Intelligent collaborative control method and system for whole process of complex appearance part forming

Abstract

The application discloses a full-process intelligent collaborative control method and system for complex appearance part forming, and relates to the field of intelligent control, and comprises the following steps: generating an appearance defect sample set by constructing a multi-physics field injection simulation model, training a graph attention network to predict the defect probability of a feature area, and using an adversarial domain adaptive technology to migrate the simulation model to an online inference process parameter-appearance mapping model; in actual production, the appearance risk is dynamically evaluated by real-time acquisition of process data, and the injection, pressure maintaining and temperature parameters are adjusted by a collaborative controller to suppress defects; after demolding, defects are identified by image recognition and incremental samples are collected, and the mapping model and control logic are fine-tuned online. The application has the advantages that: the graph attention network is trained by generating defect samples through multi-physics field simulation, and is adapted to actual working conditions; the defect risk is evaluated in combination with real-time sensing data; injection process parameters are collaboratively controlled; appearance defects are effectively suppressed; and the forming yield and process change efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and in particular to a method and system for intelligent collaborative control of the entire process of forming complex appearance parts. Background Technology

[0002] The demand for complex curved surface parts in the automotive, consumer electronics, and home appliance industries continues to surge. These parts often have uneven wall thickness, multiple ribs, and multiple gate structures. The molding process involves strong coupling characteristics of multiple fields, including melt flow, heat transfer, and stress deformation. Strict control is required for appearance defects such as shrinkage marks, weld lines, and warpage, which places extremely high demands on the consistency of the entire molding process.

[0003] Current mainstream injection molding control methods have significant shortcomings: simulation modeling often uses simplified single-physics models, failing to perform fine mesh and boundary layer solutions for key appearance areas such as highlights and textures. This results in limited prediction accuracy for microscopic appearance defects such as weld lines and flow marks. Furthermore, the simulation domain has inherent distributional deviations from actual production conditions, making it impossible to directly support precise online control. Process sensing relies heavily on single-point, low-frequency temperature and pressure monitoring, lacking multi-field synchronous acquisition and dynamic risk classification mechanisms for appearance feature areas. Defect identification generally lags behind the molding cycle, allowing only post-mortem remediation rather than proactive intervention. Control strategies are mostly single-parameter independent adjustment modes, failing to achieve multi-dimensional collaborative optimization of injection speed, multi-stage holding pressure, and zoned mold temperature, making it difficult to simultaneously suppress multiple types of appearance defects. Moreover, they generally lack online incremental learning loops that provide feedback from finished product appearance data, making it unable to adapt to material batch fluctuations, mold aging, and other operational drifts. This leads to insufficient long-term production yield stability, and mold changeover and adjustment heavily rely on manual experience, resulting in long setup cycles and difficulty in meeting the high-precision and high-efficiency molding requirements of complex appearance parts. Summary of the Invention

[0004] To improve existing methods and systems, this paper presents a fully intelligent collaborative control method and system for the molding of complex appearance parts. This method uses multi-physics simulation to generate defect sample training graphs and attention networks, adapts them to actual working conditions, combines real-time perception data to assess defect risks, and collaboratively controls multiple injection molding process parameters to effectively suppress appearance defects and improve molding yield and process changeover efficiency.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A comprehensive intelligent collaborative control method for the molding process of complex appearance parts includes:

[0007] Obtain the 3D model and appearance quality specifications of complex appearance parts, extract the highlight area and texture area as key appearance feature areas, construct a multiphysics injection molding simulation model that couples mold thermal balance, melt filling flow and cooling shrinkage, generate an appearance defect sample set through simulation, and mark the defect type and location.

[0008] Based on the sample set of appearance defects, a graph attention network is trained with cavity mesh diagram structure data and process parameters as input to obtain the probability distribution of defect occurrence in each feature region. An execution characteristic simulator is established using historical data of injection molding machine actuator response. The simulation domain graph attention network is transferred to an online inferable process parameter-appearance mapping model through adversarial domain adaptation technology.

[0009] In the actual injection molding cycle, the mold temperature field, cavity pressure curve and melt flow front position are collected in real time. The collected information is input into the process parameter-appearance mapping model to dynamically assess the risk level of weld lines, flow marks and shrinkage marks in each appearance feature area of ​​the current cycle.

[0010] Based on the risk level of shrinkage marks, the collaborative controller synchronously generates injection speed curve parameter adjustment instructions, multi-segment holding pressure and time correction instructions, and mold zone temperature setting value offset instructions. Through feedforward-feedback composite control, the melt front speed, interface temperature, and cooling rate are maintained within the process window that suppresses appearance defects.

[0011] Multi-angle appearance images of the demolded product are acquired, and an instance segmentation network is used to identify the defect type and spatial location. The defect information and the corresponding process parameter sequence are combined into incremental training samples. The process parameter-appearance mapping model is fine-tuned online, and the mapping logic in the collaborative controller is updated.

[0012] Preferably, the construction of a multiphysics injection molding simulation model coupling mold thermal balance, melt filling flow, and cooling shrinkage, and the generation of an appearance defect sample set through simulation specifically includes:

[0013] Import the 3D model into the simulation platform, refine the local mesh in the appearance feature area, and set a prism boundary layer to solve the near-wall melt temperature gradient.

