Neural network model optimization method based on annealing process for stainless steel ultra-thin strip

A neural network model optimization method based on an annealing process for a stainless steel ultra-thin strip, wherein the stainless steel ultra-thin strip has a thickness on the order of several microns to tens of microns and exhibits a size effect, the method establishing a predictive model between annealing process parameters and material mechanical properties by training a single-hidden-layer error back-propagation neural network, wherein an error back propagation algorithm is employed to train the single-hidden-layer error back-propagation neural network, the neural network model optimization method comprising:

• step 1:
• designing a network model, and determining a number of layers of the network, and a number of nodes of an input layer, a number of nodes of an output layer and a number of nodes of a hidden layer, and defining thresholds corresponding to neurons of each layer;
• wherein heat treatment temperature, heat preservation time, and sampling direction are selected as inputs of the network model, and wherein yield strength, tensile strength, and elongation after fracture and hardness are selected as outputs of the network model;
• step 2:
• selecting transfer functions, a training method and training parameters, wherein the training method uses a Levenberg-Marquardt (LM) algorithm, wherein the training parameters needed by the single-hidden-layer error back propagation neural network comprise initial weights, initial values of the thresholds defined corresponding to the neurons of each layer, a learning rate, a momentum factor, a maximum number of iterations, and an error tolerance;
• step 3:
• selecting sample data based on step 2, dividing the sample data into a training set and a testing set, and performing data preprocessing;
• step 4:
• setting and initializing parameters of the neural network;
• step 5:
• adjusting forward propagation of a working signal of the neural network;
• step 6:
• adjusting back propagation of an error signal of the neural network;
• step 7:
• calculating an error value matrix and a Jacobian matrix;
• step 8:
• updating weights and thresholds of the neural network according to the LM algorithm and the Jacobian matrix;
• and
• step 9:
• performing error calculation and testing of the trained single hidden layer neural network.