A hot-pressing intelligent control regulation method and device based on FPC

Abstract

The application provides a hot-pressing intelligent control regulation method and device based on FPC, and belongs to the technical field of FPC hot-pressing processing. The method synchronously collects the surface image and three-dimensional topographic point cloud data of the FPC substrate through a camera and a laser displacement meter, realizes real-time detection based on conditions, intelligently switches the feature extraction mode according to the interference detection result, adopts an edge detection algorithm when no interference occurs, starts a Bayesian Dropout network to perform multiple random sampling inferences when interference occurs, activates a digital twin to predict deformation characteristics if the coefficient of variation is greater than a preset threshold, calculates the hot-pressing parameter compensation amount according to the deformation characteristics and adds a tolerance weight, and continuously optimizes the control effect through closed-loop feedback. The application effectively solves the image feature extraction distortion problem caused by optical interference, and significantly improves the hot-pressing control precision and product yield.

Description

Technical Field

[0001] This invention relates to the field of FPC hot pressing technology, and specifically to a method and device for intelligent control and regulation of hot pressing based on FPC. Background Technology

[0002] In today's electronics manufacturing industry, FPCs (flexible printed circuit boards) are widely used in numerous industries, including consumer electronics, automotive electronics, medical devices, and aerospace, due to their significant advantages such as thinness, flexibility, high wiring density, and adaptability to complex spatial layouts. In smartphones, FPCs are used to connect components such as displays, camera modules, batteries, and motherboards to achieve signal transmission and power supply; their thinness contributes to the design of slimmer and lighter phones. In automotive electronics, FPCs are used in vehicle displays, sensors, and control systems, enabling efficient wiring within limited space and improving system reliability. In medical devices, such as wearable health monitoring devices, the flexibility of FPCs allows them to better conform to the human body, enabling convenient health data collection.

[0003] As electronic products evolve towards miniaturization and high performance, higher demands are being placed on the quality and production efficiency of FPCs. Roll-to-roll transport systems, due to their ability to enable continuous production, are widely used in the mass production of FPCs. During the hot-pressing process of this system, how to accurately control the hot-pressing parameters to ensure the processing quality of the FPCs has become a key focus of the industry.

[0004] An existing intelligent control method for hot pressing based on FPC (Flexible Printed Circuit) includes the following core steps: 1) Continuously acquiring surface deformation images of the FPC substrate before hot pressing using a high-speed linear CMOS camera integrated on the feeding and receiving mechanism at a rate of no less than 50 frames per second; 2) Performing grayscale conversion and edge detection processing on the acquired surface deformation images sequentially, specifically including image noise reduction using Gaussian filtering, edge feature enhancement using the Sobel operator, extracting the FPC edge contour based on the Canny algorithm and generating a binary contour map; 3) Calculating the proportion of pixels with a curvature radius less than a critical value in the contour map as the wrinkle density; and 4) Statistically calculating the standard deviation of the curvature of the contour line segments as the deformation dispersion index. The system generates real-time deformation characteristic values ​​for the FPC; compares these real-time deformation characteristic values ​​with a preset deformation tolerance threshold range. This threshold range is determined by collecting deformation data of qualified FPCs from historical production, using the ±3σ principle to define the threshold boundary. When the characteristic value exceeds the threshold range, a control command is triggered. Based on the control command, the system calculates hot pressing compensation parameters in real time. When the pleat density exceeds the upper limit threshold, the temperature is increased; when the deformation dispersion exceeds the upper limit threshold, the pressure is increased. The calculated hot press head temperature compensation, pressure compensation, and conveying speed correction coefficient are input into the hot press control system to drive the hot pressing mechanism to perform the compensated hot pressing operation, while simultaneously adjusting the torque of the tension rollers of the feeding and receiving mechanism.

[0005] However, existing technologies have some significant drawbacks that severely impact the production quality and efficiency of FPCs. Firstly, abrupt changes in the reflective properties of the FPC material surface can cause distortion in image feature extraction. While the probability of this occurring during high-speed roll-to-roll transport is less than 0.1%, it still leads to errors in the calculation of deformation feature values ​​(wrinkle density, deformation dispersion). For example, when a bright area appears on the FPC surface due to a sudden change in reflective properties, these areas may be incorrectly identified as edge features during image grayscale conversion and edge detection. This causes deviations in the calculated wrinkle density and deformation dispersion, triggering incorrect hot-pressing parameter compensation, ultimately resulting in over- or under-pressurization of the material and leading to product quality issues.

