A method and system for predicting and regulating the bonding strength for solventless lamination

Abstract

The application discloses a kind of solventless composite-oriented bonding strength prediction regulation and control method, real-time acquisition composite process parameter is exported the initial bonding strength potential of each volume length position by preliminary soft measurement submodel;It is reconstructed as two-dimensional distribution in roll by Archimedes helical winding relationship with winding radius and axial coordinate calibration;With the coupling solution of non-steady heat conduction equation and curing kinetics equation under cylindrical coordinate system containing internal heat source, the ripening temperature, humidity history and curing conversion rate of each position in roll are reconstructed;The initial bonding strength potential and ripening history characteristics are input into fusion soft measurement model, and the output is the peel strength spatial distribution of whole roll;The predicted value is compared with the qualified threshold to identify the local risk area, and the controllable amount of composite station area and curing library area is adjusted reversely.The application reconstructs the spatial non-uniformity of roll ripening field, overcomes the spatial mismatch between predicted value and actual ripening in roll, identifies and corrects abnormalities in defect cause stage.

Description

Technical Field

[0001] This application relates to the field of quality control technology for flexible packaging composite processes, specifically to a method and system for predicting and controlling the adhesive strength of solvent-free composites. Background Technology

[0002] In solvent-free laminating flexible packaging production, two-component polyurethane adhesives are applied by gluing rollers and pressed together to form a composite structure from multiple substrates. The interlayer peel strength is a key quality indicator for measuring the bonding strength. Unlike solvent-based lamination or instant curing processes, the interlayer peel strength of adhesives used in solvent-free lamination is not formed instantly at the lamination station. Instead, it gradually matures over several hours to tens of hours after the roll material is wound into a master roll and transported to a curing warehouse, relying on the cross-linking reaction between isocyanate and polyol to achieve final stability. Because peel strength has traditionally only been obtained through offline destructive sampling after curing, it cannot be detected in real time during production. Therefore, if an abnormality in the lamination quality occurs, it is often only discovered after curing is complete and during random inspection, resulting in entire rolls or even entire batches of defective products and delayed batch scrapping.

[0003] To understand the quality of lamination at the defect initiation stage, the industry has developed a technical approach that uses real-time data collected at the lamination station to establish soft measurement models and predict the peel strength after curing at the moment of lamination. This approach typically uses variables that can be collected in real-time at the lamination station, such as adhesive temperature and its corresponding viscosity, two-component metering ratio, gluing roller speed ratio, roller gap pressure, lamination speed, substrate surface tension, and ambient temperature and humidity, as inputs. These variables are then directly mapped to the peel strength after curing using regression models or neural networks. However, this approach locks all input variables at the moment of lamination, implicitly treating the curing stage after the roll material comes off the line as a uniform, standardized process for the entire roll or even the entire batch.

[0004] However, the actual cross-linking and curing of adhesives does not occur at the lamination station itself, but rather during the time and space after lamination is completed, the roll material is wound into a large-diameter master roll, and stacked in the curing chamber. Industrial-grade master rolls are wide, often exceeding one meter in diameter, and can reach several kilometers in length. After winding, the layers of adhesive film are pressed and stacked together. The actual thermal history and chemical reaction microenvironment they experience are not equivalent to the state represented by the instantaneously measured ambient temperature and humidity at the lamination station. The cross-linking reaction of adhesives is exothermic. The tighter the master roll is wound and the closer the layers are to the core, the longer the heat dissipation path, and the easier it is for the exothermic reaction to accumulate within the roll. This results in the actual curing temperature closer to the core being higher than the actual curing temperature near the outer layers of the roll material and the exposed areas at both ends of the roll material. At the same time, the hot air circulation method and the stacking method of the roll material in the curing chamber also create a continuous spatial gradient and fluctuation in the actual curing temperature and humidity fields between different master rolls and between different axial and radial positions within the same master roll. The spatial heterogeneity of curing conditions caused by the coupling of the winding stacked structure and the exothermic effect of adhesive crosslinking is a variable system that is independent of the real-time process parameters of the composite station and cannot be sensed and collected at the composite station. However, existing prediction schemes based on the instantaneous parameters of the composite station lack characterization of this.

[0005] As a result, existing soft measurement solutions fail to characterize the actual impact of the spatial non-uniformity of the curing temperature and humidity fields distributed along the length and winding radius of the roll during storage in the curing warehouse after winding. This leads to a systematic spatial mismatch between predicted values ​​and actual peel strength values ​​sampled from different locations within the same roll and batch. It is impossible to identify, before curing is complete, specific areas within the roll that may fail to meet the acceptance threshold due to deviations from standard curing conditions. Furthermore, even if potential non-compliance trends are identified, existing solutions mostly limit intervention to parameter adjustments at the composite workstation itself. Compensatory interventions for rolls already produced during the curing stage lack corresponding control measures based on their location within the roll, making it difficult to implement differentiated control for varying curing states at different locations within the roll.

[0006] The aforementioned technical issues have rather specific application scenarios, manifesting only in flexible packaging production scenarios employing solvent-free lamination processes, requiring long curing times on the order of tens of hours, and completing curing in the form of large-diameter, heavy rolls stacked together. This differs from common scenarios involving instant curing, thin-layer unfolding curing, or curing of small-sized independent samples. Only when the roll material itself is large enough and the coupling effect of adhesive crosslinking exothermics and limited heat transfer within the roll is significant enough does the spatial heterogeneity of curing conditions have engineering significance that requires separate consideration. Currently, machine learning methods for adhesive joints, soft measurement methods for polymer molding, and curing kinetic monitoring methods for structural adhesives all base predictions on a direct mapping between formulation parameters or instantaneous processing parameters and final mechanical properties, and the effective range of curing kinetic models typically only covers the early stages of curing. While technologies related to waterproof and insulating rolls involving rolled structures focus on the structure of the roll itself and the temperature gradient within the roll, they are mainly used for film annealing and heat shrinkage control, failing to establish a technical correlation between the differentiated curing microenvironment within the roll and the online prediction of peel strength. Therefore, how to characterize and predict the spatial distribution of peel strength at different locations within the roll after winding and curing at the composite work station stage, and implement targeted control accordingly, remains an unsolved problem in this technical field. Summary of the Invention

[0007] To address the technical problem that existing soft-measurement schemes for peel strength based solely on instantaneous process parameters at the composite workstation treat the curing stage after the roll comes off the line as a uniform standard process for the entire roll or even the entire batch, failing to characterize the actual impact of the spatial non-uniformity of the curing temperature and humidity fields distributed along the length and winding radius of the roll during storage in the curing warehouse after the roll is wound, resulting in a systematic spatial mismatch between the predicted value and the actual value sampled from different locations within the roll, this invention provides a method and system for predicting and controlling the adhesive strength of solvent-free composites. The idea is to no longer directly map the single strength value of the whole roll to the instantaneous process parameters of the composite station. Instead, it introduces the thermo-chemical coupling spatiotemporal field reconstruction of winding curing. It reconstructs the curing temperature and humidity history and curing transformation process of each position in the winding radius and axial direction of the roll. Then, it makes position-resolved prediction of the peel strength after curing at each position. For local areas that deviate from the qualified threshold, it reverse calculates the controllable quantities of the composite station domain and curing warehouse domain, forming a spatial resolution closed loop of prediction and control. It transforms the adhesive strength that cannot be measured in real time after curing into a controllable quantity that can be predicted online, intervened, and has the ability to resolve spatially within the roll.

[0008] In one aspect, the present invention provides a method for predicting and controlling the adhesive strength of solvent-free composites, the method comprising the following steps.

[0009] Step S1 involves real-time acquisition of process parameters during the lamination process. The initial adhesive strength potential at each roll length position is output by a pre-trained primary soft sensing sub-model, corresponding to the acquisition time and the roll length position. Specifically, the process parameters include adhesive temperature and its corresponding viscosity, two-component metering ratio, gluing roller speed ratio, roller gap pressure, lamination speed, substrate surface tension, and ambient temperature and humidity. By corresponding each process parameter to the acquisition time and roll length position, a continuous sequence of lamination station parameters along the roll length is obtained. The initial adhesive strength potential is the baseline peel strength determined solely by the coating and pressing quality during the lamination process under the standard uniform curing assumption. It characterizes the strength foundation established during the coating and pressing stages during lamination, without considering the impact of differentiated curing after the roll comes off the production line. Preferably, the primary soft measurement sub-model adopts a gradient boosting tree or a fully connected neural network, and uses the actual amount of adhesive calculated by the mapping relationship between adhesive temperature and viscosity, the two-component distribution ratio deviation, and the ratio of the glue roller speed ratio to the lamination speed as the derived feature input, so as to output the initial bonding strength potential distributed along the length of the roll material at the lamination station stage.

[0010] Step S2: The length position of the roll material is converted into the winding radius and axial position of that section of the roll within the master roll according to the Archimedean spiral winding relationship. The initial adhesive strength potential is reconstructed into a two-dimensional distribution within the roll, calibrated by the winding radius coordinates and axial coordinates. Further, a mapping between the roll material length position and the winding geometric position is established according to the Archimedean spiral winding relationship based on the core diameter, single-layer composite structure thickness, winding tension, and its take-up tension taper. This reconstructs the initial adhesive strength potential distributed along the length direction in Step S1 into a two-dimensional distribution within the roll, calibrated by the winding radius coordinates and axial coordinates. Simultaneously, the width of the master roll, core diameter, target roll diameter, winding tension distribution, and the air supply temperature, air supply humidity, hot air circulation mode, and stacking method of the curing chamber are collected as boundary conditions and structural parameters for subsequent curing time-space reconstruction.

[0011] Step S3: Construct a spatiotemporal field reconstruction model of the curing temperature and humidity within the roll. The heat transfer within the roll is described by an unsteady thermal conductivity equation in cylindrical coordinates, containing an internal heat source. The internal heat source term is determined by the exothermic rate of the crosslinking reaction given by the curing kinetics equation. The thermal conductivity equation and the curing kinetics equation are coupled and solved to obtain the curing temperature history and curing conversion rate at each winding radius and axial position within the roll. The humidity history at each position is determined by combining the curing chamber airflow humidity boundary. Specifically, the curing kinetics equation uses a reaction rate model with an autocatalytic term. Its reaction rate constant varies with local temperature according to the Arrhenius relation, and the influence of ambient humidity on the isocyanate crosslinking reaction is taken into account. The coupling between the thermal conductivity equation and the curing kinetics equation results in local temperature increases accelerating the crosslinking reaction, and the exothermic crosslinking reaction further raising the local temperature, thus forming heat accumulation in the heat-limited roll core area. Solving this coupled model yields the curing temperature history and final curing conversion rate of each winding radius and axial position within the roll as curing time changes. The humidity history of each position is determined by the air supply humidity boundary, thus quantitatively reconstructing the spatial gradient where the actual curing temperature closer to the core is higher than that closer to the outer layer of the roll and the exposed areas at both ends of the roll.

[0012] Furthermore, the in-volume curing temperature and humidity spatiotemporal field reconstruction model in step S3 is implemented using a physical information neural network. Specifically, the residual between the cylindrical coordinate heat conduction equation containing the internal heat source and the curing kinetic equation is incorporated as a soft constraint term into the loss function of the physical information neural network. The physical information neural network is configured with two sub-networks corresponding to the temperature field and curing conversion rate, respectively, and trained step-by-step. It is calibrated using temperature and humidity measurement data from points within the curing library and offline peel strength sampling data from historical batches. Further, the equivalent curing heat history of the corresponding sampling location is calculated from the offline peel strength sampling value and injected back into the two sub-networks as an additional physical soft constraint. This ensures that the curing field estimation at unmeasured locations within the volume converges consistently with each batch, allowing the network to satisfy the natural laws of heat transfer and reaction while fitting historical curing measurement data, thus balancing physical consistency and real-time online computation. As an optional technique, the spatiotemporal field reconstruction can also employ the finite volume method to directly solve the coupled equations numerically, or use a reduced-order model to construct a response surface from the offline solution results for rapid online access.

[0013] Step S4 involves inputting the initial adhesive strength potential at each location and the curing history characteristics of that location into a fusion soft sensor model. The model outputs predicted peel strength values ​​for each winding radius and axial position within the roll after curing, resulting in a spatial distribution of peel strength covering the entire roll. Specifically, the curing history characteristics include the equivalent curing thermal history formed by the time integral of the curing temperature history, the final curing conversion rate, the peak curing temperature and its arrival time, and the cumulative amount of humidity exposure. The fusion soft sensor model uses the initial adhesive strength potential as a baseline and the curing history characteristics as correction parameters to output position-resolved predicted peel strength values ​​and their confidence intervals. Thus, the strength basis at the moment of bonding obtained in Step S1 is fused with the curing history at that location obtained in Step S3 to obtain a predicted spatial distribution of cured peel strength covering the entire roll and resolved by position within the roll.

[0014] Step S5: Compare the predicted peel strength values ​​at each location within the roll with the acceptable threshold. Identify local risk areas where the predicted value is lower than the acceptable threshold, or where the difference between the predicted value and the acceptable threshold is less than a preset margin. Then, perform reverse calculation to adjust the controllable quantities in the composite work station area and the curing warehouse area. Specifically, when identifying local risk areas, distinguish the causes of the risk, i.e., distinguish areas near the core where over-curing or under-reaction occurs due to heat accumulation, and areas near the outer layer and the ears at both ends of the roll where insufficient curing occurs due to excessive heat dissipation. Output a risk area distribution map marked by the winding radius and axial position. Further, the reverse calculation adjusts the adhesive temperature in the composite work station area and, through the mapping relationship between adhesive temperature and viscosity, adjusts the target value of the adhesive application amount, adjusts the two-component metering ratio and roller gap pressure. In the curing warehouse area, adjust the air supply temperature, air supply humidity, curing time, and stacking method. Also, for the rolls in production, adjust the winding tension distribution and winding tension taper during the winding stage to bring the predicted peel strength values ​​at each location of the entire roll back to the acceptable range.

