Self-adaptive corona printing typesetting optimization system based on neural network model

Abstract

The invention provides an adaptive corona printing typesetting optimization system based on a neural network model, and relates to the technical field of data processing, and the system comprises a detection module which is used for carrying out the three-dimensional shape scanning of the surface of a PET film after corona treatment through a confocal microscope array and a laser displacement sensor, and obtaining a surface fluctuation characteristic value set; and the feature processing module is used for calculating a surface micro-roughness change value before and after processing according to the surface fluctuation feature value set, and extracting regular bulge structure feature parameters generated by corona processing, including bulge height distribution, spacing density and a surface real contact area increment ratio. According to the invention, the surface topography data of the PET film is accurately obtained through the detection module, deep analysis and intelligent calculation are carried out through the feature processing module, the dynamic prediction module and the like, the circuit parameters are optimized in combination with adaptive typesetting, and closed-loop optimization is formed by using feedback control, so that the quality of the printed circuit and the production efficiency are effectively improved, and the adaptive capability and reliability of the system are enhanced.

Description

Technical Field

[0001] The invention relates to the technical field of data processing, in particular to an adaptive corona printing typesetting optimization system based on a neural network model. Background Art

[0002] The traditional fixed typesetting scheme has obvious defects in corona printing on flexible substrates:

[0003] First, the inability to adjust the layout based on the microscopic topography of the flexible substrate after corona treatment results in poor compatibility between the printed circuit and the substrate surface. For example, after corona treatment, microscopic bumps appear on the surface of PET film. Traditional printing methods still follow the same fixed pattern, and the circuits are prone to breakage due to stress concentration at these bumps.

[0004] Second, it can easily cause circuit breakage or resistance fluctuations, failing to meet the quality requirements of high-precision circuits. For example, when printing fine sensor circuits, traditional solutions fail to consider surface topography, leading to significant resistance fluctuations in some sections of the circuit, resulting in inaccurate sensor data collection.

[0005] In addition, the electronics manufacturing industry continues to increase its requirements for product quality and performance, and traditional solutions cannot adapt to the industry's development needs for high-quality printed conductive circuits on flexible substrates.

[0006] For example, with the rapid development of wearable devices, the circuits printed by traditional solutions are unstable in quality and cannot support the long-term stable operation of the devices. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide an adaptive corona printing typesetting optimization system based on a neural network model, which improves the accuracy and quality of printed circuits and ensures the stable performance of electronic equipment.

[0008] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0009] In a first aspect, an adaptive corona printing typesetting optimization system based on a neural network model comprises:

[0010] The detection module is used to perform three-dimensional topography scanning on the surface of the corona-treated PET film using a confocal microscope array and a laser displacement sensor to obtain a set of surface relief feature values;

[0011] The feature processing module is used to calculate the change in surface micro-roughness before and after treatment based on the surface relief feature value set, and to extract the characteristic parameters of the regular protrusion structure produced by corona treatment, including protrusion height distribution, spacing density, and surface real contact area increment ratio;

[0012] A dynamic prediction module uses a dual-channel convolutional neural network to analyze surface topography parameters and preset circuit layout information, outputting the fracture probability of each edge and the resistance fluctuation coefficient;

[0013] Adaptive layout module, which is used to dynamically adjust line parameters based on fracture probability and resistance fluctuation coefficient through a path planner including swarm incremental learning algorithm;

[0014] The feedback control module is used to monitor the resistance value of the printed circuit in real time through an online impedance meter. When the relative deviation between the actual resistance value and the theoretical value is ≥5%, the parameter feedback mechanism is triggered, that is, the coordinates of the abnormal area are fed back to the detection module for local re-measurement, and the re-measurement data is input into the dynamic prediction module to update the weight matrix.

[0015] Furthermore, based on the surface undulation characteristic value set, the surface micro-roughness change value before and after treatment is calculated, and the characteristic parameters of the regular protrusion structure produced by corona treatment are extracted, including the protrusion height distribution, spacing density and surface real contact area increment ratio, including:

[0016] The surface relief feature value set obtained after the corona treatment is integrated with the PET film surface three-dimensional topography data and the corresponding feature value set before the corona treatment to form a data comparison group;

[0017] The surface height data before and after treatment in the data comparison group were analyzed, and the root mean square roughness of the surface height before and after treatment was calculated to obtain the surface micro-roughness change value;

[0018] Based on the surface micro-roughness change value, the three-dimensional topography image constructed by the surface relief feature value set is subjected to image morphological processing operations. A threshold is set according to the height difference between the protrusion and the substrate to separate the protrusion structure from the surface topography, forming an independent protrusion structure area.

[0019] Traverse each raised structure and measure the vertical distance between the highest point and the base plane, i.e., the raised height. Summarize the height values ​​of each raised structure, group them according to a certain height interval, and calculate the proportion of the number of raised structures in each interval to the total number of raised structures to generate a raised height distribution table.

[0020] Set an effective distance threshold, count the number of adjacent protrusions within the threshold range for each protrusion structure, and obtain the average spacing density;

[0021] According to the average spacing density and surface undulation characteristic value set, three-dimensional models of the PET film surface before and after treatment were constructed using three-dimensional modeling software. By comparing the real surface contact areas in the three-dimensional models, the incremental ratio of the real surface contact areas before and after treatment was calculated.

[0022] Furthermore, a dual-channel convolutional neural network is used to analyze surface topography parameters and preset circuit layout information, outputting the fracture probability of each edge and the resistance fluctuation coefficient, including:

[0023] Collect surface topography parameters and preset circuit layout information, and convert the surface topography parameters into tensor form and the circuit layout information into a two-dimensional matrix;

[0024] Normalizing the converted surface topography parameters and the preset circuit layout information, mapping the converted surface topography parameters and the preset circuit layout information to the [0, 1] interval to obtain normalized data;

[0025] Based on the normalized data, a dual-channel convolutional neural network is constructed, and the normalized data is divided into a training set and a validation set. The training set is input into the dual-channel convolutional neural network to extract the multi-scale features of the surface morphology parameters and the local features of the line connection;

[0026] The multi-scale features of the surface morphology parameters and the local features of the line connection are spliced ​​to obtain the combined features, which are then processed through the fully connected layer to learn the relationship between the features;

[0027] The combined features processed by the fully connected layer are input into the output layer of the dual-channel convolutional neural network. One neuron is used in conjunction with a normalized exponential function to calculate the fracture probability of each edge of the printed circuit. The fracture probability output value is in the interval [0, 1] and the sum of all edge probabilities is 1. Another neuron is used in conjunction with a linear activation function to obtain the resistance fluctuation coefficient.

[0028] Define the cross entropy and mean square error loss functions, calculate the gradient of the loss function with respect to the parameters of the two-channel convolutional neural network, and update the parameters of the two-channel convolutional neural network at the same time;

[0029] Use the validation set to evaluate the parameters of the dual-channel convolutional neural network, and adjust the dual-channel convolutional neural network structure and hyperparameters based on the evaluation results until the preset number of iterations is reached to obtain a trained and optimized dual-channel convolutional neural network;

[0030] The normalized surface morphology parameters and preset circuit layout information are input into the trained and optimized dual-channel convolutional neural network, and the fracture probability of each edge and the resistance fluctuation coefficient are output.

[0031] Furthermore, the dual-channel convolutional neural network includes:

[0032] The first channel: inputs the surface morphology parameter tensor in the training set and extracts the multi-scale features of the surface morphology parameters through a three-layer convolutional network;

[0033] Second channel: Input a two-dimensional matrix of preset line layout information, aggregate node neighborhood information through the graph attention network, and generate edge feature vectors.

[0034] Furthermore, based on the fracture probability and resistance fluctuation coefficient, the line parameters are dynamically adjusted through a path planner that includes a swarm incremental learning algorithm, including:

[0035] Build a path planner based on the swarm incremental learning algorithm and set the parameters of the swarm incremental learning algorithm, including the swarm size, learning rate, and incremental learning step size;

[0036] Organize the preset line layout information, determine the line connection relationship and node location, and determine the adjustment range of line parameters, including the minimum and maximum values ​​of line width, thickness and spacing parameters;

[0037] Based on the preset line layout information and the adjustment range of the line parameters, an initial group of line parameter combinations is randomly generated, each combination representing a line parameter setting scheme;

[0038] For each line parameter combination in the group, the corresponding adaptation index is calculated by combining the fracture probability and resistance fluctuation coefficient;

[0039] At the beginning of each iteration, new line performance data is obtained from the actual printing process, and the new line performance data is combined with the line parameters in the current group and the corresponding adaptation index and input into the path planner to update the path planner to obtain an updated path planner;

[0040] According to the updated path planner, the line parameter combination in the group is adjusted to obtain an adjusted line parameter combination, and the adaptation index of the adjusted line parameter combination is recalculated;

[0041] The process of acquiring new line performance data, updating the path planner, adjusting the line parameter combination, and recalculating the adaptation index is repeated until the preset maximum number of iterations is reached. The final line parameter combination is determined from the population based on the adaptation index of the line parameter combination;

[0042] According to the final circuit parameter combination, the actual parameters of the printed circuit are dynamically adjusted.

[0043] Furthermore, the line parameter combination includes line width, line thickness, line spacing, line length, conductor material, insulation material, solder joint quality and via parameters.

