Injection mold for multi-color injection molding and production process thereof

Abstract

The invention discloses an injection mold for multicolor injection molding and a production process thereof, a set of machine vision on-line inspection system with pattern detection and surface flatness detection is integrated, pattern detection is performed on a specific part of a formed part A through optical screen detection, horizontal and vertical coordinates of an actual contour are compared with a preset range, and the pattern of the specific part A is detected. Then, the surface flatness of the part is objectively evaluated by analyzing reflected light distribution on the surface of the part, whether a qualified matrix is provided for second injection or not is judged, and key quality attribute judgment on the intermediate product A before second injection is achieved; according to the method, the waste of raw materials, energy and working hours caused by the fact that known waste products continue to be subjected to second injection is effectively prevented, meanwhile, the position precision and visual definition of final multi-color products at the joints of different colors are guaranteed, the defects of boundary line distortion, blooming and the like are overcome, and the production yield and the resource utilization rate are remarkably increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of injection molding, in particular to an injection molding mold for multi-color injection molding and a production process thereof. BACKGROUND

[0002] As disclosed in the published patent CN112549439B, the multi-color injection molding process is currently the leading injection molding process in the world. The bright and colorful packaging provides consumers with unlimited visual enjoyment, and it is a revolutionary upgrade for cosmetic, daily chemical, food and beverage packaging, and other daily necessities packaging. It also brings a great breakthrough in the anti-counterfeiting technology of injection molded products.

[0003] Compared with single-color injection molded parts, using two or more colors on the same injection molded part, distinguishing different positions on the part becomes the first choice for people, which puts forward higher requirements for injection molded products, especially multi-color injection molded parts, which are the process technology that the current injection molding industry urgently needs to overcome. Using different colors on the same product poses a great challenge to the injection molding process. In order to overcome the problem of multi-color injection molding of a product, one-time injection molding process and two or more times injection molding process are currently used. The one-time injection molding process requires very high injection molding conditions, and at most only two to three colors can be used on the same injection molded part. The two or more times injection molding process can use multiple colors, but the injection molding equipment and process structure are complex and difficult to operate, which is not conducive to widespread use.

[0004] In related technologies, although multi-color injection molding molds, multi-color injection molding machines and injection molding processes are provided, they meet the requirements of using different colors on the same injection molded part to some extent. However, the existing technical solutions solve the problem of blending different colors of the same injection material, and cannot achieve the problem of mutual blending of different injection materials and different colors, as well as the problem of avoiding color bleeding between different colors. More importantly, whether the division line at the junction of different injection materials and different colors meets the design requirements and whether it can ensure that the division line does not shift, distort or have steps.

[0005] In summary, in the multi-color injection molding process, multiple injection operations are required. In related technologies, the product quality is generally detected after multiple injection moldings. However, the applicant points out that in the actual processing process, after the first color injection, if the product A has problems, it needs to be evaluated whether the second color injection is needed. Because the problem of product A cannot be overcome, after the second color injection, product B will also have problems with a high probability. In the case of the first color injection being a waste product, the second color injection not only wastes raw materials but also wastes labor costs. SUMMARY

[0006] The present application is to overcome the above-mentioned deficiencies, and aims to provide a production process for multi-color injection molding which can solve the above-mentioned problems.

[0007] To achieve the above-mentioned purposes, the present application provides the following technical solutions: A production process for multi-color injection molding, comprising the following steps: S100: performing a first injection to form a molded part A; S200: performing machine vision inspection on the molded part A, including a light screen detection process, wherein the light screen detection process comprises: S210: performing pattern detection on a specific part of the molded part A, including horizontal coordinate comparison and vertical coordinate comparison, wherein the horizontal coordinate comparison is to compare the actual horizontal coordinate of the specific part with the preset horizontal coordinate range to detect pattern deviation, and the vertical coordinate comparison is to compare the actual vertical coordinate of the specific part with the preset vertical coordinate range to detect pattern deviation; S220: performing surface flatness detection on the molded part A by analyzing the distribution of reflected light on the surface of the molded part A to evaluate whether the surface flatness meets the requirements of the second injection; S300: based on the results of the machine vision inspection, if the pattern matching rate or the surface flatness is detected to be lower than the threshold, stopping the subsequent injection process; otherwise, performing a second injection.

[0008] As a further scheme of the present application, the step S100 comprises the following steps: S110: drying and heating the thermoplastic material required for the first injection to a molten state; S120: closing the mold to form a cavity of the molded part A, and controlling the mold temperature; S130: injecting the molten material into the cavity, and performing pressure maintaining after filling is completed; S140: opening the mold after the molded part A is cooled and shaped in the cavity.

[0009] As a further scheme of the present application, the step S210 comprises the following steps: S211: before injection, driving a digital twin model of the molded part A to perform synchronous thermomechanics-structural mechanics simulation according to the real-time process parameter flow of the first injection, to predict and generate a dynamic three-dimensional reference model containing material shrinkage and cooling deformation under the specific production cycle in real time; S212: using a scanning device combined with white light interference to perform three-dimensional scanning on the specific part to obtain a two-dimensional image of the surface of the molded part A; A scanning device combining spectral domain optical coherence tomography is used to perform three-dimensional scanning on the specific part, penetrating the surface layer of the transparent or semi-transparent first-electrode material to obtain the internal structure profile image within a depth range of 0.1-0.5 mm below the subsurface, thereby simultaneously capturing surface contour and subsurface defect information. S213: Fuse the surface 3D point cloud data obtained in step S212 with the subsurface profile image data to construct an extended 3D dataset containing surface and subsurface information; The extended 3D dataset is aligned with the dynamic 3D benchmark model generated in step S211 in four-dimensional spatiotemporal coordinates and time dimension based on injection timing. S214: Input the aligned extended 3D dataset into a pre-trained deep learning convolutional neural network; The network analyzes surface geometric features, subsurface structural features and their spatial correlations, and directly outputs a comprehensive confidence score regarding whether the style of the specific part is qualified or not. The horizontal coordinate comparison and vertical coordinate comparison are embedded as primary features in the underlying perception module of the network. The network's decision goes beyond simple coordinates and is based on a deep understanding of the overall shape and the integrity of the internal structure. S215: The comprehensive confidence score obtained in step S214 and the identified defect feature types are fed back to the injection molding machine control system in real time; if the pattern is qualified, the second injection preparation is triggered. If there is a compensable pattern deviation, a feedforward control command is generated to automatically fine-tune the injection parameters of the second shot to actively compensate for the minor deformations present in the first shot during the second shot injection; at the same time, all data from this test are recorded to the digital twin model to optimize the accuracy of subsequent simulation predictions.

[0010] As a further aspect of the present invention, step S220 includes the following steps: S221: After the molded part A completes its first injection and enters the inspection station, the following operations are performed simultaneously: S221a: Structured light in the short-wave infrared band is projected onto the surface to be inspected, and its deformation image is acquired; S221b: Using a high-resolution thermal imaging camera, the temperature field distribution map of the surface of the part is acquired to obtain its real-time thermodynamic state during the cooling process; S222: The short-wave infrared structured light image acquired in S221a is used to reconstruct a high-sensitivity three-dimensional micro-topography point cloud through phase calculation; the topography point cloud is then fused with the surface temperature field data acquired in S221b at the pixel level to generate a multi-physics point cloud model containing the coordinates (X, Y, Z) of each spatial point, the temperature (T), and the local shrinkage stress (σ) of the material derived from the temperature. S223: Input the multiphysics point cloud model into a second injection molding process simulator. The simulator uses the precise morphology, temperature and stress state of the current first injection surface as initial conditions to simulate the entire process of the second injection melt flowing, filling, cooling and bonding with the first injection, and outputs the predicted bonding line strength, coverage integrity and potential weld line location and other key quality indicators. S224: Make dynamic decisions based on the simulation prediction results of S223: If the predicted quality of the second shot fully meets the design requirements, then even if there is a slight deviation in geometric flatness, it is judged as qualified. If the prediction results show that the bonding quality is at a critical state, the system will automatically generate a personalized flatness tolerance for the specific part after tightening, and make a second fine judgment. If the prediction results show that the bonding quality is unqualified, it will be directly judged as unqualified; S225: For parts judged as qualified but with suboptimal predicted bonding quality, the system reverse-engineers the simulation results to generate a set of optimized second injection parameters, including but not limited to injection speed, melt temperature, and holding pressure. These parameters are then fed forward to the injection molding machine control system for the upcoming second injection, in order to proactively improve the quality of the final product. At the same time, all data from this test and the simulation prediction results are compared with the actual product quality after the second injection for continuous training and optimization of the simulator's prediction accuracy.