[0006] Secondly, uneven coating thickness or foreign matter adhesion (such as oil stains or dust) on the FPC material surface can also cause interference. Under high frame rate (≥50fps) camera capture, localized changes in reflectivity can lead to bright noise or false shadows in the grayscale image. While Gaussian filtering can suppress this interference to some extent under normal operating conditions, when the interfering area overlaps with the actual wrinkle edge, the Sobel operator amplifies the false edge, and the Canny algorithm misidentifies it as a valid contour. Although the FPC coating process fluctuation rate is typically below 0.05%, the cumulative probability can rise to 5%-8% when a roll-to-roll system operates continuously for thousands of hours. This is significant in large-scale production and can result in a large number of defective products.

[0007] Furthermore, existing extraction algorithms rely on gradient information to calculate curvature distribution, making them vulnerable to interference. When false edges appear, the calculated radius of curvature based on gradients leads to abnormally low values. For example, in actual production, false edges caused by foreign matter adhering to the surface can cause the calculated radius of curvature to be much smaller than the true value, resulting in an artificially high wrinkle density (e.g., a measured value of 0.35 when the actual value is 0.12). Simultaneously, the standard deviation of the contour line segment curvature also increases due to noise interference. When subjected to noise interference caused by factors such as uneven surface coating thickness, the curvature calculation of the contour line segment is affected, leading to increased deformation dispersion and triggering overcompensation of the pressure compensation. This overcompensation can cause deviations in pressure regulation during hot pressing, affecting the hot pressing quality of the FPC and reducing the product yield. Summary of the Invention

[0008] The purpose of this invention is to provide a method and device for intelligent control of hot pressing based on FPC. By means of multimodal sensor fusion, intelligent interference detection and Bayesian uncertainty modeling, this invention solves the technical problems of image feature extraction distortion and inaccurate hot pressing parameter compensation caused by optical interference in the existing FPC hot pressing intelligent control technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A method for intelligent control and regulation of hot pressing based on FPC includes the following steps:

[0011] S1: Multimodal data synchronous acquisition: Simultaneously acquire surface images and three-dimensional topographic point cloud data of FPC substrate through a camera and a laser displacement meter;

[0012] S2: Real-time detection of optical interference: The acquired FPC surface image is converted to grayscale, and the bright noise areas in the image are identified and their area proportions are calculated. A n The standard deviation of surface undulation amplitude is calculated from the three-dimensional point cloud data obtained from the laser displacement gauge. s h When the difference between the area ratio of the highlighted region and the standard deviation satisfies Under certain conditions, optical interference is determined to have occurred, where k For calibration coefficients, d This is the tolerance threshold;

[0013] S3: Robust Extraction of Deformation Features: The feature extraction mode is switched based on the interference detection results. When no interference occurs, the edge detection algorithm is used to extract the image edge contour and calculate the wrinkle density. D w and deformation dispersion σc When interference occurs, the Bayesian Dropout network is activated to perform multiple random sampling and inference on the image, and outputs... D w and σc The distribution mean m And the coefficient of variation (CV). If the CV is greater than a preset threshold, the digital twin's predicted deformation features are activated.

[0014] S4: Dynamic compensation of hot pressing parameters: based on D w and σc Calculate the temperature compensation amount of the hot press head ΔT Pressure compensation amount ΔP and transmission speed correction factor Kv When the eigenvalues ​​come from a digital twin, the compensation amount is weighted with tolerance. or ;

[0015] S5: Closed-loop execution and feedback: Drive the hot pressing mechanism to execute compensation parameters and collect the flatness data of the FPC after hot pressing in real time to update the parameters of the digital twin model.

[0016] Furthermore: the camera is equipped with an infrared light source and a polarizing filter, the laser displacement meter is installed colinearly with the camera with an installation accuracy of ±1μm, the angle θ between the laser displacement meter and the camera optical axis is ≤5°, and the point cloud data sampling density is not less than 200 points / cm².

[0017] Furthermore: the tolerance threshold d This was determined by statistical analysis of production samples under normal historical conditions. ,in μA n This represents the average proportion of the area of ​​bright noise regions in an image under normal conditions. ms h This represents the average standard deviation of FPC surface morphology fluctuations under normal conditions.

[0018] Furthermore, the Bayesian Dropout network introduces a Dropout mechanism into the convolutional layers during the inference phase, randomly discarding 20% ​​of the convolutional kernel connections, performing 100 forward propagations, calculating the coefficient of variation (CV), and using the distribution mean μ when CV ≤ 0.3, and activating digital twin prediction when CV > 0.3.