[0015] Furthermore, for rolled materials already produced, compensatory regulation is implemented primarily based on the controllable amount in the curing warehouse area. For identified areas of over-curing of the core or insufficient curing of the ear section, the air supply temperature distribution and curing time are adjusted to reduce curing differences within the roll. For rolled materials in production, feedforward regulation is implemented primarily based on the controllable amount in the composite workstation area. At the defect initiation stage, coating and pressing parameters are adjusted to offset the impact of foreseeable spatial heterogeneity in the curing space on peel strength. Preferably, the reverse calculation constructs and solves an optimization problem with the constraint that the predicted peel strength value at each location within the roll is not lower than the qualified threshold, and the objective of minimizing the controllable amount adjustment range. This yields the controllable amount adjustment values ​​for the composite workstation area and the curing warehouse area, thereby minimizing the regulation cost while satisfying the overall roll qualification constraint.

[0016] In another aspect, this invention provides a system for predicting and controlling the adhesive strength of solvent-free composites, comprising a parameter acquisition module, a winding geometry mapping module, a curing spatiotemporal field reconstruction module, a peel strength prediction module, a threshold determination module, and a dual-domain control module. The parameter acquisition module is used to acquire process parameters of the composite process in real time and, according to the acquisition time and the roll length position, outputs the initial adhesive strength potential at each roll length position from a pre-trained primary soft sensing sub-model. The winding geometry mapping module is used to convert the roll length position into the winding radius and axial position of that section of roll within the parent roll according to the Archimedean spiral winding relationship, reconstructing the initial adhesive strength potential into a two-dimensional distribution within the roll calibrated by the winding radius coordinates and axial coordinates. The curing spatiotemporal field reconstruction module is used to describe the heat transfer within the roll using an unsteady heat conduction equation containing an internal heat source in cylindrical coordinates. The internal heat source term is determined by the exothermic rate of the crosslinking reaction given by the curing kinetic equation. The thermal conductivity equation is coupled with the curing kinetic equation and solved to obtain the curing temperature history and curing conversion rate of each winding radius and axial position within the roll. The humidity history of each position is determined by combining the humidity boundary of the curing chamber air supply. The peel strength prediction module is used to input the initial adhesive strength potential of each position and the curing history characteristics of that position into the fusion soft measurement model, and output the peel strength prediction value of each winding radius and axial position within the roll after curing. The threshold determination module is used to compare the peel strength prediction value of each position within the roll with the qualified threshold, and identify local risk areas where the predicted value is lower than the qualified threshold or the difference between the predicted value and the qualified threshold is less than a preset margin. The dual-domain control module is used to reverse calculate and adjust the controllable quantities of the composite work station domain and the curing chamber domain. The modules sequentially transmit data and form a closed loop in the order of parameter acquisition, winding geometry mapping, curing spatiotemporal field reconstruction, peel strength prediction, threshold determination and dual-domain control. The parameter acquisition module and winding geometry mapping module are responsible for steps S1 and S2, the curing spatiotemporal field reconstruction module is responsible for step S3, the peel strength prediction module is responsible for step S4, and the threshold determination module and dual-domain control module are responsible for step S5. Thus, the positional prediction of the peel strength within the roll and the dual-domain reverse control are completed in the composite stage.

[0017] This invention introduces thermo-chemical coupling spatiotemporal field reconstruction of the curing process during the composite stage. This transforms the adhesive strength, which cannot be measured in real-time after curing and exhibits a spatially non-uniform distribution within the roll, into a controllable quantity that can be predicted during the composite stage, distinguished by its location within the roll, and intervened in both the composite stage and the curing reservoir domains. This allows anomalies to be identified and corrected at the defect initiation stage. Compared to existing soft measurement schemes that treat the curing stage as a uniform standard process for the entire roll, this invention can quantitatively reconstruct the curing temperature and humidity field distributions along the winding radius and axial direction within the roll. This overcomes the spatial mismatch between predicted values ​​and the actual non-uniform curing process within the roll. Before curing is complete, it can identify localized areas within the roll that may not reach the acceptable threshold due to deviations from standard curing conditions. Based on this, compensatory controls are implemented for rolls already produced and feedforward controls are implemented for rolls in production, thereby reducing the batch scrapping of entire rolls or even entire batches and improving the consistency of composite strength across different locations within the roll and between different batches. Attached Figure Description

[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] Figure 1 This is an overall flowchart of the adhesive strength prediction and control method for solvent-free composites according to the present invention.

[0020] Figure 2 This is a schematic diagram of the two-dimensional distribution within the roll, showing the mapping between the roll length position, winding radius, and axial position according to the Archimedean spiral winding relationship.

[0021] Figure 3 This is a data flow diagram showing the coupling solution of the unsteady heat conduction equation containing an internal heat source and the curing kinetic equation in the spatiotemporal field reconstruction model of the in-volume curing temperature and humidity of this invention, as well as a network structure diagram of the spatiotemporal field reconstruction using a physical information neural network.

[0022] Figure 4 This is a structural block diagram of the adhesive strength prediction and control system for solvent-free composites according to the present invention.

[0023] Figure 5 This is a diagram illustrating the mechanism of position-resolved intensity synthesis based on the composite instantaneous baseline with winding curing correction of the present invention.

[0024] Figure 6 This is a pictographic diagram of the layout of the solventless laminating machine production line of this invention.

[0025] Figure 7 This is a pictographic diagram of the spatial distribution of the curing temperature field within the roll of this invention.

[0026] Figure 8 This is a comparison chart of the prediction accuracy of peel strength at different positions within the roll in this invention.

[0027] Figure 9 This is a comparison curve of the peel strength distribution along the winding radius within the roll of this invention. Detailed Implementation

[0028] To make the objectives and technical solutions of this invention clearer, the embodiments of this invention will be further described in detail below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of this invention and should not be construed as limiting the scope of protection of this invention.

[0029] Example 1

[0030] This embodiment provides a method for predicting and controlling the adhesive strength of solvent-free composites, the overall process of which is as follows: Figure 1 As shown, the process sequentially includes steps S1 to S5. By introducing thermo-chemical coupling spatiotemporal field reconstruction of winding curing during the composite work station stage, the curing temperature and humidity history and curing transformation process of each position inside the roll material along the winding radius and axial direction are reconstructed. Then, the post-curing peel strength of each position is predicted with positional resolution. Furthermore, for local areas deviating from the qualified threshold, the controllable quantities of the composite work station domain and the curing reservoir domain are calculated in reverse, forming a spatially resolved closed loop for prediction and control. To facilitate understanding by those skilled in the art, this embodiment first describes the online execution process completed by the trained model, and then separately describes the training process of each model.

[0031] Step S1: Real-time acquisition of process parameters for the composite process and corresponding acquisition time with the roll length position, outputting the initial bonding strength potential at each roll length position by the pre-trained primary soft measurement sub-model.

[0032] Specifically, sensors are installed at the gluing unit, pressing unit, and winding unit of the solventless laminating machine to collect process parameters in real time, including adhesive temperature and its corresponding viscosity, two-component metering ratio, gluing roller speed ratio, roller gap pressure, laminating speed, substrate surface tension, and ambient temperature and humidity. The adhesive temperature is measured by thermocouples at the gluing pan, and its corresponding viscosity is calculated from a pre-calibrated temperature-viscosity curve or directly measured by an online viscometer. The two-component metering ratio is given by the flow rate ratio of the two-component adhesive supply pump. The gluing roller speed ratio is the ratio of the metering roller speed to the coating roller speed. The roller gap pressure is measured by pressure sensors at both ends of the pressing roller. The laminating speed is taken from the linear velocity feedback of the main drive of the unit. The substrate surface tension is obtained by an online dyne value detection device after online corona treatment. Ambient temperature and humidity are given by temperature and humidity transmitters located near the machine. The symbols, data sources, units, typical ranges, sampling periods, and normalization methods for the above process parameters are as follows: The first item is the adhesive temperature, denoted as T. gThe first item is the temperature reading, measured by a thermocouple at the top of the adhesive tray, in degrees Celsius, with a typical range of 25 to 60. The sampling period is 100 milliseconds, and maximum-minimum normalization is used. The second item is the adhesive viscosity, denoted as η, measured by an online viscometer or calculated from a temperature-viscosity curve, in millipascal-seconds, with a typical range of 200 to 2000. The sampling period is 100 milliseconds, and maximum-minimum normalization is used after taking the logarithm. The third item is the two-component metering ratio, denoted as R. act The first term is the flow ratio of the two-component glue supply pump, a dimensionless mass ratio with a typical range of 100:70 to 100:110, a sampling period of 100 milliseconds, and zero-mean unit variance normalization. The second term is the glue application roller speed ratio, denoted as i, which is the ratio of the speeds of the metering roller and the coating roller. It is dimensionless, with a typical range of 0.30 to 1.00, a sampling period of 50 milliseconds, and maximum-minimum normalization. The third term is the roller gap pressure, denoted as p. nip The following parameters are used for measurement: 1) Pressure sensor at both ends of the pressing roller, measured in N / mm linear pressure, with a typical range of 20 to 80, a sampling period of 50 milliseconds, and normalized to maximum and minimum; 2) Composite velocity, denoted as v, is obtained from the linear velocity feedback of the main drive of the unit, measured in meters per minute, with a typical range of 100 to 450, a sampling period of 50 milliseconds, and normalized to maximum and minimum; 3) Substrate surface tension, denoted as γ, is collected by an online dyne detection device, measured in millinewtons per meter, with a typical range of 34 to 52, a sampling period of 1 second, and normalized to maximum and minimum; 4) Ambient temperature, denoted as T. a The temperature and humidity data are collected by the machine's temperature and humidity transmitter, in degrees Celsius, with a typical range of 15 to 35 degrees Celsius and a sampling period of 1 second. Zero-mean unit variance normalization is used. The ninth item is the ambient relative humidity, denoted as RH. a The relative humidity is collected by the machine's temperature and humidity transmitter, with the unit being percentage relative humidity. The typical range is 30 to 80%, and the sampling period is 1 second. Zero-mean unit variance normalization is used. In addition, the roll length position is recorded as L, obtained by pulse length measurement of the main drive encoder, with the unit being meters. The typical range is 0 to 8000, and it is collected synchronously with the above parameters.

[0033] The nine parameters, from the first to the ninth item, constitute the direct input to the model. The roll length position L is used to map each parameter to a specific location on the roll, serving as a position index without participating in normalization. The normalization coefficients of each parameter are determined by the training set statistics and are fixed along with the model. They are directly called during online inference without being updated, ensuring consistency in feature distribution between online and offline methods.

[0034] like Figure 6As shown, the solventless lamination production line of the present invention includes, in sequence along the lamination conveyor belt direction, an adhesive application unit, a pressing unit, and a winding unit. The adhesive application unit is equipped with an adhesive application tray, a metering roller, and a coating roller. A thermocouple is arranged at the adhesive application tray to measure the adhesive temperature. A two-component adhesive supply pump and a flow meter control the metering ratio of the two components. The rotational speed ratio of the metering roller and the coating roller determines the amount of adhesive applied. After adhesive application, the substrate moves forward along the conveyor belt direction, undergoing online corona treatment along the way, and the surface tension of the substrate is detected by dyne value. The ambient temperature and humidity are collected by a temperature and humidity transmitter, and the main drive encoder of the unit outputs pulses to count the length and determine the roll length position. The pressing unit is equipped with a pair of pressing rollers, with pressure sensors arranged at both ends to collect the roller gap pressure. The multi-layer substrate is pressed into a composite structure at the roller gap. The laminated substrate is wound into a large-diameter master roll in the winding unit. The master roll has a defined target roll diameter, core, and winding tension. The signals from each sensor constitute the input of the parameter acquisition stage.

[0035] During the data acquisition process, the encoder pulses of the main drive of the unit are used to accumulate and count the length of the roll material, thus assigning a unique roll material length position coordinate to each acquisition moment. After mapping each process parameter to the roll material length position according to the acquisition moment, a composite station parameter sequence continuously distributed along the roll material length direction is obtained. Specifically, the roll material length can be discretized at equal length intervals, for example, taking a position node every one meter or every preset sampling length. The average value of the process parameters in the neighborhood of this node is used as the feature vector of that length position, so that subsequent processing is performed on a unified length grid.

[0036] After obtaining the parameter sequence of the lamination station, the feature vectors of each length position are input into a pre-trained primary soft sensing sub-model, outputting the initial adhesive strength potential at that length position. This initial adhesive strength potential is the baseline peel strength determined solely by the coating and pressing quality during the lamination process, under the standard uniform curing assumption. It characterizes the strength foundation established by the coating and pressing stages during lamination, without considering the impact of differentiated curing after the roll material leaves the production line. The reason for outputting a baseline quantity without curing influence at this stage is that the coating and pressing quality during lamination and the subsequent in-roll curing process are physically independent. Decoupling and modeling them separately allows for a clear distinction between the controllable quantities of the lamination station and the controllable quantities of the curing library in subsequent adjustments.

[0037] To ensure that the initial bond strength potential is a measurable and reproducible quantifiable quantity, this embodiment provides an operational definition based on the following standard curing conditions: Samples are cut at preset length intervals at the lamination station. The samples are placed flat in a single-layer manner in a constant temperature and humidity chamber and cured for 48 hours at 45°C and 50% RH. After curing, the interlayer peel strength is measured according to the T-type peel test method specified in GB / T 8808, with a sample width of 15 mm and a peeling speed of 300 mm per minute. The measured peel strength value is defined as the initial bond strength potential at that length, expressed in Newtons per 15 mm. The reason for specifying single-layer flat curing is that it ensures consistent heat dissipation conditions throughout the sample. There are no differences in heat dissipation paths caused by winding and stacking, nor is there any accumulation of crosslinking heat within the roll. This practically realizes the standard uniform curing assumption of this invention, ensuring that the quantity reflects only the coating and pressing quality at the moment of lamination and is independent of the position within the roll. The temperature, humidity, and duration in the above-mentioned standard curing conditions are a set of specific values ​​for this embodiment. In practice, other agreements may also be made, as long as the same agreed conditions are maintained in the acquisition of training labels and subsequent calibration. The values ​​themselves do not affect the implementation of this invention.