[0011] As a further aspect of the present invention: in step S200, if the result of the machine vision inspection is in a critical state or the confidence level is lower than the threshold, the augmented reality-assisted human visual inspection step S300 is simultaneously initiated, wherein S300 includes: S310: When the quality inspector wears augmented reality glasses and looks at the molded part A, the glasses perform the following steps: S311: Track and locate parts in real time through image recognition, and overlay the machine vision results from steps S210 and S220 onto the corresponding positions of the physical object in the form of highlighted outlines, color codes, or virtual arrows. S312: Display an optimized virtual inspection path in the field of view to guide the quality inspector's gaze to systematically cover the entire area to be inspected, ensuring no omissions; S321: The eye tracker built into the augmented reality glasses tracks the eye movement trajectory and fixation time of the quality inspector in real time. The system automatically records the attention duration of the defective area indicated by the machine and compares it with the non-indication area. S322: When a quality inspector makes a judgment on a defect, the system records the reaction time from observation to decision-making and uses this data as a quantitative indicator of the judgment complexity. S323: Quality inspectors wear lightweight EEG sensors, and the system monitors the EEG signals of their prefrontal cortex, assesses their focus and cognitive load during the testing process, and issues a reminder when their focus drops below a threshold to ensure the reliability of the judgment. S331: Quality inspectors can operate the virtual defect menu superimposed on the physical object through predefined gestures (such as air selection and swiping) to complete the input of instructions such as "confirm", "reject", and "classify" without operating the physical terminal; S332: For complex or novel defects, quality inspectors can describe them by voice input. The system will automatically convert the voice into text and bind and store it with the defect image and machine data in the current field of view to build a searchable defect knowledge graph. S341: Link the entire process data of this manual inspection—including eye movement trajectory, gaze heat map, decision reaction time, final judgment result, and voice annotation—with the production batch of the part and the original machine vision data to form a complete and traceable digital inspection thread; S342: Perform bidirectional optimization of the machine vision model: a) Forward optimization: The defect data that has been finally confirmed by humans is used as an incremental labeled dataset to periodically retrain the deep learning network in step S214 to improve its automatic recognition ability. b) Backward calibration: Analyze the rejection data of quality inspectors for machine false alarm areas to dynamically relax the detection threshold in specific scenarios and reduce the false alarm rate; at the same time, analyze the defect data that the machine missed but was discovered by humans to dynamically tighten the detection sensitivity of relevant features. c) Human Factors Reliability Assessment: By integrating eye movement and reaction time data, a reliability model is established for each quality inspector. For high-difficulty defects, priority is given to quality inspectors with high reliability scores, thereby optimizing the allocation of human resources.

[0012] As a further aspect of the present invention: after completing the second injection and forming the molded part B, the following steps are performed: S600: Perform interfacial quality optical screening on the molded part B, the screening including at least the staining detection of the interface between the first and second injection materials; S610: Using an industrial camera equipped with a multi-band light source, an image of the interface region of the molded part B is acquired under a spectrum including visible light and a specific near-infrared band; wherein the specific near-infrared band is configured to have high contrast sensitivity to the minute color diffusion of the first and second emission materials. S621: Extract the spectral features of the pure first emission material region and the pure second emission material region from the multispectral image and use them as reference spectra; S622: Employ a spectral unmixing algorithm to analyze each pixel on the interface and calculate the mixing ratio of the first emission material reference spectrum and the second emission material reference spectrum in its spectrum; S623: Based on the mixing ratio, generate a color diffusion distribution map, which quantitatively characterizes the degree of diffusion and spatial distribution of the first color in the second color in the second color. S631: Compare the staining distribution map with the pre-stored staining-mechanical property correlation model; the correlation model is established based on historical data and describes the influence of different staining degrees on key mechanical properties such as the overall structural strength and bonding force of the part. S632: If the current bleeding state is predicted by the model to have no significant impact on the mechanical properties of the part, it is judged as qualified; if the prediction is that it will cause the mechanical properties to drop below the threshold, or the bleeding is located in the critical stress area, it is judged as unqualified. S640: The bleeding detection results are correlated with the machine vision inspection results after the first injection. If a statistical correlation is found between the bleeding defect and the specific surface flatness or pattern deviation of the first injection part, a process adjustment suggestion is generated and fed back to the injection parameter control system of the first or second injection.

[0013] As a further aspect of the present invention: after completing the second injection and forming the molded part B, the following steps are performed: S700: Perform optical screening on the degree of interfacial adhesion of the molded part B; S710: A pulsed laser is used to locally scan and excite the interface area of ​​the molded part B. The laser energy is absorbed by the material to generate high-frequency ultrasonic waves. At the same time, a laser interferometric vibrometer is used in conjunction with the laser interferometric vibration measurement principle to non-contactly and in the whole field measure the minute vibration displacement field induced by the ultrasonic waves on the surface of the part. S721: Extract the local vibration mode parameters of the interface region from the full-field vibration data obtained in step S710, including the resonance frequency, mode shape and damping ratio; S722: Input the extracted measured modal parameters into a pre-trained inversion model; the inversion model is established by combining finite element analysis and machine learning, and can accurately invert the effective adhesive stiffness of the interface layer and the size and location of the unbonded area (debonding) based on the vibration characteristics. S731: Input the adhesive stiffness and debonding information obtained from the inversion into a micromechanical model to predict the peel strength and shear strength of the interface under actual working conditions. S732: Compare the predicted bond strength with the minimum strength threshold required by the product design to make a functional judgment: if the predicted strength is higher than the threshold, it is judged as qualified; if it is lower than the threshold, or a continuous critical debonding area is detected, it is judged as unqualified. S740: The system performs multivariate correlation analysis on the adhesion degree detection results, the surface flatness data after the first injection, and the injection parameters of the second injection to establish a mapping relationship between adhesion quality and process parameters. If a poor adhesion trend is detected, the system automatically generates a set of optimized parameters for the second injection speed, temperature, or pressure, and feeds them forward to the next production cycle to achieve adaptive adjustment of the process.

[0014] As a further aspect of the present invention: after the molded part B is demolded, the following steps are performed: S800: Perform overall appearance defect optical screening on the molded part B; S810: In the integrated detection unit, the surface information of the molded part B is collected synchronously in the following manner: S811: Uses an integrating sphere or surrounding LED light source to acquire images under uniform diffused light conditions for detecting color uniformity, blemishes, and macroscopic defects; S812: Uses a low-angle linear light source to illuminate the surface in a specific direction to highlight micro-texture defects such as scratches, knife lines, and orange peel texture; S813: Using orthogonal polarizers placed in front of the light source and camera respectively, the specular reflection component and diffuse reflection component in the surface reflected light are separated and collected to eliminate reflection interference and detect defects such as subsurface stress lines and internal stress whitening. S821: Register and fuse the multimodal images obtained in S810 to construct a multi-channel feature image cube; S822: The U-Net deep learning model based on the attention mechanism is used to perform pixel-level segmentation on the feature image cube, automatically identify and label all potential defect areas, and pre-classify them according to their feature vectors. S831: Input the segmented defect areas and their features into a defect semantic network. This network comprehensively considers the type, size, quantity, contrast, location of defects, and degree of conflict with the product design aesthetic language. S832: The semantic network outputs a non-binary quality compliance vector, which contains scores of multiple dimensions, providing a quantitative basis for the final disposal decision; S841: Based on the quality compliance vector of S830, dynamically sort part B to the qualified area, special release area, rework area, and scrap area. S842: Spatiotemporally correlates all detection data with the real-time process parameter stream of the injection molding machine. Through big data analysis, it uncovers potential causal chains between specific appearance defects and process parameters such as mold status, injection curve, and temperature control, and generates maintenance and optimization warnings.

