Compensation control method based on the prediction of the injection weld line position of connector inserts

Abstract

This invention relates to a compensation control method for predicting the location of weld lines in connector insert injection molding, aiming to overcome the problem of inaccurate predictions caused by insufficient perception of local thermal inertia and rheological state in traditional weld line prediction methods. The method deploys a multimodal micro-sensor array at key locations in the mold to collect multi-channel signals of pressure, temperature, and micro-deformation in real time. Combined with three-dimensional thermal inertia modeling and experimental calibration of the insert, it achieves dynamic labeling of the thermophysical behavior of each batch of inserts. Based on a graph neural network, it aggregates rheological fingerprint features and thermal boundary information to output a refined two-dimensional probabilistic thermogram of the weld line, which is then spatially weighted with the compensated temperature field data to obtain the final location prediction result. The system automatically adjusts the gate switching and holding pressure curve in a closed loop, and recursively optimizes parameters based on molding quality feedback to achieve high-precision and adaptive weld line control, improving product structural integrity and production stability.

Description

Technical Field

[0001] This invention relates to the field of quality prediction and compensation control technology in injection molding process, and in particular to a compensation control method based on the prediction of the location of injection weld lines in connector inserts. Background Technology

[0002] The formation and distribution of weld lines during injection molding directly affect the mechanical strength and appearance quality of the product. Therefore, accurate prediction and dynamic compensation control of weld line locations are of significant practical importance for complex injection molding processes. In the current technological field, mainstream methods primarily focus on data-driven machine learning models, finite element simulation, and physical quantity sensing and monitoring to achieve weld line location prediction and optimized control. Specifically, the industry generally adopts the following technical approach: Deep learning prediction paradigms based on multimodal data fusion are widely used in injection molding defect detection and quality optimization. Typical approaches rely on historical production data, collecting large-scale process parameters such as melt pressure, temperature, and displacement to train multimodal models (convolutional neural networks, recurrent neural networks, or models incorporating attention mechanisms) to predict weld line distribution and formation trends. Some solutions combine finite element simulation results, using model structure, material properties, and typical process parameters as input to output the probability distribution of weld line regions in a black-box manner. In recent years, emerging models such as graph neural networks have been introduced into spatial feature learning in complex flow channel systems, but end-to-end large-scale model training remains the primary technical approach.

[0003] At the physical level, pressure and temperature sensors are widely deployed in injection molds for real-time monitoring. However, the collected data is mostly used as the basis for subsequent statistical analysis and quality traceability, without forming a real-time dynamic compensation capability. Thermal uniformity optimization is usually achieved by adjusting cooling water circuits, adjusting holding pressure curves, and fine-tuning temperature control. However, these control measures often rely on operational experience or linear regression-like static models, making it difficult to adapt to complex disturbances such as material fluctuations, mold wear, and batch differences in inserts.

[0004] Existing deep learning-based prediction methods have the following typical technical characteristics and applicable scope: They fully leverage the advantages of big data, enabling the discovery of nonlinear mapping relationships of weld lines under complex data distributions, making them suitable for large injection molding companies with large production batches, limited process parameter variations, and ample data accumulation. Finite element simulation and discrete physical models, on the other hand, focus on reproducing the melt flow and heat conduction processes under ideal working conditions, often used for mold structure optimization and preliminary setting of molding parameters, but they are less sensitive to dynamic disturbances in actual working conditions. While the above technologies can meet the weld line prediction requirements of some standard products, their generalization ability and response speed are limited when dealing with dynamic conditions such as complex inserts, frequent material changes, and long-term mold operation (wear), making it difficult to achieve closed-loop online compensation.

[0005] Specifically, the existing technology has the following main shortcomings: First, weld line prediction models based on deep learning or traditional statistical methods are highly dependent on historical training data. Once material properties fluctuate, the internal structure of the mold experiences localized wear, or the preheating state of the insert changes, the model output will deviate significantly. The black-box nature of this end-to-end approach results in a lack of physical interpretability regarding the dynamic relationship between melt rheology and thermophysical properties, limiting the model's transferability and making it unable to adapt to the complex and frequently changing disturbances in the production environment.

[0006] Secondly, existing methods often employ single or static thermal boundary assumptions (such as ignoring the thermal inertia of inserts) or rely solely on fuzzy adjustments based on local temperature feedback, making it difficult to achieve high-precision dynamic correction of the overall thermal field of the cavity. Furthermore, when faced with time-varying factors such as batch material differences and decreased mold cooling efficiency, they cannot accurately predict the spatial offset effect of local thermal field disturbances on weld line formation, leading to unstable compensation results.

[0007] Furthermore, the fusion and alignment of multimodal data lack clear physical guidance and coordination, and data noise and sampling delays are easily mislearned by the model, affecting the reliability of the prediction results. For large connector molds with complex flow channel structures and heterogeneous sensor node distributions, there is a lack of targeted modeling methods for effectively combining spatial topology information and temporal response.

[0008] In addition, existing technologies mostly focus on one-time prediction and post-event optimization, lacking a real-time dynamic compensation mechanism, and often cannot achieve accurate calibration of weld line position and closed-loop adaptive adjustment of process parameters within a single injection molding cycle.

[0009] In summary, existing methods for predicting and controlling injection molding weld lines urgently need to address the following core technical issues: This enhances the model's generalization ability under dynamic conditions such as batch material variations, mold thermal resistance fluctuations, and insert condition differences, breaking through the strong dependence of traditional data-driven models on historical data and endowing the prediction system with high adaptability and physical interpretability. This technology enables precise modeling and real-time correction of dynamic disturbances in the cavity thermal field, breaking through the limitations of coarse-grained adjustment methods based on single boundary conditions. It effectively compensates for the influence of insert thermal inertia and local heat flow variations on the micro-solidification process of weld lines. Establish a deep fusion mechanism between physical layer multimodal sensing data and spatial topology to improve prediction accuracy, response speed, and compensation stability, thereby meeting the actual needs of high-cycle injection molding production for process quality optimization. Currently, the industry lacks a complete technical system capable of achieving physically interpretable dynamic weld line prediction and supporting online feedforward process compensation without relying on large-scale historical samples. To address these shortcomings, there is an urgent need to propose a novel real-time prediction and closed-loop compensation control method that integrates melt rheological fingerprint migration, insert thermal inertia dynamic compensation, and spatial physical field weighted correction. This method would provide effective technical support for high-precision control of injection molding weld lines under complex dynamic disturbance conditions. Summary of the Invention

[0010] This application provides a compensation control method based on the prediction of the injection weld line position of connector inserts, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0011] The compensation control method based on the prediction of the injection weld line position of connector inserts provided in this application specifically includes: S1: Simultaneously collect transient pressure peaks, temperature rise slopes, and local cavity micro-deformations of the melt as it passes through the injection molding machine nozzle to the main runner inlet section, the end of the branch runner, and the leading edge of each gate to generate the original rheological response sequence. S2: Process the original rheological response sequence to obtain the melt rheological fingerprint vector; S3: Based on the three-dimensional geometry and material properties of the connector insert, a thermal inertia model is established. Using the equivalent heat capacity and surface heat transfer coefficient data calibrated by offline experiments, an insert thermal inertia tag is generated that is bound to each batch of inserts. S4: Read the thermal inertia tag of the current batch of inserts and call the pre-stored thermal boundary correction function to dynamically adjust the mold cooling water circuit control strategy and the local thermal compensation power output during the holding pressure stage, and generate cavity temperature field distribution data. S5: Input the melt rheological fingerprint vector into the graph neural network model to output a two-dimensional probability heat map; S6: Spatial superposition and weighted calculation are performed on the two-dimensional probability heat map and the cavity temperature field distribution data to generate the final predicted coordinate set of weld line location; S7: Generate a real-time compensation control signal based on the predicted coordinate set of the final weld line position; S8: Compensation control for predicting the injection weld line position is completed based on the real-time compensation control signal.

[0012] The compensation control method based on the prediction of the injection weld line position of connector inserts provided in this application has the following beneficial effects: (1) By deploying no less than five sets of micro-embedded sensor arrays from the nozzle to the main runner inlet section, the end of the branch runner and the front edge of each gate of the injection molding machine, and combining high-response piezoelectric sensors, thin film thermocouples and micro-strain gauges to synchronously collect the transient pressure peak, temperature rise slope and local cavity micro-deformation when the melt passes through, a raw rheological response sequence with clear physical meaning is constructed. Furthermore, highly interpretable scalar features such as pressure decay rate, temperature integral area and strain hysteresis phase angle are extracted to form a melt rheological fingerprint vector with temporal topology, which significantly improves the characterization accuracy and dynamic perception capability of melt flow behavior. Compared with the traditional monitoring method based on statistical mean or empirical threshold, this scheme effectively overcomes the process instability caused by material batch fluctuations and differences in insert thermal state, realizes the pre-emptive and refined capture of weld line generation conditions, and provides a high-fidelity input basis for subsequent prediction and control.

[0013] (2) A three-dimensional thermal inertia modeling mechanism is introduced for connector inserts. The equivalent heat capacity and surface heat transfer coefficient parameter sets under different preheating temperatures, insertion times and material batches are calibrated through offline experiments and encoded as uniquely associated insert thermal inertia tags and uploaded to the central database. Before each mold closing, the corresponding thermal boundary correction function is automatically read and called to dynamically adjust the cooling water circuit control strategy and the local thermal compensation power during the pressure holding stage, so that the cavity thermal field distribution actively approaches the reference state without inserts. This effectively solves the problem of local temperature imbalance and uneven cooling caused by the introduction of inserts in the existing technology. This feedforward compensation mechanism can achieve accurate thermal field reconstruction without relying on a large amount of historical data for training, avoiding the high dependence of deep learning models on labeled samples and the lack of generalization ability. At the same time, it supports rapid changeover adaptation in multi-variety small-batch production scenarios, which greatly improves the robustness and engineering practicality of the system.

