Adhesive tape coating regulation and control optimization method based on intelligent image analysis

Abstract

The invention relates to an adhesive tape coating regulation and control optimization method and system based on intelligent image analysis, and the method comprises the following steps: S1, obtaining a product specification, current situation data and production line capability, building a quantitative index system and constraint conditions, determining a security domain based on historical data and a physical model, and obtaining a constraint set and the security domain; s2, constructing a multi-modal sensing network based on the constraint set and the security domain; s3, based on a multi-mode sensing network, collecting images and process data of each station, and preprocessing the images and the process data; s4, constructing a multi-task network analysis model, and performing analysis based on the preprocessed image and process data of each station to obtain an analysis result; and S5, based on an analysis result, adopting reinforcement learning to control strategy generation, and obtaining an optimal control sequence. The intelligent quality control efficiency and reliability of film coating are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and in particular to a method for optimizing tape coating based on intelligent image analysis. Background Technology

[0002] The coating and drying process of adhesive tapes (pressure-sensitive adhesive tapes, functional film tapes, etc.) is characterized by "complex materials, diverse indicators, strong coupling, and long hysteresis". Compared with general film coating, tape coating also has the added requirement of strict requirements for viscoelasticity-rheology, surface energy and interfacial adhesion of pressure-sensitive adhesive (PSA) system, making the control more complex.

[0003] High-end tapes (such as optical tapes for OLED displays and insulating tapes for lithium battery tabs) require coating thickness tolerances controlled within ±2~5μm, while also requiring transverse thickness profile (CD direction) fluctuations of less than 3%, which traditional open-loop processes struggle to consistently meet. This is especially true for medical tapes and electronic tapes, where issues like edge "fogging," "adhesive creep," and "adhesive overflow" directly impact product performance and aesthetics, necessitating increasingly stringent requirements for edge geometric precision.

[0004] Defects such as microbubbles, impurities, missed coatings, and adhesive residues have a significant impact on product performance. Moreover, maintaining quality stability under high-speed coating is a long-term challenge, especially for solvent-based adhesives, which are prone to defects such as bubbles and sagging at high speeds. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a tape coating control and optimization method based on intelligent image analysis, which effectively improves the efficiency and reliability of intelligent quality control for thin film coating.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for controlling and optimizing tape coating based on intelligent image analysis includes the following steps:

[0008] S1: Obtain product specifications, current status data, and production line capabilities; establish a quantitative indicator system and constraints; and determine the safety domain based on historical data and physical models, thereby obtaining the constraint set and safety domain.

[0009] S2: Construct a multimodal perception network based on constraint sets and security domains;

[0010] S3: Based on a multimodal sensing network, images and process data from each workstation are collected and preprocessed;

[0011] S4: Construct a multi-task network analysis model, and analyze the preprocessed images and process data of each workstation to obtain the analysis results;

[0012] S5: Based on the analysis results, a reinforcement learning control strategy is used to generate and obtain the optimal control sequence.

[0013] Furthermore, obtain product specifications, current status data, and production line capabilities, and establish a quantitative indicator system and constraints, as detailed below:

[0014] The product specifications include the target thickness tsp and tolerance ±tol; current data includes yield Y, defect spectrum distribution P(d), and process capability index Cpk; production line capability includes width W and line speed range v∈[v min ,v max [Actuator adjustable range and resolution;]

[0015] Establish a quantitative indicator system, including thickness fluctuation and contour indicators, defect density and severity, edge uniformity and glue buildup index, energy consumption and response time;

[0016] The thickness fluctuation and contour indicators are as follows:

[0017] Let the in-plane pixel thickness map be T, the horizontal direction be defined as x∈[0,W], and the machine direction be y;

[0018] Average thickness: ;

[0019] in, Let it be the expected function;

[0020] Thickness standard deviation: ;

[0021] Tolerance satisfies constraints:

[0022] ;

[0023] Among them, t sp t represents the target thickness; tol represents the thickness tolerance; t min ,t max These represent the minimum and maximum thickness, respectively.

[0024] Banner outline: ;

[0025] Where y0 is the initial machine direction; L is the banner length;

[0026] Contour basis function decomposition:

[0027] ;

[0028] in, c are the basis functions for contour decomposition; i The coefficients of the contour basis functions characterize the intensity of the contour shape components; The average thickness profile in the horizontal direction;

[0029] Process capability index C pk :

[0030] ;

[0031] USL and LSL are the upper and lower specification limits, respectively.

[0032] The specific defect density and severity are as follows:

[0033] Let N be the number of defect events within area A. d The chance defect number DPMO is:

[0034] ;

[0035] Where a0 is the unit opportunity area;

[0036] Severity Index:

[0037] ;

[0038] Where, α tk βs is the weight of defect type k. k For size grading, γ pk For functional area location weights; a k Let be the area of ​​the k-th defect;

[0039] Objectives and constraints:

[0040] DPMO≤DPMO max ,D≤D max ;

[0041] Among them, DPMO max D max These are the preset thresholds for the number of opportunity defects and the severity index, respectively;

[0042] The edge alignment and adhesive buildup index are as follows:

[0043] Let the edge position function be E(y), then the homogeneity deviation is:

[0044] ;

[0045] in, This represents the average value at the edge positions.