[0014] Construct a mold thermal model that includes conformal cooling water channels, apply zoned temperature setting boundaries to the mold, and set the initial temperature of each zone according to the distribution of appearance feature areas;

[0015] The equations of incompressible non-isothermal flow and material state are coupled to solve the shear rate field, temperature field and free surface evolution of the melt during the filling process, and the volume shrinkage and residual stress field are calculated during the pressure holding and cooling stage.

[0016] In post-processing, weld lines, flow marks, and shrinkage marks are marked by combining the weld confluence angle, peak surface shear stress, and local shrinkage rate, generating a structured defect sample dataset with process parameter labels.

[0017] Preferably, the training of the graph attention network and the transfer learning specifically include:

[0018] The simulation mesh is transformed into graph data according to the spatial connection relationship. The node attributes include node coordinates, temperature, pressure, shear rate and wall heat flux, and the edge attributes are normalized flow direction vector and distance between nodes.

[0019] A prediction network consisting of multiple stacked graph attention layers is constructed, taking node features and global process condition vectors as input, and outputting the defect category probability of each node;

[0020] A long short-term memory network is trained using historical operating data of an injection molding machine as an execution characteristic simulator. The simulator predicts the timing of physical actions that actually reach the mold based on ideal process instructions.

[0021] An adversarial domain adaptive training mechanism is introduced, using a small amount of labeled data from actual production as the target domain. The difference in feature distribution between the simulation domain and the actual domain is reduced by using a gradient inversion layer, and the simulation domain prediction network is transferred to a process parameter-appearance mapping model.

[0022] Preferably, the adversarial domain adaptive training mechanism specifically includes:

[0023] Construct a domain discriminator network, using the output of the intermediate layer of the feature extraction network as input, to determine whether the data originates from the simulation domain or the actual domain;

[0024] By alternating training to minimize defect prediction loss and maximize domain discrimination loss, the feature representation is made domain invariant.

[0025] Given the extremely limited number of actual domain samples, data augmentation techniques are employed to expand the actual defect samples through geometric and color transformations, thereby assisting the domain adaptive process.

[0026] Preferably, the real-time acquisition of mold temperature field, cavity pressure curve, and melt flow front position, and the input of the acquired information into the process parameter-appearance mapping model to dynamically assess the risk level of weld lines, flow marks, and shrinkage marks in each appearance feature area of ​​the current cycle specifically includes:

[0027] Temperature measurement holes are opened on the back of each appearance feature area of ​​the mold cavity surface, and multiple pairs of K-type armored thermocouples are embedded. High sampling rate modules are used to collect data synchronously to form a time series temperature distribution tensor.

[0028] A piezoelectric pressure sensor is flush-mounted at the end of the cavity filling to measure pressure changes in real time and extract the pressure rise rate and pressure peak as features.

[0029] By opening an infrared window on the fixed side of the mold and setting up a near-infrared high-speed camera, real-time images of the melt front are captured, and the arrival time and instantaneous velocity of the front are obtained through sub-pixel edge extraction and spatiotemporal tracking algorithms.

[0030] The temperature distribution tensor, pressure characteristics, and frontal velocity information are input into the process parameter-appearance mapping model to calculate the dynamic risk index of each feature region, and the perception strategy is updated based on the risk index.

[0031] Preferably, the step of updating the perception strategy based on the risk index specifically includes:

[0032] A risk index grading threshold is set. When the risk index of a certain feature area exceeds the medium risk threshold, the sampling frequency of the temperature sensor near the area is dynamically increased, and the optical imaging probe is activated to scan the local solidified layer morphology.

[0033] The encrypted data is fed into the model in real time for secondary evaluation, refining the defect boundaries and risk trends.

[0034] The time step of the subsequent regulation cycle will be adjusted based on the refined risk trends.

[0035] Preferably, the step of synchronously generating injection speed curve parameter adjustment instructions, multi-segment holding pressure and time correction instructions, and mold zone temperature setpoint bias instructions by the collaborative controller according to the shrinkage mark risk level, and maintaining the melt front velocity, interface temperature, and cooling rate within the process window that suppresses appearance defects through feedforward-feedback composite control specifically includes:

[0036] Receive the risk level vector and main cause identifier of each appearance feature area output by the defect risk assessment module, and construct a collaborative optimization problem with the goal of minimizing risk. The decision variables include the injection speed curve control point, the holding pressure timing node and the mold partition temperature offset.

[0037] A model-based predictive control architecture is adopted, using the process parameter-appearance mapping model as the predictive model. Multi-objective optimization is solved in the rolling time domain to generate a coordinated control sequence that satisfies the physical constraints of the actuators. The first control command is then issued to the injection servo motor, the pressure holding servo valve, and the zoned mold temperature actuator.

[0038] In the next control cycle, the deviation between the actual melt response and the output of the prediction model is calculated using real-time sensing data, and the linearization offset term in the prediction model is corrected in real time.

[0039] Preferably, the solution process for the predictive control architecture specifically includes:

[0040] Using the current real-time status and process parameter-appearance mapping model, with the injection speed curve Bezier control point increment, holding pressure inflection point increment, and mold temperature zone increment as decision variables, a cost function including risk index minimization and energy consumption weighting is established.