[0014] (3) A lightweight graph neural network is adopted with the geometric topology of the mold flow channel as the graph structure. Each sensor node is used as a graph node and the rheological fingerprint is used as a node attribute. The upstream and downstream flow information is aggregated by the message passing mechanism to output a two-dimensional probability heat map of the weld line on the cavity plane. Then, it is spatially superimposed and weighted with the temperature field after thermal inertia compensation to generate the final prediction coordinate set and feed it back to the injection molding machine closed-loop control system in real time to trigger the gate switching or pressure holding curve replanning action. The entire process is inferred at the edge and the single time is less than 80 milliseconds, which fully meets the requirements of high-speed injection molding cycle time. Unlike the existing black box deep learning prediction paradigm, this solution emphasizes the decoupling of physical process and the driving of measurable parameters, avoiding complex paths such as multimodal data alignment, attention weight allocation, GAN enhancement, and LoRA adaptation. It not only improves the model running efficiency and response sensitivity, but also enhances the interpretability of the decision-making process and the transparency of the process. It builds an integrated adaptive closed-loop system of "perception-modeling-compensation-prediction-control", providing an efficient, reliable and low-latency technical solution for precision injection molding.

[0015] In summary, this solution integrates key technologies such as multi-source embedded sensing, physically interpretable feature extraction, insert thermal inertia feedforward compensation, and lightweight graph neural network spatial inference to achieve high-precision, low-latency, and robust prediction of weld line locations and real-time process intervention. It overcomes the limitations of traditional methods in terms of dynamic adaptability, computational efficiency, and integration of physical mechanisms, and has good scalability and industrial application value. It is especially suitable for precision injection molding manufacturing scenarios with high complexity and high consistency requirements. Attached Figure Description

[0016] Figure 1 This is the main flowchart of a compensation control method based on the prediction of the injection weld line location of connector inserts.

[0017] Figure 2 This is a sub-flowchart of a compensation control method based on the prediction of the injection weld line location of connector inserts.

[0018] Figure 3 This is another sub-flowchart of the compensation control method based on the prediction of the injection weld line position of connector inserts. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0021] like Figure 1 As shown, this application provides a compensation control method based on the prediction of the injection molding weld line position of connector inserts, specifically including: S1: Simultaneously collect transient pressure peaks, temperature rise slopes, and local cavity micro-deformations of the melt as it passes through the injection molding machine nozzle to the main runner inlet section, the end of the branch runner, and the leading edge of each gate to generate the original rheological response sequence. S2: Process the original rheological response sequence to obtain the melt rheological fingerprint vector; S3: Based on the three-dimensional geometry and material properties of the connector insert, a thermal inertia model is established. Using the equivalent heat capacity and surface heat transfer coefficient data calibrated by offline experiments, an insert thermal inertia tag is generated that is bound to each batch of inserts. S4: Read the thermal inertia tag of the current batch of inserts and call the pre-stored thermal boundary correction function to dynamically adjust the mold cooling water circuit control strategy and the local thermal compensation power output during the holding pressure stage, and generate cavity temperature field distribution data. S5: Input the melt rheological fingerprint vector into the graph neural network model to output a two-dimensional probability heat map; S6: Spatial superposition and weighted calculation are performed on the two-dimensional probability heat map and the cavity temperature field distribution data to generate the final predicted coordinate set of weld line location; S7: Generate a real-time compensation control signal based on the predicted coordinate set of the final weld line position; S8: Compensation control for predicting the injection weld line position is completed based on the real-time compensation control signal.

[0022] Step S1: Simultaneously collect transient pressure peaks, temperature rise slopes, and local cavity micro-deformations of the melt as it passes through the injection molding machine nozzle to the main runner inlet section, the end of the branch runner, and the leading edge of each gate, generating the original rheological response sequence. Specifically, this includes: S1.1: Based on the geometric topology of the mold runner, determine the layout scheme of the sensor nodes. Plan the installation coordinates of no less than five sets of micro embedded sensor arrays at the injection molding machine nozzle to the main runner inlet section, the end of the branch runner and the front edge of each gate, and generate a sensor node spatial distribution mapping table to establish the physical boundary conditions for multi-source heterogeneous data acquisition.

[0023] A 3D CAD geometric model of the connector insert injection mold was obtained, and the central axis topology of the main runner, branch runners, and gate system was extracted to construct a runner network connectivity diagram. Based on melt flow simulation data, potential weld line regions with the highest probability of melt front convergence were identified, and the coordinates of key monitoring points at the gate front were determined. A first sensing node was installed at the section from the injection machine nozzle to the main runner inlet to capture the initial injection pressure and temperature baseline. Second and third sensing nodes were installed at the branch ends of the branch runners to monitor pressure loss and heat dissipation during the flow splitting process. Fourth and fifth sensing nodes were installed at each gate inlet to capture abrupt changes in rheological state at the filling end. The Euclidean distance between each sensing node and the runner friction coefficient were calculated, and the signal transmission delay and attenuation characteristics were evaluated. Based on the sensor physical dimensions and mold steel strength requirements, the wall thickness allowance at each preset installation position was verified to ensure that the integrity of the mold structure was not compromised. A spatial distribution mapping table of sensing nodes, including node ID, 3D spatial coordinates, runner level, and associated gate number, was generated. Through the above layout planning, the geometric model of the previous step is transformed into a sensor network topology with physical constraints, establishing a spatial benchmark for multi-source heterogeneous data acquisition and achieving precise coverage of each key rheological node.

[0024] For example, in a precision USB connector mold, the main runner is 80mm long, and the branch runners adopt a balanced layout. The first sensing node is located 10mm after the nozzle interface, with coordinates (0,0,0). The second and third nodes are located at the end of the branch runner, 40mm from the center of the main runner, with coordinates (-40,20,0) and (40,20,0) respectively. The fourth and fifth nodes are located at the entrances of the two pinhole gates, with coordinates (-40,35,0) and (40,35,0) respectively. The spacing between each node meets the minimum installation spacing requirement of 5mm, and the wall thickness allowance is greater than 3mm. The generated mapping table contains information on the five nodes, with clear relationships, providing precise guidance for subsequent hardware integration and significantly improving the spatial representativeness of the data acquisition.

[0025] S1.2: Perform hardware integration operation according to the spatial distribution mapping table of the sensing nodes, embed a miniature embedded sensing array containing a high-response piezoelectric sensor, a thin-film thermocouple and a micro-strain gauge into the surface of a predetermined mold cavity, construct a multimodal sensing network with a synchronous triggering mechanism, and generate a set of physical layer data acquisition channels.

[0026] Read the spatial distribution mapping table of sensor nodes, parse the installation coordinates, depth parameters and sensor type configuration information of each micro embedded sensor array, and establish the physical benchmark for hardware integration.

[0027] Precision machining is performed on the surface of the mold cavity according to the preset coordinates. Blind hole slots that match the shape of the sensor are prepared by electrical discharge wire cutting technology to ensure that the dimensional tolerance of the slots is controlled within ±0.01mm, so as to eliminate signal distortion caused by assembly gaps.

[0028] High-response piezoelectric sensors, thin-film thermocouples, and micro-strain gauges are embedded in their respective blind hole slots and filled and cured with high-temperature resistant epoxy resin to ensure that the sensor's sensitive surface is flush with the inner wall of the mold cavity, thus creating a disturbance-free physical sensing interface.

[0029] Shielded twisted-pair wires are soldered to the pins of each sensor, and the cables are led out to the external junction box of the mold through the reserved channel of the mold cooling water channel. Grounding is used to suppress high-frequency electromagnetic interference at the injection site and form a stable analog signal transmission link.

[0030] A synchronous trigger controller is installed on the moving mold side of the mold. The injection molding machine's mold closing completion signal is used as the global time reference. Synchronous acquisition commands are sent to each sensor node through hardware interrupts to eliminate time jitter in multi-channel data acquisition.

[0031] Impedance matching and insulation resistance verification were performed on the integrated sensor network to confirm that the signal-to-noise ratio of each channel was higher than 60dB, and a physical layer data acquisition channel set containing channel ID, sensor type, gain coefficient and zero offset was generated.

[0032] Through the aforementioned hardware integration and calibration process, the spatial layout scheme of the previous step is transformed into a set of physical layer data acquisition channels with high-fidelity synchronous acquisition capabilities, enabling accurate capture of multimodal rheological signals and laying the hardware foundation for the subsequent generation of the original rheological response sequence.

[0033] For example, in a USB connector mold, a Kistler 6157B piezoelectric sensor is embedded 10mm after the nozzle, and a PT100 thin-film thermocouple is embedded at the end of the manifold. A blind hole with a depth of 2.5mm is machined, and Epoxy-Tek 377 adhesive is used for curing at 150°C for 2 hours. The synchronous trigger controller is set to a sampling delay of 0ms and a sampling frequency of 10kHz. Testing shows that the pressure channel signal-to-noise ratio is 65dB, the temperature response time is less than 5ms, the generated channel set contains 5 nodes and 15 data channels, and the zero-point drift is less than 0.5%FS, meeting the requirements for transient signal acquisition in high-speed injection molding.

[0034] S1.3: The transient physical quantities during the molten plastic filling process are captured in real time using the physical layer data acquisition channel set. The transient pressure peak signal, temperature rise slope signal and local cavity micro-deformation signal at each sensing node are acquired simultaneously to generate a multi-channel original analog voltage signal set.

[0035] Based on the physical layer data acquisition channel set, the synchronous trigger signals of each micro-embedded sensor array are activated to initiate real-time monitoring of the molten plastic filling process. High-response piezoelectric sensors sense the transient mechanical stress as the melt front passes through, converting it into an analog voltage signal proportional to the pressure, capturing the pressure peak and its dynamic rate of change. Thin-film thermocouples sense the heat flux changes at the contact interface in real time, outputting a millivolt-level thermoelectric potential signal reflecting the temperature rise slope, characterizing the local enthalpy exchange rate. Micro-strain gauges detect the micron-level elastic deformation of the cavity wall caused by high-pressure injection, generating a resistance change voltage signal linearly related to the local cavity micro-deformation. Each sensing node converts the above multimodal physical quantities into independent analog electrical signals and transmits them to the signal conditioning module via shielded twisted-pair cables. The signal conditioning module performs impedance matching and low-noise amplification on each analog signal to eliminate environmental electromagnetic interference introduced by long-distance transmission. A high-speed parallel sample-and-hold circuit is used to lock the instantaneous voltage values ​​of all channels on the same clock edge, ensuring the time consistency of multi-source data. The processed analog voltage signals are fed into a multi-channel data acquisition card, forming a multi-channel raw analog voltage signal set that includes pressure, temperature, and strain dimensions. Through multi-sensor synchronous acquisition and signal conditioning, the physical deployment of the previous step is transformed into a quantifiable multi-channel raw analog voltage signal set, achieving high-fidelity, low-latency panoramic capture of the melt rheological state and providing a high-quality data foundation for subsequent digital signal processing.

