Method for predicting warping of plastic plate forming based on multi-dimensional thermal parameters and stress calculation

Abstract

The present application relates to the technical field of plastic extrusion molding, and discloses a plastic plate forming warping prediction method based on multi-dimensional thermal parameters and stress calculation, comprising: receiving a sensing data stream to track material displacement and map it into a particle thermal history set; monitoring a die pressure gradient to add a pollution label; reading the thermal history set to calculate thermal strain and output a residual stress tensor integral to obtain physical characteristics; inputting the physical characteristics and plate parameters into a neural network to generate a two-dimensional deflection grid prediction result; calculating an energy self-consistent residual to perform physical verification, fusing actual deflection errors to generate a reward signal to update an action network, and outputting corrected environmental parameters. Through the combination of physical mechanism and neural network, and the construction of a closed loop based on energy verification and hybrid rewards, the lag problem of quality inspection data is overcome, online adaptive calibration under dynamic working conditions is realized, and the accuracy and physical consistency of warping prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of plastic extrusion molding technology, specifically to a method for predicting warpage during plastic sheet molding based on multidimensional thermal parameters and stress calculations. Background Technology

[0002] During continuous extrusion molding of plastic sheets, macroscopic warping deformation is prone to occur due to non-isothermal cooling and uneven internal stress distribution. To control product quality, the industry currently uses warping prediction methods to guide the molding process. Existing prediction methods are mainly divided into two categories: one is based on rheological and thermodynamic mechanism simulation calculations, which analyzes the relationship between residual stress and geometric deformation by constructing mathematical models; the other is based on machine learning data-driven models, which use sensor data such as temperature and pressure collected on the production line to train neural networks and then directly predict the final surface warping.

[0003] Existing prediction methods face multiple challenges in practical continuous production applications. On the one hand, sensor data from extrusion production lines is continuously collected at fixed spatial locations, while the residual stress of plastic sheets is the result of the accumulation of thermal history along with the material itself. Directly using data from a fixed space makes it difficult to accurately characterize the true thermal history of a specific material cross-section, and occasional abrupt changes in boundary parameters, such as mechanical adjustments during production, can interfere with feature extraction from the model. On the other hand, conventional mechanistic simulation calculations are complex and time-consuming, making it difficult to meet the needs of real-time online prediction. Furthermore, models that rely solely on data-driven approaches, lacking constraints from fluid mechanics and thermodynamics, cannot guarantee the physical consistency of the mapping relationship between stress accumulation and macroscopic warping when faced with complex dynamic operating conditions, easily leading to prediction biases.

[0004] The macroscopic warpage of plastic sheets typically requires accurate measurement data only after the material has completely cooled and solidified and has been transported to downstream quality inspection equipment. This results in a significant time lag in quality inspection feedback. During the operation of the prediction system, the lack of real-time online physical verification indicators makes it difficult to identify and intercept abnormal prediction results that do not conform to physical laws in the early stages. Limited by the delayed feedback of quality inspection data, existing prediction systems can usually only passively output results, making it difficult to effectively utilize multi-source error feedback to perform closed-loop correction of the calculation boundary conditions. This leads to environmental parameter drift after long-term continuous operation, making it difficult to maintain stable prediction accuracy. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting molding warpage of plastic sheets based on multidimensional thermal parameters and stress calculations. This method solves the problems of existing molding warpage prediction methods in plastic sheet extrusion production, which are unable to simultaneously handle dynamic changes in operating conditions and predict physical consistency, rely solely on data-driven methods without physical constraints that easily lead to prediction deviations, lack online real-time adjustment capabilities in conventional mechanism simulations, lack online physical verification methods, and have difficulty using quality inspection data with time lags for system closed-loop calibration.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculations, comprising the following steps: S10 receives the sensor data stream from the continuous extrusion production line of plastic sheet, establishes a ring buffer and a first-in-first-out queue, tracks the physical displacement of a specific material cross section on the production line based on the integral result of the traction line velocity, and maps and combines discrete observation data collected from different spatial locations into a Lagrange point thermal history set of the same virtual material cross section.

[0007] S20 synchronously monitors the temporal gradient change of the melt pressure signal in the die head. When the pressure temporal gradient triggers the mechanical intervention threshold, it adds an artificial intervention contamination label to the Lagrange point thermal history set of the current mapping combination.

[0008] S30, read the Lagrange point thermal history set, use the time-temperature equivalent equation and rheological model to calculate the thermal strain of the point section during the non-isothermal cooling and shaping process, output the integral value of the equivalent residual stress deviator tensor through discrete integration, and retain the calculation result of the three-dimensional residual stress tensor corresponding to the integral value of the equivalent residual stress deviator tensor to obtain the physical feature vector characterizing the cumulative state of residual stress inside the plastic sheet.

[0009] S40, input the physical feature vector and the basic parameters of the board into the deep feedforward neural network, perform forward propagation, output a discrete two-dimensional deflection grid node matrix, and generate the prediction result of plastic board molding warpage.

[0010] S50, based on the three-dimensional residual stress tensor, the thermal stress release potential energy at the microscale is solved. The central finite difference method is used to perform spatial second-order differentiation on the two-dimensional deflection grid node matrix to solve the geometric deformation energy at the macroscale. The difference between the thermal stress release potential energy and the geometric deformation energy is calculated, and an energy self-consistency residual is generated for physical consistency verification. The energy self-consistency residual is extracted as an instantaneous dense reward signal. Combined with the delayed sparse reward signal generated by the actual deflection distribution error sent by the quality inspection equipment, and after removing abnormal samples according to the label filtering rules, the action network is updated using a reinforcement learning algorithm. The corrected boundary environment parameters are output and the corrected thermodynamic deduction boundary conditions are fed back.

[0011] Preferably, tracking the physical displacement of a specific material cross-section on the production line includes: A circular buffer is constructed by allocating contiguous memory space for each sensor, and Euler observation data with absolute timestamps are continuously written into the corresponding circular buffer.

[0012] The cumulative physical displacement of the virtual material cross-section is calculated by taking the moment when the virtual material cross-section leaves the extruder die as the starting point of integration, and by compensating for the traction linear velocity by combining the empirical shrinkage rate parameter of the material.

[0013] The cumulative physical displacement is compared with the physical distance coordinates of each sensor. When the difference between the two falls within the preset distance tolerance range, the arrival time is recorded. The observation data pair with the smallest difference between the timestamp and the arrival time is extracted from the corresponding circular buffer and pushed into the first-in-first-out queue.

[0014] Preferably, the monitoring of the time gradient change of the melt pressure signal at the die head and the addition of a tag includes: The first-order time derivative of the melt pressure in the die head is calculated using forward differential logic.

[0015] The absolute value of the first-order time derivative is compared with a preset mechanical intervention safety threshold, the value of which is determined based on the standard deviation of the pressure derivative distribution within a historical normal production cycle.

[0016] If the absolute value of the first-order time derivative is greater than the mechanical intervention safety threshold, locate the first-in-first-out queue that has not yet completed assembly at the current moment, set the reserved Boolean state position inside the queue to true, and generate the artificial intervention pollution label.

[0017] Preferably, the calculation of thermal strain using the time-temperature equivalent equation includes: Separate the non-isothermal cooling temperature time series recorded within the Lagrange point thermal history set.

[0018] The preset Williams-Landel-Ferry equation is invoked to calculate the time-temperature equivalent shift factor based on the transient temperature and the set material reference temperature.

[0019] Numerical integration is performed along the physical time axis on the reciprocal of the time-temperature equivalent shift factor to convert the physical time coordinates into reduced time coordinates at the reference temperature, thus mapping the non-isothermal cooling boundary to an isothermal relaxation process at a constant reference temperature.

[0020] Preferably, the step of outputting the integral value of the equivalent residual stress deviatoric tensor through discrete integration includes: The stress relaxation modulus of the plastic sheet material decays with time using the generalized Maxwell viscoelastic rheological model.

[0021] The pure thermal shrinkage strain increment is calculated based on the linear expansion coefficient of the material and the temperature gradient. The pure thermal shrinkage strain increment is then substituted into the linear viscoelastic genetic integral equation, and discrete integral calculation is performed along the reduced time axis to obtain the three-dimensional residual stress tensor.

[0022] A hydrostatic stress tensor is constructed based on the average value of the three normal stress components in the three-dimensional residual stress tensor. The hydrostatic stress tensor is then subtracted from the three-dimensional residual stress tensor, and the deviatoric stress tensor is extracted.

[0023] The scalar result of the second invariant of the deviatoric stress tensor is calculated to obtain the integral value of the equivalent residual stress deviatoric tensor.

[0024] Preferably, before inputting the physical feature vector and the basic parameters of the board material into the deep feedforward neural network, the method further includes: Obtain the system's preset basic parameters for the board material, including the design thickness, nominal width, and room temperature elastic modulus of the plastic board.

[0025] The range transformation method is used to perform normalization preprocessing on the integral value of the equivalent residual stress deviator tensor and the basic parameters of the plate material, with a mechanism to prevent the decimalization of the bias constant.

[0026] The normalized data are concatenated and combined to construct a one-dimensional high-dimensional input feature tensor with a uniform numerical scale.