[0038] It should be noted that the initial bond strength potential is not a quantity "separated" from the cured strength by adjusting certain process parameters, but rather a quantity uniquely determined by the process parameters at the moment of bonding and predicted by the primary soft sensing sub-model based on those parameters. Its value is determined at the moment of bonding. The aforementioned standard curing peel test is only used to obtain the true value of this quantity offline as a training label for the model. This test is not required during online operation; instead, the trained model directly infers the value from the real-time collected process parameters. Therefore, those skilled in the art can either measure this quantity using the aforementioned standard curing peel test or calculate it from the process parameters using the model described below. These two methods correspond to the offline calibration and online prediction stages, respectively.

[0039] Preferably, the primary soft sensing sub-model employs a gradient boosting tree or a fully connected neural network. When using a gradient boosting tree, it exhibits good fitting ability for nonlinear relationships and feature interactions in process parameters and is insensitive to feature dimensions; when using a fully connected neural network, it facilitates unified deployment with subsequent fused soft sensing models. Furthermore, to improve the physical interpretability and accuracy of the model, in addition to directly collected process parameters, derived features are used as input. These derived features include the actual coating viscosity obtained from the mapping relationship between adhesive temperature and viscosity, the two-component ratio deviation obtained from the deviation between the actual and theoretical ratios of the two components, and the actual adhesive amount calculated from the ratio of the gluing roller speed to the lamination speed. The actual adhesive amount reflects the adhesive layer thickness per unit area of ​​the substrate and is a key quantity determining the bonding foundation at the moment of lamination. Explicitly using it as a derived feature input allows the model to output a physically meaningful initial bonding strength potential distributed along the length of the roll material at the lamination stage.

[0040] The calculation methods for the three derived features are as follows. Actual coating viscosity T g The measured adhesive temperature, η0, and B are temperature-viscosity curve constants obtained beforehand by calibrating a rotational viscometer at several temperature points; the two-component distribution ratio deviation. , where R act The measured mass flow ratio of curing agent to main agent, R, is... set Set the mixing ratio for the process; actual glue volume Where i is the speed ratio between the metering roller and the coating roller, and D c With ω c Here, v represents the coating roller diameter and rotational speed, respectively; v is the compounding speed; and k is the coating roller diameter and rotational speed, respectively. c G is the adhesive transfer factor determined by the anilox volume ratio of the coating roller, and its unit is grams per square meter. Additionally, it is determined by the ambient temperature T. a relative humidity (RH) a The dew point temperature T obtained by converting according to the dew point formula d It is also used as an input to characterize the impact of substrate surface condensation risk on adhesion.

[0041] The specific structure and training hyperparameters of the primary soft sensing sub-model are further given. The input of the model is a thirteen-dimensional feature vector, namely the nine directly acquired parameters from the first to the ninth item mentioned above, plus the four derived features mentioned above: actual coating viscosity, two-group distribution ratio deviation, actual adhesive amount, and dew point temperature (where the actual coating viscosity coincides with item 2 when it is obtained from the temperature-viscosity curve, and in this case, the difference between the measured value and the converted value of the online viscometer is used as the input of this dimension). The output is a single scalar, namely the initial adhesive strength potential at this length position, in Newtons per 15 millimeters.

[0042] When using gradient boosting trees, the base learner is a regression decision tree with 500 trees, a maximum tree depth of 6, a learning rate of 0.05, a feature sampling ratio of 0.8 per tree, a sample subsampling ratio of 0.8, and a minimum number of samples per leaf node of 20. The loss function is the squared error, and early stopping is determined by the validation set loss not decreasing for 30 consecutive rounds. When using a fully connected neural network, the network consists of an input layer, three hidden layers, and an output layer connected sequentially. The input layer has a dimension of 13, and the number of neurons in the three hidden layers are 128, 64, and 32, respectively. Each hidden layer is followed by a modified linear unit activation function and a randomly deactivated layer with a deactivation ratio of 0.1. The output layer is a single neuron without an activation function. All layers are fully connected without cross-layer connections. The training hyperparameters are described later in the section on model training.

[0043] To facilitate understanding of the correspondence between the above inputs and outputs by those skilled in the art, a set of numerical examples is provided. Assume that at a length L = 1250 meters of a roll of material, the following data is collected: adhesive temperature 40.2℃; actual coating viscosity calculated from the temperature-viscosity curve is approximately 820 mPa·s; the measured mass flow ratio of the two components is 100:78.4, while the process-set ratio is 100:80 (corresponding to a two-component ratio deviation of approximately −2.0%); the gluing roller speed ratio is 0.62; the roller gap pressure is 48 N / mm; the lamination speed is 300 m / min; the substrate surface tension is 40 mN / m; the ambient temperature is 26℃; and the ambient relative humidity is 55%RH (corresponding to a dew point temperature of approximately 16.3℃). Based on this, the actual adhesive content is estimated to be approximately 1.35 grams per square meter. After normalizing the above thirteen-dimensional feature vector according to the aforementioned parameter normalization methods, it is input into the trained primary soft sensing sub-model. The model outputs that the initial adhesive strength potential at this length position is approximately 3.6 N per 15 mm. If the adhesive temperature is lowered to 35.0℃ while other conditions remain unchanged, the actual coating viscosity increases and the actual amount of adhesive decreases to about 1.20 g / m². The initial bond strength potential output by the model is correspondingly reduced to about 3.3 N / 15 mm, which reflects the inherent correlation between process parameters and this amount.

[0044] Step S2: Convert the length position of the roll material into its winding radius and axial position within the master roll according to the Archimedes spiral winding relationship. Reconstruct the initial adhesive strength potential into a two-dimensional distribution within the roll, calibrated by the winding radius and axial coordinates, as shown below. Figure 2 As shown.

[0045] After lamination, the roll material is wound into a large-diameter master roll in the winding unit. Any length position on the roll corresponds to a specific geometric position within the master roll after winding. Furthermore, based on the core diameter, the thickness of the single-layer composite structure, the winding tension, and its winding tension taper, a mapping between the roll length position and the winding geometric position is established according to the Archimedean spiral winding relationship. Specifically, assuming the core radius is the initial radius, the actual compressed thickness of the single-layer composite structure under winding tension varies with the winding tension and the winding tension taper. With each additional layer of the roll material, the winding radius increases according to this actual compressed thickness, thus forming an Archimedean spiral relationship between the cumulative length of the roll material from the core and its corresponding winding radius. Based on this, the winding radius within the master roll can be calculated from any length position. Simultaneously, the position of the roll material along its width directly corresponds to the axial coordinate of the master roll, and the edges of the roll material correspond to the exposed ear areas at both ends of the master roll.

[0046] Through the above mapping, the initial adhesive strength potential, which is distributed one-dimensionally along the length direction in step S1, is reconstructed into a two-dimensional distribution within the roll, calibrated by the winding radius coordinates and the axial coordinates. This reconstruction allows the strength basis obtained from the composite workstation to be arranged according to the actual geometric position within the roll, providing a spatial coordinate reference for subsequent position-resolved curing field reconstruction and strength prediction. Furthermore, considering that the winding tension taper will make the outer layer winding relatively loose, a layer thickness correction that varies with the winding radius can be introduced into the mapping, making the conversion of the winding radius more consistent with the actual winding shape of the master roll.

[0047] While establishing the geometric mapping, the width of the master roll, the core diameter, the target roll diameter, the winding tension distribution, and the supply air temperature, supply air humidity, hot air circulation mode, and stacking method of the curing chamber are collected as boundary conditions and structural parameters for subsequent curing spatiotemporal field reconstruction. Among them, the width of the master roll and the target roll diameter determine the spatial solution domain of the reconstruction model, the winding tension distribution affects the interlayer contact thermal resistance within the roll, the supply air temperature and supply air humidity of the curing chamber constitute the convective heat transfer boundary and humidity boundary of the solution domain, and the hot air circulation mode and stacking method determine the intensity of convective heat transfer in each direction of the outer surface of the master roll and the mutual shading between adjacent master rolls.

[0048] Step S3: Construct a spatiotemporal field reconstruction model of the curing temperature and humidity within the roll. The heat transfer within the roll is described by an unsteady thermal conductivity equation in cylindrical coordinates, containing an internal heat source. The internal heat source term is determined by the exothermic rate of the crosslinking reaction given by the curing kinetics equation. The thermal conductivity equation and the curing kinetics equation are coupled and solved to obtain the curing temperature history and curing conversion rate at each winding radius and axial position within the roll. The humidity history at each position is determined by combining this with the humidity boundary of the curing chamber air supply. The coupled solution data stream is as follows: Figure 3 As shown.

[0049] Specifically, the master roll is considered as an axisymmetric solution domain in cylindrical coordinates. A two-dimensional solution domain is formed by the winding radius coordinates and the axial coordinates. An unsteady-state heat conduction equation containing an internal heat source describes the heat transfer within the roll. That is, the temperature change at any location within the roll over time is determined by the heat conduction along the radial and axial directions at that location, as well as by the internal heat source at that location. The internal heat source term represents the unit volume exothermic rate of the adhesive crosslinking reaction, which is equal to the product of the exothermic enthalpy of the reaction and the rate of change of the curing conversion rate over time. The rate of change of the curing conversion rate over time is given by the curing kinetic equation.

[0050] The cylindrical coordinate unsteady heat conduction equation containing the internal heat source is as follows:

[0051]

[0052] Where T is the temperature inside the roll, t is the curing time, r is the winding radius coordinate, z is the axial coordinate; ρ is the equivalent density of the composite structure, c is the equivalent specific heat capacity; λ r With λ z λ and y are the equivalent thermal conductivity along the winding radius and axial direction, respectively. These two values ​​are not equal to reflect the anisotropy of heat transfer in the wound laminated structure. r It is significantly smaller than λ due to the cumulative interlayer contact thermal resistance. z This is precisely the physical root cause of the difficulty in heat dissipation of the coil core; q v The internal heat source intensity is given by the following formula:

[0053]

[0054] Where ΔH is the total exothermic enthalpy per unit mass of the adhesive crosslinking reaction, and ρ g α represents the mass of adhesive contained in a unit volume of the composite structure, and α represents the curing conversion rate. The following curing kinetic equations provide the values. Typical ranges for the above physical properties are: ρ = 900–1100 kg / m³, c = 1500–2000 J / kg Kelvin, and λ = 2000 kg / kg Kelvin. r Take 0.05~0.15 watts per meter Kelvin, λ z Take 0.20~0.40 watts per meter Kelvin, ΔH; take 120~200 kJ per kilogram of adhesive, ρ g The amount of adhesive applied is calculated from the thickness of the single-layer composite structure, and is usually taken as 8~20 kg per cubic meter. The above parameters are calibrated by differential scanning calorimetry and thermal property tests of roll materials, or determined by inversion of the measured curing temperature of historical batches.

[0055] The initial and boundary conditions of the equation are as follows: the initial condition is that at time t=0 (i.e., the time when the roll material is transferred from the production line to the curing chamber), the temperature at various points inside the roll is taken as the production line temperature T0, and the curing conversion rate is taken as the initial conversion rate α0 at the production line time; on the outer surface of the mother roll (r=R) o The third type of convective heat transfer boundary condition is applied to the ear ends (z=0 and z=W) of the mother roll, i.e. Where n is the outward normal direction of the boundary, λ is the equivalent thermal conductivity in that direction, and T air The temperature of the curing chamber air supply is given by r, and h is the convective heat transfer coefficient of the surface. Its value is determined by the hot air circulation method and stacking method of the curing chamber, with a larger value for the windward side and a smaller value for the surface blocked by adjacent mother rolls. A typical value range is 5~25 Kelvin per square meter. On the core surface (r=R... i Insulation boundary measurements are taken based on whether the core is ventilated. Or the convective boundary between the parent roll and the air inside the core cavity. Since the parent roll is axisymmetric, the solution domain is taken as the axial-radial two-dimensional cross-section of the parent roll, and the circumferential direction is not discretized.

[0056] The curing kinetic equation preferably employs a reaction rate model with an autocatalytic term, meaning the rate of change in curing conversion is equal to the product of the reaction rate constant and the autocatalytic function related to the conversion rate. This autocatalytic function reflects the autocatalytic characteristic of accelerating the reaction in the initial stage as more products are generated and slowing it down in the later stage as reactants are depleted. The reaction rate constant varies with local temperature according to the Arrhenius relation, i.e., it increases exponentially with increasing local temperature. Furthermore, the influence of ambient humidity on the isocyanate crosslinking reaction is taken into account; humidity is introduced as a correction factor into the reaction rate constant to reflect the role of moisture in the isocyanate reaction on the crosslinking process.

[0057] The curing kinetic equation with the autocatalytic term is:

[0058]

[0059] Where α is the curing conversion rate, ranging from 0 to 1, where α=0 indicates that the crosslinking reaction has not yet started, and α=1 indicates that the crosslinking reaction is complete; m is the reaction order of the autocatalytic term, and n is the reaction order of the reactant consumption term. In this embodiment, m is taken as 0.3~0.8 and n is taken as 1.2~1.8, determined by fitting the isothermal curing curve from the differential scanning calorimetry experiment. The reaction slows down in the later stages as reactants are depleted. The term "autocatalytic term" reflects the promoting effect of the increased number of urethane groups generated in the early stages of the reaction on subsequent cross-linking. Together, these two terms constitute the autocatalytic characteristic of initial acceleration followed by deceleration. k1 and k2 are the rate constants for non-autocatalytic and autocatalytic reactions, respectively, both expressed according to the Arrhenius relation and varying with local temperature, taking into account the influence of humidity.