[0019] Furthermore, the digital twin is constructed based on a thermo-mechanical coupling physical model, including transient heat conduction control equations and deformation calculation expressions under temperature coupling. The equivalent deformation field of the entire FPC is calculated using a finite element solver, and the maximum gradient value of the deformation field is extracted as the wrinkle density. D w Alternative indices are used to calculate the variability index of the deformation field distribution as the deformation dispersion. σc .

[0020] Furthermore: the tolerance weight η is constructed using an exponential function based on the cumulative time t of the continuous calls to the digital twin model. Where γ is the forgetting factor, controlling or The rate of increase over time.

[0021] Furthermore: when the coefficient of variation output by the Bayesian Dropout network is greater than a preset threshold, the LSTM time series prediction module is invoked, using the historical feature sequence under normal conditions within the previous 10 minutes as input to predict the deformation features at the current moment.

[0022] The present invention also provides a hot-pressing intelligent control and regulation device based on FPC, comprising:

[0023] The data acquisition module includes a camera and a laser displacement meter. The camera is used to acquire surface images of the FPC substrate, and the laser displacement meter is used to acquire three-dimensional topographic point cloud data.

[0024] The interference detection module is used to perform grayscale processing on the surface image, identify bright noise areas and calculate the area ratio, and determine optical interference by combining the surface undulation standard deviation of the point cloud data.

[0025] The feature extraction module includes an edge detection unit, a Bayesian Dropout network unit, and a digital twin prediction unit, which switch working modes to extract deformation features based on interference detection results.

[0026] The parameter compensation module is used to calculate the temperature compensation amount, pressure compensation amount, and conveying speed correction coefficient of the hot press head based on the deformation characteristics.

[0027] The control and execution module is used to drive the hot pressing mechanism to execute compensation parameters and collect feedback data;

[0028] The model update module is used to update the parameters of the digital twin model based on feedback data.

[0029] Furthermore: the camera is equipped with an 850nm / 940nm dual-band infrared light source and a polarizing filter, the laser displacement meter is installed colinearly with the camera with an installation accuracy of ±1μm, and the angle θ between the laser displacement meter and the camera optical axis is ≤5°.

[0030] Furthermore, the digital twin prediction unit is constructed based on a thermo-mechanical coupled physical model, including a finite element solver for solving the transient heat conduction equations and stress-strain equations, and calculating the equivalent deformation field distribution across the entire FPC domain.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] I. This invention effectively solves the problem of image feature extraction distortion caused by optical interference, significantly improving the accuracy of hot-press control. Existing technologies rely solely on camera image acquisition, which is susceptible to factors such as abrupt changes in the reflective properties of the FPC surface, uneven coating thickness, or foreign object adhesion, leading to errors in feature value calculation. This invention employs multimodal data synchronous acquisition technology, simultaneously acquiring surface images and 3D topographic point cloud data through a camera and a co-linearly mounted laser displacement meter. The camera is equipped with an infrared light source and a polarizing filter. Combined with 3D data from the laser displacement meter for verification, the accurate determination of optical interference is achieved through the condition |An-k×σh|>δ, reducing the impact of optical interference at its source and ensuring data reliability.

[0033] Second, robust extraction of deformation features is achieved, avoiding errors in feature value calculation due to interference. Existing extraction algorithms rely on gradient information to calculate curvature distribution, which is significantly vulnerable to interference and prone to overcompensation. This invention intelligently switches feature extraction modes based on interference detection results. When no interference occurs, a conventional edge detection algorithm is used. When interference occurs, a Bayesian Dropout network is activated to perform multiple random sampling inferences. The prediction uncertainty is assessed by calculating the coefficient of variation (CV). If the CV is greater than a preset threshold, the digital twin is activated to predict deformation features, thus constructing a multi-level feature extraction redundancy mechanism, effectively improving the anti-interference capability of feature extraction.

[0034] Third, this invention ensures the accuracy and stability of hot-pressing parameter compensation, thereby improving product yield. Existing technologies suffer from interference-induced errors in feature value calculation, leading to deviations in hot-pressing parameter compensation and resulting in a large number of defective products. In the dynamic compensation of hot-pressing parameters, this invention adds a tolerance weight η=1-e^(-γt) when the feature value comes from the digital twin. This weight is dynamically adjusted using an exponential function based on the cumulative time of continuous calls to the digital twin model, realizing a fusion strategy that dynamically adjusts the model's reliability over time. This makes the control commands more robust under uncertain conditions. Simultaneously, through a closed-loop execution and feedback mechanism, the flatness data of the FPC after hot pressing is collected in real time to update the digital twin model parameters, further optimizing the compensation effect and effectively improving product yield.