[0027] Preferably, the number of nodes in the output layer of the deep feedforward neural network matches the total number of grid nodes in the discretized plastic sheet surface. After performing forward propagation, the one-dimensional prediction column vector generated by the output layer is reshaped into the two-dimensional deflection grid node matrix according to the preset topology dimensions. The internal element values ​​of the two-dimensional deflection grid node matrix represent the local deflection prediction value of the corresponding physical coordinate point relative to the standard horizontal plane.

[0028] Preferably, the deep feedforward neural network is trained offline using mean squared error as the loss function before operation. The mean squared error is calculated based on the actual deflection label values ​​of the grid nodes of the molded plastic sheet obtained by the quality inspection equipment using a 3D laser scanner and the corresponding node deflection prediction values ​​output by the network.

[0029] Preferably, the solution of the thermal stress release potential energy at the microscale includes: Based on the normal stress components extracted from the three-dimensional residual stress tensor, and combined with the room temperature elastic modulus and Poisson's ratio of the material, the elastic strain energy density inside each discrete micro-element is solved.

[0030] The micro-cumulative elastic strain energy is calculated by performing discrete volume integral accumulation along the material computation domain on the elastic strain energy density.

[0031] Preferably, the method for solving the geometric deformation energy at the macroscopic scale includes: Extract the physical spacing between grid nodes in the two-dimensional deflection grid node matrix.

[0032] The local curvature is obtained by calculating the second-order partial derivatives at each discrete coordinate point using the central finite difference method.

[0033] By substituting the local curvature and the bending stiffness parameters of the plastic sheet into the thin plate bending energy equation, the macroscopic bending strain energy of the thin plate can be obtained.

[0034] Preferably, the process of generating and verifying the energy self-consistency residual includes: The Euclidean distance between the additional bias constant is calculated based on the micro-accumulated elastic strain energy and the macro-thin plate bending strain energy, generating a dimensionless energy self-consistency residual index.

[0035] When the energy self-consistency residual index is less than or equal to the confidence tolerance threshold, the current prediction result is determined to be physically consistent, and the prediction result is output.

[0036] When the energy self-consistency residual index is greater than the confidence tolerance threshold and less than or equal to the outlier isolation threshold, the batch of samples is retained as low-confidence samples and the output of the prediction results is intercepted.

[0037] When the energy self-consistency residual index is greater than the outlier isolation threshold, a physical consistency anomaly label is generated, and the current batch is transferred to the specific isolation sample library.

[0038] Preferably, the step of removing samples based on label filtering rules includes: Read the tag variables carried in the data queue, including the human intervention pollution tag, the equipment status abnormality tag, and the physical consistency abnormality tag.

[0039] Batch data carrying the aforementioned artificial intervention pollution label, equipment status abnormality label, or physical consistency abnormality label will be removed from the training data stream and temporarily stored in a specific isolation sample library, and will not participate in online reward calculation and gradient update of the action network.

[0040] Preferred, the delayed processing mechanism for the reward signal includes: Configure a corresponding batch identifier for each virtual material section.

[0041] When the actual deflection distribution data of the corresponding batch has not yet been returned by the quality inspection equipment, the instantaneous dense reward signal is generated based on the energy self-consistency residual, and the state, action vector and prediction result of the current batch are written into the delayed reward buffer.

[0042] After the quality inspection equipment returns the actual deflection distribution data with the same batch identifier, the mean square error of each node is calculated to generate the delayed sparse reward signal, and the delayed sparse reward signal is backfilled into the delayed reward buffer of the corresponding batch.

[0043] Preferably, a hybrid signal fusing the immediate dense reward and the delayed sparse reward is constructed using the following formula: ; In the formula, The reward signal value is a mixture. and These represent the microscopic cumulative elastic strain energy and macroscopic thin plate bending strain energy calculated at the mechanistic layer for current high-quality samples, respectively. The minimum bias constant is set to avoid the computational dead zone when the denominator approaches zero. This represents the total number of nodes in the deflection grid. and The first The actual deflection label value of each node and the network prediction value; This refers to the weighting coefficient for dense rewards; For sparse reward weighting coefficients, , The sum of the two is one.

[0044] Preferably, the reinforcement learning algorithm utilizes a policy network as the action network, and the output layer of the policy network employs a hyperbolic tangent activation function to constrain the boundary range of the output actions. The forward propagation output of the policy network includes the convective heat transfer coefficient correction for each independent cooling section of the continuous extrusion production line and the overall traction speed compensation value.

[0045] Preferably, the reinforcement learning algorithm employs a near-end policy optimization algorithm, which performs short-cycle updates in conjunction with the immediate dense reward signal before the actual deflection distribution data is returned, and performs supplementary updates in conjunction with the mixed reward signal after the actual deflection distribution data is returned, calculates the network gradient, and updates the network weights.

[0046] Preferably, the boundary conditions in the modified thermodynamic deduction process include: The correction amount for the convective heat transfer coefficient and the compensation value for the traction speed are subjected to amplitude limiting and smoothing processing.

[0047] The corresponding environmental input parameters in the linear viscoelastic genetic integral equation are replaced with the processed correction values, and the traction speed integral tracking variables in the physical displacement calculation process are updated with the processed traction speed compensation values.

[0048] A second aspect of the present invention provides a plastic sheet molding warpage prediction system based on multidimensional thermal parameters and stress calculations, which applies the above-mentioned prediction method and includes: The data acquisition nodes are located from the extruder die head to the cooling and shaping traction section, and are used to collect multi-dimensional thermal parameters and traction motion parameters.

[0049] The controller communicates with the acquisition nodes and is used to receive and transmit parameter signals.

[0050] Quality inspection equipment, installed on the production line, is used to measure the macroscopic dimensions and actual deflection distribution of the molded plastic sheets.

[0051] The server communicates with the controller and quality inspection equipment to reassemble sensor data, extract stress features, generate warpage prediction results through neural networks, perform physical consistency verification, and perform online closed-loop correction of predicted boundary conditions through an action network.

[0052] This invention provides a method for predicting warpage during plastic sheet molding based on multidimensional thermal parameters and stress calculations. It offers the following advantages: 1. This invention constructs a ring-shaped buffer and combines it with traction linear velocity to track the physical displacement of the material cross-section, reconstructing discrete sensor data at fixed locations into a Lagrangian thermal history set of particles following the material. Simultaneously, it monitors the temporal gradient of the die head melt pressure signal to attach contamination tags. This method accurately matches data with physical location and effectively identifies unrecorded mechanical adjustment operations during production, avoiding interference from abrupt changes in boundary parameters on feature extraction and improving the reliability of model input data.

[0053] 2. This invention utilizes the time-temperature equivalent equation and a rheological model to calculate the thermal strain during non-isothermal cooling, extracts the residual stress deviator tensor integral value as a physical feature, and combines it with the input of the plate's basic parameters into a deep feedforward neural network to generate deflection prediction results. This approach, combining physical mechanics mechanisms with neural networks, ensures that the data-driven model is constrained by physical laws during operation, solving the problem of prediction bias that easily arises when a single model faces changes in operating conditions, and improving the accuracy of macroscopic warping deformation prediction.

[0054] 3. This invention generates an energy self-consistency residual by calculating the microscopic elastic strain energy and the macroscopic thin plate bending strain energy, then fuses this residual with the lagged actual deflection error to construct a hybrid reward signal, and uses a reinforcement learning algorithm to update the action network output correction parameters online. This mechanism utilizes the residual to achieve real-time verification of the physical consistency of the prediction results and interception of abnormal samples, while overcoming the feedback delay caused by the time lag of quality inspection data, and realizing adaptive closed-loop calibration of environmental boundary parameters in continuous production. Attached Figure Description

[0055] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a schematic diagram illustrating the sensor data reconstruction and anomaly monitoring principle of the present invention. Figure 4 This is a schematic diagram illustrating the principle of thermal history calculation and stress feature extraction for the mechanism layer in this invention. Figure 5 This is a schematic diagram of the feedforward neural network prediction principle of the present invention; Figure 6 This is a schematic diagram illustrating the principle of cross-scale virtual and real elastic strain energy self-consistency verification in this invention. Figure 7 This is a schematic diagram of the closed-loop calibration principle of the hybrid reward-driven and sample-isolated method of the present invention. Figure 8 This is a three-dimensional waterfall plot of the spatiotemporal evolution of the predicted residuals in this invention; Figure 9 This is a three-dimensional topological diagram of residual stress release and macroscopic deformation reconstruction according to the present invention.

[0056] Among them, 110 is the acquisition node; 111 is the temperature sensor; 112 is the pressure sensor; 113 is the photoelectric encoder; 120 is the controller; 130 is the server; 140 is the quality inspection equipment; 210 is the reconstruction module; 220 is the extraction module; 230 is the prediction module; 240 is the verification module; and 250 is the calibration module. Detailed Implementation

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] See Figure 1 , Figure 1 This is a schematic diagram of the system architecture according to an embodiment of the present invention. The present invention provides a plastic sheet molding warpage prediction system based on multi-dimensional thermal parameters and stress calculation. This system relies on the underlying hardware environment of a continuous extrusion production line and is used to reconstruct and calculate the evolution of residual stress inside the sheet during the extrusion, cooling, and traction molding processes, and to predict the macroscopic warpage morphology of the molded plastic sheet. The system includes: a data acquisition node 110, a controller 120, a server 130, and quality inspection equipment 140.

