Plastic accessory molding process parameter optimization regulation and control system

Abstract

The invention relates to a plastic part forming process parameter optimization regulation and control system, in particular to the field of plastic part forming, through the synergistic effect of multiple modules, the intelligent level and stability of a forming process are remarkably improved, and the system can quickly respond to the physical property change of a new material or a new batch; intelligently generating high-adaptability initial parameter suggestions based on historical data; by fusing physical simulation and real-time sensor data, dynamic behavior characteristics of a material are accurately captured, and an enhanced characteristic vector is constructed; on the premise of strictly following the thermodynamic safety boundary, the process parameter combination is efficiently optimized to meet the complex rheological constraint; and meanwhile, closed-loop precipitation and self-improvement of process knowledge are realized according to a continuous evolution anti-disturbance control strategy of production data, rapid and accurate regulation and control in a high-frequency switching scene are finally achieved, and the consistency of forming quality and production stability of plastic accessories are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of plastic parts molding, and more specifically, to a plastic parts molding process parameter optimization and control system. Background Art

[0002] In modern injection molding manufacturing, especially in the production of high-end precision plastic components such as automotive parts, electronic appliances, and medical devices, orders are increasingly trending towards smaller batches, higher varieties, and greater customization. At the same time, to achieve cost savings and environmental protection goals, the incorporation of large quantities of recycled plastics or specific property modifications (such as the addition of high proportions of masterbatches and toughening agents) has become the norm. This forces single injection molding lines to face frequent production switching tasks: switching between different plastic types, such as switching from polypropylene to acrylonitrile-butadiene-styrene copolymer, or switching between batches of the same plastic substrate from different suppliers, with different recycled content ratios, and with different additive formulations. These switching operations can occur multiple times daily or even hourly, becoming a critical factor affecting the overall operational efficiency and resource utilization of the production line. The switching process is also accompanied by external uncertainties such as natural fluctuations in the workshop temperature and humidity, further exacerbating the complexity of process control. This high-frequency switching scenario requires systems with extremely high agility and intelligence to adapt to rapidly changing production needs. Otherwise, production setup time, resource consumption, and defective product rates will be significantly increased.

[0003] The prior art system has significant technical bottlenecks in dealing with such switching challenges, and the current mainstream solution relies heavily on a pre-set process parameter template library. The operator needs to manually input the material brand or batch code according to the switching material to retrieve the pre-set parameter combination. However, the key physical properties of the material, such as melt index, melt viscosity, thermal stability, and shrinkage rate which ultimately affects the dimensional accuracy, not only vary significantly between different resin brands, but also vary significantly between different batches of the same brand due to fluctuations in raw material suppliers, changes in the proportion of recycled materials, and differences in color masterbatch or additive dispersion. Small shifts in these core physical property parameters can significantly change the physical behavior of the plastic during melting in the injection molding cylinder, flowing and filling in the mold cavity, and cooling during the holding pressure stage, ultimately directly affecting product quality. Unfortunately, the existing system lacks real-time online sensing capability for changes in the micro-characteristics of new materials or new batches of materials, and cannot quickly and accurately obtain or estimate dynamic parameters such as transient melt viscosity and real-time shrinkage feedback. At the same time, during the initial stage of production start-up, sensor data related to the complex flow behavior of the melt under high pressure and high speed, such as multi-point pressure distribution in the cavity and temperature field gradient, are often not fully obtained or difficult to analyze in real time. This results in the system being unable to accurately and dynamically adjust core process parameters such as injection speed, injection pressure, barrel temperature distribution, and holding pressure curve based on the actual performance of the material in the first time after switching. Usually, it can only use pre-set values or rely on operator experience for manual trial adjustment. This trial-and-error-based debugging often requires the production of several or more invalid trial samples to find the appropriate parameter combination, which is time-consuming and labor-intensive and has high uncertainty. Therefore, how to overcome the dynamic sensing problem of material properties when starting a new material or a new batch, establish a real-time online mapping model of material properties and optimal process parameters, and achieve second-level or single-mold-level adaptive optimization and stable control of process parameters has become a key problem that needs to be broken through in the current technical system, aiming to eliminate waste products during the transition period and significantly improve switching efficiency and resource utilization. SUMMARY

[0004] The present application is directed to the technical problems existing in the prior art, and provides a plastic accessory forming process parameter optimization and control system, which significantly improves the intelligent level and stability of the forming process through the cooperation of multiple modules to solve the problems raised in the above background art.

[0005] The technical solution of the present application to solve the above technical problems is as follows: specifically includes: an optimal transmission migration module, a physical residual feature generation module, a collaborative optimization decision module, and an incremental adversarial reinforcement module, wherein,

[0006] Optimal transmission migration module: in response to new material or new batch property input instructions, call the material molecular characteristics and corresponding optimal process parameters in the historical process database, according to the molecular characteristic space distribution difference between new material and historical material, execute optimal transmission mapping calculation, generate pre-adaptation process parameter prediction model for new batch, output initial parameter suggestion;

[0007] Physical residual feature generation module: based on the initial parameter suggestion to drive the injection molding physical simulation model, output the melt front position time sequence curve and the cavity pressure distribution prediction curve; Real-time acquisition of the first injection cycle screw displacement, nozzle pressure and mold cavity temperature signal, wavelet packet transform and energy entropy feature extraction of the time domain residual signal of the pressure curve predicted by the physical model and the measured pressure curve, fusion of the original sensor features to generate enhanced feature vectors;

[0008] Collaborative optimization decision module: receive the enhanced feature vectors and construct a composite objective function containing physical constraints, perform global particle swarm search and projected gradient descent operation in parallel in the parameter space, where the projected gradient descent operation is strictly limited to moving within the material thermal decomposition temperature threshold range, output the optimal parameter combination that meets the injection rheological constraints;

[0009] Incremental adversarial reinforcement module: according to the optimal parameter combination, execute continuous injection molding production, collect multi-cycle product quality index data and input into the adversarial training environment, incrementally update based on the historical strategy through the proximal policy optimization algorithm, generate robust control strategy resistant to environmental disturbances and store to the historical process database;

[0010] In a preferred embodiment, in the optimal transmission migration module, the specific operation of executing optimal transmission mapping calculation is:

[0011] Construct a historical feature space based on the molecular weight distribution analysis coefficients of historical materials and the principal components of thermal stability quantification, the molecular weight distribution analysis coefficients are obtained by analyzing the gel permeation chromatography data and fitting the molecular weight probability distribution function, and the principal components of thermal stability quantification are the characteristic vectors of the thermogravimetric analysis curve of the material in a specific temperature range; map the melt index, density, and enthalpy of the new batch of materials to the target feature space according to the predefined nonlinear function; calculate the correlation strength matrix of the historical feature space and the target feature space, the matrix elements are obtained by correcting the edge weight of the historical material process correlation graph; solve the optimal transport matrix from the historical feature space to the target feature space, the objective function includes the Euclidean distance between the historical features and the target features and the KL divergence regularization term of the process parameter distribution, and the optimal transport model is obtained by iterative calculation through entropy regularization; perform transfer transformation on the full connection layer weight matrix of the historical process parameter prediction model based on the optimal transport matrix, wherein the transfer weight decays exponentially according to the depth of the neural network level, and the convolution kernel weight is tensor contracted and combined with one-dimensional adaptive filtering operation; finally output the initial process parameter combination that satisfies the melt temperature being at least 10 degrees Celsius lower than the material thermal decomposition temperature.