[0060]

[0061]

[0062] Where A1 and A2 are pre-exponential factors, in units of seconds; E1 and E2 are activation energies of the reaction, in this embodiment E1 is 50-70 kJ / mol and E2 is 40-60 kJ / mol; R is the molar gas constant, taken as 8.314 J / mol Kelvin; T is the local absolute temperature at this location, that is, the value of the temperature field given by the above heat conduction equation at this location and at this time. This is a humidity correction factor used to account for the effect of moisture participation in the isocyanate reaction on the crosslinking process. In this embodiment, it is set to:

[0063]

[0064] Where RH is the relative humidity percentage at that location, determined by the humidity boundary of the curing chamber supply air and the moisture diffusion within the roll; RH ref c is the reference relative humidity used during calibration. w The humidity sensitivity coefficient ranges from 0.05 to 0.20 in absolute value. It should be noted that moisture has a dual effect on isocyanate crosslinking: appropriate amounts of moisture participating in the reaction can accelerate the formation of the crosslinked network, thus increasing the concentration of c... w When the value is positive, excess moisture will react with isocyanate groups to form urea bonds, release carbon dioxide, and consume cross-linking sites, thus reducing the concentration of carbon dioxide. w Taking a negative value, therefore c w The sign and value should be determined by the measured data of historical batches during the ripening process, and should not be fixed in advance.

[0065] The above-mentioned curing kinetics equation and the aforementioned heat conduction equation form a two-way coupled system of equations: the internal heat source term q in the heat conduction equation v The conversion rate depends on the curing kinetic equation. The rate constants k1 and k2 in the curing kinetic equation depend on the local temperature T given by the heat conduction equation. The two must be solved simultaneously and cannot be solved separately.

[0066] The coupling relationship between the thermal conductivity equation and the curing kinetics equation is as follows: the increase in local temperature increases the reaction rate constant according to the Arrhenius relation, thereby accelerating the crosslinking reaction; the exothermic reaction of the crosslinking reaction further raises the local temperature as an internal heat source. In the region near the core, due to the long heat dissipation path and the accumulation of interlayer contact thermal resistance, the exothermic reaction is difficult to dissipate in time, resulting in heat accumulation. This makes the actual curing temperature closer to the core higher than the actual curing temperature near the outer layer of the roll and the exposed areas at both ends of the roll. Therefore, by combining the thermal conductivity equation and the curing kinetics equation, and applying initial boundary conditions with supply air temperature and supply air humidity as convection boundaries, the coupled model is solved. This yields the curing temperature history and final curing conversion rate of each winding radius and axial position within the roll as a function of curing time. The humidity history at each position is determined by the supply air humidity boundary, allowing the spatial gradient of curing temperature along the winding radius and axial directions to be quantitatively reconstructed.

[0067] like Figure 7 As shown, the grayscale distribution of the curing temperature field within the roll is superimposed on the axial-radial cross-section of the master roll as the carrier. The horizontal axis represents the axial coordinate z, with the two ends being the ears; the vertical axis represents the winding radius r, pointing from the core to the outer layer. The grayscale levels from white to black correspond to curing temperatures from low to high. The core area has the longest heat dissipation path and accumulates heat, resulting in the highest curing temperature and a risk of over-curing or under-reaction. Heat dissipation gradually increases outward along the winding radius, leading to lower curing temperatures in the outer layer. The exposed ear areas at both ends of the master roll dissipate heat the fastest, resulting in the lowest curing temperature and insufficient curing. The spatial non-uniformity of the curing temperature field within the roll is the fundamental reason for the systematic deviation in the peel strength of the entire roll along the winding radius and axial direction. Based on this, the present invention performs position-resolved strength prediction and regional control.

[0068] Furthermore, the in-volume ripening temperature and humidity spatiotemporal field reconstruction model in step S3 is preferably implemented using a physical information neural network, the network structure of which is as follows: Figure 3As shown. Specifically, the winding radius coordinates, axial coordinates, and curing time are used as network inputs. The residuals of the cylindrical coordinate heat conduction equation containing the internal heat source and the curing kinetic equation are incorporated as soft constraints into the loss function of the physical information neural network. This allows the network to satisfy the natural laws of heat transfer and reaction while fitting historical curing measurement data, balancing physical consistency and real-time online computation. The physical information neural network is configured with two sub-networks corresponding to the temperature field and curing conversion rate, respectively, and is trained in stages. First, the temperature field sub-network fits the temperature and humidity measurement data in the curing library and initially satisfies the residuals of the heat conduction equation. Then, the curing conversion rate sub-network is introduced to fine-tune the residuals of the curing kinetic equation, thereby mitigating the training instability caused by the large difference in dimensions between the two types of physical quantities. As an optional technique, the spatiotemporal field reconstruction can also use the finite volume method to directly solve the coupled equations numerically, or use a reduced-order model to construct a response surface from the offline solution results for rapid online access, further improving the online response speed when computational resources are limited.

[0069] Step S4: Input the initial adhesive strength potential at each position and the curing history characteristics at that position into the fusion soft measurement model, and output the predicted peel strength value after curing for each winding radius and axial position within the roll, to obtain the spatial distribution of peel strength covering the entire roll.

[0070] Specifically, the curing history features of the location are extracted from the temperature history, humidity history, and curing conversion rate obtained in step S3. These curing history features include the equivalent curing heat history (the integral of curing temperature history over curing time), the final curing conversion rate, the peak curing temperature and its arrival time, and the cumulative humidity exposure. The equivalent curing heat history comprehensively reflects the amount of heat experienced by the location throughout the curing process; the final curing conversion rate reflects the degree of completion of the crosslinking reaction; the peak curing temperature and its arrival time reflect whether the area has experienced excessive heat or a rapid reaction; and the cumulative humidity exposure reflects the influence of moisture on crosslinking.

[0071] The initial adhesive strength potential at the location obtained in step S1 is concatenated with the aforementioned curing history features to form a feature vector, which is then input into the fusion soft measurement model. The fusion soft measurement model uses the initial adhesive strength potential as a baseline and the curing history features as correction parameters to output a location-resolved peel strength prediction value and its confidence interval. The reason for using a baseline plus correction structure is that the initial adhesive strength potential already characterizes the strength basis at the moment of bonding, while the curing history features characterize the gain or loss caused by differentiated curing on this basis. Adding the two together yields the actual peel strength at that location after curing. Therefore, by calculating the peel strength spatial distribution prediction for each winding radius and axial position within the roll, covering the entire roll and resolving to the position within the roll, the final cured peel strength prediction is obtained. The specific structure of the fusion soft measurement model is as follows: its input is a six-dimensional feature vector, namely, the initial adhesion strength potential at that location (one dimension), and five dimensions of curing history features: equivalent curing heat history, final curing conversion rate, peak curing temperature, peak curing temperature arrival time, and cumulative humidity exposure. The network consists of an input layer, two hidden layers, and an output layer, all fully connected sequentially. The number of neurons in the two hidden layers is 64 and 32 respectively, and the activation function of the hidden layers uses modified linear units. The output layer does not directly provide the peel strength, but instead provides a correction amount Δy relative to the baseline. The final position-resolved peel strength prediction value is calculated according to... Synthesis, where y pot (r,z) represents the initial adhesion strength potential at this location. By outputting a correction value instead of directly outputting the strength, the aforementioned baseline plus correction decoupling structure is forced to be implemented at the network structure level, preventing the network from degenerating into an end-to-end regression of all inputs during training and losing this decoupling property.

[0072] The confidence interval is implemented using quantile regression. Specifically, the output layer of the fused soft sensor model has three parallel output heads, trained at quantile loss at 0.05, 0.50, and 0.95 quantiles. The correction value from the 0.50 quantile output head, combined with the baseline, serves as the predicted peel strength at that location. The combined values ​​from the 0.05 and 0.95 quantile output heads constitute the lower and upper bounds of the predicted value at that location, respectively, and the difference between them is the width of the 90% confidence interval. Alternatively, during the inference phase, Monte Carlo random inactivation can be applied to the hidden layer of the model, and the forward inference can be repeated several times (e.g., 50 times) with the same input, using the standard deviation of the multiple outputs to characterize the prediction uncertainty. Alternatively, several models independently trained with different random initializations can be integrated, with the variance of each model's output characterizing the prediction uncertainty. The confidence interval provides a risk margin reference for subsequent threshold determination.

[0073] like Figure 5As shown, the spatial distribution of the overall roll peel strength described in this invention is synthesized from the composite instantaneous baseline and the winding curing correction amount. Specifically, the initial adhesive strength potential serves as the baseline, determined by the instantaneous coating and pressing quality, and its fluctuations are gentle along the roll position. The curing history characteristic correction amount applies gain or loss based on position on this baseline. The correction sources include equivalent curing thermal history, final curing conversion rate, peak curing temperature and its arrival time, and cumulative humidity exposure. The baseline and correction amount are superimposed to obtain the position-resolved peel strength prediction value, which fluctuates more significantly along the roll position. It is resolved by the winding radius r and axial coordinate z and includes a confidence interval, covering the entire roll and eliminating the spatial mismatch between the existing assumption of overall roll uniformity and the actual distribution within the roll when only a single strength value is given. This constitutes the distinguishing technical feature of this invention.

[0074] Step S5: Compare the predicted peel strength values ​​at each location within the roll with the qualified threshold, identify local risk areas where the predicted values ​​are lower than the qualified threshold or the difference between the predicted values ​​and the qualified threshold is less than a preset margin, and reverse calculate and adjust the controllable amount of the composite work station domain and the curing warehouse domain.

[0075] Specifically, the predicted peel strength values ​​at each location within the roll obtained in step S4 are compared with the acceptable threshold. Locations where the predicted value is lower than the acceptable threshold, or where the predicted value is higher than the acceptable threshold but the difference between the predicted value and the acceptable threshold is less than the confidence interval, are identified as local risk areas. When identifying local risk areas, the causes of risk are distinguished, specifically areas near the core where over-curing or under-reaction is caused by heat accumulation, and areas near the outer layer and the ear ends of the roll where insufficient curing is caused by excessive heat dissipation. A risk area distribution map, marked by the winding radius and axial position, is output. The basis for distinguishing the causes is that risks in the core area are usually accompanied by higher peak curing temperatures and abnormal curing conversion rates, while risks in the outer layer and ear areas are usually accompanied by lower curing temperatures and insufficient curing conversion rates. Based on this, it can be determined whether local areas should be adjusted towards enhanced curing or suppression of over-curing.

[0076] Furthermore, the reverse calculation adjusts the adhesive temperature in the composite work station domain and, through the mapping relationship between adhesive temperature and viscosity, adjusts the target value of adhesive application, the metering ratio of the two components, and the roller pressure. In the curing warehouse domain, it adjusts the air supply temperature, air supply humidity, curing time, and stacking method. For the rolls in production, it adjusts the winding tension distribution and winding tension taper during the winding stage, so that the predicted peel strength values ​​at all positions of the entire roll fall back to the acceptable range. Specifically, for rolls already produced, compensatory control is implemented primarily based on the controllable amount in the curing warehouse domain. For identified areas of over-curing of the core or insufficient curing of the ear, the air supply temperature distribution and curing time are adjusted to reduce curing differences within the roll. For rolls in production, feedforward control is implemented primarily based on the controllable amount in the composite work station domain. At the defect cause stage, the coating and pressing parameters are adjusted to offset the impact of foreseeable spatial heterogeneity in the curing space on peel strength. Preferably, the reverse calculation constructs and solves an optimization problem with the constraint that the predicted peel strength value at each position within the roll is not lower than the qualified threshold, and with the objective of minimizing the controllable adjustment range. This yields the controllable adjustment values ​​for the composite workstation domain and the maturation warehouse domain, thereby minimizing the control cost while satisfying the overall roll qualification constraint. To ensure that the reverse calculation has a specific form that can be implemented by those skilled in the art, this embodiment constructs it as a minimum adjustment optimization problem with nonlinear constraints, and solves it using a sequential quadratic programming algorithm within a rolling time domain framework. The decision variables, objective function, constraints, solution algorithm, and their selection criteria are described below.

[0077] First, define the decision variable vector u. u is composed of controllable quantities from the composite workstation domain and the maturation warehouse domain, as follows:

[0078]

[0079] The first four items belong to the composite workstation area, namely, the adhesive temperature setting, the two-component metering ratio setting, the roller gap pressure setting, and the winding tension taper; the latter four items belong to the curing area, namely, the air supply temperature, the air supply humidity, the curing time, and the stacking method. The adjustable range, the upper limit of the single adjustment rate, and whether they are applicable to already rolled materials are as follows: Adhesive temperature T g This is a composite workstation area with an adjustable range of 30 to 55 degrees Celsius. The maximum adjustment rate per cycle is 2 degrees Celsius. It is not suitable for rolls of material that have already been produced. The two-component metering ratio R... set This is a composite workstation area, with an adjustable range of 100:75 to 100:100. The maximum adjustment rate per cycle is 1.5%. It is not applicable to rolls that have already been rolled off the production line. Roller pressure p nip This is a composite workstation area, with an adjustable range of 25 to 70 N / mm, and a maximum single adjustment rate of 5 N / mm. It is not applicable to rolls that have already been unloaded. The winding tension taper τ... w This is a composite workstation area, with an adjustable range of 0% to 40%, and a maximum single adjustment rate of 5%. It is not applicable to rolls of material that have already been produced. The supply air temperature is T.air This is a curing storage area, with an adjustable range of 35 to 55 degrees Celsius and a maximum single adjustment rate of 2 degrees Celsius. It can be used for rolled materials that have already been produced. Supply air humidity (RH) air This is a curing warehouse area with an adjustable relative humidity range of 30% to 70%, and a single adjustment rate capped at 5% relative humidity. It can be used for rolls that have already been produced. Curing time: t cure This is a maturation storage area, with an adjustable range of 24 to 96 hours and a maximum single adjustment rate of 6 hours. It can be used for rolled materials that have already been produced. The stacking method is s. stack It belongs to the maturation warehouse area, and its selectable gears are three levels: close-packed, spaced-out, and single-packed. The maximum adjustment rate at one time is level 1, and it can be used for rolls that have already been taken off the production line.