[0059] The data acquisition node 110 is deployed along the line from the extruder die head to the cooling and shaping traction section, and is used to collect multi-dimensional thermal parameters and traction motion parameters during the plastic sheet molding process. The data acquisition node 110 includes a temperature sensor 111, a pressure sensor 112, and a photoelectric encoder 113 mounted on the traction roller. The temperature sensor 111 includes an ambient temperature sensor and a sheet surface temperature sensor, used to collect the production line ambient temperature and the plastic sheet surface temperature, respectively; the pressure sensor 112 includes a cooling water pressure sensor and a die head melt pressure sensor, used to collect the cooling water pressure and the die head melt pressure, respectively; the photoelectric encoder 113 is used to collect the traction line speed signal.

[0060] The controller 120 is communicatively connected to the acquisition node 110 to receive physical signals collected by various sensors and transmit these signals to the server 130 via an industrial Ethernet network. The server 130 establishes a communication connection with the controller 120 through a standard data interface to run a plastic sheet molding warpage prediction and calibration program. The quality inspection equipment 140 is located at the end of the production line or at an offline inspection station to measure the macroscopic dimensions and actual deflection distribution of the molded plastic sheet and send the actual deflection distribution data to the server 130.

[0061] The plastic sheet molding warpage prediction and calibration program running inside server 130 is divided into a reassembly module 210, an extraction module 220, a prediction module 230, a verification module 240, and a calibration module 250 at the logic layer.

[0062] The reassembly module 210 continuously receives the sensor data stream forwarded by the controller 120 and establishes a circular buffer and a first-in-first-out queue in memory. Based on the integral result of the traction line velocity, the reassembly module 210 tracks the physical displacement of a specific material cross-section on the production line, mapping and combining discrete observation data collected from different spatial locations into a Lagrange point thermal history set of the same virtual material cross-section, thereby reconstructing the non-isothermal cooling thermal history experienced by the same material cross-section during the plastic sheet molding process. Simultaneously, the reassembly module 210 calculates the first-order time derivative of the die head melt pressure signal in real time. When the first-order time derivative of the pressure exceeds a preset mechanical operation safety threshold, it determines that physical intervention such as manual screen replacement or manual adjustment of the die head bolts has occurred on the production line, and adds a manual intervention contamination label to the Lagrange point thermal history set established during that time period.

[0063] Extraction module 220 is connected to reconstruction module 210. It reads the Lagrange particle thermal history set and, using the time-temperature equivalent equation and rheological model, performs time-scale discrete integral calculations on the thermal strain of the particle cross-section during non-isothermal cooling and shaping, obtaining the equivalent residual stress partial tensor integral value. This equivalent residual stress partial tensor integral value characterizes the residual stress state inside the plastic sheet and serves as the physical feature vector for predicting plastic sheet molding warpage. Simultaneously, extraction module 220 retains the three-dimensional residual stress tensor calculation results corresponding to the equivalent residual stress partial tensor integral value and establishes an index association between the two for subsequent energy self-consistency verification by verification module 240.

[0064] The prediction module 230 is connected to the extraction module 220 and is used to obtain physical feature vectors and basic parameters of the sheet material, and input them into a preset deep feedforward neural network. The prediction module 230 performs forward propagation of the neural network and outputs a discrete two-dimensional deflection grid node matrix. The two-dimensional deflection grid node matrix contains the predicted deflection values ​​at each discrete coordinate point of the plastic sheet, which is used to characterize the macroscopic warpage prediction morphology of the molded plastic sheet.

[0065] The verification module 240 is connected to both the extraction module 220 and the prediction module 230, and is used to verify the energy self-consistency of the predicted warpage of the plastic sheet molding. The verification module 240 calculates the thermal stress release potential energy at the microscale based on the three-dimensional residual stress tensor indexed and associated with the integral value of the equivalent residual stress partial tensor, and the corresponding integral value of the equivalent residual stress partial tensor. It then uses the central finite difference method to perform spatial second-order differentiation on the two-dimensional deflection mesh node matrix to solve for the geometric deformation energy at the macroscale. The verification module 240 calculates the difference between the thermal stress release potential energy and the geometric deformation energy, generating an energy self-consistency residual.

[0066] The calibration module 250 is connected to the verification module 240, the quality inspection device 140, and the extraction module 220, and embeds a reinforcement learning proxy algorithm. The calibration module 250 extracts the energy self-consistency residual into an immediate dense reward signal and extracts the actual deflection distribution error provided by the quality inspection device 140 into a delayed sparse reward signal. When the actual deflection distribution data for the corresponding batch has not yet been returned, the calibration module 250 writes the state, action vector, and prediction result of the corresponding batch into the delayed reward buffer, and generates the delayed sparse reward signal upon receiving subsequent actual deflection distribution data with the same batch identifier.

[0067] Before calculating the reward signal, the calibration module 250 initiates a label filtering mechanism to proactively intercept and remove sample data carrying labels indicating human intervention or abnormal equipment status. It also isolates extreme anomalous samples where the energy self-consistency residual exceeds the outlier isolation threshold. This isolates abnormal physical boundary mutations caused by unrecorded human intervention, preventing reward misjudgments by the reinforcement learning algorithm. Based on the filtered reward signal, the calibration module 250 updates its internal action network, adjusts and outputs the corrected boundary environment parameters to the extraction module 220, and updates the boundary conditions in the thermodynamic deduction process after amplitude limiting and smoothing, improving the accuracy and stability of the plastic sheet molding warpage prediction results.

[0068] See Figure 2 , Figure 2 This is a flowchart of a method according to an embodiment of the present invention. The present invention also provides a method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculations, applied to the aforementioned plastic sheet molding warpage prediction system. The method includes the following steps: S10 receives sensor data streams from the continuous extrusion production line of plastic sheet, establishes a ring buffer and a first-in-first-out queue, tracks the physical displacement of a specific material cross-section on the production line based on the integral result of the traction line velocity, maps and combines discrete observation data collected from different spatial locations into a Lagrange point thermal history set of the same virtual material cross-section, so as to reconstruct the non-isothermal cooling thermal history of the same material cross-section during the molding process of plastic sheet; synchronously monitors the time gradient change of the die head melt pressure signal, and when the pressure time gradient triggers the mechanical intervention threshold, adds a manual intervention contamination label to the currently mapped and combined Lagrange point thermal history set; S20 reads the Lagrange particle thermal history set, uses the time-temperature equivalent equation and rheological model to calculate the thermal strain of the particle cross section during the non-isothermal cooling and shaping process, and outputs the integral value of the equivalent residual stress partial tensor through discrete integration. At the same time, it retains the calculation results of the three-dimensional residual stress tensor corresponding to the integral value of the equivalent residual stress partial tensor to obtain the physical feature vector used to characterize the cumulative state of residual stress inside the plastic sheet, and provides the stress tensor data basis for subsequent calculation of thermal stress release potential energy. S30: Input the integral value of the equivalent residual stress deviator tensor and the basic parameters of the sheet into the deep feedforward neural network, perform forward propagation of the neural network, and output a discrete two-dimensional deflection grid node matrix. The two-dimensional deflection grid node matrix is ​​used to characterize the deflection prediction value of the molded plastic sheet at different coordinate positions, thereby generating the prediction result of the plastic sheet molding warpage. S40, based on the three-dimensional residual stress tensor and its corresponding equivalent residual stress partial tensor integral value, solves the thermal stress release potential energy at the microscale, and uses the central finite difference method to perform spatial second derivative on the two-dimensional deflection grid node matrix to solve the geometric deformation energy at the macroscale, calculates the difference between the thermal stress release potential energy and the geometric deformation energy, and generates the energy self-consistency residual to verify the physical consistency of the plastic sheet molding warpage prediction results. S50: Extract the energy self-consistency residual as the instantaneous dense reward signal. Receive the actual deflection distribution data sent by the quality inspection equipment 140. Match the actual deflection distribution data with the corresponding two-dimensional deflection grid node matrix according to the batch identifier of the virtual material cross section to calculate the actual deflection distribution error. Use the actual deflection distribution error as the delayed sparse reward signal. When the actual deflection distribution data of the corresponding batch has not yet returned, write the batch identifier, status, action vector, action network output information, and prediction results of that batch into the delayed reward buffer. After the actual deflection distribution data returns, calculate the actual deflection distribution error and generate the delayed sparse reward signal. Before fusing the reward signal, remove queue samples with attached artificial intervention contamination tags or abnormal equipment status tags according to the tag filtering rules. Isolate extreme abnormal samples whose energy self-consistency residual exceeds the outlier isolation threshold to shield feature contamination caused by human operation and severely distorted physical boundary samples. The action network inside the calibration module 250 is updated based on the filtered reward signal using a reinforcement learning algorithm. The corrected boundary environment parameters are output to the extraction module 220. After the corrected boundary environment parameters are subjected to amplitude limiting and smoothing processing, the boundary conditions in the thermodynamic deduction process are corrected and used for subsequent calculation of residual stress characteristics of plastic sheet and prediction of two-dimensional deflection grid node matrix, thereby realizing online calibration and accuracy improvement of the prediction results of plastic sheet molding warpage.

[0069] See Figure 3 , Figure 3 This is a schematic diagram of sensor data reconstruction and anomaly monitoring according to an embodiment of the present invention. To achieve multi-dimensional sensor data reconstruction and anomaly intervention monitoring, in this embodiment, S10 is further refined into the following specific sub-steps. In an actual extrusion production environment, since the sensors are spatially fixed, the information they collect constitutes typical Euler observation coordinate system data. To accurately track the stress evolution of the material during the molding process, it is necessary to convert the observation values ​​at fixed points into Lagrange point thermal history data that follows the movement of the material.