[0012] In a preferred embodiment, the specific operation of the transfer transformation is:

[0013] A sparse constraint based on the process correlation graph is applied to the optimal transport matrix, only the transmission paths of the historical features with a similarity of the top 30% to the target feature space are retained; a hierarchical attenuation factor is applied to the full connection layer weight transfer process, the attenuation coefficient decreases according to the negative exponential law with the increase of the number of neural network layers, and the attenuation ratio from the input layer to the output layer is set to 1.0 to 0.2; the convolution kernel weight transfer adopts a third-order tensor contraction operation, the contraction direction is performed along the input channel dimension, and the contracted tensor is smoothed and corrected by a three-layer one-dimensional adaptive filter; the transferred historical process parameter prediction model weight is used as a pre-trained network, which runs forward propagation under the constraint condition of the material thermal decomposition safety boundary, and the safety boundary is defined as the feasible region in which the melt temperature peak does not exceed the material thermal decomposition temperature by 10 degrees Celsius; if the propagation result is out of boundary, the projection gradient method is used to pull the parameters back to the feasible region, and the corrected initial process parameter suggestion value is output, which includes the melt temperature, the injection speed, and the holding pressure.

[0014] In a preferred embodiment, in the physical residual feature generation module, the specific operation of driving the injection molding physical simulation model is:

[0015] Drive the modified Cross-WLF melt rheology model, input melt temperature, injection speed, and holding pressure parameters, and output the cavity pressure distribution prediction curve; perform five-layer wavelet packet decomposition on the time domain residual signal, and use the Daubechies-4 wavelet basis function to perform five-layer wavelet packet decomposition to generate thirty-two sub-bands; screen the sub-band nodes with an energy share greater than 1% as the valid node set, calculate the energy value of each node signal and normalize it to the energy distribution ratio; calculate the Shannon entropy value based on the energy distribution ratio of 1, and extract the time domain mutation index of each sub-band signal; combine the node energy value, Shannon entropy value and time domain mutation index into a wavelet packet feature vector with dimension 16.

[0016] In a preferred embodiment, the operation of fusing original sensor features to generate an enhanced feature vector is:

[0017] A filling uniformity index is extracted from the screw displacement signal, and this index is obtained by calculating the integral of the absolute deviation between the actual displacement velocity and the theoretical average velocity; the maximum temperature gradient value is extracted from the cavity temperature signal; the filling uniformity index, the maximum temperature gradient value, the nozzle pressure rise rate and the integral value of the pressure curve are combined into the original sensor feature vector; the original sensor feature vector and the wavelet packet feature vector are linearly weighted fused according to a preset weight matrix to generate a twenty-two-dimensional enhanced feature vector, which stores or outputs the cavity pressure prediction curve, the measured nozzle pressure curve, the screw displacement curve and the cavity temperature curve; when the maximum cavity temperature exceeds the safety threshold of the material thermal decomposition temperature minus five degrees Celsius, the twenty-second dimensional component of the enhanced feature vector is forcibly set to infinity to trigger the melt temperature anomaly protection mechanism; the preset weight matrix is ​​obtained through training of historical injection molding defect samples, and the melt temperature safety threshold is updated in conjunction with the material thermal decomposition temperature database.

[0018] In a preferred embodiment, in the collaborative optimization decision module, the operation of constructing a composite objective function containing physical constraints includes:

[0019] An adaptive weight matrix is ​​dynamically generated based on the enhanced eigenvector, and the matrix is ​​obtained by normalizing the product of the enhanced eigenvector and the predefined target association matrix through the Softmax function; a rheological stability objective function is defined, which includes the second-order norm distance between the measured pressure change rate and the predicted pressure change rate and the absolute deviation weighted by the average injection speed and the ideal speed; a thermodynamic safety objective function is defined, whose value is the square of the excess of the maximum temperature of the mold cavity over the thermal decomposition temperature of the material minus fifteen degrees Celsius; a quality prediction objective function is defined, which is output by a pre-trained warpage deformation neural network model; the three objective functions are weighted and summed according to the adaptive weight matrix to generate a composite objective function; the target association matrix is ​​obtained by training with historical injection molding defect data, and the ideal speed is set according to the material rheological property library.

[0020] In a preferred embodiment, the parallel execution of global particle swarm search and projected gradient descent operations is specifically as follows:

[0021] Fifty sets of process parameter combinations are initialized as particle swarm populations, and ten generations of evolutionary calculations are iteratively performed. In each generation, the top five solutions with the smallest composite objective function value are retained as the elite solution set. At the same time, starting from the elite solution set, the algorithm moves along the negative gradient direction of the composite objective function within the thermal decomposition temperature constraint boundary, and the step size is dynamically scaled by the inverse of the current gradient norm. After each gradient descent, a projection operation is performed to force the parameters to meet the constraint that the melt temperature does not exceed the thermal decomposition temperature of the material minus ten degrees Celsius. The particle swarm elite solution is compared with the gradient descent solution, and the solution with the smallest composite objective function value is selected as the candidate optimization parameter.

[0022] In a preferred embodiment, the operation of outputting an optimized parameter combination that satisfies the injection molding rheological constraints includes:

[0023] Rheological verification is performed on the candidate optimization parameters, and the melt shear stress is calculated by modifying the Cross-WLF model. If the critical shear stress of the material is exceeded, the injection speed is proportionally reduced. When the melt temperature is in the critical zone between the thermal decomposition temperature of the material minus twelve degrees Celsius and minus ten degrees Celsius, a conservative constraint that the melt temperature does not exceed the thermal decomposition temperature minus fifteen degrees Celsius is added and re-projected. The final output is a process parameter package containing melt temperature, average injection speed, holding pressure and cooling time, where the average injection speed is calculated by dividing the maximum screw displacement by the filling time.

[0024] In a preferred embodiment, in the incremental adversarial reinforcement module, the process of constructing the adversarial training environment is as follows:

[0025] A physical constraint disturbance generator is generated based on the optimized process parameter combination. The generator outputs an environmental disturbance vector based on the material property historical distribution histogram and normal distribution random noise. The environmental disturbance vector includes the material viscosity fluctuation of plus or minus eight percent and the ambient temperature drift of plus or minus three degrees Celsius. The disturbance vector is injected into the melt viscosity and barrel temperature parameters, where the barrel temperature is forced to meet the safety constraint of not exceeding the material thermal decomposition temperature minus ten degrees Celsius after the disturbance. The design reward function is the sliding pass rate divided by the two-norm of the quality index offset vector plus one minus the defect penalty term. The sliding pass rate statistics are the average product pass rates of ten consecutive production cycles. The quality index offset vector includes the dimensional tolerance and the quantitative value of the appearance defect. The defect penalty term takes the value of one when flash or short shot occurs and is zero otherwise.