[0035] Fourth, an adaptive optical interference judgment mechanism has been established, reducing false positives and false negatives. Existing technologies do not scientifically set the optical interference judgment threshold, making them susceptible to environmental influences. This invention, through statistical analysis of production samples under historical normal conditions, employs an adaptive threshold setting method of δ=0.15×(μAn+μσh), enabling the threshold to adapt to the normal fluctuation range of different batches of materials and ambient light. This improves the accuracy of identifying abnormal events such as abnormal reflections, dust adhesion, or measurement deviations, avoiding interference from false positives and false negatives in hot-pressing control. Attached Figure Description

[0036] Figure 1 This is an overall architecture diagram of the FPC-based hot-press intelligent control and regulation method and device of the present invention;

[0037] Figure 2 This is a schematic diagram of the optical interference detection principle of the present invention;

[0038] Figure 3 This is a flowchart of the robust extraction process for deformation features in this invention;

[0039] Figure 4 This is a schematic diagram of the dynamic compensation and closed-loop feedback of hot pressing parameters in this invention;

[0040] Figure 5This is a schematic diagram of the structure of the FPC-based hot-press intelligent control and regulation device of the present invention;

[0041] Figure 6 This is a structural side view of the data acquisition module of the present invention. Detailed Implementation

[0042] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0044] like Figure 1 As shown, the FPC-based intelligent control and regulation method for hot pressing provided by this invention includes five main steps: synchronous acquisition of multimodal data, real-time detection of optical interference, robust extraction of deformation features, dynamic compensation of hot pressing parameters, and closed-loop execution and feedback.

[0045] In the multimodal data synchronous acquisition step, a high-speed linear array CMOS camera was selected as the camera, which has an image acquisition rate of no less than 50 frames per second, meeting the high-speed acquisition requirements for surface deformation images of FPC substrates. The camera is equipped with an 850nm / 940nm dual-band infrared light source and a polarizing filter, which can effectively suppress reflection and penetrate the oil stain layer to obtain the underlying texture. A laser displacement meter with an accuracy of ±1μm was selected. During installation, it was ensured that the angle θ between the laser displacement meter and the camera's optical axis was ≤5°, and the point cloud data sampling density was no less than 200 points / cm², to provide high-precision three-dimensional topographic point cloud data. The camera and laser displacement meter were colinearly installed on the take-up and untake-up mechanism of the roll-to-roll transport system to ensure that both could synchronously acquire FPC substrate surface information in the same area.

[0046] like Figure 2As shown, in the real-time optical interference detection step, the system performs grayscale processing on the acquired FPC surface image, converting it into a single-channel brightness image to remove color interference and reduce the computational complexity of subsequent processing. Subsequently, a fixed threshold or adaptive threshold method is used to identify areas in the grayscale image whose brightness is significantly higher than the surrounding background. These areas are usually caused by external infrared reflection, dust, or oil stains, manifesting as bright noise points in the image. After binarizing these noise areas, their area proportion in the entire image is calculated, denoted as An, which reflects the relative scale of the interfered area in the image. Simultaneously, the Z-axis height information of the current frame FPC surface is extracted from the 3D point cloud data obtained from the laser displacement meter, and the standard deviation of its fluctuation amplitude is calculated, denoted as... s h This parameter is used to quantify the degree of fluctuation in the surface morphology of the substrate. When the difference between the area ratio of the high-brightness region and the standard deviation meets the following condition...

[0047]

[0048] Under certain conditions, the system determines that abnormal bright spots in the image are likely caused by optical interference rather than actual deformation structures. Here, k is a calibration coefficient obtained based on historical data fitting, used to match the orders of magnitude of the two types of data, and δ is the maximum tolerance threshold set by the system.

[0049] The tolerance threshold δ is set using an adaptive method, which calculates the proportion of bright noise area by statistically analyzing production samples under historical normal conditions. A n Standard deviation of point cloud fluctuation s h The mean, denoted as μA n and ms h These two statistics reflect the typical performance of image and 3D data under undisturbed conditions, respectively. Subsequently, a baseline tolerance level is constructed based on the weighted average of the two, using the formula:

[0050]

[0051] The constant coefficient 0.15 represents the empirically adjusted tolerance control ratio, which is used to limit the maximum permissible inconsistency between the image and the topographic information.