[0015] An injection mold for multi-color injection molding, wherein the injection mold adopts the above-described production process.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates a machine vision online inspection system for multi-color injection molding production, which includes pattern detection and surface flatness detection. The system uses optical sieve detection to perform pattern detection on specific parts of the molded part A, comparing the horizontal and vertical coordinates of the actual contour with a preset range to identify geometric deviations such as offset or distortion of the dividing lines. Subsequently, the surface flatness is objectively evaluated by analyzing the distribution of reflected light on the part's surface, determining whether it provides a qualified substrate for the second injection. This allows for the determination of key quality attributes of the intermediate product A before the second injection, effectively preventing the waste of raw materials, energy, and labor caused by continuing a second injection on a known defective product. Simultaneously, it ensures the positional accuracy and visual clarity of the final multi-color product at the junctions of different colors, avoiding defects such as distorted dividing lines and color bleeding, significantly improving production yield and resource utilization. Attached Figure Description

[0017] Figure 1 This is a flowchart of S100-S300 in this invention; Figure 2 This is a schematic diagram of the horizontal and vertical coordinates in step S210 of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-2 A production process for multi-color injection molding includes the following steps: S100: Perform the first injection to form molded part A; S200: Perform machine vision inspection on the molded part A, including a light sieving inspection process, wherein the light sieving inspection process includes: The multi-color injection molding production process in this invention achieves a shift from "post-injection inspection" to "in-process control" by introducing an integrated machine vision online inspection system after the first injection. As described in the published patent CN220075362U, "An Injection Molding Device for a Multi-color Injection Mold", during the multi-color injection molding process, after the first injection molded part (part A) cools and solidifies in the cavity, the mold for the first injection will open and then rotate, so that part A can be rotated from the previous injection station to the next injection station. During the process of part A moving from the previous injection station to the next injection station, part A is exposed, so an inspection station can be added between the previous injection station and the next injection station. During this process, we can systematically perform optical screening. First, such as Figure 2 As shown, this invention uses "optical screening" to capture and compare the horizontal and vertical coordinates of specific parts of part A, which can capture minute offsets, distortions, or deformations of the dividing lines, ensuring the accuracy of the geometry of the base for the second injection. At the same time, by analyzing the distribution of reflected light to evaluate the surface flatness, it can determine whether the surface of part A can provide an ideal substrate for the second injection material, thereby preventing appearance defects such as uneven bonding lines, color bleeding, steps, or overflow caused by uneven substrates, ensuring that the final product has clear, sharp, and design-intention-compliant dividing lines at the joints of different colors. S210: Perform style detection on a specific part of the molded part A, including horizontal coordinate comparison and vertical coordinate comparison. The horizontal coordinate comparison is performed by comparing the actual horizontal coordinate of the specific part with a preset horizontal coordinate range to detect style deviation. The vertical coordinate comparison is performed by comparing the actual vertical coordinate of the specific part with a preset vertical coordinate range to detect style deviation. S220: Perform surface flatness testing on the molded part A, and evaluate whether its surface flatness meets the requirements of the second injection by analyzing the reflected light distribution on the surface of the molded part A. S300: Based on the results of the machine vision inspection, if the pattern matching rate or surface flatness is detected to be lower than the threshold, the subsequent injection molding process is stopped; otherwise, a second injection is performed. Secondly, in terms of production efficiency and cost control, this intermediate inspection step can immediately identify defective products that have irreparable defects even after the first injection and stop the subsequent process immediately. This avoids wasting raw materials for the second injection, energy consumption of the injection molding machine, and operating hours on defective products, thus achieving cost savings. At the same time, it reduces the workload of conducting full inspection of complex finished products, and reduces losses such as customer complaints and returns caused by defective products. This invention integrates a machine vision online inspection system for multi-color injection molding production, which includes pattern detection and surface flatness detection. The system uses optical sieve detection to perform pattern detection on specific parts of the molded part A, comparing the horizontal and vertical coordinates of the actual contour with a preset range to identify geometric deviations such as offset or distortion of the dividing lines. Subsequently, the surface flatness is objectively evaluated by analyzing the distribution of reflected light on the part's surface, determining whether it provides a qualified substrate for the second injection. This allows for the determination of key quality attributes of the intermediate product A before the second injection, effectively preventing the waste of raw materials, energy, and labor caused by continuing a second injection on a known defective product. Simultaneously, it ensures the positional accuracy and visual clarity of the final multi-color product at the junctions of different colors, avoiding defects such as distorted dividing lines and color bleeding, significantly improving production yield and resource utilization.

[0020] In this embodiment of the invention, step S100 includes the following steps: S110: Dry and heat the thermoplastic material required for the first shot to a molten state; S120: Close the mold to form the cavity of the molded part A and control the mold temperature; S130: Inject molten material into the cavity, and hold pressure after filling; S140: After the molded part A has cooled and solidified in the cavity, the mold is opened; First, the thermoplastic material for the first injection is dried to remove moisture and heated to a molten state to ensure material flowability and molding quality. Then, the mold is closed to form a cavity, and the mold temperature is actively controlled to provide a stable molding environment for the molten material. Next, the material is injected into the cavity, and pressure is maintained after filling to compensate for shrinkage and enhance material density. Finally, the part is allowed to cool and solidify fully in the cavity before the mold is opened. By eliminating moisture in the raw material, stabilizing molding conditions, compensating for volume shrinkage, and ensuring full solidification, the geometric dimensions and stable molecular structure of the molded part A are guaranteed, with uniform internal stress distribution and qualified surface quality, thus providing a stable matrix for the subsequent second injection.

[0021] In this embodiment of the invention, step S210 includes the following steps: S211: Before injection, based on the real-time process parameters of the first injection, drive the digital twin model of the molded part A to perform synchronous thermodynamic-structural mechanics simulation, predict and generate a dynamic three-dimensional reference model containing material shrinkage and cooling deformation under this specific production cycle in real time. S212: A scanning device incorporating white light interference is used to perform a three-dimensional scan on the specific area to obtain a two-dimensional image of the surface of the molded part A; A scanning device combining spectral domain optical coherence tomography is used to perform three-dimensional scanning on the specific part, penetrating the surface layer of the transparent or semi-transparent first-electrode material to obtain the internal structure profile image within a depth range of 0.1-0.5 mm below the subsurface, thereby simultaneously capturing surface contour and subsurface defect information. S213: Fuse the surface 3D point cloud data obtained in step S212 with the subsurface profile image data to construct an extended 3D dataset containing surface and subsurface information; The extended 3D dataset is aligned with the dynamic 3D benchmark model generated in step S211 in four-dimensional spatiotemporal coordinates and time dimension based on injection timing. S214: Input the aligned extended 3D dataset into a pre-trained deep learning convolutional neural network; The network analyzes surface geometric features, subsurface structural features and their spatial correlations, and directly outputs a comprehensive confidence score regarding whether the style of the specific part is qualified or not. The horizontal coordinate comparison and vertical coordinate comparison are embedded as primary features in the underlying perception module of the network. The network's decision goes beyond simple coordinates and is based on a deep understanding of the overall shape and the integrity of the internal structure. S215: The comprehensive confidence score obtained in step S214 and the identified defect feature types are fed back to the injection molding machine control system in real time; if the pattern is qualified, the second injection preparation is triggered. If there is a compensable pattern deviation, a feedforward control command is generated to automatically fine-tune the injection parameters of the second shot to actively compensate for the minor deformations present in the first shot during the second shot injection; at the same time, all data from this test are recorded to the digital twin model to optimize the accuracy of subsequent simulation predictions. The spectral domain optical coherence tomography technique can non-destructively detect microstructures (such as internal stress cracks, microbubbles, uneven filler distribution, etc.) 0.1-0.5 mm below the subsurface of transparent or semi-transparent materials. These internal defects may not be visible on the surface at present, but under the thermal and mechanical action of the second injection, they are prone to develop into faults such as insufficient bonding strength, surface collapse, or stress concentration. The system achieves source quality control by discovering and intercepting these hidden faults in advance. The working principle of the spectral domain optical coherence tomography technique can be found in the CNKI paper "[1] Wang Changming, Gao Wanrong. Measurement of glass subsurface defect scattering coefficient based on micron SDOCT [J]. Acta Optica Sinica, 2021, 41(07): 182-187.". Unlike using a fixed, ideal CAD model as a benchmark, this solution generates a dynamic three-dimensional benchmark model for each production cycle using digital twin technology. This model reflects the expected shrinkage and deformation under the current specific process parameters (such as material temperature, pressure, and cooling rate) in real time, ensuring that the detection standard is consistent with the actual behavior in the physical world. This time-varying comparison benchmark eliminates misjudgments caused by process fluctuations, making the detection results more in line with actual production. The operating principle of the digital twin technology and dynamic three-dimensional benchmark model can be found in the published patent "3D Printer Modeling Method and Model System Based on Digital Twin Three-Dimensional Model" with publication number CN111046597B. When a "compensable style deviation" is detected, the system does not simply judge it as a defective product. Instead, it generates a feedforward control command to automatically fine-tune the injection parameters of the second injection (such as injection speed, pressure, and starting position) to actively "cater to" or "counteract" the deficiencies of the first injection. For example, if a slight warping of part A is detected, the system can instruct the moving side of the second injection mold to fine-tune the injection parameters to ensure that the material of the second injection can fill the gap caused by the warping. On the basis of meeting the production requirements, it can make up for the defects and save some of the "defective products" that should have been scrapped as "qualified products", thereby improving the yield rate. By entrusting the complex multi-source data fusion and decision-making tasks to a pre-trained deep learning network, millisecond-level online automatic decision-making and process adjustment are achieved, avoiding production interruptions caused by manual intervention, offline measurement, and repeated debugging required in related technologies, and ensuring that the production line can operate continuously and smoothly. Each detection's data (including process parameters, simulation predictions, actual scanning results, and network decisions) is recorded and fed back to the digital twin model. This process is equivalent to establishing a continuously expanding "experience database" for the system. Using this data, the parameters of the thermodynamic-structural mechanics simulation model can be continuously calibrated and optimized, making its next prediction more accurate. At the same time, deep learning convolutional neural networks can learn the complex, nonlinear relationship between surface geometric features and subsurface structural features. This means that its decision goes beyond simple coordinate range comparison and can identify potential defects with "strange" shapes even if the coordinates are within the range, based on a deep understanding of the overall morphology of "healthy" and "pathological". This invention constructs a closed-loop "detection-prediction-compensation" system, integrating digital twins, multimodal 3D scanning, and deep learning. Before injection, real-time process parameters drive the digital twin model for synchronous simulation, generating a dynamic 3D benchmark model that predicts material shrinkage and cooling deformation within the current cycle. Subsequently, a scanning device combining white light interferometry and spectral domain optical coherence tomography (OCT) is used to simultaneously acquire high-precision 3D morphology of the surface of the molded part A and internal structural profiles within a specific depth range below the subsurface. This surface and subsurface information is then fused into an extended 3D dataset. Next, this dataset is precisely aligned with the dynamic benchmark model in four-dimensional spacetime and input into a pre-trained deep learning network for comprehensive analysis. This network goes beyond simple coordinate comparison, outputting a comprehensive confidence score and identifying defect features based on a deep understanding of the overall morphology and internal structural integrity. Finally, the system executes decisions based on the judgment results: if qualified, the next process is triggered; if there are compensable deviations, feedforward control commands are automatically generated to fine-tune the second injection parameters to actively compensate for the deformation of the first injection, while all data is fed back to the digital twin model to continuously optimize simulation accuracy.

