Control method of pressure control equipment for uniformly distributing multiple layers of composite copper foils

Abstract

The invention provides a control method of pressure control equipment for uniform distribution of multiple layers of composite copper foils, which comprises the following steps: acquiring material attribute parameters of each layer of the multiple layers of composite copper foils, and establishing a composite copper foil material attribute database; analyzing stress distribution and deformation conditions of each layer of the copper foil under different pressure conditions through a finite element analysis method in combination with a multi-layer structure and material attribute data of the composite copper foil, so as to obtain a quantitative evaluation index of pressure distribution uniformity; according to the quantitative evaluation index of the pressure distribution uniformity, the mapping relation between the pressure distribution uniformity and the technological parameters and the mapping relation between the pressure distribution uniformity and the equipment mechanical characteristics are analyzed, and an optimal technological parameter combination and an equipment control strategy for achieving uniform pressure distribution are obtained; and when the process requirement changes, a historical optimization case similar to the current process requirement is extracted from the knowledge base as an initial optimization scheme, and optimization is performed through a reinforcement learning algorithm to respond to the process requirement change.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to a control method for a pressure control device for uniformly distributing a multi-layer composite copper foil. Background Art

[0002] In the manufacturing process of multi-layer composite copper foil, the multi-layer structure and material properties of the copper foil lead to the problem of uneven pressure distribution when applying pressure. Due to the differences in material properties between the layers of copper foil, such as thickness, hardness and surface roughness, the deformation and stress distribution between different layers are not consistent under the same pressure. This uneven pressure distribution will lead to differences in the bonding strength between the layers of copper foil, which in turn affects the overall performance and quality of the composite copper foil. In addition, in the actual production process, due to changes in process parameters and the mechanical characteristics of the equipment, it is difficult for the pressure application device to achieve precise control and dynamic adjustment of pressure. This requires a control method that can achieve uniform pressure distribution and dynamic adjustment according to the multi-layer structure and material properties of the copper foil. At the same time, the control method also needs to take into account the changes in process requirements and the mechanical characteristics of the equipment during the production process to ensure that accurate control and rapid response to pressure can be achieved under different working conditions, thereby ensuring the production quality and efficiency of the composite copper foil. Summary of the invention

[0003] The present invention provides a control method for a pressure control device for uniformly distributing a multi-layer composite copper foil, which mainly includes:

[0004] Obtaining material property parameters of each layer of the multi-layer composite copper foil and establishing a composite copper foil material property database;

[0005] By using the finite element analysis method, combined with the multi-layer structure and material property data of the composite copper foil, the stress distribution and deformation of each layer of the copper foil under different pressure conditions are analyzed to obtain a quantitative evaluation index for the uniformity of pressure distribution.

[0006] According to the quantitative evaluation index of pressure distribution uniformity, the mapping relationship between pressure distribution uniformity and process parameters and equipment mechanical characteristics is analyzed to obtain the optimal process parameter combination and equipment control strategy to achieve uniform pressure distribution;

[0007] The optimal process parameter combination and equipment control strategy for achieving uniform pressure distribution are applied to the copper foil production line. During the production of composite copper foil, the process parameters and equipment operation status data are collected in real time to dynamically predict the pressure distribution under the current working conditions. If the prediction results meet the requirements of the quantitative evaluation index of pressure distribution uniformity, the current process parameters and equipment control strategy are maintained unchanged.

[0008] If the predicted pressure distribution does not meet the uniformity requirements, the process parameters and equipment control strategies are dynamically adjusted through the reinforcement learning algorithm to make the pressure distribution gradually uniform, and the optimized process parameter combination and control strategy are updated to the knowledge base;

[0009] When process requirements change, historical optimization cases similar to the current process requirements are extracted from the knowledge base as the initial optimization plan, and then optimized through the reinforcement learning algorithm to respond to changes in process requirements.

[0010] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0011] The invention discloses a control method for a pressure control device for achieving uniform distribution of multi-layer composite copper foil. The method establishes a composite copper foil material property database, establishes a pressure distribution model through finite element analysis, and uses a machine learning algorithm to establish a mapping relationship model between pressure distribution uniformity and process parameters. In the production process, an online thickness gauge and a multi-directional pressure detection device are used to collect copper foil thickness and pressure data in real time, and a multi-physical field coupling simulation method is combined to simulate transient pressure distribution under high-speed motion. The pressure roller parameters are adjusted in real time through an adaptive control algorithm to ensure the stability of the copper foil thickness under dynamic conditions. A reinforcement learning algorithm is used to dynamically optimize process parameters and equipment control strategies so that the pressure distribution gradually tends to be uniform. When the process requirements change, similar historical cases are extracted from the knowledge base and fine-tuned to quickly respond to process changes. The invention realizes precise control and dynamic optimization of pressure distribution in the production process of composite copper foil, effectively improves the uniformity of copper foil thickness and product quality stability, and provides technical support for the large-scale production of high-performance composite copper foil. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flow chart of a control method of a pressure control device for uniformly distributing a multi-layer composite copper foil. DETAILED DESCRIPTION

[0013] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0014] like Figure 1 The control method of the pressure control device for uniformly distributing the multi-layer composite copper foil in this embodiment may specifically include:

[0015] S101, obtaining material property parameters of each layer of a multi-layer composite copper foil, and establishing a composite copper foil material property database.

[0016] Obtain scanning electron microscope photos of each layer of the composite copper foil structure; obtain surface morphology point cloud data based on the scanning electron microscope photo measurement data; use a Gaussian filter to process the point cloud data to obtain a surface contour value; apply a cyclic load on the surface of the composite copper foil according to the surface contour value, obtain the Young's modulus and hardness value of each layer through load-displacement curve calculation, and establish a composite copper foil material property database.