[0052] like Figure 3 As shown, in the robust deformation feature extraction step, the system switches the feature extraction mode based on the interference detection results. When no interference occurs, the Sobel-Canny algorithm is used to extract the image edge contours, and the wrinkle density Dw and deformation dispersion are calculated. σcSpecifically, this includes using Gaussian filtering to reduce image noise, enhancing edge features through the Sobel operator, extracting FPC edge contours based on the Canny algorithm and generating a binary contour map, then calculating the proportion of pixels with a radius of curvature less than a critical value in the contour map as the wrinkle density, and statistically analyzing the standard deviation of the curvature of the contour line segments as the deformation dispersion index.

[0053] When interference occurs, the system initiates a Bayesian Dropout network to perform multiple random sampling inferences on the image. This network is built on the U-Net architecture, taking a grayscale FPC surface image as input and outputting a probability heatmap of the deformed regions. To achieve Bayesian inference characteristics, a Dropout mechanism is introduced into the convolutional layers during the inference phase, randomly discarding 20% ​​of the convolutional kernel connections. This creates structural perturbations in each forward propagation, resulting in statistical volatility in the output. Under the condition of the same input image, the network performs 100 forward propagations, obtaining 100 sets of deformed region segmentation results. From these, the corresponding wrinkle density and deformation dispersion features are extracted, forming a feature set.

[0054]

[0055] in The wrinkle density is calculated from the i-th sampling result. This represents the deformation dispersion.

[0056] The specific architecture of the Bayesian Dropout network is as follows: the encoder part contains 4 convolutional blocks, each containing two 3×3 convolutional layers, with the number of filters being 32, 64, 128, and 256 respectively. Each convolutional layer is followed by a ReLU activation function and batch normalization. The decoder part adopts a symmetrical structure, upsampling through transposed convolution, and establishing skip connections with the corresponding layers of the encoder. A Dropout layer is added after the second convolutional layer of each convolutional block, with a dropout rate set to 0.2.

[0057] Subsequently, the mean and standard deviation of these 100 sets of features were statistically analyzed, and the coefficient of variation was calculated to assess the prediction uncertainty. Specifically,

[0058]

[0059] in σD w and μD w These represent the standard deviation and mean of the fold density, respectively. ssc and msc Here, represents the standard deviation and mean of the deformation dispersion. When any coefficient of variation is greater than 0.3, it indicates that the neural network output is highly volatile, the prediction results are not reliable, and the system determines that this set of features is unreliable.

[0060] When a feature is deemed unreliable, the system automatically invokes the digital twin prediction module. This module first constructs a thermo-mechanical coupled physical model to simulate the temperature field evolution of the FPC during hot pressing and the structural response caused by thermal expansion and external forces. The temperature field satisfies the transient heat conduction control equation.

[0061]

[0062] in r For material density, C p For specific heat capacity, k Thermal conductivity, Qheater This refers to the heating power of the hot press head. h The thermal convection coefficient, T∞ The ambient air temperature. T This represents the current temperature distribution of the FPC. The equation is discretized and solved using a finite element thermal module, yielding the temperature distribution at each time step and spatial point.

[0063] Subsequently, a deformation calculation expression under temperature coupling is introduced.

[0064]

[0065] in α The coefficient of thermal expansion is... T 0 is the initial reference temperature. F p The normal pressure applied to the hot press head, E The elastic modulus of the material. A For the area of ​​force application, e eff The equivalent linear deformation is used. This model considers both thermal and mechanical strain and establishes a three-dimensional structural response prediction over the entire FPC domain. The currently acquired hot compressor temperature is input into this model. T ,pressure F p and transmission speed v The equivalent deformation field of the entire FPC region was numerically calculated using the finite element solver based on the heat dissipation conduction equation and the stress-strain equation set. e eff .

[0066] To extract characteristic indicators, the gradient of the spatial distribution of the deformation field is calculated, and the maximum gradient value is obtained.

[0067]

[0068] This value represents the intensity of local deformation abrupt changes and serves as a proxy for wrinkle density (Dw), accurately describing the intensity of areas with concentrated local wrinkles or depressions. Simultaneously, to assess the dispersion of the overall deformation field distribution, its variability index is calculated.

[0069]

[0070] This value reflects the degree of uniformity in deformation of the FPC after hot pressing. A larger standard deviation and a smaller mean indicate better performance. σc A higher value indicates more severe fluctuations in local strain. Through the extraction of these three quantities, the system ultimately outputs the deformation characteristic value used for thermocompression compensation control, namely, the wrinkle density. D w and deformation dispersion σc This enables highly reliable predictions based on digital twin modeling in interference environments.