[0070] S110 is configured with a high-frequency circular buffer and writes Euler observation data with absolute timestamps.

[0071] The reorganization module 210 allocates independent, contiguous memory space in the server's memory for each sensor included in the acquisition node 110, thereby constructing multiple circular buffers. The capacity of each circular buffer is set to an integer multiple of the product of the corresponding sensor's sampling frequency and the production line's maximum tracking cycle, to prevent unread physical data from being prematurely overwritten by new data.

[0072] Temperature sensor 111 and pressure sensor 112 continuously acquire temperature and pressure physical values ​​at fixed locations on the production line. When receiving these physical values, controller 120 timestamps the values ​​uploaded by each sensor based on a unified system clock or industrial Ethernet clock synchronization mechanism, generating an absolute timestamp corresponding to the current system time. The observation data pairs, containing the absolute timestamps and physical values, are then continuously written into the corresponding circular buffers in sequence. The pointer loop overwrite and overflow control mechanisms within the circular buffers can be implemented using low-level methods based on standard data structures, and will not be elaborated upon here.

[0073] S120 tracks physical displacement based on the integral of the traction linear velocity and retrieves the corresponding timestamp to generate a Lagrange point thermal history set.

[0074] Thermodynamic stress accumulation in polymer materials depends on the thermal history experienced by a specific cross-section. Data spliced ​​from different spatial points at the same moment cannot accurately reflect the material's state changes. Therefore, a data queue with responsive characteristics needs to be reconstructed. The reconstructing module 210 discretizes the continuously flowing material into multiple independent virtual material cross-sections at fixed time intervals. The photoelectric encoder 113 acquires the rotational speed of the variable frequency traction roller and calculates the traction linear velocity signal by combining it with the physical circumference of the traction roller. Recombination module 210 as the first The moment when a virtual material cross-section leaves the extruder die. Using the starting point of integration, calculate the virtual material cross-section at time t using the following displacement formula. cumulative physical displacement : ; In the formula, For integration variables The velocity of the traction line at any given time. Calculation. The empirical shrinkage rate parameter of the material is introduced to compensate for the speed reduction, thereby improving the accuracy of physical displacement tracking.

[0075] The extrusion production line is arranged sequentially along the material flow direction. There are 1 sensor node, and the physical distance coordinates between it and the mold opening are denoted as . ,in The value range is 1 to , The specific value is determined by the total number of sensors actually installed on site. During system operation, continuous comparisons are performed. Physical distance coordinates to each sensor When the difference between the two falls within the preset distance tolerance range, the virtual material cross-section is determined to have reached the [number missing]. The location of each sensor is recorded, along with its corresponding arrival time. .

[0076] By calling the memory query instruction, the program in the... Perform a retrieval within the circular buffer corresponding to each sensor, extracting the timestamp parameter and arrival time. The observation data pairs with the smallest difference and within the allowable tolerance range are selected. The reorganization module 210 establishes an independent FIFO queue for each virtual material section. The retrieved observation data pairs are sequentially pushed into the FIFO queue according to the spatial topology of the sensor deployment. As the virtual material section moves towards the end of the production line, the corresponding FIFO queue is gradually filled with data, ultimately combining to generate a Lagrange mass thermal history set matrix for a specific material section. In the formula, Representing the The spatial identifier parameter of a virtual material cross section is determined by the number of discrete particles and the length of the time series. The matrix records the time series experienced by a single material cross section throughout the extrusion and cooling shaping stages, as well as the corresponding temperature and pressure distribution evolution.

[0077] S130 calculates the first time derivative of the melt pressure signal at the die head, determines abnormalities based on the mechanical intervention threshold, and adds a contamination label.

[0078] In the production environment, operators' localized mechanical adjustments to equipment often lack electronic recording by the control system. Changes in physical boundaries can easily be misjudged by the system as abnormal material properties, thus contaminating the feature extraction of subsequent models. To avoid evaluation dead zones in the algorithm logic, this embodiment introduces a monitoring mechanism based on rheological properties. The recombination module 210 extracts the time series of the die melt pressure from the buffer corresponding to the pressure sensor 112. The first-order time derivative of the melt pressure at the die head is calculated using forward differential logic. The formula for calculating the derivative is as follows: ; In the formula, This represents the fixed time interval between two adjacent data samples set by the controller 120; The melt pressure value collected at the current moment. This is the melt pressure value collected at the previous sampling time.

[0079] The reassembly module 210 has a pre-set mechanical intervention safety threshold. The threshold represents the maximum allowable pressure fluctuation gradient in a production line under stable closed-loop control. Safety threshold. The value is defined as three to five times the standard deviation of the pressure derivative distribution within a historical normal production cycle. The system executes numerical comparison logic to determine the absolute value of the first-order time derivative. Is it greater than the safety threshold? .

[0080] Manually replacing filters online or manually tightening die lip bolts during production alters the physical resistance of the polymer melt flow channel, causing sudden changes in melt pressure. When the absolute value of the first-order time derivative... Exceeding the safety threshold for mechanical intervention At this point, the system determines that an unrecorded human intervention event has occurred on the production line. The reassembly module 210 then locates all unassembled FIFO queues within the space between the extruder die and the cooling / shaping traction section, setting the reserved Boolean state positions within these queues to true. This adds a human intervention contamination label to the Lagrange point thermal history set generated during this time period. The data structure with the contamination label clearly identifies distorted thermodynamic boundary parameters in the batch data, ensuring the purity of the data source and providing a fundamental basis for sample isolation in the subsequent reinforcement learning module.

[0081] See Figure 4 , Figure 4 This is a schematic diagram illustrating the principle of thermal history calculation and stress feature extraction for the mechanistic layer according to an embodiment of the present invention. In this embodiment, S20 of the thermal history calculation and stress feature extraction for the mechanistic layer is further refined into the following specific sub-steps. Under actual extrusion molding conditions, the plastic sheet undergoes non-isothermal cooling and viscoelastic state transition during the cooling and shaping process, and physical time cannot accurately measure the changes in the relaxation properties inside the material. To decouple the interaction between time and temperature, this embodiment introduces the time-temperature equivalence principle from polymer rheology to construct a calculation benchmark, transforming the temperature-variable relaxation process into an isothermal equivalent process, providing a physical scale for subsequent stress calculations.

[0082] S210 applies the time-temperature equivalence principle to perform the physical scale conversion of temperature sequence to reduced time.

[0083] Extraction module 220 establishes a data call connection with recombination module 210, reads the Lagrange particle thermal history set matrix, and separates the non-isothermal cooling temperature time series recorded within the matrix. Extraction module 220 calls the preset Williams-Landel-Ferry equation to calculate the time-temperature equivalent shift factor. The Williams-Landel-Ferry equation outputs the shift factor value at the corresponding time based on the difference between the current transient temperature and the set material reference temperature, combined with empirical constants characterizing the relaxation dependence of the material. To ensure the completeness of the algorithm logic and avoid computational dead zones caused by the denominator of the equation approaching zero, the set material reference temperature is defined within a numerical range of 10°C to 50°C above the glass transition temperature of the corresponding polymer material. For the testing and calibration of empirical constants, those skilled in the art can use a dynamic thermomechanical analyzer to perform multi-frequency linear sweep tests. The parameter acquisition method is a well-known technology in the field and will not be described in detail here.

[0084] After obtaining the shift factor, the extraction module 220 performs numerical integration along the physical time axis on the reciprocal of the shift factor, converting the physical time coordinates into reduced time coordinates at the reference temperature. After the conversion, the stress relaxation behavior under non-isothermal cooling boundary is mapped to an isothermal relaxation process at a constant reference temperature, eliminating the risk of calculation divergence caused by drastic temperature changes in subsequent integration formulas.

[0085] S220 performs discrete integral calculations of thermal strain during non-isothermal cooling and shaping processes based on a rheological model.

[0086] Extraction module 220 uses a generalized Maxwell viscoelastic rheological model to characterize the stress relaxation modulus of the plastic sheet material. The stress relaxation modulus characterizes the physical properties of the material's elasticity and viscosity decaying with reduced time. The generalized Maxwell model, through the parallel connection of multiple basic units, covers the microscopic relaxation behavior of polymer materials at multiple time scales. Under the dual effects of track traction and cooling water immersion, the internal cross-sectional particles of the plastic sheet undergo volume shrinkage due to temperature drop. Extraction module 220 calculates the pure thermal shrinkage strain increment based on the material's linear expansion coefficient and temperature gradient.

[0087] Substituting the thermal shrinkage strain increment into the linear viscoelastic genetic integral equation, and performing continuous discrete integral calculations along the reduced time axis with respect to the thermal strain, the three-dimensional residual stress tensor inside the particle is solved. The integral formula for the three-dimensional residual stress tensor is as follows: ; In the formula, for The three-dimensional residual stress tensor components at time t, subscript and These represent the three orthogonal directions of a three-dimensional Cartesian coordinate system. The stress relaxation modulus function decays over time; The time is the reduced time after scaling. The integral traversal variable for physical time; The total mechanical strain component is determined by the traction linear velocity, plate thickness constraint, cooling and shaping section support constraint, and preset plane stress or plane strain boundary conditions, and is input as a thermodynamic boundary condition into the linear viscoelastic genetic integral equation. The strain component is the pure thermal shrinkage strain. The extraction module 220 uses the backward Euler difference algorithm to transform the continuous integral into a discrete time-domain algebraic sum form, and calculates the stress tensor results at each node by combining the preset small discrete time steps.

