Automatic control system for wrapping rate of polyimide sintering line

Abstract

The invention relates to the technical field of sintered wire wrapping rate control, in particular to an automatic polyimide sintered wire wrapping rate control system, which automatically detects surface defects of copper flat wires by combining a moving light source and an image recognition technology, ensures that only the copper flat wires with good surface states enter a wrapping stage, and avoids the problem of uneven wrapping from the source. The system dynamically adjusts the wrapping rate and the propelling rate through real-time feedback, adapts to the change of production conditions, and ensures the accurate wrapping of the polyimide film. By utilizing the image processing and artificial intelligence technology, the system can accurately identify the wrapping rate of the polyimide film and the copper flat wire, and the detection and verification module detects the insulation rate of the copper flat wire to ensure that the electrical performance of a final product reaches the standard. A decision model based on reinforcement learning is trained by introducing a near-end strategy optimization algorithm (PPO), closed-loop control is achieved, the packet stacking rate is minimized on the premise that the insulation performance is met, material use is remarkably reduced, and cost optimization is achieved.

Description

Technical Field

[0001] The invention relates to the technical field of sintering line stacking rate control, in particular to an automatic control system for the stacking rate of a polyimide sintering line. Background Art

[0002] In the field of modern electronics and electrical engineering, with the miniaturization of equipment and the improvement of performance requirements, higher requirements are placed on the insulation performance and mechanical stability of wire materials. Polyimide materials are widely used in the insulation layer of electronic components due to their excellent high temperature resistance, chemical corrosion resistance, electrical insulation and mechanical strength, especially in the application of high-performance copper flat wires. However, in the production process, how to effectively control the overlap rate of polyimide film on the surface of copper flat wires to ensure insulation performance while achieving the economy of materials has become a major challenge facing the industry.

[0003] The traditional polyimide film stacking process relies on mechanical settings and manual monitoring, which makes it difficult to cope with the variable changes in complex processes. This method often leads to too high a stacking rate, resulting in material waste, or too little stacking, which affects the insulation performance. Therefore, developing a system that can automatically control the stacking rate of polyimide sintering wires has become the key to improving production efficiency and product quality. Summary of the invention

[0004] In order to solve the above problems, the present invention provides an automatic control system for the overlapping rate of polyimide sintered wires. By integrating image recognition, semantic recognition technology and deep learning technology, the adjustment decisions of the advancement rate and winding rate of the copper flat wire are controlled, thereby minimizing the overlapping rate and achieving the goal of cost optimization while ensuring the insulation quality of the copper flat wire.

[0005] To achieve the above object, the technical solution adopted by the present invention is: A polyimide sintering line stacking rate automatic control system, comprising a pre-check module, a wrapping module, a stacking rate monitoring module, a sintering module, an annealing and cooling module, a detection and verification module and a stacking rate control module which are connected in sequence, wherein the stacking rate control module is also connected to the detection and verification module and the wrapping module respectively; The pre-check module is used to identify the copper flat wire by moving the light source and image recognition to confirm the characterization state of the copper flat wire, wherein the characterization state includes the presence of a characterization defect and the absence of a characterization defect; The wrapping module is used to control the wrapping rate of the polyimide film, and is also used to control the advancing rate of the copper flat wire whose characteristic state is that there is no characteristic defect; The overlap rate monitoring module is used to perform semantic segmentation on the polyimide film and the copper flat wire in the image data of the wrapping process based on the semantic segmentation model, and calculate the overlap rate of the polyimide film by the overlap rate of the polyimide film width on the copper flat wire; The sintering module is used to control the sintering temperature of the copper flat wire that has been laminated with a polyimide film, and the annealing and cooling module is used to control the temperature of the annealing and cooling of the copper flat wire that has been sintered; the detection and verification module is used to detect the insulation rate of the copper flat wire that has been annealed and cooled, and calculate the difference between the insulation rate of the copper flat wire and the preset standard insulation rate; The overlapping rate control module is used to construct a decision model through deep learning based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire, and generate adjustment decisions on the advancement rate and wrapping rate of the copper flat wire through the decision model.

[0006] Furthermore, the characterization and identification of the copper flat wire by moving the light source and image recognition to confirm the characterization state of the copper flat wire includes the following steps: Control the moving light source to move along a preset trajectory and illuminate the surface of the copper flat wire; Acquire image data of several angles of the surface of the copper flat wire when the moving light source moves; The color value trend is determined for the image data of several angles on the surface of the copper flat wire, and the abnormal points that do not conform to the color value trend are extracted; Determine whether there are characteristic defects on the surface of the copper flat wire based on the size of the abnormal point.