[0022] In this embodiment of the invention, step S220 includes the following steps: S221: After the molded part A completes its first injection and enters the inspection station, the following operations are performed simultaneously: S221a: Structured light in the short-wave infrared band is projected onto the surface to be inspected, and its deformation image is acquired; S221b: Using a high-resolution thermal imaging camera, the temperature field distribution map of the surface of the part is acquired to obtain its real-time thermodynamic state during the cooling process; S222: The short-wave infrared structured light image acquired in S221a is used to reconstruct a high-sensitivity three-dimensional micro-topography point cloud through phase calculation; the topography point cloud is then fused with the surface temperature field data acquired in S221b at the pixel level to generate a multi-physics point cloud model containing the coordinates (X, Y, Z) of each spatial point, the temperature (T), and the local shrinkage stress (σ) of the material derived from the temperature. S223: Input the multiphysics point cloud model into a second injection molding process simulator. The simulator uses the precise morphology, temperature and stress state of the current first injection surface as initial conditions to simulate the entire process of the second injection melt flowing, filling, cooling and bonding with the first injection, and outputs the predicted bonding line strength, coverage integrity and potential weld line location and other key quality indicators. S224: Make dynamic decisions based on the simulation prediction results of S223: If the predicted quality of the second shot fully meets the design requirements, then even if there is a slight deviation in geometric flatness, it is judged as qualified. If the prediction results show that the bonding quality is at a critical state, the system will automatically generate a personalized flatness tolerance for the specific part after tightening, and make a second fine judgment. If the prediction results show that the bonding quality is unqualified, it will be directly judged as unqualified; S225: For parts judged as qualified but predicted to have suboptimal quality, the system reverse-engineers the simulation results to generate a set of optimized second injection parameters, including but not limited to injection speed, melt temperature and holding pressure. These parameters are then fed forward to the injection molding machine control system for the upcoming second injection, in order to proactively improve the quality of the final product. At the same time, all data from this test and the simulation prediction results are compared with the actual product quality after the second injection for continuous training and optimization of the simulator's prediction accuracy. In related technologies, optical flatness detection can only determine whether a surface "looks flat." However, this system, by integrating short-wave infrared structured light (which is highly sensitive to nanoscale undulations) and thermal imaging (which reflects real-time thermodynamic states), can construct a multiphysics model that includes local stress (σ). The system can identify a surface that is "geometrically flat" but has "highly uneven internal stress." Such surfaces are prone to warping, streaks, or even cracking under the thermal shock and pressure of the second-shot melt, resulting in insufficient bonding strength. This system can predict this failure risk based on material state in advance. The operating principle of the multiphysics point cloud model is described in the published patent "A Parametric Simulation Model Design Method and System for Precision Molds" with announcement number CN118430720B. The multiphysics point cloud model no longer mechanically applies a fixed flatness tolerance (such as ±0.01mm). Instead, it dynamically evaluates the actual impact of the current surface condition on the final bonding quality (strength, coverage) through a second injection molding process simulator. For example, a part with a small depression in a non-critical area will be judged as qualified if the simulation predicts that the second injection melt can fill it and form a high-strength bond. This "function-based tolerance" avoids misjudging functionally qualified parts as scrap, thereby improving the yield rate. For "critical parts" whose simulation prediction results are on the edge of the pass line, the system does not arbitrarily accept or reject them, but generates a "personalized, tightened flatness tolerance" for secondary judgment. This is equivalent to launching a stricter, tailor-made inspection protocol for high-risk areas. This refined management reduces unnecessary waste. For parts that are deemed qualified but whose "bonding quality is not optimal" according to simulation predictions, the system can reverse-engineer and generate a set of optimized second injection parameters. For example, if the simulation shows that the current surface temperature of the first injection is too low and may cause weld lines, the system will automatically increase the melt temperature or injection speed of the second injection to compensate. This is equivalent to equipping each semi-finished product with a dedicated "process prescription" to ensure that it can ultimately achieve better quality. Through the aforementioned intelligent tolerance and feedforward compensation, the system reduces the scrap rate caused by minor, non-functional defects in the first shot; the system compares all detection data, simulation predictions, and actual quality results after the second shot, and uses this information to continuously train and optimize the simulator. In related technologies, the production process relies on the "feel" and "experience" of experienced workers to adjust the process to deal with subtle situations. This system transforms this difficult-to-express "tacit knowledge" into a quantifiable, calculable, and executable "explicit algorithm" through multi-physics sensing and physical simulation. This invention constructs an intelligent decision-making system based on real-time multiphysics perception and forward-looking process simulation. After the molded part A enters the inspection station, short-wave infrared structured light and high-resolution thermal imaging are used simultaneously to capture the three-dimensional micro-morphology, which is extremely sensitive to minute fluctuations, and the temperature field, which reflects the real-time thermodynamic state of the part, respectively. Then, the morphology, temperature, and the derived local shrinkage stress data are fused at the pixel level to generate a multiphysics point cloud model. Using this model as the initial condition, the second injection molding process simulator is driven to forward-lookingly simulate and output key quality indicators such as bond line strength and encapsulation integrity. Finally, the system makes dynamic judgments based on this high-fidelity simulation results—not only can it make a "qualified / unqualified" judgment, but it can also initiate a secondary fine judgment based on personalized tolerances for parts in a critical state, and generate a set of optimized second injection parameters for all qualified but not optimal parts, feeding them forward to the injection molding machine to actively improve the final quality, while feeding all data back in a closed loop to continuously optimize the simulator accuracy.

