A New High Generalization Response Surface Estimation Approach Based On Deep Learning For Complex Product Simulation Optimization

The level of functional integration and intelligence of mechatronics products is constantly improving,and the system composition of products is becoming ever more complex.In the complex product design process,the specific approximation function for computer simulation and optimization will present significant features such as multidisciplinary,multi-influence factors,multi-output response and strong coupling nonlinearity.At the same time,for the traditional response surface method,it still needs to face the challenge of lack of model training sample data.Due to the weak generalization of the response surface,it will be difficult to support the accurate characterization of complex relationships between multiple factors and multiple responses involved in complex product design.In this regard,the deep response surface based on deep learning can theoretically have better model performance potential than the shallow model.However,a deep network model with neurons that select inappropriate activation functions or models that have mismatched model data sets will perform worse than the shallow model when approximating a specific function.At the same time,the deep network model will become very slow in the training convergence speed,and even the model training may not converge.To this end,in this paper,a deep response surface estimation method based on the deep full-connected feedforward neural network is proposed.First of all,the model should have a relatively deep depth(at least 3 or more hidden layers),and to rationally configure the activation function for each layer of the model with the function ReLU with a constant gradient and a function such as sigmoid with strong nonlinear characteristics.There are multiple layers of hidden layers in which the neuron activation function is set to a strong nonlinear function such as sigmoid,so that the input data of the model can be acquired with a higher level of abstract features after multiple space mappings,thereby improving the nonlinear expression ability of the model and ultimately improving its feature learning ability.Moreover,the neuron activation function in most hidden layers in the model is set to the ReLU function,and the constant gradient of the activation function can ensure that the model has a better convergence in training.Additionally,poor training sample data set can greatly expand the number of training samples by increasing the Gaussian noise.The Gaussian noise added during the model training process is amplified by the multi-layer network and transmitted to the output layer to achieve the effect of regularization of the network training,which makes the model over-fitting problem effectively mitigated,and the model training convergence speed is also improved.Finally,the model can directly implement effective training only through the BP algorithm,and construct a deep network model with stronger generalization ability,and the whole model training operation is simple and convenient.In the end,the algorithm and model are implemented under TensorFlow,and the performance test is conducted by the test function approximated hard and designed in this paper,and applied to the sensor correction problem of linear encoder example in the design of precision numerical control equipment under small-scale data and the optimization example of the single leg section size ratio of the hexapod robot.The results show that the depth response surface proposed in this paper is nearly 3 times higher than the traditional response surface on the Adjusted R-square,and it has also improved the fitting accuracy evaluation standard,RMSE,by at least 30%.