[0007] Furthermore, the packet overlap rate monitoring module is used to perform the following steps: Acquire image data of the copper flat wire being wrapped by the polyimide film during the wrapping process; Preprocess the image data, including image denoising, contrast enhancement and distortion correction; The preprocessed image data is input into the pre-trained semantic segmentation model to classify the polyimide film and the copper flat wire in the image, and the segmentation masks of the polyimide film and the copper flat wire are obtained; Extracting the width of the polyimide film through a segmentation mask of the polyimide film; The overlap rate of the polyimide film is calculated by the width of the overlapped area of ​​the polyimide film on the copper flat wire and the width of the polyimide film.

[0008] Furthermore, the training step of the semantic segmentation model includes: Obtain an image dataset of the polyimide film and copper flat wire wrapping process, and perform pixel-level data annotation on the polyimide film and copper flat wire in the image; Constructing a semantic segmentation model, wherein the semantic segmentation model includes an encoder and a decoder, wherein the encoder downsamples and extracts features of an input image, reduces the spatial resolution of a feature map and increases the number of channels of the feature map through convolution operations and pooling operations, and captures semantic information of the image; The decoder performs upsampling and semantic segmentation on the features extracted by the encoder, restores the spatial resolution of the feature map through deconvolution layers and skip connections, and maps the semantic information extracted by the encoder to pixel-level segmentation results; The image dataset with pixel-level data annotation is input into the semantic segmentation model for training, with the cross entropy loss function as the optimization target, and the model parameters are iteratively updated through the stochastic gradient descent algorithm.

[0009] Furthermore, the formula of the convolution operation is as follows: ; in, is the input image; For the input image In the The feature output of the layer; Is the connection Layer and The weight matrix of the kth convolution kernel of the layer; For the input image In the The feature output of the layer; It is Bias of the layer; is the activation function; is the number of convolution kernels.

[0010] Furthermore, the sintering temperature control and the annealing cooling temperature control are respectively performed by presetting the sintering working temperature and the annealing cooling temperature and their corresponding preset times.

[0011] Furthermore, the construction of a decision model through deep learning based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire includes the following steps: Define the state space, including the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire; define the action space, including the adjustment of the wrapping rate and the advancement rate; generate a reward function based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate and the overlap rate; A decision model is built based on the state space, action space and reward function, and reinforcement learning training is performed using historical data. The proximal strategy optimization algorithm is used to limit the difference between the new and old strategies during the iterative training process.

[0012] Furthermore, the objective function formula of the proximal strategy optimization algorithm is as follows: ; in, is the objective function; represents the expected value at time step t; is the probability ratio of the new strategy to the old strategy at time step t; is the advantage function, which is calculated by the difference between the actual return and the baseline; is a hyperparameter; To limit the probability ratio Function used to clip Restricted to and between; Indicates the minimum value operation.

[0013] Furthermore, the reward function includes: If the difference between the insulation rate of the copper flat wire and the preset standard insulation rate is a positive value or zero, the reward function is a positive reward and is negatively correlated with the overlap rate; If the difference between the insulation rate of the copper flat wire and the preset standard insulation rate is a negative value, the reward function is a negative reward and is irrelevant to the overlapping rate.

[0014] The beneficial effect of the present invention is that the present invention automatically detects possible defects on the surface of the copper flat wire by combining a mobile light source with image recognition technology, ensuring that in the production process, only the copper flat wire with a good surface condition enters the wrapping stage, thereby avoiding the problem of uneven lapping caused by uneven surface from the source. By dynamically adjusting the wrapping rate and the advancement rate of the copper flat wire, combined with the data provided by the pre-check module, the accurate lapping of the polyimide film is ensured. This process no longer relies on fixed mechanical settings, but adjusts the process parameters through real-time feedback, so that the wrapping process can adapt to changes in production conditions and ensure the uniformity and accuracy of each wrapping. In addition, by processing the real-time captured image data, the images of the polyimide film and the copper flat wire are accurately identified and segmented, and their overlap rate is calculated. This technology based on image processing and artificial intelligence can overcome the limitations of traditional methods under complex lap patterns and ensure accurate control of the lap rate. The detection and verification module detects the insulation rate of the copper flat wire that has completed the entire process, compares the actual insulation rate with the preset standard, and ensures that the electrical performance of the final product meets the standard. The stacking rate control module establishes a closed-loop control system based on reinforcement learning by introducing the proximal policy optimization algorithm (PPO). The system can autonomously learn and adjust key parameters such as wrapping rate and advancement rate under multi-dimensional data input. In this process, the system will continuously optimize the decision model based on actual production data to ensure that the stacking rate is minimized while meeting the insulation performance, thereby significantly reducing material usage and achieving cost optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1The invention discloses a structural schematic diagram of an automatic control system for the stacking rate of a polyimide sintering line.