[0046] Edge adhesive buildup index:

[0047] ;

[0048] Where Δx is the edge bandwidth used to evaluate edge effects;

[0049] The energy consumption and response time are as follows:

[0050] Energy consumption per unit area E area :

[0051] ;

[0052] Among them, P oven (t),P fan (t),P pump (t) represents the instantaneous power of the oven, fan, and pump, respectively; For integration time windows;

[0053] Closed-loop response time:

[0054]

[0055] Where e T For thickness error;

[0056] The constraints unify process, equipment, and safety limitations into a single set of constraints:

[0057] Actuator physical boundaries:

[0058] u min ≤u≤u max , |Δu|≤Δu max ;

[0059] Where u is the actuator; u min ,u max These represent the upper and lower physical limits of the actuator; Δu is the control increment; Δu max The maximum allowable control increment;

[0060] Production line operating boundaries:

[0061] ;

[0062] Among them, v min ,v max These represent the minimum and maximum linear velocities; This represents the upper limit of the temperature in the i-th drying zone; Let be the temperature of the i-th drying zone; LEL(t) is the ratio of the solvent concentration to the lower explosive limit concentration; This is a threshold value representing the ratio of the preset solvent concentration to the lower explosive limit concentration.

[0063] Furthermore, security domains are determined based on historical data and physical models, as follows:

[0064] Mahalanobis distance safety ellipsoid based on historical normal operation data:

[0065] ;

[0066] Where z is the joint observation vector; These are the mean vector and covariance matrix of historical normal data, respectively. Here, p represents the chi-square distribution quantile, p is the dimension, and p is the confidence level. ;

[0067] Open set detection boundary:

[0068] ;

[0069] Based on physical constraints, the rheological, coating, and drying coupled boundaries are obtained, and the approximate relationship between the volumetric flow rate and thickness through the slit die is obtained:

[0070] ;

[0071] Constraints include ;

[0072] Improving coating stability by utilizing capillary number boundaries:

[0073] ;

[0074] Drying mass transfer and LEL boundary:

[0075] ;

[0076] Thermo-rheological coupling threshold of edge contour:

[0077] ;

[0078] The intersection of the union of the data domain and the physical domain is used as the security domain:

[0079] ;

[0080] Where S data For data-driven security domains; S phys Security domains defined for physical constraints.

[0081] Furthermore, based on constraint sets and security domains, a multimodal sensing network is constructed, as follows:

[0082] Based on the indicators and safety domain obtained by S1, the thickness accuracy, minimum defect size, edge homogeneity, and line speed / width requirements are quantified into parameters such as resolution, frame rate, SNR, optical mode, and measurement range, and the thickness gauge and safety sensor are determined as the reference and interlock.

[0083] The system is connected in series with four major stations: pre-coating, die head, drying end, and pre-winding. Backlight, coaxial, dark field, polarization, multispectral, and telecentric lenses are configured for different imaging of wet / dry films.

[0084] On the data stream, a spatiotemporally consistent data chain is constructed with PTP synchronization and encoder hard triggering.

[0085] Furthermore, based on a multimodal sensing network, images and process data from each workstation are collected and preprocessed, as follows:

[0086] Based on the constructed multimodal sensing network, images and process data from each workstation are collected; and the collected images and process data are preprocessed respectively.

[0087] Finally, by matching the timestamp with the meter standard, an integrated index of spatial location, time, and parameters is established. Specifically:

[0088] Using the encoder position s as a spatial reference, images and data from different workstations are mapped to their relative positions. Given the known distances d1, d2, d3, d4 between each workstation and the coating point, and the delay in the same batch of material appearing at different workstations, let d be the distance between the workstations. a / v, constructs a two-dimensional graph line by line using linear arrays:

[0089] Images and data from different workstations are aligned using the 's' method to construct a full-process view of the same cross-section; timestamps of all equipment are made consistent through clock synchronization provided by PTP.

[0090] Modeling the index, key structure: key = (s, t, batch) id , position id ), where s is the spatial location, t is the timestamp, and the value structure includes multimodal data packets, containing image fragments, process parameter vectors, control commands, and quality judgments.

[0091] Furthermore, the multi-task network analysis model adopts a fusion architecture of shared backbone, multi-task heads, and process conditionation, and embeds physical priors and uncertainty estimation, as detailed below:

[0092] The shared backbone adopts ConvNeXt-T, and combined with the geometric characteristics of linear array data, a 1D / 2D hybrid convolutional front end is designed to extract information in both the horizontal direction (MD) and the machine direction (CD).

[0093] The multi-task head includes a thickness regression head, a defect detection segmentation head, and an edge homogeneity head; the thickness regression head predicts the thickness field and uncertainty; the defect detection segmentation head, through instance segmentation, outputs the category, mask, size, and position; the edge homogeneity head, using sub-pixel edge line regression, outputs the edge curve, homogeneity, and adhesive stacking index.

[0094] Physical priors and constraints are embedded by using physical consistency regularization to constrain the Ca interval, Re upper bound, and dryness conservation, and soft regularization is added to the loss.

[0095] Shape prioritization is implemented by prioritizing low-frequency parameters such as edge smoothness and thickness of the horizontal profile, and by reducing the weight of samples not within Csafety or triggering a degradation path during inference.