[0023] In this embodiment of the invention, in step S200, if the result of the machine vision inspection is in a critical state or the confidence level is below a threshold, the augmented reality-assisted human visual inspection step S300 is simultaneously initiated, wherein S300 includes: S310: When the quality inspector wears augmented reality glasses and looks at the molded part A, the glasses perform the following steps: S311: Track and locate parts in real time through image recognition, and overlay the machine vision results from steps S210 and S220 onto the corresponding positions of the physical object in the form of highlighted outlines, color codes, or virtual arrows. S312: Display an optimized virtual inspection path in the field of view to guide the quality inspector's gaze to systematically cover the entire area to be inspected, ensuring no omissions; S321: The eye tracker built into the augmented reality glasses tracks the eye movement trajectory and fixation time of the quality inspector in real time. The system automatically records the attention duration of the defective area indicated by the machine and compares it with the non-indication area. S322: When a quality inspector makes a judgment on a defect, the system records the reaction time from observation to decision-making and uses this data as a quantitative indicator of the judgment complexity. S323: Quality inspectors wear lightweight EEG sensors, and the system monitors the EEG signals of their prefrontal cortex, assesses their focus and cognitive load during the testing process, and issues a reminder when their focus drops below a threshold to ensure the reliability of the judgment. S331: Quality inspectors can operate the virtual defect menu superimposed on the physical object through predefined gestures (such as air selection and swiping) to complete the input of instructions such as "confirm", "reject", and "classify" without operating the physical terminal; S332: For complex or novel defects, quality inspectors can describe them by voice input. The system will automatically convert the voice into text and bind and store it with the defect image and machine data in the current field of view to build a searchable defect knowledge graph. S341: Link the entire process data of this manual inspection—including eye movement trajectory, gaze heat map, decision reaction time, final judgment result, and voice annotation—with the production batch of the part and the original machine vision data to form a complete and traceable digital inspection thread; S342: Perform bidirectional optimization of the machine vision model: a) Forward optimization: The defect data that has been finally confirmed by humans is used as an incremental labeled dataset to periodically retrain the deep learning network in step S214 to improve its automatic recognition ability. b) Backward calibration: Analyze the rejection data of quality inspectors for machine false alarm areas to dynamically relax the detection threshold in specific scenarios and reduce the false alarm rate; at the same time, analyze the defect data that the machine missed but was discovered by humans to dynamically tighten the detection sensitivity of relevant features. c) Human factors reliability assessment: Based on eye movement and reaction time data, a reliability model is established for each quality inspector. For high-difficulty defects, priority is given to quality inspectors with high reliability scores to achieve optimal allocation of human resources. In related technologies, manual quality inspection relies on the experience and condition of inspectors, has a long training cycle, and has large performance fluctuations. This system, through AR overlay, directly "sees through" the defect location, risk level, and even microscopic morphology heat map that is invisible to the naked eye, predicted by machine vision and simulation, onto the physical object, and guides the eye with a virtual inspection path. This is equivalent to equipping each quality inspector with a guide, which greatly reduces the skill threshold and training cost. In related technologies, quality inspection only manages the "results" (pass / fail), while this system, through eye trackers, reaction time recording, and EEG monitoring, achieves quality control over the "inspection process" itself for the first time. Eye track analysis ensures that inspectors do indeed follow the system's guidance and give sufficient attention to key risk areas, preventing missed inspections due to distraction or skipping inspections. EEG focus monitoring can detect the decline in inspector attention due to fatigue in real time and promptly remind them before making unreliable judgments, ensuring from a physiological perspective that every judgment is made in the best cognitive state. This elevates the reliability of on-duty quality inspection from "moral self-discipline" to the level of "scientific assurance." Through gesture and voice recognition, all judgments (confirmation, rejection, classification) and descriptions (voice recording) by inspectors can be completed while looking at the parts. The system records not only the final result, but also the closed-loop data of the entire inspection process, including: what the part looked like at the time (machine data) - what the system prompted (AR overlay) - where the inspector looked and for how long (eye movement) - whether he was focused when making the judgment (EEG) - what decision he ultimately made (gestures / voice). This provides complete data traceability capabilities for quality disputes or defect analysis. Forward optimization (machine learning from humans): Defect data ultimately confirmed by humans, especially complex and novel defects that the machine initially "hesitated" (critical state) or completely missed, are fed back to the deep learning model as "annotated data." This allows the machine vision system to continuously learn from inspectors' judgments, becoming increasingly intelligent and with stronger recognition capabilities. Backward calibration (human learning from machines / machine self-calibration): By analyzing inspectors' rejection data of areas where the machine "falsely reports" the system, the system can dynamically relax the detection threshold in certain specific, harmless scenarios, thereby effectively reducing the false alarm rate and minimizing unnecessary interference with skilled workers. At the same time, it analyzes defects that the machine misses but are discovered by humans, using this data to dynamically tighten the detection sensitivity of relevant features and improve the detection rate. Through long-term collection of eye movement, reaction time, and EEG data, the system can build personalized reliability models for each quality inspector. When extremely difficult judgment tasks arise, the system can prioritize assigning them to "high-quality quality inspectors" who have historically demonstrated higher focus and more stable judgment, thereby achieving refined management of the company's human resources. When machine vision inspection results are critical or lack sufficient confidence, the system automatically initiates an augmented reality (AR) assistance process. The AR glasses worn by the quality inspector first precisely locate the part through image recognition, and then overlay the machine vision and simulation analysis results (such as defect outlines and risk points) onto the corresponding physical location in real time using highlighted outlines and color coding. Simultaneously, the system displays the optimal virtual inspection path to guide the eye to systematically cover the area to be inspected. Building upon this, the system integrates biosignal monitoring, using an eye tracker to track gaze trajectory and dwell time to quantify attention, and an EEG sensor to assess focus and cognitive load, issuing alerts when focus decreases, thus ensuring the reliability of manual judgment. Quality inspectors interact with virtual information through predefined gestures or voice to complete decision input and knowledge accumulation. Ultimately, the system associates all process data (eye movement trajectory, reaction time, judgment results, etc.) with production data to form a complete digital inspection thread, and uses this data to perform bidirectional adaptive optimization of the machine vision model. On the one hand, the defect data confirmed by humans is used as an incremental dataset to train the deep learning network to improve the recognition ability (forward optimization). On the other hand, the detection threshold and sensitivity are dynamically calibrated based on the rejection of false alarms and the discovery of missed detections by humans (backward calibration). At the same time, a quality inspector reliability model is established by integrating human factors data to achieve optimized allocation of human resources.

[0024] In this embodiment of the invention, after the second injection is completed and the molded part B is formed, the following steps are performed: S600: Perform interfacial quality optical screening on the molded part B, the screening including at least the staining detection of the interface between the first and second injection materials; S610: Using an industrial camera equipped with a multi-band light source, an image of the interface region of the molded part B is acquired under a spectrum including visible light and a specific near-infrared band; wherein the specific near-infrared band is configured to have high contrast sensitivity to the minute color diffusion of the first and second emission materials. S621: Extract the spectral features of the pure first emission material region and the pure second emission material region from the multispectral image and use them as reference spectra; S622: Employ a spectral unmixing algorithm to analyze each pixel on the interface and calculate the mixing ratio of the first emission material reference spectrum and the second emission material reference spectrum in its spectrum; S623: Based on the mixing ratio, generate a color diffusion distribution map, which quantitatively characterizes the degree of diffusion and spatial distribution of the first color in the second color in the second color. S631: Compare the staining distribution map with the pre-stored staining-mechanical property correlation model; the correlation model is established based on historical data and describes the influence of different staining degrees on key mechanical properties such as the overall structural strength and bonding force of the part. S632: If the current bleeding state is predicted by the model to have no significant impact on the mechanical properties of the part, it is judged as qualified; if the prediction is that it will cause the mechanical properties to drop below the threshold, or the bleeding is located in the critical stress area, it is judged as unqualified. S640: The bleeding detection results are correlated with the machine vision inspection results after the first injection. If a statistical correlation is found between the bleeding defect and the specific surface flatness or pattern deviation of the first injection part, a process adjustment suggestion is generated and fed back to the injection parameter control system of the first or second injection. In related technologies, detection relies on the human eye judging whether color "bleeds" under standard light sources. This is affected by subjective factors, lighting conditions, and background color. This system, through multi-band light sources (especially specific near-infrared bands sensitive to dye molecule diffusion) and spectral demixing algorithms, can penetrate surface visual interference and directly analyze the proportion of material chemical composition of each pixel, generating a "bleeding distribution map" that uses specific values ​​to represent the degree of bleeding. This transforms the vague description of "some bleeding" into precise quantitative data of "x% of the first-emission material is mixed in at a certain position." The operating principle of the multi-band light source is described in the published patent "An Adjustable Single-Shot Multi-Band Light Source System" with publication number CN117191341A; the operating principle of the spectral demixing algorithm is described in the published patent "A Hyperspectral Demixing Algorithm Based on Denoising Three-Dimensional Convolutional Autoencoder Network" with publication number CN111260576B. The "bleeding-mechanical property correlation model" introduced by the system is based on a large amount of historical experimental data (such as tensile and shear tests). It can predict how bleeding at a specific degree and location will affect the key functional attributes of the product, such as structural strength, bonding force, and impact resistance. This means that a part that is "imperfect" in appearance can be scientifically judged as a qualified product as long as the model predicts that its mechanical properties fully meet the design requirements. For some products where functional requirements are higher than appearance requirements, this avoids unnecessary scrapping of functional products due to overly conservative appearance standards, thus improving the yield rate. Some slight staining that is difficult to detect with the naked eye but is located in critical stress areas may be "weak points" in the bonding interface strength. The detection technology in related technologies may miss these "hidden defects", leading to early failure of the product during user use. This system can identify and intercept these parts with potential performance risks. By correlating the final bleeding detection results with the machine vision inspection results after the first injection (such as surface flatness and pattern deviation), the system can perform big data mining to discover causal patterns such as "when the first injection part has a small dent in a certain area, the probability of severe bleeding in that area after the second injection increases significantly." Once such statistical correlation is found, the system can automatically generate specific process adjustment suggestions (such as "fine-tuning the holding pressure of the first injection in area Z" or "increasing the melt temperature of the second injection") and directly feed them back to the injection molding parameter control system. This is equivalent to equipping the entire production line with a closed-loop control link, which can deduce the root cause of the process parameters from the result of quality defects and provide solutions, thus realizing feedforward control of the production process. By performing optical screening of the interface quality of molded part B after the second injection, an intelligent color bleeding evaluation system based on spectral analysis and performance prediction was constructed. First, images of the interface area were acquired using multi-band light sources (including specific near-infrared bands) sensitive to color diffusion. Using a spectral unmixing algorithm, the mixing ratio was calculated pixel by pixel based on the spectra of the pure first and second injection materials, thereby generating a color bleeding distribution map that quantitatively characterizes the degree and range of color bleeding. Subsequently, the system compared this distribution map with a pre-stored color bleeding-mechanical property correlation model, scientifically predicted the actual impact of the current color bleeding state on the structural strength and bonding force of the part based on historical data, and made a pass / fail judgment based on mechanical properties. Only when the prediction has no significant impact on performance was the part deemed passable. Finally, the system performed correlation analysis between the color bleeding results and the detection data after the first injection. If a statistical correlation was found between color bleeding and specific surface defects of the first injection, process adjustment suggestions were automatically generated and fed back to the injection molding parameter control system, constructing a causal chain from the final defect back to the root cause of the preceding process, forming a closed loop.