[0041] The particle swarm optimization algorithm is used to quickly search for the optimal coordination sequence for several future control cycles within the constraint space, and a partial solution space is pre-defined by explicit parameterization rules to accelerate online solution.

[0042] The optimal instructions in the first cycle are synchronously sent to each actuator via the EtherCAT bus, and the actuator characteristics simulator is used to perform one-order advance compensation to eliminate instruction latency.

[0043] Furthermore, a full-process intelligent collaborative control system for the molding of complex appearance parts is proposed, including:

[0044] Simulation and Sample Generation Module: Based on the 3D model and quality specifications, a multiphysics simulation model coupling heat, fluid and solid is constructed to simulate the injection molding process and generate a batch of structured sample datasets of defect types and locations;

[0045] Mapping Model Construction and Transfer Module: Use simulation samples to train a graph attention network to predict defects, and combine it with historical data of injection molding machines to build an execution characteristic simulator. Through adversarial domain adaptation technology, the simulation-trained model is transferred and adapted into a model that infers the mapping relationship between process parameters and appearance quality in actual production.

[0046] Perception and Dynamic Assessment Module: Collects physical field data in real time and inputs it into the process parameter-appearance mapping model to dynamically assess the risk level of appearance defects in each feature area during the current production cycle;

[0047] Collaborative optimization control module: Based on the dynamic risk assessment results, it optimizes and generates coordinated control commands for injection speed curves, multi-segment holding pressure parameters, and zoned mold temperature settings. Through feedforward-feedback composite control, it stabilizes the process state within the window for suppressing defects.

[0048] Online learning and model update module: performs image acquisition and instance segmentation on demolded products, identifies defect information, combines it with corresponding process parameters to form incremental samples, fine-tunes the process parameter-appearance mapping model online, and synchronously updates the mapping logic of the collaborative controller;

[0049] Process execution and closed-loop interface module: It sends the optimization instructions of the collaborative controller to the field actuators and collects their response feedback to form the basic data link for closed-loop control.

[0050] Compared with the prior art, the advantages of the present invention are:

[0051] A simulation system coupled with multiphysics is constructed to generate defect samples in batches. By combining graph attention networks and adversarial domain adaptation technology, the simulation model can be transferred to the production scenario with a small number of actual samples, solving the problem of accurate modeling caused by the scarcity of actual defect samples. By collecting mold temperature field, cavity pressure and melt front information in real time through multi-source sensors, the defect risk of each feature area is dynamically assessed and the perception strategy is adaptively adjusted to achieve advance prediction of appearance defect risk. A model predictive control architecture is adopted to synchronously regulate injection speed, multi-stage holding pressure and zoned mold temperature. Combined with actuator characteristic compensation and feedforward-feedback composite mechanism, the process is accurately constrained within the defect suppression window. At the same time, an online incremental learning closed loop is built based on finished product inspection data to continuously iterate the model and control logic. This can effectively adapt to working condition drift, significantly improve the molding yield and process stability of complex high-gloss and textured appearance parts, and greatly shorten the product changeover and mold adjustment cycle. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the intelligent collaborative control method for the entire process of forming complex appearance parts proposed in this invention.

[0053] Figure 2 This is a schematic diagram of the multiphysics injection molding simulation model proposed in this invention;

[0054] Figure 3 This is a schematic diagram of the training and transfer of the graph attention network proposed in this invention. Detailed Implementation

[0055] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0056] A fully intelligent collaborative control system for the molding of complex appearance parts, including:

[0057] Simulation and Sample Generation Module: Based on the 3D model and quality specifications, a multiphysics simulation model coupling heat, fluid and solid is constructed to simulate the injection molding process and generate a batch of structured sample datasets of defect types and locations;

[0058] Mapping Model Construction and Transfer Module: Use simulation samples to train a graph attention network to predict defects, and combine it with historical data of injection molding machines to build an execution characteristic simulator. Through adversarial domain adaptation technology, the simulation-trained model is transferred and adapted into a model that infers the mapping relationship between process parameters and appearance quality in actual production.

[0059] Perception and Dynamic Assessment Module: Collects physical field data in real time and inputs it into the process parameter-appearance mapping model to dynamically assess the risk level of appearance defects in each feature area during the current production cycle;

[0060] Collaborative optimization control module: Based on the dynamic risk assessment results, it optimizes and generates coordinated control commands for injection speed curves, multi-segment holding pressure parameters, and zoned mold temperature settings. Through feedforward-feedback composite control, it stabilizes the process state within the window for suppressing defects.

[0061] Online learning and model update module: performs image acquisition and instance segmentation on demolded products, identifies defect information, combines it with corresponding process parameters to form incremental samples, fine-tunes the process parameter-appearance mapping model online, and synchronously updates the mapping logic of the collaborative controller;

[0062] Process execution and closed-loop interface module: It sends the optimization instructions of the collaborative controller to the field actuators and collects their response feedback to form the basic data link for closed-loop control.

[0063] See Figure 1 As shown, the intelligent collaborative control method for the entire process of forming complex appearance parts includes:

[0064] Obtain the 3D model and appearance quality specifications of complex appearance parts, extract the highlight area and texture area as key appearance feature areas, construct a multiphysics injection molding simulation model that couples mold thermal balance, melt filling flow and cooling shrinkage, generate an appearance defect sample set through simulation, and mark the defect type and location.