[0096] Furthermore, the shared backbone adopts ConvNeXt-T, and combined with the geometric characteristics of linear array data, a 1D / 2D hybrid convolutional front end is designed to extract information in both the horizontal direction (MD) and the machine direction (CD), as follows:

[0097] The input is a striped image stitched together row by row, with width x in the CD direction and positions s accumulated in the MD direction; first, a 1D convolution F is applied. 1D Extract features along the path in the MD direction, then use 2D convolution F. 2D Obtaining the local spatial context:

[0098] ;

[0099] ;

[0100] Where x, y are image coordinates; s is the spatial position of the material in the MD (machine feed direction), obtained by encoder integration; I(x, s,:) is the linear array stitched strip image, a two-dimensional image accumulated with s; The kernel is a one-dimensional convolution along the MD direction, with kernel size k. s ; For a two-dimensional convolution kernel, the horizontal kernel k x Along the process, core k s ;F l For the feature map of layer l, the number of channels C l ;

[0101] The main branch uses ConvNeXt-T hierarchical feature extraction to maintain high-resolution branches for edge and defect tasks;

[0102] Project the process vector p into the channel modulation parameters:

[0103] ;

[0104] Where, γ l ,β l These are channel scaling and offset, respectively, generated from the process vector; MLP l This represents the mapping from the process vector to the modulation parameters.

[0105] Features F in the l-th layer l FiLM modulation is performed on the above:

[0106]

[0107] Employing SE attention and process gating fusion:

[0108] ;

[0109] GAP stands for Global Average Pooling; These are the Sigmoid function and the activation function, respectively.

[0110] Furthermore, the multi-task head includes a thickness regression head, a defect detection segmentation head, and an edge homogeneity head, as detailed below:

[0111] The thickness regression head H thick Predicting the thickness field With pixel / contour level uncertainty Output parameterization:

[0112] ;

[0113] ;

[0114] in, σ is the mean of the thickness predictions. T (x,y) represents the standard deviation of the thickness prediction;

[0115] Depend on Export metrics:

[0116] ;

[0117] The defect detection segmentation head H defect Built on anchor-free architecture, outputting a collection of instances. , including mask m k Category C k Size s k Location box b k Confidence level q k ;

[0118] Severity-weighted output:

[0119] ;

[0120] in, This is a weighting factor for defect severity.

[0121] The edge alignment head H edgeRegress the marginal curve E(y) and calculate the homogeneity U. edge With the rubber pile index B edge ;

[0122] ;

[0123] ;

[0124] ;

[0125] in, B is a basis function. j The coefficients of the edge curve on the basis functions;

[0126] Furthermore, the physical priors and constraint embeddings are as follows:

[0127] Capillary number and Reynolds number interval regularization, given feasible region I Ca ,I Re :

[0128] ;

[0129] ;

[0130] in, , This represents the degree to which the distance interval exceeds its bounds.

[0131] μ is the liquid viscosity, γ is the surface tension, ρ is the density, H is the film thickness; Ca is the capillary number, which characterizes the viscous force, and Re is the Reynolds number, which characterizes the viscosity ratio.

[0132] Relationship between the conservation of drying mass and solid content:

[0133] ;

[0134] Among them, T wet ,T dry These refer to the wet film thickness and the dry film thickness, respectively. This is the solvent volume fraction;

[0135] For thickness profiles, low-frequency priority is given; for width profiles, Fourier low-frequency emphasis is applied.

[0136] ;

[0137] in, The average thickness profile in the horizontal direction; This refers to the frequency components after the low-frequency emphasis of the Fourier transform.

[0138] A tape coating control and optimization system based on intelligent image analysis includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the tape coating control and optimization method based on intelligent image analysis as described above.

[0139] The present invention has the following beneficial effects:

[0140] 1. This invention constructs a full-process multimodal sensing network using linear / area array cameras, multispectral imaging, and process sensors. It achieves sub-millisecond cross-station data alignment based on PTP clock synchronization and high-precision encoders, and establishes a mapping relationship for the same material point at different stations using a unified spatial-temporal-parameter index, thus solving the problem of delayed alignment between processes such as coating, drying, and winding.

[0141] 2. This invention adopts a ConvNeXt-T shared backbone + multi-task head fusion architecture. Through the FiLM / SE mechanism, process parameters are conditionalized into the feature space to achieve deep fusion of image and process data. The 1D / 2D hybrid convolution dedicated to linear array data makes full use of the geometric characteristics of the MD / CD direction, while the thickness regression head, defect detection segmentation head and edge homogeneity head share features, enabling the model to learn the intrinsic correlation of the process more efficiently.

[0142] 3. This invention achieves intelligent quality control of thin film coating through multimodal perception and spatiotemporal integration, multi-task network and physical prior, uncertainty-driven and safety domain constraint, and rolling optimization. It can effectively solve the problems of traditional methods in multivariable coupled control, delay compensation, and generalization to new working conditions. Attached Figure Description

[0143] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0144] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0145] refer to Figure 1 In this embodiment, a method for controlling and optimizing tape coating based on intelligent image analysis is provided, including the following steps:

[0146] S1: Obtain product specifications (target thickness, tolerance), current status data (yield, defect spectrum), production line capabilities (width, line speed), establish a quantitative indicator system and constraints, and determine the safety domain based on historical data and physical models, and obtain the constraint set and safety domain;

[0147] S2: Construct a multimodal perception network based on constraint sets and security domains;

[0148] S3: Based on a multimodal sensing network, images and process data from each workstation are collected and preprocessed;

[0149] S4: Construct a multi-task network analysis model, and analyze the preprocessed images and process data of each workstation to obtain the analysis results;

[0150] S5: Based on the analysis results, a reinforcement learning control strategy is used to generate and obtain the optimal control sequence.

