ABSTRACT: The key process of producing oxygen-free copper strip by submerged-flow horizontal continuous casting billet-high precision cold rolling is melting-casting process. For the casting process, if the parameters do not match properly, there will be some defects such as coarse grain, micro-crack, shrinkage hole, shrinkage porosity and segregation in the billet, which will reduce the density. Through the orthogonal experiment of continuous casting process parameters, the neural network description between billet density and main process parameters is established. The combination of network model and Genetic Algorithm is optimized, the structure quality and density of the billet are improved.

Key words: oxygen-free Copper Strip billet; horizontal continuous casting; process parameter optimization oxygen-free copper processing material plays an important role in information, electronics, electric power, refrigeration and other industries because of its excellent conductivity and thermal conductivity, especially in high-power electric vacuum devices is irreplaceable [1,2] . Densification and oxygen content are the two most important internal quality indexes of oxygen-free copper processing material, which directly affect the service performance of the material. The production of oxygen-free copper strip by submerged-flow horizontal continuous casting ingot-high precision cold rolling process avoids oxygen permeation in the process because there is no hot rolling process, if the technological parameters and their combination are not reasonable, there will be some defects such as coarse grain, micro-crack, shrinkage hole, shrinkage porosity and segregation in the slab, these defects lead to low density and instability, and are also the main factors that affect the yield and service performance of the Strip. In this paper, the casting temperature, cooling strength, drawing parameters and their matching, which affect the quality of the casting billet, have been studied experimentally, the optimum process parameters and combination were obtained, the structure and quality of the billet were improved and stabilized, and the density of the strip billet was increased, which provided a feasible method and theoretical basis for production.

1. Process flow and experimental factor selection

 1.1 process flow

The complete process of producing oxygen-free copper strip by horizontal continuous casting billet-high precision cold rolling technology is as follows: continuous horizontal casting billet is milled, roughed and rolled, among them, horizontal continuous casting strip billet is the key stage which affects the inner quality of Strip. This experiment is completed on the actual production line. The melting-heat preservation triplet furnace, crystallizer, tractor and system control of continuous casting unit are produced by our company. Figures 1 and 2 are the schematic diagram of continuous casting process and crystallizer respectively. The crystallizer is a double casting channel, the length of cooling water is 240mm, the length of cooling water is 240mm, the length of graphite mould is 290mm.

1.2 The factor selection experiment is designed to explore the optimized process parameters and combinations of continuous casting parts, based on production experience, theoretical analysis and equipment Operability, the casting temperature (melt temperature in the holding furnace) , cooling water inlet temperature (water temperature measured into the crystallizer) , water pressure of the system, residence time during the drawing cycle, back-push time, back-push pause time and drawing time were selected as seven parameters

To investigate the factors, the levels of each factor are casting temperature 1180,1240,1280 °C, cooling water inlet temperature 18,25,30 °C, water pressure 0.4,0.6,0.8 MPA, respectively, traction pause time is 3.50,4.50,5.50,6.20,6.80,7.60 S, retraction time is 0.08,0.16,0.25 S, retraction pause time is 0.15,0.20,0.25 s, retraction time is 0.35,0.50,0.70 S. The examination index is the compact density of the Strip. Drawing speed is a control variable, and the optimized parameters and combinations belong to specific drawing speed to facilitate production. The orthogonal experiment scheme L18(6136)[5] under the condition of 17.5 mm/s drawing speed, and the experimental and simulation results are shown in table 1.

2. Network model 

2.1 network structure three-layer Feed-forward Networks can approximate any nonlinear relationship with any precision. In this paper, the experimental model of the network structure is shown in Fig. 3. There are 7 neurons in the input layer, 14 neurons in the hidden layer and one neuron in the output layer. A row vector xj = [ x 1, ... , xn ] t is used to represent the input to the hidden layer or the output layer, where n is the number of the next neuron, and a row vector WJ = [ J 1, j 2, ... , Ji, 177

Jn ] represents the connecting weight vector of Neuron J in hidden layer or output layer, JI represents the first input of Neuron j, the total input of Neuron J, SJ = ∑ Ni = 1 Jixi + J = Wjxj + J, J is the threshold, and F (SJ) = 11 + e-sj is the transfer function, the output of Neuron J is yj = F (SJ) = [1 + EXP (- WjXj + J)]-1. 2.2 the network learning takes the casting temperature, traction pause time, traction pause time, traction pause time, cooling water inlet temperature, cooling water pressure and density of each experimental group of orthogonal experiment as an input sample pair for network training. Let the system have p sample pairs, and the error function is defined as: EP = 12∑ MJ = 1(djp-yjp)2, DJP and YJP are the actual output and supervisory signals of the J output element respectively. In the network learning stage, if the network output error is less than the specified precision, it will end, otherwise, it will go to the error back Propagation (ErrorBack Propagation) , that is to say, the error signal will be corrected by the weight vector of each neuron along the original connection path, causes the error signal to achieve the specified precision so far. The basic iterative formula is [7] : (K + 1) = (K)-(JTJ + ul)-1g where G is the gradient of the error function to the weight vector, J is the Jacobian Matrix of the differential of the error function to the weight vector, and I is the Unit Matrix, u is an adjustable nonnegative number. After learning, the intrinsic rule of data is stored in the network in the form of neuron weight vector value for the use of the network stage. This experiment takes 0.001 as the training accuracy, that is, the end of network learning criterion, the learning speed changes step length, the initial is 0.01, the sample data obtained from three groups of parallel experiments are trained, after 18 test samples are tested, the simulation results are shown in the last column of Table 1. 

3. The optimization and experiment are trained and verified by the network model used to optimize the prediction. In the optimization stage, the input (a combination of factors) of the network is generated by genetic algorithms (Ga) , and the output of the network is the fitness function value of Ga.. Each factor is encoded by a 4-bit binary code, 28-bit binary string composed of 7 process parameters is an individual, and 20 individuals are randomly generated to form the initial population Y (0) = { Y1(0) , y 2(0) , ... , Y20(0)} , so that K = 0; Calculate the fitness function value of each individual in population y (k)(input an individual as the factor level of an experiment into the network, the network output is the fitness function value of that individual) ; Some pairs of Y (K) were selected to cross (the probability of PC = 0.6) , and each new individual was mutated according to the probability of pm = 0.04, and the mutated individual was regarded as the individual of the next generation of y (K + 1) , k = K + 1; The individual with the largest fitness value in the current generation is the optimal process parameter until the difference of the average fitness of the individual in the fourth successive generation is ≤0.01. In this experiment, when the casting speed is 17.5 mm/s, the optimum density is 8.94 g/cm3 in the casting speed range, and the corresponding parameters are: casting temperature 1230 °C, traction pause time 5.2 s, retraction time 0.17 s; The retracting pause time is 0.19 s, the traction time is 0.48 s, the inlet temperature of cooling water is 19 °C, and the pressure of cooling water is 0.53 MPA. The experimental results show that the actual density is 8.93 g/cm3 ± 2% and the copper content is 99.97% ~ 99.99% . 4. Conclusion the density and oxygen content of the Strip are the two most important internal quality indexes of the oxygen-free copper processing material, which directly affect the performance of the material, the density depends on the process conditions and the composition and content of the alloy itself. By means of orthogonal experiment and network modeling and optimization, the discontinuity of experiment points in single factor or orthogonal experiment is compensated to a great extent, a set of better process parameters can be obtained under certain equipment and production environment, and at the same time, the actual experiment frequency and cost can be reduced. 

Source: Chinanews.com, by Guo Mingen