[0026] In a preferred embodiment, the operation of incremental updating by the proximal policy optimization algorithm includes:

[0027] Load the historical control policy as the initial policy network; construct a policy update constraint to limit the KL divergence between the new policy and the historical policy to no more than 0.01; execute the proximal policy optimization algorithm to calculate the policy loss function, which includes the policy ratio clipping advantage function and the KL divergence penalty term; the policy ratio clipping operation limits the ratio of the new policy to the historical policy to between 0.8 and 1.2; the advantage function uses the generalized advantage estimation algorithm to calculate the discounted sum of the rewards for the next 20 steps; perform adversarial training with a dual-loop structure, the inner loop fixes the policy network to update the perturbation generator to maximize the expected reward, and the outer loop fixes the perturbation generator to update the policy network to maximize the policy loss function; evaluate the policy anti-disturbance ratio after each round of training, and if the anti-disturbance ratio is less than 1.05, fall back to the historical policy;

[0028] The operations of generating and storing the robust control strategy against environmental disturbances are:

[0029] Perform knowledge distillation on the trained policy network to generate a lightweight policy network that satisfies the absolute value of the error between the output action and the original policy network in any state is less than 0.01; store the lightweight policy network in combination with the anti-disturbance ratio evaluation value and the applicable domain tag; the applicable domain tag limits the ambient temperature fluctuation range to plus or minus two degrees Celsius and the material viscosity fluctuation range to plus or minus five percent; add an environmental disturbance vector to the material property historical distribution histogram database; and trigger a policy version switching alarm when the actual production environment parameters exceed the applicable domain.

[0030] The beneficial effects of the present invention are: through the synergistic effect of multiple modules, the intelligence level and stability of the molding process are significantly improved, the system can quickly respond to changes in the physical properties of new materials or new batches, and intelligently generate highly adaptable initial parameter recommendations based on historical data; by integrating physical simulation and real-time sensor data, the dynamic behavior characteristics of the material are accurately captured and enhanced feature vectors are constructed; under the premise of strictly following the thermodynamic safety boundary, the process parameter combination is efficiently optimized to meet complex rheological constraints; at the same time, the anti-disturbance control strategy is continuously evolved based on production data to realize the closed-loop precipitation and self-improvement of process knowledge, and finally achieve rapid and precise regulation in high-frequency switching scenarios, greatly improving the consistency of plastic accessories molding quality and production stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Flow chart of the method of the present invention;

[0032] Figure 2 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0033] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, any person skilled in the art can obtain all other embodiments without creative work, which shall fall into the scope of protection of the present application.

[0034] In the description of the present application, the terms "first", "second" are used only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0035] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can recognize that the present application can be implemented without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope consistent with the principles and characteristics disclosed in the present application.

[0036] Embodiment 1

[0037] The present embodiment provides a plastic accessory forming process parameter optimization and control system as shown in Figures 1-2 The present embodiment provides a plastic accessory forming process parameter optimization and control system as shown in

[0038] The optimal transmission migration module: in response to the input instruction of new material or new batch property, the material molecular characteristics and the corresponding optimal process parameters in the historical process database are called, the optimal transmission mapping calculation is performed according to the molecular characteristic space distribution difference between the new material and the historical material, the pre-adaptation process parameter prediction model facing the new batch is generated, and the initial parameter suggestion is output. In the scene of multi-batch rapid switching of plastic accessories, due to the lack of dynamic modeling ability of the microscopic difference of material properties in the existing system, the switching efficiency is low. The module realizes the cross-material knowledge transfer through the construction of molecular characteristic space and the optimal transmission theory (OT), so as to solve the problem of new batch process parameter initialization;

[0039] Physical residual feature generation module: drives the injection molding physical simulation model based on initial parameter recommendations and outputs the melt front position timing curve and cavity pressure distribution prediction curve ; Real-time acquisition of screw displacement during the first injection cycle , nozzle pressure and cavity temperature signal , the time-domain residual signal of the pressure curve predicted by the physical model and the measured pressure curve is subjected to wavelet packet transform and energy entropy feature extraction, and the original sensor features are fused to generate an enhanced feature vector. This module breaks through the representation limitations of traditional sensor data by converting the physical simulation residual into multi-scale time-frequency features;

[0040] Collaborative Optimization Decision Module: This module receives the enhanced feature vector and constructs a composite objective function that includes physical constraints. It then performs a global particle swarm search and a projected gradient descent operation in the parameter space in parallel. The projected gradient descent operation is strictly limited to the material's thermal decomposition temperature threshold, and outputs an optimized parameter combination that satisfies the injection molding rheological constraints. This module overcomes the limitations of traditional single-objective optimization by integrating adaptive multi-objective aggregation and a dual-track parallel search strategy.

[0041] Incremental adversarial reinforcement module: Continuous injection molding production is performed based on the optimized parameter combination, multi-cycle product quality index data is collected and input into the adversarial training environment, and the proximal strategy optimization algorithm is used to incrementally update the historical strategy to generate a robust control strategy that is resistant to environmental disturbances and store it in the historical process database. This module uses the "virtual disturbance-incremental reinforcement" dual-loop architecture to simulate extreme working conditions in the adversarial training environment, and combines the PPO algorithm constrained by historical strategies to achieve dynamic evolution of the control strategy.

[0042] In this embodiment, it should be specifically explained that in the optimal transmission migration module, the specific operations for performing the optimal transmission mapping calculation are:

[0043] A historical feature space is constructed based on the analytical coefficient of the molecular weight distribution of historical materials and the main component of thermal stability quantification. The construction expression of the molecular feature space of historical materials is:

[0044] ;

[0045] in, The feature vector representing the historical material (historical feature space), Indicates the The distribution coefficient of historical materials, Indicates the The distribution width parameter of historical materials ( and Derived from GPC (gel permeation chromatography) data fitting), Indicates the TGA principal component vectors of historical materials, Indicates the number of historical material types (take an integer and ), Indicates the number index of historical materials ( ), the molecular weight distribution analysis coefficient is obtained by analyzing the gel permeation chromatography data and fitting the molecular weight probability distribution function. The molecular weight distribution function formula is:

[0046] ;

[0047] in, represents the molecular weight distribution function, It represents the molecular weight variable, representing the mass of the polymer chain segment (its data source is GPC chromatography data), and its value range is (0, +∞). It represents the shape parameter of molecular weight distribution, and its value range is (1,10). It represents the number average molecular weight (measured directly by GPC), and its value range is , Represents the gamma function, which is used to normalize the molecular weight distribution function. This formula is used to characterize the molecular weight polydispersity of polymers. The larger the value, the narrower the molecular weight distribution. The main component of thermal stability quantification is the characteristic vector of the material thermogravimetric analysis curve in a specific temperature range; the melt index of the new batch of material is ,density , thermal decomposition enthalpy change The physical property parameters are mapped to the target feature space according to a predefined nonlinear function. The mapping function formula for the new batch of materials is:

[0048] ;

[0049] in, represents the target feature vector for the new batch of materials, Indicates the melt index (the data source is the physical property tester), the value range is , Indicates density (measured by a densitometer), with a value range of , Indicates the thermal decomposition enthalpy change (the data source is TGA-DSC combined instrument), and the value range is , Represents the pre-trained mapping weight (the data source is historical data regression analysis), which is used to adjust the contribution ratio of physical property parameters. In this formula, set Used to compress the melt index to the (0,1) interval, eliminate the dimension difference, and set Used to enhance the density's ability to express nonlinear features, set Used to limit the thermal decomposition enthalpy change to (−1,1) to avoid gradient explosion; calculate the correlation strength matrix between the historical feature space and the target feature space , the matrix elements are composed of historical material process association maps The edge weights are corrected to obtain the optimal transmission matrix from the historical feature space to the target feature space (Calculated by the Sinkhorn algorithm), its objective function includes the Euclidean distance between historical features and target features and the KL divergence regularization term of the process parameter distribution of the two. The adaptive distance calculation formula is:

[0050] ;

[0051] in, Representing historical materials With target material The characteristic differences, Indicates the Characteristic vectors of historical material (historical feature space), represents the Hadamard product (matrix element-by-element multiplication), Represents the historical material-process correlation map The extracted row vector, represents the target feature vector (target feature space) of the new batch of materials, It represents the process similarity regularization coefficient (the weight used to balance the feature distance and process distribution difference), and its value range is [0.1, 0.5]. represents the process parameter distribution divergence, Representing historical materials Process parameter distribution (source domain), Indicates target material The process parameter distribution (target domain) is obtained by iterative calculation through the entropy regularized optimal transfer model. The entropy regularized OT solution formula is:

[0052] ;

[0053] in, Represents the feature distance matrix (given by Calculation generation), represents the matrix transpose operator, represents the optimal transfer matrix to be solved (the weight distribution scheme for migrating historical material characteristics to new batch target materials), Representation matrix and The inner product of ), Indicates the entropy regularization strength (control matrix smoothness), and its value range is (affects the solution speed and accuracy), represents the entropy of the transmission matrix, which is calculated as , Representing historical materials With target material The knowledge transfer intensity ranges from , represents the probability distribution of historical / target features, where is the weight distribution of historical features, is the feature weight distribution of the new batch; based on the optimal transfer matrix, the fully connected layer weight matrix of the historical process parameter prediction model is transferred, where the transfer weight decays exponentially according to the depth of the neural network layer. The fully connected layer transfer formula is:

[0054] ;

[0055] in, represents the weight matrix of the fully connected layer after migration, represents the optimal transmission matrix (transition probability matrix with row and column sums equal to 1), Represents the historical process parameter prediction model The weight matrix of the fully connected layer, It represents the interlayer migration attenuation factor, with a value range of (0,1] and is calculated as follows: , represents the attenuation rate hyperparameter, Represents a diagonal matrix operation, while performing tensor reduction on the convolution kernel weights and combining it with a one-dimensional adaptive filtering operation; the final output is an initial process parameter combination that satisfies the requirement that the melt temperature is at least 10 degrees Celsius lower than the thermal decomposition temperature of the material;

[0056] The specific operations of the migration transformation are:

[0057] A sparsification constraint based on the process association graph is applied to the optimal transfer matrix, retaining only the historical feature transfer paths with the top 30% similarity to the target feature space. A hierarchical attenuation factor is applied to the weight migration process of the fully connected layer. The attenuation coefficient decreases exponentially with the increase in the number of neural network layers, and the attenuation ratio from the input layer to the output layer is set to 1.0 to 0.2. The convolution kernel weight migration adopts a third-order tensor contraction operation, and the contraction direction is performed along the input channel dimension. The contracted tensor is smoothed and corrected by a three-layer one-dimensional adaptive filter. The convolution kernel migration formula is:

[0058] ;

[0059] in, represents the convolution kernel weight after migration, denotes the history process parameter prediction model, denotes the layer convolution kernel weight tensor, denotes the tensor contraction operation (sum along a specific dimension), denotes the 3x1 adaptive filter kernel, denotes the one-dimensional convolution operation; the migrated history process parameter prediction model weight serves as a pre-trained network, which runs forward propagation under the material thermal decomposition safety boundary constraint, expressed as:

[0060] ;

[0061] wherein, denotes the initial parameter estimation value under the constraint, denotes the gradient of the feasible region constraint (ensuring that the parameter update direction does not violate the thermal boundary), denotes the parameter variation amount (its value needs to satisfy the gradient projection condition), denotes the set of migrated model weights (the migrated model weight serves as a pre-trained network, and ), denotes the target feature vector of the new batch of materials, and the safety boundary is defined as the feasible region in which the melt temperature peak does not exceed the material thermal decomposition temperature by 10 degrees Celsius, expressed as:

[0062] ;

[0063] wherein, denotes the actual melt temperature (measured value of the injection barrel), with a value range of , denotes the material thermal decomposition temperature (inherent property of the material property database), with a value range of , denotes the safety temperature redundancy (its value is set to ≥ 10 degrees Celsius); if the propagation result is out of bounds, the projection gradient method is used to pull the parameters back into the feasible region, expressed as:

[0064] ;

[0065] wherein, denotes the initial process parameter after projection optimization, denotes the squared Euclidean distance (minimizing parameter deviation), denotes the feasible region indicator function (if , output , forcing the solution to be within the feasible region, denotes the process parameter vector, outputting the corrected initial process parameter suggestion , the initial process parameter suggestion includes the melt temperature , injection speed , holding pressure Three core process parameters.

[0066] In this embodiment, it should be specifically explained that in the physical residual feature generation module, the specific operation of driving the injection molding physical simulation model is:

[0067] Drive the modified Cross-WLF melt rheological model, input melt temperature, injection speed, and holding pressure parameters, and output the cavity pressure distribution prediction curve. The expression of the driven modified Cross-WLF rheological model is:

[0068] ;

[0069] in, represents the melt shear viscosity, represents the Cross-WLF model constant (intrinsic parameter), which is used to calibrate the sensitivity of viscosity to temperature / shear rate. represents the zero shear viscosity / characteristic stress / power law exponent (which is an intrinsic parameter) and describes the behavior of non-Newtonian fluids. represents the material reference temperature (WLF equation parameter), Indicates the shear rate (melt shear rate), reflects the melt flow deformation rate, and is related to the injection speed. Indicates the real-time temperature of the melt during the flow process (varies with position and time), and characterizes the melt viscosity Distribution with temperature and shear rate Dynamic changes; perform five-layer wavelet packet decomposition on the time domain residual signal (the calculation process of the time domain residual signal is: based on the difference between the cavity pressure distribution prediction curve and the measured nozzle pressure signal, superimpose the acceleration compensation term, where the acceleration compensation coefficient is used to suppress the high-frequency oscillation noise of the model). The formula for calculating the time domain residual is:

[0070] ;

[0071] in, represents the time domain residual signal, represents the acceleration compensation coefficient, represents the cavity pressure prediction curve, Represents the measured nozzle pressure (real-time pressure collected by the piezoelectric sensor), The second-order derivative (acceleration term) of the predicted pressure is expressed by using the Daubechies-4 wavelet basis function to perform five-layer wavelet packet decomposition to generate thirty-two sub-bands. The sub-band nodes with an energy share greater than 1% are selected as the valid node set. The expression is:

[0072] ;

[0073] in, represents the valid node index set (the set of nodes whose energy accounts for more than 1%), represents the sub-band node index, Represents the sub-band node signal ( sub-band wavelet coefficient vector), Indicates the node energy value (sub-band signal The square norm of represents the total energy of all sub-band nodes, 0.01 represents the energy ratio threshold, the energy value of each node signal is calculated and normalized to the energy distribution ratio; the Shannon entropy value is calculated based on the energy distribution ratio of 1, and the time domain mutation index of each sub-band signal is extracted. The Shannon entropy value calculation formula is:

[0074] ;

[0075] in, It represents the Shannon entropy value (the maximum value of the first-order derivative of the sub-band signal), which characterizes the process stability in each monitoring area (e.g. cavity partition) (the higher the entropy, the greater the melt flow fluctuations), Represents the total energy value of the node, which is calculated as , Indicates the The first sub-band node signal discrete points ( ), represents the total number of sub-band node signals. In this formula, the larger the Shannon entropy value, the more dispersed the energy distribution (the Shannon entropy value suddenly changes when abnormal filling occurs);

[0076] The expression of the time domain mutation index is:

[0077] ;

[0078] in, represents the time domain mutation index, represents the time derivative, Indicates the sub-band node signals, represents the absolute value of the sub-band node signal, Represents the time variable, that is, the time axis of the injection molding process. In this formula, high value Pressure spikes during filling Indicates short shot or flash defect, low value Smooth flow The material has good flowability; the node energy value, Shannon entropy value and time domain mutation index are combined into a wavelet packet characteristic vector with 16 dimensions ;

[0079] The operation of fusing the original sensor features to generate the enhanced feature vector is as follows:

[0080] The filling uniformity index is extracted from the screw displacement signal, which is obtained by calculating the absolute deviation integral of the actual displacement speed and the theoretical average speed; the maximum temperature gradient value is extracted from the mold cavity temperature signal; the filling uniformity index, the maximum temperature gradient value, the nozzle pressure rise rate and the pressure curve integral value are combined into the original sensor feature vector, and the expression of the filling uniformity index is as follows:

[0081] ;

[0082] Among them, the filling uniformity index is represented by, the theoretical average filling speed (process set value) is represented by, , the maximum screw displacement is represented by, reflecting the total travel of the melt into the mold cavity in a single injection cycle (measuring method: the magnetic grating ruler records the displacement extreme value of the screw from the starting position to the end of filling in real time), the filling time is represented by, the instantaneous displacement speed of the screw (its data source is the magnetic grating sensor), which quantifies the stability of the melt filling (the smaller the value, the more uniform), reflects the consistency of material flow;

[0083] The expression of the maximum temperature gradient value is as follows:

[0084] ;

[0085] Among them, the mold cavity temperature gradient (maximum temperature gradient value) is represented by, the mold cavity temperature at the moment (mold cavity temperature signal) is represented by, the function of this formula is to characterize the mold thermal equilibrium state (the larger the value, the more uneven the cooling), and to predict the product warping risk;

[0086] The expression of the original sensor feature vector is as follows:

[0087] ;

[0088] Among them, the original sensor feature vector is represented by, the maximum nozzle pressure rise rate is represented by, represents the pressure integral of the filling stage; the original sensor feature vector and the wavelet packet feature vector are linearly weighted and fused according to a preset weight matrix to generate a twenty-two-dimensional enhanced feature vector, and a cavity pressure prediction curve, a measured nozzle pressure curve, a screw displacement curve and a mold cavity temperature curve are stored or output, and the expression of the enhanced feature vector is:

[0089] ;

[0090] wherein, represents a vector splicing operation, represents a preset weight matrix, represents a wavelet packet feature vector, represents an original sensor feature vector, represents an enhanced feature vector; when the highest temperature of the mold cavity exceeds a safety threshold of five degrees Celsius less than the material thermal decomposition temperature, the twenty-second dimension of the enhanced feature vector is forced to be infinite to trigger the melt temperature abnormality protection mechanism (if , represents the highest temperature of the mold cavity, represents the mold cavity temperature, represents the material thermal decomposition temperature, and -5 represents a safety margin threshold (more stringent than -10 in the foregoing, real-time melt-down early warning); wherein the preset weight matrix is obtained by training historical injection defect samples, and the melt temperature safety threshold is updated in linkage with the material thermal decomposition temperature database.

[0091] In the present embodiment, it is specifically necessary to explain that in the collaborative optimization decision module, the operation of constructing a composite objective function containing physical constraints includes:

[0092] An adaptive weight matrix is dynamically generated based on the enhanced feature vector, which is obtained by normalizing the product of the enhanced feature vector and a pre-defined target association matrix through a Softmax function, and the expression of the adaptive weight matrix is:

[0093] ;

[0094] wherein, represents an enhanced feature vector, represents a target association matrix (trained by historical defect data), represents the transpose matrix of , represents an adaptive weight matrix, represents a normalization exponential function; a rheological stability objective function is defined, which includes a second-order norm distance of the measured pressure change rate and the predicted pressure change rate and a weighted item of the absolute deviation of the average injection speed and the ideal speed, and the expression of the rheological stability objective function is:

[0095] ;

[0096] wherein, represents process parameter vector (such as melt temperature, injection speed), represents the measured nozzle pressure rate of change, represents the predicted cavity pressure rate of change, represents the second order norm (measuring pressure change deviation), represents the theoretical average filling speed (process set value), represents the ideal injection speed (set by material rheological property library), represents the speed deviation weight coefficient (range [0.1, 1.0]); define the thermodynamic safety objective function, whose value is the square of the amount of the maximum temperature of the mold cavity exceeding the material thermal decomposition temperature minus fifteen degrees Celsius, and the expression of the thermodynamic safety objective function is:

[0097] ;

[0098] wherein, represents the maximum temperature of the mold cavity, represents the material thermal decomposition temperature, -15 represents the safety margin, represents the penalty activated only when overheating; define the quality prediction objective function, which is output by the pre-trained warping deformation neural network model, and the expression of the quality prediction objective function is:

[0099] ;

[0100] wherein, represents the warping prediction neural network model; the three objective functions are weighted and summed according to the adaptive weight matrix to generate a composite objective function, and the expression of the composite objective function is:

[0101] ;

[0102] wherein, represents the adaptive weight matrix, represents the enhanced feature vector, represents the rheological stability objective function, the thermodynamic safety objective function, and the quality prediction objective function, represents the weighted sum; wherein the target correlation matrix is obtained by training historical injection defect data, and the ideal speed is set according to the material rheological property library;

[0103] The global particle swarm search and the projection gradient descent operation are executed in parallel, and specifically:

[0104] Fifty sets of process parameter combinations are initialized as the particle swarm population, and ten generations of evolutionary calculations are iteratively executed, and the top five solutions with the smallest composite objective function values in each generation are reserved as the elite solution set , and the speed update formula is:

[0105] ;

[0106] in, Represents particles In the The velocity vector of the generation, control parameter search direction and step size, Represents particles In the The position vector of the generation characterizes the current process parameter combination, Indicates the inertia weight, and its value range is , balance global exploration and local development capabilities, represents the learning factor ( ), controls the weight of individual and group experience, Represents a random number uniformly distributed between [0,1], which avoids premature convergence by introducing randomness. Represents particles The historical optimal position of the individual optimal solution is recorded. It represents the global historical optimal position of the population and guides the population to move towards the optimal area. At the same time, starting from the elite solution set, it moves along the negative gradient direction of the composite objective function within the thermal decomposition temperature constraint boundary, and the step size is dynamically scaled by the inverse of the current gradient norm. After each gradient descent, a projection operation is performed to force the parameters to meet the constraint that the melt temperature does not exceed the material thermal decomposition temperature minus ten degrees Celsius. The projection gradient fine search formula is:

[0107] ;

[0108] in, represents the current process parameter vector, represents the updated parameter vector, represents the thermal decomposition temperature constraint projection ( ), Represents the adaptive step size (the step size is dynamically scaled by the inverse of the gradient norm), and the value range is , Represents the gradient of the composite objective function (the negative gradient direction of the composite objective function); compare the particle swarm elite solution with the gradient descent solution, select the solution with the smallest composite objective function value as the candidate optimization parameter, and superimpose the total variation penalty term of parameter change to suppress oscillation. The dynamic solution optimization formula is:

[0109] ;

[0110] in, represents the candidate optimization parameter vector, represents the gradient descent solution set, represents the elite solution set of the particle swarm, denotes the total variation of parameter changes (total variation of parameter change penalty term), denotes the oscillation suppression weight (penalty coefficient of "suppressing oscillation"), whose value is set to 0.3;

[0111] The operation of outputting the optimal parameter combination satisfying the injection rheological constraint includes:

[0112] Performing rheological verification on the candidate optimal parameters, calculating the melt shear stress by modifying the Cross-WLF model, and proportionally reducing the injection speed if the critical shear stress of the material is exceeded, and the Cross-WLF rheological verification formula is:

[0113] ;

[0114] wherein, denotes the calculation of the melt viscosity (melt shear stress), denotes the shear rate, denotes the candidate melt temperature, denotes the wall shear stress; when the melt temperature is in the critical zone of the material thermal decomposition temperature minus twelve degrees Celsius to minus ten degrees Celsius, a conservative constraint that the melt temperature does not exceed the thermal decomposition temperature minus fifteen degrees Celsius is added and re-projected, and the expression is:

[0115] ;

[0116] wherein, denotes the candidate optimal parameter vector, denotes the conservative constraint domain (( )) ; finally output the process parameter package containing the melt temperature, average injection speed, holding pressure and cooling time, and the expression is:

[0117] ;

[0118] wherein, denotes the process parameter package, denotes the holding pressure, denotes the theoretical average injection speed, denotes the melt temperature, denotes the cooling time, wherein the average injection speed is calculated by dividing the maximum screw displacement by the filling time.

[0119] In this embodiment, it is specifically necessary to explain that in the incremental adversarial reinforcement module, the construction process of the adversarial training environment is as follows:

[0120] A physical constraint disturbance generator is generated based on the optimal process parameter combination, which outputs an environmental disturbance vector according to the material physical property historical distribution histogram and normal distribution random noise, and the formula of the physical constraint disturbance generator is:

[0121] ;

[0122] wherein, represents the environmental disturbance vector, represents the disturbance generator, represents the random noise, which is in the range of 0 to 1, represents the process parameter package (optimized process parameter combination), represents the material property distribution histogram (through historical database historical fluctuation statistics); the environmental disturbance vector contains material viscosity positive and negative percentage eight fluctuation and environmental temperature positive and negative three degrees Celsius drift; the disturbance vector is injected into the melt viscosity and the barrel temperature parameter, wherein the barrel temperature disturbance is forced to satisfy the safety constraint of not more than the material thermal decomposition temperature minus ten degrees Celsius, and the melt viscosity formula is:

[0123] ;

[0124] wherein, represents the melt viscosity after disturbance, represents the viscosity disturbance component, represents the nominal melt viscosity (its data source is the material property database);

[0125] The barrel temperature formula is:

[0126] ;

[0127] wherein, represents the barrel temperature after disturbance, represents the temperature disturbance component, represents the barrel temperature;

[0128] The safety constraint expression is:

[0129] ; the design reward function is the sliding qualified rate divided by the two norm of the quality index offset vector plus one minus the defect penalty term, and the expression is:

[0130] ;

[0131] wherein, represents the quality index offset two norm, represents the defect penalty term (flash / short shot event), represents the penalty coefficient (set by process experts), the sliding qualified rate statistics the average of product qualified rate in continuous ten production cycles, the quality index offset vector contains the size tolerance and the appearance defect quantitative value, the defect penalty term takes the value of one when flash or short shot occurs, otherwise it is zero;

[0132] The operation of incrementally updating by the proximal policy optimization algorithm includes:

[0133] The historical control policy is loaded as an initial policy network; a policy update constraint condition is constructed to limit the KL divergence between the new policy and the historical policy to not more than 0.01, and the policy update constraint formula is:

[0134] ;

[0135] wherein, the KL divergence loss is represented by, the KL divergence is represented by, which measures the probability distribution difference between the new policy and the historical policy, the maximum policy offset is represented by, and the value is set to , the new policy network parameter is represented by, the historical policy network is represented by, which is used to constrain the deviation degree of the new policy; the proximal policy optimization algorithm is executed to calculate the policy loss function, which includes the policy ratio clipping advantage function and the KL divergence penalty term; the policy ratio clipping operation limits the ratio of the new policy to the historical policy to be between 0.8 and 1.2, and the proximal policy optimization loss function formula is:

[0136] ;