[0151] In this embodiment, preferably, product specifications, current status data, and production line capabilities are obtained, and a quantitative indicator system and constraints are established, as follows:

[0152] The product specifications include the target thickness tsp and tolerance ±tol; current data includes yield Y, defect spectrum distribution P(d), and process capability index Cpk; production line capability includes width W and line speed range v∈[v min ,v max ], Actuator adjustable range and resolution;

[0153] Establish a quantitative indicator system, including thickness fluctuation and contour indicators, defect density and severity, edge uniformity and glue buildup index, energy consumption and response time;

[0154] The thickness fluctuation and contour indicators are as follows:

[0155] Let the in-plane pixel thickness map be T, the horizontal direction be defined as x∈[0,W], and the machine direction be y;

[0156] Average thickness: ;

[0157] in, Let it be the expected function;

[0158] Thickness standard deviation (3σ control): ;

[0159] Tolerance satisfies constraints:

[0160] ;

[0161] Among them, t sp t represents the target thickness; tol represents the thickness tolerance; t min ,t max These represent the minimum and maximum thickness, respectively.

[0162] Banner outline: ;

[0163] Where y0 is the initial machine direction; L is the banner length;

[0164] Contour basis function decomposition (PCA / Fourier):

[0165] ;

[0166] in, c are the basis functions for contour decomposition; i The coefficients of the contour basis functions characterize the intensity of the contour shape components; The average thickness profile in the horizontal direction;

[0167] Process capability index C pk :

[0168] ;

[0169] USL and LSL are the upper and lower specification limits, respectively.

[0170] Preferably, set target setting C pk ≥1.33 (mass production), or C pk ≥1.67 (High reliability).

[0171] The specific defect density and severity are as follows:

[0172] Let N be the number of defect events within area A. d The chance defect number DPMO is:

[0173] ;

[0174] Where a0 is the unit opportunity area (set according to the minimum inspectable size);

[0175] Severity Index (area, contrast, location weighted):

[0176] ;

[0177] Where, α tk βs is the weight of defect type k. k For size grading, γ pk For functional area location weights; a k Let be the area of ​​the k-th defect;

[0178] Objectives and constraints:

[0179] DPMO≤DPMO max ,D≤D max ;

[0180] Among them, DPMO max D max These are the preset thresholds for the number of opportunity defects and the severity index, respectively;

[0181] The edge alignment and adhesive buildup index are as follows:

[0182] Let the edge position function be E(y), then the homogeneity deviation (based on sub-pixel fitting):

[0183] ;

[0184] in, This represents the average value at the edge positions.

[0185] Edge adhesive buildup index (relative excess thickness in the edge area):

[0186] ;

[0187] Where Δx is the edge bandwidth used to evaluate edge effects;

[0188] The energy consumption and response time are as follows:

[0189] Energy consumption per unit area E area (Drying as the main method):

[0190] ;

[0191] Among them, P oven (t),P fan (t),P pump (t) represents the instantaneous power of the oven, fan, and pump, respectively; For integration time windows;

[0192] Closed-loop response time (arrival and maintenance within the tolerance band):

[0193]

[0194] Where e T For thickness error;

[0195] The constraints unify process, equipment, and safety limitations into a single set of constraints:

[0196] Actuator physical boundaries (hard constraints):

[0197] u min ≤u≤u max , |Δu|≤Δu max ;

[0198] Where u represents the actuator, including the die / blade zone opening, pump speed, linear speed, tension, and drying zone temperature / air velocity; u min ,u max These represent the upper and lower physical limits of the actuator (corresponding dimensions for each channel); Δu is the control increment; Δu max The maximum allowable control increment;

[0199] Production line operating boundaries:

[0200] ;

[0201] Among them, v min ,v max These represent the minimum and maximum linear velocities; This represents the upper limit of the temperature in the i-th drying zone; Let be the temperature of the i-th drying zone; LEL(t) is the ratio of the solvent concentration to the lower explosive limit concentration; This is a threshold value representing the ratio of the preset solvent concentration to the lower explosive limit concentration.

[0202] In this embodiment, preferably, the security domain is determined based on historical data and a physical model, as follows:

[0203] Mahalanobis distance safety ellipsoid based on historical normal operation data:

[0204] ;

[0205] Where z is the joint observation vector; These are the mean vector and covariance matrix of historical normal data, respectively. Here, p represents the chi-square distribution quantile, p is the dimension, and p is the confidence level. ;

[0206] Open set detection boundary (density threshold):

[0207] ;

[0208] Based on physical constraints, obtain the rheological, coating, and drying coupled boundaries, and approximate the relationship between volumetric flow rate and thickness through a slit die (or comma cutter) (steady-state, Newtonian approximation):

[0209] ;

[0210] Constraints include ;

[0211] Improving coating stability by utilizing capillary number boundaries:

[0212] ;

[0213] Among them, Ca is prone to pulsation / bubble entrainment / air entrainment when it exceeds the boundary;

[0214] Drying mass transfer and LEL boundary:

[0215] ;

[0216] It also links to the upper limits of exhaust / fan speed / temperature;

[0217] Thermo-rheological coupling threshold of edge contour:

[0218] ;

[0219] The intersection of the union of the data domain and the physical domain is used as the security domain:

[0220] ;

[0221] Where S data For data-driven security domains (Mahavioran distance ellipsoids or density threshold sets); S phys Security domains defined for physical constraints.

[0222] In this embodiment, preferably, a multimodal sensing network is constructed based on constraint sets and security domains, as follows:

[0223] Based on the indicators and safety domain obtained by S1, the thickness accuracy, minimum defect size, edge homogeneity, and line speed / width requirements are quantified into parameters such as resolution, frame rate, SNR, optical mode, and measurement range, and the thickness gauge and safety sensor are determined as the reference and interlock.