[0016] Figure 2 It is a flow chart of the training steps of the semantic segmentation model in the present invention. DETAILED DESCRIPTION

[0017] See also Figure 1-2 As shown, the present invention relates to an automatic control system for the stacking rate of a polyimide sintering line, comprising a pre-check module, a wrapping module, a stacking rate monitoring module, a sintering module, an annealing and cooling module, a detection and verification module and a stacking rate control module which are connected in sequence, and the stacking rate control module is also connected to the detection and verification module and the wrapping module respectively; The pre-check module is used to identify the copper flat wire by moving the light source and image recognition to confirm the characterization state of the copper flat wire, wherein the characterization state includes the presence of a characterization defect and the absence of a characterization defect; The wrapping module is used to control the wrapping rate of the polyimide film, and is also used to control the advancing rate of the copper flat wire whose characteristic state is that there is no characteristic defect; The overlap rate monitoring module is used to perform semantic segmentation on the polyimide film and the copper flat wire in the image data of the wrapping process based on the semantic segmentation model, and calculate the overlap rate of the polyimide film by the overlap rate of the polyimide film width on the copper flat wire; The sintering module is used to control the sintering temperature of the copper flat wire that has been laminated with a polyimide film, and the annealing and cooling module is used to control the temperature of the annealing and cooling of the copper flat wire that has been sintered; the detection and verification module is used to detect the insulation rate of the copper flat wire that has been annealed and cooled, and calculate the difference between the insulation rate of the copper flat wire and the preset standard insulation rate; The overlapping rate control module is used to construct a decision model through deep learning based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire, and generate adjustment decisions on the advancement rate and wrapping rate of the copper flat wire through the decision model.

[0018] In some embodiments, the system first realizes accurate detection of the surface state of the copper flat wire through the pre-check module. The module includes a high-resolution industrial camera and an adjustable mobile light source. The light source can evenly illuminate the surface of the copper flat wire at multiple angles, while the industrial camera captures multi-angle image data in real time. Combined with an embedded processor, the system uses advanced image recognition algorithms to analyze this data to identify possible minor defects on the surface, such as unevenness, cracks or contaminants. The pre-check module is integrated in an integrated framework, which can seamlessly connect with the production line and transmit the detection results to the control system in real time through a standard communication protocol. The wrapping module is mainly composed of a high-precision servo motor, a guide rail system and a tension controller. The servo motor controls the wrapping speed of the polyimide film, and the guide rail system accurately controls the advancement speed of the copper flat wire. The tension controller is used to ensure that the tension of the film during the wrapping process remains within a preset range to prevent the wrapping process from being too loose or too tight. The system realizes dynamic adjustment of the wrapping rate and advancement rate through close linkage with the pre-check module to ensure uniform stacking of the film. All hardware of the module is monitored and adjusted in real time by the central controller to ensure response speed and accuracy when production conditions change. The stacking rate monitoring module includes an industrial camera and an image processing unit. The image processing unit uses GPU acceleration and can run deep learning models in real time to perform semantic segmentation on images. Through the collaborative work of these hardware, the system can quickly calculate the stacking rate and transmit the data to the central control system for further optimization and adjustment. The sintering module is used in combination with the annealing and cooling module to control the sintering and cooling processes of the copper flat wire respectively. The sintering module uses precision heating elements and temperature sensors to ensure that the polyimide film is evenly sintered on the surface of the copper flat wire. The temperature control system of this module can maintain an accurate temperature curve in a high temperature environment to ensure the consistency of material properties. The annealing and cooling module uses a cooling fan and a liquid cooling system to gradually cool the sintered copper flat wire to prevent thermal stress from occurring inside the material, thereby improving the mechanical strength and insulation performance of the copper flat wire. The detection and verification module integrates a high-precision insulation resistance tester and a dielectric strength tester to comprehensively evaluate the insulation performance of the copper flat wire. The insulation resistance tester measures the resistance value of the polyimide film on the outer layer of the copper flat wire by applying a certain DC voltage. This resistance value reflects the insulation effect of the polyimide film. If the insulation resistance value is lower than the preset standard, it indicates that the film is defective or the stacking is insufficient. In order to ensure the accuracy of the measurement, the system strictly controls the environmental conditions during the test. The test environment temperature and humidity are within the standard range, and a shielding box is used to prevent external electromagnetic interference from affecting the measurement results. During the test, the system first applies a low voltage for pre-detection to ensure that there is no arc discharge or obvious insulation failure, and then gradually increases the voltage until the specified test voltage is reached. After the test is completed, the system automatically uploads the measurement data to the central control system.The central control system compares the actual measured insulation resistance and breakdown voltage with the preset standard and calculates the difference between the insulation performance and the standard. If the difference is positive, it means that the batch of copper flat wires is qualified.