[0065] Based on the sample set of appearance defects, a graph attention network is trained with cavity mesh diagram structure data and process parameters as input to obtain the probability distribution of defect occurrence in each feature region. An execution characteristic simulator is established using historical data of injection molding machine actuator response. The simulation domain graph attention network is transferred to an online inferable process parameter-appearance mapping model through adversarial domain adaptation technology.

[0066] In the actual injection molding cycle, the mold temperature field, cavity pressure curve and melt flow front position are collected in real time. The collected information is input into the process parameter-appearance mapping model to dynamically assess the risk level of weld lines, flow marks and shrinkage marks in each appearance feature area of ​​the current cycle.

[0067] Based on the risk level of shrinkage marks, the collaborative controller synchronously generates injection speed curve parameter adjustment instructions, multi-segment holding pressure and time correction instructions, and mold zone temperature setting value offset instructions. Through feedforward-feedback composite control, the melt front speed, interface temperature, and cooling rate are maintained within the process window that suppresses appearance defects.

[0068] Multi-angle appearance images of the demolded product are acquired. An instance segmentation network is used to identify defect types and spatial locations. Defect information is combined with corresponding process parameter sequences to form incremental training samples. The process parameter-appearance mapping model is fine-tuned online, and the mapping logic in the collaborative controller is updated. The specific construction process is as follows: After demolding, product appearance images are acquired through a multi-angle camera system. After detection by an instance segmentation network (using ResNet-50 as the backbone, fusing feature pyramids, and outputting defect categories and pixel-level masks), the 2D mask contour is back-projected into 3D space using calibrated camera-fixture extrinsic parameters and registered with the product CAD model. The defect coordinates are mapped to the corresponding appearance feature regions to obtain structured defect labels. It contains defect category c and a set of grid node indices. Simultaneously, the complete process parameter timing data recorded during this injection molding cycle, including the injection speed profile, is retrieved from the production execution database. Holding pressure curve Temperature settings for each zone of the mold and the cavity pressure sensor records The above process data are aligned according to the same time base, and the feature sequences of the filling section, holding section, and cooling section are extracted through a sliding window to form a global process state vector. Subsequently, and Along with the period number Encapsulated as an incremental sample tuple All tuples are accumulated in production time sequence to form an incremental training set. This structured approach ensures a precise correlation between the spatial distribution of defects and process dynamics, providing a clear input-output mapping for subsequent fine-tuning.

[0069] See Figure 2 As shown, the 3D model of the target complex appearance part is imported into the injection molding simulation platform. Unstructured tetrahedral or hexahedral meshes are used for overall meshing, and local mesh refinement is performed for key appearance feature areas such as highlight and texture regions. Simultaneously, prism boundary layer meshes are generated at the cavity wall to accurately solve for the temperature gradient and shear rate of the melt near the wall. The prism layer thickness is set according to the estimated thermal boundary layer thickness to ensure that the y+ value corresponding to the first layer mesh node on the wall is in the viscous sublayer region.

[0070] A mold thermal model incorporating conformal cooling water channels is constructed. Based on the distribution and thermal sensitivity differences of the appearance feature areas, the mold cavity surface is divided into multiple independent temperature control zones. Temperature setting boundaries are applied to each zone, and the initial temperature is preset with different cooling or heating strategies based on the location of the feature areas.

[0071] During the filling and holding cooling stages, the incompressible non-isothermal flow equations and the material state equations are solved in a coupled manner. The governing equations for melt flow are as follows, where the momentum equation considers pressure gradient, viscous forces, and inertial forces, and the energy equation couples shear heat generation and heat conduction:

[0072]

[0073]

[0074] Where ρ is the melt density, v is the velocity vector, p is the pressure, and η is the melt viscosity. γ is the specific heat capacity, T is the temperature, k is the thermal conductivity, and γ˙ is the shear rate.

[0075] Melt viscosity is described using the Cross-WLF model, and its expression is as follows:

[0076]

[0077] Among them, zero shear viscosity Depends on temperature and pressure, the formula is:

[0078]

[0079] in, For temperature, For pressure, Reference temperature Zero shear viscosity at the following conditions , Here are the material constants for WLF. The pressure coefficient, Let be the critical shear stress, and n be the power-law exponent. This model accurately captures the rheological behavior of the melt under shear-thinning effects during the filling process.

[0080] During the pressure holding and cooling stage, the local volumetric shrinkage caused by the change in specific volume is calculated based on the material's PVT equation of state. The volumetric shrinkage rate is mapped to the mesh nodes, and the residual stress distribution is calculated using continuum mechanics. Volumetric shrinkage rate It can be estimated using the following formula:

[0081]

[0082] In the formula, V is the specific volume obtained based on the material's PVT curve. This is the demolding temperature. Atmospheric pressure , The average temperature and pressure during the pressure holding phase. Larger areas are more prone to shrinkage marks.

[0083] After simulation, each mesh node is marked with a defect during post-processing. Weld lines are determined based on the angle between the leading edges of the two melt streams; when the angle is below a critical value, it is marked as a visible weld line risk node. The critical value is an angle equal to... Flow mark nodes are identified based on whether the surface shear stress peak experienced by the node during the filling stage exceeds a material-related threshold, which is 0.8 times the tensile yield stress of the material. Shrinkage mark nodes are determined by combining the local shrinkage rate being 1.2 times higher than the global average shrinkage rate with the surface depression displacement field. Finally, the defect label of each grid node is combined with the corresponding process parameter vector, which includes the injection speed curve, holding pressure sequence, mold zone temperature setting, and injection time, to generate a structured defect sample dataset with process parameter labels.