[0017] For example, the three-dimensional digital point cloud data of the copper foil surface morphology is obtained based on the high-magnification scanning electron microscope photo measurement data of each layer in the composite copper foil structure, the root mean square value of the surface roughness of each layer and the surface morphology characteristic value are calculated through the point cloud data, and the surface profile value of each layer is obtained by processing the point cloud data with a Gaussian filter with a cutoff wavelength of 0.8 microns. A nanoindenter is used to apply cyclic loads to the surface of each layer of the composite copper foil, and the elastic-plastic deformation behavior of each layer of the material is quantitatively analyzed through the load-displacement curve. The Young's modulus and hardness value of each layer are calculated according to the elastic modulus equation in the Oliver-Fahr model, and a composite copper foil material property database is established. According to the surface profile value and the load-strain relationship data, a digital model of the cross-sectional profile of each layer of the composite copper foil is established, and the contact mechanics equation is used to calculate the interlayer stress distribution state. For the interlayer stress distribution data, an acoustic emission detector is used to record the interface stress wave signal under the action of the pressure load, and the signal is decomposed and reconstructed into five layers through the db6 wavelet basis function. The fuzzy clustering algorithm is used to classify and identify the stress wave characteristics, and the interface stress concentration area is determined. Based on the surface profile values ​​of each layer, the load-strain relationship and the interface stress concentration area data, the finite element analysis software was used to construct the mechanical model of the composite copper foil multilayer structure, and the deformation of each layer was calculated by the nonlinear stress-strain relationship. According to the deformation data of each layer, the strain energy density criterion J = (1 / 2E)(σ^2+τ^2) was used to judge the interface stress concentration position, and the deformation stress distribution law of each layer of the composite copper foil under pressure was obtained. In the surface morphology analysis of the composite copper foil structure, the high-power scanning electron microscope measurement used a 5000-fold magnification to obtain the surface microscopic morphology image, and the surface morphology profile data points were extracted by image processing software to form a three-dimensional point cloud data matrix. For the measurement of surface roughness, a 10×10 micron scanning area was selected, the sampling interval was 0.1 micron, and the height data of 100×100 measurement points were obtained. The original data was filtered using a Gaussian filter, and the filter cutoff wavelength was selected as 0.8 microns, which can effectively remove high-frequency noise. After filtering, the surface roughness root mean square value Ra was calculated to be 0.35 microns. During the material mechanical properties test, the nanoindenter uses a triangular pyramidal diamond indenter with a maximum load set to 500 millinewtons and a loading rate of 10 millinewtons per second. A 5×5 dot matrix is ​​selected on the sample surface for indentation testing. The load-displacement curve is analyzed by the Oliver-Fahr model, where the elastic modulus calculation formula is E=0.5(dP / dh)(π / A)^0.5, P is the load, h is the indentation depth, and A is the contact area. The elastic modulus of the surface copper foil is calculated to be 110 GPa and the hardness value is 1.2 GPa. In the interface stress analysis, the acoustic emission detection uses a piezoelectric sensor array arrangement, and the sensor frequency response range is 100-1000 kHz. The collected acoustic emission signal is processed by wavelet transform, and the db6 wavelet basis function is selected for 5-layer decomposition to extract energy characteristic parameters.The acoustic emission signal features were classified by fuzzy clustering algorithm, and the number of cluster centers was set to 3, corresponding to the three states of intact interface, micro-damaged interface and interface peeling. In the finite element analysis, a multilayer structural mechanical model was established according to the measured material parameters. The unit type was selected as 8-node hexahedral unit, and the mesh size was 0.1 micron. Under the action of pressure load, the stress concentration position was judged by the strain energy density criterion J = (1 / 2E)(σ^2+τ^2), where σ is the normal stress and τ is the shear stress. When the strain energy density J exceeds the critical value of 0.5 joules per cubic millimeter, it is judged as a stress concentration area. The calculation results show that stress concentration is easy to occur at the interface due to the sudden change of material properties, and the maximum strain energy density value reaches 0.8 joules per cubic millimeter.

[0018] S102. By using the finite element analysis method, combined with the multi-layer structure and material property data of the composite copper foil, the stress distribution and deformation of each layer of the copper foil under different pressure conditions are analyzed to obtain a quantitative evaluation index of the uniformity of pressure distribution.

[0019] A three-dimensional finite element geometric structure is established according to the number of layers and material density of the composite copper foil, a tetrahedral mesh is generated by a mesh divider, and the surface morphology characteristic values ​​of the copper foil are obtained from the scanned data; the contact area is calculated using the surface morphology characteristic values ​​of the copper foil, and the interlayer friction stress value is obtained by the Coulomb friction law; the node displacement field is calculated according to the interlayer friction stress value by the virtual work principle, and if the local pressure deviation of the node displacement field exceeds a preset threshold, the residual iteration method is used to obtain the corrected node displacement field; the unit strain energy is calculated by the strain energy density function using the corrected node displacement field, and the integral value of the unit strain energy of each layer is obtained by three-point Gaussian integration.

[0020] For example, a three-dimensional finite element geometric structure is established based on the number of composite copper foil layers, material density, interlayer contact state and interface bonding strength data, a mesher is used to generate a tetrahedral mesh, the mesh node displacement is assigned by a four-node bilinear interpolation function, and the copper foil surface morphology and roughness characteristic values ​​are obtained from the scanning data. The contact area is calculated using the copper foil surface morphology data, and the interlayer friction stress is calculated using the Coulomb friction law τ=μσ, where τ is the friction stress, μ is the friction coefficient, and σ is the normal stress. The stress distribution of each layer of the composite copper foil structure is solved using the piecewise cubic Hermite interpolation function. Based on the stress distribution data of each layer, the displacement field of each layer node is calculated using the virtual work principle, and the strain coordination equation is used. Determine the deformation variables of each layer, ε is the strain, u is the displacement in the x direction, and v is the displacement in the y direction. If the local pressure deviation exceeds the preset threshold, the residual iteration method is used to correct the displacement field, and the stress equilibrium equation is used. Solve the node force balance, σx is the normal stress in the x direction, and τxy is the shear stress. According to the corrected displacement field data, the strain energy density function U=1 / 2(σε+τγ) is used to calculate the unit strain energy, where U is the strain energy density, ε is the normal strain, and γ is the shear strain. The integral value of the unit strain energy of each layer is obtained by the three-point Gaussian integral. The pressure distribution uniformity evaluation index σ=√[Σ(Ui-Uavg)^2 / n] is calculated using the strain energy integral value, where Ui is the unit strain energy, Uavg is the average strain energy, and n is the total number of units. In the modeling of the pressure distribution of the composite copper foil, the copper foil surface is first meshed. For the analysis area of ​​0.1 square millimeters, tetrahedral units are used for meshing. The unit size is set to 5 microns, and a total of 20,000 mesh units are generated. Based on the scanning data showing that the root mean square value of the surface roughness is 0.35 microns, the displacement of the mesh nodes is assigned by a four-node bilinear interpolation function, and the interpolation function uses a shape function expression. In the interlayer contact analysis, the surface morphology data shows that the actual contact area is about 75% of the nominal contact area, the friction factor μ is taken as 0.3, and when the normal pressure is 100 MPa, the interlayer friction stress τ is calculated to be 30 MPa according to the Coulomb friction law. The stress field is interpolated by the piecewise cubic Hermite interpolation function, and the interpolation interval [-1,1] is divided into 50 subintervals. In the deformation calculation process, the node displacement field is calculated by the virtual work principle δW=Σ(σδε+τδγ), where σ is the normal stress, τ is the shear stress, ε is the normal strain, and γ is the shear strain. The calculation results show that the maximum node displacement is 2.5 microns, which occurs in the boundary load action area. The strain coordination equation is used to verify the continuity of the deformation, and the maximum strain is calculated to be 0.025. When it is found that the local pressure deviation exceeds the set threshold of 10%, the residual iteration method is used to correct the displacement field, and the iterative convergence criterion is set to 0.001. The node force balance is solved by the stress balance equation, and the calculation shows that the maximum stress concentration factor is reduced to 2.1 after correction. Then, the unit strain energy is calculated using the strain energy density function U=1 / 2(σε+τγ), and the Gaussian integration point coordinates are selected as ±0.774597, and the weight coefficient is 0.555556 for numerical integration. Finally, the pressure distribution uniformity evaluation index σ=√[Σ(Ui-Uavg)^2 / n] is obtained through strain energy calculation, where the unit average strain energy Uavg is 0.45 joules per cubic millimeter, and the standard deviation σ is 0.08 joules per cubic millimeter, indicating that the unevenness of the pressure distribution is 17.8%.

[0021] S103. Analyze the mapping relationship between pressure distribution uniformity and process parameters and equipment mechanical characteristics based on the quantitative evaluation index of pressure distribution uniformity, and obtain the optimal process parameter combination and equipment control strategy for achieving pressure uniform distribution.