[0036] For example, in the injection molding scenario of a precision USB connector, the sampling frequency is set to 100kHz. The piezoelectric sensor has a sensitivity of 5mV / MPa, capturing a peak voltage of 2.5V when the melt passes through, corresponding to a transient pressure of 50MPa. The thin-film thermocouple has a time constant of less than 1ms, detecting a temperature rise slope of 15℃ / s, and outputting a voltage signal amplitude of 0.8mV. The micro-strain gauge has a range of 1000με, measuring a cavity micro-deformation of 50με, corresponding to a voltage change of 0.2mV. Five sets of sensor arrays, totaling 15 channels, synchronously output analog signals. After processing by an amplifier with a gain of 100, the signal-to-noise ratio is improved to over 60dB. The generated multi-channel raw analog voltage signal set completely records the dynamic changes within 0.5 seconds of the filling stage, with no data loss and a phase alignment error of less than 1μs, significantly improving the accuracy of rheological feature extraction.

[0037] S1.4: Perform analog-to-digital conversion and timing alignment processing on the multi-channel original analog voltage signal set, use a high-precision sampling clock to unify the timestamp reference of each sensor data, eliminate the phase difference caused by transmission delay, and generate a time-synchronized multi-dimensional digital sensing data stream.

[0038] S1.5: Based on the time-synchronized multidimensional digital sensing data stream, perform data structured encapsulation, combine discrete data points into continuous time series segments according to a preset sampling frequency, and generate an original rheological response sequence containing complete filling cycle information, which serves as the standard input object for the subsequent feature extraction module.

[0039] Step S2: Process the original rheological response sequence to obtain the melt rheological fingerprint vector. Specifically, this includes: S2.1: Perform adaptive sliding window segmentation processing on the original rheological response sequence. Based on the characteristic time constant of the melt front through the sensor array, set the window length and overlap step size to convert the continuous time domain signal into multiple local time series data segments to obtain a standardized time series data segment set containing the complete transient response process.

[0040] The system receives the time-synchronized multidimensional digital sensing data stream generated by S1.5, which contains continuous time-series signals from three channels: pressure, temperature, and micro-strain. It analyzes the arrival time of the melt front corresponding to each sensing node, calculates the standard residence time of the melt through a single sensing area based on historical batch data, and sets the initial sliding window width to 1.5 times the characteristic time constant to ensure complete coverage of the transient response peak and decay process.

[0041] An adaptive step-size method is used to calculate the overlap rate of adjacent windows. Based on the signal sampling frequency and processor computing power constraints, the overlap step-size is set to 50% of the window width to balance the real-time performance of feature extraction and data redundancy. The original signal sequence is truncated to extract multi-channel data segments within the first complete window, forming a preliminary local time-series data block.

[0042] The signal-to-noise ratio (SNR) of the pressure signal within the detection window is measured. If the SNR is lower than a preset threshold, the window width is dynamically expanded to include more background noise samples for subsequent filtering; otherwise, the standard window size is maintained. Zero-mean normalization is performed on the extracted local time-series data blocks to eliminate the dimensional effects caused by differences in reference levels between different sensors, generating standardized time-series data segments.

[0043] Iterate through all sensor nodes, repeating the windowing and standardization operations described above, until all valid data segments of the entire filling cycle are covered. Store the processed data segments from each node into a cache queue in timestamp order, constructing a set of standardized time-series data segments containing the complete transient response process.

[0044] By using adaptive sliding window segmentation and standardization, continuous non-stationary sensing signals are transformed into discrete and aligned local time-series segments, enabling accurate capture of transient features of melt rheology and providing a high-quality data foundation for subsequent high-dimensional feature extraction.

[0045] For example, in the injection molding of LCP material connectors, the sampling frequency is set to 10kHz. The characteristic time constant of the melt passing through the gate sensing area is approximately 2ms, the window width is set to 3ms (i.e., 30 sampling points), and the overlap step is 1.5ms. The pressure signal is Z-score normalized, with the mean set to zero and the standard deviation set to one. If the detected pressure peak fluctuation exceeds three times the baseline standard deviation, it is considered a valid filling window. Finally, a normalized data matrix of 5 nodes × 100 time slices is generated, significantly improving the signal-to-noise ratio and consistency of feature extraction.

[0046] S2.2: Based on the pressure component data in the standardized time-series data set, the linear regression fitting method is used to calculate the descent slope after the pressure peak within the window, and the normalization operation is performed in combination with the melt viscosity characteristic parameters to obtain the scalar characteristic of the pressure decay rate that characterizes the change in melt flow resistance.

[0047] Pressure component data sequences are extracted from a standardized time-series data set, and the transient pressure peak point and its corresponding time index are located within the window. Data from the pressure peak point to the end of the window are extracted to construct the independent variable time vector and dependent variable pressure vector for linear regression analysis. The pressure drop segment is fitted with a first-order linear model using the least squares method, and the slope of the fitted line is calculated. This slope represents the instantaneous pressure decay rate of the melt within a specific flow channel segment. The melt reference viscosity parameter and standard injection rate parameter set in the current injection molding process are obtained as normalization benchmark factors. A pressure decay rate normalization calculation model is constructed, dividing the original slope value by the product of the reference viscosity and the injection rate to eliminate the interference of process setting fluctuations on the feature value. Through the above processing method, the results of the previous step are transformed into a dimensionless pressure decay rate scalar feature, realizing a quantitative characterization of the melt flow resistance variation characteristics and improving the robustness of the feature to material batch fluctuations.

[0048] For example, in the injection molding scenario of LCP material connectors, the sampling frequency is set to 10kHz and the sliding window length is set to 20ms. In the pressure sequence captured by a certain sensor node, a peak pressure of 85MPa is identified, followed by a linear decrease in pressure to 70MPa within 15ms. The original pressure drop slope is obtained by least squares fitting: -1.0 MPa / ms. The reference viscosity of this batch of LCP material is known to be 150 Pa·s, and the standard injection rate is 100 mm / s. Substituting these values ​​into the normalization formula, the scalar characteristic value of the pressure decay rate is obtained as -6.67 × 10⁻⁶. -5 This eigenvalue eliminates the influence of absolute pressure level, accurately reflects the rheological resistance state of the melt in this flow channel section, and provides stable node attribute inputs for subsequent graph neural networks.

[0049] S2.3: Using the temperature component data in the standardized time-series data set, perform a numerical integration method to calculate the area of ​​the closed region between the temperature rise curve and the baseline temperature within the window period, so as to obtain the temperature integral area scalar characteristic reflecting the local enthalpy accumulation effect.

[0050] The temperature component data sequence is read from the standardized time-series data fragment set, and the temperature values ​​of discrete sampling points of the thin-film thermocouple within the sliding window time domain are extracted. The mold cavity baseline temperature is set as the reference zero point, and the baseline temperature is taken as the steady-state background temperature value at the point where the melt front reaches the front sensing node.

[0051] A difference function between the temperature rise curve and the baseline temperature is constructed. The baseline temperature is subtracted from the temperature measurement value at each sampling time within the window to obtain the transient temperature rise deviation sequence, thereby eliminating the interference of ambient temperature drift on enthalpy calculation.

[0052] The composite trapezoidal numerical integration method is used to calculate the area accumulation of the transient temperature rise deviation sequence. The continuous time domain is divided into several small intervals, and the sum of the trapezoidal areas formed by adjacent sampling points is calculated.

[0053] The integral area scalar characteristic of temperature is calculated using the following formula: in, Let N be the temperature integral area, and N be the total number of sampling points within the window. Let i be the temperature value at the i-th sampling point. Baseline temperature This represents the sampling time interval.

[0054] The calculated temperature integral area is normalized by dividing it by the product of the total window duration and the maximum allowable temperature rise. This eliminates the magnitude effect caused by the difference in filling speed during different injection cycles, generating a dimensionless scalar feature of the temperature integral area.

[0055] Through the above numerical integration and normalization processing, the temperature time series data of the previous step is transformed into a temperature integral area scalar feature that reflects the local enthalpy accumulation effect. This realizes the quantitative characterization of the total heat exchange when the melt flows through the sensing node, providing key thermodynamic dimension input for the subsequent construction of melt rheological fingerprint vector.

[0056] For example, in the injection molding process of LCP material connectors, the sampling frequency was set to 10kHz, the sliding window length was set to 5ms, and 50 sampling points were included. The baseline temperature was set to 80℃. Within a certain sampling window, the temperature rapidly increased from 80℃ to 320℃ and then slowly decreased. The calculated trapezoidal integral sum of the transient temperature rise deviation sequence was 1250℃·ms. The maximum allowable temperature rise was set to 300℃, and the total window duration was 5ms. The normalized temperature integral area scalar characteristic value was 1250 / (300×5)=0.833. This characteristic value accurately reflects the high enthalpy state of the melt at this node. Compared with the unnormalized data, it eliminates the absolute area difference caused by injection speed fluctuations and significantly improves the consistency of characteristics between different batches.

[0057] S2.4: For the micro-strain component data and pressure component data in the standardized time-series data fragment set, the cross-correlation analysis method is applied to determine the time delay between the peaks of the two components, and the time delay is converted into a phase angle representation to obtain the strain hysteresis phase angle scalar characteristic that characterizes the hysteresis response properties of viscoelastic materials.

[0058] The micro-strain component sequence and pressure component sequence in the standardized time series data fragment set are read as the input signal source for cross-correlation analysis.

[0059] The micro-strain component sequence and pressure component sequence are subjected to mean-removal preprocessing to eliminate the interference of DC bias on correlation calculation and generate zero-mean strain signal vector and pressure signal vector.

[0060] Based on the zero-mean signal, the discrete cross-correlation function is used to calculate the similarity measure of two signals under different time lags, and a cross-correlation coefficient sequence is constructed.

[0061] Traverse the sequence of cross-correlation coefficients, search for the hysteresis index corresponding to the global maximum value, and determine the time delay of the strain response relative to the pressure excitation.