[0088] S230, extract the deviatoric stress tensor invariants to construct a physical feature vector characterizing the internal stress.

[0089] The three-dimensional residual stress tensor contains hydrostatic stress components that cause changes in material volume and deviatoric stress components that cause distortion in material shape; macroscopic warping deformation is driven by deviatoric stress. Inputting the complete tensor matrix into the subsequent neural network increases computational resource consumption and interferes with the convergence of feature extraction in the prediction model. Extraction module 220 decomposes the three-dimensional residual stress tensor, constructs a hydrostatic stress tensor based on the average of the three normal stress components in the three-dimensional residual stress tensor, and subtracts this hydrostatic stress tensor from the three-dimensional residual stress tensor to extract the deviatoric stress tensor. After separation, extraction module 220 calculates the scalar result of the second invariant of the deviatoric stress tensor and outputs the equivalent residual stress deviatoric tensor integral value, calculated as follows: ; In the formula, This is the integral value of the equivalent deviatoric residual stress tensor in scalar form; To filter out the deviatoric stress tensor components after removing hydrostatic stress, This indicates that the squares of each component of the deviatoric stress tensor are summed according to the tensor summation convention. The extraction module 220 concatenates the equivalent residual stress deviatoric tensor integral values ​​calculated at each thickness level of the discrete mass cross-section into an array, constructing a low-dimensional array without redundant volumetric deformation interference. This low-dimensional array is flattened into a one-dimensional physical feature vector, directly serving as data characterizing the non-uniform stress distribution within the plastic sheet. This provides the downstream prediction module 230's feedforward neural network with mechanism-compressed input features, ensuring that the subsequent neural network has physical constraint boundaries during processing. Simultaneously, the extraction module 220 establishes an index association between the intermediate calculation results of the three-dimensional residual stress tensor and the corresponding equivalent residual stress deviatoric tensor integral values, sending them to the verification module 240 or storing them in the temporary data area of ​​the server 130 for subsequent energy self-consistency verification.

[0090] See Figure 5 , Figure 5 This is a schematic diagram of a feedforward neural network prediction principle according to an embodiment of the present invention. In this embodiment, S30 in the plastic sheet molding warpage prediction method is further refined into the following specific sub-steps. A complex nonlinear mapping relationship exists between the microscopic residual stress and the macroscopic warpage morphology during the molding process of polymer materials, making accurate decoupling difficult with conventional linear algebraic equations. To solve this problem, this embodiment employs a deep feedforward neural network to perform quantitative deduction of geometric deformation, aiming to transform the physical characteristics calculated by the aforementioned mechanism into macroscopic prediction indicators that can be verified in geometric space.

[0091] S310: Obtain the physical feature vector and basic parameters of the board material, and perform feature space preprocessing.

[0092] The prediction module 230 is communicatively connected to the extraction module 220, reading the physical feature vector characterizing the non-uniform distribution of internal stress in the plastic sheet. The feature vector is a one-dimensional array composed of flattened integral values ​​of the partial tensors of equivalent residual stress at each discrete particle cross-section; its dimension is defined by the number of extracted particles and the number of thickness levels in the sheet. Besides internal residual stress, the macroscopic deformation of the plastic sheet is also constrained by geometric boundaries and the inherent stiffness of the material. The prediction module 230 simultaneously acquires the system's preset basic parameters for the sheet, including the design thickness, nominal width, and room temperature elastic modulus of the plastic sheet.

[0093] Because the stress integral value differs from the fundamental parameters in both physical dimensions and numerical magnitude, directly inputting it into the neural network can easily lead to weight update bias and gradient divergence during training. As a preferred approach, the prediction module 230 employs the range transformation method to perform normalization preprocessing on the various types of input data. The normalization calculation formula is as follows: ; In the formula, These are the normalized dimensionless eigenvalues; The original value of the input variable; and These represent the maximum and minimum values ​​obtained statistically from the historical sample dataset for the corresponding input variables; The set minimum bias constant is used to avoid when equal The computational dead zone where the denominator is zero occurs is avoided to ensure the integrity of the algorithm logic. The normalized physical feature vector and the basic parameters of the board are concatenated and combined to construct a one-dimensional high-dimensional input feature tensor with a uniform numerical scale, which serves as the initial input data for the prediction model.

[0094] At the model deployment level, the deep feedforward neural network undergoes offline training using supervised learning before going live to determine the weight matrix and bias vector within the network. Training samples are derived from physical feature vectors calculated by module 220 and corresponding basic parameters of the sheet metal extracted from historical production batches. Training labels are derived from quality inspection equipment 140 located at the end of the production line or at offline inspection stations.

[0095] The quality inspection equipment 140 uses a 3D laser scanner to scan the surface morphology of the cooled and shaped real molded plastic sheet, extracting the actual vertical displacement at each grid node to generate actual deflection distribution data as label data. The server 130 assigns a unique batch identifier to each virtual material cross-section, and the prediction module 230 records this batch identifier simultaneously when outputting the two-dimensional deflection grid node matrix. When uploading the actual deflection distribution data, the quality inspection equipment 140 includes the corresponding batch identifier and resamples or interpolates the actual deflection distribution data according to the same grid topology size as the prediction module 230, ensuring a one-to-one correspondence between the actual deflection label value and the predicted node in the two-dimensional deflection grid node matrix. The prediction module 230 uses mean squared error as the key loss function for model training; the loss function calculation formula is as follows: ; In the formula, This represents the mean squared error loss value; This represents the total number of nodes contained in the two-dimensional deflection mesh. The first one provided for quality inspection equipment 140 The actual deflection label value of each node; The first output of the neural network The algorithm calculates the network gradient based on the mean squared error loss value and uses an adaptive moment estimation optimization algorithm to backpropagate along the network to update the weights until the loss value converges to the preset lower bound of error.

[0096] S320 constructs a deep feedforward neural network and performs forward propagation computation and data reshaping of multi-layer neurons.

[0097] In this embodiment, the prediction module 230 uses a deep feedforward neural network based on a multilayer perceptron architecture. The network's internal hierarchical structure consists of one input layer, three to five hidden layers, and one output layer, with fully connected modules connecting each layer. The number of neurons in the hidden layers decreases in a stepwise manner, and the data flow between adjacent layers is unidirectional forward propagation, without feedback loops. To enhance the model's ability to fit high-dimensional nonlinear surfaces and alleviate gradient vanishing, the hidden layers use a linear rectified function as the activation function. After receiving the constructed one-dimensional high-dimensional input feature tensor, the network sequentially performs linear weighted summation and nonlinear activation calculations at each layer node. The formula for single-layer forward propagation is as follows: ; In the formula, For the first The activation vector output by the hidden layer; This is the input vector passed from the previous layer; To connect the first Layer and First Layer weight matrix; For the first Layer bias vector; This represents a linear rectified activation function. Data passes through the hidden layer and is eventually derived to reach the output layer.

[0098] The number of nodes in the output layer of the deep feedforward neural network is set to match the total number of grid nodes in the discretized plastic sheet surface. The output layer outputs a one-dimensional prediction column vector after a linear transformation. To achieve physical spatial mapping of the data structure, the prediction module 230 reshapes the one-dimensional prediction column vector into a two-dimensional deflection grid node matrix according to a preset topology size. The business meaning of the two-dimensional deflection grid node matrix points to the macroscopic spatial morphology of the molded plastic sheet. The row and column indices of the matrix map to the discrete physical coordinate points of the plastic sheet along the extrusion traction direction and width direction, respectively. The element values ​​within the matrix represent the vertical displacement difference of the corresponding physical coordinate point relative to the standard horizontal plane, i.e., the local deflection prediction value. By generating the two-dimensional deflection grid node matrix, the system realizes the digital representation of the macroscopic warping prediction morphology of the molded plastic sheet in three-dimensional space.

[0099] See Figure 6 , Figure 6 This is a schematic diagram illustrating the principle of cross-scale virtual-real elastic strain energy self-consistency verification according to an embodiment of the present invention. In this embodiment, S40 of the cross-scale virtual-real elastic strain energy self-consistency verification is further refined into the following specific sub-steps. The two-dimensional deflection grid output by the neural network is essentially a data-driven spatial geometric prediction result, which may deviate from physical reality when dealing with specific working conditions. In order to abandon the traditional pure data loss calculation, this embodiment establishes an internal evaluation scale based on the physical energy conservation law, and evaluates the reliability of the prediction result in terms of physical mechanism by verifying the energy self-consistency between microscopic residual stress and macroscopic bending deformation.

[0100] S410 performs algebraic solution of the mechanism of microscale thermal stress release potential energy based on the extracted normal stress components and material properties.