[0019] Furthermore, the characterization and identification of the copper flat wire by moving the light source and image recognition to confirm the characterization state of the copper flat wire includes the following steps: Control the moving light source to move along a preset trajectory and illuminate the surface of the copper flat wire; Acquire image data of several angles of the surface of the copper flat wire when the moving light source moves; The color value trend is determined for the image data of several angles on the surface of the copper flat wire, and the abnormal points that do not conform to the color value trend are extracted; Determine whether there are characteristic defects on the surface of the copper flat wire based on the size of the abnormal point.

[0020] In some embodiments, when the light source moves along a set trajectory, a high-resolution camera captures image data of the surface of the copper flat wire in real time. These image data are directly transmitted to the image analysis algorithm through the data stream processing module. First, the system preprocesses each frame of the image, including noise filtering, contrast enhancement, and geometric correction, to ensure that the input image data has high quality and consistency. Next, the system analyzes the image through a color value trend determination algorithm. Based on statistical and machine learning techniques, the algorithm identifies areas with abnormal color values ​​by comparing the chromaticity histograms of multi-angle images. The algorithm determines which areas have color value distributions that deviate from the normal range based on a preset color value model. For areas with significant deviations, the system marks them as abnormal points. Further, the algorithm performs cluster analysis on these abnormal points and calculates their size and distribution in the image. Clustering algorithms such as K-means or DBSCAN (density-based clustering algorithm) cluster dense points into possible defect-characterizing areas based on the spatial distribution of abnormal points. The size of these defect areas is determined by calculating their pixel area in the image. If the area exceeds the set threshold, the system marks the area as a potential defect. In addition, the system will conduct a secondary screening of the clustering results, using a classification algorithm based on rules or machine learning to determine the actual impact of these areas. For example, the shape, texture and distribution characteristics of these areas are further analyzed using a support vector machine (SVM) or a convolutional neural network (CNN), so as to accurately determine whether the area is a true characterization defect. Finally, the system generates a characterization report that lists in detail all detected abnormal points and their corresponding characterization defect areas. This report will be used to determine whether the copper flat wire is suitable for subsequent processing. If serious defects are found, the system will automatically notify the operator to take appropriate measures, such as adjusting process parameters or replacing materials.

[0021] Furthermore, the packet overlap rate monitoring module is used to perform the following steps: Acquire image data of the copper flat wire being wrapped by the polyimide film during the wrapping process; Preprocess the image data, including image denoising, contrast enhancement and distortion correction; The preprocessed image data is input into the pre-trained semantic segmentation model to classify the polyimide film and the copper flat wire in the image, and the segmentation masks of the polyimide film and the copper flat wire are obtained; Extracting the width of the polyimide film through a segmentation mask of the polyimide film; The overlap rate of the polyimide film is calculated by the width of the overlapped area of ​​the polyimide film on the copper flat wire and the width of the polyimide film.