[0044] Furthermore, for each line parameter combination in the group, the corresponding adaptation index is calculated by combining the fracture probability and resistance fluctuation coefficient, including:

[0045] Obtain the actual parameter values ​​and corresponding adjustment ranges of each line unit's fracture probability, resistance fluctuation coefficient, line width, line thickness, and line spacing;

[0046] According to the fracture probability of each line unit, the contribution of the fracture probability part is determined; according to the resistance fluctuation coefficient of each line unit, the contribution of the resistance fluctuation coefficient is obtained;

[0047] For each line unit, determining the middle position of the line width and line thickness adjustment range, and determining the degree of deviation of the line width from the middle position and the degree of deviation of the line thickness from the middle position based on the middle position;

[0048] Determine the middle position of the line spacing adjustment range and calculate the degree of deviation of the actual line thickness from the middle position;

[0049] The contribution of the fracture probability, the contribution of the resistance fluctuation coefficient, the deviation of the line width from the middle position, the deviation of the line thickness from the middle position, and the deviation of the actual line thickness from the middle position are integrated to obtain the adaptation index of the line parameter combination.

[0050] Furthermore, the resistance value of the printed circuit is monitored in real time through an online impedance meter. When the relative deviation between the actual resistance value and the theoretical value is ≥5%, the parameter feedback mechanism is triggered, that is, the coordinates of the abnormal area are fed back to the detection module for local retesting, and the retested data is input into the dynamic prediction module to update the weight matrix, including:

[0051] Connect the online impedance meter to the printed circuit. The online impedance meter measures the circuit resistance value in real time according to the set sampling frequency, and continuously obtains the actual resistance value data of the printed circuit.

[0052] Obtain the theoretical resistance design value of the printed circuit from the database, and calculate the relative deviation between the actual resistance value measured in real time and the theoretical value;

[0053] Compare the relative deviation value with the set threshold of 5%. If the relative deviation value is less than 5%, the real-time resistance value monitoring will continue; if the relative deviation value is ≥5%, the parameter feedback mechanism will be triggered immediately.

[0054] When the parameter feedback mechanism is triggered, the coordinates of the abnormal resistance value area are located based on the layout information of the printed circuit and the measurement point position of the online impedance meter;

[0055] The detection module receives the coordinates of the abnormal resistance area, starts the local retest program, controls the detection equipment to move to the abnormal area, and retests the resistance of the printed circuit in the abnormal area at multiple points and angles. During the retest process, the environmental parameters are collected and the retested resistance data and environmental parameters are integrated into a retest data set.

[0056] The retest data set is input into the dynamic prediction module to analyze the impact of the retest data on the current dual-channel convolutional neural network. The weight matrix is ​​updated by recalculating the weights of each parameter in the dual-channel convolutional neural network, including the weights of the convolutional layer and the fully connected layer.

[0057] In a second aspect, a computing device includes:

[0058] one or more processors;

[0059] The storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the system.

[0060] According to a third aspect, a computer-readable storage medium stores a program, which implements the system when executed by a processor.

[0061] The above solution of the present invention includes at least the following beneficial effects:

[0062] The inspection module uses a confocal microscope array and laser displacement sensors to perform three-dimensional topographic scanning of the PET film surface, accurately capturing a set of surface relief characteristic values. Compared to traditional inspection methods, it can capture surface structural changes at the micron and even nanometer levels, ensuring a comprehensive understanding of the film's surface microstructure.

[0063] Based on a collection of surface undulation eigenvalues, the feature processing module deeply calculates surface micro-roughness variations and extracts key parameters such as bump height distribution, pitch density, and the ratio of the actual surface contact area increment. These quantitative indicators systematically analyze the impact of corona treatment on the film surface, providing a scientific basis for understanding the effects of corona treatment and evaluating surface quality. This allows production personnel to determine from multiple dimensions whether the film surface meets the requirements for printed circuits.

[0064] The dynamic prediction module uses a dual-channel convolutional neural network to combine surface topography parameters with pre-set circuit layout information to output the fracture probability of each edge and the resistance fluctuation coefficient. This module simulates circuit performance under different surface conditions and circuit layouts, predicting potential fracture risks and resistance fluctuations in advance. This model changes the traditional empirical judgment model and provides forward-looking guidance for circuit layout optimization through a data-driven approach.

[0065] The adaptive layout module dynamically adjusts circuit parameters based on breakage probability and resistance fluctuation coefficient using a path planner incorporating a swarm incremental learning algorithm. This optimization approach responds in real time to surface topography and circuit performance requirements, proactively avoiding circuit quality defects caused by film surface issues, such as breakage and resistance instability, during the circuit design phase. This improves the yield and reliability of printed circuits, while reducing design rework due to improper parameters and enhancing production efficiency.

[0066] The feedback control module monitors the printed circuit resistance in real time using an online impedance meter. When an anomaly occurs, a parameter feedback mechanism is triggered, enabling precise retesting of the abnormal area. This retested data is then used to update the weight matrix of the dynamic prediction module. This feedback mechanism forms a closed-loop system from production monitoring and problem location to model optimization, enabling the system to continuously learn and adapt to changes in actual production, continuously optimizing the neural network model and circuit layout strategy, ensuring that printed circuit quality remains stable and gradually improves over the long term, enhancing the company's market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of an adaptive corona printing typesetting optimization system based on a neural network model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0068] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0069] like Figure 1 As shown, an embodiment of the present invention provides an adaptive corona printing typesetting optimization system based on a neural network model, comprising:

[0070] Detection module 1 is used to perform three-dimensional topography scanning on the surface of the corona-treated PET film using a confocal microscope array and a laser displacement sensor to obtain a set of surface relief feature values;

[0071] Feature processing module 2 is used to calculate the change in surface micro-roughness before and after treatment based on the surface relief feature value set, and extract the characteristic parameters of the regular protrusion structure generated by corona treatment, including protrusion height distribution, spacing density and surface real contact area increment ratio;

[0072] Dynamic prediction module 3 is used to analyze surface morphology parameters and preset circuit layout information using a dual-channel convolutional neural network, and output the fracture probability of each edge and the resistance fluctuation coefficient;

[0073] Adaptive layout module 4, used to dynamically adjust line parameters according to the fracture probability and resistance fluctuation coefficient through a path planner including a swarm incremental learning algorithm;

[0074] The feedback control module 5 is used to monitor the resistance value of the printed circuit in real time through an online impedance meter. When the relative deviation between the actual resistance value and the theoretical value is ≥5%, the parameter feedback mechanism is triggered, that is, the coordinates of the abnormal area are fed back to the detection module for local re-measurement, and the re-measurement data is input into the dynamic prediction module to update the weight matrix.

[0075] In this embodiment of the present invention, the detection module uses a confocal microscope array and a laser displacement sensor to perform three-dimensional topographic scanning of the PET film surface, enabling high-precision, comprehensive acquisition of a set of surface relief feature values. Compared to traditional detection methods, this approach can capture more subtle surface topographic changes, ensuring accurate control of the film's surface condition.

[0076] Based on data collected by the detection module, the feature processing module not only calculates surface micro-roughness changes but also extracts characteristic parameters of the regular raised structures produced by corona treatment, such as raised height distribution, pitch density, and the ratio of the actual surface contact area increment. These parameters provide a deeper understanding of the impact of corona treatment on film surfaces, providing a key basis for optimizing printing layouts and enabling targeted adjustments to surface properties after corona treatment.

[0077] The dynamic prediction module uses a dual-channel convolutional neural network to analyze surface topography parameters and preset circuit layout information, outputting the fracture probability of each edge and the resistance fluctuation coefficient. Leveraging the neural network's powerful data analysis and prediction capabilities, this module can quickly and accurately assess the potential risks of different circuit layouts based on the current film surface state. Compared to traditional empirical judgment or simple model predictions, this module improves the accuracy and reliability of predictions, providing scientific quantitative indicators for circuit optimization.

[0078] The Adaptive Layout Module dynamically adjusts circuit parameters based on the fracture probability and resistance fluctuation coefficient output by the Dynamic Prediction Module, using a path planner incorporating a swarm incremental learning algorithm. This module responds in real time to film surface conditions and circuit risk assessment results, automatically optimizing circuit parameters to effectively prevent circuit breakage and resistance fluctuations caused by microscopic irregularities on the film surface. This ensures that printed circuits meet the stringent requirement of ≤10μm precision, improving printed circuit quality and stability.

[0079] The feedback control module monitors the printed circuit resistance in real time using an online impedance meter. When the actual resistance value deviates from the theoretical value by 5% or more, a parameter feedback mechanism is triggered, feeding the coordinates of the abnormal area back to the detection module for local retesting. This retested data is then fed into the dynamic prediction module to update the weight matrix. This closed-loop feedback mechanism enables real-time monitoring and dynamic adjustment of printed circuit quality, enabling timely identification and resolution of circuit quality issues, continuous optimization of the system model, and continuous adaptation to changes in the production process. This ensures long-term stability of printed circuit quality, improving production efficiency and product qualification rates.