[0071] The specific implementation of the digital twin prediction module includes: geometric modeling using tetrahedral meshes to divide the FPC geometric domain, with a unit size of 0.1 mm; boundary conditions set to apply temperature boundary conditions to the hot press head contact surface, and natural convection boundaries to the remaining surfaces; material property parameters are set as follows: density ρ = 1200 kg / m³, specific heat capacity Cp = 1400 J / (kg·K), thermal conductivity k = 0.3 W / (m·K), and coefficient of thermal expansion. elastic modulus Poisson's ratio ν = 0.35.

[0072] The thermo-mechanical coupling solution employs a sequential coupling method: first, the transient heat conduction equation is solved to obtain the temperature field distribution. Then, the temperature field is used as a thermal load in the structural mechanics equations. Perform stress-strain analysis and iterate until convergence. The convergence criterion is that the relative error of the temperature field between two consecutive iterations is less than 1. .

[0073] The specific algorithm for extracting deformation features from the finite element solution results is as follows: calculate the displacement gradient of each node. The wrinkle density Dw is equal to the maximum value of the displacement gradient modulus divided by the average element size, and the deformation dispersion σc is equal to the standard deviation of the displacement gradient modulus divided by its mean. Finally, a Gaussian filter with a standard deviation of 0.5 is used to smooth the abrupt change points.

[0074] When the coefficient of variation output by the Bayesian Dropout network exceeds a preset threshold, the system can also call the LSTM time series prediction module as a backup. This module takes historical feature sequences under normal conditions from the 10 minutes prior to the current time as input, including...

[0075]

[0076] After extracting the temporal trend through a multi-layer LSTM encoder, the deformation characteristics at the current moment are predicted, and the output value is... ,

[0077] The prediction results, which incorporate historical evolution information, are highly robust and are ultimately used to replace unreliable Dropout output values, ensuring the continuity and stability of subsequent hot-press compensation parameters.

[0078] The complete algorithm flow for robust extraction of deformation features is as follows:

[0079] Step 1: Input the grayscale FPC surface image of the current frame. And laser displacement meter 3D point cloud data;

[0080] Step 2: Calculate the area ratio An of bright noise regions in the image and the standard deviation σh of point cloud surface undulation;

[0081] Step 3: Judgment Is it greater than the tolerance threshold δ? If not, proceed to step 4; if yes, proceed to step 5.

[0082] Step 4: Extract the contour using the Sobel-Canny edge detection algorithm, calculate the wrinkle density Dw and deformation dispersion σc, and then proceed to step 8;

[0083] Step 5: Start the Bayesian Dropout network and perform 100 forward propagations to obtain the feature distribution;

[0084] Step 6: Calculate the coefficient of variation (CV). If CV ≤ 0.3, use the distribution mean as the output and proceed to step 8.

[0085] Step 7: If CV > 0.3, call the digital twin module or LSTM time series prediction module to obtain the feature value;

[0086] Step 8: Output the final deformation characteristics (Dw, σc) for subsequent hot pressing parameter compensation.

[0087] During the training of the LSTM time series prediction module, to balance the accuracy of the deformation feature values ​​themselves with their fitting effect on the time evolution trend, the system uses a weighted loss function for optimization. This loss function consists of two parts: the absolute error between the predicted and true values, and the mean absolute error of their first-order differences. For each training sample, the... i The true value is y i The corresponding model prediction value is The loss function is defined as

[0088]

[0089] The first term is the mean absolute error between the predicted and actual values, ensuring the numerical accuracy of the deformation characteristics; the second term is the mean absolute error of the first-order difference, where... = Indicates the first i The change at each time step relative to the previous time step is used to ensure that the trend of the predicted sequence is consistent with the actual evolution.

[0090] Where N is the number of samples, The second term introduces the difference operator. Then, the differences in changes between adjacent time points in the sequence are compared to measure the degree of matching between the predicted curve and the actual trend, thereby enhancing the model's ability to fit characteristic fluctuation patterns. (Difference operator) This indicates that the first difference of the time series reflects the rate of change of deformation characteristics over time.

[0091] like Figure 4 As shown, in the dynamic compensation step of hot pressing parameters, the system calculates the fold density. D w and deformation dispersion σc The temperature compensation amount of the hot press head is calculated according to the preset compensation algorithm. ΔT Pressure compensation amount ΔP and transmission speed correction factor Kv When the wrinkle density exceeds the upper limit threshold, the temperature compensation is increased; when the deformation dispersion exceeds the upper limit threshold, the pressure compensation is increased. When the eigenvalue comes from the digital twin, the compensation is weighted by a tolerance weight η. This weight is constructed using an exponential function based on the cumulative time t of the digital twin model being continuously called.