[0224] The system is connected in series with four major stations: pre-coating, die head, drying end, and pre-winding. Backlight, coaxial, dark field, polarization, multispectral, and telecentric lenses are configured for different imaging of wet / dry films.

[0225] Preferably, in this embodiment, the arrangement is as follows:

[0226] Before coating (incoming material control): Linear array + telecentric backlight: substrate pores, cracks, width, and deviation; tension / deviation sensing.

[0227] Dark field supplement: micro-particles / scratches; electrostatic / dust control interface (ionization / vacuuming).

[0228] Die head exit (key to wet film): Area array + coaxial / polarized: wet film uniformity, bubbles / fisheye, early stripes; high frame rate short exposure to suppress blur.

[0229] Structured light can be added: wet film slope and edge line shape; mold head / blade temperature probe monitors thermal drift.

[0230] Drying end (dry film formation): linear array + dark field / multispectral: pinholes, gel, orange peel, streaks.

[0231] Several sampling points of the thickness reference (white light interferometry / laser) are used for S4 regression calibration.

[0232] Temperature / wind speed array banners are used to monitor the uniformity of the drying zone and constrain the "dog bone" outline.

[0233] Before winding (final quality inspection and measurement): Linear array + telecentric lens + backlight: edge alignment, width and adhesive buildup index; full-frame thickness regression map registration.

[0234] Area array sampling inspection: high-resolution screenshots of key areas are archived; encoders / meter wheels provide spatial reference;

[0235] On the data stream, a spatiotemporally consistent data chain is constructed with PTP synchronization and encoder hard triggering.

[0236] In this embodiment, preferably, images and process data from each workstation are collected based on a multimodal sensing network and preprocessed, as follows:

[0237] Based on the constructed multimodal sensing network, images and process data from each workstation are collected;

[0238] In this embodiment, the acquisition of images and process data at each workstation specifically includes:

[0239] Image stream acquisition:

[0240] Line scan camera (full-frame detection): Each station continuously samples at a line rate vv to form a two-dimensional image stream with latitude and longitude registration;

[0241] Area scan camera (close-up view): Triggered by encoder pulses / fixed distance, it captures high-precision snapshots of the wet film and key areas;

[0242] Original image formats: 16-bit TIFF / HDR PNG / proprietary formats, preserving full dynamic range; tiled for parallel processing;

[0243] Process data acquisition:

[0244] Process parameters: pump speed (Hz / rpm), die / blade gap (μm), linear speed (m / min), tension (N / m), drying zone temperature array (°C), air velocity (m / s), exhaust valve position (%);

[0245] Actuator feedback: The actual executed values ​​of all control variables are compared with the command values ​​to verify the S5 control effect;

[0246] Sampling frequency: 10–100 Hz for fast variables, 1–10 Hz for slow variables, to ensure capture of transient dynamics;

[0247] Batch and quality data:

[0248] Batch / Formula Metadata: Product Code, Width, Thickness Specifications, Formulation Parameters, Shift / Operator;

[0249] Offline quality assessment: Laboratory measurements of thickness points, tensile strength, peel strength, initial adhesion, etc., are used for supervised learning and annotation.

[0250] Contextual information: ambient temperature and humidity, season / day / night, production change time, and abnormal event markers.

[0251] The acquired images and process data are preprocessed separately;

[0252] In this embodiment, the preprocessing is as follows:

[0253] The image is processed through distortion correction, flat and dark field compensation, light source aging LUT correction, and median / bandpass / spatiotemporal noise reduction, histogram equalization (CLAHE), highlight / glare suppression, and finally edge / functional band cropping to build an image pyramid to adapt to multi-task analysis.

[0254] Process data preprocessing includes noise reduction (Butterworth / Kalman) and unit unification (SI).

[0255] Outlier correction (Hampel / IQR), missing interpolation (spline / Kalman), aligning process / control samples to a unified PTP time axis, and recording delays and jitter.

[0256] Finally, by matching the timestamp with the meter standard, an integrated index of spatial location, time, and parameters is established. Specifically:

[0257] Using the encoder position s as a spatial reference, images and data from different workstations are mapped to their relative positions. Given the known distances d1, d2, d3, d4 between each workstation and the coating point, and the delay in the same batch of material appearing at different workstations, let d be the distance between the workstations. a / v, constructs a two-dimensional graph line by line using linear arrays:

[0258] Images and data from different workstations are aligned using the 's' method to construct a full-process view of the same cross-section; timestamps of all equipment are made consistent through clock synchronization provided by PTP.

[0259] Modeling the index, key structure: key = (s, t, batch) id , position id ), where s is the spatial location, t is the timestamp, and the value structure includes multimodal data packets, containing image fragments, process parameter vectors, control commands, and quality judgments.

[0260] In this embodiment, preferably, the multi-task network analysis model adopts a fusion architecture of shared backbone, multi-task heads, and process conditionalization, and embeds physical priors and uncertainty estimation, as follows:

[0261] The shared backbone adopts ConvNeXt-T, and combined with the geometric characteristics of linear array data, a 1D / 2D hybrid convolutional front end is designed to extract information in both the horizontal direction (MD) and the machine direction (CD).

[0262] The multi-task head includes a thickness regression head, a defect detection segmentation head, and an edge homogeneity head; the thickness regression head predicts the thickness field and uncertainty; the defect detection segmentation head, through instance segmentation, outputs the category, mask, size, and position; the edge homogeneity head, using sub-pixel edge line regression, outputs the edge curve, homogeneity, and adhesive stacking index.