[0022] In some embodiments, the preprocessing stage uses a variety of image processing algorithms, including denoising, contrast enhancement, and distortion correction. Denoising uses a convolution-based filtering algorithm, such as Gaussian filtering or median filtering, to remove random noise in the image. Contrast enhancement algorithms, such as histogram equalization, can improve the contrast between the polyimide film and the copper flat wire in the image, ensuring that they are clearly visible in subsequent analysis. Distortion correction uses a geometric transformation algorithm to correct the deformation caused by optical or mechanical reasons in the image to ensure the accuracy and consistency of the image. The preprocessed image data is then input into a pre-trained semantic segmentation model. The model uses a convolutional neural network (CNN) architecture, especially a deep neural network (such as U-Net or SegNet) to achieve accurate classification of polyimide film and copper flat wire in the image. By learning a large amount of labeled data, the model can identify and segment the polyimide film and copper flat wire areas in the image and generate corresponding segmentation masks. The segmentation mask is a binary image in which each pixel belongs to either a polyimide film or a copper flat wire. This mask provides basic data for subsequent analysis. With the generated segmentation mask, the system is able to extract the width of the polyimide film. The width extraction algorithm is based on image processing technology, usually using edge detection and connected domain analysis methods to accurately measure the width of the polyimide film in the mask. This process requires consideration of the spatial resolution and geometric distortion correction of the image to ensure the accuracy of the measurement. Finally, the system determines the overlap ratio by calculating the ratio of the width of the overlap area of ​​the polyimide film on the copper flat wire to the total width of the film.

[0023] Furthermore, the training step of the semantic segmentation model includes: Obtain an image dataset of the polyimide film and copper flat wire wrapping process, and perform pixel-level data annotation on the polyimide film and copper flat wire in the image; Constructing a semantic segmentation model, wherein the semantic segmentation model includes an encoder and a decoder, wherein the encoder downsamples and extracts features of an input image, reduces the spatial resolution of a feature map and increases the number of channels of the feature map through convolution operations and pooling operations, and captures semantic information of the image; The decoder performs upsampling and semantic segmentation on the features extracted by the encoder, restores the spatial resolution of the feature map through deconvolution layers and skip connections, and maps the semantic information extracted by the encoder to pixel-level segmentation results; The image dataset with pixel-level data annotation is input into the semantic segmentation model for training, with the cross entropy loss function as the optimization target, and the model parameters are iteratively updated through the stochastic gradient descent algorithm.

[0024] In some embodiments, the process of pixel-level data annotation of image data is usually manually completed by professionals using annotation tools, and each pixel point must be accurately classified as polyimide film, copper flat wire or background. This fine annotation provides an accurate supervision signal for the model, ensuring that the model can learn the tiny features in the image during training. In terms of the construction of the semantic segmentation model, this embodiment adopts an encoder-decoder architecture. The encoder part is responsible for downsampling and feature extraction of the input image. Specifically, the encoder extracts local features of the image through a series of convolution operations, and reduces the spatial resolution of the feature map through pooling operations. This processing not only reduces the amount of calculation, but also helps the model capture more abstract semantic information. The number of channels of the feature map will increase as the convolution layer deepens, so that more complex features can be represented. The task of the decoder part is to upsample the high-level features extracted by the encoder and restore them to the same spatial resolution as the input image. The decoder usually enlarges the feature map through a deconvolution operation and uses skip connections to introduce low-level detail information in the encoder into the decoder. This design can combine high-level semantic information and low-level detail features while maintaining spatial resolution, thereby generating accurate pixel-level segmentation results. During model training, the annotated image dataset is input into the semantic segmentation model. The core of the training is to use the cross-entropy loss function as the optimization target. The cross-entropy loss function is used to measure the difference between the model's prediction results and the actual annotations. The smaller the loss value, the higher the prediction accuracy of the model. In each iteration, the system updates the parameters of the model through the stochastic gradient descent (SGD) algorithm. SGD gradually approaches the optimal solution by calculating the gradient of the loss function relative to the model parameters and adjusting the parameters in the direction of gradient descent in each training step.

[0025] Furthermore, the formula of the convolution operation is as follows: ; in, is the input image; For the input image In the The feature output of the layer; Is the connection Layer and The weight matrix of the kth convolution kernel of the layer; For the input image In the The feature output of the layer; It is Bias of the layer; is the activation function; is the number of convolution kernels.

[0026] Specifically, This feature represents the input image In the The feature extraction result of each convolution kernel is a fixed-size matrix (e.g., 3x3 or 5x5) used to extract features of local regions in the input feature map. The bias term is a scalar or matrix that linearly adjusts the result after the convolution operation to improve the expressiveness of the model. The activation function uses a nonlinear function, such as ReLU (Rectified Linear Unit). The activation function introduces nonlinearity so that the model can learn more complex features. Indicates how many different convolution kernels are used in the current layer. Each convolution kernel extracts different features, so the output feature map is also multi-channel.

[0027] Furthermore, the sintering temperature control and the annealing cooling temperature control are respectively performed by presetting the sintering working temperature and the annealing cooling temperature and their corresponding preset times.