[0025] In this embodiment of the invention, after the second injection is completed and the molded part B is formed, the following steps are performed: S700: Perform optical screening on the degree of interfacial adhesion of the molded part B; S710: A pulsed laser is used to locally scan and excite the interface area of ​​the molded part B. The laser energy is absorbed by the material to generate high-frequency ultrasonic waves. At the same time, a laser interferometric vibrometer is used in conjunction with the laser interferometric vibration measurement principle to non-contactly and in the whole field measure the minute vibration displacement field induced by the ultrasonic waves on the surface of the part. S721: Extract the local vibration mode parameters of the interface region from the full-field vibration data obtained in step S710, including the resonance frequency, mode shape and damping ratio; S722: Input the extracted measured modal parameters into a pre-trained inversion model; the inversion model is established by combining finite element analysis and machine learning, and can accurately invert the effective adhesive stiffness of the interface layer and the size and location of the unbonded area (debonding) based on the vibration characteristics. S731: Input the adhesive stiffness and debonding information obtained from the inversion into a micromechanical model to predict the peel strength and shear strength of the interface under actual working conditions. S732: Compare the predicted bond strength with the minimum strength threshold required by the product design to make a functional judgment: if the predicted strength is higher than the threshold, it is judged as qualified; if it is lower than the threshold, or a continuous critical debonding area is detected, it is judged as unqualified. S740: Perform multivariate correlation analysis on the adhesion degree detection results, surface smoothness data after the first injection, and injection parameters of the second injection to establish a mapping relationship between adhesion quality and process parameters; if a poor adhesion trend is detected, the system automatically generates a set of optimized parameters for the second injection speed, temperature, or pressure, and feeds them forward to the next production cycle to achieve adaptive adjustment of the process; Destructive sampling in related technologies (such as tensile testing after cutting) not only wastes products but also fails to achieve full inspection. Other non-destructive methods (such as ultrasonic testing) usually require coupling agents and are difficult to apply to complex shapes. This system uses pulsed laser excitation and laser interferometric vibration measurement to achieve non-contact "full-field" detection. It can excite and capture ultrasonic waves propagating at the interface, just like performing an "ultrasound" on the parts. Debonding and loose bonding areas inside the product will significantly change the wave propagation characteristics (such as resonant frequency and damping), thus being captured by the system. The working principle of the laser interferometric vibration measurement principle is described in the published patent "A Real-time Online Monitoring Device and Method for Laser Peeling Cracks of Semiconductor Materials Based on Laser Ultrasound" with publication number CN120102464A. The system does not merely "discover" a defect, but rather uses a pre-trained inversion model to accurately invert the "effective adhesive stiffness" of the interface and the "size and location" of the debonding region based on vibration mode parameters. This provides quantitative data: for example, we not only know "this part is not bonded well," but also "the adhesive stiffness here has decreased by 50%" and "the diameter of this debonding point is 2 mm." This quantification provides quantitative data for subsequent performance prediction and judgment. The operating principle of the inversion model is described in the published patent "An Inversion Method for Arbitrary Variable Density Interfaces" with announcement number CN116661014B. Through the micromechanical model, the detected adhesive stiffness and other parameters are directly converted into predicted values ​​of peel strength and shear strength, which are key to the product in actual operation. This makes the quality judgment no longer based on an indirect, empirical "adhesive area ratio" threshold, but on a direct, physically clear functional standard of "whether the predicted strength meets the design requirements". Many parts with localized poor adhesion may appear "good" under static visual inspection, but during user use, especially under impact, vibration, or cyclic loading, they can crack from these weak points, leading to premature failure. This system can identify and intercept these "hidden dangers," improving the structural reliability of manufactured products and avoiding costly after-sales claims and recall risks. Because it is non-contact and non-destructive, it can theoretically screen the interface strength of every part leaving the factory, replacing the sampling inspection mode in related technologies that requires destroying samples to obtain strength data. The system performs multivariate correlation analysis on the adhesion test results with the surface condition of the first injection and the injection parameters of the second injection, which is equivalent to performing a deep "problem retrospective." It can discover process problems such as "when the melt temperature of the second injection is too low, even if the surface of the first injection is flat, the adhesion stiffness will systematically decrease." By constructing a non-contact interface adhesion detection system based on laser ultrasound and full-field vibration measurement, after the molding of part B is completed, a pulsed laser is used to scan and excite the interface area to generate high-frequency ultrasonic waves. Combined with the principle of laser interferometry, the system non-contactly captures the minute vibration displacement field induced by the ultrasonic waves on the material surface. Subsequently, the system extracts the local vibration mode parameters (such as resonance frequency, mode shape, and damping ratio) of the interface from the full-field vibration data and inputs them into an inversion model based on finite element method and machine learning pre-training. This accurately inverts the effective adhesion stiffness of the interface layer and the specific size and location of the debonding area. Then, these inverted structural parameters are input into a micromechanical model to predict the peel strength and shear strength of the interface under actual working conditions. The system judges the functionality based on whether the predicted strength is higher than the product design threshold. Finally, the system performs multivariate correlation analysis on the adhesion degree detection results with the surface flatness data of the first injection and the injection parameters of the second injection. If a poor adhesion trend is found, a targeted set of optimized process parameters for the second injection is automatically generated and fed forward to the next production cycle.