[0062] The time delay is converted into a phase angle and normalized using the characteristic frequency of melt filling.

[0063] The strain hysteresis phase angle is calculated using the following formula: in, For strain hysteresis phase angle, To characterize the characteristic frequencies of the melt front passing through the sensing nodes, This represents the time delay.

[0064] The calculated phase angle is subjected to interval constraint processing to ensure that it falls within the effective physical range of 0 to 2π, thus generating the final strain hysteresis phase angle scalar feature.

[0065] By using cross-correlation time delay estimation and frequency domain phase mapping, the multidimensional time series data from the previous step is transformed into a strain hysteresis phase angle scalar feature that characterizes the hysteresis response of viscoelastic materials. This enables a quantitative description of the equilibrium state between viscous dissipation and elastic energy storage in the rheological behavior of melt, significantly improving the sensitivity of rheological fingerprints to batch fluctuations in materials.

[0066] S2.5: The pressure decay rate scalar feature, temperature integral area scalar feature, and strain hysteresis phase angle scalar feature are vectorized and encoded according to the preset sensor spatial topology order to generate a melt rheological fingerprint vector with clear physical meaning and temporal correlation structure, which is used as the node attribute input of the downstream graph neural network model.

[0067] like Figure 2 As shown, step S3 involves establishing a thermal inertia model based on the three-dimensional geometry and material properties of the connector insert, and generating an insert thermal inertia tag bound to each batch of inserts using equivalent heat capacity and surface heat transfer coefficient data calibrated offline. Specifically, this includes: S3.1: Obtain the three-dimensional geometric structure data and basic material property parameters of the connector insert, and discretize the three-dimensional geometric structure data based on the finite element mesh generation technology to generate a discrete mesh model of the insert space containing node coordinates and element topology relationships, which serves as the geometric reference for subsequent thermal field simulation calculations.

[0068] Read the basic parameters from the connector insert's 3D CAD geometric model file and material property database, including density, specific heat capacity, and thermal conductivity, to establish the initial physical boundary for thermal inertia modeling.

[0069] Perform topological integrity checks on the 3D geometric model, repair non-manifold edges and tiny gaps, ensure the closure of geometric entities, and provide a defect-free computational domain for mesh generation.

[0070] A tetrahedral-dominated hybrid mesh generation method is adopted, and the local mesh size gradient is set according to the rate of change of the insert wall thickness. The mesh nodes are densified in the thin-walled region to capture rapid temperature gradient changes.

[0071] Generate a discretized mesh data structure containing a node spatial coordinate matrix and a cell connection table, define the thermal degree of freedom index for each node, and construct an insert space discrete mesh model.

[0072] The basic material property parameters are mapped to the centroid of the discrete mesh element, and the thermophysical property field of the element is initialized to form the geometric and physical reference for transient heat conduction finite element calculation.

[0073] Through the discretization process described above, continuous geometric entities are transformed into computable finite element mesh models, achieving an accurate mapping from physical entities to the digital simulation domain, and providing a high-fidelity geometric foundation for subsequent thermal inertia parameter calibration.

[0074] For example, for a certain model of USB Type-C connector metal shielding insert, a 3D model in STEP format was imported, with a minimum mesh size of 0.05 mm, a maximum mesh size of 0.5 mm, and a growth rate of 1.2. After meshing, a discrete mesh model containing 12,500 nodes and 68,000 tetrahedral elements was generated. The mesh quality was checked, and the Jacobian ratio was greater than 0.7, meeting the simulation convergence requirements. The density of copper alloy (8900 kg / m³), specific heat capacity (385 J / (kg·K)), and thermal conductivity (400 W / (m·K)) were assigned to all elements to complete the geometric datum construction.

[0075] S3.2: Using the discrete mesh model of the insert space as the computational domain, and combining the basic material property parameters, a transient heat conduction control equation is constructed. The internal temperature gradient distribution under different preheating temperature conditions is solved by numerical iteration to generate an initial three-dimensional thermal inertial model characterizing the heat storage and transfer capability inside the insert.

[0076] Using an embedded spatial discrete mesh model as the computational domain, the spatial coordinates of each mesh node and the element connection relationships are read to construct the geometric topological foundation for solving the three-dimensional heat conduction problem. Density, specific heat capacity, and thermal conductivity from the basic material properties are mapped to the corresponding mesh elements, assigning each discrete element an independent thermophysical property identifier. Transient heat conduction control equations are established based on Fourier's law of heat conduction to describe the evolution of the internal temperature field of the embedding over time. Several typical preheating temperature conditions are set as initial boundary conditions, including room temperature, medium temperature, and high temperature preheating states, and the temperature values ​​of the mesh nodes are initialized accordingly. The implicit finite difference method is used to discretize the time domain, selecting a fixed time step that satisfies stability conditions to ensure the convergence of the numerical iteration process. An alternating direction implicit scheme is used to solve the spatial domain, sequentially scanning along the X, Y, and Z coordinate axes to decouple the three-dimensional heat conduction equations and reduce computational complexity. Within each time step, a linear equation solution method is executed to update the temperature values ​​of all mesh nodes until the preset simulation termination time is reached. Temperature gradient distribution data within the insert at different times were extracted to record the dynamic path and rate characteristics of heat transfer from the surface to the core. Temperature evolution sequences under multiple operating conditions were integrated to construct an initial three-dimensional thermal inertial model incorporating both temporal and spatial dimensions, characterizing the insert's heat storage and transfer capabilities during unsteady-state heat exchange. Through the aforementioned numerical iterative solution process, discrete geometric meshes and material parameters were transformed into a dynamic thermal response model, achieving a digital reconstruction of the insert's thermal inertial behavior and providing a high-precision simulation benchmark for subsequent parameter calibration.

[0077] S3.3: Based on offline injection molding experiments with multiple batches of different materials and insertion timing, real-time temperature response curves and pressure change sequences of the interface between the melt and the insert are collected. The real-time temperature response curves are subjected to nonlinear fitting processing to extract the equivalent heat capacity scalar value and the surface heat transfer coefficient scalar value that reflect the actual heat transfer efficiency.

[0078] For the initial three-dimensional thermal inertial model generated by S3.2, a multivariate orthogonal experimental matrix is ​​constructed, covering the combination of working conditions in three dimensions: batch difference of insert material, preheating temperature gradient and insertion time delay.

[0079] Before the injection molding machine closes the mold, the insert to be tested is placed in the designated position in the mold cavity, and the transient response data of the interface temperature during the melt filling stage is collected in real time using a micro thin film thermocouple array pre-embedded on the surface of the insert.

[0080] The high-frequency pressure sensor signal located near the gate is read synchronously to obtain the pressure fluctuation sequence at the moment when the melt front contacts the insert, forming a multimodal experimental dataset that is strictly aligned with the temperature data timestamp.

[0081] The acquired temperature response curves were subjected to noise reduction preprocessing. A Savitzky-Golay filter was used to eliminate high-frequency electrical noise interference while retaining the low-frequency trend components that reflect the nature of heat exchange.

[0082] Based on Newton's law of cooling, a nonlinear heat exchange differential equation is constructed. The equivalent heat capacity and surface heat transfer coefficient are set as parameters to be identified, and a residual function between the theoretical temperature evolution model and the actual measured value is established.

[0083] The Levenberg-Marquardt iterative method is used to minimize the residual function, and the step size factor is dynamically adjusted to balance the convergence characteristics of gradient descent and Gauss-Newton method.

[0084] In each iteration, the Jacobian matrix approximation is calculated, and the estimated values ​​of equivalent heat capacity and surface heat transfer coefficient are updated until the residual change is lower than the preset convergence threshold.

[0085] The optimal set of parameters after convergence is extracted and used as the scalar values ​​of equivalent heat capacity and surface heat transfer coefficient under this specific working condition.

[0086] Through the aforementioned nonlinear fitting and parameter identification processing methods, the transient thermodynamic response data collected in offline experiments are transformed into precise scalar parameters characterizing the thermophysical properties of the insert, thereby realizing the quantitative calibration of the thermal inertial behavior of the insert and providing a reliable data foundation for the subsequent generation of high-precision thermal inertial tags.

[0087] S3.4: The equivalent heat capacity scalar value and the surface heat transfer coefficient scalar value are used as correction factors and substituted into the initial three-dimensional thermal inertial model for parameter assimilation and update, eliminating the deviation between theoretical calculation and actual working conditions, and generating a high-precision calibrated three-dimensional thermal inertial model that can accurately reproduce thermal behavior under dynamic disturbances.

[0088] The initial three-dimensional thermal inertial model mesh data generated in S3.2 and the equivalent heat capacity scalar value and surface heat transfer coefficient scalar value extracted in S3.3 are obtained as input references for parameter assimilation and update.

[0089] The theoretical material specific heat capacity parameter in the initial model is replaced with the equivalent heat capacity scalar value after nonlinear fitting correction, and a corrected heat capacity matrix reflecting the actual heat absorption capacity of the insert is constructed.

[0090] Based on the deviation ratio between the equivalent heat capacity scalar value and the theoretical value, the thermal diffusivity of the grid nodes is locally weighted and adjusted to eliminate the prediction error of thermal conductivity rate caused by batch differences in materials.

[0091] The scalar value of the surface heat transfer coefficient is mapped to the boundary mesh element of the interface between the insert and the melt, replacing the original constant heat transfer boundary condition, and a dynamic convection heat transfer boundary equation is established.

[0092] The updated transient heat conduction control equations are solved using the implicit finite difference method, and the temperature evolution sequence of each node inside the insert during the filling cycle is calculated iteratively.

[0093] By comparing the interface temperature field obtained from iterative calculations with the actual temperature response curves collected from offline experiments, the root mean square error is calculated to evaluate the model fit.

[0094] If the error exceeds the preset convergence threshold, the gradient descent algorithm is used to fine-tune the weight distribution of the equivalent heat capacity and the surface heat transfer coefficient, and the iterative solution process is re-executed.

[0095] The process continues until the deviation between the model's output temperature field and the measured data meets the accuracy requirements, at which point the current parameter combination is solidified, and a high-precision calibrated three-dimensional thermal inertial model is generated.