[0022] A median filter is used to perform noise reduction processing on the pressure distribution data collected by the pressure sensor, and the noise-reduced data is decomposed to obtain a pressure distribution uniformity characteristic vector; based on the pressure distribution uniformity characteristic vector, a Pearson correlation coefficient method is used to establish a correlation matrix with process parameters, and a radial basis kernel function support vector machine is used to perform nonlinear mapping on the correlation matrix to obtain a parameter coupling coefficient matrix; for the parameter coupling coefficient matrix, an orthogonal design method is used to generate a process parameter test plan, and the test plan is evaluated and calculated through a hierarchical analysis method to obtain a parameter optimization data set; a neural network based on error back propagation is used to train the parameter optimization data set, and if the weight matrix meets the convergence condition after training, the process parameters are combined and optimized through a genetic algorithm of single-point crossover and Gaussian mutation to obtain a pressure uniform distribution control parameter combination.

[0023] Exemplarily, according to the real-time pressure data collected by the pressure sensor, the pressure distribution data is subjected to noise reduction by a median filter, and the noise-reduced data is subjected to five-layer decomposition by the db4 wavelet basis function, and the pressure distribution uniformity eigenvector is obtained from the reconstruction coefficient matrix. The Pearson correlation coefficient method is used to establish the correlation matrix R(i, j) between the pressure distribution uniformity eigenvector and the process parameters and mechanical characteristic parameters, and the radial basis kernel function support vector machine is used to perform nonlinear mapping on the correlation matrix data to determine the coupling coefficient value between the parameters. According to the parameter coupling coefficient matrix, the orthogonal design method is used to generate the L16(4^5) process parameter test scheme, and the hierarchical analysis method is used to evaluate and calculate each combination scheme, and a parameter optimization data set is established based on the calculation results. The parameter optimization data set is trained by a neural network based on error back propagation, and the corresponding relationship between the process parameters and the equipment parameters is fitted by the hidden layer neurons, and the parameter mapping relationship is obtained from the weight matrix after training. According to the parameter mapping relationship matrix, the genetic algorithm of single-point crossover and Gaussian mutation is used to combine and optimize the process parameters, and the optimal solution is calculated by the minimum variance fitness function to obtain the control parameter combination under the condition of uniform pressure distribution. In the pressure distribution data processing, a 5×5 window median filter is used to reduce the noise of the original pressure data. The standard deviation of the data before filtering is 0.85 MPa, and it is reduced to 0.32 MPa after filtering. Then the db4 wavelet basis function is used for 5-layer decomposition. The high-frequency coefficients obtained after decomposition reflect the pressure fluctuation characteristics, and the low-frequency coefficients reflect the pressure distribution trend. The eigenvalues ​​of the reconstructed coefficient matrix are 2.45, 1.87, 1.32, 0.95 and 0.68 respectively. In the correlation analysis between the pressure distribution uniformity eigenvector and the process parameters, the Pearson correlation coefficient method is used to construct the correlation matrix. The calculation results show that the correlation coefficient between pressure and temperature is 0.82, the correlation coefficient with speed is 0.75, and the correlation coefficient with equipment vibration is 0.68. The radial basis kernel function is selected as the kernel function of the support vector machine, where σ takes a value of 1.5. After cross-validation, the classification accuracy reaches 92.5%. In the parameter combination optimization stage, the L16 (4^5) orthogonal table design test scheme was used, including 4 levels of temperature factors of 120 degrees, 140 degrees, 160 degrees, and 180 degrees, 4 levels of pressure factors of 8 MPa, 10 MPa, 12 MPa, and 14 MPa, and 4 levels of speed factors of 5 meters per minute, 8 meters per minute, 11 meters per minute, and 14 meters per minute. The weights of each scheme were calculated by the hierarchical analysis method, and the consistency ratio CR was 0.047, which met the consistency requirements. The neural network training adopted a BP network structure with two hidden layers, with 15 neurons in the first hidden layer and 8 neurons in the second hidden layer. The learning rate was set to 0.05 and the momentum factor was 0.8. After 2000 iterations on 1000 training samples, the training error converged to 0.0025, and the prediction accuracy of the test set reached 95.3%.Finally, in the genetic algorithm optimization, the population size is set to 100, the crossover probability is 0.85, and the mutation probability is 0.05. The minimum variance fitness function f=1 / sqrt(Σ(xi-μ)^2 / n) is used for fitness evaluation, where xi is the single-point pressure value, μ is the average pressure, and n is the number of measurement points. After 200 generations of evolution, the fitness value of the optimal solution increased from the initial 0.62 to 0.93, and the corresponding optimal process parameter combination was a temperature of 155 degrees, a pressure of 11.5 MPa, a speed of 9.5 meters per minute, an equipment vibration frequency of 35 Hz, and a roller gap of 0.8 mm.

[0024] The equipment control strategy includes: collecting copper foil thickness data in real time through an online thickness gauge on a multi-layer composite copper foil production line, processing the collected data through a data analysis method to obtain a spatial distribution map of the copper foil thickness, judging the uniformity of the thickness distribution, and adjusting the pressure roller pressure of the corresponding area when abnormal thickness is found in a local area.

[0025] The data collected by the thickness gauge are processed by a sliding mean filter, and the copper foil thickness distribution value is obtained by fitting through the least square method; according to the copper foil thickness distribution value, a bicubic interpolation algorithm is used to perform spatial transformation reconstruction, and the reconstructed data is gridded by the Kriging interpolation method to obtain a thickness distribution map; for the thickness distribution map, a support vector regression algorithm is used to establish a mapping relationship between pressure and thickness, and the thickness uniformity of the copper foil is judged by the variance calculation formula. If the variance of a local area exceeds a preset standard deviation threshold, the grid coordinates of the area are obtained; according to the grid coordinates, a recursive least squares formula is used to update the pressure compensation value, and a PI controller is used to perform closed-loop regulation on the pressure roller pressure to obtain a dynamic pressure adjustment amount.

[0026] Exemplarily, based on the original data collected by the multi-point scanning thickness gauge, a sliding mean filter with a window length of 128 points is used to preprocess the thickness measurement data, and a third-order polynomial curve is fitted by the least squares method to obtain the copper foil thickness distribution value from the fitting curve. A 16×16-point bicubic interpolation algorithm is used to spatially transform and reconstruct the thickness distribution value, and the reconstructed data is gridded by the Kriging interpolation method to obtain the copper foil thickness spatial distribution map from the gridded data. According to the thickness spatial distribution map, a Gaussian kernel function support vector regression algorithm is used to establish the pressure and thickness mapping relationship, where the kernel function parameter σ is determined by cross-validation. The variance calculation formula is used. Calculate the thickness uniformity index, xi is the thickness of the i-th measurement point, is the average thickness of all measurement points, n is the number of measurement points, and if the variance of a local area exceeds the preset standard deviation threshold, the grid coordinates of the area are recorded. For the grid coordinates of the exceeding area, the recursive least squares formula is used to update the pressure compensation value. According to the pressure compensation value, a PI controller with a proportional coefficient Kp=0.8 and an integral time Ti=0.5 is used to perform closed-loop adjustment on the pressure roller pressure, and the dynamic pressure adjustment amount is obtained from the feedback data. During the copper foil thickness measurement process, the scanning thickness gauge continuously scans the copper foil at a sampling frequency of 100 Hz, and the raw data has a random fluctuation of ±0.15 microns. A 128-point sliding mean filter is used to preprocess the data, and the measurement noise is reduced to after filtering.