[0080] In this invention, the components of the composite workstation domain are adjustable only for in-production rolls still in the composite workstation. For rolls that have already been wound and rolled off the line, the composite process is complete and cannot be reversed. Therefore, for rolls that have already rolled off the line, only the components of the curing warehouse domain are adjustable. This distinction is the basis for the implementation of the domain-based control of this invention. stack It is a discrete quantity, represented by a finite number of bits.

[0081] Next, we define the objective function and constraints. The objective is to minimize the weighted sum of the normalized adjustment magnitudes of each controllable quantity relative to the current setpoint. The constraint is that the predicted peel strength at each location across the entire roll is not lower than the sum of the qualified threshold and the preset margin, as follows:

[0082] Objective function:

[0083] Constraint (1): For any position (r, z) in the solution domain,

[0084] Constraint (2):

[0085] Constraint (3):

[0086] Constraint (4): For the rolled material that has been removed from the production line, each component of the composite work station domain is fixed as follows:

[0087] Among them, u 0 Set the current process value; u i min with u i max Δu represents the upper and lower limits of the adjustable range of each of the aforementioned decision variables. i max This represents the upper limit of the single adjustment rate for each of the aforementioned decision variables; To reflect the weight of the adjustment costs of each controllable quantity, quantities with a wide impact or high adjustment costs (such as the need for re-stacking of stacking methods or the impact of curing time on turnover efficiency) are given a larger weight, while quantities that can be smoothly and continuously adjusted (such as air supply temperature) are given a smaller weight; S th δ is the acceptable threshold; δ is the preset margin, which can be taken as half the width of the confidence interval at this position during implementation. The predicted peel strength at that position is obtained by substituting u into the forward prediction chain formed by steps S1 to S4. That is, the constraint is related to the decision variable through the entire prediction chain rather than an explicit analytical expression. Constraint (1) is applied point by point on the discrete grid within the volume in implementation, or the minimum value min ŷ of the predicted value on the grid is taken to construct a single constraint to reduce the size of the optimization problem.

[0088] The selection criteria and specific implementation methods of the solution algorithm are explained again. The controlled object is a batch process with multiple variables, spatial distribution constraints, and large time delays on the order of tens of hours. Proportional-integral-derivative control (PIDC) relies on feedback adjustment based on a single scalar deviation, which cannot handle multi-point constraints distributed within the roll, nor can it cope with the long time delays of the ripening process, so it is not suitable. Reinforcement learning methods require a large number of online trial-and-error interactions to obtain strategies, but in actual production, the cost of each trial and error is the material of the entire roll or even the entire batch, making it impossible to provide a feasible training environment, so it is also not suitable. Therefore, this embodiment selects constrained nonlinear programming and implements it under the rolling time domain (i.e., model predictive control) framework: using the forward prediction chain as the process model, the above optimization problem is resolved based on the latest acquired data in each control cycle, and only the adjustment amount corresponding to the current control cycle is executed. The problem is then resolved based on the updated prediction results in the next control cycle, thereby suppressing the cumulative effects of model mismatch and external disturbances through repeated resolving. The solver employs a sequential quadratic programming algorithm. The gradients of the objective and constraint functions with respect to u are obtained analytically via automatic differentiation along the forward chain of the physical information neural network when the spatiotemporal field reconstruction model is implemented. When implemented using the finite volume method for direct numerical solution or a reduced-order model, the gradients are obtained numerically for each decision variable component using the central difference method. Discrete quantity s stack Instead of participating in continuous optimization, each of the selectable levels is solved as a continuous subproblem, and the solution with the smallest objective function value is taken as the final solution. The convergence criterion is that the change in the objective function between two consecutive iterations is less than 1 × 10⁻⁶. -4 And the constraint violation is less than 1×10 -3The typical number of iterations is 20 to 60, and the time taken for a single solution on a server equipped with a graphics processor is in the seconds range, which can meet the real-time requirements of online control of composite workstations. When the above optimization problem has no feasible solution within the given adjustable range, it indicates that adjusting the controllable quantity alone is not enough to make all positions of the entire roll reach the qualified threshold. At this time, an infeasibility flag is output and turned into an alarm prompt, allowing process personnel to decide whether to downgrade the roll, rather than giving an invalid adjustment value that exceeds the adjustable range.

[0089] Finally, a set of numerical examples are given to illustrate the execution process of the above reverse calculation. After the prediction in step S4, the peel strength of a certain position near the core of a certain roll that has been taken off the production line is predicted to be 3.2 N per 15 mm, which is lower than the qualified threshold of 3.5 N per 15 mm. The peak curing temperature at this position is too high and the final curing conversion rate is abnormal. Based on this, it is determined that the cause is over-curing due to the heat accumulation of the core. Since the roll has been taken off the production line, the components of the composite station domain are fixed according to the constraint condition (4). The above optimization problem is solved only in the curing library domain. The adjustment values ​​are: the air supply temperature is reduced from 48℃ to 46℃, the curing time is extended from 72 hours to 78 hours, the stacking method is changed from close stacking to spaced stacking, and the air supply humidity is kept constant at 50%RH. Substituting the adjusted values ​​back into the forward prediction chain for recalculation, the predicted peel strength at this location increases to 3.6 N / 15 mm, and the minimum predicted value at each location of the entire roll increases from 3.2 to 3.55 N / 15 mm, all of which are not lower than the qualified threshold. At this time, the objective function value is 0.31. For the same specification of in-production roll material, the constraint condition (4) is removed and the solution is recalculated. The adjusted values ​​of the composite station domain are: the adhesive temperature is increased by 1.5℃ and the target value of the amount of adhesive applied is increased from 1.35 to 1.45 g / m² through the mapping relationship between adhesive temperature and viscosity; the two-component metering ratio is adjusted from 100:80 to 100:82; and the roller gap pressure is increased from 48 to 52 N / mm. Based on this, the feedforward offsetting of the foreseeable heterogeneity of the curing space within the roll on the peel strength is achieved in the defect cause stage.

[0090] Thus, the location-resolved prediction of intravolume peeling intensity and the dual-domain inverse control are completed in the composite stage, forming a spatially resolved closed loop of prediction and control.

[0091] To enable those skilled in the art to understand the input data, output data, and their inherent relationships in each stage of the above online execution process, the algorithm steps of the process are as follows.

[0092] Input: Composite station parameter sequence P = {p j}, where j is the index of the roll length location node, p j The feature vector is composed of the aforementioned nine process parameters; the parent roll structure parameter Θ = {R i , R oW, winding tension distribution, winding tension taper τ w}; Boundary condition for the maturation library B = {T air (t), RH air (t), h, s stack};Qualified threshold S th With the preset margin δ.

[0093] Output: Intra-volume position-resolved peel intensity prediction field Its confidence interval; risk region set Ω; controllable adjustment value Δu.

[0094] Step 1 (corresponding to S1): For each length position node j, by p j The four derived features—actual coating viscosity, two-component distribution ratio deviation, actual adhesive content, and dew point temperature—were calculated and concatenated with directly acquired parameters, then normalized to form a thirteen-dimensional feature vector x. j Input the primary soft measurement sub-model to obtain The unit is 15 millimeters per niut.

[0095] Step 2 (corresponding to S2): Using Θ according to the Archimedean spiral winding relationship, the length coordinates are... Mapped to winding geometric coordinates The two-dimensional potential distribution y within the volume is obtained. pot (r, z).

[0096] Step 3 (corresponding to S3): Using Θ as the solution domain structure parameter and B as the initial boundary condition, the temperature field T(r, z, t) and the curing conversion rate field α(r, z, t) are obtained by forward reasoning from the in-roll curing temperature and humidity spatiotemporal field reconstruction model. Time t covers the entire curing period from the roll material being removed from the production line to the end of curing, and the humidity field RH(r, z, t) is determined by the supply air humidity boundary.

[0097] Step 4 (corresponding to the first part of S4): Extract the maturation history feature vector at this position from the above three fields. ,in For equivalent curing heat history, For the final solidification conversion rate, and These are the peak ripening temperature and its arrival time, This represents the cumulative amount of humidity exposure.

[0098] Step 5 (corresponding to the latter part of S4): The lower bound of the confidence interval is given by the output headers of the three quantiles. With the upper realm .

[0099] Step 6 (corresponding to the S5 front end): ,or The positions are also included in this set; for each position within Ω, according to its T... peak With α f Determine the cause, T peak High and α f Abnormalities are classified into the over-ripened core subset, T peak Low and α f Those that are insufficient are classified into the outer layer and ear maturation insufficiency subset.

[0100] Step 7 (corresponding to the end of S5): If Ω is not empty, solve Δu according to the aforementioned minimum adjustment optimization problem and send it to the corresponding actuator via the fieldbus; if Ω is empty, maintain the current process settings and repeat steps 1 to 7 with the newly acquired data in the next control cycle.

[0101] As can be seen from the above algorithm steps, the composite station process parameters are transformed into a position-independent strength baseline in step 1, the in-roll spatial coordinates are obtained in step 2, and the curing history corrections are transformed into position-related values ​​in steps 3 and 4. The two are combined into a position-resolved strength prediction value in step 5, and then transformed into adjustment values ​​for controllable quantities in the composite station domain and the curing library domain in steps 6 and 7. The input and output data of each link are thus linked step by step and closed into a loop.

[0102] The following describes the training process of each soft measurement sub-model and the spatiotemporal field reconstruction model in this embodiment. The primary soft measurement sub-model, the fused soft measurement model, and the physical information neural network are all trained offline based on historical production data. After training, they are solidified and deployed to the online execution process for use in steps S1, S3, and S4.

[0103] For the training dataset, composite workstation parameter sequences, master roll winding and curing chamber structural parameters, temperature and humidity measurement data within the curing chamber, and offline peel strength data sampled from within the roll after curing of each batch were collected from historical batch production records. The data scale ranges from tens to hundreds of batches, with each batch covering thousands of length position nodes. The offline peel strength sampling values ​​were measured by standard peel tests, and the winding radius and axial coordinates corresponding to the sampling positions were labeled. Data preprocessing included removing invalid records from sensor anomalies and downtime sections, resampling the parameter sequences according to the aforementioned length grid, normalizing each feature, and dividing the dataset into training, validation, and test sets by batch, with the preferred ratio being 7:1.5:1.5 to avoid information leakage caused by the same batch of data appearing simultaneously in both the training and test sets.

[0104] The training of the primary soft-sensing sub-model uses the composite workstation parameter sequence and its derived features as input, and the baseline peel strength at the corresponding position under standard curing conditions as the label for supervised learning. The training of the fusion soft-sensing model uses the initial adhesion strength potential and curing history features as input, and the offline peel strength sampling value as the label for supervised learning. Both models are implemented using fully connected neural networks, built on a deep learning framework and trained on a graphics processing unit (GPU). The loss function is the mean squared error, the optimizer is an adaptive moment estimation optimizer, and the initial learning rate is set to 2.3 × 10⁻⁶. -4 The cosine annealing decay strategy is adopted, the batch size is 128, and the number of training rounds is determined by the early stopping criterion that the validation set loss no longer decreases for several consecutive rounds. It usually converges within 200 rounds, and weight decay regularization is applied to suppress overfitting.

[0105] The training of the physical information neural network adopts a step-by-step strategy. Its loss function consists of three parts: the fitting error of the temperature and humidity measurement data of the curing chamber, the physical residual between the thermal conductivity equation and the curing kinetic equation, and the deviation of the initial boundary conditions. These three parts are weighted and summed. The optimizer adopts an adaptive moment estimation optimizer, with an initial learning rate of 1.2 × 10⁻⁶. -3 The training process decays exponentially, with batch size set to 256 sampling points. Training terminates when both the physical residual and data fitting error fall below a preset convergence criterion. Furthermore, the equivalent curing heat history of the corresponding sampling location is calculated from the offline stripping strength sampling values ​​and used as an additional physical soft constraint to reinject the temperature field and curing conversion rate sub-networks, ensuring that the curing field estimation for unmeasured locations within the roll converges consistently with each batch. For evaluation metrics, the root mean square error and mean absolute error are used to measure the deviation between predicted and sampled values ​​for each soft sensor model. The physical consistency of the physical information neural network is measured by the magnitude of the physical residual. The decrease in prediction error based on location within the roll compared to the baseline scheme treating the entire roll as a uniform standard curing process is used as a comparative metric for evaluating location resolution capability. After successful validation on the test set, each trained model is solidified and deployed for use in the aforementioned online execution process.

[0106] Example 2

[0107] In a preferred embodiment of the present invention, a further refined implementation method is provided for the reconstruction of the spatiotemporal field of curing temperature and humidity within the roll. It is based on the coupling of the unsteady heat conduction equation with internal heat source in cylindrical coordinate system and the curing kinetic equation with self-catalytic term. The physical information neural network is used as the solution and calibration carrier of the coupled equation. At the same time, a closed-loop calibration method is introduced to invert the continuous curing field within the roll using sparse offline sampling data. Thus, under the conditions of limited measurement points in the curing library and sparse offline peel strength sampling data, the curing temperature history, humidity history and curing conversion rate of each winding radius and axial position within the roll are reconstructed.