[0092]

[0093] in c As a forgetting factor, it controls the rate at which η increases over time. t This represents the cumulative time, in seconds, during which digital twins have been continuously used as a source of deformation features. This formula constitutes a typical ascending exponential function, as... t Gradually increase or Approaching 1 means that the system's dependence on the digital twin's output is constantly increasing.

[0094] In the closed-loop execution and feedback step, the calculated hot-pressing compensation parameters are input into the hot-pressing machine control system to drive the hot-pressing mechanism to perform the compensated hot-pressing operation, while simultaneously adjusting the torque of the tension rollers of the feeding and take-up mechanism. After the hot-pressing process is completed, the flatness data of the FPC after hot pressing is collected in real time and fed back into the digital twin model to update the model parameters. This allows the digital twin to more accurately simulate the hot-pressing process of the FPC, providing a more reliable basis for the next hot-pressing parameter compensation.

[0095] like Figure 5-6 As shown, this invention also provides a hot-pressing intelligent control and regulation device based on FPC, including a data acquisition module, an interference detection module, a feature extraction module, a parameter compensation module, a control execution module, and a model update module. The data acquisition module includes a camera and a laser displacement meter. The camera is used to acquire surface images of the FPC substrate, and the laser displacement meter is used to acquire three-dimensional topographic point cloud data. The interference detection module performs grayscale processing on the surface image, identifies high-brightness noise areas and calculates their area proportions, and determines optical interference by combining the surface undulation standard deviation of the point cloud data. The feature extraction module includes an edge detection unit, a Bayesian Dropout network unit, and a digital twin prediction unit, switching working modes to extract deformation features based on the interference detection results. The parameter compensation module calculates the temperature compensation amount, pressure compensation amount, and conveying speed correction coefficient of the hot-pressing head based on the deformation features. The control execution module drives the hot-pressing mechanism to execute the compensation parameters and collects feedback data. The model update module updates the digital twin model parameters based on the feedback data.

[0096] The technical parameters of the laser displacement meter include: laser wavelength 655nm, measurement range ±2mm, resolution 0.1μm, and sampling frequency 2kHz; the infrared light source wavelengths of 850nm and 940nm are selected in a dual-band manner because this band has good penetration to the FPC substrate and can effectively suppress surface reflection interference; the polarization filter adopts linear polarization, polarization angle 45°, and transmittance ≥90%, which is used to further reduce the influence of specular reflection.

[0097] The key parameters in the Bayesian Dropout network are set based on the following: a dropout rate of 20% is a balance between ensuring the network's expressive power and introducing uncertainty, determined through grid search; 100 forward propagations ensure the stability of the statistics while also taking into account computational efficiency; and a coefficient of variation threshold of 0.3 is a confidence threshold determined through offline data statistics.

[0098] To verify the effectiveness of the technical solution of this invention, a practical application case test was conducted. On a roll-to-roll transport system hot-pressing production line, 1000 FPC substrates were selected for hot-pressing experiments. During the experiment, optical interference such as abrupt changes in the reflective properties of the FPC material surface, uneven coating thickness, and foreign object adhesion were artificially simulated. When using existing technology for hot-pressing control, optical interference caused distortion in image feature extraction, leading to errors in hot-pressing parameter compensation, ultimately resulting in 150 defective pieces and a yield rate of only 85%. However, by adopting the technical solution of this invention, through a three-level defense strategy of multimodal sensor fusion and predictive uncertainty assessment, the optical interference problem was effectively solved, and hot-pressing parameters were accurately calculated, ultimately resulting in only 30 defective pieces and a yield rate increased to 97%.

[0099] Under different working conditions, such as strong interference conditions where there are 2mm diameter oil stains on the surface of the FPC material, the wrinkle density calculated by existing technology... D w An error rate as high as 142% leads to severe deviations in hot-pressing parameters, resulting in quality problems such as over-pressure or under-pressure in the product. This invention, however, utilizes the synergistic effect of a Bayesian Dropout network and a digital twin to... D w With the error controlled within 9%, the FPC products after hot pressing have good flatness, stable performance, and meet production requirements.