[0028] Specifically, first, the sintering temperature control module sets the sintering working temperature within a certain range according to the material characteristics and process requirements. The temperature setting is determined by the pre-set experimental data and the thermal performance analysis of the material. It is usually required to ensure that the polyimide film can form a uniform and firm adhesion layer on the surface of the copper flat wire without damaging the copper flat wire. In order to achieve precise temperature control, the system adopts a closed-loop control strategy, that is, the temperature of the sintering area is monitored in real time by a high-precision temperature sensor installed in the sintering furnace, and the data is fed back to the temperature controller. The temperature controller adjusts the power output of the heater in real time according to the deviation between the actual temperature and the preset temperature to ensure that the temperature is maintained near the set value. The sintering process is usually accompanied by a preset heating and insulation stage, and the specific time arrangement is determined according to the thermal stability of the polyimide film and the thermal conductivity of the copper flat wire. In the heating stage, the system raises the temperature to the target sintering temperature at a preset rate, and maintains this temperature for a period of time in the insulation stage to ensure that the polyimide film is completely melted and fully attached to the surface of the copper flat wire. After sintering is completed, the system immediately enters the annealing cooling stage. The annealing cooling module also sets a preset cooling temperature and corresponding cooling time. The main purpose of annealing cooling is to improve the mechanical strength and durability of polyimide film by gradually reducing the temperature and eliminating the stress inside the material. At this stage, the temperature controller adopts a staged cooling strategy based on the thermodynamic characteristics of the annealing process to prevent the accumulation of thermal stress in the material caused by too fast cooling. The precise control of the cooling process not only relies on the real-time data of the temperature sensor, but also combines the calculation results of the heat conduction model to ensure that the rate of temperature drop matches the cooling characteristics of the material. The entire sintering and annealing cooling process is strictly executed according to the preset time nodes on the timeline.

[0029] Furthermore, the construction of a decision model through deep learning based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire includes the following steps: Define the state space, including the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire; define the action space, including the adjustment of the wrapping rate and the advancement rate; generate a reward function based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate and the overlap rate; A decision model is built based on the state space, action space and reward function, and reinforcement learning training is performed using historical data. The proximal strategy optimization algorithm is used to limit the difference between the new and old strategies during the iterative training process.

[0030] In some embodiments, first, a state space is defined. The state space includes key parameters in the production process, such as the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the current wrapping rate, the width of the polyimide film, and the advancement rate of the copper flat wire. These parameters together describe the current state of the production process and are the basic data for the input of the decision model. The definition of the state space ensures that the model can fully perceive the various factors in the production process, so as to make accurate process adjustments. Next, the action space is defined. The action space includes the adjustment range of the wrapping rate and the advancement rate. For each state, the model can choose to adjust the wrapping rate or the advancement rate, or keep it unchanged. The definition of the action space directly affects the adjustment strategies that the model can adopt in different states, which will further affect the production results. In order to enable the model to self-learn and improve during the optimization process, a reward function based on the insulation rate difference and the overlap rate is constructed. The reward function is used to quantify the effect of the model after selecting a certain action in a specific state. Specifically, when the insulation rate of the copper flat wire approaches or reaches the preset standard and the overlap rate is reduced, a positive reward is given; if the insulation rate deviates from the standard or the overlap rate is high, a negative reward is given. The design of the reward function aims to guide the model to minimize material waste and optimize process parameters while ensuring insulation performance. During the model training process, the proximal policy optimization algorithm (PPO) in reinforcement learning was adopted. First, the system generates a large number of state-action pairs through historical production data and simulation environments for the model to conduct preliminary learning. In each iteration, the PPO algorithm limits the magnitude of model updates by calculating the difference between the current strategy and the old strategy to avoid instability caused by excessive changes in the strategy. Through careful adjustment of the strategy, the PPO algorithm ensures that the model's strategy is gradually optimized after each iteration and eventually converges to a stable and effective strategy.

[0031] The training process of the model is continuously iterated and optimized, so that the model can accurately predict and execute the optimal process adjustment strategy under different production conditions. With the continuous increase of training data, the decision-making ability of the model is gradually improved, and it can realize real-time optimization and adjustment in the production process, ensuring the continuous improvement of production quality and efficiency.

[0032] Furthermore, the objective function formula of the proximal strategy optimization algorithm is as follows: ; in, is the objective function; represents the expected value at time step t; is the probability ratio of the new strategy to the old strategy at time step t; is the advantage function, which is calculated by the difference between the actual return and the baseline; is a hyperparameter; To limit the probability ratio Function used to clip Restricted to and between; Indicates the minimum value operation.