[0101] The verification module 240 establishes a communication connection with the extraction module 220 to obtain a three-dimensional residual stress tensor indexed and associated with the integral value of the equivalent residual stress deviator tensor. This three-dimensional residual stress tensor is an intermediate calculation result retained by the extraction module 220 during the discrete integral calculation of thermal strain. The potential energy accumulated inside the material during non-uniform cooling theoretically corresponds to the spatial integral of the elastic strain energy density within each micro-element volume. As a preferred method, the verification module 240, based on the extracted normal stress components and combined with the material's elastic modulus and Poisson's ratio, solves for the theoretical potential energy accumulated inside the micro-element particles. The verification module 240 determines the discrete micro-element volume based on the geometric dimensions of the particle division, performs discrete accumulation along the overall computational domain of the material, and calculates the micro-accumulated elastic strain energy. The formula for calculating the micro-accumulated elastic strain energy is as follows: ; In the formula, This is for the accumulation of elastic strain energy at the microscopic level; The total number of discrete infinitesimal elements; The room temperature elastic modulus of the material; The Poisson's ratio of the material; , and The first The three normal stress components of a microelement in a three-dimensional Cartesian coordinate system; For the first The physical volume of each infinitesimal element. The energy value obtained by solving the mechanism algebraically represents the total internal energy accumulated inside the plastic sheet due to non-uniform cooling, providing a reference energy input for subsequent physical verification.

[0102] S420 introduces the central finite difference method to perform spatial second-order differentiation, and completes the numerical derivation and solution of geometric deformation energy.

[0103] The verification module 240 receives the reshaped and transformed two-dimensional deflection mesh node matrix from the prediction module 230. Based on the classical thin plate bending theory, the macroscopic bending geometric strain energy consumed by the plastic sheet under the drive of residual stress release during macroscopic warping deformation can be obtained through the deflection distribution function of the plate surface. Since the matrix output by the neural network is discretized data, the verification module 240 introduces a discrete difference operator to transform the two-dimensional deflection prediction matrix into spatial second-order partial derivatives. Specifically, after extracting the physical spacing of the mesh nodes, the second-order partial derivatives at each discrete coordinate point are calculated using the central finite difference method to obtain the local curvature. The verification module 240 substitutes the derivative results into the thin plate bending energy equation to calculate the macroscopic thin plate bending strain energy. The calculation formula is as follows: ; In the formula, For macroscopic thin plate bending strain energy; This refers to the bending stiffness parameter of the plastic sheet; This represents the total number of discrete elements contained in the two-dimensional deflection mesh. For the first Predicted vertical deflection values ​​for each grid node; For deflection Laplace operator; and These represent the physical spacing of the grid along the longitudinal and transverse directions, respectively. For bending stiffness... The specific value can be derived by those skilled in the art using the plate thickness and elastic modulus through formulas, and will not be elaborated here.

[0104] S430 calculates the characteristic distance of the energy self-consistency residual based on Euclidean distance and triggers the decision mechanism.

[0105] Based on the energy conversion and magnitude correspondence between residual stress release and macroscopic bending deformation, if the predicted shape of the neural network closely matches physical reality, the macroscopic thin plate bending strain energy should numerically approach the internal microscopic accumulated elastic strain energy. The verification module 240 calculates the Euclidean distance between the thermal stress release potential energy and the geometric deformation energy, generating a residual index logic to measure the consistency between the mechanism calculation and the network's physical prediction. To avoid a computational dead zone where the denominator is zero when the microscopic accumulated elastic strain energy approaches zero, this embodiment adds a minimal bias constant to the denominator. The formula for calculating the energy self-consistency residual index is as follows: ; In the formula, It is a dimensionless energy self-consistency residual index; This is the set minimum bias normal number. The verification module 240 internally sets a confidence tolerance threshold and an outlier isolation threshold higher than the confidence tolerance threshold. The confidence tolerance threshold is defined as ranging from 5% to 15%, with the specific value determined by the on-site process requirements and the comprehensive measurement accuracy of the multi-dimensional sensors. The outlier isolation threshold is used to identify extreme anomalies that severely deviate from physical constraints, with values ​​exceeding the confidence tolerance threshold. As a preferred approach, the outlier isolation threshold can be set to 1.5 to 3 times the confidence tolerance threshold, or determined based on the upper quantile of the statistical distribution of the energy self-consistency residual index in historical normal production batches. The system executes a logical comparison and judgment instruction to determine whether the energy self-consistency residual index exceeds the confidence tolerance threshold and whether it further exceeds the outlier isolation threshold.

[0106] When the energy self-consistency residual index When the prediction result is less than or equal to the confidence tolerance threshold, the verification module 240 determines that the prediction result of the current batch conforms to the physical consistency constraint and outputs a warpage warning signal with high confidence to the controller 120 or the production line monitoring interface. When the energy self-consistency residual index... When the confidence tolerance threshold is greater than the threshold but less than or equal to the outlier isolation threshold, the verification module 240 determines that the physical confidence of the current batch prediction result has decreased, intercepts the direct transmission of the prediction result to the controller 120 or the production line monitoring interface, and retains the batch sample as a low-confidence sample that can be used for reinforcement learning reward evaluation.

[0107] If the residual index further exceeds the outlier isolation threshold, the system determines that the neural network model may be overfitting when handling the current specific operating conditions, or that the boundary conditions of the mechanism layer have shifted. At this time, the verification module 240 marks the feature tensor of the current batch as an outlier, adds a physical consistency anomaly label to the corresponding batch samples, and transfers them to the specific isolation sample library. At the same time, it intercepts the transmission of distorted prediction results and pushes model decay alarm information to the central control room. The abnormal data isolation mechanism prevents the underlying control equipment from adjusting process parameters based on distorted prediction deflection, ensuring the safety and stability of the closed-loop control of the extrusion production line.

[0108] See Figure 7 , Figure 7 This is a schematic diagram of a closed-loop calibration principle based on a hybrid reward-driven and sample-isolated approach according to an embodiment of the present invention. In this embodiment, S50 of the closed-loop calibration with hybrid reward-driven and sample-isolated approach is further refined into the following specific sub-steps. In actual polymer extrusion processes, factors such as environmental temperature fluctuations and equipment aging can easily cause invisible boundary parameters to drift over time. In addition, human intervention is present on-site, making static prediction models prone to failure under such time-varying environments. This embodiment utilizes reinforcement learning combined with a contamination interception mechanism to establish an online adaptive closed-loop calibration logic that integrates system mechanistic constraints and quality inspection feedback.

[0109] S510 implements a contamination label filtering and abnormal sample isolation mechanism before the data queue enters the reinforcement learning reward evaluation process.

[0110] The calibration module 250 establishes a communication connection with the external control bus and the quality inspection equipment 140, receiving the data queue of the molding batch in real time. Unrecorded mechanical operations exist on the continuous extrusion line, such as manual online filter replacement and manual adjustment of the die lip bolts, which can cause sudden changes in the die melt pressure, as well as abnormal adjustments to the traction track speed or cooling water pressure, which can be identified by abnormal equipment status codes or process parameters. These random interventions generate contaminated samples that deviate from thermodynamic laws. Directly using such samples for model iteration will lead to biased network weight updates.

[0111] To mitigate misjudgments of rewards due to unrecorded mechanical operations, the calibration module 250 actively reads the attached tag variables from the data queue. As a preferred approach, the tag variables include human intervention contamination tags, equipment status anomaly tags, and physical consistency anomaly tags. Among them, the human intervention contamination tag is generated by the time gradient triggering the melt pressure of the die head; the equipment status anomaly tag is generated by the abnormal status of equipment status codes or process parameters such as traction speed and cooling water pressure uploaded by the controller 120; and the physical consistency anomaly tag is generated when the energy self-consistency residual index exceeds the outlier isolation threshold.

[0112] The system identifies batches of data carrying labels indicating human intervention contamination, abnormal equipment status, or abnormal physical consistency based on preset filtering rules. For samples carrying these labels, the calibration module 250 removes them from the main training data stream and temporarily stores them in a specific isolation sample library for offline review; these samples do not participate in current online reward calculations or action network updates. For samples without these labels and whose energy self-consistency residuals do not exceed the outlier isolation threshold, the calibration module 250 retains them for subsequent reward evaluation and converts the residual magnitude into an instantaneous dense reward signal. Only when the energy self-consistency residual exceeds the outlier isolation threshold are the corresponding samples transferred to the specific isolation sample library.

[0113] S520 constructs a hybrid signal that integrates immediate dense rewards and delayed sparse rewards based on mechanism calculations and end-of-line or offline quality inspection results.

[0114] The convergence of reinforcement learning algorithms depends on the design of the reward function. Relying solely on terminal quality inspection data as feedback suffers from time lag, making it difficult to achieve online model adaptation. Before performing reward fusion, server 130 determines whether quality inspection equipment 140 has uploaded the actual deflection distribution data for the corresponding batch based on the unique batch identifier of the virtual material cross-section. For batches that have already uploaded actual deflection distribution data, server 130 matches the two-dimensional deflection grid node matrix output by prediction module 230 with the actual deflection distribution data uploaded by quality inspection equipment 140, and performs grid resampling or interpolation processing on the actual deflection distribution data to ensure a one-to-one correspondence between the actual deflection label value and the predicted deflection node value.

[0115] When the actual deflection distribution data for the corresponding batch has not yet been returned by the quality inspection equipment 140, the calibration module 250 first generates an instantaneous dense reward based on the energy self-consistency residual, and writes the batch identifier, status, action vector, action network output information and prediction results of the batch into the delayed reward cache area; after the quality inspection equipment 140 returns the actual deflection distribution data with the same batch identifier, the actual deflection distribution error is calculated and a delayed sparse reward is generated, and the delayed sparse reward is backfilled into the delayed reward cache area of ​​the corresponding batch for supplementary updates of the action network.