[0080] In a preferred embodiment of the present invention, based on the surface relief characteristic value set, the surface micro-roughness change value before and after treatment is calculated, and the characteristic parameters of the regular protrusion structure generated by the corona treatment are extracted, including the protrusion height distribution, spacing density and surface real contact area increment ratio, which may include:

[0081] The surface relief feature value set obtained after the corona treatment is integrated with the PET film surface three-dimensional topography data and the corresponding feature value set before the corona treatment to form a data comparison group;

[0082] The surface height data before and after treatment in the data comparison group were analyzed, and the root mean square roughness of the surface height before and after treatment was calculated to obtain the surface micro-roughness change value;

[0083] Based on the surface micro-roughness change value, the three-dimensional topography image constructed by the surface relief feature value set is subjected to image morphological processing operations. A threshold is set according to the height difference between the protrusion and the substrate to separate the protrusion structure from the surface topography, forming an independent protrusion structure area.

[0084] Traverse each raised structure and measure the vertical distance between the highest point and the base plane, i.e., the raised height. Summarize the height values ​​of each raised structure, group them according to a certain height interval, and calculate the proportion of the number of raised structures in each interval to the total number of raised structures to generate a raised height distribution table.

[0085] Set an effective distance threshold, count the number of adjacent protrusions within the threshold range for each protrusion structure, and obtain the average spacing density;

[0086] According to the average spacing density and surface undulation characteristic value set, three-dimensional models of the PET film surface before and after treatment were constructed using three-dimensional modeling software. By comparing the real surface contact areas in the three-dimensional models, the incremental ratio of the real surface contact areas before and after treatment was calculated.

[0087] In an embodiment of the present invention, the detection module uses a confocal microscope array and a laser displacement sensor to obtain a set of surface undulation feature values ​​of the PET film after corona treatment. The three-dimensional surface topography data and corresponding feature value set of the PET film before corona treatment are collected in advance using the same detection equipment and methods and stored in a system database before the film undergoes corona treatment. The pre- and post-treatment data are matched and merged using the same coordinate system or measurement point locations. For example, if both measurements establish a coordinate system with a corner of the film as the origin, the feature values ​​of the corresponding points are linked based on the coordinate positions to form a data set containing the pre- and post-treatment data, i.e., a data comparison set.

[0088] Root mean square roughness (Rq) is an important indicator to measure the surface micro-roughness, and its calculation formula is: where z i is the surface height of the ith measurement point, is the average of the surface heights at all measurement points, and n is the total number of measurement points.

[0089] Specifically, surface height data before and after treatment were extracted from the comparison data set. For each data set, the average surface height of all measured points was calculated, and then the root mean square roughness (RMS) was calculated using the formula above. Finally, the RMS roughness before treatment was subtracted from the RMS roughness after treatment to obtain the surface micro-roughness change.

[0090] Image processing software is used to process the three-dimensional topography image constructed from a set of surface undulation feature values. Operations include erosion, dilation, opening, and closing. For example, the opening operation can first erode the image to remove some small bumps and noise, and then perform a dilation operation to restore the larger bump structure, making the bump structure clearer and more independent. A suitable threshold is set based on the surface micro-roughness change value and the actual height difference between the bump and the substrate. The portion of the three-dimensional topography image with a height greater than the threshold is determined to be a bump structure, and the portion with a height less than the threshold is determined to be the substrate. In this way, the bump structure is separated from the surface topography to form an independent bump structure area.

[0091] Traverse each isolated protrusion structure and find its highest point. Using the base plane as a reference, measure the vertical distance between the highest point and the base plane. This distance represents the height of the protrusion structure. Determine appropriate height intervals, for example, you can divide the protrusion height into intervals such as 0-1μm, 1-2μm, and 2-3μm. After summing the height values ​​of all protrusion structures, count the number of protrusions within each interval. Then, calculate the proportion of the number of protrusions in each interval to the total number of protrusions. Finally, organize the height intervals and corresponding proportions into a table, the protrusion height distribution table. Based on actual research needs and the distribution of protrusion structures, set an effective distance threshold. For example, you can set the threshold to 5μm, indicating that the number of adjacent protrusions within a circular area with a radius of 5μm centered on the protrusion structure is counted. For each protrusion structure, count the number of adjacent protrusions within the effective distance threshold. Sum the number of adjacent protrusions for all protrusion structures and divide it by the total number of protrusions to obtain the average spacing density.

[0092] Using professional 3D modeling software such as SolidWorks and AutoCAD, three-dimensional models of the PET film surface before and after corona treatment are constructed based on the average spacing density and surface undulation feature value set. During the modeling process, information such as the height, position, and distribution of the raised structures is accurately reflected in the model. Using the measurement function of the 3D modeling software, the true surface contact area in the 3D model before and after treatment is calculated. The true surface contact area refers to the area of ​​actual contact between the film surface and other objects, taking into account the microscopic undulations of the surface. The true surface contact area increment ratio is obtained by subtracting the true surface contact area before treatment from the true surface contact area after treatment, and then dividing by the true surface contact area before treatment.

[0093] Assume that there is a piece of PET film. Before corona treatment, a confocal microscope array and a laser displacement sensor are used to perform a three-dimensional topography scan of its surface, and surface height data containing 1000 measurement points are obtained. After calculation, the root mean square roughness R of the surface height before treatment is q1 After the corona treatment, the film surface is scanned again to obtain a new surface relief characteristic value set, which also contains 1000 measurement points. After integrating the data before and after treatment to form a data comparison group, the root mean square roughness R of the surface height after treatment is calculated. q2 is 0.8 μm, so the surface micro-roughness change value is 0.8-0.5=0.3 μm.

[0094] Traverse these 200 protrusion structures, measure the height of each protrusion, and then group the heights into three intervals: 0-1μm, 1-2μm, and 2-3μm. Assume that there are 120 protrusions in the 0-1μm interval, 60 protrusions in the 1-2μm interval, and 20 protrusions in the 2-3μm interval. The proportion of the number of protrusions in each interval to the total number of protrusions is: 120 ÷ 200 = 60%, 60 ÷ 200 = 30%, and 20 ÷ 200 = 10%. The resulting protrusion height distribution table is as follows:

[0095]

[0096]

[0097] The effective distance threshold was set to 5 μm, and the number of adjacent protrusions within the threshold range of each protrusion structure was counted. After calculation, the average spacing density was 2 adjacent protrusions around each protrusion. 3D modeling software was used to construct 3D models of the PE film surface before and after treatment, and the actual contact area of ​​the surface before treatment was measured to be 100 mm. 2 , the actual contact area of ​​the surface after treatment is 120mm 2 The actual surface contact area increment ratio is (120-100)÷100=20%.

[0098] By integrating the data before and after treatment to form a comparison group and calculating the change in surface micro-roughness, we can fully and accurately understand the impact of corona treatment on the surface microstructure of PET film. These data provide a basis for the subsequent extraction and analysis of the protrusion structure characteristics, which will help to further study the effects and mechanisms of corona treatment. Accurately extracting the characteristic parameters of the regular protrusion structure produced by corona treatment, such as protrusion height distribution, spacing density and surface real contact area increment ratio, can provide a key basis for printing layout optimization. For example, the protrusion height distribution can help determine the extent of the protrusion's impact on the printed circuit, the spacing density can affect the layout and connection method of the circuit, and the surface real contact area increment ratio can reflect the adhesion performance between the film surface and the conductive ink, thereby providing scientific guidance for adjusting the circuit parameters.

[0099] Based on these accurate surface feature parameters, the adaptive layout module can more precisely adjust circuit parameters, avoiding problems such as breakage or resistance fluctuations caused by raised structures on the film surface. This improves the accuracy and stability of printed circuits and meets the stringent quality requirements of electronic devices. Through in-depth analysis of the surface characteristics of the film after corona treatment, corona treatment process parameters can be optimized, improving treatment effectiveness and reducing defective product rates caused by surface defects. Furthermore, rational circuit layout optimization can reduce production costs, improve production efficiency, and enhance a company's market competitiveness.

[0100] In a preferred embodiment of the present invention, a dual-channel convolutional neural network is used to analyze surface topography parameters and preset circuit layout information, and output the fracture probability of each edge and the resistance fluctuation coefficient, including:

[0101] Collect surface topography parameters and preset circuit layout information, and convert the surface topography parameters into tensor form and the circuit layout information into a two-dimensional matrix;

[0102] Normalizing the converted surface topography parameters and the preset circuit layout information, mapping the converted surface topography parameters and the preset circuit layout information to the [0, 1] interval to obtain normalized data;

[0103] Based on the normalized data, a dual-channel convolutional neural network is constructed, and the normalized data is divided into a training set and a validation set; the training set is input into the dual-channel convolutional neural network to extract multi-scale features of surface morphology parameters and local features of line connections; the dual-channel convolutional neural network includes:

[0104] The first channel: inputs the surface morphology parameter tensor in the training set and extracts the multi-scale features of the surface morphology parameters through a three-layer convolutional network;

[0105] Second channel: Input a two-dimensional matrix of preset line layout information, aggregate node neighborhood information through the graph attention network, and generate edge feature vectors;

[0106] The multi-scale features of the surface morphology parameters and the local features of the line connection are spliced ​​to obtain the combined features, which are then processed through the fully connected layer to learn the relationship between the features;

[0107] The combined features processed by the fully connected layer are input into the output layer of the dual-channel convolutional neural network. One neuron is used in conjunction with a normalized exponential function to calculate the fracture probability of each edge of the printed circuit. The fracture probability output value is in the interval [0, 1] and the sum of all edge probabilities is 1. Another neuron is used in conjunction with a linear activation function to obtain the resistance fluctuation coefficient.