[0084] See Figure 3 As shown, the 3D mesh generated by injection molding simulation is transformed into graph structure data G=(V,E) according to the spatial topological connection relationship, where each node in the node set V... attribute vector Includes normalized physical quantities such as node spatial coordinates, peak temperature during the filling stage, peak pressure, peak shear rate, and average wall heat flux; each edge in the edge set E The attributes include the normalized flow direction vectors between nodes and the Euclidean distance. The node attributes and edge attributes constitute the initial input of the model.

[0085] A defect prediction network consisting of multiple stacked graph attention layers is constructed. In each graph attention layer, node i's attention weights are assigned to its neighboring nodes j. The attention coefficients are calculated through a joint mapping of node and edge features. Normalized attention coefficients are obtained after applying LeakyReLU activation and softmax normalization. :

[0086]

[0087] In the formula, , Let i be the feature vectors of nodes i and j. Let W be the edge feature vector. Let be the linear transformation weight matrices for node features and edge features, respectively, and let a be the attention weight vector. This represents a vector concatenation operation. (i) represents the neighborhood set of node i. Features of neighboring nodes are weighted and aggregated based on attention coefficients, and their expressive power is enhanced through multi-head attention concatenation. After multiple stacking layers, the defect category probability distribution of each node is output. The network uses weighted cross-entropy as the defect prediction loss. Optimization was carried out to address the issue of imbalanced defect category samples.

[0088] By utilizing the ideal process instruction sequence recorded in the injection molding machine's historical operation database and the physical action timing sequence collected by actual sensors, a long short-term memory network is trained as an execution characteristic simulator. This simulator takes the desired injection speed curve, holding pressure, and mold temperature setting sequence as inputs, and predicts the actual time sequence of physical quantities reaching the mold cavity after being affected by factors such as actuator hysteresis and dynamic response, providing a feedforward compensation basis for subsequent coordinated control.

[0089] To transfer graph attention networks trained in the simulation domain to the real production environment, an adversarial domain adaptive training mechanism is introduced. A gradient inversion layer is inserted between the feature extraction network and the defect classifier, and the node embedding features output from the intermediate layer are fed into the domain discriminator network D. The domain discriminator takes the node embedding vector as input and outputs the binary classification probability of whether the node data originates from the simulation domain or the real domain. The domain discrimination loss is calculated using binary cross-entropy. .

[0090] Total loss function for adversarial domain adaptation The defect prediction loss and the domain discrimination loss are weighted and combined using the hyperparameter λ.

[0091]

[0092] The gradient reversal layer implements an identity mapping during forward propagation and multiplies the gradient from the domain discriminator by a negative constant -λ during backpropagation. This ensures that the parameter update direction of the feature extraction network simultaneously minimizes the defect prediction loss and maximizes the domain discrimination loss, thereby learning a domain-invariant feature representation. The training process is performed alternately: first, the parameters of the feature extraction network are fixed to optimize the domain discriminator, and then the parameters of the domain discriminator are fixed to optimize both the feature extraction network and the defect classifier.

[0093] To address the issue of extremely limited labeled defect samples in actual production, a physical-based illumination color transformation and geometric affine transformation are employed to augment the actual defect images, expanding the target domain sample size to assist the domain adaptive training process. After transfer training, a process parameter-appearance mapping model is obtained, which can directly infer the probability distribution of defects in each appearance feature region online based on real-time acquired multimodal physical quantities during the actual injection molding cycle.

[0094] Temperature measurement holes are opened on the back of each appearance feature area facing the mold cavity, and multiple pairs of K-type armored thermocouples are embedded therein. A high sampling rate synchronization module is used for millisecond-level multi-channel temperature acquisition. Spatial interpolation is performed on the discrete temperature values ​​measured at the temperature measurement points to reconstruct the temperature distribution tensor of the cavity surface. The interpolation can be based on a radial basis function method, constructing a continuous temperature field with the coordinates of each temperature measurement point and the corresponding real-time temperature to ensure the accuracy of temperature representation of feature areas such as highlights and textures.

[0095] Meanwhile, a piezoelectric pressure sensor is installed flush with the flow channel at the end of the cavity filling to measure the cavity pressure p(t) in real time, and the pressure rise rate and pressure peak value are extracted as characteristic quantities for evaluating the filling state.

[0096] To capture the dynamics of the melt flow front, an infrared window was opened on the fixed side of the mold, and a near-infrared high-speed camera was installed to capture a sequence of images of the melt front. Subpixel edge detection was used for each frame to obtain the contour of the flow front, and the arc length position s(t) of the front was tracked along the flow direction. The instantaneous velocity of the melt front was calculated using the position difference between consecutive frames.

[0097]

[0098] In the formula, Δt is the camera frame interval. This velocity reflects the flow stability and is an important input for predicting defects such as flow marks.