[0100] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent control and regulation of hot pressing based on FPC, characterized in that, Includes the following steps: S1: Multimodal data synchronous acquisition: Simultaneously acquire surface images and three-dimensional topographic point cloud data of FPC substrate through a camera and a laser displacement meter; S2: Real-time detection of optical interference: The acquired FPC surface image is converted to grayscale, the bright noise areas in the image are identified and their area proportion An is calculated, and the standard deviation σh of the surface undulation amplitude is calculated from the 3D point cloud data obtained from the laser displacement meter. When the difference between the area proportion of the bright area and the standard deviation meets the condition... Under certain conditions, optical interference is determined to occur, where k is the calibration coefficient and δ is the tolerance threshold. S3: Robust extraction of deformation features: The feature extraction mode is switched based on the interference detection results. When no interference occurs, the edge detection algorithm is used to extract the image edge contour and calculate the wrinkle density Dw and deformation dispersion σc. When interference occurs, the Bayesian Dropout network is started to perform multiple random sampling inferences on the image and output the distribution mean μ of Dw and σc and the coefficient of variation CV. If CV is greater than the preset threshold, the digital twin is activated to predict deformation features. S4: Dynamic compensation of hot pressing parameters: Calculate the temperature compensation amount ΔT, pressure compensation amount ΔP and transmission speed correction coefficient Kv of the hot pressing head according to Dw and σc. When the characteristic value comes from the digital twin, the compensation amount is given an additional tolerance weight η. S5: Closed-loop execution and feedback: Drive the hot pressing mechanism to execute compensation parameters and collect the flatness data of the FPC after hot pressing in real time to update the parameters of the digital twin model.

2. The method according to claim 1, characterized in that, The camera is equipped with an infrared light source and a polarizing filter. The laser displacement meter is installed colinearly with the camera with an installation accuracy of ±1μm and an angle θ≤5° with the camera's optical axis. The point cloud data sampling density is not less than 200 points / cm².

3. The method according to claim 1, characterized in that, The tolerance threshold δ is determined statistically from production samples under historical normal conditions. , where μAn represents the average proportion of the area of ​​the bright noise region in the image under normal conditions, and μσh is the average standard deviation of the surface morphology fluctuation of the FPC under normal conditions.

4. The method according to claim 1, characterized in that, The Bayesian Dropout network introduces a Dropout mechanism into the convolutional layers during the inference phase, randomly discarding 20% ​​of the convolutional kernel connections, performing 100 forward propagations, calculating the coefficient of variation (CV), and using the distribution mean μ when CV ≤ 0.3, and activating digital twin prediction when CV > 0.

3.

5. The method according to claim 1, characterized in that, The digital twin is constructed based on a thermo-mechanical coupling physical model, including transient heat conduction control equations and deformation calculation expressions under temperature coupling. The equivalent deformation field of the entire FPC is calculated through a finite element solver. The maximum gradient value of the deformation field is extracted as a substitute index for the wrinkle density Dw, and the variation index of the deformation field distribution is calculated as the deformation dispersion σc.

6. The method according to claim 1, characterized in that, The tolerance weight η is constructed using an exponential function based on the cumulative time t of the digital twin model being continuously invoked. , where γ is the forgetting factor, which controls the rate at which η increases over time.

7. The method according to claim 1, characterized in that, When the coefficient of variation output by the Bayesian Dropout network exceeds a preset threshold, the LSTM time series prediction module is invoked. Using the historical feature sequence under normal conditions within the previous 10 minutes as input, the deformation features at the current moment are predicted.

8. A hot-pressing intelligent control and regulation device based on FPC, characterized in that, The apparatus is used to perform the method according to any one of claims 1 to 7, comprising: The data acquisition module includes a camera and a laser displacement meter. The camera is used to acquire surface images of the FPC substrate, and the laser displacement meter is used to acquire three-dimensional topographic point cloud data. The interference detection module is used to perform grayscale processing on the surface image, identify bright noise areas and calculate the area ratio, and determine optical interference by combining the surface undulation standard deviation of the point cloud data. The feature extraction module includes an edge detection unit, a Bayesian Dropout network unit, and a digital twin prediction unit, which switch working modes to extract deformation features based on interference detection results. The parameter compensation module is used to calculate the temperature compensation amount, pressure compensation amount, and conveying speed correction coefficient of the hot press head based on the deformation characteristics. The control and execution module is used to drive the hot pressing mechanism to execute compensation parameters and collect feedback data; The model update module is used to update the parameters of the digital twin model based on feedback data.

9. The apparatus according to claim 8, characterized in that, The camera is equipped with an 850nm / 940nm dual-band infrared light source and a polarizing filter. The laser displacement meter is installed colinearly with the camera, with an installation accuracy of ±1μm and an angle θ≤5° between it and the camera's optical axis.

10. The apparatus according to claim 8, characterized in that, The digital twin prediction unit is built based on a thermo-mechanical coupling physical model and includes a finite element solver for solving transient heat conduction equations and stress-strain equations to calculate the equivalent deformation field distribution across the entire FPC domain.

Images