[0096] By using parameter assimilation and update processing, the theoretical model is transformed into a high-precision calibrated three-dimensional thermal inertial model that can accurately reproduce thermal behavior under dynamic disturbances, thereby achieving the expected technical effect of eliminating the deviation between theoretical calculations and actual operating conditions.

[0097] S3.5: Read the key feature parameter set in the high-precision calibrated three-dimensional thermal inertial model, perform hash encoding processing in combination with the unique identification code of the current production batch, generate an embedded thermal inertial tag containing complete thermophysical property information, and write the embedded thermal inertial tag into the central database to achieve permanent binding with the embedded parts of a specific batch.

[0098] like Figure 3 As shown, step S4 involves: reading the thermal inertia tags of the current batch of inserts and calling the pre-stored thermal boundary correction function to dynamically adjust the mold cooling water circuit control strategy and the local thermal compensation power output during the holding pressure stage, thereby generating cavity temperature field distribution data. Specifically, this includes: S4.1: Based on the insert thermal inertia tag of the current production batch, retrieve the corresponding equivalent heat capacity parameter set and surface heat transfer coefficient parameter set from the central database, and load the retrieved equivalent heat capacity parameter set and surface heat transfer coefficient parameter set as input variables into the thermal boundary correction function execution unit to generate a thermal boundary correction instruction sequence containing local thermal resistance correction factors.

[0099] Read the insert thermal inertia tag of the current production batch and parse the unique batch identification code and pre-stored thermophysical parameter index contained therein.

[0100] Based on the batch's unique identifier, a structured query request is initiated to the central database to retrieve the set of equivalent heat capacity scalar values ​​and the set of surface heat transfer coefficient scalar values ​​associated with that batch.

[0101] The completeness of the retrieved equivalent heat capacity scalar value set is verified to confirm that it covers the heat capacity distribution data of all key geometric partitions of the insert.

[0102] The validity of the retrieved scalar values ​​of surface heat transfer coefficients is verified to ensure that the boundary heat transfer characteristic parameters corresponding to different preheating temperature conditions are correct.

[0103] The set of equivalent heat capacity scalar values ​​that have passed the verification is used as the first input variable and loaded into the memory buffer of the thermal boundary correction function execution unit.

[0104] The set of verified surface heat transfer coefficient scalar values ​​is used as the second input variable and synchronously loaded into the parameter register of the thermal boundary correction function execution unit.

[0105] The pre-stored thermal boundary correction function is invoked, and the equivalent thermal capacity scalar value is mapped to a local thermal resistance correction factor based on the embedding geometry topology mapping relationship.

[0106] The local thermal resistance correction factor is nonlinearly weighted and corrected using the scalar value of the surface heat transfer coefficient, and the thermal boundary compensation coefficient of each mold cavity mesh node is calculated.

[0107] The thermal boundary compensation coefficient is serialized into a standard control command format to generate a local thermal resistance correction factor thermal boundary correction command sequence containing spatial coordinate index and correction value.

[0108] Through the above data processing flow, discrete thermal inertia tags are transformed into an executable sequence of thermal boundary correction instructions, thus realizing accurate feedforward compensation preparation for dynamic disturbances in the mold thermal field.

[0109] For example, for a connector insert batch B202310, its thermal inertia tag is parsed to obtain the ID code. The equivalent heat capacity is retrieved from the database as 2.5 J / (g·K), and the surface heat transfer coefficient is 800 W / (m²·K). These values ​​are then input into the thermal boundary correction function. : in It is 0.8. It is 0.05. For equivalent heat capacity, Surface heat transfer coefficient. The calculated local thermal resistance correction factor is 2.04. An instruction sequence containing this factor is generated for subsequent PID regulation of the cooling water circuit, significantly improving the response accuracy of the thermal field control.

[0110] S4.2: Using the thermal boundary correction instruction sequence containing the local thermal resistance correction factor, perform PID adjustment calculations on the initial flow rate setpoint and initial water temperature setpoint of the mold cooling water circuit to generate a real-time control parameter group for the cooling water circuit with dynamic response characteristics. The real-time control parameter group for the cooling water circuit directly inherits and applies the compensation logic for heat exchange efficiency of the local thermal resistance correction factor.

[0111] Receive a sequence of thermal boundary correction instructions containing a local thermal resistance correction factor, and parse the target temperature deviation and flow compensation coefficient for each independent cooling loop.

[0112] Based on the topology of the mold cooling water circuit, the global thermal resistance correction factor is mapped to the input of the PID control method of the specific cooling circuit, and the temperature setpoint correction term at the current moment is set.

[0113] The system collects real-time water temperature sensor data and flow meter feedback signals from the inlet and outlet of the cooling water circuit, and calculates the instantaneous error between the current actual water temperature and the target water temperature.

[0114] Based on the instantaneous error value, and combined with preset proportional gain, integral time, and derivative time parameters, discrete PID calculation is performed to generate the valve opening adjustment amount.

[0115] The calculated valve opening adjustment is limited to ensure it is within the physically feasible range of the actuator, preventing overshoot or saturation.

[0116] The limited control quantity is converted into a PWM drive signal or an analog voltage signal and sent to the proportional control valve actuator to dynamically adjust the cooling medium flow rate.

[0117] The power setting value of the water heater is updated synchronously, and the feedforward compensation logic is used to offset the heat exchange lag effect caused by the change in flow rate.

[0118] Through the combined effect of PID regulation and feedforward compensation, the thermal boundary correction command is transformed into a set of real-time control parameters for the cooling water circuit with dynamic response characteristics, thereby achieving rapid and precise control of the thermal field distribution in the cavity.

[0119] S4.3: Based on the real-time control parameter group of the cooling water circuit, drive the proportional valve and the heating rod actuator to operate. Simultaneously combine the preheating temperature deviation value in the thermal inertia tag of the insert to perform feedforward compensation calculation on the heater power baseline during the pressure holding stage, so as to generate a local thermal compensation power output curve containing spatial distribution characteristics, ensuring that the local thermal compensation power output curve can offset the additional heat capacity load introduced by the insert.

[0120] The preheating temperature deviation value between the real-time control parameter group of the cooling water circuit and the thermal inertia tag of the insert is read and used as the initial boundary condition for local thermal compensation calculation.

[0121] The preheating temperature deviation is analyzed, and the heat capacity distribution of each node in the embedded space discrete mesh model is combined to calculate the instantaneous heat flux density required to offset the additional heat load.

[0122] A feedforward control method is adopted to construct a basic power compensation matrix based on the spatial power allocation ratio of the heating rod array according to the heat flux density distribution.

[0123] A time decay factor is introduced to dynamically correct the basic power compensation matrix in order to match the nonlinear variation characteristics of the melt solidification rate during the pressure holding stage.

[0124] The instantaneous power values ​​of each region are calculated and integrated according to the time series to generate a local thermal compensation power output curve that includes spatial distribution characteristics and temporal dynamic characteristics.

[0125] By using a feedforward compensation calculation method, static thermal inertia parameters are converted into dynamic power execution commands, thereby achieving the expected technical effect of offsetting the heat capacity load of the insert.

[0126] For example, for a certain type of connector insert, the preheating temperature deviation is -5℃, and the time decay constant is set to 0.02s. -1 At the 10th second of the holding pressure stage, the calculated compensation power for the heating area near the critical gate was 120W, which is 30% higher than the baseline power. This power curve effectively offsets the local rapid cooling caused by the low temperature of the insert, significantly improving the uniformity of the cavity temperature field and reducing the predicted position deviation of the weld line to within 0.05mm.

[0127] S4.4: The local thermal compensation power output curve and the real-time control parameter set of the cooling water circuit after PID regulation are applied together to the finite element model of heat conduction in the mold cavity. The unsteady heat conduction equation is solved iteratively to generate the original temperature field grid data of the cavity that reflects the transient heat distribution under actual working conditions. The original temperature field grid data of the cavity fully integrates the heat source terms provided by the local thermal compensation power output curve and the boundary conditions provided by the real-time control parameter set of the cooling water circuit.

[0128] S4.5: Perform spatial smoothing filtering and outlier removal on the original temperature field grid data of the cavity, and map the processed original temperature field grid data of the cavity into a continuous temperature scalar field to generate the final thermal inertia-compensated cavity temperature field distribution data that can be used for weld line location prediction calculation, thereby realizing the standardized conversion from discrete grid nodes to continuous physical field.

[0129] Step S5: Input the melt rheological fingerprint vector into the graph neural network model to output a two-dimensional probabilistic heatmap. Specifically, this includes: S5.1: Based on the 3D CAD model of the mold runner, extract the central axis and connection relationship, construct the geometric topology graph structure of the mold runner containing the set of nodes and the set of edges, map the physical coordinates of each micro embedded sensor array to graph nodes, and define the runner connectivity as graph edges, generating injection mold runner graph data objects with spatial topology attributes.

[0130] The process involves reading the 3D CAD model data of the mold runner, analyzing its solid geometry and topological connections, and extracting the coordinate sequences of the central axes of the main runner, branch runners, and gates. Based on these coordinate sequences, an undirected graph node set is constructed, defining the injection molding machine nozzle outlet, main runner bifurcation point, branch runner end, and each gate inlet as key topological nodes. An edge set is established based on the physical connectivity of the runners, mapping runner segments between adjacent nodes to graph edges, and calculating the Euclidean distance of each edge as its weight, generating an initial mold runner topology graph. The physical installation coordinates of the miniature embedded sensor array deployed in S1.2 in the mold coordinate system are obtained. Using a spatial nearest neighbor matching method, each sensor array is mapped to the nearest graph node in the topology graph, establishing a sensor-graph node index mapping table. For intermediate topology nodes where sensors are not directly deployed, linear interpolation is used to derive their virtual coordinates based on the geometric positions of adjacent sensor nodes, ensuring the graph structure covers the complete runner network. The spatial 3D coordinates of the nodes are encoded into node attribute vectors, combined with edge weight information, and encapsulated to generate injection runner graph data objects with spatial topological attributes. Through the above processing method, the physical sensor layout of the previous step is transformed into a standardized graph data structure that can be processed by graph neural networks, realizing digital topological mapping of the melt flow path and providing a geometric benchmark for subsequent feature aggregation.