[0099] The temperature distribution tensor obtained within the same period is compressed into the average temperature and temperature gradient of each appearance feature region, along with pressure features and frontal velocity sequence. Spatiotemporal alignment and normalization are performed to form a real-time state vector. This vector is then input into the transferred process parameter-appearance mapping model, which outputs the defect category probability for each feature region. A dynamic risk index for region k is defined. The maximum probability of all defect categories in this region:

[0100]

[0101] Where C is the set of defect categories, The model provides the feature embedding for region k, and g is the global process condition vector. A higher value indicates a greater risk of defects occurring.

[0102] After obtaining the risk index, implement adaptive perception updates. Set low, medium, and high risk thresholds. When the risk index of a certain area exceeds the medium-risk threshold... At the same time, dynamically increase the sampling frequency of nearby thermocouples in that area. The adjusted sampling frequency... Determined by the following formula:

[0103]

[0104] In the formula, The default sampling frequency is set to γ, which is the gain coefficient. Simultaneously, a preset optical imaging probe is activated to scan the local solidification layer thickness in the region, acquiring the solidification front morphology and the growth state of the solidified layer. The densely collected temperature and solidification layer data are fed into the model in real time for secondary inference, refining the defect boundary and calculating the risk index change rate. The control cycle step size for subsequent collaborative control is adaptively adjusted based on the risk trend.

[0105] Risk index of each appearance feature area The risks are aggregated into a global risk vector R. The process steps at which risks arise are identified based on the main contributing factors, and high-risk areas are prioritized. A collaborative optimization problem is constructed with the goal of minimizing risk. Decision variables include the location of control points on the injection speed curve, the inflection points of the holding pressure sequence, and the temperature offsets of each mold zone. The injection speed is parameterized using a Bézier curve, with the speed-displacement profile represented by a control point vector; the holding pressure is described by piecewise linear inflection point vectors and corresponding pressure values; and the mold temperature offset is represented by a zone vector. These three types of decision variables are integrated into the optimization variables. .

[0106] A model predictive control architecture is adopted, using the process parameter-appearance mapping model as the predictive model to solve the multi-objective coordinated control problem in the rolling time domain. Starting from the current cycle number k, the future... The evolution of defect risk over a control cycle is defined by the cost function J, which is a weighted sum of a risk index term and a control action smoothing term.

[0107]

[0108] In the formula, M represents the total number of appearance feature regions. Let i be the risk index of the (k+j+1)th period region i predicted at the current time k. This refers to the weighting coefficient for regional importance. To control the incremental sequence, Q is a positive definite weight matrix. To predict the time domain, To control the time domain.

[0109] The optimization process must satisfy actuator physical constraints, including injection speed limits, holding pressure change rates, and mold temperature adjustment rates. Within the rolling time-domain optimization framework, the currently sensed state variables are substituted into the prediction model for recursion. A particle swarm optimization algorithm is used to search for the control sequence that minimizes the cost function J within the constraint space. To accelerate online solving, a partial solution space boundary is predefined based on the causal markers of the appearance feature regions, and the initial particle population is biased towards the neighborhood of the historical optimal solution.

[0110] After obtaining the optimal coordinated control sequence, only the first control cycle command is issued to the injection servo motor, the pressure holding servo valve, and the zoned mold temperature actuator.

[0111] During the instruction issuance phase, a constructed execution characteristic simulator is used to perform one-order advance compensation on the instructions. Based on the executor response delay time output by the simulator, the instruction timing is compensated in advance, so that the actual physical action time experienced by the melt is consistent with the expectation.

[0112] In the next control cycle, the deviation between the actual melt response and the output of the prediction model is calculated using real-time sensing data. The linearization offset term d(k) in the prediction model is corrected in real time.

[0113]

[0114] In the formula, To offset the updated gain matrix, a diagonal attenuation coefficient is used to ensure stable convergence of the correction process. The updated offset term is added to the prediction model for the next cycle to achieve adaptive compensation for model mismatch, maintain the synchronization and accuracy of multi-actuator commands, and keep the melt front velocity, interface temperature, and cooling rate continuously stable within the process window that suppresses appearance defects.

[0115] After the product is demolded, an image sequence covering all highlight surfaces and textured areas is acquired. A pre-trained instance segmentation network is used to detect defects in each frame. This network uses ResNet combined with a feature pyramid as its backbone, and outputs the category and pixel-level segmentation mask for each defect instance through region proposal and mask prediction branches. The network outputs the detection results as an instance set.

[0116] Using the calibrated intrinsic and extrinsic parameters of the camera-fixture system, the two-dimensional defect mask is transmitted through multiple views. Figure 3 The reconstruction is elevated to three-dimensional space. Specifically, based on the camera projection matrix and fixture rotation parameters, the mask contour pixels are back-projected into a 3D point cloud, and an iterative nearest-point algorithm is used to register it with the product CAD model. By indexing the registered defect point cloud to the model's appearance feature regions, a correspondence between defects and feature regions is established, resulting in mapped structured defect information.

[0117] Based on the injection molding cycle timestamp, the complete sequence of process parameters and mold status time-series data recorded within that cycle are retrieved from the production database. Process parameters include injection speed curves, holding pressure curves, and temperature setpoints for each mold zone; mold status includes cavity temperature field time-series data and pressure curves. Structured defect information and process status data are encapsulated into an incremental sample tuple to construct a structured incremental training dataset.