[0108] Specifically, the unsteady-state thermal conductivity equation containing an internal heat source is established in cylindrical coordinates. Its unknowns are the temperature field within the roll, and its independent variables are the winding radius coordinates, axial coordinates, and curing time. The radial and axial thermal conductivity coefficients are taken according to the interlayer thermal conductivity characteristics of the roll material to reflect the anisotropy of heat transfer in the radial and axial directions of the winding structure. The internal heat source term in the equation equals the heat of reaction released per unit volume of adhesive layer per unit time due to the crosslinking reaction of isocyanate and polyol. Its value is equal to the product of the total heat released by the reaction and the rate of change of the curing conversion rate at that location with respect to time. The curing kinetic equation with an autocatalytic term describes the rate of evolution of the curing conversion rate with curing time. This rate is jointly determined by the reaction rate constant, the power term of the current conversion rate, and the autocatalytic term. The autocatalytic term reflects the promoting effect of the generated urethane groups on subsequent crosslinking reactions. The reaction rate constant increases with increasing local temperature according to the Arrhenius relation, and a correction factor that varies with ambient relative humidity is used to account for the influence of moisture on isocyanate side reactions and crosslinking network formation. The thermal conductivity equation and the curing kinetics equation are bidirectionally coupled through local temperature and curing conversion rate. That is, the internal heat source term in the thermal conductivity equation depends on the conversion rate given by the curing kinetics equation, while the reaction rate constant in the curing kinetics equation depends on the local temperature given by the thermal conductivity equation. The two reinforce each other in the heat dissipation-limited core area, forming a positive feedback loop where the temperature rises, accelerates the reaction, and the heat released by the reaction further raises the temperature, thereby accumulating heat near the core.

[0109] Furthermore, a physical information neural network is used to solve and calibrate the aforementioned coupled equations. The physical information neural network consists of two structurally independent subnetworks. The first subnetwork takes the winding radius coordinates, axial coordinates, and curing time as input and the curing temperature at that location and time as output. The second subnetwork takes the same coordinates and time as input and the curing conversion rate at that location and time as output. The two subnetworks do not share hidden layer weights in the forward computation; they are coupled only through the physical residual term at the loss function level. Preferably, each subnetwork employs a five- to eight-layer fully connected structure, with the number of neurons in each layer ranging from sixty-four to two hundred and fifty-six. The hidden layer activation function is a hyperbolic tangent function or a sinusoidal activation function to ensure that the higher-order derivatives of the spatiotemporal field can be analytically obtained. Before entering the first layer, the network's input coordinates are normalized to specified intervals according to the winding radius range and curing time range, respectively, to improve training convergence under different dimensional variables. The two sub-networks adopt a step-by-step training strategy. First, the second sub-network is fixed, and the first sub-network is trained mainly with temperature measurement data and the residual of the heat conduction equation to achieve initial convergence of the temperature field. Then, the first sub-network is fixed, and the second sub-network is trained mainly with the residual of the solidification kinetic equation to achieve convergence of the solidification conversion rate field. Finally, the two sub-networks are unfixed and jointly fine-tuned to make the coupling terms self-consistently transfer between the two fields.

[0110] The loss function of the physical information neural network is a weighted sum of a data fitting term and a physical residual soft constraint term. The data fitting term includes a temperature data term and a sampling data term. The temperature data term is the mean square error between the predicted temperature and the measured temperature at the temperature and humidity measurement points in the curing chamber by the first sub-network. The sampling data term is the equivalent curing heat history constraint constructed in the following inversion calibration step. The physical residual soft constraint term includes a thermal conductivity residual term and a curing residual term. The thermal conductivity residual term is the mean square value of the residual obtained by substituting the outputs of the first and second sub-networks into the thermal conductivity equation containing the internal heat source at the spatiotemporal sampling points. The curing residual term is the mean square value of the residual obtained by substituting the outputs of the two sub-networks into the curing kinetic equation at the spatiotemporal sampling points. It also includes boundary condition residual terms and initial condition residual terms, used to constrain the convective heat transfer boundaries at the outer layer of the roll material and the exposed surfaces at both ends of the roll material, as well as the initial temperature and initial conversion rate at the time of unloading. The loss function is: Where L is the total loss, L temp L is the fitting term for the temperature data. peel For sampling and inversion of data items, L pde,heat L is the residual term of the heat conduction equation. pde,cure L is the residual term of the curing kinetic equation. bc L represents the boundary condition residual term. ic is the initial condition residual term, and w1 to w6 are the weighting coefficients of each term.

[0111] The specific expressions for each residual term are as follows. Let the output of the first subnetwork be T. θ (r, z, t), the output of the second subnetwork is α ψ (r, z, t), where θ and ψ are the trainable parameters of the two sub-networks, respectively:

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119]

[0120] Among them, R heat With R cureThese are the residuals obtained by substituting the outputs of the two subnetworks into the aforementioned cylindrical coordinate heat conduction equation containing the internal heat source and the solidification kinetic equation with the self-catalytic term, respectively; subscript f traverses the spatiotemporal collocation points, k traverses the temperature and humidity measurement point samples in the curing chamber, s traverses the offline peel strength sampling samples, b traverses the boundary collocation points, 0 traverses the initial time collocation points, and N... f N d N s N b N0 and T are the corresponding sample sizes, respectively; k meas The actual temperature measured at the measuring point; For output by the first subnetwork The calculated equivalent aging heat history, H eq,s inv The equivalent ripening heat history is calculated by inverse mapping of the offline peel strength measured at the sampling location using a fusion soft-sensor model. The difference between the two constitutes the aforementioned sampling inversion constraint. The partial derivatives of each order in the above formula... , , , , and All input coordinates are analytically derived using the automatic differentiation of the subnetwork, rather than through numerical difference. This is precisely why the aforementioned hidden layer activation functions must be hyperbolic tangent or sinusoidal activation functions, ensuring the existence and continuity of the second derivative of the network output with respect to the input coordinates, allowing the residuals of the heat conduction equation to be analytically calculated. Preferably, in the initial training phase, the weights of the data fitting terms are set greater than those of the physical residual terms, allowing the network to initially fit the measured data. As the number of training epochs increases, the weights of the physical residual terms are gradually increased, enabling the network to satisfy the natural constraints of heat transfer and reaction while fitting the data. The weights can also be dynamically balanced using adaptive normalization based on the gradient norm of each residual term to prevent any single residual term from dominating the gradient direction during training.

[0121] Furthermore, a closed-loop calibration step is introduced to invert the continuous curing field within the roll using sparse offline sampling data. Since only a limited number of temperature and humidity measurement points can be placed at the roll ends and within the curing chamber, and peel strength can only be obtained by sampling specific locations on the roll material after curing and through destructive testing, the curing history of most locations within the roll is not directly observed. Therefore, the measured peel strength value obtained from each offline sampling is inversely mapped using a fusion soft sensor model to calculate the equivalent curing heat history that the sampling location should undergo to achieve the required peel strength. This equivalent curing heat history is a scalar feature formed by the time integral of the curing temperature history over the curing period. Then, the deviation between the equivalent curing heat history calculated from the outputs of the first and second sub-networks at the sampling location and the aforementioned inversely calculated value is used to construct a sampling data item and back-inject the loss functions of the two sub-networks. Therefore, a small number of destructive sampling checks are not only used to correct the output of the fusion soft measurement model, but also act as additional physical soft constraints to reverse the spatiotemporal field reconstruction itself, so that the estimation of the curing temperature field and the curing conversion rate field at locations where no measurement points are placed within the volume converges self-consistently with batch accumulation. Specifically, after each batch of curing is completed and the sampling results are obtained, the newly added sampling inversion constraints are incorporated into the training set to incrementally fine-tune the physical information neural network. A small learning rate and fewer training rounds are used to avoid forgetting the calibration results of existing batches, thereby forming an in-volume curing field reconstruction capability that is updated continuously with production batches.

[0122] Preferably, the physical information neural network is implemented within a general deep learning framework. Training is performed offline on computing nodes with GPU acceleration. Spatiotemporal sampling points are obtained by sampling points along three dimensions: winding radius, axis, and curing time. The scale of a single batch of sampling points ranges from several thousand to tens of thousands of points. The initial learning rate ranges from one ten-thousandth to one thousandth, coupled with a learning rate decay strategy that decreases with each training epoch. The number of training epochs ranges from several thousand to tens of thousands, and an early stopping mechanism is set up when the validation set loss no longer decreases. Weight decay regularization is introduced into the hidden layers of the two sub-networks to suppress overfitting. After training, the network is solidified and deployed in the online computing unit of the composite workstation. For newly produced rolls, only the winding geometry and curing library boundary conditions need to be input to perform forward inference to obtain the curing temperature history, humidity history, and curing conversion rate at each position within the roll. The computation time for forward inference is much less than the time required for direct numerical solution of coupled partial differential equations, thus balancing physical consistency and real-time online computing.

[0123] Beneficial Effects. Even with limited temperature and humidity measurement points within the curing chamber and sparse offline peel strength sampling data, the above implementation method can still reconstruct the curing temperature history and curing conversion rate at each winding radius and axial position within the roll, improving the spatial continuity and batch consistency of the curing field estimation within the roll. Incorporating the physical residuals of heat transfer and curing kinetics as soft constraints into the loss function ensures that even when historical curing measurement data is insufficient to uniquely determine the continuous field within the roll, the network remains constrained by heat transfer and reaction laws, providing a physically self-consistent estimate and reducing multiple solutions and extrapolation distortion caused by data sparsity. By using the equivalent curing heat history reinjection temperature subnetwork and the curing conversion rate subnetwork obtained from offline sampling peel strength inversion, a small number of destructive samplings exert a reverse constraint on the spatiotemporal field reconstruction. The curing field estimation at unmeasured locations within the roll converges with batch rolling updates, reducing the required sampling sample size for calibration. The two sub-networks are trained step-by-step and structurally independent, ensuring that the convergence processes of the temperature field and the curing conversion rate field do not interfere with each other. Joint fine-tuning then achieves self-consistent transfer of the coupling terms, improving the numerical stability of the training process. After training, forward inference replaces the online numerical solution of the coupled partial differential equations, allowing the intravolume curing field reconstruction to be completed online during the compositing stage, providing real-time input for subsequent position-resolved peel strength prediction.

[0124] Principle Analysis. The aforementioned beneficial effects are achieved because, during the curing storage period, the internal heat transfer of the master roll after winding is constrained by the winding and stacking structure, exhibiting anisotropy in the radial and axial directions. The exothermic reaction from the cross-linking process accumulates in the core region where heat dissipation is limited. This coupling results in a spatially non-uniform distribution of the curing temperature field and the curing conversion rate field within the roll along the winding radius and axial direction. This distribution cannot be described by instantaneous parameters of the composite workstation or by the assumption of uniformity across the entire roll. Incorporating the residuals of the heat conduction equation containing the internal heat source and the curing kinetic equation with the self-catalytic term as soft constraints into the loss function of the physical information neural network is equivalent to filling the observation gaps in the spatiotemporal region where data is insufficient with the physical laws of heat transfer and reaction. This constrains the network output within the solution space that satisfies the relationship between energy conservation and reaction rate, thus preventing field estimates that violate physical laws even with sparse measurement points. The offline sampling peeling intensity is inversely mapped to an equivalent curing heat history using fused soft measurement and then reinjected into the two sub-networks. Essentially, this transforms the sparse but reliable intensity observation of destructive sampling into an indirect constraint on the curing field. Since peeling intensity has a monotonically dependent relationship with the curing heat history, this inversion constraint can provide effective field correction information for in-volume locations lacking direct temperature measurements, thus enabling the estimation of the continuous curing field within the volume to converge consistently with each batch. The two sub-networks are structurally independent and coupled only through physical residuals, allowing the influence of temperature on curing and the influence of curing exothermics on temperature to be gradually transferred through the loss function rather than being forcibly fitted within a single network, avoiding numerical divergence caused by positive feedback coupling in the early stages of training. It should be noted that when the offline sampling locations in a batch are too concentrated within the volume, or when the curing library measurement points do not cover the axial direction sufficiently, the accuracy of the curing field estimation in areas far from the measurement points and sampling locations within the volume may decrease. In such cases, it is advisable to appropriately increase the number of axial measurement points or adjust the in-volume distribution of the sampling locations to maintain the accuracy of field reconstruction.

[0125] Example 3

[0126] In another embodiment of the present invention, an alternative solution is provided that differs from the aforementioned method of reconstructing the spatiotemporal field of intravolume ripening using a physical information neural network. This alternative solution addresses the same technical problem of spatial mismatch in peel strength prediction caused by spatial non-uniformity of intravolume ripening, but adopts a technical approach of directly solving the coupled equations numerically and combining a reduced-order model with data-driven residual correction. This alternative solution can be used to replace the aforementioned implementation method in situations where computational resources are limited or a differentiable solution environment is lacking.

[0127] As an alternative, the unsteady thermal conductivity equation containing an internal heat source and the curing kinetics equation with an autocatalytic term in cylindrical coordinates are discretized and then directly solved numerically using the finite volume method. Specifically, the cross-section of the mother roll is divided into control volume meshes according to the winding radius direction and the axial direction. Unequal spacing meshes are used in the core region and the outer layer region to refine the mesh in the outer layer and the exposed ear area where the heat dissipation gradient is large. For each control volume, the thermal conductivity equation is discretized by energy conservation and the curing kinetics equation is discretized by time progression. The internal heat source term is given at each time step by the product of the current curing conversion rate advancement of the control volume and the total heat release of the reaction. The reaction rate constant is updated according to the current temperature of the control volume through the Arrhenius relation, thereby alternately updating the temperature field and curing conversion rate field in the time step until the entire curing period is covered, obtaining the curing temperature history, humidity history and curing conversion rate of each control volume in the roll. The finite volume method direct numerical solution can give physically conserved field solutions without relying on historical curing data, and its accuracy is determined by the mesh density and time step.

[0128] Furthermore, considering the high computational cost of performing full-time direct numerical solutions for each newly produced roll material and the difficulty in meeting the real-time requirements of online prediction at composite workstations, a reduced-order model is constructed as an alternative to the aforementioned direct numerical solutions. Specifically, during the offline phase, using different core diameters, target roll diameters, winding tension distributions, supply air temperatures, and supply air humidity as samples, the finite volume method is used to perform direct numerical solutions in batches, obtaining a solution set of the in-roll curing field covering typical working conditions. The dominant spatial modes are extracted from this solution set by intrinsic orthogonal decomposition, and the in-roll curing temperature field and curing conversion rate field are approximated by a linear combination of a few dominant modes. The response surface is constructed with boundary conditions and structural parameters as independent variables and the combination coefficients of each dominant mode as dependent variables. In the online phase, the boundary conditions and structural parameters of the newly produced roll material only need to be substituted into the response surface to quickly obtain the combination coefficients of each mode, thereby reconstructing the curing field within the roll. The online calculation time is much less than the direct numerical solution of the coupled equations throughout the entire time period. Thus, the response surface is constructed from the offline solution results and can be quickly called online to meet the real-time requirements.