[0033] Specifically, first, the objective function takes into account the change between the new strategy and the old strategy, that is, by calculating the probability ratio of their selection of a certain action in a specific state. This ratio is used to evaluate whether the new strategy deviates significantly from the old strategy, thereby helping the model maintain a stable strategy update. In order to prevent excessive policy adjustment, PPO introduces a clipping mechanism. Specifically, the objective function limits the probability ratio of the new and old strategies to a specific range when calculating the strategy update. This range is determined by a small hyperparameter control, usually set to 0.1 or 0.2. This means that if the adjustment of the new strategy exceeds this range, the algorithm will clip it back to the predetermined range to avoid drastic changes during the update process. In addition, the PPO algorithm introduces an advantage function to measure the difference between the actual performance of the current strategy and the expected performance. The advantage function determines whether an action is better than the average level by comparing the actual return of the model in a certain state with the baseline return. This process helps the model to more accurately evaluate the effectiveness of each action during the optimization process. By combining the clipping mechanism and the advantage function in the objective function, the PPO algorithm achieves stability and effectiveness in reinforcement learning. This design enables the model to gradually improve its decision-making ability during training without causing instability or poor training results due to too fast or too slow policy updates. Ultimately, the PPO algorithm enables the model to learn to make optimal decisions in complex environments through multiple iterative processes, thereby improving the automation and optimization level of the production process.

[0034] Furthermore, the reward function includes: If the difference between the insulation rate of the copper flat wire and the preset standard insulation rate is a positive value or zero, the reward function is a positive reward and is negatively correlated with the overlap rate; If the difference between the insulation rate of the copper flat wire and the preset standard insulation rate is a negative value, the reward function is a negative reward and is irrelevant to the overlapping rate.

[0035] Specifically, when the difference between the insulation rate of the copper flat wire and the preset standard insulation rate is positive or zero: this situation means that the insulation performance of the copper flat wire meets or exceeds the preset standard. In this case, the reward function gives a positive reward, indicating positive feedback to the current strategy. At the same time, the reward value is negatively correlated with the overlap rate, that is, the lower the overlap rate, the higher the reward value. This design is intended to encourage the model to minimize the overlap rate while ensuring insulation performance, thereby reducing material costs. When the difference between the insulation rate of the copper flat wire and the preset standard insulation rate is a negative value: at this time, the insulation performance of the copper flat wire does not meet the preset standard. In order to guide the model to prioritize insulation performance, the reward function gives a negative reward in this case, indicating that the current strategy needs to be adjusted. In this scenario, the overlap rate will not affect the reward value, because the primary goal of the model at this time is to adjust the process parameters to improve the insulation rate and ensure that product quality meets the standards.

[0036] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An automatic control system for the stacking rate of a polyimide sintering line, characterized in that: It includes a pre-check module, a wrapping module, a stacking rate monitoring module, a sintering module, an annealing and cooling module, a detection and verification module and a stacking rate control module which are connected in sequence, and the stacking rate control module is also connected to the detection and verification module, the stacking rate monitoring module and the wrapping module; The pre-check module is used to identify the copper flat wire by moving the light source and image recognition to confirm the characterization state of the copper flat wire, wherein the characterization state includes the presence of a characterization defect and the absence of a characterization defect; The wrapping module is used to control the wrapping rate of the polyimide film, and is also used to control the advancing rate of the copper flat wire whose characteristic state is that there is no characteristic defect; The overlap rate monitoring module is used to perform semantic segmentation on the polyimide film and the copper flat wire in the image data of the wrapping process based on the semantic segmentation model, and calculate the overlap rate of the polyimide film by the overlap rate of the polyimide film width on the copper flat wire; The sintering module is used to control the sintering temperature of the copper flat wire that has been laminated with a polyimide film, and the annealing and cooling module is used to control the temperature of the annealing and cooling of the copper flat wire that has been sintered; the detection and verification module is used to detect the insulation rate of the copper flat wire that has been annealed and cooled, and calculate the difference between the insulation rate of the copper flat wire and the preset standard insulation rate; The overlapping rate control module is used to construct a decision model through deep learning based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire, and generate adjustment decisions on the advancement rate and wrapping rate of the copper flat wire through the decision model.