[0027] ±0.05 microns. The filtered data was fitted with a third-order polynomial, and the fitting equation was y = 2.35 × 10^-6x^3-4.12 × 10^-4x^2+0.0183x+12.53, where x is the measurement position coordinate, y is the copper foil thickness value, and the goodness of fit R^2 reached 0.985. In the spatial data reconstruction, the 16 × 16 point bicubic interpolation algorithm was used to resample the thickness data, the interpolation function weight coefficient matrix was 4 × 4, and the sampling interval was 2 mm. The Kriging interpolation used the spherical semivariogram model γ(h) = C0 + C[1.5(h / a)-0.5(h / a)^3], where C0 is the nugget value and takes 0.01, C is the base value and takes 1.2, a is the range and takes 50 mm, and the grid resolution is increased to 0.5 mm after interpolation. The support vector regression algorithm uses a Gaussian kernel function, and the kernel parameter σ=0.8 and the penalty factor C=10 are determined by five-fold cross validation. The training data contains 5000 sets of pressure-thickness sample pairs, and the predicted root mean square error after training is 0.08 microns. The thickness uniformity evaluation uses variance calculation. For a 100×100 mm area divided into 400 grid units, the calculated standard deviation threshold is 0.12 microns. In the pressure compensation control, the forgetting factor λ of the recursive least squares algorithm is set to 0.95, the initial covariance matrix P0=100I, and the initial value of the estimated parameter θ0=[0.5,0.3]^T. When it is detected that the thickness standard deviation of a certain area exceeds 0.12 microns, the pressure compensation calculation is triggered and the parameter estimate θ(k) is updated. The pressure closed-loop regulation uses a PI controller, the proportional coefficient Kp=0.8 is used for fast response, and the integral time Ti=0.5 seconds is used to eliminate steady-state errors. The measured data show that when the pressure fluctuation amplitude is ±0.5 MPa, the thickness control accuracy is maintained within the range of ±0.1 micron. Through data analysis, the thickness distribution of the copper foil shows that the middle area is slightly thicker and the edge area is slightly thinner, with the maximum thickness difference of about 0.25 micron. The comparison data before and after pressure compensation shows that after adopting dynamic pressure regulation, the thickness uniformity is significantly improved, and the standard deviation is reduced from 0.15 micron to 0.08 micron, meeting the requirements of high-precision copper foil production.

[0028] The equipment control strategy includes: collecting the pressure data of the copper foil during the production process in real time through a pressure sensor, establishing a mapping relationship between pressure and thickness through a machine learning algorithm, predicting the deformation of the copper foil under different pressure conditions, and guiding the dynamic adjustment of the pressure roller to suppress the thickness difference caused by anisotropy.

[0029] A Butterworth low-pass filter is used to filter the pressure signal collected by the multi-directional pressure sensor array, and the filtered pressure signal is decomposed to obtain the pressure distribution feature matrix from the reconstruction coefficient; based on the pressure distribution feature matrix, feature extraction is performed through a convolutional neural network, and the mapping relationship between pressure and deformation is obtained from the last layer of convolution feature map; for the mapping relationship between pressure and deformation, a principal component analysis model is constructed using the covariance matrix, and the principal components of deformation in each direction are obtained from the eigenvalues; based on the principal components of deformation in each direction, the deformation feature weight coefficient is calculated through singular value decomposition, and an adaptive PID controller is used to compensate and adjust the pressure roller, wherein the proportional coefficient is determined by the deformation feature weight coefficient.

[0030] Exemplarily, according to the pressure signal collected by the multi-directional pressure sensor array, a Butterworth low-pass filter is used to pre-process the pressure data in different directions, and the filtered data is decomposed into three layers by the db4 wavelet basis function to obtain the pressure distribution feature matrix from the reconstruction coefficient. A four-layer convolutional neural network is used to extract the features of the pressure distribution feature matrix, wherein the first layer of convolution kernel size is 5×5, the second layer is 3×3, the third layer is 3×3, and the fourth layer is 2×2, and the mapping relationship between pressure and deformation is obtained from the last layer of convolution feature map. According to the pressure and deformation mapping data, the covariance matrix is ​​used to construct the principal component analysis model U=XV, where X is the standardized deformation data matrix, V is the eigenvector matrix, and the deformation principal components in each direction are obtained from the eigenvalue λi. For the deformation principal component data, the deformation eigenvalue is calculated by singular value decomposition W=USV^T, and the deformation sensitivity in each direction is judged from the singular value sequence Si to obtain the deformation feature weight coefficient. According to the deformation feature weight coefficient, an adaptive PID controller is used to compensate and adjust the pressure roller, in which the proportional coefficient Kp is determined by the deformation sensitivity, the integral time Ti is determined by the response characteristics, and the differential time Td is determined by the dynamic characteristics. The recursive least squares method is used to update the PID parameters in real time, and the compensation effect is evaluated by the prediction error criterion, and the pressure dynamic compensation parameters are obtained from the evaluation results. In the multi-directional pressure detection of copper foil, the pressure sensor array is arranged along four directions of 0 degrees, 45 degrees, 90 degrees, and 135 degrees, with 8 sensors arranged in each direction, and the sampling frequency is 200 Hz. The original pressure signal has a random fluctuation of ±0.2 MPa. A Butterworth fourth-order low-pass filter with a cutoff frequency of 50 Hz is used for preprocessing, and the signal-to-noise ratio is increased to 35 dB after filtering. The db4 wavelet basis function is used to decompose the filtered data into three layers to obtain high-frequency coefficients d1, d2, d3 and low-frequency coefficient a3 reflecting pressure changes at different scales. In the structural design of the four-layer convolutional neural network, the first layer has 16 5×5 convolution kernels to extract local features, the second layer has 32 3×3 convolution kernels to extract medium-scale features, the third layer has 64 3×3 convolution kernels to extract large-scale features, and the fourth layer has 128 2×2 convolution kernels to integrate features of all scales. The ReLU activation function is used after each layer of convolution, and the pooling uses 2×2 maximum pooling. The network training uses 5000 sets of pressure-deformation data pairs, and the training error converges to 0.015 after 2000 iterations. In the principal component analysis, the deformation data is first standardized, and the eigenvalues ​​of the covariance matrix are calculated to be

[0031] [2.85, 1.62, 0.38, 0.15], and the corresponding cumulative contribution rate is [56.8%, 89.2%, 96.8%, 100%]. The first two principal components are selected as the main deformation features, and their eigenvectors are

[0032] [0.82, 0.45, 0.21, 0.12] and [0.15, 0.76, 0.52, 0.23]. The singular value decomposition results show that the singular values ​​in the four directions are 6.82, 4.35, 1.85 and 0.92, respectively, indicating that the deformation sensitivity in the 0-degree and 45-degree directions is relatively high. Based on this, the adaptive PID controller parameters are set, the proportional coefficient Kp in the 0-degree direction is 1.2, the integral time Ti is 0.8 seconds, and the differential time Td is 0.2 seconds; the control parameters in the 45-degree direction are 1.0, 0.6 seconds and 0.15 seconds, respectively. The recursive least squares method uses the forgetting factor λ=0.95 for parameter update, and the prediction error criterion uses the root mean square error RMSE. When the RMSE exceeds 0.1 MPa, the parameter adaptive adjustment is triggered. The measured data show that after adopting dynamic pressure compensation, the thickness variation coefficient in the 0-degree direction dropped from 4.2% to 2.1%, in the 45-degree direction from 3.8% to 1.9%, in the 90-degree direction from 2.5% to 1.4%, and in the 135-degree direction from 2.3% to 1.3%. The overall pressure distribution uniformity is significantly improved.