[0137] wherein, the policy ratio (the ratio of the new policy action probability to the historical policy) is represented by, the advantage function (calculated by the GAE algorithm, used to predict the deviation of future rewards and baseline values) is represented by, the clipping function (policy ratio clipping) is represented by, which is used to limit the policy update amplitude and ensure training stability, the clipping range radius is represented by, and the value is set to 0.2, the KL divergence penalty coefficient is represented by, which is used to balance the policy update and the constraint strength, and the value range is 0.5-2.0; the advantage function uses the generalized advantage estimation algorithm to calculate the future twenty-step reward discount; the double-ring structure is used to perform adversarial training, and the inner ring is fixed policy network the disturbance generator is updated to maximize the expected reward , and the outer ring is fixed disturbance generator to update the policy network to maximize the policy loss function ; after each round of training, the policy disturbance ratio is evaluated, and if the disturbance ratio is less than 1.05, the historical policy is returned, and the disturbance performance evaluation formula is:

[0138] ;

[0139] wherein, represents the disturbance rejection ratio (data source is the reward ratio of the new strategy and the historical strategy), represents the number of disturbance samples, represents the reward function, represents the new strategy (the control strategy output by training), represents the historical strategy network, represents the disturbance sample;

[0140] The operation of generating a robust control strategy against environmental disturbance and storing is as follows:

[0141] Perform knowledge distillation on the trained strategy network to generate a lightweight strategy network that meets the condition that the absolute value of the error between the output action under any state and the original strategy network is less than 0.01, expressed as:

[0142] ;

[0143] wherein, represents the lightweight strategy; the lightweight strategy network is stored in combination with the disturbance rejection ratio evaluation value and the applicable domain label; the applicable domain label limits the environmental temperature fluctuation range to plus or minus 2 degrees Celsius and the material viscosity fluctuation range to plus or minus 5%; the applicable domain expression is:

[0144] ;

[0145] wherein, represents the applicable domain, represents the temperature fluctuation (measured by a temperature and humidity sensor), represents the viscosity fluctuation (measured by rheometer data); an environmental disturbance vector is added to the material physical property historical distribution histogram database ; when the actual production environment parameters exceed the applicable domain, a strategy version switching alarm is triggered (when the environmental parameters exceed the applicable domain, an interrupt signal is sent to the injection molding machine control system).

[0146] It should be noted that in the above embodiments, the description of each embodiment is focused on, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0147] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0148] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0149] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0151] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0152] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A plastic parts molding process parameter optimization and control system, characterized in that: Specifically include: Optimal transmission migration module, physical residual feature generation module, collaborative optimization decision module and incremental adversarial reinforcement module, among which, Optimal Transfer Migration Module: In response to input instructions for new material or new batch physical properties, it calls the material molecular characteristics and corresponding optimal process parameters in the historical process database. Based on the spatial distribution differences of the molecular characteristics of the new material and the historical materials, it performs optimal transfer mapping calculations, generates a pre-adaptive process parameter prediction model for the new batch, and outputs initial parameter recommendations. Physical residual feature generation module: This module drives the injection molding physical simulation model based on the initial parameter recommendations and outputs a melt front position timing curve and a cavity pressure distribution prediction curve. It also collects the screw displacement, nozzle pressure, and cavity temperature signals of the first injection cycle in real time. It performs wavelet packet transformation and energy entropy feature extraction on the time-domain residual signal between the pressure curve predicted by the physical model and the measured pressure curve, and fuses the original sensor features to generate an enhanced feature vector. Collaborative optimization decision module: receives the enhanced feature vector and constructs a composite objective function containing physical constraints, performs global particle swarm search and projected gradient descent operations in parallel in the parameter space, wherein the projected gradient descent operation is strictly limited to move within the material thermal decomposition temperature threshold range, and outputs an optimized parameter combination that meets the injection molding rheological constraints; Incremental adversarial reinforcement module: Continuous injection molding production is performed according to the optimization parameter combination, multi-cycle product quality index data is collected and input into the adversarial training environment, and the proximal strategy optimization algorithm is used to incrementally update the historical strategy to generate a robust control strategy that is resistant to environmental disturbances and store it in the historical process database.

2. The system for optimizing and controlling molding process parameters of plastic parts according to claim 1, characterized in that: In the optimal transmission migration module, the specific operations for performing optimal transmission mapping calculation are: A historical feature space is constructed based on the analytical coefficient of the molecular weight distribution of historical materials and the principal component of thermal stability quantification. The molecular weight distribution analytical coefficient is obtained by analyzing gel permeation chromatography data and fitting the molecular weight probability distribution function, and the principal component of thermal stability quantification is the eigenvector of the material thermogravimetric analysis curve in a specific temperature range; the melt index, density, and thermal decomposition enthalpy change physical parameters of the new batch of materials are mapped to the target feature space according to a predefined nonlinear function; the correlation strength matrix between the historical feature space and the target feature space is calculated, and the elements of this matrix are obtained by modifying the edge weights of the historical material process association map; the optimal transfer matrix from the historical feature space to the target feature space is solved, and its objective function includes the Euclidean distance between the historical features and the target features and the KL divergence regularization term of the process parameter distributions of both, and is calculated iteratively through an entropy regularized optimal transfer model; based on the optimal transfer matrix, a migration transformation is performed on the fully connected layer weight matrix of the historical process parameter prediction model, where the migration weight is exponentially decayed according to the depth of the neural network layer, and the convolution kernel weight is tensor-shrunk and combined with a one-dimensional adaptive filtering operation; the final output is an initial process parameter combination that satisfies the requirement that the melt temperature is at least 10 degrees Celsius lower than the thermal decomposition temperature of the material.

3. The system for optimizing and controlling molding process parameters of plastic parts according to claim 2, characterized in that: The specific operations of the migration transformation are: A sparsification constraint based on the process association graph is imposed on the optimal transfer matrix, and only the historical feature transfer paths with the top 30% similarity to the target feature space are retained; A hierarchical attenuation factor is applied to the weight migration process of the fully connected layer. The attenuation coefficient decreases exponentially with the number of neural network layers, and the attenuation ratio from the input layer to the output layer is set to 1.0 to 0.

2. The convolution kernel weight migration adopts a third-order tensor contraction operation, and the contraction direction is performed along the input channel dimension. The contracted tensor is smoothed and corrected by a three-layer one-dimensional adaptive filter. The weights of the migrated historical process parameter prediction model are used as a pre-trained network, and forward propagation is performed under the material thermal decomposition safety boundary constraint condition. The safety boundary is defined as the feasible domain where the peak temperature of the melt does not exceed the material thermal decomposition temperature by 10 degrees Celsius. If the propagation result exceeds the limit, the projected gradient method is used to pull the parameters back to the feasible domain, and the revised initial process parameter recommended values ​​are output. The initial process parameter recommended values ​​include three core process parameters: melt temperature, injection speed, and holding pressure.