[0108] Define the cross entropy and mean square error loss functions, calculate the gradient of the loss function with respect to the parameters of the two-channel convolutional neural network, and update the parameters of the two-channel convolutional neural network at the same time;

[0109] Use the validation set to evaluate the parameters of the dual-channel convolutional neural network, and adjust the dual-channel convolutional neural network structure and hyperparameters based on the evaluation results until the preset number of iterations is reached to obtain a trained and optimized dual-channel convolutional neural network;

[0110] The normalized surface morphology parameters and preset circuit layout information are input into the trained and optimized dual-channel convolutional neural network, and the fracture probability of each edge and the resistance fluctuation coefficient are output.

[0111] In an embodiment of the present invention, surface topography parameters, including protrusion height distribution, spacing density, and surface actual contact area increment ratio, are obtained from a feature processing module. Preset circuit layout information, such as circuit connection relationships and node positions, is obtained from a preset parameter library. The surface topography parameters are converted into a tensor form. A tensor is a multidimensional array that can be easily calculated and processed in a neural network. For example, parameters such as protrusion height distribution and spacing density are combined into a tensor according to a certain dimensional order. The preset circuit layout information is converted into a two-dimensional matrix. The elements in the matrix can represent the connection relationship between circuit nodes, such as in the form of an adjacency matrix, where 1 indicates a connection between two nodes and 0 indicates no connection.

[0112] For the converted surface morphology parameter tensor and circuit layout information two-dimensional matrix, normalization methods are used to map them to the [0, 1] interval. Common normalization methods include minimum-maximum normalization. For each feature in the data set (each dimension in the tensor or each element in the matrix), the formula is used. Calculate, where x is the original data, x min and x max The minimum and maximum values ​​of the feature, respectively, are used to normalize all data to a specified interval, bringing the data to the same scale, which aids in neural network training and convergence. A dual-channel convolutional neural network is constructed, consisting of two independent but interconnected channels. The normalized data is divided into a training set and a validation set according to a specific ratio (e.g., 70% for training and 30% for validation). The training set is used to train the neural network to learn the features and patterns in the data; the validation set is used to evaluate the model's performance during training and prevent overfitting.

[0113] First channel: The surface topography parameter tensor from the training set is input into the first channel, which consists of three convolutional layers. The convolutional layers perform convolution operations by sliding convolution kernels over the input tensor, extracting features at different scales. For example, the first convolutional layer can extract small-scale local features, while subsequent layers can extract larger-scale, more abstract surface topography features by increasing the receptive field of the convolution kernels or stacking more convolutional layers.

[0114] Second channel: The two-dimensional matrix of preset line layout information is input into the second channel and the Graph Attention Network (GAT) is used. The GAT calculates the attention weights between nodes and aggregates node neighborhood information to generate edge feature vectors. These edge feature vectors can reflect the local characteristics of line connections, such as the closeness and importance of the connection between nodes.

[0115] The multi-scale features of surface topography parameters extracted by the first channel and the local features of line connections generated by the second channel are concatenated. This means that the two feature vectors are sequentially connected into a longer vector to obtain a combined feature. The combined feature is input into a fully connected layer. Each neuron in the fully connected layer is connected to all neurons in the previous layer. By performing weighted summation and activation operations on the combined features, the relationship between different features is learned, and a more advanced feature representation is further extracted.

[0116] Suppose that after the combined features processed by the fully connected layer are input to the output layer, the input value corresponding to the I-th line edge is z I , there are N line edges (I=1,2,…,N). The Softmax function calculates the break probability F of the I-th line edge I The formula is Where e is a natural constant (approximately equal to 2.71828), Indicates the corresponding values ​​for all N line edges Sum, The base is the natural constant e, z J is the result of the exponential operation of the exponent, z J It is the value of the Jth line edge after the combined features processed by the fully connected layer are input to the output layer. That is to say, for each line edge, after being processed by the previous neural network layers (including the fully connected layer), a corresponding value z will be obtained. J In the calculation process of the Softmax function, each line edge is calculated Then set the value of all line edges Add to get the denominator By F I Satisfying 0≤p I ≤1, and This clearly reflects the probability distribution of a break on each line edge. Another neuron uses a linear activation function to obtain the resistance fluctuation coefficient. The linear activation function directly outputs a real number representing the resistance fluctuation coefficient of the corresponding line edge, reflecting the degree of resistance fluctuation.

[0117] Define the cross entropy loss function (used to calculate the difference between the predicted fracture probability and the true label, suitable for classification problems) and the mean square error loss function (used to calculate the error between the predicted resistance fluctuation coefficient and the true value, suitable for regression problems). Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the parameters of the dual-channel convolutional neural network (such as the convolution kernel weights of the convolutional layer, the weights and biases of the fully connected layer, etc.). Update the network parameters using the gradient descent algorithm or its variants (such as Adam, Adagrad, and other optimization algorithms), continuously adjust the parameter values, and gradually reduce the value of the loss function, thereby improving the model's prediction accuracy.

[0118] The validation set is used to evaluate the parameters of the dual-channel convolutional neural network during training, and the loss value and other evaluation indicators (such as accuracy, mean square error, etc.) on the validation set are calculated. According to the evaluation results, the structure of the dual-channel convolutional neural network (such as adding or reducing convolutional layers, adjusting the number of layers of the graph attention network, etc.) and hyperparameters (such as learning rate, convolution kernel size, parameters of the attention mechanism, etc.) are adjusted. The training, evaluation and adjustment process is repeated until the preset number of iterations (such as 1000 iterations) is reached to obtain the trained and optimized dual-channel convolutional neural network. The normalized surface morphology parameters and the preset circuit layout information are input into the trained and optimized dual-channel convolutional neural network. The network outputs the fracture probability and resistance fluctuation coefficient of each edge, providing a quantitative basis for subsequent circuit layout optimization.

[0119] Consider a piece of PET film. The surface topography parameters obtained by the feature processing module include: a bump height distribution of [60% for 0-1μm, 30% for 1-2μm, and 10% for 2-3μm], a spacing density of an average of two adjacent bumps around each bump, and a 20% increase in the actual surface contact area. These parameters are converted into tensor form, such as [0.6, 0.3, 0.1, 2, 0.2]. The default circuit layout information is a simple 4-node circuit, whose adjacency matrix is ​​represented as:

[0120]

[0121] The above data are normalized, and the surface morphology parameter tensor and the circuit layout information two-dimensional matrix are mapped to the [0, 1] interval. Then, a two-channel convolutional neural network is constructed, and the normalized data are divided into a training set and a validation set.

[0122] During training, the first channel's three-layer convolutional network processes the surface topography parameter tensor to extract surface features at different scales. The second channel's graph attention network processes the two-dimensional matrix of circuit layout information to generate edge feature vectors. The extracted features are concatenated and then passed through a fully connected layer to learn feature relationships. The output layer then calculates the fracture probability of each edge (e.g., 0.1 for edge 1-2, 0.05 for edge 2-3), as well as the resistance fluctuation coefficient for each edge (e.g., 0.2 for edge 1-2). Network parameters are continuously updated by defining cross-entropy and mean squared error loss functions. After multiple iterations and evaluation and adjustments using a validation set, a trained and optimized two-channel convolutional neural network is obtained, which can be used to predict edge fracture probabilities and resistance fluctuation coefficients for new surface topologies and circuit layouts.

[0123] A dual-channel convolutional neural network employs different feature extraction methods for surface topography and circuit layout information. The first channel utilizes a convolutional network to extract multi-scale features of the surface topography, while the second channel utilizes a graph attention network to capture local features of circuit connections. This approach efficiently and comprehensively extracts key information, resulting in more targeted and richer features than a single network architecture. By fusing these two features and training them with appropriate activation and loss functions, the system accurately outputs the fracture probability of each edge and the resistance fluctuation coefficient. This provides precise quantitative metrics for circuit layout optimization, enabling the adaptive layout module to more accurately adjust circuit parameters and effectively reduce the risk of circuit fracture and resistance fluctuation.

[0124] The network can be trained and optimized based on different surface topography and circuit layout information, adapting to the needs of various real-world production scenarios. Whether it's flexible substrates of varying materials or circuit layouts of varying complexity, effective analysis and prediction can be achieved by adjusting data and network parameters. Accurate prediction results can help manufacturers proactively identify potential circuit quality issues, optimize corona treatment processes and printed circuit layout plans, reduce defective product rates due to circuit failures, improve production efficiency, reduce production costs, and enhance market competitiveness.

[0125] In a preferred embodiment of the present invention, dynamically adjusting line parameters based on the fracture probability and the resistance fluctuation coefficient by a path planner including a swarm incremental learning algorithm may include:

[0126] Build a path planner based on the swarm incremental learning algorithm and set the parameters of the swarm incremental learning algorithm, including the swarm size, learning rate, and incremental learning step size;

[0127] Organize the preset line layout information, determine the line connection relationship and node location, and determine the adjustment range of line parameters, including the minimum and maximum values ​​of line width, thickness and spacing parameters;

[0128] Based on the preset line layout information and the adjustment range of the line parameters, an initial group of line parameter combinations is randomly generated, each combination representing a line parameter setting scheme;

[0129] For each line parameter combination in the group, the corresponding adaptation index is calculated by combining the fracture probability and the resistance fluctuation coefficient; the line parameter combination includes line width, line thickness, line spacing, line length, conductor material, insulation material, solder joint quality and via parameters;

[0130] At the beginning of each iteration, new line performance data is obtained from the actual printing process, and the new line performance data is combined with the line parameters in the current group and the corresponding adaptation index and input into the path planner to update the path planner to obtain an updated path planner;

[0131] According to the updated path planner, the line parameter combination in the group is adjusted to obtain an adjusted line parameter combination, and the adaptation index of the adjusted line parameter combination is recalculated;

[0132] The process of acquiring new line performance data, updating the path planner, adjusting the line parameter combination, and recalculating the adaptation index is repeated until the preset maximum number of iterations is reached. The final line parameter combination is determined from the population based on the adaptation index of the line parameter combination;

[0133] According to the final circuit parameter combination, the actual parameters of the printed circuit are dynamically adjusted.