[0263] Physical priors and constraints are embedded by using physical consistency regularization to constrain the Ca interval, Re upper bound, and dryness conservation, and soft regularization is added to the loss.

[0264] Shape prioritization is achieved through low-frequency preference of edge line smoothness and thickness of the horizontal profile (Fourier / wavelet sparse constraints), and samples not within Csafety are deweighted or trigger a degradation path during inference.

[0265] In this embodiment, preferably, the shared backbone adopts ConvNeXt-T. Combining the geometric characteristics of the linear array data, a 1D / 2D hybrid convolutional front-end is designed to extract information in both the horizontal direction (MD) and the machine direction (CD), as follows:

[0266] The input is a striped image stitched together row by row, with width x in the CD direction and positions s accumulated in the MD direction; first, a 1D convolution F is applied. 1D Extract features along the path in the MD direction, then use 2D convolution F. 2D Obtaining the local spatial context:

[0267] ;

[0268] ;

[0269] Where x, y are image coordinates; s is the spatial position of the material in the MD (machine feed direction), obtained by encoder integration; I(x, s,:) is the linear array stitched strip image, a two-dimensional image accumulated with s; The kernel is a one-dimensional convolution along the MD direction, with kernel size k. s ; For a two-dimensional convolution kernel, the horizontal kernel k x Along the process, core k s ;F l For the feature map of layer l, the number of channels C l ;

[0270] The main branch uses ConvNeXt-T hierarchical feature extraction to maintain high-resolution branches for edge and defect tasks;

[0271] Project the process vector p into the channel modulation parameters:

[0272] ;

[0273] Where, γ l ,β l These are channel scaling and offset, respectively, generated from the process vector; MLP l This represents the mapping from the process vector to the modulation parameters.

[0274] Features F in the l-th layer l FiLM modulation is performed on the above:

[0275]

[0276] Employing SE attention and process gating fusion:

[0277] ;

[0278] GAP stands for Global Average Pooling; These are the Sigmoid function and the activation function, respectively.

[0279] In this embodiment, preferably, the multi-task head includes a thickness regression head, a defect detection segmentation head, and an edge homogeneity head, as detailed below:

[0280] The thickness regression head H thick Predicting the thickness field With pixel / contour level uncertainty Output parameterization:

[0281] ;

[0282] ;

[0283] in, σ is the mean of the thickness predictions. T (x,y) represents the standard deviation of the thickness prediction;

[0284] Depend on Export metrics:

[0285] ;

[0286] The defect detection segmentation head H defect Built on anchor-free architecture, outputting a collection of instances. , including mask mk Category C k Size s k Location box b k Confidence level q k ;

[0287] Severity-weighted output (consistent with the S1 metric):

[0288] ;

[0289] in, Weighting factor for defect severity (by type, size, and location importance);

[0290] The edge alignment head H edge Regress the marginal curve E(y) and calculate the homogeneity U. edge With the rubber pile index B edge ;

[0291] ;

[0292] ;

[0293] ;

[0294] in, B is a basis function. j The coefficients of the edge curve on the basis functions;

[0295] In this embodiment, the preferred method for embedding physical priors and constraints is as follows:

[0296] Capillary number and Reynolds number interval regularization, given feasible region I Ca ,I Re :

[0297] ;

[0298] ;

[0299] in, , This represents the degree to which the distance interval exceeds its bounds.

[0300] μ is the liquid viscosity, γ is the surface tension, ρ is the density, H is the film thickness; Ca is the capillary number, which characterizes the viscous force, and Re is the Reynolds number, which characterizes the viscosity ratio.

[0301] Relationship between the conservation of drying mass and solid content:

[0302] ;

[0303] Among them, Twet ,T dry These refer to the wet film thickness and the dry film thickness, respectively. This is the solvent volume fraction;

[0304] For thickness profiles, low-frequency priority is given; for width profiles, Fourier low-frequency emphasis is applied.

[0305] ;

[0306] in, The average thickness profile in the horizontal direction; This refers to the frequency components after the low-frequency emphasis of the Fourier transform.

[0307] In this embodiment, preferably, based on the analysis results, a reinforcement learning control strategy is used to generate and obtain the optimal control sequence, as follows:

[0308] Define control objectives and constraints: unify the cost function (thickness, defects, capacity, energy consumption) and constraint set. Establish a state-action-dynamic model to clarify the relationship between controllable variables, observed states, and system dynamics.

[0309] Control objective and cost function:

[0310] Quality objectives: Minimize thickness deviation and variation, and achieve edge homogeneity U. edge With glue B edge Reduce and minimize defect severity.

[0311] Production capacity / energy consumption: Maximize linear velocity v and minimize unit energy consumption.

[0312] Execution smoothing: motion change rate and jitter constraint.

[0313] Unified cost (minimization):

[0314] Constraint set:

[0315] ;

[0316] State-Action-Dynamic Model

[0317] state s t : Concatenate the structured output and key process quantities of S4, including uncertainty and safety region scores: ;

[0318] Action u t Multi-actuator cooperative control vector ;

[0319] Environmental Dynamics: Black Box / Semi-Physical Model It also provides a safety constraint to determine whether g(st,ut)≤0.

[0320] Training reinforcement learning strategy: Offline RL (BCQ / CQL) is adopted, combined with simulation and historical data.

[0321] Implement a rolling optimization mechanism: Generate candidate actions based on a policy network, and select the optimal sequence by combining model predictions.

[0322] A tape coating control and optimization system based on intelligent image analysis includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the tape coating control and optimization method based on intelligent image analysis as described above.