[0026] In this embodiment of the invention, after the molded part B is demolded, the following steps are performed: S800: Perform overall appearance defect optical screening on the molded part B; S810: In the integrated detection unit, the surface information of the molded part B is collected synchronously in the following manner: S811: Uses an integrating sphere or surrounding LED light source to acquire images under uniform diffused light conditions for detecting color uniformity, blemishes, and macroscopic defects; S812: Uses a low-angle linear light source to illuminate the surface in a specific direction to highlight micro-texture defects such as scratches, knife lines, and orange peel texture; S813: Using orthogonal polarizers placed in front of the light source and camera respectively, the specular reflection component and diffuse reflection component in the surface reflected light are separated and collected to eliminate reflection interference and detect defects such as subsurface stress lines and internal stress whitening. S821: Register and fuse the multimodal images obtained in S810 to construct a multi-channel feature image cube; S822: The U-Net deep learning model based on the attention mechanism is used to perform pixel-level segmentation on the feature image cube, automatically identify and label all potential defect areas, and pre-classify them according to their feature vectors. S831: Input the segmented defect areas and their features into a defect semantic network. This network comprehensively considers the type, size, quantity, contrast, location of defects, and degree of conflict with the product design aesthetic language. S832: The semantic network outputs a non-binary quality compliance vector, which contains scores of multiple dimensions, providing a quantitative basis for the final disposal decision; S841: Based on the quality compliance vector of S830, dynamically sort part B to the qualified area, special release area, rework area, and scrap area. S842: Spatiotemporally correlate all detection data with the real-time process parameter stream of the injection molding machine. Through big data analysis, it can uncover the potential causal chain between specific appearance defects and process parameters such as mold status, injection curve, and temperature control, and generate maintenance and optimization warnings. In related technologies, visual inspection typically relies on a single light source, making it difficult to simultaneously capture defects of different types and contrasts. This system employs three optical paths working in tandem: uniformly diffused light provides a shadowless reference image to evaluate macroscopic defects such as color uniformity, blemishes, and shrinkage marks; a low-angle linear light source elongates and magnifies the shadows of minute texture defects such as scratches, knife lines, and orange peel texture through an extremely shallow incident angle, making them clearly visible in the image; and orthogonally polarized light filters out dazzling specular reflections from the surface to observe subsurface defects such as "stress whitening" and microcracks caused by internal stress, which are difficult for the human eye and conventional cameras to detect. The U-Net model based on the attention mechanism can perform pixel-level segmentation of the fused multi-channel feature image, much like an experienced quality inspector scans parts with a magnifying glass. It can not only find defects but also outline their contours and pre-classify them based on their optical features (e.g., this is a scratch, that is a stain), reducing false alarms (mistaking reflections for defects) and false negatives. It also provides geometric data (e.g., the precise area and shape of defects) for subsequent quantitative evaluation. The working principle of the U-Net deep learning model based on the attention mechanism is described in the published patent "A Dam Crack Detection Method Based on U-net Network and SC-SAM Attention Mechanism" with announcement number CN112232391B. By introducing "aesthetic semantic understanding" and simulating the perspectives of customers and designers, the defect semantic network introduces multi-dimensional and business-logical evaluation criteria. For example, it simulates the end user's tolerance for defects. A minor defect on a non-A-side (inconspicuous location) may not affect the use or overall aesthetics at all. It also assesses the degree to which defects damage the original design intent of the product. A defect that damages the clarity of the brand logo is much more serious than a similar defect on a smooth surface. Based on the type and location of the defect, it determines whether it is easy to repair. The working principle of the defect semantic network is described in the published patent "PCB Defect Image Generation Method Based on Semantic Editing and Generative Adversarial Network" with publication number CN119006267A. Based on the above multi-dimensional scoring, the system can perform dynamic sorting and set up a "special release zone". This means that parts that are fully functional, have only minor appearance defects and do not affect the customer experience can be released, avoiding unnecessary scrapping based on overly conservative standards. The system spatially and temporally correlates all appearance defect data with the real-time process parameter stream of the injection molding machine. Through big data analysis, it can uncover deep-seated patterns invisible to the naked eye, discovering that "when the mold temperature is below X℃, the probability of orange peel texture appearing on the side of the product increases significantly," and that "the injection speed curve fluctuates in the Y stage, which is strongly correlated with the appearance of surface flow marks." Through these analyses, the system can generate specific maintenance and optimization warnings, such as "it is recommended to clean the mold venting groove" or "optimize the second stage injection speed." This makes quality control no longer the end point of the production line, but another starting point for process optimization, forming a continuous improvement closed loop that drives process parameters and mold maintenance from the detection results. This invention constructs an appearance inspection system integrating multimodal optical imaging, deep learning, and semantic understanding. After part B is demolded, the integrated inspection unit first simultaneously uses three optical modes—uniform diffuse light, low-angle linear light, and orthogonally polarized light—to collect multi-dimensional surface information such as macroscopic defects, microscopic textures, and subsurface stress. These multimodal images are then registered and fused into a feature image cube, which is input into a U-Net network based on an attention mechanism for pixel-level segmentation and pre-classification to identify potential defects. The segmentation results are then input into a defect semantic network, which goes beyond simple binary judgment. By comprehensively evaluating the type, size, location, and degree of conflict with aesthetic language of defects, it outputs a quality compliance vector containing multi-dimensional scores such as "customer acceptability" and "deviation from design standards." Finally, the system dynamically sorts parts based on this vector and correlates all inspection data with real-time process parameters. Through big data mining, the causal chain between appearance defects and process parameters is generated to generate optimization warnings.

[0027] An injection mold for multi-color injection molding, wherein the injection mold adopts the above-mentioned production process; The mold itself, as a physical carrier, has its cavity design, temperature control system, and moving mechanism pre-configured to respond to and execute the aforementioned optimization instructions. For example, the mold's cavity profile is highly matched with the dynamic benchmark model of the digital twin simulation to accommodate expected material shrinkage. The mold's cooling circuit can be precisely temperature-controlled based on multiphysics detection feedback to optimize the interface bonding state. The mold's moving components (such as sliders or rotating mechanisms) can receive more precise stroke control or wait for parameter fine-tuning under critical conditions based on augmented reality-assisted judgment or adhesive strength prediction results. At the same time, the mold's fixed frame or the mold itself has reserved standardized interfaces to facilitate the integration of sensing units such as pulsed lasers, multi-band light sources, and polarization optical components, providing a stable physical benchmark and detection environment for online optical sieve detection.

[0028] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A production process for multi-color injection molding, characterized in that, Includes the following steps: S100: Perform the first injection to form molded part A; S200: Perform machine vision inspection on the molded part A, including a light sieving inspection process, wherein the light sieving inspection process includes: S210: Perform style detection on a specific part of the molded part A, including horizontal coordinate comparison and vertical coordinate comparison. The horizontal coordinate comparison is performed by comparing the actual horizontal coordinate of the specific part with a preset horizontal coordinate range to detect style deviation. The vertical coordinate comparison is performed by comparing the actual vertical coordinate of the specific part with a preset vertical coordinate range to detect style deviation. S220: Perform surface flatness testing on the molded part A, and evaluate whether its surface flatness meets the requirements of the second injection by analyzing the reflected light distribution on the surface of the molded part A. S300: Based on the results of the machine vision inspection, if the pattern matching rate or surface flatness is detected to be lower than the threshold, the subsequent injection molding process is stopped; otherwise, a second injection is performed.

2. The production process for multi-color injection molding according to claim 1, characterized in that, Step S100 includes the following steps: S110: Dry and heat the thermoplastic material required for the first shot to a molten state; S120: Close the mold to form the cavity of the molded part A and control the mold temperature; S130: Inject molten material into the cavity, and hold pressure after filling; S140: After the molded part A has cooled and solidified in the cavity, the mold is opened.

3. The production process for multi-color injection molding according to claim 2, characterized in that, Step S210 includes the following steps: S211: Before injection, based on the real-time process parameters of the first injection, drive the digital twin model of the molded part A to perform synchronous thermodynamic-structural mechanics simulation, predict and generate a dynamic three-dimensional reference model containing material shrinkage and cooling deformation under this specific production cycle in real time. S212: A scanning device incorporating white light interference is used to perform a three-dimensional scan on the specific area to obtain a two-dimensional image of the surface of the molded part A; A scanning device combining spectral domain optical coherence tomography is used to perform three-dimensional scanning on the specific part, penetrating the surface layer of the transparent or semi-transparent first-electrode material to obtain the internal structure profile image within a depth range of 0.1-0.5 mm below the subsurface, thereby simultaneously capturing surface contour and subsurface defect information. S213: Fuse the surface 3D point cloud data obtained in step S212 with the subsurface profile image data to construct an extended 3D dataset containing surface and subsurface information; The extended 3D dataset is aligned with the dynamic 3D benchmark model generated in step S211 in four-dimensional spatiotemporal coordinates and time dimension based on injection timing. S214: Input the aligned extended 3D dataset into a pre-trained deep learning convolutional neural network; The network analyzes surface geometric features, subsurface structural features and their spatial correlations, and directly outputs a comprehensive confidence score regarding whether the style of the specific part is qualified or not. The horizontal coordinate comparison and vertical coordinate comparison are embedded as primary features in the underlying perception module of the network. The network's decision goes beyond simple coordinates and is based on a deep understanding of the overall shape and the integrity of the internal structure. S215: The comprehensive confidence score obtained in step S214 and the identified defect feature types are fed back to the injection molding machine control system in real time; if the pattern is qualified, the second injection preparation is triggered. If there is a compensable pattern deviation, a feedforward control command is generated to automatically fine-tune the injection parameters of the second shot to actively compensate for the minor deformations present in the first shot during the second shot injection; at the same time, all data from this test are recorded to the digital twin model to optimize the accuracy of subsequent simulation predictions.

4. The production process for multi-color injection molding according to claim 3, characterized in that, Step S220 includes the following steps: S221: After the molded part A completes its first injection and enters the inspection station, the following operations are performed simultaneously: S221a: Structured light in the short-wave infrared band is projected onto the surface to be inspected, and its deformation image is acquired; S221b: Using a high-resolution thermal imaging camera, the temperature field distribution map of the surface of the part is acquired to obtain its real-time thermodynamic state during the cooling process; S222: The short-wave infrared structured light image acquired in S221a is used to reconstruct a high-sensitivity three-dimensional micro-topography point cloud through phase calculation; the topography point cloud is then fused with the surface temperature field data acquired in S221b at the pixel level to generate a multi-physics point cloud model containing the coordinates (X, Y, Z) of each spatial point, the temperature (T), and the local shrinkage stress (σ) of the material derived from the temperature. S223: Input the multiphysics point cloud model into a second injection molding process simulator. The simulator uses the precise morphology, temperature and stress state of the current first injection surface as initial conditions to simulate the entire process of the second injection melt flowing, filling, cooling and bonding with the first injection, and outputs the predicted bonding line strength, coverage integrity and potential weld line location and other key quality indicators. S224: Make dynamic decisions based on the simulation prediction results of S223: If the predicted quality of the second shot fully meets the design requirements, then even if there is a slight deviation in geometric flatness, it is judged as qualified. If the prediction results show that the bonding quality is at a critical state, the system will automatically generate a personalized flatness tolerance for the specific part after tightening, and make a second fine judgment. If the prediction results show that the bonding quality is unqualified, it will be directly judged as unqualified; S225: For parts judged as qualified but with suboptimal predicted bonding quality, the system reverse-engineers the simulation results to generate a set of optimized second injection parameters, including but not limited to injection speed, melt temperature, and holding pressure. These parameters are then fed forward to the injection molding machine control system for the upcoming second injection, in order to proactively improve the quality of the final product. At the same time, all data from this test and the simulation prediction results are compared with the actual product quality after the second injection for continuous training and optimization of the simulator's prediction accuracy.