[0033] The equipment control strategy also includes: establishing a mathematical model of the dynamic contact process between the pressure roller and the copper foil according to the production line speed and the deformation characteristics of the copper foil material, using a multi-physical field coupling simulation method to simulate the transient pressure distribution under high-speed motion, obtain the change pattern of the copper foil thickness with time and position, and adjust the operating parameters of the pressure roller in real time to ensure the stability of the copper foil thickness under dynamic conditions.

[0034] The second-order Newmark time integration algorithm is used to solve the dynamic equation to obtain the displacement field distribution data of the contact process between the pressure roller and the copper foil; the stress field distribution is calculated according to the displacement field distribution data through the stress-strain constitutive equation of the viscoelastic model, and the dynamic contact pressure data is obtained from the stress field distribution; for the dynamic contact pressure data, an autoencoder neural network with three hidden layers is used to extract features, and the pressure and deformation mapping relationship data are obtained from the feature extraction results; based on the pressure and deformation mapping relationship data, the fourth-order Runge-Kutta method is used to solve the copper foil thickness differential equation group, and the thickness spatiotemporal distribution characteristic data are obtained from the numerical solution.

[0035] For example, according to the production line operation data collected by the speed sensor, the second-order Newmark time integration algorithm is used to solve the dynamic equation M(d 2u / dt2)+C(du / dt)+Ku=F(t) is solved to obtain the displacement field distribution of the contact process between the pressure roller and the copper foil. The stress-strain constitutive equation σ=Dε+ηdε / dt is used to calculate the stress field distribution during the deformation of the copper foil. The material deformation characteristics are described by the viscoelastic model, and the dynamic contact pressure distribution is obtained from the stress field data. σ is the strain tensor, D is the elastic modulus matrix, ε is the strain tensor, and η is the viscosity coefficient. An autoencoder neural network with three hidden layers is used to extract features from the dynamic pressure distribution data. The number of neurons in each hidden layer is 64, 32, and 16, respectively. The mapping relationship between pressure and deformation is obtained through the encoder. According to the pressure-deformation mapping relationship, the fourth-order Runge-Kutta method is used to solve the copper foil thickness differential equation group dx / dt=f(x,t), where x is the thickness and the integral step t is taken as one-fourth of the control period, and the spatiotemporal distribution characteristics of the thickness are obtained from the numerical solution. The temperature field distribution during rolling is solved by the thermomechanical coupling equations, the temperature stress is calculated by the thermoelastic constitutive relationship, and the deformation correction of the copper foil is obtained from the coupling field data. The Kalman filter state equation is constructed according to the deformation correction, and the pressure roller parameters are estimated and predicted by the state observation equation. The adaptive pole configuration controller is used to adjust the pressure roller parameters in real time, and the controller transfer function adopts the second-order form G(s)=ωn2 / (s 2+2ξωns+ωn2), s is the complex frequency variable in the Laplace transform, and the pressure compensation parameters are obtained from the closed-loop response. During the dynamic rolling process of copper foil, the second-order Newmark time integration algorithm is used to numerically solve the dynamic equation, where the mass matrix M reflects the density distribution of the copper foil material, the damping matrix C represents the internal resistance of the material, the stiffness matrix K describes the elastic properties of the material, the external force vector F(t) represents the pressure roller force, and u is the displacement vector. When the production line speed is 20 meters per minute, the integration time step is 0.001 seconds, and the calculation results show that the maximum displacement of the pressure action area reaches 0.15 mm. During the deformation of the material, the viscoelastic constitutive equation is used to describe the stress-strain relationship, where the elastic modulus D is 110 GPa and the viscosity coefficient η is 0.8. The calculated pressure distribution in the dynamic contact area shows typical asymmetric characteristics, with the maximum pressure in the inlet area being 180 MPa and the outlet area being 150 MPa. The feature extraction is performed by autoencoder neural network. The 64 neurons in the input layer correspond to the pressure sensor data. The three hidden layers in the middle are set with 64, 32 and 16 neurons respectively. After 2000 iterations of training, the reconstruction error is reduced to 0.025. The fourth-order Runge-Kutta method is used to solve the differential equation of copper foil thickness, and the integral step is set to 0.00025 seconds. Calculations show that when the rolling speed fluctuates by ±5%, the thickness change shows obvious periodic characteristics, and the fluctuation period is about 0.1 seconds. At the same time, the thermal-mechanical coupling calculation shows that the temperature in the rolling zone rises by about 25 degrees Celsius, resulting in a thermal expansion deformation of 0.03 mm. This thermal effect causes the actual thickness deviation to increase by 15% compared with the pure mechanical calculation result. In the state equation of the Kalman filter, the state transfer matrix A is obtained by system identification, the control input matrix B is determined by the actuator characteristics, and the covariance matrices of the process noise w(k) and the measurement noise v(k) are Q=diag[0.01,0.01] and R=0.04 respectively. The filtering results show that the root mean square error of pressure prediction is 0.8 MPa, and the prediction lead time is 0.05 seconds. The characteristic frequency ωn of the adaptive pole configuration controller is set to 20 Hz, and the damping ratio ξ is 0.7. After closed-loop control, the pressure fluctuation amplitude is reduced by 60%, and the thickness control accuracy is improved to ±0.5 microns. This multi-physics field coupling method comprehensively considers mechanical deformation, thermal effects and dynamic response characteristics, and realizes precise control of copper foil thickness during high-speed rolling. Measured data show that under the condition of a production line speed of 20 meters per minute, the standard deviation of the copper foil thickness uniformity index is reduced from 2.8 microns to 1.2 microns, and the dynamic rolling accuracy is significantly improved.

[0036] S104. Apply the optimal process parameter combination and equipment control strategy for achieving uniform pressure distribution to the copper foil production line. Collect process parameter and equipment operation status data in real time during the composite copper foil production process, and dynamically predict the pressure distribution under the current working conditions. If the prediction result meets the requirements of the quantitative evaluation index of pressure distribution uniformity, maintain the current process parameters and equipment control strategy unchanged.

[0037] According to the real-time data stream collected by the pressure sensor, the Kalman filter of the state equation is used to perform state estimation, and a preprocessed data sequence is obtained from the recursive formula; for the preprocessed data sequence, the memory unit and hidden state in the long short-term memory network are used to extract the parameter evolution law, and the pressure distribution prediction data is obtained from the output layer; according to the pressure distribution prediction data, a Gaussian kernel function is used to construct a pressure distribution probability density curve, the kernel function bandwidth parameter is determined by the cross-validation method, and the pressure distribution variance value is calculated from the probability density function; for the pressure distribution prediction data and process parameters, a matrix decomposition method is used to perform singular value decomposition, the parameter importance sequence is obtained by eigenvalue sorting, and the parameter sensitivity coefficient is extracted from the eigenvector matrix.