[0134] In an embodiment of the present invention, a programming tool (such as Python combined with a relevant machine learning library) is used to implement a path planner based on a swarm incremental learning algorithm.

[0135] Setting parameters:

[0136] Group size: determines the number of line parameter combinations involved in learning and optimization. For example, setting the group size to 50 means that 50 different line parameter setting schemes participate in the iterative optimization process of the algorithm.

[0137] Learning rate: controls the step size of each parameter update, such as setting the learning rate to 0.01.

[0138] Incremental learning step size: determines the amount of new data introduced for learning each time. For example, if the incremental learning step size is set to 10, incremental learning is performed every time 10 sets of new line performance data are obtained.

[0139] Detailed analysis of the pre-set circuit layout information is performed to clarify the connection relationships between each node in the circuit. For example, this information can be recorded using an adjacency matrix or other graph structure representation method. Furthermore, the position coordinates of each node on the PET film are precisely determined. Based on the production process and product requirements, the adjustment range of circuit parameters is determined. For circuit width, the minimum value is 5μm and the maximum value is 20μm; for circuit thickness, the minimum value is 1μm and the maximum value is 5μm; and for circuit spacing, the minimum value is 10μm and the maximum value is 30μm.

[0140] Based on the preset circuit layout information and the determined circuit parameter adjustment range, a random number generation method (such as a uniformly distributed random number generator) is used to randomly select values ​​for each circuit parameter (circuit width, circuit thickness, circuit spacing, circuit length, conductor material, insulation material, solder joint quality, and via parameters) within its adjustment range. These values ​​are then combined to form a circuit parameter combination. This process is repeated to generate a population of initial circuit parameter combinations equal to the population size. For example, if the population size is 50, 50 different circuit parameter combinations are generated, each representing an initial circuit parameter setting. Each circuit parameter combination is specified to include information such as circuit width, circuit thickness, circuit spacing, circuit length, conductor material, insulation material, solder joint quality, and via parameters. For each circuit parameter combination in the population, the adaptation index calculation formula (such as the formula that comprehensively considers various factors) is used to calculate the adaptation index for each circuit parameter combination, combining the fracture probability and resistance fluctuation coefficient output by the dual-channel convolutional neural network. This results in a corresponding adaptation index. The adaptation index reflects the quality of the circuit parameter combination considering factors such as fracture probability and resistance fluctuation coefficient.

[0141] At the beginning of each iteration, new circuit performance data is obtained from the actual printing process through online monitoring equipment (such as online impedance meter, microscope, etc.), including but not limited to resistance value, circuit morphology and other information. The newly acquired circuit performance data is input into the path planner together with the circuit parameter combination in the current group and their corresponding adaptation indicators. The path planner analyzes the relationship between new data and existing data based on the group incremental learning algorithm, updates the internal model and parameters, and obtains an updated path planner. According to the updated path planner, the circuit parameter combination in the group is adjusted. The adjustment method can be an algorithm-based optimization strategy, such as crossover and mutation operations in genetic algorithms, or gradient descent methods, to change certain parameter values ​​in the circuit parameter combination to obtain an adjusted circuit parameter combination.

[0142] For the adjusted line parameter combination, the fracture probability and resistance fluctuation coefficient are combined again, and the adaptation index is recalculated according to the previous adaptation index calculation formula to evaluate the pros and cons of the adjusted line parameter combination. Repeat the process of obtaining new line performance data, updating the path planner, adjusting the line parameter combination and recalculating the adaptation index until the preset maximum number of iterations (such as 100 iterations) is reached. After reaching the maximum number of iterations, the line parameter combination is screened from the population based on the adaptation index, and the line parameter combination with the best adaptation index (such as the smallest adaptation index value) is selected as the final line parameter combination. According to the final line parameter combination determined, the parameters such as line width, line thickness, and line spacing are applied to the actual printed circuit production. By adjusting the parameter settings of the printing equipment, the actual parameters of the printed circuit are dynamically adjusted to optimize the quality and performance of the printed circuit.

[0143] Suppose you want to print a simple sensor circuit on a PET film. The preset circuit layout is a network structure with 5 nodes and 4 edges.

[0144] A path planner was built using Python and related libraries, with a population size of 30, a learning rate of 0.005, and an incremental learning step size of 8. Circuit layout information was organized to determine the connection relationships and node positions between nodes. The circuit parameter adjustment range was set as follows: the minimum circuit width was 8μm and the maximum was 18μm; the minimum circuit thickness was 2μm and the maximum was 4μm; and the minimum circuit spacing was 12μm and the maximum was 25μm. Thirty initial circuit parameter combinations were obtained through random number generation. For example, one combination included a circuit width of 12μm, a circuit thickness of 3μm, a circuit spacing of 18μm, a circuit length (determined based on the layout), a conductor material (assumed to be silver), an insulating material (assumed to be polyimide), good solder joint quality, and via parameters (e.g., a via diameter of 0.2mm). The adaptation index of each line parameter combination is calculated by combining the edge fracture probability of each edge (such as the edge 1-2 fracture probability is 0.1, the edge 2-3 fracture probability is 0.08, etc.) and the resistance fluctuation coefficient (such as the edge 1-2 resistance fluctuation coefficient is 0.15, etc.) output by the dual-channel convolutional neural network.

[0145] At the beginning of each iteration, new resistance value data is obtained through an online impedance meter. This data, along with the current 30 line parameter combinations and their adaptation indicators, is input into the path planner to update the path planner. Based on the updated path planner, the line parameter combinations are adjusted, such as adjusting the line width of a certain combination to 14μm, and then recalculating the adaptation indicators. After 50 iterations, the combination with the best adaptation indicator (such as the smallest adaptation indicator value) is selected from the 30 line parameter combinations. Assume that the final combination is a line width of 13μm, a line thickness of 3.2μm, and a line spacing of 20μm. The parameters of the final line parameter combination are applied to the printing equipment, and the actual parameters of the printed circuit are adjusted for production.

[0146] Using a swarm incremental learning algorithm and path planner, the system dynamically adjusts circuit parameter combinations based on continuously acquired circuit performance data, enabling them to adapt to changes and requirements during the actual printing process, improving circuit stability and reliability. By calculating an adaptation index based on the fracture probability and resistance fluctuation coefficient, and adjusting circuit parameters accordingly, the system effectively reduces the risk of fracture and the degree of resistance fluctuation, thereby improving printed circuit quality and meeting high-precision circuit requirements (e.g., circuit accuracy ≤ 10μm). This system avoids the high defective rate and low production efficiency associated with unreasonable circuit parameters in traditional fixed layout solutions. Through continuous iterative optimization, the final circuit parameter combination is quickly found, reducing debugging time and material waste during production and improving production efficiency. The system can adapt to different circuit layouts and production environments, optimizing printed circuits for different PET film materials and printing process requirements by adjusting the circuit parameter combination, enhancing the system's versatility and adaptability. Decisions based on circuit performance data acquired during the actual printing process make the optimization process more scientific and rational, reducing the subjectivity and uncertainty of human judgment and improving the accuracy and reliability of decisions.

[0147] In another preferred embodiment of the present invention, for each line parameter combination in the group, the corresponding adaptation index is calculated in combination with the fracture probability and the resistance fluctuation coefficient, including:

[0148] Obtain the actual parameter values ​​and corresponding adjustment ranges of each line unit's fracture probability, resistance fluctuation coefficient, line width, line thickness, and line spacing;

[0149] According to the fracture probability of each line unit, the contribution of the fracture probability part is determined; according to the resistance fluctuation coefficient of each line unit, the contribution of the resistance fluctuation coefficient is obtained;

[0150] For each line unit, determining the middle position of the line width and line thickness adjustment range, and determining the degree of deviation of the line width from the middle position and the degree of deviation of the line thickness from the middle position based on the middle position;

[0151] Determine the middle position of the line spacing adjustment range and calculate the degree of deviation of the actual line thickness from the middle position;

[0152] The contribution of the fracture probability, the contribution of the resistance fluctuation coefficient, the deviation of the line width from the middle position, the deviation of the line thickness from the middle position, and the deviation of the actual line thickness from the middle position are integrated to obtain the adaptation index of the line parameter combination.

[0153] In the embodiment of the present invention, high-precision measuring equipment, such as optical microscopes and electron probes, is used to accurately measure the line width W, line thickness T, and line spacing S of each line unit, thereby obtaining its actual parameter values. These measuring equipment can meet the requirements for high-precision measurement of line parameters, ensuring that the acquired data is accurate and reliable. The adjustment range of the line parameters, that is, the minimum line width W min =5μm, maximum value W max =20μm, minimum line thickness T min =1μm, maximum value T max =5μm, minimum line spacing S min =10μm, maximum value S max =30μm, which is determined by the production process requirements and product design specifications.