[0116] The formula for calculating the hybrid reward signal is as follows: ; In the formula, The reward signal value is a mixture. and These represent the microscopic cumulative elastic strain energy and macroscopic thin plate bending strain energy calculated at the mechanistic layer for current high-quality samples, respectively. The minimum bias constant is set to avoid the computational dead zone when the denominator approaches zero. This represents the total number of nodes in the deflection grid. and The first The actual deflection label value of each node and the network prediction value; This refers to the weighting coefficient for dense rewards; This refers to the sparse reward weighting coefficient. The dense reward term, based on real-time mechanism algebraic solution, provides continuous physical direction guidance for the agent; the sparse reward term ensures the accuracy of the final geometric mapping. The values ​​of both the dense and sparse reward weighting coefficients are defined between 0.1 and 0.9, and their sum is strictly set to one. The specific values ​​are configured by technicians based on the acquisition frequency of the current process data stream.

[0117] The S530 drives the motion network to perform high-frequency iterations and outputs the corrected values ​​of thermodynamic boundary environment parameters in reverse to calibrate the system in a closed loop.

[0118] The calibration module 250 internally incorporates an action network for adaptive adjustment of boundary environment parameters. Specifically, the action network employs a policy network based on a multilayer perceptron architecture, containing an input layer, three to five fully connected hidden layers, and an output layer. The number of neurons in the hidden layers decreases in a stepwise manner, and the hidden layers use a linear rectified activation function. The output layer uses a hyperbolic tangent activation function to constrain the boundary range of the output action. During each iteration, the policy network receives the normalized physical feature vector of the current time step and a high-dimensional feature tensor formed by concatenating the basic parameters of the board material as its state input.

[0119] After forward propagation calculations, the network generates continuous action vectors representing process corrections at the output layer. The business meaning of the action vectors points to the boundary parameter correction terms in the extrusion cooling and shaping process. To establish strict structural constraints, the number of nodes in the action network output layer is strictly equal to the number of independent cooling sections in the production line plus one, mapping the convective heat transfer coefficient correction of each cooling section and an overall traction speed compensation value. The traction speed compensation value, used by the extraction module 220 when performing physical displacement tracking, total mechanical strain boundary calculation, and residual stress characteristic calculation, is not directly equivalent to the control commands issued to the underlying traction equipment.

[0120] The calibration module 250, acting as the training engine for the reinforcement learning agent, receives parameter correction terms output by the action network. Before the actual deflection distribution data is returned, it performs short-cycle updates by combining the immediate dense reward signal. After the actual deflection distribution data is returned and a delayed sparse reward signal is generated, it combines the backfilled hybrid reward signal. A supplementary update is performed, employing a near-end policy optimization algorithm to calculate the network gradient. The algorithm propagates back along the policy network, updating the weight matrix and bias vector within the network, prompting the network to converge towards a parameter space that achieves a higher mixed reward value, i.e., minimizing both physical and geometric errors. After completing a single action network iteration, the calibration module 250 backfeeds the output thermodynamic boundary environment parameter correction term to the preceding mechanism layer thermal history calculation module, performs amplitude limiting and smoothing processing on the thermodynamic boundary environment parameter correction term, and then updates the corresponding original boundary parameters.

[0121] In subsequent batch calculations, the extraction module 220 re-executes thermal strain integration and residual stress feature extraction based on the corrected boundary environment parameters, and the prediction module 230 generates a new two-dimensional deflection mesh node matrix based on the updated physical feature vectors, thus forming a closed-loop calibration link for the prediction results of plastic sheet molding warpage. Through the above closed-loop execution logic, the system can adaptively track changes in production line conditions and maintain the reliability of the warpage prediction algorithm without relying on manual shutdown calibration.

[0122] In this embodiment, the plastic sheet molding warpage prediction system is deployed on a continuous extrusion production line, using polycarbonate polymer material. The set basic sheet parameters include a nominal thickness of 5 mm, a nominal width of 1200 mm, a room temperature modulus of elasticity of 2300 MPa, and a reference temperature of 160 degrees Celsius. Three independently controlled cooling water sections are configured along the extrusion production line.

[0123] During system operation, acquisition node 110 obtains production parameters, specifically ambient temperature and plastic sheet surface temperature via temperature sensor 111, cooling water pressure and die head melt pressure via pressure sensor 112, and traction line speed via photoelectric encoder 113. The aforementioned sensor data streams are transmitted to server 130 via controller 120. The reassembly module 210 running within server 130 converts the sensor data streams into absolute timestamp format and maps and combines discrete Euler observation data into a Lagrange point thermal history set based on an integral displacement mechanism. If an unrecorded intervention operation occurs during production—such as manual fine-tuning of the die head lip bolts—and the reassembly module 210 detects that the first-order time derivative of the die head melt pressure exceeds the set mechanical intervention safety threshold, the system immediately adds a manual intervention contamination label to the Lagrange point thermal history set generated within the corresponding time window.

[0124] The extraction module 220 reads the thermal history set of particles, substitutes it into the time-temperature equivalent equation and the generalized Maxwell viscoelastic rheological model, and calculates the equivalent residual stress partial tensor integral value of the polycarbonate sheet during the non-isothermal cooling phase transformation process, constructing a physical feature vector characterizing the non-uniform distribution of internal stress. The prediction module 230 performs range normalization processing on the physical feature vector and the sheet's basic parameters, and then inputs it into a deep feedforward neural network. The network's forward propagation outputs a two-dimensional deflection mesh node matrix characterizing spatial deformation.

[0125] The calibration module 240 extracts the microscale thermal stress tensor and calculates the microscale cumulative elastic strain energy. Simultaneously, it uses the central finite difference method to perform spatial second-order differentiation on the two-dimensional deflection grid node matrix to calculate the macroscopic thin-plate bending strain energy, and calculates the difference between the two to generate an energy self-consistency residual. For samples carrying artificially intervened contamination labels, the calibration module 250 directly imports them into a specific isolated sample library. For high-quality samples, the calibration module 250 converts the energy self-consistency residual into an instantaneous dense reward signal, and generates a delayed sparse reward signal after receiving the actual deflection distribution data returned by the 3D laser scanner (quality inspection equipment 140) at the end of the production line. The reinforcement learning agent calculates the strategy network gradient based on the mixed reward signal and outputs the correction amount for the convective heat transfer coefficient and the traction speed compensation value for the three cooling water sections. The correction parameters are fed back to the extraction module 220 to complete the boundary condition replacement, achieving online unsupervised closed-loop calibration.

[0126] This embodiment constructs a comparative experiment of continuous production line operation for 72 hours. The experimental environment incorporates two boundary drift caused by cooling water temperature fluctuations and one unrecorded mechanical intervention operation of the die head. The experiment is divided into three comparative groups: the first group uses a static mechanism analysis method, relying on fixed thermodynamic boundary parameters to deduce residual stress; the second group uses a purely data-driven deep feedforward neural network prediction method; and the third group uses the fusion energy self-consistency verification and hybrid reward-driven closed-loop calibration method provided in this invention. The quantitative evaluation indicators are set as mean absolute error and maximum prediction error. The experimental statistical results are shown in the table below.

[0127] Comparison Methods Mean Absolute Error Maximum prediction error Abnormal sample interception rate Static mechanism analysis method 1.35 mm 2.85 mm not applicable Pure data-driven deep feedforward neural network prediction method 1.05 mm 2.45 mm 0% The method provided by this invention 0.82 mm 1.65 mm 100% Static mechanism analysis methods lack adaptive correction mechanisms when encountering production line boundary parameter drift, making it difficult for stress-derived boundary conditions to accurately reflect actual operating conditions, resulting in a high overall error base. Pure data-driven deep feedforward neural network prediction methods exhibit some convergence during normal operation, but mechanical intervention introduces contaminated samples into the weight update stream, triggering localized error amplification in the model. The method provided in this invention intercepts abnormal interference data through physical consistency verification and label filtering logic, combined with a hybrid reward signal driving the action network to output boundary correction parameters, dynamically compensating for thermodynamic parameter drift caused by cooling water temperature fluctuations. During the comparative testing period, the average absolute error of the method in this invention was controlled at 0.82 mm, achieving a certain degree of accuracy improvement compared to pure data-driven methods and effectively suppressing the maximum prediction error under extreme operating conditions.

[0128] See Figure 8 , Figure 8 This is a three-dimensional waterfall plot illustrating the spatiotemporal evolution of the predicted residuals according to an embodiment of the present invention. In the plot, the X-axis is defined as a long-period continuous production time series, the Y-axis discretizes the predicted grid nodes representing the transverse width of the plastic sheet, and the Z-axis quantifies the absolute value of the difference between the network-predicted displacement and the measured deflection by the three-dimensional laser scanner. The continuous spatiotemporal surface constructed in the plot visually presents the dynamic evolution characteristics of the system's closed-loop calibration process. During the stable production cycle in the early part of the time axis, without external intervention, the surface undulates gently, and the Z-axis amplitude remains within the low error band. When the time series progresses to a specific time window recording cooling water temperature fluctuations and mold head mechanical intervention, the Z-axis elevation experiences a local spike, forming a residual peak, indicating that the static prediction model encounters physical boundary failure under such time-varying conditions. With the intervention of the reinforcement learning mechanism within the calibration module and the high-frequency iterative updates of the action network, the surface topology rapidly falls back to the error reference plane and eventually converges and reconstructs into a stable band, confirming the convergence stability of the system's fusion of hybrid reward and sample isolation mechanisms when tracking nonlinear time-varying disturbances from a spatial evolution perspective.