5. A production process for multi-color injection molding according to claim 1, characterized in that, In step S200, if the result of the machine vision inspection is in a critical state or the confidence level is below a threshold, the augmented reality-assisted human visual inspection step S300 is initiated simultaneously, wherein S300 includes: S310: When the quality inspector wears augmented reality glasses and looks at the molded part A, the glasses perform the following steps: S311: Track and locate parts in real time through image recognition, and overlay the machine vision results from steps S210 and S220 onto the corresponding positions of the physical object in the form of highlighted outlines, color codes, or virtual arrows. S312: Display an optimized virtual inspection path in the field of view to guide the quality inspector's gaze to systematically cover the entire area to be inspected, ensuring no omissions; S321: The eye tracker built into the augmented reality glasses tracks the eye movement trajectory and fixation time of the quality inspector in real time. The system automatically records the attention duration of the defective area indicated by the machine and compares it with the non-indication area. S322: When a quality inspector makes a judgment on a defect, the system records the reaction time from observation to decision-making and uses this data as a quantitative indicator of the judgment complexity. S323: Quality inspectors wear lightweight EEG sensors, and the system monitors the EEG signals of their prefrontal cortex, assesses their focus and cognitive load during the testing process, and issues a reminder when their focus drops below a threshold to ensure the reliability of the judgment. S331: Quality inspectors can operate the virtual defect menu superimposed on the physical object through predefined gestures to complete the command input without operating the physical terminal; S332: For complex or novel defects, quality inspectors can describe them by voice input. The system will automatically convert the voice into text and bind and store it with the defect image and machine data in the current field of view to build a searchable defect knowledge graph. S341: Link the entire process data of this manual inspection with the production batch of the part and the original machine vision data to form a complete and traceable digital inspection thread; S342: Perform bidirectional optimization of the machine vision model: a) Forward optimization: The defect data that has been finally confirmed by humans is used as an incremental labeled dataset to periodically retrain the deep learning network in step S214 to improve its automatic recognition ability. b) Backward calibration: Analyze the rejection data of quality inspectors for machine false alarm areas to dynamically relax the detection threshold in specific scenarios and reduce the false alarm rate; at the same time, analyze the defect data that the machine missed but was discovered by humans to dynamically tighten the detection sensitivity of relevant features. c) Human Factors Reliability Assessment: By integrating eye movement and reaction time data, a reliability model is established for each quality inspector. For high-difficulty defects, priority is given to quality inspectors with high reliability scores, thereby optimizing the allocation of human resources.

6. The production process for multi-color injection molding according to claim 1, characterized in that, After completing the second injection and forming the molded part B, the following steps are performed: S600: Perform interfacial quality optical screening on the molded part B, the screening including at least the staining detection of the interface between the first and second injection materials; S610: Using an industrial camera equipped with a multi-band light source, an image of the interface region of the molded part B is acquired under a spectrum including visible light and a specific near-infrared band; wherein the specific near-infrared band is configured to have high contrast sensitivity to the minute color diffusion of the first and second emission materials. S621: Extract the spectral features of the pure first emission material region and the pure second emission material region from the multispectral image and use them as reference spectra; S622: Employ a spectral unmixing algorithm to analyze each pixel on the interface and calculate the mixing ratio of the first emission material reference spectrum and the second emission material reference spectrum in its spectrum; S623: Based on the mixing ratio, generate a color diffusion distribution map, which quantitatively characterizes the degree of diffusion and spatial distribution of the first color in the second color in the second color. S631: Compare the staining distribution map with the pre-stored staining-mechanical property correlation model; the correlation model is established based on historical data and describes the influence of different staining degrees on key mechanical properties such as the overall structural strength and bonding force of the part. S632: If the current bleeding state is predicted by the model to have no significant impact on the mechanical properties of the part, it is judged as qualified; if the prediction is that it will cause the mechanical properties to drop below the threshold, or the bleeding is located in the critical stress area, it is judged as unqualified. S640: The bleeding detection results are correlated with the machine vision inspection results after the first injection. If a statistical correlation is found between the bleeding defect and the specific surface flatness or pattern deviation of the first injection part, a process adjustment suggestion is generated and fed back to the injection parameter control system of the first or second injection.

7. A production process for multi-color injection molding according to claim 1, characterized in that, After completing the second injection and forming the molded part B, the following steps are performed: S700: Perform optical screening on the degree of interfacial adhesion of the molded part B; S710: A pulsed laser is used to locally scan and excite the interface area of ​​the molded part B. The laser energy is absorbed by the material to generate high-frequency ultrasonic waves. At the same time, a laser interferometric vibrometer is used in conjunction with the laser interferometric vibration measurement principle to non-contactly and in the whole field measure the minute vibration displacement field induced by the ultrasonic waves on the surface of the part. S721: Extract the local vibration mode parameters of the interface region from the full-field vibration data obtained in step S710, including the resonance frequency, mode shape and damping ratio; S722: Input the extracted measured modal parameters into a pre-trained inversion model; this inversion model is established by combining finite element analysis and machine learning, and can accurately invert the effective adhesive stiffness of the interface layer and the size and location of the unbonded area based on the vibration characteristics. S731: Input the adhesive stiffness and debonding information obtained from the inversion into a micromechanical model to predict the peel strength and shear strength of the interface under actual working conditions. S732: Compare the predicted bond strength with the minimum strength threshold required by the product design to make a functional judgment: if the predicted strength is higher than the threshold, it is judged as qualified; if it is lower than the threshold, or a continuous critical debonding area is detected, it is judged as unqualified. S740: Perform multivariate correlation analysis on the adhesion degree test results, the surface smoothness data after the first injection, and the injection parameters of the second injection to establish the mapping relationship between adhesion quality and process parameters; If a poor adhesion trend is detected, the system automatically generates a set of optimized parameters for the second injection speed, temperature, or pressure, and feeds them forward to the next production cycle to achieve adaptive adjustment of the process.

8. A production process for multi-color injection molding according to claim 1, characterized in that, After the molded part B is demolded, the following steps are performed: S800: Perform overall appearance defect optical screening on the molded part B; S810: In the integrated detection unit, the surface information of the molded part B is collected synchronously in the following manner: S811: Uses an integrating sphere or surrounding LED light source to acquire images under uniform diffused light conditions for detecting color uniformity, blemishes, and macroscopic defects; S812: Uses a low-angle linear light source to illuminate the surface in a specific direction to highlight micro-texture defects such as scratches, knife lines, and orange peel texture; S813: Using orthogonal polarizers placed in front of the light source and camera respectively, the specular reflection component and diffuse reflection component in the surface reflected light are separated and collected to eliminate reflection interference and detect defects such as subsurface stress lines and internal stress whitening. S821: Register and fuse the multimodal images obtained in S810 to construct a multi-channel feature image cube; S822: The U-Net deep learning model based on the attention mechanism is used to perform pixel-level segmentation on the feature image cube, automatically identify and label all potential defect areas, and pre-classify them according to their feature vectors. S831: Input the segmented defect areas and their features into a defect semantic network. This network comprehensively considers the type, size, quantity, contrast, location of defects, and degree of conflict with the product design aesthetic language. S832: The semantic network outputs a non-binary quality compliance vector, which contains scores of multiple dimensions, providing a quantitative basis for the final disposal decision; S841: Based on the quality compliance vector of S830, dynamically sort part B to the qualified area, special release area, rework area, and scrap area. S842: Spatiotemporally correlates all detection data with the real-time process parameter stream of the injection molding machine. Through big data analysis, it uncovers potential causal chains between specific appearance defects and process parameters such as mold status, injection curve, and temperature control, and generates maintenance and optimization warnings.

9. An injection mold for multi-color injection molding, characterized in that, The injection mold is manufactured using the production process described in any one of claims 1-8.

Images