[0154] For example, considering the precision limitations of the printing process, the current printing equipment and technology are difficult to accurately print circuits with a width of less than 5μm. When the circuit width exceeds 20μm, it will affect the circuit integration and the miniaturization design of the product. From the perspective of the physical properties of the materials, the conductive ink and substrate used may not be able to form a stable conductive circuit when the thickness is less than 1μm, while a thickness exceeding 5μm will increase the cost and weight of the circuit and may also affect the flexibility of the product. Based on the requirements of circuit performance, a circuit spacing less than 10μm is likely to increase the risk of short circuits between circuits, while a spacing exceeding 30μm will increase the resistance of the circuit and affect signal transmission. Taking these factors into consideration, a reasonable value range is set for each circuit parameter as shown above to ensure that the produced circuit can meet the quality and performance standards.

[0155] First, the surface morphology parameters (such as the height distribution of protrusions on the surface of the PET film after corona treatment, spacing density, and the incremental ratio of the actual surface contact area) are collected and converted into tensor form; at the same time, the preset circuit layout information (such as the connection relationship of the circuit, the position of the node, etc.) is collected and converted into a two-dimensional matrix. The converted data is normalized, mapped to the interval [0, 1], and then input into a dual-channel convolutional neural network. The first channel of the network extracts the multi-scale features of the surface morphology parameters through a three-layer convolutional network, and the second channel uses a graph attention network to aggregate the node neighborhood information of the circuit layout information to generate an edge feature vector. After the features extracted by the two channels are spliced, the relationship between the features is learned through the fully connected layer and then input into the output layer. A neuron in the output layer uses a normalized exponential function (Softmax function) to calculate the fracture probability of each circuit unit. in, The base is the natural constant e, z I is the result of the exponential operation of the exponent, z Iis the value corresponding to the I-th line edge (line unit), e is a natural constant (approximately equal to 2.71828), N is the total number of line units, and J is an index variable. The base is the natural constant e, z J is the result of the exponential operation of the exponent, z J After the combined features processed by the fully connected layer are input to the output layer, the value of the J-th line edge (line unit) is obtained by another neuron with the linear activation function to obtain the resistance fluctuation coefficient R I =k×y+b, where k is the slope parameter, b is the intercept parameter, and y is the eigenvalue. In this way, the powerful data analysis capabilities of neural networks are leveraged to indirectly determine the fracture probability and resistance fluctuation coefficient of the line unit.

[0156] The fracture probability F of all line units I Add up and get Then multiply it by the weight w of the fracture probability F , and the contribution of the fracture probability to the adaptation index is obtained Weight w F It reflects the importance of the fracture probability in the overall assessment. The larger the weight, the more critical the impact of the fracture probability on line quality.

[0157] Similarly, the resistance fluctuation coefficient R of all line units is I Add together and get Then multiply it by the weight w of the resistance fluctuation coefficient R , and the contribution of the resistance fluctuation coefficient is obtained Weight w R This shows the importance of resistance fluctuation coefficient in evaluating line quality.

[0158] For each line unit, first calculate the middle position of the line width adjustment range, that is, Subtract the middle position value from the actual width W of the line unit and take the absolute value of the difference Divide by half the width adjustment range difference Get a relative deviation ratio Multiply this ratio by the influence coefficient of line width on line performance And sum the results of all line units Finally multiply it by the line width weight w W , get the contribution value of line width deviation from the middle position to the adaptation index Influence coefficient This value indicates the difference in the impact of line width changes on line performance for different line units. Assuming there are three line units, the specific values ​​are set as follows:

[0159] For line unit 1, if it is located in a relatively flat area on the film surface and is a common signal transmission line with general functional importance, it can be set This means that the change in line width of the line unit has relatively little impact on line performance.

[0160] If the line unit 2 bears the key power transmission function, and the film surface in the area has many micro-protrusions and complex surface morphology, in order to highlight the greater impact of the line width change on the performance, it is set

[0161] If line unit 3 belongs to an auxiliary line and the surface morphology of the area is relatively stable, set

[0162] Same as line width calculation method, first calculate the middle position of line thickness adjustment range Get the deviation ratio between the actual thickness T and the middle position Multiply by the line thickness influence coefficient Then sum it up and multiply it by the line thickness weight w T , get the contribution value of line thickness deviation from the middle position to the adaptation index Line thickness influence coefficient The value can be set based on factors such as the function of the circuit in the circuit, the stress conditions in the area, etc. Assuming there are 4 circuit units on a printed circuit board, the specific value settings are as follows:

[0163] Line unit 1 is a high-frequency signal transmission line, which is sensitive to changes in line thickness and is located in an area prone to stress concentration. This shows that changes in line thickness have a great impact on line performance.

[0164] Line unit 2 is mainly used for power transmission. The thickness stability is related to the power carrying capacity. The stress in the area is moderate. This shows that changes in line thickness have a significant impact on its performance.

[0165] Line unit 3 is a control signal line, with general functional importance and low stress in the area where it is located. This means that changes in line thickness have a moderate impact on its performance.

[0166] Line unit 4 is an auxiliary line, which has little impact on performance and is located in a stable environment. This indicates that changes in line thickness have little effect on its performance.

[0167] Calculate the middle position of the line spacing adjustment range Difference between the actual line spacing S and the middle position value and take the absolute value Divide by half the difference in the spacing adjustment range Get the relative deviation ratio Multiply this ratio by the line spacing influence coefficient Then sum it up and multiply it by the line spacing weight w S , get the contribution value of line spacing deviation from the middle position to the adaptation index Line spacing influence coefficient The value can be set based on factors such as circuit function, surrounding electromagnetic environment, and layout complexity. Assuming that a flexible electronic device circuit contains three circuit units, the specific value settings are as follows:

[0168] Line unit 1: This line unit is responsible for high-frequency signal transmission and is located in a densely populated area. Changes in line spacing can easily cause signal crosstalk and electromagnetic interference, significantly affecting line performance. Highlight the significant impact of spacing changes on performance.

[0169] Line unit 2: This line unit is used for power transmission. There are many sensitive components around it. The line spacing must take into account insulation and electromagnetic compatibility. This indicates that the change in spacing has a moderate impact on performance.

[0170] Line unit 3: As an auxiliary line, it is mainly responsible for simple control signal transmission. The layout of the area is loose and the electromagnetic environment is good. The change of line spacing has little effect on performance.

[0171] Add the fracture probability contribution, resistance fluctuation coefficient contribution, line width deviation contribution, line thickness deviation contribution, and line spacing deviation contribution obtained above, and you have:

[0172] The result A is the adaptation index of the line parameter combination, which comprehensively reflects the overall quality level of the line in terms of fracture risk, resistance stability, and parameter compliance. The lower the value, the better the line parameter combination.

[0173] Among them, w F This reflects the importance attached to the line break probability factor when calculating the adaptation index. The larger the weight, the greater the impact of the break probability on the final result when evaluating the line parameter combination. F The value range is between [0, 1] and satisfies w F +w R +w W +w T +w S =1. RThis reflects the degree of attention paid to line resistance fluctuations in the adaptation index calculation. The larger the weight value, the more significant the impact of the resistance fluctuation coefficient on the final adaptation index. W It is used to measure the importance of the line width deviation from the center value of its adjustment range in the adaptation index. T It reflects the degree of influence of the line thickness deviation from the center value of its adjustment range on the adaptation index. S Indicates the importance of the line spacing deviation from the center value of its adjustment range when calculating the adaptation index.

[0174] The adaptation index comprehensively considers fracture probability, resistance fluctuation coefficient, and deviations in parameters such as line width, thickness, and spacing. It evaluates line parameter combinations from multiple dimensions, avoiding the one-sidedness of single-index evaluation and providing a comprehensive and accurate reflection of the actual line quality and performance. By quantifying the contribution of each factor to the adaptation index, the degree of influence of different parameters on line quality is clarified. When optimizing line parameters, the weights and parameter values ​​can be adjusted based on the adaptation index to better meet actual requirements, effectively reducing problems such as line fracture and resistance anomalies, improving line accuracy and stability, and meeting stringent requirements such as line accuracy ≤10μm. By accounting for the adjustable range of line parameters and the influence coefficients of different line elements, the adaptation index can adapt to complex situations such as variations in the micromorphology of the film surface after corona treatment and varying production process requirements. Whether considering changes in material properties or fluctuations in process parameters, this index can be used to assess the rationality of line parameter combinations, enhancing the system's adaptability in various production scenarios. Accurate adaptation index calculation helps screen optimal line parameter combinations before production, reducing trial-and-error costs and material waste caused by inappropriate parameters. At the same time, it reduces the defective rate, shortens the production cycle, and improves production efficiency, thereby reducing the company's production costs and enhancing the company's competitiveness in the market.

[0175] In a preferred embodiment of the present invention, the resistance value of the printed circuit is monitored in real time by an online impedance meter. When the relative deviation between the actual resistance value and the theoretical value is ≥5%, the parameter feedback mechanism is triggered, that is, the coordinates of the abnormal area are fed back to the detection module for local re-measurement, and the re-measurement data is input into the dynamic prediction module to update the weight matrix, which may include:

[0176] Connect the online impedance meter to the printed circuit. The online impedance meter measures the circuit resistance value in real time according to the set sampling frequency, and continuously obtains the actual resistance value data of the printed circuit.

