simulation optimization of complex products is mainly multi-objective optimization.The internal correlation among each objective makes it impossible to obtain the absolute optimal solution.Traditional multi-objective optimization methods usually take the entire population or neighborhood solution set as mating pool to randomly select parent individuals to generate new solutions.The population size is maintained by environmental selection based on dominant relationships.And the lack of consideration of the structural characteristics of the solution and the characteristics of the problem is not conducive to maintaining the diversity of the solution in the decision space Therefore,Spectral clustering and reference point set are introduced based on MOEA/D,and then the algorithm MOEA/D-DP(MOEA/D with Decision Preference)is proposed.In each generation,MOEA/D-DP first uses spectral clustering to explore the structural characteristic of the solution.Based on this characteristic and the set of design reference points given by the decision maker,the solutions which are close to the design reference points and greatly different from each other are selected as the parent generation individuals.And a variety of different mutation operators are guided with a certain probability to generate new solutions,which make the new solutions more diversified and automatically carry the preference information of decision makers.Furthermore,in order to maintain the population size and the diversity of the solutions in the decision space,the structural characteristics and the dominant relationships of the solutions are combined to select the environment.Furthermore,in the environmental selection phase,an external file is designed to preserve the solutions that are closest to the design reference points and that dominates the current design reference points,in order to facilitate the designer’s decision and the implementation of the decision scheme MOEA/D-DP is combined with the deep agent model to achieve efficient simulation optimization of complex electromechanical products(4)Application of simulation and optimization for complex electromechanical products based on deep surrogate model.Based on the above theoretical research,and combined with the transverse project "Digital Prototype of Telescopic Boom Forklift",the design optimization of a telescopic Boom Forklift is carried out:1)Optimize its hydraulic control system performance of the boom action characteristics under the lifting operation;2)Optimize its safety and comfort under the driving operation.Realization:1)Reduce the maximum fluctuation of amplitude cylinder pressure by 46%in the boom expansion process;2)Reduce the roll angle of the frame by 46.5%on average and the Z-axis amplitude of the cab base by 18%on average in the process of driving.