[0038] Exemplarily, based on the real-time data stream collected by the pressure sensor and the equipment status monitor, the Kalman filter of the state equation is used for state estimation, and the data sequence is smoothed by the recursive formula θ(k)=θ(k-1)+K(k)[y(k)-θ(k-1)], where θ(k) is the state estimate at the current moment, θ(k-1) is the state estimate at the previous moment, K(k) is the weight factor of 1 / k+1, and y(k) is the measured value at the current moment. The time series characteristics of the process parameters are obtained from the preprocessed data. A long short-term memory network containing two hidden layers is used to dynamically predict the time series characteristics of the process parameters, with the number of hidden layer units being 128 and 64 respectively. The parameter evolution law is extracted through the memory unit ct and the hidden state ht, and the pressure distribution prediction data is obtained from the output layer. According to the pressure distribution prediction data, the pressure distribution probability density curve is constructed using the Gaussian kernel function, the kernel function bandwidth parameter h is determined by the cross-validation method, and the pressure distribution variance is calculated from the probability density function. The matrix decomposition method is used to perform singular value decomposition U=WSV^T on the process parameters and pressure distribution data, and the parameter importance ranking is obtained by eigenvalue sorting, and the parameter sensitivity coefficient is extracted from the eigenvector matrix. If the pressure distribution variance is less than the variance threshold, the autoregressive moving average method is used to test the stability of the current process parameters, and the parameter stability is judged by the unit root test, and the parameter maintenance strategy is determined from the test results. The principal component regression method is used to establish the process parameter optimization equation group Y=XB+E, where Y is the pressure uniformity index, X is the process parameter matrix, and the parameter compensation weight is obtained from the regression coefficient B. In the real-time monitoring of the composite copper foil production line, the pressure sensor collects pressure data at a sampling frequency of 100 Hz, and the equipment status monitor collects speed, temperature, vibration and other parameters. The Kalman filter is used for state estimation. The state vector contains three components: pressure, velocity and temperature. The state transfer matrix A is obtained through system identification. The process noise covariance Q = diag[0.01, 0.02, 0.015] and the measurement noise covariance R = diag[0.02, 0.03, 0.025]. The forgetting factor λ of the recursive least squares method is set to 0.95, and the data smoothing window length is 50 sampling points. The long short-term memory network adopts a two-layer hidden layer structure, with 128 memory units in the first layer and 64 memory units in the second layer. The input layer contains 10 process parameter nodes. 5000 sets of historical data are used for network training, and each set of data contains a sequence of 100 time steps. After training, the root mean square error of the prediction on the validation set is 0.085, the prediction lead time is 0.5 seconds, and the correlation coefficient reaches 0.92. The Gaussian kernel function is used for the probability density estimation of pressure distribution, and the kernel function bandwidth is determined to be 0.8 through five-fold cross validation. The calculation results show that when the pressure distribution is uniform, the probability density curve presents a single-peak symmetrical feature with a variance less than 0.1; when the pressure is uneven, the curve shows a skewed or bimodal feature, and the variance increases to more than 0.25.Perform singular value decomposition on the process parameter matrix to obtain the eigenvalue sequence.

[0039] [8.52, 4.31, 2.15, 1.08, 0.54, 0.27], indicating that the cumulative contribution rate of the first three principal components reached 85%. Eigenvector analysis showed that rolling speed, temperature and pressure were the parameters that most significantly affected uniformity, with sensitivity coefficients of 0.82, 0.65 and 0.58, respectively. In the parameter stability test, the augmented Dickey-Fuller test method was used, and the test statistic ADF was -3.85, which was less than the critical value of -3.43 at the 1% significance level, indicating that the parameter sequence had good stability. The results of principal component regression analysis showed that there was a significant nonlinear relationship between the pressure uniformity index and the process parameters, and the multivariate correlation coefficient R 2 Reached 0.89. The regression coefficient matrix indicates that the compensation weight of the pressure parameter is 0.45, the compensation weight of the speed parameter is 0.32, and the compensation weight of the temperature parameter is 0.23, which provides a quantitative basis for the precise regulation of process parameters. In actual production, when the pressure distribution variance is less than 0.1 and the parameter sequence passes the stationarity test, the current process parameters are maintained unchanged; when an abnormality is detected, dynamic adjustment is performed according to the parameter compensation weight, which significantly improves the uniformity and stability of the pressure distribution.

[0040] S105. If the predicted pressure distribution does not meet the uniformity requirement, the process parameters and equipment control strategy are dynamically adjusted through the reinforcement learning algorithm to make the pressure distribution gradually uniform, and the optimized process parameter combination and control strategy are updated to the knowledge base.

[0041] A working condition feature space is constructed according to the state vector, and a working condition state representation is obtained by principal component analysis. A parameter optimization function is constructed according to the working condition state representation by using the policy gradient method, and a policy gradient value is obtained by Monte Carlo sampling. An adaptive learning rate optimizer is used to perform iterative calculations according to the policy gradient value, and a parameter adjustment amount is obtained by Taylor expansion. A neural network model is constructed according to the parameter adjustment amount by using the deep deterministic policy gradient method, and an optimal process parameter combination is obtained from the policy network.

[0042] Exemplarily, according to the deviation between the pressure distribution prediction value and the uniformity threshold, the state vector

[0043] [x1,x2,...,xn] constructs the working condition feature space, reduces the dimension of the state data through principal component analysis, and selects the eigenvalues ​​corresponding to the eigenvalues ​​with cumulative contribution rates greater than 95% to construct the working condition state representation. Based on the working condition state representation, the policy gradient method is used to construct the parameter optimization function θ is the policy parameter, Rτ is the total return of trajectory τ, and the policy gradient is calculated by 500 Monte Carlo samplings, and the parameter adjustment direction is obtained from the historical database. The adaptive learning rate optimizer is used to iteratively update the process parameters, and the parameter adjustment amount Δθ=f'(x) / 1! +f”(x) / 2! +f”'(x) / 3! is calculated by the third-order Taylor expansion. The reward value r(s,a) is calculated for the degree of improvement in pressure uniformity. The time difference objective function is constructed according to the reward value, and the deep deterministic policy gradient method is used to reinforce the learning of the compensation parameters. The critic network adopts a three-layer structure [256,128,64], and the actor network adopts a symmetric structure [64,128,256]. The experience replay buffer pool is used to store the optimization trajectory, and the high-value experience is selected for learning through the priority sampling method. The optimal process parameter combination is obtained from the trained policy network. According to the optimal process parameter combination, the knowledge graph structured method is used to build the process knowledge base, and the optimization experience is described by semantic triples (working conditions, parameters, effects), and the process parameter control specifications are obtained from rule reasoning. In the composite copper foil pressure optimization control, the working condition feature space contains 10 state variables, including pressure unevenness, velocity fluctuation rate, temperature distribution, vibration amplitude, etc. The principal component analysis results show that the eigenvalue sequence is