[0129] Furthermore, to compensate for the discrepancies between the physical property parameters and curing kinetic parameters relied upon by the finite volume method and actual production, data-driven residual correction is applied to the output of the reduced-order model using offline peel strength sampling data from historical batches. Specifically, the in-volume curing field reconstructed by the reduced-order model is mapped to a peel strength prediction value via a fusion soft measurement model. The deviation between this prediction and the offline measured peel strength at the corresponding sampling location is taken as the residual sample. A residual regression model is trained using boundary conditions, structural parameters, and in-volume location as independent variables. The residual regression model employs gradient boosting tree or Gaussian process regression. In the case of Gaussian process regression, the uncertainty of the residual estimate can also be provided. In the online stage, the strength prediction output of the reduced-order model is added to the residual correction amount given by the residual regression model to obtain the corrected in-volume location-resolved peel strength prediction value. Therefore, the physical mechanism solution undertakes the main reconstruction of the spatial distribution of the curing field within the volume, while the data-driven residual correction undertakes the compensation for the systematic modeling bias. The two work together to enable the alternative solution to give a position-resolved peeling intensity prediction within the volume without relying on the differentiable physical information neural network. This prediction can be used to identify local risk areas and adjust the controllable quantities of the composite work station domain and the curing library domain according to the aforementioned threshold judgment and dual-domain reverse calculation process.

[0130] In the above alternative, the winding geometry mapping, threshold determination, and dual-domain reverse calculation and control steps can still use the processing method of the aforementioned implementation method. That is, the length position of the roll is still converted into the winding radius and axial position according to the Archimedes spiral winding relationship. The predicted peel strength value of each position in the roll is still compared with the qualified threshold to identify local risk areas where the predicted value is lower than the qualified threshold or the difference between the predicted value and the qualified threshold is less than the preset margin. It also distinguishes between areas near the core that are over-cured or under-reacted due to heat accumulation and areas near the outer layer and the ears at both ends of the roll that are under-cured due to excessive heat dissipation. The adhesive temperature, two-component metering ratio, and roller pressure are still adjusted in the composite station area. The air supply temperature, air supply humidity, and curing time are adjusted in the curing warehouse area. The winding tension distribution of the roll in production is adjusted during the winding stage to make the predicted peel strength value of each position of the entire roll fall back to the qualified range. The applicability of this alternative solution lies in the fact that when the production conditions exceed the range of boundary conditions and structural parameters covered by the offline solution, the extrapolation of the response surface of the reduced-order model may introduce errors. In this case, it is advisable to supplement the direct numerical solution samples of the corresponding conditions to expand the solution set and reconstruct the response surface. When the historical batch sampling data does not cover a certain type of condition, the compensation capability of data-driven residual correction for that type of condition will decrease accordingly.

[0131] Example 4

[0132] This embodiment provides a system for predicting and controlling the adhesive strength of solvent-free composites, the structure of which is as follows: Figure 4As shown, it includes a parameter acquisition module, a winding geometry mapping module, a curing spatiotemporal field reconstruction module, a peel strength prediction module, a threshold determination module, and a dual-domain control module. Each module corresponds one-to-one with steps S1 to S5 of the method described in Example 1, and sequentially undertakes the functions of parameter acquisition and initial adhesive strength potential output, winding geometry mapping, curing spatiotemporal field reconstruction, peel strength position resolution prediction, threshold determination, and dual-domain reverse control. In the compounding stage, it completes the position resolution prediction and dual-domain reverse control of the peel strength within the roll, forming a spatial resolution closed loop of prediction and control.

[0133] The parameter acquisition module is used to acquire process parameters of the composite process in real time and, according to the acquisition time and the position of the roll length, output the initial bonding strength potential of each roll length position by the pre-trained primary soft measurement sub-model. Specifically, the module collects process parameters, including adhesive temperature and its corresponding viscosity, two-component metering ratio, gluing roller speed ratio, roller gap pressure, lamination speed, substrate surface tension, and ambient temperature and humidity, through sensors arranged at the gluing unit, pressing unit, and winding unit of the solventless laminating machine. It also accumulates the roll length by counting the encoder pulses of the main drive of the unit, assigning a unique roll length position coordinate to each collection moment, thereby obtaining a continuous sequence of lamination station parameters along the roll length direction. The module is further configured to use the actual coating viscosity obtained from the mapping relationship between adhesive temperature and viscosity, the two-component metering ratio deviation, and the actual amount of adhesive calculated from the ratio of gluing roller speed ratio to lamination speed as derived features, and input them together with the directly collected process parameters into the primary soft measurement sub-model, outputting the initial adhesive strength potential of this length position under the standard uniform curing assumption, which is determined only by the coating and pressing quality at the moment of lamination.

[0134] The winding geometry mapping module is used to convert the length position of the roll material into the winding radius and axial position of that section of roll material within the master roll according to the Archimedean spiral winding relationship, and to reconstruct the initial adhesive strength potential into a two-dimensional distribution within the roll, calibrated by the winding radius coordinates and axial coordinates. Specifically, this module is configured to establish a mapping between the length position of the roll material and the winding geometry position based on the core diameter, the thickness of the single-layer composite structure, the winding tension, and the tapered winding tension. It can inversely solve for the winding radius of any length position, and correspond the position of the roll material along the width direction to the axial coordinates of the master roll, thereby reconstructing the initial adhesive strength potential distributed one-dimensionally along the length direction into a two-dimensional distribution within the roll. This module also collects the width of the master roll, the core diameter, the target roll diameter, the winding tension distribution, and the air supply temperature, air supply humidity, hot air circulation mode, and stacking mode of the curing chamber as boundary conditions and structural parameters for subsequent curing time-space field reconstruction.

[0135] The curing spatiotemporal field reconstruction module is used to describe the heat transfer within the roll using an unsteady thermal conduction equation with an internal heat source in cylindrical coordinates. The internal heat source term is determined by the exothermic rate of the crosslinking reaction given by the curing kinetic equation. The thermal conduction equation and the curing kinetic equation are coupled and solved to obtain the curing temperature history and curing conversion rate of each winding radius and axial position within the roll. The humidity history of each position is determined by combining the humidity boundary of the curing chamber air supply. Specifically, this module treats the master roll as an axisymmetric solution domain in cylindrical coordinates, uses a reaction rate model with an autocatalytic term as the curing kinetic equation, and its reaction rate constant varies with local temperature according to the Arrhenius relation and takes into account the influence of ambient humidity on the isocyanate crosslinking reaction. Through the coupling relationship of local temperature increase accelerating the crosslinking reaction and crosslinking reaction exothermicly raising the local temperature, the spatial gradient of the heat-limited core area, where the actual curing temperature is higher than that of the outer layer and the exposed ear area due to heat accumulation, is reconstructed. This module is preferably configured to realize the spatiotemporal field reconstruction using a physical information neural network, and the residual between the cylindrical coordinate heat conduction equation containing the internal heat source and the curing kinetic equation is included as a soft constraint term in the loss function, and two sub-networks corresponding to the temperature field and curing conversion rate are set up for step-by-step training.

[0136] The peel strength prediction module is used to input the initial adhesive strength potential at each location and the curing history features at that location into the fusion soft measurement model, and output the peel strength prediction value after curing for each winding radius and axial position within the roll. Specifically, this module is configured to extract curing history features, including equivalent curing thermal history, final curing conversion rate, peak curing temperature and its arrival time, and cumulative humidity exposure, from the temperature history, humidity history, and curing conversion rate obtained by the curing spatiotemporal field reconstruction module. These features are then concatenated with the initial adhesive strength potential to form a feature vector, which is then input into the fusion soft measurement model. Using the initial adhesive strength potential as a baseline and the curing history features as correction parameters, the model outputs the location-resolved peel strength prediction value and its confidence interval, thereby obtaining the spatial distribution of peel strength covering the entire roll.

[0137] The threshold determination module is used to compare the predicted peel strength values ​​at various locations within the roll with a qualified threshold, and identify local risk areas where the predicted values ​​are lower than the qualified threshold, or where the difference between the predicted values ​​and the qualified threshold is less than a preset margin. Specifically, this module is configured to identify locations where the predicted values ​​are lower than the qualified threshold, or where the predicted values ​​are higher than the qualified threshold but the difference between them and the qualified threshold is less than the confidence interval, as local risk areas. It further distinguishes between areas near the core that are over-cured or under-cured due to heat accumulation, and areas near the outer layer and the ear ends of the roll that are under-cured due to excessive heat dissipation, and outputs a risk area distribution map marked by the winding radius and axial position. The dual-domain control module is used to reverse-calculate and adjust the controllable quantities in the composite work station domain and the curing warehouse domain. It is configured to adjust the adhesive temperature in the composite work station domain and adjust the target value of the amount of adhesive applied in conjunction with the mapping relationship between adhesive temperature and viscosity, adjust the metering ratio of the two components and the roller gap pressure, adjust the air supply temperature, air supply humidity, curing time and stacking method in the curing warehouse domain, and adjust the winding tension distribution and winding tension taper of the roll material in production during the winding stage. With the peel strength prediction value at each position in the roll not lower than the qualified threshold as a constraint and the minimum controllable quantity adjustment range as the objective, an optimization problem is constructed and solved to make the peel strength prediction value at each position of the entire roll fall back to the qualified range.

[0138] The modules sequentially transmit data and form a closed loop in the order of parameter acquisition, winding geometry mapping, curing spatiotemporal field reconstruction, peel strength prediction, threshold determination, and dual-domain control. Specifically, the composite station parameter sequence and initial adhesive strength potential output by the parameter acquisition module flow to the winding geometry mapping module; the two-dimensional distribution within the roll and boundary conditions output by the winding geometry mapping module flow to the curing spatiotemporal field reconstruction module; the curing history characteristics output by the curing spatiotemporal field reconstruction module flow to the peel strength prediction module; the peel strength spatial distribution output by the peel strength prediction module flows to the threshold determination module; the risk area distribution identified by the threshold determination module flows to the dual-domain control module; and the controllable adjustment value calculated by the dual-domain control module is reinjected into the composite station domain and the curing reservoir domain via the field actuator, thereby closing the prediction and control loop. In terms of hardware deployment, the parameter acquisition module can be carried by an edge computing device arranged on the machine to complete data acquisition and initial soft measurement nearby. The curing spatiotemporal field reconstruction module and the peeling strength prediction module involve solving coupled equations and neural network inference, so they can be deployed on a server equipped with a graphics processor to meet the real-time online computing requirements. The controllable adjustment value of the dual-domain control module is sent to the actuators of the composite workstation and the curing library via the fieldbus. The modules are interconnected by an industrial communication network to form a distributed predictive control system.

[0139] Example 5

[0140] This embodiment uses the production of a batch of wide-width solvent-free composite flexible packaging substrates for winding and curing as an example to illustrate the implementation process and technical effects of the method described in Example 1. This batch of master rolls has a large width, a large target roll diameter, and a small core diameter. A two-component polyurethane adhesive is used for solvent-free lamination, followed by winding into master rolls and then transferring them to a curing warehouse for storage and curing. Because the master roll has a large diameter and a small core, the heat dissipation paths along the winding radius vary significantly within the roll. Heat dissipation is difficult near the core due to accumulated interlayer contact thermal resistance, while it dissipates heat faster near the outer layers and the ear-shaped ends of the roll. Furthermore, the exothermic autocatalytic nature of the cross-linking reaction results in significant differences in the actual curing temperature and curing process at different locations within the roll. Simultaneously, the curing process continues in the curing chamber after the roll is produced, making it impossible to measure the peel strength in real time. If conventional soft measurement methods that treat the entire roll as a uniform standard curing process are used, there is a systematic spatial mismatch between the predicted values ​​and the actual values ​​sampled from different locations within the roll. This makes it difficult to identify localized defective areas within the roll before curing is complete. Therefore, the location-resolved prediction and dual-domain control method of this invention is required.

[0141] During implementation, the parameter acquisition module first collects process parameters in real time at the lamination station, including adhesive temperature and its corresponding viscosity, two-component metering ratio, gluing roller speed ratio, roller gap pressure, lamination speed, substrate surface tension, and ambient temperature and humidity. These parameters are then discretized at one node per meter along the roll length corresponding to the encoder pulse. Combined with derived characteristics such as actual coating viscosity, two-component metering ratio deviation, and actual adhesive amount, the initial adhesive strength potential distributed along the roll length is output by the primary soft-sensor sub-model. Subsequently, the length positions are transformed according to the Archimedean spiral winding relationship. The coordinates are calculated as the winding radius and axial coordinates, and reconstructed into a two-dimensional distribution within the roll. Then, the curing spatiotemporal field reconstruction model is coupled with the unsteady thermal conduction equation in cylindrical coordinates containing internal heat sources and the curing kinetic equation with self-catalytic terms to obtain the curing temperature history, humidity history and curing conversion rate of each position within the roll. Then, the fusion soft measurement model uses the initial adhesive strength potential as the baseline and the curing history characteristics as the correction amount to output the predicted peel strength value of each position within the roll after curing. Finally, the threshold judgment module and the dual-domain control module identify local risk areas and reverse calculate the controllable quantity.