[0323] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0324] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0325] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0326] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0327] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for controlling and optimizing tape coating based on intelligent image analysis, characterized in that, Includes the following steps: S1: Obtain product specifications, current status data, and production line capabilities; establish a quantitative indicator system and constraints; and determine the safety domain based on historical data and physical models, thereby obtaining the constraint set and safety domain. S2: Construct a multimodal perception network based on constraint sets and security domains; S3: Based on a multimodal sensing network, images and process data from each workstation are collected and preprocessed; S4: Construct a multi-task network analysis model, and analyze the preprocessed images and process data of each workstation to obtain the analysis results; S5: Based on the analysis results, a reinforcement learning control strategy is used to generate and obtain the optimal control sequence.

2. The tape coating control and optimization method based on intelligent image analysis according to claim 1, characterized in that, The acquisition of product specifications, current status data, and production line capabilities, and the establishment of a quantitative indicator system and constraints, are detailed below: The product specifications include the target thickness tsp and tolerance ±tol; current data includes yield Y, defect spectrum distribution P(d), and process capability index Cpk; production line capability includes width W and line speed range v∈[v min ,v max [Actuator adjustable range and resolution;] Establish a quantitative indicator system, including thickness fluctuation and contour indicators, defect density and severity, edge uniformity and glue buildup index, energy consumption and response time; The thickness fluctuation and contour indicators are as follows: Let the in-plane pixel thickness map be T, the horizontal direction be defined as x∈[0,W], and the machine direction be y; Average thickness: ; in, Let it be the expected function; Thickness standard deviation: ; Tolerance satisfies constraints: ； Among them, t sp t represents the target thickness; tol represents the thickness tolerance; t min ,t max These represent the minimum and maximum thickness, respectively. Banner outline: ; Where y0 is the initial machine direction; L is the banner length; Contour basis function decomposition: ； in, c are the basis functions for contour decomposition; i The coefficients of the contour basis functions characterize the intensity of the contour shape components; The average thickness profile in the horizontal direction; Process capability index C pk : ； USL and LSL are the upper and lower specification limits, respectively. The specific defect density and severity are as follows: Let N be the number of defect events within area A. d The chance defect number DPMO is: ； Where a0 is the unit opportunity area; Severity Index: ； Where, α tk βs is the weight of defect type k. k For size grading, γ pk For functional area location weights; a k Let be the area of ​​the k-th defect; Objectives and constraints: DPMO≤DPMO max ,D≤D max ; Among them, DPMO max D max These are the preset thresholds for the number of opportunity defects and the severity index, respectively; The edge alignment and adhesive buildup index are as follows: Let the edge position function be E(y), then the homogeneity deviation is: ； in, This represents the average value at the edge positions. Edge adhesive buildup index: ； Where Δx is the edge bandwidth used to evaluate edge effects; The energy consumption and response time are as follows: Energy consumption per unit area E area : ； Among them, P oven (t),P fan (t),P pump (t) represents the instantaneous power of the oven, fan, and pump, respectively; For integration time windows; Closed-loop response time: ； Where e T For thickness error; The constraints unify process, equipment, and safety limitations into a single set of constraints: Actuator physical boundaries: in min Oh, oh. max ,∣Δu∣≤Δu max ; Where u is the actuator; u min ,u max These represent the upper and lower physical limits of the actuator; Δu is the control increment; Δu max The maximum allowable control increment; Production line operating boundaries: ； Among them, v min ,v max These represent the minimum and maximum linear velocities; This represents the upper limit of the temperature in the i-th drying zone; Let be the temperature of the i-th drying zone; LEL(t) is the ratio of the solvent concentration to the lower explosive limit concentration; This is a threshold value representing the ratio of the preset solvent concentration to the lower explosive limit concentration.

3. The tape coating control and optimization method based on intelligent image analysis according to claim 2, characterized in that, The determination of the security domain based on historical data and physical models is as follows: Mahalanobis distance safety ellipsoid based on historical normal operation data: ； Where z is the joint observation vector; These are the mean vector and covariance matrix of historical normal data, respectively. Here, p represents the chi-square distribution quantile, p is the dimension, and p is the confidence level. ; Open set detection boundary: ； Based on physical constraints, the rheological, coating, and drying coupled boundaries are obtained, and the approximate relationship between the volumetric flow rate and thickness through the slit die is obtained: ； Constraints include ; Improving coating stability by utilizing capillary number boundaries: ； Drying mass transfer and LEL boundary: ； Thermo-rheological coupling threshold of edge contour: ； The intersection of the union of the data domain and the physical domain is used as the security domain: ； Where S data For data-driven security domains; S phys Security domains defined for physical constraints.

4. The tape coating control and optimization method based on intelligent image analysis according to claim 1, characterized in that, The multimodal perception network is constructed based on constraint sets and security domains, as detailed below: Based on the indicators and safety domain obtained by S1, the thickness accuracy, minimum defect size, edge homogeneity, and line speed / width requirements are quantified into parameters such as resolution, frame rate, SNR, optical mode, and measurement range, and the thickness gauge and safety sensor are determined as the reference and interlock. The system is connected in series with four major stations: pre-coating, die head, drying end, and pre-winding. Backlight, coaxial, dark field, polarization, multispectral, and telecentric lenses are configured for different imaging of wet / dry films. On the data stream, a spatiotemporally consistent data chain is constructed with PTP synchronization and encoder hard triggering.