[0118] In injection molding production intervals or real-time background threads, the process parameter-appearance mapping model is fine-tuned online in small batches using accumulated incremental samples. Let the current model parameters be Θ, the incremental sample batch be B, and the fine-tuning optimization objective be to minimize a combination function of incremental loss and historical knowledge retention loss.

[0119]

[0120] In the formula, The cross-entropy loss is used for defect classification on incremental samples, and the right-hand side is the elastic weight consolidation regularization term. To fine-tune the model parameters, For parameters The diagonal element estimate of the Fisher information matrix measures the importance of this parameter to historical tasks. is the regularization strength coefficient. This regularization term effectively suppresses catastrophic forgetting during incremental learning, enabling the model to maintain its ability to identify existing defects while adapting to new defect patterns.

[0121] The diagonal elements of the Fisher information matrix are estimated using the expected value of the squared gradient of the model on historical samples before fine-tuning:

[0122]

[0123] In the formula, H represents the historical sample distribution. After fine-tuning, the updated model weights replace the current online inference model, and the prediction model instance called in the collaborative controller is updated synchronously, so that the mapping logic continuously adapts to the current production conditions and completes the self-evolution of the quality closed loop.

[0124] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0125] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fully intelligent collaborative control method for the molding process of complex appearance parts, characterized in that, include: Obtain the 3D model and appearance quality specifications of complex appearance parts, extract the highlight area and texture area as key appearance feature areas, construct a multiphysics injection molding simulation model that couples mold thermal balance, melt filling flow and cooling shrinkage, generate an appearance defect sample set through simulation, and mark the defect type and location. Based on the sample set of appearance defects, a graph attention network is trained with cavity mesh diagram structure data and process parameters as input to obtain the probability distribution of defect occurrence in each feature region. An execution characteristic simulator is established using historical data of injection molding machine actuator response. The simulation domain graph attention network is transferred to an online inferable process parameter-appearance mapping model through adversarial domain adaptation technology. In the actual injection molding cycle, the mold temperature field, cavity pressure curve and melt flow front position are collected in real time. The collected information is input into the process parameter-appearance mapping model to dynamically assess the risk level of weld lines, flow marks and shrinkage marks in each appearance feature area of ​​the current cycle. Based on the risk level of shrinkage marks, the collaborative controller synchronously generates injection speed curve parameter adjustment instructions, multi-segment holding pressure and time correction instructions, and mold zone temperature setting value offset instructions. Through feedforward-feedback composite control, the melt front speed, interface temperature, and cooling rate are maintained within the process window that suppresses appearance defects. Multi-angle appearance images of the demolded product are acquired, and an instance segmentation network is used to identify the defect type and spatial location. The defect information and the corresponding process parameter sequence are combined into incremental training samples. The process parameter-appearance mapping model is fine-tuned online, and the mapping logic in the collaborative controller is updated.

2. The intelligent collaborative control method for the entire process of forming complex appearance parts according to claim 1, characterized in that, The construction of a multiphysics injection molding simulation model that couples mold thermal balance, melt filling flow, and cooling shrinkage, and the generation of an appearance defect sample set through simulation, specifically includes: Import the 3D model into the simulation platform, refine the local mesh in the appearance feature area, and set a prism boundary layer to solve the near-wall melt temperature gradient. Construct a mold thermal model that includes conformal cooling water channels, apply zoned temperature setting boundaries to the mold, and set the initial temperature of each zone according to the distribution of appearance feature areas; The equations of incompressible non-isothermal flow and material state are coupled to solve the shear rate field, temperature field and free surface evolution of the melt during the filling process, and the volume shrinkage and residual stress field are calculated during the pressure holding and cooling stage. In post-processing, weld lines, flow marks, and shrinkage marks are marked by combining the weld confluence angle, peak surface shear stress, and local shrinkage rate, generating a structured defect sample dataset with process parameter labels.

3. The intelligent collaborative control method for the entire process of forming complex appearance parts according to claim 1, characterized in that, The training and transfer of the graph attention network specifically includes: The simulation mesh is transformed into graph data according to the spatial connection relationship. The node attributes include node coordinates, temperature, pressure, shear rate and wall heat flux, and the edge attributes are normalized flow direction vector and distance between nodes. A prediction network consisting of multiple stacked graph attention layers is constructed, taking node features and global process condition vectors as input, and outputting the defect category probability of each node; A long short-term memory network is trained using historical operating data of an injection molding machine as an execution characteristic simulator. The simulator predicts the timing of physical actions that actually reach the mold based on ideal process instructions. An adversarial domain adaptive training mechanism is introduced, using a small amount of labeled data from actual production as the target domain. The difference in feature distribution between the simulation domain and the actual domain is reduced by using a gradient inversion layer, and the simulation domain prediction network is transferred to a process parameter-appearance mapping model.

4. The intelligent collaborative control method for the entire process of forming complex appearance parts according to claim 3, characterized in that, The adversarial domain adaptive training mechanism specifically includes: Construct a domain discriminator network, using the output of the intermediate layer of the feature extraction network as input, to determine whether the data originates from the simulation domain or the actual domain; By alternating training to minimize defect prediction loss and maximize domain discrimination loss, the feature representation is made domain invariant. Given the extremely limited number of actual domain samples, data augmentation techniques are employed to expand the actual defect samples through geometric and color transformations, thereby assisting the domain adaptive process.