[0131] For example, for a precision connector mold, the flow channel axis containing 12 key nodes is extracted from the CAD model. The nozzle outlet is set as node V1, the ends of the two branch channels are set as V6 and V7, and the four gate inlets are set as V9-V12. A graph structure containing 12 nodes and 15 edges is constructed, with the edge weights equal to the actual flow channel length. Five sets of sensor arrays are mapped to nodes V1, V4, V6, V9, and V10, respectively. For the remaining nodes V2, V3, etc., virtual coordinates are generated by linear interpolation based on the coordinates of V1 and V4. Finally, an injection molding flow channel graph data object containing 12 node attribute vectors (x, y, z coordinates) and 15 weighted edges is generated. The node spacing error is controlled within 0.1 mm, ensuring a high degree of consistency between the topology and the physical flow channel.

[0132] S5.2: For each graph node in the injection runner graph data object, associate the melt rheological fingerprint vector generated in the previous step as the initial node feature matrix, and introduce the insert thermal inertia label as the global context vector. Perform feature splicing and normalization processing to generate an enhanced node feature representation with local rheological state and global thermal boundary conditions.

[0133] S5.3: Based on the enhanced node feature representation, a lightweight graph convolution operator is used to execute a multi-layer message passing mechanism, aggregating the pressure decay rate, temperature integral area, and strain hysteresis phase angle features of adjacent graph nodes, updating the hidden layer state vector of the current graph node, and generating a high-dimensional node embedding vector containing the upstream and downstream melt merging trend.

[0134] The system receives the enhanced node feature representation generated by S5.2, which integrates local rheological fingerprints and global thermal inertia context. A lightweight graph convolutional network layer is constructed, defining a message-passing function to aggregate neighborhood node information. For each graph node, all connected nodes in its first-order neighborhood are identified, and the hidden state vectors of these neighboring nodes are extracted. The state vectors of the neighboring nodes are linearly transformed and mapped to a unified feature space through a learnable weight matrix. The attention coefficients between the current node and each neighboring node are calculated, quantifying the influence weights of melt flow in different directions on the current node's state. The attention coefficients are normalized using the Softmax function to ensure the sum of the weights is 1, forming a standardized attention distribution matrix. The normalized attention weights are weighted and summed with the corresponding neighboring node feature vectors to generate aggregated neighborhood feature information. A residual connection mechanism is introduced to concatenate or add the original features of the current node with the aggregated neighborhood features, preventing the gradient vanishing problem in deep networks. The fused feature vectors are mapped using the nonlinear activation function LeakyReLU to enhance the model's ability to fit complex rheological nonlinear relationships. Multi-layer iterative updates are performed, with the output of each layer serving as the input for the next, progressively expanding the receptive field to capture long-range dependencies. In the final graph convolution operation, the output contains a high-dimensional node embedding vector that includes the merging trend of upstream and downstream melts.

[0135] Through the aforementioned multi-layer message passing mechanism, the dispersed local rheological features are transformed into high-dimensional node embedding vectors containing global topological relationships, thereby achieving accurate characterization of the weld line formation path.

[0136] S5.4: Apply the readout function to the high-dimensional node embedding vector to perform graph-level feature pooling operation, map the scattered node states to the weld line formation probability value of the cavity planar grid unit, and fill the grid gaps through spatial interpolation method to generate two-dimensional probability heat map data of weld line position covering the entire cavity area.

[0137] The system receives a set of high-dimensional node embedding vectors output from S5.3, which represents the melt rheological state and upstream and downstream merging trend information at each sensing node. A two-dimensional regular mesh matrix corresponding to the physical dimensions of the mold cavity plane is constructed, and the mesh resolution is set to balance computational accuracy and real-time requirements, establishing a spatial reference coordinate system for probability mapping.

[0138] A global readout function is used to aggregate the node features output by the graph neural network, mapping the discretely distributed node embedding vectors to a continuous spatial domain. Using bilinear interpolation, the weight contribution value of each sensing node in adjacent grid cells is calculated based on its physical coordinates on the cavity plane, achieving a coordinate transformation from graph topology space to Euclidean space.

[0139] Define a weld line formation probability activation function to calculate the initial weld line existence probability of each grid cell based on the aggregated node feature vectors.

[0140] To address the sparse data gaps generated during the gridding process, the Kriging spatial interpolation method is applied to fill in the probability values ​​of the uncovered areas. By combining the probability gradient information of neighboring already calculated grid nodes, the probability distribution trend of the blank areas is estimated, ensuring the spatial continuity and smoothness of the heatmap data.

[0141] The generated preliminary probability matrix is ​​normalized to map the probability values ​​to the range of 0 to 1, eliminating the influence of dimensional differences on subsequent threshold determination. A two-dimensional probability heatmap of weld line locations covering the entire cavity area is generated, visually presenting the spatial distribution density of possible weld lines.

[0142] By employing the aforementioned readout aggregation, coordinate mapping, probabilistic activation, and spatial interpolation methods, the dispersed high-dimensional node embedding vectors from the previous step are transformed into two-dimensional probabilistic heatmap data with clear physical spatial meaning, achieving the expected technical effect of transforming abstract graph features into a visualized distribution of weld line risks.

[0143] For example, the cavity plane mesh resolution is set to 100×100, with a total of 10,000 mesh elements. The number of sensor nodes, N, is 5, and the physical coordinates of each node are known. The Gaussian kernel function bandwidth parameter is set to 5 mm. The calculated mesh element probability value is 0.85 in the central region and 0.12 in the edge region. After Kriging interpolation, the heatmap has no data holes, and the probability distribution transitions smoothly. After normalization, the maximum probability value is 1.0, and the minimum is 0.0. This heatmap clearly identifies the high-risk area where the two melt fronts converge, providing accurate input for the thermal field superposition in step S6.

[0144] S5.5: Based on the two-dimensional probability heat map data of the weld line location, extract the centroid coordinates of the probability density peak region as prediction candidate points, and filter low probability noise points by combining the confidence threshold, and output the final two-dimensional probability heat map of the weld line on the cavity plane as the input basis for subsequent spatial superposition weighted calculation.

[0145] Receive two-dimensional probability heat map data of weld line location covering the cavity region generated by S5.4. This data characterizes the macroscopic melt convergence probability distribution inferred based on rheological fingerprint.

[0146] The connected component analysis method is applied to the two-dimensional probability heatmap to identify local bright areas in the graph with probability values ​​higher than a preset confidence threshold. Isolated low-probability noise pixels caused by sensor noise or model edge effects are removed, and a set of candidate weld line formation regions is generated.

[0147] For each candidate region, calculate the probability-weighted centroid coordinates of all pixels within it, and determine the region centroid using the following formula: in, The x-coordinate of the centroid The x-coordinate is the pixel coordinate. This represents the probability value for the corresponding pixel.

[0148] Similarly, the ordinate is calculated to converge the irregular probability distribution region into a single geometric feature point, generating a preliminary set of candidate prediction points.

[0149] The peak probability intensities of each candidate point are sorted, and the point with the highest probability density and spatial topology that conforms to the flow channel merging logic is selected as the primary prediction point, while the remaining points are marked as secondary reference points.

[0150] The probability distribution data of the main prediction point and its neighborhood are standardized and encapsulated to output the final two-dimensional probability heat map of the weld line on the cavity plane and the core candidate coordinates, which serve as the input basis for subsequent spatial superposition and weighted calculation with the thermal inertia compensation temperature field.

[0151] By extracting the probability peak and locating the centroid, the discrete probability field generated in the previous step is transformed into spatial coordinate data with clear physical orientation, realizing the transformation from statistical probability to deterministic location indicators, significantly reducing the computational complexity of subsequent multi-source data fusion and improving positioning accuracy.

[0152] Step S6: Spatially superimpose and weight the two-dimensional probability heat map with the cavity temperature field distribution data to generate the final predicted coordinate set of weld line location. Specifically, this includes: S6.1: Perform gridded resampling on the two-dimensional probability heatmap to unify the spatial resolution and generate a standard discrete probability matrix, ensuring that it is strictly aligned with the cavity temperature field distribution data in the spatial dimension.

[0153] S6.2: Based on the standard discrete probability matrix and the cavity temperature field distribution data after thermal inertia compensation, the thermal disturbance correction coefficient of each grid node is calculated using the thermal-fluid coupling weight function to quantify the influence of the local temperature gradient on the microscopic offset of the melt solidification front.

[0154] Read the standard discrete probability matrix and the cavity temperature field distribution data after thermal inertia compensation, and extract the temperature value of each grid node and its gradient vector in the spatial domain.

[0155] The temperature gradient modulus at each grid node is calculated to characterize the severity of local thermal field changes, serving as a fundamental physical quantity for measuring the intensity of disturbance at the melt solidification front.

[0156] A reference temperature threshold is introduced, which is set as a weighted average of the glass transition temperature of the polymer material and the temperature of the mold wall, to define the effective thermal disturbance region.

[0157] A heat-fluid coupling weight function is constructed, which is based on an exponential decay model and maps the deviation between the temperature gradient modulus and the reference temperature threshold into a nonlinear correction coefficient.

[0158] The calculated correction coefficients are normalized to ensure that their values ​​are between 0 and 1, reflecting the suppression or enhancement effect of the local thermal field on the offset of the weld line position.

[0159] The normalized thermal perturbation correction coefficients are assigned to the corresponding grid nodes to form a thermal perturbation correction coefficient matrix with the same dimensions as the standard discrete probability matrix.

[0160] By using nonlinear mapping processing of the thermal-fluid coupling weight function, the temperature field data from the previous step is transformed into thermal disturbance correction coefficients that quantify the influence of local micro solidification offset, thus achieving a precise mathematical description of the positional deviation caused by the thermal inertia of the insert.

[0161] S6.3: Apply the thermal perturbation correction coefficient to the standard discrete probability matrix to perform a point-by-point weighted fusion operation to counteract the non-uniform cooling effect caused by the thermal inertia of the insert, and generate a comprehensive confidence map of weld lines containing micro solidification offset correction information.