2. The automatic control system for the stacking rate of a polyimide sintering line according to claim 1, characterized in that: The characterization and identification of the copper flat wire by moving the light source and image recognition to confirm the characterization state of the copper flat wire comprises the following steps: Control the moving light source to move along a preset trajectory and illuminate the surface of the copper flat wire; Acquire image data of several angles of the surface of the copper flat wire when the moving light source moves; The color value trend is determined for the image data of several angles on the surface of the copper flat wire, and the abnormal points that do not conform to the color value trend are extracted; Determine whether there are characteristic defects on the surface of the copper flat wire based on the size of the abnormal point.

3. The automatic control system for the stacking rate of a polyimide sintering line according to claim 1, characterized in that: The packet overlap rate monitoring module is used to perform the following steps: Acquire image data of the copper flat wire being wrapped by the polyimide film during the wrapping process; Preprocess the image data, including image denoising, contrast enhancement and distortion correction; The preprocessed image data is input into the pre-trained semantic segmentation model to classify the polyimide film and the copper flat wire in the image, and the segmentation masks of the polyimide film and the copper flat wire are obtained; Extracting the width of the polyimide film through a segmentation mask of the polyimide film; The overlap rate of the polyimide film is calculated by the width of the overlapped area of ​​the polyimide film on the copper flat wire and the width of the polyimide film.

4. The automatic control system for the stacking rate of a polyimide sintering line according to claim 3, characterized in that: The training steps of the semantic segmentation model include: Obtain an image dataset of the polyimide film and copper flat wire wrapping process, and perform pixel-level data annotation on the polyimide film and copper flat wire in the image; Constructing a semantic segmentation model, wherein the semantic segmentation model includes an encoder and a decoder, wherein the encoder downsamples and extracts features of an input image, reduces the spatial resolution of a feature map and increases the number of channels of the feature map through convolution operations and pooling operations, and captures semantic information of the image; The decoder performs upsampling and semantic segmentation on the features extracted by the encoder, restores the spatial resolution of the feature map through deconvolution layers and skip connections, and maps the semantic information extracted by the encoder to pixel-level segmentation results; The image dataset with pixel-level data annotation is input into the semantic segmentation model for training, with the cross entropy loss function as the optimization target, and the model parameters are iteratively updated through the stochastic gradient descent algorithm.

5. The automatic control system for the stacking rate of a polyimide sintering line according to claim 4, characterized in that: The formula for the convolution operation is as follows: ； in, is the input image; For the input image In the The feature output of the layer; Is the connection Layer and The weight matrix of the kth convolution kernel of the layer; For the input image In the The feature output of the layer; It is Bias of the layer; is the activation function; is the number of convolution kernels.

6. The automatic control system for the stacking rate of a polyimide sintering line according to claim 1, characterized in that: The sintering temperature control and the annealing cooling temperature control are respectively performed by presetting the sintering working temperature and the annealing cooling temperature and their corresponding preset times.

7. The automatic control system for the stacking rate of a polyimide sintering line according to claim 1, characterized in that: The method of constructing a decision model through deep learning based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire includes the following steps: Define the state space, including the difference between the insulation rate of the copper flat wire and the preset standard insulation rate, the wrapping rate, the width of the polyimide film and the advancement rate of the copper flat wire; define the action space, including the adjustment of the wrapping rate and the advancement rate; generate a reward function based on the difference between the insulation rate of the copper flat wire and the preset standard insulation rate and the overlap rate; A decision model is built based on the state space, action space and reward function, and reinforcement learning training is performed using historical data. The proximal strategy optimization algorithm is used to limit the difference between the new and old strategies during the iterative training process.

8. The automatic control system for the stacking rate of a polyimide sintering line according to claim 7, characterized in that: The objective function formula of the proximal strategy optimization algorithm is as follows: ； in, is the objective function; represents the expected value at time step t; is the probability ratio of the new strategy to the old strategy at time step t; is the advantage function, which is calculated by the difference between the actual return and the baseline; is a hyperparameter; To limit the probability ratio Function used to clip Restricted to and between; Indicates the minimum value operation.

9. The automatic control system for the stacking rate of a polyimide sintering line according to claim 7, characterized in that: The reward function includes: If the difference between the insulation rate of the copper flat wire and the preset standard insulation rate is a positive value or zero, the reward function is a positive reward and is negatively correlated with the overlap rate; If the difference between the insulation rate of the copper flat wire and the preset standard insulation rate is a negative value, the reward function is a negative reward and is irrelevant to the overlapping rate.

Images