[0177] Obtain the theoretical resistance design value of the printed circuit from the database, and calculate the relative deviation between the actual resistance value measured in real time and the theoretical value;

[0178] Compare the relative deviation value with the set threshold of 5%. If the relative deviation value is less than 5%, the real-time resistance value monitoring will continue; if the relative deviation value is ≥5%, the parameter feedback mechanism will be triggered immediately.

[0179] When the parameter feedback mechanism is triggered, the coordinates of the abnormal resistance value area are located based on the layout information of the printed circuit and the measurement point position of the online impedance meter;

[0180] The detection module receives the coordinates of the abnormal resistance area, starts the local retest program, controls the detection equipment to move to the abnormal area, and retests the resistance of the printed circuit in the abnormal area at multiple points and angles. During the retest process, the environmental parameters are collected and the retested resistance data and environmental parameters are integrated into a retest data set.

[0181] The retest data set is input into the dynamic prediction module to analyze the impact of the retest data on the current dual-channel convolutional neural network. The weight matrix is ​​updated by recalculating the weights of each parameter in the dual-channel convolutional neural network, including the weights of the convolutional layer and the fully connected layer.

[0182] In an embodiment of the present invention, the online impedance meter is electrically connected to the designated measurement point of the printed circuit by means of a probe or a fixture to ensure the integrity and accuracy of the measurement loop. For example, for a complex multi-layer printed circuit board, it is necessary to accurately connect to the test pad of the target circuit layer. According to the operating frequency of the printed circuit, the speed at which the resistance value may change, and the production efficiency requirements, a suitable sampling frequency is set in the control interface or supporting software of the online impedance meter. For example, for high-frequency and rapidly changing circuits, the sampling frequency can be set to 100 times / second to ensure that small fluctuations in the resistance value can be captured in time; for low-frequency and stable circuits, the sampling frequency can be set to 10 times / second to reduce the amount of data processing while ensuring the monitoring effect. The online impedance meter continuously measures the resistance value of the printed circuit according to the set sampling frequency, and transmits the actual resistance value data obtained from each measurement in real time and stores it in the system's data cache or database.

[0183] The theoretical resistance design value of the printed circuit is retrieved from the database. This value is a standard reference value obtained through simulation based on the circuit design parameters (such as circuit length, width, material resistivity, etc.). For example, during the design phase, the theoretical resistance value is calculated based on the copper foil thickness, trace length and width of the circuit, combined with the copper resistivity, and stored in the corresponding circuit design parameter table in the database. The relative deviation between the actual resistance value measured in real time and the theoretical value is calculated using the formula: The calculated relative deviation value is compared with a pre-set threshold of 5%. This process is automatically executed by the system's logic judgment module and implemented through conditional statements (such as if-else statements).

[0184] If the relative deviation value is less than 5%, it means that the resistance value of the current printed circuit is within the normal fluctuation range, and the real-time resistance value monitoring will continue according to the set sampling frequency; if the relative deviation value is ≥5%, it is determined that the resistance value is abnormal, and the parameter feedback mechanism is immediately triggered to notify the system to enter the abnormal handling process. Combined with the layout information of the printed circuit (such as the topology of the circuit, node coordinates, routing path, etc.) and the measurement point position information of the online impedance meter (the coordinates or number corresponding to each measurement point), a spatial positioning algorithm is used for analysis. For example, the connection relationship between the coordinates of the measurement point and the circuit layout is used to determine which line segments are associated with abnormal measurement values. Based on the analysis results, the coordinate range of the area with abnormal resistance values ​​is accurately calculated. For a regular rectangular circuit area, the coordinates of its upper left corner and lower right corner can be determined; for an irregular area, its boundary can be described by the coordinates of multiple key points.

[0185] After receiving the coordinates of the abnormal resistance value area, the detection module automatically starts the local retest program. The motion control system of the detection equipment (such as a high-precision microscope, probe station, etc.) is driven by control instructions, so that the measurement probe or lens of the detection equipment moves to the designated position of the abnormal area. Within the abnormal area, the resistance value is retested at multiple points and angles according to the preset measurement strategy. For example, within a rectangular abnormal area, multiple measurement points are selected in a grid-like distribution, and the resistance value of each measurement point is measured from different angles (such as vertical, 45° tilt, etc.) to obtain more comprehensive resistance value data. During the retest process, environmental parameters such as ambient temperature, humidity, and air pressure are collected simultaneously. These environmental parameters may affect the resistance value and therefore need to be recorded. Environmental parameters can be collected in real time using devices such as temperature and humidity sensors and air pressure sensors connected to the system. The resistance value data obtained from the retest is integrated with the collected environmental parameters to form a retest data set containing multi-dimensional information. For example, the resistance value, measurement time, measurement angle, and corresponding ambient temperature, humidity, and other information of each measurement point are combined into a data record, and all records constitute the retest data set.

[0186] The retest data set is input into the dynamic prediction module, which first preprocesses and extracts features from the retest data. It then analyzes the differences between the retest data and the predictions of the current dual-channel convolutional neural network model and assesses the impact of the retest data on the model. For example, by calculating the statistical features of the retest data (such as the mean and variance) and comparing them with the corresponding features predicted by the model, it is determined whether the model's predictive accuracy has been affected. Based on the analysis results, an optimization algorithm (such as stochastic gradient descent or the Adam algorithm) is used to recalculate the weights of each parameter in the dual-channel convolutional neural network, including the convolution kernel weights of the convolutional layers and the weights of the fully connected layers. The specific calculation process involves adjusting the weights according to a certain learning rate based on the gradient of a loss function (such as the cross-entropy loss function or the mean squared error loss function) to minimize the model's prediction error on the retest data. Through continuous iterative calculation and updating, an updated weight matrix is ​​ultimately obtained, enabling the model to better adapt to changes in actual production and improving the accuracy and reliability of predictions.

[0187] Assume that a printed circuit for a sensor is produced on a flexible printed circuit production line.

[0188] An online impedance meter was connected to a test point on a printed circuit board, with a sampling frequency of 50 times per second. The meter continuously measured the circuit resistance. For example, at a specific moment, the actual resistance value measured was 12.3Ω, and this data was stored in a database. The theoretical design resistance value for the printed circuit board was retrieved from the database as 10Ω. According to the formula, the relative deviation was calculated to be 23%. Since 23% was greater than 5%, the parameter feedback mechanism was immediately triggered, regular monitoring was stopped, and the exception handling process began. Combining the circuit layout information and the measurement point locations, the abnormal area was determined to be located in a certain section of the circuit, with coordinates ranging from the upper left corner (10mm, 15mm) to the lower right corner (20mm, 25mm). After receiving the coordinates, the detection module controlled the high-precision probe station to move to the abnormal area. Nine measurement points were selected within this area, and resistance values ​​were measured at both vertical and 45° angles at each point. Data was also collected for an ambient temperature of 25°C and a humidity of 50%. All measured resistance values ​​were integrated with the environmental parameters to form a retest data set. This retest data set was fed into the dynamic prediction module, which analyzed the model and found significant deviations in the predictions for this area. The Adam algorithm is used to recalculate the weights of the dual-channel convolutional neural network. After multiple iterative adjustments, the model's weight matrix is ​​updated, making the model's resistance value prediction for this type of circuit more accurate.

[0189] By measuring resistance in real time at a set frequency, an online impedance meter can promptly detect subtle changes in printed circuit resistance. This improves real-time monitoring and accuracy compared to manual spot checks or timed measurements, ensuring continuous monitoring of circuit quality. When the relative deviation between the actual and theoretical resistance values ​​reaches a set threshold, a parameter feedback mechanism is immediately triggered, quickly locating the abnormal area and preventing the problem circuit from being passed to the next process step. This reduces production delays and resource waste caused by circuit quality issues, thereby improving production efficiency. Based on circuit layout and measurement point information, the abnormal area is accurately located, and re-measurements are performed at multiple locations and angles. Simultaneously, environmental parameters are collected to comprehensively and accurately capture the actual situation in the abnormal area, avoiding misjudgments caused by a single measurement point or angle, and improving the accuracy of problem diagnosis. This re-measurement data is fed into a dynamic prediction module to update the weight matrix, enabling the dual-channel convolutional neural network to continuously learn and adapt to actual changes in the production process. This improves the accuracy and reliability of the model's circuit resistance predictions, providing a more scientific basis for circuit layout optimization and production process adjustments, fundamentally improving the quality and stability of printed circuits. This process forms a complete closed-loop quality control system from real-time monitoring, anomaly detection, precise positioning to model optimization, which can continuously improve the production process, continuously improve product quality, and enhance the company's competitiveness in the market.