[0129] See Figure 9 , Figure 9This is a three-dimensional topological diagram of residual stress release and macroscopic deformation reconstruction according to an embodiment of the present invention. The base mesh in the diagram is orthogonally divided into two-dimensional planes strictly according to the physical dimensions of the sheet material set in production and the preset finite difference discretization accuracy. The vertical deflection absolute component output by the prediction module is mapped positively to the Z-axis of the three-dimensional spatial coordinate system of each discrete mesh node, reconstructing a macroscopic deformation surface that conforms to the actual formed appearance. The pseudo-color spectrum matrix rendered on the surface layer deeply integrates the equivalent residual stress partial tensor integral value calculated by the bottom extraction module. The color band spectrum transitions continuously from cool to warm tones, locating the high microscopic stress accumulation area inside the plastic sheet caused by non-isothermal phase change shrinkage. Through cross-scale coupling mapping of geometric curvature and physical stress spectrum, the diagram intuitively reveals the physical consistency of the transformation of stress release potential energy into thin sheet bending deformation energy. The outer edge of the topological morphology exhibits an asymmetric saddle-shaped raised edge, while the central region of the sheet material, constrained by traction and tensile tension, forms a gently concave structure, highly restoring the macroscopic evolution morphology of polycarbonate polymer materials under complex thermodynamic asymmetric boundary constraints.

[0130] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation, characterized in that, Includes the following steps: S10 receives the production line sensor data stream, tracks the physical displacement of the material cross section based on the traction line speed, and maps and combines the discrete observation data into a Lagrange mass thermal history set of the same virtual material cross section. S20, monitor the time gradient of the melt pressure signal of the die head, and when the time gradient triggers a preset intervention threshold, add an artificial intervention contamination label to the corresponding mapping combination of the Lagrange particle thermal history set. S30, read the Lagrange particle thermal history set, calculate the thermal strain using the time-temperature equivalent equation and rheological model, and obtain the physical feature vector by outputting the integral value of the equivalent residual stress partial tensor and the three-dimensional residual stress tensor through discrete integration. S40, input the physical feature vector and the basic parameters of the plate into the deep feedforward neural network, output the two-dimensional deflection grid node matrix, and generate the molding warping prediction result. S50, calculate the energy self-consistency residual based on the three-dimensional residual stress tensor and the two-dimensional deflection grid node matrix for physical verification, fuse the energy self-consistency residual with the actual deflection error to generate a reward signal, remove samples with additional labels, update the action network and output the corrected boundary environment parameters.

2. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation as described in claim 1, characterized in that, In step S10, receiving the production line sensor data stream, tracking the physical displacement of the material cross-section based on the traction line speed, and mapping and combining discrete observation data into a Lagrange mass thermal history set of the same virtual material cross-section includes: A circular buffer is constructed by allocating contiguous memory space, and the discrete observation data with absolute timestamps are continuously written into the circular buffer. After compensating for the traction line speed by combining empirical shrinkage rate, numerical integration is performed to calculate the cumulative physical displacement of the virtual material cross section; The cumulative physical displacement is compared with the physical distance coordinates of the sensor. The absolute timestamp when the difference between the cumulative physical displacement and the physical distance coordinates of the sensor falls within a preset tolerance range is recorded as the arrival time. The data with the smallest difference between the timestamp and the arrival time is extracted from the circular buffer and pushed into the first-in-first-out queue.

3. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation as described in claim 1, characterized in that, In step S20, the time gradient of the monitored die melt pressure signal, when the time gradient triggers a preset intervention threshold, adds an artificial intervention contamination label to the corresponding mapping combination's Lagrange particle thermal history set, including: The first-order time derivative of the die head melt pressure signal is calculated using forward differential logic. Determine whether the absolute value of the first-order time derivative is greater than the mechanical intervention safety threshold, wherein the mechanical intervention safety threshold is defined based on the standard deviation of the pressure derivative distribution within the historical production cycle, and the preset intervention threshold is equal to the mechanical intervention safety threshold; If the absolute value of the first-order time derivative is greater than the mechanical intervention safety threshold, the reserved Boolean state position is set to true, and the artificial intervention pollution label is generated.

4. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation as described in claim 1, characterized in that, In step S30, the calculation of thermal strain using the time-temperature equivalent equation and rheological model, and the output of the integral value of the equivalent residual stress deviator tensor and the three-dimensional residual stress tensor through discrete integration, includes: The time-temperature equivalent shift factor is calculated based on the transient temperature and the material reference temperature, and the physical time coordinate is converted into a reduced time coordinate at the reference temperature. The pure thermal shrinkage strain increment is calculated based on the linear expansion coefficient and temperature gradient. The pure thermal shrinkage strain increment is then substituted into the linear viscoelastic genetic integral equation and integral calculation is performed along the reduced time coordinate to obtain the three-dimensional residual stress tensor. A hydrostatic stress tensor is constructed based on the average value of the three normal stress components in the three-dimensional residual stress tensor, and a deviatoric stress tensor is extracted by subtracting the hydrostatic stress tensor from the three-dimensional residual stress tensor. The second invariant scalar result of the deviatoric stress tensor is calculated to obtain the integral value of the equivalent residual stress deviatoric tensor.

5. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation as described in claim 1, characterized in that, In step S40, before inputting the physical feature vector and the basic parameters of the board material into the deep feedforward neural network, the following steps are included: Obtain the preset basic parameters of the board material, including the design thickness, nominal width, and room temperature elastic modulus of the material; The integral value of the equivalent residual stress deviator tensor and the basic parameters of the plate are normalized with bias constants using the range transformation method. The normalized data are concatenated and combined to construct an input feature tensor with a uniform numerical scale, which serves as the physical feature vector.

6. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation according to claim 1, characterized in that, In step S50, the physical verification of the energy self-consistency residual calculated based on the three-dimensional residual stress tensor and the two-dimensional deflection mesh node matrix includes: Based on the normal stress component extracted from the three-dimensional residual stress tensor, combined with the elastic modulus and Poisson's ratio, the elastic strain energy density inside the discrete micro-element is solved and the discrete volume integral is accumulated to obtain the micro-accumulated elastic strain energy. The local curvature at the coordinate points of the two-dimensional deflection grid node matrix is ​​calculated using the central finite difference method. The local curvature and bending stiffness parameters are substituted into the thin plate bending energy equation to obtain the macroscopic thin plate bending strain energy. The Euclidean distance is calculated based on the microscopic cumulative elastic strain energy and the macroscopic thin plate bending strain energy to generate the energy self-consistency residual.

7. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation as described in claim 6, characterized in that, In step S50, fusing the energy self-consistency residual with the actual deflection error to generate a reward signal includes: The absolute value of the difference between the micro-accumulated elastic strain energy and the macro-thin plate bending strain energy is divided by twice the sum of the micro-accumulated elastic strain energy and the bias constant, and then multiplied by the dense reward weight coefficient after subtracting the quotient by one to obtain the dense reward component. Calculate the sum of squares of the differences between the actual deflection label value of each node and the corresponding value of the molding warp prediction result. Divide the sum of squares by the total number of nodes and take the negative value. Then multiply the sum by the sparse reward weight coefficient to obtain the sparse reward component. The dense reward component and the sparse reward component are summed to construct a hybrid reward signal.

8. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation according to claim 7, characterized in that, In step S50, during the process of fusing and generating the reward signal, the following logic control flow is executed: Configure a corresponding batch identifier for the virtual material cross-section; When the actual deflection distribution data corresponding to the batch identifier is not returned, the dense reward component is generated as an instantaneous dense reward signal based on the energy self-consistency residual, and the prediction state, action vector and the forming warp prediction result are written into the delayed reward buffer. After receiving actual deflection distribution data with the same batch identifier, the sparse reward component is calculated as a delayed sparse reward signal, the delayed sparse reward signal is backfilled into the delayed reward buffer, and the mixed reward signal is calculated.

9. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation according to claim 1, characterized in that, In S50, updating the action network after removing samples with additional labels includes: Read the label variables carried in the data queue, remove the batch data carrying the artificial intervention pollution label from the training data stream and transfer it to the specific isolation sample library; The near-end strategy optimization algorithm is used to calculate the network gradient and update the weights of the action network by combining the reward signal corresponding to the unremoved samples, thereby generating the correction amount of the convective heat transfer coefficient and the traction speed compensation value of the cooling section.

10. The method for predicting warpage in plastic sheet molding based on multidimensional thermal parameters and stress calculation according to claim 9, characterized in that, In step S50, the output of the corrected boundary environment parameters includes: The correction amount for the convective heat transfer coefficient and the compensation value for the traction speed are subjected to amplitude limiting and smoothing processing; The processed convective heat transfer coefficient correction is used to replace the rheological model environment input parameters, and the processed traction speed compensation value is used to update the integral tracking variable in the process of tracking the physical displacement of the material cross section.

Images