[0162] The standard discrete probability matrix and thermal perturbation correction coefficient matrix are read to ensure a strict one-to-one correspondence between them in the spatial grid index. The coordinates of each node in the cavity plane grid are traversed, extracting the base probability value of the weld line formation and the corresponding thermal perturbation correction coefficient at that location. A nonlinear weighted fusion model is constructed, using the base probability value as the multiplicand and the thermal perturbation correction coefficient as the adjustment factor, performing point-by-point multiplication to generate local corrected probability values. For high thermal resistance regions at the insert edge, if the thermal perturbation correction coefficient is greater than 1, the confidence of weld line formation in that region is enhanced to compensate for the prediction underestimation caused by excessively rapid cooling. For low thermal resistance regions in the center of the main channel, if the thermal perturbation correction coefficient is less than 1, the background noise probability in that region is suppressed to eliminate the prediction overestimation caused by uniform temperature. The local corrected probability values ​​of all grid nodes are reassembled to construct a comprehensive confidence map of weld lines covering the entire cavity plane. By using a point-by-point weighted fusion processing method, the probability distribution and thermal field deviation of the previous step are transformed into comprehensive confidence data containing micro solidification offset correction information, thereby achieving high-precision spatial mapping of weld line positions under dynamic working conditions.

[0163] For example, for a cavity mesh with a resolution of 100x100, the base probability value of a node is 0.65, and the thermal perturbation correction factor is 1.12. The corrected probability value is calculated to be 0.728. The same operation is performed on all 10,000 nodes of the mesh to generate a comprehensive confidence map. At the corner of the insert, the correction factor is increased to 1.25, significantly enhancing the detectability of weak signals and eliminating the prediction blind spot caused by local rapid cooling.

[0164] S6.4: Perform a global extreme value search based on the comprehensive confidence map of the weld line to locate the confidence peak region and extract its geometric center coordinates to generate a preliminary set of candidate weld line locations.

[0165] A comprehensive confidence profile of weld lines after thermal inertia compensation is obtained. This profile represents the probability density distribution of weld lines at various locations on the cavity plane in the form of a discrete grid.

[0166] Morphological filtering preprocessing is performed on the comprehensive confidence map, and Gaussian smoothing operator is used to suppress high-frequency noise interference introduced by grid discretization, while preserving the global topological structure features of the probability distribution.

[0167] Construct a global extreme value search window and set the search radius based on the connector insert geometry to ensure coverage of the entire spatial range where potential weld lines may form.

[0168] Traverse all grid nodes within the search window, extract local probability density maxima, and form a preliminary set of candidate peaks.

[0169] For each candidate peak point, calculate the connected regions in its neighborhood with probability values ​​higher than a preset threshold, and define the spatial boundaries where weld lines may exist.

[0170] The centroid algorithm is used to calculate the geometric center coordinates of each connected region, mapping the spatial centroid of the probability density distribution to specific two-dimensional plane coordinate points.

[0171] Redundancy is eliminated from the generated candidate point set, adjacent candidate points with a spatial distance less than the minimum resolution are merged, and representative points with the highest cumulative probability value are retained.

[0172] By using global extremum search and centroid localization, the comprehensive confidence map from the previous step is transformed into a preliminary set of candidate weld line locations containing spatial location information, realizing a structured transformation from a probability field to discrete coordinate points, and providing an initial solution for subsequent sub-pixel level precise localization.

[0173] For example, the resolution of the comprehensive confidence map is set to 100x100 grids, and the standard deviation of the Gaussian smoothing operator is set to 1.5. The search radius is set to 5mm, corresponding to 10 grid cells at the grid step size. The preset confidence threshold is set to 0.6. After traversing the map, three local maxima regions are identified, located near coordinates (25, 30), (75, 80), and (50, 50), respectively. For the first region, 12 connected grids with a probability value greater than 0.6 are extracted, and its weighted centroid coordinates are calculated to be (25.4, 30.2). The centroids of the other two regions are calculated similarly. After redundancy detection, the distance between the three candidate points is greater than the minimum resolution of 2mm, and they are retained. The final output contains a preliminary candidate point set with three coordinates, which significantly reduces the search space for the weld line location and improves the convergence speed of subsequent precise localization.

[0174] S6.5: Perform sub-pixel level interpolation optimization on the preliminary weld line location candidate point set to smooth the discrete grid error and accurately locate the point with the maximum probability density, and finally output the final weld line location prediction coordinate set.

[0175] The preliminary weld line location candidate point set and its neighborhood grid data in the comprehensive confidence map are obtained as input objects for sub-pixel level interpolation optimization.

[0176] A local search window is constructed with the candidate point as the center. The coordinate values ​​of each discrete grid node within the window and the corresponding comprehensive confidence scores are extracted to form a local probability distribution dataset.

[0177] A biquadratic polynomial surface fitting method is used to establish a nonlinear mapping relationship between grid coordinates and comprehensive confidence, and to construct an overdetermined system of equations to be solved.

[0178] The overdetermined system of equations is solved using the least squares method, and the polynomial coefficient matrix is ​​calculated, thereby determining the analytical expression of the local probability surface.

[0179] Take the first-order partial derivatives of the analytical expression with respect to the horizontal and vertical coordinates, and set the partial derivatives to zero to construct a system of equations for locating the extreme points.

[0180] Solving the above system of equations yields the coordinates of the sub-pixel extreme points, which are the maximum probability density points after eliminating discrete grid quantization errors.

[0181] The calculated subpixel-level coordinates are encapsulated into a standard floating-point number format to generate the final predicted coordinate set of weld line positions.

[0182] By using a biquadratic polynomial surface fitting and analytical extremum solution process, the discrete grid candidate points from the previous step are transformed into high-precision coordinate data in continuous space, achieving the expected technical effect of improving the resolution of weld line location prediction from the millimeter level to the micrometer level.

[0183] Step S7: Generate a real-time compensation control signal based on the predicted coordinate set of the final weld line position. Specifically, this includes: S7.1: Obtain the final predicted coordinate set of the weld line position and the preset standard position data, calculate the spatial position deviation vector between the two based on the Euclidean distance method, and generate the actual offset value of the weld line.

[0184] S7.2: Read the actual offset value of the weld line and the preset deviation threshold parameter, use a multi-level logic comparator to perform threshold judgment operation, and generate a gate switching trigger flag or a pressure holding curve replanning enable signal.

[0185] Read the actual offset value of the weld line generated by S7.1 and input it as the variable to be judged into the multi-level logic comparator. Load the pre-stored deviation threshold parameter set, which includes a first-level early warning threshold and a second-level intervention threshold, corresponding to slight deviation and severe defect conditions, respectively.

[0186] The first-level logic judgment is performed using a comparator, comparing the actual offset value with the first-level warning threshold. If the actual offset is less than the first-level warning threshold, the current weld line position is determined to be within the allowable tolerance range, and a silent flag is generated to maintain the original process parameters.

[0187] If the actual offset is greater than or equal to the first-level warning threshold but less than the second-level intervention threshold, the second-level logic judgment branch is triggered. The system identifies the microscopic solidification offset as caused by moderate thermal field disturbance, activates the pressure holding curve replanning enable signal, and marks that the pressure gradient during the pressure holding stage needs to be locally corrected.

[0188] If the actual offset is greater than or equal to the secondary intervention threshold, the third-level logic judgment branch is triggered. The system identifies it as an abnormal macroscopic rheological path or a severe mismatch in insert thermal inertia, activates the gate switching trigger flag, and marks that the backup gate opening or main gate closing action must be executed immediately to change the melt confluence point.

[0189] The final control flag state is determined using the following logical expression: in, Select factors for control mode. This represents the actual offset of the weld line. The threshold for Level 1 warning is... This is the threshold for secondary intervention.

[0190] Based on the value of the control mode selection factor w, a specific execution instruction flag is generated. When w is 1, the pressure holding curve replanning enable signal is output to the pressure control module; when w is 2, the gate switching trigger flag is output to the valve needle drive module.

[0191] By using multi-level threshold judgment and logical branch processing, continuous spatial position deviations are transformed into discrete control decision signals, enabling graded responses to weld line offsets of varying severity, thus ensuring the accuracy and timeliness of compensation control.

[0192] S7.3: Based on the gate switching trigger flag, call the pre-stored gate timing mapping table, or based on the pressure holding curve replanning enable signal, load the local pressure correction function to generate the target gate switching instruction sequence or the target pressure holding curve replanning parameter set.

[0193] The gate switching trigger flag or the pressure holding curve replanning enable signal output from S7.2 is used as the basis for judging the execution logic branch.

[0194] If the gate switching trigger flag is detected to be in a valid state, the system immediately retrieves the pre-stored gate timing mapping table from the central database. This table records the optimal gate opening time sequence corresponding to different weld line offset directions.

[0195] Based on the offset vector direction determined by the predicted coordinate set of the final weld line position, the index of the closest standard compensation strategy is matched in the gate timing mapping table, and the corresponding target gate opening delay time and duration parameters are extracted.

[0196] The extracted time parameters are encapsulated into a standardized digital instruction sequence, and the action timing of each auxiliary gate hydraulic valve is explicitly specified to form a target gate switching instruction sequence to change the melt confluence point position.

[0197] If the pressure holding curve replanning enable signal is detected to be in an effective state, the system loads a local pressure correction function, which is constructed based on the pressure decay rate characteristics in the melt rheological fingerprint vector.

[0198] The Euclidean distance deviation between the predicted weld line location and the standard location is calculated and used as the input variable of the correction function to dynamically adjust the pressure holding time and peak pressure value during the pressure holding stage.

[0199] The obtained pressure correction is superimposed on the original holding pressure reference curve to generate a new holding pressure-time curve that includes local pressure enhancement or weakening characteristics, forming the target holding pressure curve reprogramming parameter set.

[0200] By using table lookup matching or function operation, the logical judgment result of the previous step is transformed into a specific set of process execution parameters, thereby achieving active compensation control of the weld line position.

[0201] S7.4: Integrate the target gate switching command sequence and the target holding pressure curve replanning parameter set, encapsulate them into a standardized control frame through the industrial fieldbus protocol, and generate a real-time compensation control signal containing specific execution action information.

[0202] Step S8: Compensation control for the predicted injection weld line position is completed based on the real-time compensation control signal. Specifically, this includes: S8.1: Obtain the gate switching command or the local replanning action parameters of the holding pressure curve from the real-time compensation control signal, and map the action parameters to the drive level signal of the underlying actuator based on the communication protocol of the injection molding machine closed-loop control system, and generate a real-time process adjustment command set containing timestamps and action types.