5. The intelligent collaborative control method for the entire process of forming complex appearance parts according to claim 1, characterized in that, The real-time acquisition of mold temperature field, cavity pressure curve, and melt flow front position, and input of the acquired information into the process parameter-appearance mapping model, dynamically assesses the risk level of weld lines, flow marks, and shrinkage marks in each appearance feature area of ​​the current cycle, specifically including: Temperature measurement holes are opened on the back of each appearance feature area of ​​the mold cavity surface, and multiple pairs of K-type armored thermocouples are embedded. High sampling rate modules are used to collect data synchronously to form a time series temperature distribution tensor. A piezoelectric pressure sensor is flush-mounted at the end of the cavity filling to measure pressure changes in real time and extract the pressure rise rate and pressure peak as features. By opening an infrared window on the fixed side of the mold and setting up a near-infrared high-speed camera, real-time images of the melt front are captured, and the arrival time and instantaneous velocity of the front are obtained through sub-pixel edge extraction and spatiotemporal tracking algorithms. The temperature distribution tensor, pressure characteristics, and frontal velocity information are input into the process parameter-appearance mapping model to calculate the dynamic risk index of each feature region, and the perception strategy is updated based on the risk index.

6. The intelligent collaborative control method for the entire process of forming complex appearance parts according to claim 5, characterized in that, The strategy for updating perception based on the risk index specifically includes: A risk index grading threshold is set. When the risk index of a certain feature area exceeds the medium risk threshold, the sampling frequency of the temperature sensor near the area is dynamically increased, and the optical imaging probe is activated to scan the local solidified layer morphology. The encrypted data is fed into the model in real time for secondary evaluation, refining the defect boundaries and risk trends. The time step of the subsequent regulation cycle will be adjusted based on the refined risk trends.

7. The intelligent collaborative control method for the entire process of forming complex appearance parts according to claim 1, characterized in that, The process of synchronously generating injection speed curve parameter adjustment instructions, multi-segment holding pressure and time correction instructions, and mold zone temperature setpoint bias instructions based on the risk level of shrinkage marks, using a collaborative controller, and maintaining the melt front velocity, interface temperature, and cooling rate within the process window that suppresses appearance defects through feedforward-feedback composite control, specifically includes: Receive the risk level vector and main cause identifier of each appearance feature area output by the defect risk assessment module, and construct a collaborative optimization problem with the goal of minimizing risk. The decision variables include the injection speed curve control point, the holding pressure timing node and the mold partition temperature offset. A model-based predictive control architecture is adopted, using the process parameter-appearance mapping model as the predictive model. Multi-objective optimization is solved in the rolling time domain to generate a coordinated control sequence that satisfies the physical constraints of the actuators. The first control command is then issued to the injection servo motor, the pressure holding servo valve, and the zoned mold temperature actuator. In the next control cycle, the deviation between the actual melt response and the output of the prediction model is calculated using real-time sensing data, and the linearization offset term in the prediction model is corrected in real time.

8. The intelligent collaborative control method for the entire process of forming complex appearance parts according to claim 7, characterized in that, The solution process for the predictive control architecture specifically includes: Using the current real-time status and process parameter-appearance mapping model, with the injection speed curve Bezier control point increment, holding pressure inflection point increment, and mold temperature zone increment as decision variables, a cost function including risk index minimization and energy consumption weighting is established. The particle swarm optimization algorithm is used to quickly search for the optimal coordination sequence for several future control cycles within the constraint space, and a partial solution space is pre-defined by explicit parameterization rules to accelerate online solution. The optimal instructions in the first cycle are synchronously sent to each actuator via the EtherCAT bus, and the actuator characteristics simulator is used to perform one-order advance compensation to eliminate instruction latency.

9. A full-process intelligent collaborative control system for the molding of complex appearance parts, used to implement the full-process intelligent collaborative control method for the molding of complex appearance parts as described in any one of claims 1-8, characterized in that, include: Simulation and Sample Generation Module: Based on the 3D model and quality specifications, a multiphysics simulation model coupling heat, fluid and solid is constructed to simulate the injection molding process and generate a batch of structured sample datasets of defect types and locations; Mapping Model Construction and Transfer Module: Uses simulation samples to train a graph attention network to predict defects, and combines historical data of injection molding machines to build an execution characteristic simulator. Through adversarial domain adaptation technology, the simulation-trained model is transferred and adapted into a model that infers the mapping relationship between process parameters and appearance quality in actual production. Perception and Dynamic Assessment Module: Collects physical field data in real time and inputs it into the process parameter-appearance mapping model to dynamically assess the risk level of appearance defects in each feature area during the current production cycle; Collaborative optimization control module: Based on the dynamic risk assessment results, it optimizes and generates coordinated control commands for injection speed curves, multi-segment holding pressure parameters, and zoned mold temperature settings. Through feedforward-feedback composite control, it stabilizes the process state within the window for suppressing defects. Online learning and model update module: performs image acquisition and instance segmentation on demolded products, identifies defect information, combines it with corresponding process parameters to form incremental samples, fine-tunes the process parameter-appearance mapping model online, and synchronously updates the mapping logic of the collaborative controller; Process execution and closed-loop interface module: It sends the optimization instructions of the collaborative controller to the field actuators and collects their response feedback to form the basic data link for closed-loop control.

Images