4. The plastic parts molding process parameter optimization and control system according to claim 3, characterized in that: In the physical residual feature generation module, the specific operations of driving the injection molding physical simulation model are: Drive the modified Cross-WLF melt rheology model, input melt temperature, injection speed, and holding pressure parameters, and output a cavity pressure distribution prediction curve. Perform five-layer wavelet packet decomposition on the time domain residual signal using the Daubechies-4 wavelet basis function to generate 32 sub-bands. Nodes in the sub-band with an energy ratio greater than 1% are selected as the valid node set, and the energy value of each node signal is calculated and normalized to the energy distribution ratio. The Shannon entropy value is calculated based on the energy distribution ratio of 1, and the time domain mutation index of each sub-band signal is extracted at the same time; the node energy value, Shannon entropy value and time domain mutation index are combined into a wavelet packet feature vector with dimension 16.

5. The plastic parts molding process parameter optimization and control system according to claim 4, characterized in that: The operation of fusing the original sensor features to generate the enhanced feature vector is: Extracting a filling uniformity index from the screw displacement signal, which is obtained by calculating the integral of the absolute deviation between the actual displacement velocity and the theoretical average velocity; extracting the maximum temperature gradient value from the cavity temperature signal; combining the filling uniformity index, the maximum temperature gradient value, the nozzle pressure rise rate, and the integral value of the pressure curve into the original sensor feature vector; performing linear weighted fusion of the original sensor feature vector and the wavelet packet feature vector according to a preset weight matrix to generate a twenty-two-dimensional enhanced feature vector, which stores or outputs the cavity pressure prediction curve, the measured nozzle pressure curve, the screw displacement curve, and the cavity temperature curve; when the maximum cavity temperature exceeds the safety threshold of the material thermal decomposition temperature minus five degrees Celsius, the twenty-second dimension of the enhanced feature vector is forcibly set to infinity to trigger the melt temperature anomaly protection mechanism; The preset weight matrix is ​​obtained through training of historical injection molding defect samples, and the melt temperature safety threshold is updated in conjunction with the material thermal decomposition temperature database.

6. The plastic parts molding process parameter optimization and control system according to claim 5, characterized in that: In the collaborative optimization decision module, the operations of constructing a composite objective function including physical constraints include: An adaptive weight matrix is ​​dynamically generated based on the enhanced eigenvector, and the matrix is ​​obtained by normalizing the product of the enhanced eigenvector and the predefined target association matrix through the Softmax function; a rheological stability objective function is defined, which includes the second-order norm distance between the measured pressure change rate and the predicted pressure change rate and the absolute deviation weighted by the average injection speed and the ideal speed; a thermodynamic safety objective function is defined, whose value is the square of the excess of the maximum temperature of the mold cavity over the thermal decomposition temperature of the material minus fifteen degrees Celsius; a quality prediction objective function is defined, which is output by a pre-trained warpage deformation neural network model; the three objective functions are weighted and summed according to the adaptive weight matrix to generate a composite objective function; the target association matrix is ​​obtained by training with historical injection molding defect data, and the ideal speed is set according to the material rheological property library.

7. The plastic parts molding process parameter optimization and control system according to claim 6, characterized in that: The parallel execution of global particle swarm search and projected gradient descent operations is specifically as follows: Fifty sets of process parameter combinations are initialized as particle swarm populations, and ten generations of evolutionary calculations are iteratively performed. In each generation, the top five solutions with the smallest composite objective function value are retained as the elite solution set. At the same time, starting from the elite solution set, the algorithm moves along the negative gradient direction of the composite objective function within the thermal decomposition temperature constraint boundary, and the step size is dynamically scaled by the inverse of the current gradient norm. After each gradient descent, a projection operation is performed to force the parameters to meet the constraint that the melt temperature does not exceed the thermal decomposition temperature of the material minus ten degrees Celsius. The particle swarm elite solution is compared with the gradient descent solution, and the solution with the smallest composite objective function value is selected as the candidate optimization parameter.

8. The plastic parts molding process parameter optimization and control system according to claim 7, characterized in that: The operation of outputting an optimized parameter combination that satisfies the injection molding rheological constraints includes: Rheological verification is performed on the candidate optimization parameters, and the melt shear stress is calculated by modifying the Cross-WLF model. If the critical shear stress of the material is exceeded, the injection speed is proportionally reduced. When the melt temperature is in the critical zone between the thermal decomposition temperature of the material minus twelve degrees Celsius and minus ten degrees Celsius, a conservative constraint that the melt temperature does not exceed the thermal decomposition temperature minus fifteen degrees Celsius is added and re-projected. The final output is a process parameter package containing melt temperature, average injection speed, holding pressure and cooling time, where the average injection speed is calculated by dividing the maximum screw displacement by the filling time.

9. The plastic parts molding process parameter optimization and control system according to claim 8, characterized in that: In the incremental adversarial reinforcement module, the process of constructing the adversarial training environment is as follows: A physical constraint disturbance generator is generated based on the optimized process parameter combination. The generator outputs an environmental disturbance vector based on the material property historical distribution histogram and normal distribution random noise. The environmental disturbance vector includes the material viscosity fluctuation of plus or minus eight percent and the ambient temperature drift of plus or minus three degrees Celsius. The disturbance vector is injected into the melt viscosity and barrel temperature parameters, where the barrel temperature is forced to meet the safety constraint of not exceeding the material thermal decomposition temperature minus ten degrees Celsius after the disturbance. The design reward function is the sliding pass rate divided by the two-norm of the quality index offset vector plus one minus the defect penalty term. The sliding pass rate statistics are the average product pass rates of ten consecutive production cycles. The quality index offset vector includes the dimensional tolerance and the quantitative value of the appearance defect. The defect penalty term takes the value of one when flash or short shot occurs and is zero otherwise.

10. The plastic parts molding process parameter optimization and control system according to claim 9, characterized in that: The operation of incremental updating through the proximal policy optimization algorithm includes: Load the historical control policy as the initial policy network; construct a policy update constraint to limit the KL divergence between the new policy and the historical policy to no more than 0.01; execute the proximal policy optimization algorithm to calculate the policy loss function, which includes the policy ratio clipping advantage function and the KL divergence penalty term; the policy ratio clipping operation limits the ratio of the new policy to the historical policy to between 0.8 and 1.2; the advantage function uses the generalized advantage estimation algorithm to calculate the discounted sum of the rewards for the next 20 steps; perform adversarial training with a dual-loop structure, the inner loop fixes the policy network to update the perturbation generator to maximize the expected reward, and the outer loop fixes the perturbation generator to update the policy network to maximize the policy loss function; evaluate the policy anti-disturbance ratio after each round of training, and if the anti-disturbance ratio is less than 1.05, fall back to the historical policy; The operations of generating and storing the robust control strategy against environmental disturbances are: Perform knowledge distillation on the trained policy network to generate a lightweight policy network that satisfies the absolute value of the error between the output action and the original policy network in any state is less than 0.01; store the lightweight policy network in combination with the anti-disturbance ratio evaluation value and the applicable domain tag; the applicable domain tag limits the ambient temperature fluctuation range to plus or minus two degrees Celsius and the material viscosity fluctuation range to plus or minus five percent; add an environmental disturbance vector to the material property historical distribution histogram database; and trigger a policy version switching alarm when the actual production environment parameters exceed the applicable domain.

Images