[0203] S8.2: Receive the real-time process adjustment instruction set and drive the hydraulic servo valve or electric screw of the injection molding machine to perform corresponding actions. Simultaneously collect the finished product image data of the connector insert after the current injection molding cycle ends. Use machine vision measurement methods to extract the actual physical coordinate position of the weld line in the finished product and generate the actual molding quality feature vector of the weld line.

[0204] S8.3: Read the final weld line position prediction coordinate set generated in this cycle and the actual forming quality feature vector of the weld line, calculate the Euclidean distance deviation and angular offset between the two on the cavity plane, and construct a cycle quality difference dataset containing prediction error amplitude and direction information.

[0205] S8.4: Based on the aforementioned periodic quality difference dataset and the insert thermal inertia labels of the current batch, the equivalent heat capacity correction coefficient and surface heat transfer coefficient correction factor in the thermal boundary correction function are retrieved using the recursive least squares estimation method, and a thermal boundary parameter update gradient vector is generated for the next period calibration.

[0206] The periodic quality difference dataset constructed in S8.3 is read, and the prediction error magnitude vector and angular offset contained therein are extracted as target observation variables for recursive least squares estimation. The insert thermal inertia label corresponding to the current production batch is loaded, and the initial equivalent heat capacity scalar value and surface heat transfer coefficient scalar value are parsed and combined into the thermal boundary parameter state vector to be corrected. A first-order linearized observation equation based on Taylor expansion is constructed, mapping the weld line position deviation to a linear function relationship of thermal boundary parameter perturbation, establishing the mathematical model foundation for parameter identification. The covariance matrix and forgetting factor of the recursive least squares method are initialized, and the forgetting factor value is set to give higher weight to recent data, enhancing the model's ability to track dynamic operating condition changes. The Kalman gain matrix is ​​calculated using the position deviation residual of the current cycle, quantifying the correction strength of the new observation data on the parameter estimates, achieving effective weighting of the new information. Based on the calculated Kalman gain and observation residual, the estimated values ​​of the equivalent heat capacity correction coefficient and the surface heat transfer coefficient correction factor are updated, completing the single-step parameter iteration. The covariance matrix of the parameter estimation error is updated synchronously to reflect the convergence trend of parameter uncertainty, providing prior information for gain calculation in the next cycle. The updated correction coefficients and correction factors are combined to generate a thermal boundary parameter update gradient vector pointing to the optimal thermal boundary conditions. Through the recursive least squares estimation method, the mass difference data from the previous step is transformed into accurate thermophysical parameter correction gradients, realizing adaptive calibration of the thermal boundary model and continuous optimization of prediction accuracy.

[0207] S8.5: Use the thermal boundary parameters to update the gradient vector and perform weighted iterative correction on the pre-stored thermal boundary correction function model parameters, output the updated thermal boundary correction function version, and write the updated thermal boundary correction function version into the central database to overwrite the old version parameters, thus completing the adaptive closed-loop optimization process.

[0208] Receive the thermal boundary parameter update gradient vector generated by S8.4, which includes the increment of the equivalent heat capacity correction coefficient and the increment of the surface heat transfer coefficient correction factor. Load the pre-stored thermal boundary correction function model and read the current version's baseline parameter set, including the initial equivalent heat capacity value and the initial surface heat transfer coefficient value.

[0209] A weighted iterative correction method is employed, multiplying the updated gradient vector by the learning rate decay factor to calculate the parameter adjustment step size for this iteration. The learning rate decay factor is dynamically determined based on historical convergence curves, ensuring that the step size is reduced as the solution approaches optimal to avoid oscillations.

[0210] The calculated parameter adjustment steps are then superimposed onto the current equivalent heat capacity and surface heat transfer coefficient values ​​to generate an updated thermophysical parameter set. The parameter update is performed using the following formula: in, For the updated thermal boundary parameters, Given the current thermal boundary parameters, For adaptive learning rate, Update the gradient vector for the thermal boundary parameters.

[0211] Physical constraint verification is performed on the updated thermophysical parameter set to ensure that the equivalent heat capacity is within the material's theoretical range and that the surface heat transfer coefficient meets the convective heat transfer limit condition. If a parameter exceeds the physically feasible region, it is truncated to the boundary value and an anomaly flag is recorded to prevent model divergence.

[0212] The validated set of thermophysical parameters is encapsulated into a new version of the thermal boundary correction function, and assigned a unique version number and timestamp. The new version function is written to the central database via the industrial Ethernet protocol, overwriting the old version parameter index, thus completing the persistent storage of the model parameters.

[0213] By using weighted iterative correction and physical constraint verification, the gradient vector from the previous step is transformed into an updated thermal boundary correction function with physical consistency, thereby achieving adaptive closed-loop optimization of the thermal field prediction model and significantly improving the prediction accuracy of weld line locations in subsequent cycles.

[0214] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0215] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0216] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A compensation control method based on the prediction of the injection molding weld line position of connector inserts, characterized in that, Specifically, it includes: S1: Simultaneously collect transient pressure peaks, temperature rise slopes, and local cavity micro-deformations of the melt as it passes through the injection molding machine nozzle to the main runner inlet section, the end of the branch runner, and the leading edge of each gate to generate the original rheological response sequence. S2: Process the original rheological response sequence to obtain the melt rheological fingerprint vector; S3: Based on the three-dimensional geometry and material properties of the connector insert, a thermal inertia model is established. Using the equivalent heat capacity and surface heat transfer coefficient data calibrated by offline experiments, an insert thermal inertia tag is generated that is bound to each batch of inserts. S4: Read the thermal inertia tag of the current batch of inserts and call the pre-stored thermal boundary correction function to dynamically adjust the mold cooling water circuit control strategy and the local thermal compensation power output during the holding pressure stage, and generate cavity temperature field distribution data. S5: Input the melt rheological fingerprint vector into the graph neural network model to output a two-dimensional probability heat map; S6: Spatial superposition and weighted calculation are performed on the two-dimensional probability heat map and the cavity temperature field distribution data to generate the final predicted coordinate set of weld line location; S7: Generate a real-time compensation control signal based on the predicted coordinate set of the final weld line position; S8: Compensation control for predicting the injection weld line position is completed based on the real-time compensation control signal.

2. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 1, characterized in that, The synchronous acquisition is achieved by deploying miniature embedded sensor arrays at the nozzle to the main runner inlet section, the end of the branch runner, and the leading edge of each gate.

3. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 1, characterized in that, The process of processing the original rheological response sequence to obtain the melt rheological fingerprint vector includes: The original rheological response sequence is segmented into sliding windows, and three scalar features—pressure decay rate, temperature integral area, and strain hysteresis phase angle—are extracted from the window and combined to generate a melt rheological fingerprint vector with a temporal topology.

4. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 1, characterized in that, The step of inputting the melt rheological fingerprint vector into a graph neural network model and outputting a two-dimensional probability heatmap includes: Based on the geometric topology of the mold flow channel, each acquisition node is mapped as a graph node. The melt rheological fingerprint vector is used as the node attribute input to the graph neural network model. The features of upstream and downstream nodes are aggregated through the message passing mechanism, and a two-dimensional probabilistic heat map of the weld line on the cavity plane is output.

5. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 1, characterized in that, The step of generating a real-time compensation control signal based on the predicted coordinate set of the final weld line position includes: The deviation between the predicted coordinate set of the final weld line position and the preset standard position is calculated. If the deviation exceeds the set threshold, the gate switching command and the local replanning action of the holding pressure curve are triggered. Real-time compensation control signal is obtained based on the gate switching command and the local replanning action of the holding pressure curve.

6. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 1, characterized in that, The S3 specifically includes: The three-dimensional geometric structure data and basic material property parameters of the connector insert are obtained. The three-dimensional geometric structure data are discretized based on the finite element mesh generation technology to generate a discrete mesh model of the insert space containing node coordinates and element topology relationships. An initial three-dimensional thermal inertial model is obtained using the embedded spatial discrete mesh model and the basic material property parameters; Based on offline injection molding experiments with multiple batches of different materials and insertion timing, real-time temperature response curves and pressure change sequences at the interface between the melt and the insert were collected. The real-time temperature response curves and pressure change sequences were subjected to nonlinear fitting processing to extract the equivalent heat capacity scalar value and the surface heat transfer coefficient scalar value that reflect the actual heat transfer efficiency. The equivalent heat capacity scalar value and the surface heat transfer coefficient scalar value are used as correction factors and substituted into the initial three-dimensional thermal inertial model for parameter assimilation and update, generating a high-precision calibrated three-dimensional thermal inertial model. An insert thermal inertial label is generated based on the high-precision calibrated three-dimensional thermal inertial model.

7. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 6, characterized in that, The process of obtaining the initial three-dimensional thermal inertia model using the embedded spatial discrete mesh model and the basic material property parameters includes: Using the discrete mesh model of the insert space as the computational domain, and combining the basic material property parameters, a transient heat conduction control equation is constructed. The internal temperature gradient distribution under different preheating temperature conditions is solved by numerical iteration, and an initial three-dimensional thermal inertial model characterizing the heat storage and transfer capacity inside the insert is generated.

8. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 6, characterized in that, The step of generating an insert thermal inertial tag based on the high-precision calibrated three-dimensional thermal inertial model includes: The key feature parameter set in the high-precision calibrated three-dimensional thermal inertial model is read, and hash encoding is performed in combination with the unique identification code of the current production batch to generate an embedded thermal inertial tag containing complete thermophysical property information. The embedded thermal inertial tag is then written into the central database.

9. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 1, characterized in that, The step of spatially superimposing and weighting the two-dimensional probability heat map with the cavity temperature field distribution data to generate the final weld line location prediction coordinate set includes: The cavity temperature field distribution data is generated by finite element superposition and iteration, and aligned with the two-dimensional probabilistic thermogram space of the weld line location. After meshing, the data is fused point by point and output sub-pixel level accurate coordinates. The sub-pixel level accurate coordinates are then fitted with biquadratic polynomial surfaces and subjected to extreme value analysis to obtain the final predicted coordinate set of the weld line location.

10. The compensation control method based on the prediction of the injection molding weld line position of connector inserts according to claim 1, characterized in that, The graph neural network model contains at least three layers of graph convolution, the weight matrix is ​​initialized with a normal distribution, and the node features are normalized and regularized.

Images