[0142] The prediction accuracy was evaluated based on the offline peel strength test results obtained from random sampling within the roll. Conventional soft measurement schemes, which treat the entire roll as a uniform standard curing process, exhibited an average absolute error of approximately 0.42 N / 15 mm and a coefficient of determination of approximately 0.71 for peel strength prediction at roll location resolution. Significant systematic bias was observed near the core and ear area, and the recall rate for identifying locally defective areas within the roll was only about 0.63. After adopting the complete method of this invention, the average absolute error of peel strength prediction at roll location resolution was reduced to approximately 0.18 N / 15 mm, the coefficient of determination increased to approximately 0.93, the recall rate for identifying locally risky areas increased to approximately 0.90, the precision was approximately 0.88, and the F1 score was approximately 0.89. The systematic bias at the core and ear area locations was effectively eliminated. To verify the contribution of each innovative step, an ablation comparison was conducted: When the spatiotemporal field reconstruction of curing in step S3 was removed and the overall roll strength was directly predicted using only composite station parameters, the mean absolute error rose to approximately 0.39 N / 15 mm, and the recall rate dropped to approximately 0.66, indicating that the spatiotemporal field reconstruction of curing temperature and humidity within the roll was key to eliminating spatial mismatch; When the spatiotemporal field reconstruction was retained but empirical interpolation was used instead of the physical information neural network, and no physical residual constraints of heat transfer and curing kinetics were applied, the mean absolute error was approximately 0.27 N / 15 mm, and the coefficient of determination was approximately 0.84, indicating that the consistency of curing field estimation at unmeasured locations within the roll decreased, indicating that the soft constraints of physical residuals contributed to the self-consistent convergence of unmeasured locations; Further, when the baseline plus correction structure in step S4, which used curing history characteristics as the correction amount, was removed, and the initial bond strength potential and curing history characteristics were directly regressed end-to-end, the mean absolute error was approximately 0.24 N / 15 mm, and the stability decreased on small sample batches, indicating that the decoupled structure of baseline plus correction helps to improve interpretability and generalization ability. The above values ​​fluctuate to some extent between different batches. For batches with larger roll diameters and smaller cores, the curing temperature gradient within the roll is steeper, and the differences in error and recall rate between the various schemes are more pronounced.

[0143] like Figure 8 As shown, the horizontal axis, from left to right, represents the conventional whole-roll uniform scheme, reconstruction without S3 ripening field, reconstruction with field retention but without physical residual constraints, end-to-end regression with baseline removal and correction, and the complete method of this invention. The left vertical axis bars represent the mean absolute error, and the right vertical axis broken line represents the coefficient of determination R. 2 With the gradual introduction of ripening spatiotemporal field reconstruction, physical residual constraints, and composite baseline plus ripening correction, the mean absolute error gradually decreased from approximately 0.42 N / 15 mm in the conventional scheme to approximately 0.18 N / 15 mm in the complete method of this invention, and the coefficient of determination R... 2 The increase from approximately 0.71 to approximately 0.93 indicates that the present invention significantly improves the prediction accuracy of intravolume position-resolved peeling intensity through the synergy of ripening spatiotemporal field reconstruction and physical information constraints.

[0144] like Figure 9 As shown, the horizontal axis represents the winding radius r, normalized from the core to the outer layer, and the vertical axis represents the peel strength. Offline sampling shows that the measured values ​​are higher in the core area and gradually decrease along the radius towards the outer layer and the ear area, presenting a true spatial fluctuation. The predicted value of this invention closely follows the measured curve, accurately reproducing the systematic difference between the core and the ear area. In contrast, the predicted value of the conventional whole-roll uniform scheme is close to a horizontal straight line, and deviates significantly from the measured value in the core area and the outer ear area, resulting in spatial mismatch. Accordingly, the conventional scheme misses the risk area below the qualified threshold in the ear area due to overestimation of strength. The predicted value of this invention is consistent with the measured value and can identify the risk area, thereby effectively eliminating the systematic bias in the prediction of peel strength within the roll.

[0145] Regarding the control effect, the threshold judgment module identified an over-cured area near the core and an under-cured area near the ear of one end of the roll in a certain batch of master rolls, and output a risk area distribution map marked by the winding radius and axial position. For the master roll that has been produced, compensatory control was implemented mainly based on the controllable amount of curing in the curing warehouse area. For the over-cured area of ​​the core, the air supply temperature was appropriately reduced and the curing time was shortened. For the under-cured area of ​​the ear, the air supply temperature distribution was adjusted to reduce the curing difference within the roll. The predicted peel strength values ​​at all locations of the entire roll fell back to the qualified range. For the in-production rolls of the same specification, feedforward control was implemented mainly based on the controllable amount of the composite station area. During the composite stage, the adhesive temperature was appropriately increased and the target value of adhesive application was increased through the linkage of adhesive temperature and viscosity mapping. The metering ratio of the two components and the roller gap pressure were finely adjusted to offset the impact of the foreseeable heterogeneity of the curing space within the roll on the peel strength. After the above dual-domain regulation, the consistency of peel strength at different locations within the master roll was significantly improved. The range of peel strength between locations within the roll was reduced from approximately 1.1 N per 15 mm before regulation to approximately 0.4 N per 15 mm. Local unqualified areas were basically eliminated, verifying the feasibility and technical effectiveness of the present invention in completing the positional prediction of peel strength within the roll and dual-domain reverse regulation during the composite stage.

[0146] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting and controlling the adhesive strength of solvent-free composites, characterized in that, include: Step S1: Real-time acquisition of process parameters of the composite process and corresponding acquisition time with the length position of the roll material, outputting the initial bonding strength potential of each roll material length position by the pre-trained primary soft measurement sub-model. Step S2: Convert the length position of the roll material into the winding radius and axial position of the roll material within the mother roll according to the Archimedes spiral winding relationship, and reconstruct the initial bonding strength potential into a two-dimensional distribution within the roll calibrated by the winding radius coordinates and axial coordinates. Step S3: Construct a spatiotemporal field reconstruction model of curing temperature and humidity within the roll. The heat transfer within the roll is described by an unsteady thermal conductivity equation containing an internal heat source in cylindrical coordinates. The internal heat source term is determined by the exothermic rate of the crosslinking reaction given by the curing kinetic equation. The thermal conductivity equation and the curing kinetic equation are coupled and solved to obtain the curing temperature history and curing conversion rate of each winding radius and axial position within the roll. The humidity of the curing chamber air supply is used as the convection boundary, and humidity boundary conditions are applied to the outer surface of the mother roll and the ear end faces at both ends of the mother roll. The relative humidity of each winding radius and axial position within the roll is determined by the curing chamber air supply humidity boundary and the moisture diffusion within the roll, thus obtaining the humidity history of each position. Step S4: Extract the curing history features of the location from the temperature history, humidity history, and curing conversion rate obtained in step S3. The curing history features include the equivalent curing heat history formed by the time integral of the curing temperature history, the final curing conversion rate, the peak curing temperature and its arrival time, and the cumulative amount of humidity exposure. Concatenate the initial bond strength potential of each location with the curing history features of that location into a feature vector and input it into the fusion soft measurement model. The input of the fusion soft measurement model is a six-dimensional feature vector, namely, the initial bond strength potential of the location (one dimension), and the equivalent curing heat history, final curing conversion rate, peak curing temperature, peak curing temperature arrival time, and cumulative amount of humidity exposure (five dimensions of curing history features). Output the predicted peel strength values ​​of each winding radius and axial position within the roll after curing, and obtain the spatial distribution of peel strength covering the entire roll. Step S5: Compare the predicted peel strength values ​​at each location within the roll with the qualified threshold, identify local risk areas where the predicted values ​​are lower than the qualified threshold or the difference between the predicted values ​​and the qualified threshold is less than a preset margin, and reverse calculate and adjust the controllable amount of the composite work station domain and the curing warehouse domain.

2. The method according to claim 1, characterized in that, The process parameters in step S1 include adhesive temperature and its corresponding viscosity, two-component metering ratio, gluing roller speed ratio, roller gap pressure, lamination speed, substrate surface tension, and ambient temperature and humidity. The primary soft measurement sub-model adopts a gradient boosting tree or a fully connected neural network, and uses the actual amount of adhesive calculated from the mapping relationship between adhesive temperature and viscosity, the two-component metering ratio deviation, and the ratio of gluing roller speed ratio to lamination speed as derived feature inputs. The initial bond strength potential is the baseline peel strength determined solely by the coating and pressing quality at the moment of lamination under the standard uniform curing assumption.

3. The method according to claim 1, characterized in that, The curing kinetic equation in step S3 adopts a reaction rate model with an autocatalytic term. Its reaction rate constant varies with local temperature according to the Arrhenius relation, and the influence of ambient humidity on the isocyanate crosslinking reaction is taken into account. The coupling between the thermal conductivity equation and the curing kinetic equation is that the local temperature rise accelerates the crosslinking reaction, and the exothermic crosslinking reaction further raises the local temperature, thereby forming heat accumulation in the core area where heat dissipation is limited.

4. The method according to claim 1, characterized in that, The in-roll curing temperature and humidity spatiotemporal field reconstruction model in step S3 is implemented using a physical information neural network. Specifically, the residual between the cylindrical coordinate heat conduction equation containing the internal heat source and the curing kinetic equation is used as a soft constraint term and incorporated into the loss function of the physical information neural network. The physical information neural network is set up with two sub-networks corresponding to the temperature field and the curing conversion rate, respectively, and is trained step by step. It is calibrated using temperature and humidity measurement data arranged in the curing library and offline peel strength sampling data of historical batches. Furthermore, the equivalent curing heat history of the corresponding sampling location is back-calculated from the offline peel strength sampling value and used as an additional physical soft constraint to back-inject the two sub-networks, so that the curing field estimation of the unmeasured location in the roll converges self-consistently with the batch.

5. The method according to claim 1, characterized in that, The fusion soft measurement model uses the initial adhesive strength potential as a baseline and the maturation history characteristics as a correction factor to output a position-resolved peel strength prediction value and its confidence interval.

6. The method according to claim 1, characterized in that, In step S5, when identifying local risk areas, the causes of risk are distinguished, that is, the areas near the core that are over-cured or under-reacted due to heat accumulation are distinguished, and the areas near the outer layer and the ears at both ends of the roll material that are under-cured due to excessive heat dissipation are distinguished, and a risk area distribution map marked by the winding radius and axial position is output.

7. The method according to claim 1, characterized in that, In step S5, the reverse calculation involves adjusting the adhesive temperature in the composite work station area and adjusting the target value of the amount of adhesive applied in conjunction with the mapping relationship between adhesive temperature and viscosity, adjusting the metering ratio of the two components and the roller gap pressure, adjusting the air supply temperature, air supply humidity, curing time and stacking method in the curing warehouse area, and adjusting the winding tension distribution and winding tension taper of the roll material in production during the winding stage, so that the predicted peel strength value at each position of the entire roll falls back to the qualified range.

8. The method according to claim 7, characterized in that, For roll materials that have already been produced, compensatory regulation is implemented mainly based on the controllable amount in the curing warehouse area. For identified areas where the core is over-cured or the ear is under-cured, the air supply temperature distribution and curing time are adjusted to reduce the curing difference within the roll. For roll materials in production, feedforward regulation is implemented mainly based on the controllable amount in the composite work station area. At the defect cause stage, the coating and pressing parameters are adjusted to offset the impact of foreseeable spatial heterogeneity in the curing space on the peel strength.

9. The method according to claim 7, characterized in that, The reverse calculation is constrained by the fact that the predicted peel strength at each position within the roll is not lower than the qualified threshold, and the objective is to construct and solve an optimization problem with the minimum controllable adjustment range, so as to obtain the controllable adjustment values ​​of the composite work station domain and the maturation library domain.

10. A system for predicting and controlling the adhesive strength of solvent-free composites, characterized in that, It includes a parameter acquisition module, a winding geometry mapping module, a curing spatiotemporal field reconstruction module, a peel strength prediction module, a threshold determination module, and a dual-domain control module; the parameter acquisition module is used to acquire the process parameters of the composite process in real time and, according to the acquisition time and the position of the roll length, output the initial bonding strength potential of each roll length position by a pre-trained primary soft measurement sub-model. The winding geometry mapping module is used to convert the length position of the roll material into the winding radius and axial position of that section of roll material within the master roll according to the Archimedean spiral winding relationship, and to reconstruct the initial adhesive strength potential into a two-dimensional distribution within the roll calibrated by the winding radius coordinates and axial coordinates; the curing spatiotemporal field reconstruction module is used to describe the heat transfer within the roll using an unsteady thermal conductivity equation containing an internal heat source in cylindrical coordinates. The internal heat source term is determined by the exothermic rate of the crosslinking reaction given by the curing kinetics equation. The thermal conductivity equation and the curing kinetics equation are coupled and solved. The curing temperature history and curing conversion rate at each winding radius and axial position within the roll are obtained. Using the curing chamber air supply humidity as the convective boundary, humidity boundary conditions are applied to the outer surface of the mother roll and the ear faces at both ends of the mother roll. The relative humidity at each winding radius and axial position within the roll is determined by the curing chamber air supply humidity boundary and the moisture diffusion within the roll, thus obtaining the humidity history for each position. The peel strength prediction module is used to extract the curing temperature history, humidity history, and curing conversion rate at that position from the temperature history, humidity history, and curing conversion rate obtained by the curing spatiotemporal field reconstruction module. The curing history features include the equivalent curing thermal history formed by the time integral of the curing temperature history, the final curing conversion rate, the peak curing temperature and its arrival time, and the cumulative amount of humidity exposure. The initial adhesive strength potential at each location is concatenated with the curing history features at that location to form a feature vector, which is then input into the fusion soft measurement model. The input of the fusion soft measurement model is a six-dimensional feature vector, namely, the initial adhesive strength potential at that location (one dimension), and the equivalent curing thermal history, final curing conversion rate, peak curing temperature, peak curing temperature arrival time, and cumulative amount of humidity exposure (five dimensions of curing history features). The model outputs the predicted peel strength values ​​after curing for each winding radius and axial position within the roll, obtaining the spatial distribution of peel strength covering the entire roll. The threshold determination module is used to compare the predicted peel strength values ​​at each location within the roll with a qualified threshold, identifying local risk areas where the predicted value is lower than the qualified threshold or the difference between the predicted value and the qualified threshold is less than a preset margin. The dual-domain control module is used to reverse-calculate and adjust the controllable quantities of the composite work station domain and the curing library domain.

Images