[0044] [4.82, 2.56, 1.33, 0.65, 0.34, 0.15, 0.08, 0.04, 0.02, 0.01], the cumulative contribution rate of the first four principal components reached 95.2%, and these principal components were selected to construct a low-dimensional working condition representation. In the parameter optimization process, the policy gradient method uses the Gaussian policy function πθ(a|s) = N(μθ(s), σ2), where the mean function μθ(s) is fitted by a three-layer neural network. The policy gradient is calculated by 500 Monte Carlo samplings. The sampling results show that the pressure parameter gradient is 0.85, the speed parameter gradient is 0.62, and the temperature parameter gradient is 0.43, indicating that pressure regulation is most sensitive to uniformity improvement. The adaptive learning rate optimization uses the third-order Taylor expansion to approximate the parameter adjustment amount, and the derivatives of each order of the expansion are f'(x) = 2.5, f" (x) = -0.8, and f"'(x) = 0.15. The calculated adjustment of the pressure parameter is 0.42 MPa, the speed parameter is 0.25 m / min, and the temperature parameter is 0.18 degrees Celsius. The reward function is designed as r(s,a)=-w1σ2-w2|Δa|, where σ2 is the pressure variance, Δa is the parameter adjustment, and the weight coefficients w1=0.7 and w2=0.3. The deep deterministic policy gradient network adopts a symmetric structure. The three hidden layers of the critic network are 256, 128, and 64 neurons respectively, and the actor network adopts a mirror structure. The capacity of the experience replay buffer pool is set to 10,000, and the priority sampling probability is proportional to the temporal difference error. After 5,000 rounds of training, the network converges to a stable strategy, and the accuracy of pressure uniformity prediction reaches 92.5%. The knowledge graph uses semantic triples to store optimization experience, and the triple structure is (operating condition type, parameter combination, optimization effect). For example, when the working condition type is "high speed and light load", the optimal parameter combination is pressure 10.5 MPa, speed 25 m / min, temperature 155 degrees Celsius, and the corresponding pressure unevenness is less than 5%. The parameter configuration specifications of typical working conditions such as "high speed-light load-low pressure" are extracted through rule reasoning to form a complete process knowledge base. Practical applications show that the parameter optimization method based on reinforcement learning and knowledge graph significantly improves the pressure uniformity of composite copper foil, and the unevenness is reduced from the original 12% to 4.5%, and the process stability is also significantly improved.

[0045] S106. When the process requirements change, historical optimization cases similar to the current process requirements are extracted from the knowledge base as the initial optimization plan, and then optimized through the reinforcement learning algorithm to respond to the changes in process requirements.

[0046] A feature vector is constructed according to process parameters, equipment parameters and pressure distribution, the similarity value between the feature vectors is calculated by cosine similarity, and the Euclidean distance hierarchical clustering method is used to obtain the closest process solution; for the closest process solution, a dual network structure including a critic network and an actor network is constructed, and the critic network calculates the parameter gradient value through the cross entropy loss function; a Gaussian kernel function is used to spatially map the parameter gradient value, and the kernel function parameters are determined through grid search, and the pressure distribution prediction value is obtained from the regression equation; a reward function including pressure deviation and variance is constructed according to the pressure distribution prediction value, and the parameters are iteratively updated using a deep substitution strategy optimization algorithm, and the optimal parameter combination is obtained through random exploration.

[0047] For example, according to the process requirement change data, the cosine similarity calculation formula is constructed using the feature vector [process parameters, equipment parameters, pressure distribution]

[0048] s im(A,B)=Σ(Ai×Bi) / sqrt(Σ(Ai 2 )×Σ(Bi 2 ), Ai is the i-th feature of vector A, Bi is the i-th feature of vector B, similar cases are grouped by Euclidean distance hierarchical clustering method, and the most similar process solution is obtained from the cluster center. According to the most similar process solution, the soft actor-critic algorithm is used to construct a dual network structure. The critic network contains three layers [256, 128, 1], and the actor network contains three layers

[0049] [128,64,32], the parameter gradient is calculated by the cross entropy loss function. The Gaussian kernel function is used to map the parameter space, and the kernel parameter σ is determined by grid search. The pressure distribution prediction value is obtained from the regression equation. According to the pressure distribution prediction value, the reward function r(s,a)=-k1×D(p)+k2×var(p) is constructed, where D(p) represents the pressure deviation and var(p) represents the variance. The parameter adjustment amount is calculated by the gradient ascent method. The deep substitution strategy optimization algorithm is used to iteratively update the parameters. The ε-greedy strategy is used to perform random exploration in the parameter space to obtain the optimal parameter combination from the exploration data. According to the optimal parameter combination, the knowledge distillation method is used to integrate the new optimization experience into the knowledge base, and the soft label cross entropy is used to realize the fusion update of new and old knowledge, and the process parameter optimization specification is formed from the updated knowledge base. In the knowledge base case retrieval, the feature vector contains 12 process parameters, such as speed, temperature, pressure, etc., 8 equipment parameters, such as equipment vibration, roller gap, etc., and 16 pressure distribution feature values. When the process requirements change from "high speed and light load" to "medium speed and heavy load", the cosine similarity distribution with the historical cases is calculated to be [0.92, 0.85, 0.78, 0.65]. A threshold of 0.8 is used for clustering, and the case with a similarity of 0.92 is selected as the initial value for optimization. In the dual network structure of the soft actor-critic algorithm, the critic network adopts a three-layer structure [256, 128, 1], uses the ReLU activation function, and outputs Q value estimation; the actor network adopts a three-layer structure [128, 64, 32] and outputs action probability distribution. The calculation results of the cross entropy loss function show that the pressure parameter gradient is 0.45, the speed parameter gradient is 0.32, and the temperature parameter gradient is 0.28, indicating that pressure regulation is the most critical. The parameter space mapping uses a Gaussian kernel function, and the kernel parameter σ=0.8 is determined by grid search. The kernel principal component regression prediction shows that the uniformity of pressure distribution has a nonlinear relationship with the process parameters, and the correlation coefficient R 2Reached 0.88. The weight coefficients k1=0.6 and k2=0.4 in the reward function, and the maximum parameter adjustment was calculated as follows: pressure adjustment ±0.5 MPa, speed adjustment ±2 m / min, and temperature adjustment ±5 degrees Celsius. The deep substitution strategy optimization uses the ε-greedy strategy for parameter exploration, and the exploration probability ε decays linearly from the initial 0.3 to 0.05. A total of 500 random samplings were performed in the parameter space, and the optimal parameter combination obtained was pressure 12.5 MPa, speed 18 m / min, and temperature 165 degrees Celsius. This set of parameters improved the pressure distribution uniformity index by 45%. During the knowledge distillation process, the soft label with temperature parameter T=3 was used for knowledge transfer, and the cross entropy loss was reduced from the initial 1.25 to 0.15. The fusion of new and old knowledge uses exponential sliding average to update the knowledge base parameters, and the update coefficient α=0.9. The updated knowledge base contains 350 sets of optimized parameter combinations for typical working conditions, covering 95% of process change scenarios. In this way, the process parameter optimization response time was shortened from 15 minutes to 2 minutes, and the optimization efficiency was significantly improved. The optimization results showed that when the process requirements changed, the pressure distribution uniformity index always remained within the range of ±5%, meeting the requirements of high-precision copper foil production.

[0050] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A control method for a pressure control device for uniformly distributing a multi-layer composite copper foil, characterized in that: The method comprises: Obtaining material property parameters of each layer of the multi-layer composite copper foil and establishing a composite copper foil material property database; By using the finite element analysis method, combined with the multi-layer structure and material property data of the composite copper foil, the stress distribution and deformation of each layer of the copper foil under different pressure conditions are analyzed to obtain a quantitative evaluation index for the uniformity of pressure distribution. According to the quantitative evaluation index of pressure distribution uniformity, the mapping relationship between pressure distribution uniformity and process parameters and equipment mechanical characteristics is analyzed to obtain the optimal process parameter combination and equipment control strategy to achieve uniform pressure distribution; The optimal process parameter combination and equipment control strategy for achieving uniform pressure distribution are applied to the copper foil production line. During the production of composite copper foil, the process parameters and equipment operation status data are collected in real time to dynamically predict the pressure distribution under the current working conditions. If the prediction results meet the requirements of the quantitative evaluation index of pressure distribution uniformity, the current process parameters and equipment control strategy are maintained unchanged. If the predicted pressure distribution does not meet the uniformity requirements, the process parameters and equipment control strategies are dynamically adjusted through the reinforcement learning algorithm to make the pressure distribution gradually uniform, and the optimized process parameter combination and control strategy are updated to the knowledge base; When process requirements change, historical optimization cases similar to current process requirements are extracted from the knowledge base as the initial optimization plan, which is then optimized through a reinforcement learning algorithm.