5. The tape coating control and optimization method based on intelligent image analysis according to claim 1, characterized in that, The multimodal sensing network is used to collect images and process data from each workstation, and these are preprocessed as follows: Based on the constructed multimodal sensing network, images and process data from each workstation are collected; and the collected images and process data are preprocessed respectively. Finally, by matching the timestamp with the meter standard, an integrated index of spatial location, time, and parameters is established. Specifically: Using the encoder position s as a spatial reference, images and data from different workstations are mapped to their relative positions. Given the known distances d1, d2, d3, d4 between each workstation and the coating point, and the delay in the same batch of material appearing at different workstations, let d be the distance between the workstations. a / v, constructs a two-dimensional graph line by line using linear arrays: Images and data from different workstations are aligned using the 's' method to construct a full-process view of the same cross-section; timestamps of all equipment are made consistent through clock synchronization provided by PTP. Modeling the index, key structure: key = (s, t, batch) id , position id ), where s is the spatial location, t is the timestamp, and the value structure includes multimodal data packets, containing image fragments, process parameter vectors, control commands, and quality judgments.

6. The tape coating control and optimization method based on intelligent image analysis according to claim 1, characterized in that, The multi-task network analysis model adopts a fusion architecture of shared backbone, multi-task heads, and process conditionalization, and embeds physical priors and uncertainty estimation, as detailed below: The shared backbone adopts ConvNeXt-T, and combined with the geometric characteristics of linear array data, a 1D / 2D hybrid convolutional front end is designed to extract information in both the horizontal direction (MD) and the machine direction (CD). The multi-task head includes a thickness regression head, a defect detection segmentation head, and an edge homogeneity head; the thickness regression head predicts the thickness field and uncertainty; the defect detection segmentation head, through instance segmentation, outputs the category, mask, size, and position; the edge homogeneity head, using sub-pixel edge line regression, outputs the edge curve, homogeneity, and adhesive stacking index. Physical priors and constraints are embedded by using physical consistency regularization to constrain the Ca interval, Re upper bound, and dryness conservation, and soft regularization is added to the loss. Shape prioritization is implemented by prioritizing low-frequency parameters such as edge smoothness and thickness of the horizontal profile, and by reducing the weight of samples not within Csafety or triggering a degradation path during inference.

7. The tape coating control and optimization method based on intelligent image analysis according to claim 6, characterized in that, The shared backbone adopts ConvNeXt-T, and combined with the geometric characteristics of linear array data, a 1D / 2D hybrid convolutional front end is designed to extract information in both the horizontal direction (MD) and the machine direction (CD), as follows: The input is a striped image stitched together row by row, with width x in the CD direction and position s accumulated in the MD direction; first, a 1D convolution F is applied. 1D Extract features along the path in the MD direction, then use 2D convolution F. 2D Obtaining the local spatial context: ； ； Where x, y are image coordinates; s is the spatial position of the material on the MD, obtained by encoder integration; I(x, s,:) is the linear array stitched strip image, a two-dimensional image accumulated with s; The kernel is a one-dimensional convolution along the MD direction, with kernel size k. s ; For a two-dimensional convolution kernel, the horizontal kernel k x Along the process, core k s ;F l For the feature map of layer l, the number of channels C l ; The main branch uses ConvNeXt-T hierarchical feature extraction to maintain high-resolution branches for edge and defect tasks; Project the process vector p into the channel modulation parameters: ； Where, γ l ,β l These are channel scaling and offset, respectively, generated from the process vector; MLP l This represents the mapping from the process vector to the modulation parameters. Features F in the l-th layer l FiLM modulation is performed on the above: ； Employing SE attention and process gating fusion: ； GAP stands for Global Average Pooling; These are the Sigmoid function and the activation function, respectively.

8. The tape coating control and optimization method based on intelligent image analysis according to claim 6, characterized in that, The multi-task head includes a thickness regression head, a defect detection segmentation head, and an edge homogeneity head, as detailed below: The thickness regression head H thick Predicting the thickness field With pixel / contour level uncertainty Output parameterization: ； ； in, σ is the mean of the thickness predictions. T (x,y) represents the standard deviation of the thickness prediction; Depend on Export metrics: ； The defect detection segmentation head H defect Built on anchor-free architecture, outputting a collection of instances. , including mask m k Category C k Size s k Location box b k Confidence level q k ; Severity-weighted output: ； in, This is a weighting factor for defect severity. The edge alignment head H edge Regress the marginal curve E(y) and calculate the homogeneity U. edge With the glue accumulation index B edge ; ； ； ； in, B is a basis function. j The coefficients of the edge curve on the basis functions; 9. The tape coating control and optimization method based on intelligent image analysis according to claim 6, characterized in that, The physical priors and constraint embedding are specifically as follows: Capillary number and Reynolds number interval regularization, given feasible region I Ca ,I Re : ; ; in, , This represents the degree to which the distance interval exceeds its bounds. μ is the liquid viscosity, γ is the surface tension, ρ is the density, H is the film thickness; Ca is the capillary number, which characterizes the viscous force, and Re is the Reynolds number, which characterizes the viscosity ratio. Relationship between the conservation of drying mass and solid content: ； Among them, T wet ,T dry These refer to the wet film thickness and the dry film thickness, respectively. This is the solvent volume fraction; For thickness profiles, low-frequency priority is given; for width profiles, Fourier low-frequency emphasis is applied. ； in, The average thickness profile in the horizontal direction; This refers to the frequency components after the low-frequency emphasis of the Fourier transform.

10. A tape coating control and optimization system based on intelligent image analysis, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the tape coating control and optimization method based on intelligent image analysis as described in any one of claims 1-9.

Images