[0190] An embodiment of the present invention further provides a computing device comprising: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, executes the system described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0191] The embodiment of the present invention further provides a computer-readable storage medium storing instructions, which, when executed on a computer, causes the computer to execute the system described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0192] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An adaptive corona printing typesetting optimization system based on a neural network model, characterized in that: include: The detection module is used to perform three-dimensional topography scanning on the surface of the corona-treated PET film using a confocal microscope array and a laser displacement sensor to obtain a set of surface relief feature values; The feature processing module is used to calculate the change in surface micro-roughness before and after treatment based on the surface relief feature value set, and to extract the characteristic parameters of the regular protrusion structure produced by corona treatment, including protrusion height distribution, spacing density, and surface real contact area increment ratio; A dynamic prediction module uses a dual-channel convolutional neural network to analyze surface topography parameters and preset circuit layout information, outputting the fracture probability of each edge and the resistance fluctuation coefficient; Adaptive layout module, which is used to dynamically adjust line parameters based on fracture probability and resistance fluctuation coefficient through a path planner including swarm incremental learning algorithm; The feedback control module is used to monitor the resistance value of the printed circuit in real time through an online impedance meter. When the relative deviation between the actual resistance value and the theoretical value is ≥5%, the parameter feedback mechanism is triggered, that is, the coordinates of the abnormal area are fed back to the detection module for local re-measurement, and the re-measurement data is input into the dynamic prediction module to update the weight matrix.

2. The adaptive corona printing typesetting optimization system based on the neural network model according to claim 1, characterized in that: Based on the surface relief characteristic value set, the surface micro-roughness change value before and after treatment is calculated, and the characteristic parameters of the regular protrusion structure produced by corona treatment are extracted, including the protrusion height distribution, spacing density and surface real contact area increment ratio, including: The surface relief feature value set obtained after the corona treatment is integrated with the PET film surface three-dimensional topography data and the corresponding feature value set before the corona treatment to form a data comparison group; The surface height data before and after treatment in the data comparison group were analyzed, and the root mean square roughness of the surface height before and after treatment was calculated to obtain the surface micro-roughness change value; Based on the surface micro-roughness change value, the three-dimensional topography image constructed by the surface relief feature value set is subjected to image morphological processing operations. A threshold is set according to the height difference between the protrusion and the substrate to separate the protrusion structure from the surface topography, forming an independent protrusion structure area. Traverse each raised structure and measure the vertical distance between the highest point and the base plane, i.e., the raised height. Summarize the height values ​​of each raised structure, group them according to a certain height interval, and calculate the proportion of the number of raised structures in each interval to the total number of raised structures to generate a raised height distribution table. Set an effective distance threshold, count the number of adjacent protrusions within the threshold range for each protrusion structure, and obtain the average spacing density; According to the average spacing density and surface undulation characteristic value set, three-dimensional models of the PET film surface before and after treatment were constructed using three-dimensional modeling software. By comparing the real surface contact areas in the three-dimensional models, the incremental ratio of the real surface contact areas before and after treatment was calculated.

3. The adaptive corona printing typesetting optimization system based on the neural network model according to claim 2 is characterized in that: A dual-channel convolutional neural network is used to analyze surface topography parameters and preset circuit layout information, outputting the fracture probability of each edge and the resistance fluctuation coefficient, including: Collect surface morphology parameters and preset circuit layout information, and convert the surface morphology parameters into tensor form and the circuit layout information into a two-dimensional matrix; Normalizing the converted surface topography parameters and the preset circuit layout information, mapping the converted surface topography parameters and the preset circuit layout information to the [0, 1] interval to obtain normalized data; Based on the normalized data, a dual-channel convolutional neural network is constructed, and the normalized data is divided into a training set and a validation set. The training set is input into the dual-channel convolutional neural network to extract the multi-scale features of the surface morphology parameters and the local features of the line connection; The multi-scale features of the surface morphology parameters and the local features of the line connection are spliced ​​to obtain the combined features, which are then processed through the fully connected layer to learn the relationship between the features; The combined features processed by the fully connected layer are input into the output layer of the dual-channel convolutional neural network. One neuron is used in conjunction with a normalized exponential function to calculate the fracture probability of each edge of the printed circuit. The fracture probability output value is in the interval [0, 1] and the sum of all edge probabilities is 1. Another neuron is used in conjunction with a linear activation function to obtain the resistance fluctuation coefficient. Define the cross entropy and mean square error loss functions, calculate the gradient of the loss function with respect to the parameters of the two-channel convolutional neural network, and update the parameters of the two-channel convolutional neural network at the same time; Use the validation set to evaluate the parameters of the dual-channel convolutional neural network, and adjust the dual-channel convolutional neural network structure and hyperparameters based on the evaluation results until the preset number of iterations is reached to obtain a trained and optimized dual-channel convolutional neural network; The normalized surface morphology parameters and preset circuit layout information are input into the trained and optimized dual-channel convolutional neural network, and the fracture probability of each edge and the resistance fluctuation coefficient are output.

4. The adaptive corona printing typesetting optimization system based on the neural network model according to claim 3 is characterized in that: The dual-channel convolutional neural network includes: The first channel: inputs the surface morphology parameter tensor in the training set and extracts the multi-scale features of the surface morphology parameters through a three-layer convolutional network; Second channel: Input a two-dimensional matrix of preset line layout information, aggregate node neighborhood information through the graph attention network, and generate edge feature vectors.

5. The adaptive corona printing typesetting optimization system based on the neural network model according to claim 4 is characterized in that: Based on the fracture probability and resistance fluctuation coefficient, the line parameters are dynamically adjusted through a path planner that includes a swarm incremental learning algorithm, including: Build a path planner based on the swarm incremental learning algorithm and set the parameters of the swarm incremental learning algorithm, including the swarm size, learning rate, and incremental learning step size; Organize the preset line layout information, determine the line connection relationship and node location, and determine the adjustment range of line parameters, including the minimum and maximum values ​​of line width, thickness and spacing parameters; Based on the preset line layout information and the adjustment range of the line parameters, an initial group of line parameter combinations is randomly generated, each combination representing a line parameter setting scheme; For each line parameter combination in the group, the corresponding adaptation index is calculated by combining the fracture probability and resistance fluctuation coefficient; At the beginning of each iteration, new line performance data is obtained from the actual printing process, and the new line performance data is combined with the line parameters in the current group and the corresponding adaptation index and input into the path planner to update the path planner to obtain an updated path planner; According to the updated path planner, the line parameter combination in the group is adjusted to obtain an adjusted line parameter combination, and the adaptation index of the adjusted line parameter combination is recalculated; The process of acquiring new line performance data, updating the path planner, adjusting the line parameter combination, and recalculating the adaptation index is repeated until the preset maximum number of iterations is reached. The final line parameter combination is determined from the population based on the adaptation index of the line parameter combination; According to the final circuit parameter combination, the actual parameters of the printed circuit are dynamically adjusted.

6. The adaptive corona printing typesetting optimization system based on the neural network model according to claim 5, characterized in that: The line parameter combination includes line width, line thickness, line spacing, line length, conductor material, insulation material, solder joint quality and via parameters.

7. The adaptive corona printing typesetting optimization system based on a neural network model according to claim 6, characterized in that: For each line parameter combination in the group, the corresponding adaptation index is calculated based on the fracture probability and resistance fluctuation coefficient, including: Obtain the actual parameter values ​​and corresponding adjustment ranges of each line unit's fracture probability, resistance fluctuation coefficient, line width, line thickness, and line spacing; According to the fracture probability of each line unit, the contribution of the fracture probability part is determined; according to the resistance fluctuation coefficient of each line unit, the contribution of the resistance fluctuation coefficient is obtained; For each line unit, determining the middle position of the line width and line thickness adjustment range, and determining the degree of deviation of the line width from the middle position and the degree of deviation of the line thickness from the middle position based on the middle position; Determine the middle position of the line spacing adjustment range and calculate the degree of deviation of the actual line thickness from the middle position; The contribution of the fracture probability, the contribution of the resistance fluctuation coefficient, the deviation of the line width from the middle position, the deviation of the line thickness from the middle position, and the deviation of the actual line thickness from the middle position are integrated to obtain the adaptation index of the line parameter combination.

8. The adaptive corona printing typesetting optimization system based on a neural network model according to claim 6, characterized in that: The resistance value of the printed circuit is monitored in real time using an online impedance meter. When the relative deviation between the actual resistance value and the theoretical value is ≥5%, the parameter feedback mechanism is triggered, which feeds the coordinates of the abnormal area back to the detection module for local retesting. The retested data is then input into the dynamic prediction module to update the weight matrix, including: Connect the online impedance meter to the printed circuit. The online impedance meter measures the circuit resistance value in real time according to the set sampling frequency, and continuously obtains the actual resistance value data of the printed circuit. Obtain the theoretical resistance design value of the printed circuit from the database, and calculate the relative deviation between the actual resistance value measured in real time and the theoretical value; Compare the relative deviation value with the set threshold of 5%. If the relative deviation value is less than 5%, the real-time resistance value monitoring will continue; if the relative deviation value is ≥5%, the parameter feedback mechanism will be triggered immediately. When the parameter feedback mechanism is triggered, the coordinates of the abnormal resistance value area are located based on the layout information of the printed circuit and the measurement point position of the online impedance meter; The detection module receives the coordinates of the abnormal resistance area, starts the local retest program, controls the detection equipment to move to the abnormal area, and retests the resistance of the printed circuit in the abnormal area at multiple points and angles. During the retest process, the environmental parameters are collected and the retested resistance data and environmental parameters are integrated into a retest data set. The retest data set is input into the dynamic prediction module to analyze the impact of the retest data on the current dual-channel convolutional neural network. The weight matrix is ​​updated by recalculating the weights of each parameter in the dual-channel convolutional neural network, including the weights of the convolutional layer and the fully connected layer.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the system according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the system according to any one of claims 1 to 8.

Images