2. The method according to claim 1, characterized in that The method of obtaining the material property parameters of each layer of the multi-layer composite copper foil and establishing a composite copper foil material property database comprises: Obtain scanning electron microscope photos of each layer of the composite copper foil structure; The surface morphology point cloud data is obtained based on the SEM photo measurement data; The point cloud data is processed by a Gaussian filter to obtain a surface profile value; A cyclic load is applied to the surface of the composite copper foil according to the surface profile value, the Young's modulus and hardness value of each layer are calculated through the load-displacement curve, and a composite copper foil material property database is established.

3. The method according to claim 1, characterized in that The finite element analysis method is combined with the multilayer structure and material property data of the composite copper foil to analyze the stress distribution and deformation of each layer of the copper foil under different pressure conditions, and obtain quantitative evaluation indicators of pressure distribution uniformity, including: A three-dimensional finite element geometric structure is established according to the number of layers and material density of the composite copper foil, a tetrahedral mesh is generated through a mesher, and the surface morphology characteristic values ​​of the copper foil are obtained from the scanned data; The contact area is calculated by using the surface morphology characteristic value of the copper foil, and the interlayer friction stress value is obtained by using the Coulomb friction law; The node displacement field is calculated according to the interlayer friction stress value by using the virtual work principle, and if the local pressure deviation of the node displacement field exceeds a preset threshold, a residual iteration method is used to obtain a corrected node displacement field; The modified node displacement field is used to calculate the unit strain energy through the strain energy density function, the integral value of the unit strain energy of each layer is obtained through three-point Gaussian integration, and the quantitative evaluation index of the pressure distribution uniformity is calculated using the strain energy integral value.

4. The method according to claim 1, characterized in that According to the quantitative evaluation index of pressure distribution uniformity, the mapping relationship between pressure distribution uniformity and process parameters and equipment mechanical characteristics is analyzed to obtain the optimal process parameter combination to achieve uniform pressure distribution, including: A median filter is used to perform noise reduction on the pressure distribution data collected by the pressure sensor, and the noise-reduced data is decomposed to obtain a pressure distribution uniformity feature vector; According to the pressure distribution uniformity characteristic vector, a correlation matrix between the pressure distribution uniformity characteristic vector and the process parameters is established by using the Pearson correlation coefficient method, and a parameter coupling coefficient matrix is ​​obtained by performing nonlinear mapping on the correlation matrix by using a radial basis kernel function support vector machine; For the parameter coupling coefficient matrix, an orthogonal design method is used to generate a process parameter test plan, and the test plan is evaluated and calculated by a hierarchical analysis method to obtain a parameter optimization data set; The parameter optimization data set is trained using a neural network based on error back propagation. If the weight matrix meets the convergence condition after training, the process parameters are combined and optimized through a genetic algorithm of single-point crossover and Gaussian mutation to obtain the optimal process parameter combination for achieving uniform pressure distribution.

5. The method according to claim 1, characterized in that The equipment control strategy includes: collecting copper foil thickness data in real time through an online thickness gauge on a multi-layer composite copper foil production line, processing the collected data through a data analysis method to obtain a spatial distribution map of the copper foil thickness, judging the uniformity of the thickness distribution, and adjusting the pressure roller pressure of the corresponding area when abnormal thickness is found in a local area.

6. The method according to claim 1, characterized in that The equipment control strategy includes: collecting the pressure data of the copper foil during the production process in real time through a pressure sensor, establishing a mapping relationship between pressure and thickness through a machine learning algorithm, predicting the deformation of the copper foil under different pressure conditions, and guiding the dynamic adjustment of the pressure roller to suppress the thickness difference caused by anisotropy.

7. The method according to claim 1, characterized in that The equipment control strategy includes: establishing a mathematical model of the dynamic contact process between the pressure roller and the copper foil according to the production line speed and the deformation characteristics of the copper foil material, using a multi-physical field coupling simulation method to simulate the transient pressure distribution under high-speed motion, obtaining the variation pattern of the copper foil thickness with time and position, and adjusting the operating parameters of the pressure roller in real time to ensure the stability of the copper foil thickness under dynamic conditions.

8. The method according to claim 1, characterized in that The optimal process parameter combination and equipment control strategy for achieving uniform pressure distribution are applied to the copper foil production line, and process parameters and equipment operation status data are collected in real time during the composite copper foil production process, and the pressure distribution under the current working conditions is dynamically predicted. If the prediction result meets the requirements of the quantitative evaluation index of pressure distribution uniformity, the current process parameters and equipment control strategy are maintained unchanged, including: According to the real-time data stream collected by the pressure sensor, the Kalman filter of the state equation is used to perform state estimation and obtain the preprocessed data sequence from the recursive formula; For the preprocessed data sequence, the memory unit and hidden state in the long short-term memory network are used to extract the parameter evolution law, and the pressure distribution prediction data is obtained from the output layer; According to the pressure distribution prediction data, a pressure distribution probability density curve is constructed using a Gaussian kernel function, a kernel function bandwidth parameter is determined by a cross-validation method, and a pressure distribution variance value is calculated from the probability density function; A matrix decomposition method is used to perform singular value decomposition on the pressure distribution prediction data and process parameters, and a parameter importance sequence is obtained by eigenvalue sorting, and parameter sensitivity coefficients are extracted from the eigenvector matrix.

9. The method according to claim 1, characterized in that: If the predicted pressure distribution does not meet the uniformity requirement, the process parameters and equipment control strategy are dynamically adjusted through the reinforcement learning algorithm to make the pressure distribution gradually uniform, and the optimized process parameter combination and control strategy are updated to the knowledge base, including: The working condition feature space is constructed according to the state vector, and the working condition state representation is obtained through the principal component analysis method; A policy gradient method is used to construct a parameter optimization function for the working condition characterization, and a policy gradient value is obtained through Monte Carlo sampling; An adaptive learning rate optimizer is used to iteratively calculate the policy gradient value, and a parameter adjustment amount is obtained through a Taylor expansion; A deep deterministic policy gradient method is used to construct a neural network model for the parameter adjustment amount, and an optimized process parameter combination is obtained from the policy network.

10. The method according to claim 1, characterized in that When the process requirements change, historical optimization cases similar to the current process requirements are extracted from the knowledge base as the initial optimization solution, and then optimized by the reinforcement learning algorithm, including: Constructing feature vectors according to process parameters, equipment parameters and pressure distribution, calculating similarity values ​​between the feature vectors by cosine similarity, and using Euclidean distance hierarchical clustering method to obtain the most similar process solution; For the most similar process solution, a dual network structure including a critic network and an actor network is constructed, wherein the critic network calculates a parameter gradient value through a cross entropy loss function; A Gaussian kernel function is used to spatially map the parameter gradient value, the kernel function parameters are determined by grid search, and the pressure distribution prediction value is obtained from the regression equation; A reward function including pressure deviation and variance is constructed according to the pressure distribution prediction value, and the parameters are iteratively updated using a deep substitution strategy optimization algorithm, and the optimal parameter combination is obtained through random exploration.

Images