Shell diaphragm digital twinning production method based on knowledge graph driving

Abstract

The invention relates to technologies of digital twinning and the like, and provides a digital twinning production method for shell diaphragms based on knowledge graph driving. According to the method, a workpiece texture image is collected through a visual imaging module, a production line controller synchronously collects process multi-source data and extracts defect features and geometric features in combination with a visual recognition model, and an observation data set with a unified time reference is formed; writing data and workpiece texture types, batches and formulas into a knowledge graph, and outputting adjustable process parameters and constraint intervals by an inference engine; a digital twin engine comprising a mechanism and a data driving model is called to carry out state estimation and parameter calibration, optimal process parameters are obtained based on multi-objective optimization, an execution controller sends the parameters to a transfer printing unit, a printing unit and a curing unit, production is carried out according to a quality gating sequence, and the production process is finished. And on-line calibration and incremental updating are carried out on the digital twinning and knowledge graph based on a recheck result. According to the invention, detection condition consistency and model adaptive updating are realized, the first-pass yield is improved, and energy consumption and rejection rate are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical fields of digital twinning, knowledge graph, shell membrane production, and the like, and particularly relates to a shell membrane digital twinning production method based on knowledge graph driving. BACKGROUND

[0002] With the wide application of electronic components, display modules and functional membrane products, roll-to-roll transfer printing and multi-layer printing processes are widely used in membrane manufacturing processes. The existing membrane production line usually includes a transfer unit, a printing unit and a curing unit, and cooperates with optical detection equipment to perform quality inspection on the product. However, the existing technology still has the following shortcomings: First, data fragmentation and time alignment are difficult. On the traditional production line, image detection data and process parameter data are often collected independently by different devices and systems, lacking a unified time reference. Even if the same workpiece is detected, it is difficult to ensure that the image frame and the process data are in the same time window, resulting in incomplete observation data and difficulty in establishing a corresponding relationship between defect features and process factors.

[0003] Second, process parameter setting relies on experience and lacks optimization constraints. Currently, key process parameters such as ultraviolet curing time and energy, lamp distance, tension, line speed and spraying volume are mostly set by manual experience or single variable experiments, lacking a systematic multi-objective optimization and parameter constraint mechanism, which can easily cause excessive energy consumption, material waste, and even new defects.

[0004] Third, product quality fluctuates and the rework rate is high. Due to differences in material characteristics between different batches, production equipment state degradation (such as light source attenuation, roller system eccentricity, bearing wear), changes in environmental temperature and humidity, and other influences, traditional control methods are difficult to timely correct process parameters, resulting in poor batch-to-batch and within-batch consistency, low first-pass yield, and high rework and scrap rates.

[0005] In addition, the detection conditions are unstable. In the existing detection link, the poses of the camera and the light source are not uniformly included in the adjustable parameter set, and the detection conditions fluctuate with the changes in the process or batch, resulting in deviations in the recognition results of similar defects under different working conditions, affecting the reliability and comparability of the detection.

[0006] Finally, there is a lack of adaptive closed-loop updating mechanism. Although existing production lines have introduced some digital detection and statistical process control methods, they often lack online calibration and knowledge updating functions based on reinspection feedback. With the long-term operation of the production process, the visual recognition model and the process prediction model are prone to drift, and if they are not updated in time, they will gradually deviate from the actual working conditions, weakening the guiding role of defect prediction and optimization.

[0007] In summary, the existing membrane transfer and printing production method has obvious deficiencies in data fusion, parameter optimization, detection consistency and model adaptation, and a new technical solution is needed to realize data and image fusion under a unified time reference, multi-objective optimization and constraint solving of process parameters, unified control of detection conditions and model adaptive update based on re-inspection feedback, thereby improving the first pass yield, reducing unit energy consumption and waste rate. SUMMARY

[0008] To solve the above problems of the prior art, the present application provides a shell membrane digital twin production method based on knowledge graph driving, which realizes data and image fusion under a unified time reference, multi-objective optimization and constraint solving of process parameters, unified control of detection conditions and model adaptive update based on re-inspection feedback, thereby improving the first pass yield, reducing unit energy consumption and waste rate.

[0009] The shell membrane digital twin production method based on knowledge graph driving provided by the present application comprises: The membrane output station of the roll-to-roll transfer machine collects the texture image of the same workpiece by the visual imaging module, the process line controller synchronously collects the process line multi-source data, obtains the defect feature set and the geometric feature set based on the visual recognition model, and forms an observation data set based on a unified time reference in combination with the process line multi-source data and the same workpiece identification; the process line comprises a transfer unit, a printing unit and a curing unit; The observation data set and the texture type, batch and formula of the workpiece are written into the defect knowledge graph, and the knowledge graph reasoning engine outputs an adjustable process parameter set and a parameter constraint interval; a digital twin engine containing a process mechanism model is called to perform state estimation and parameter calibration, multi-objective optimization is constructed, and the optimal process parameters are obtained within the constraints, and the adjustable process parameter set comprises camera pose and light source pose; The execution controller downlink the optimal process parameters to the transfer unit, the printing unit and the curing unit, implements production according to the predetermined quality gate sequence, and performs online calibration and incremental update of the digital twin and the defect knowledge graph based on the re-inspection result.

[0010] Compared with the prior art, the present application has the following advantages: The application provides a shell membrane digital twin production method based on a knowledge graph driving, which comprises the following steps: collecting a texture image of a same workpiece at a membrane output station of a roll-to-roll transfer printer by a visual imaging module, synchronously collecting process multi-source data in a production line by a production line controller, obtaining a defect feature set and a geometric feature set based on a visual recognition model, and jointly forming an observation data set based on a unified time reference with the process multi-source data and the same workpiece identification; the production line comprises a transfer unit, a printing unit and a curing unit; writing the observation data set and the texture type, batch and formula of the workpiece into a defect knowledge graph, outputting an adjustable process parameter set and a parameter constraint interval by a knowledge graph reasoning engine; calling a digital twin engine containing a process mechanism model to perform state estimation and parameter calibration, constructing a multi-objective optimization and obtaining optimal process parameters within the constraint, and the adjustable process parameter set comprises a camera pose and a light source pose; the optimal process parameters are sent to the transfer unit, the printing unit and the curing unit by an execution controller, production is implemented according to a predetermined quality gating sequence, and online calibration and incremental updating of the digital twin and the defect knowledge graph are performed based on reinspection results. The method of the application can solve the problems of image process misalignment caused by data fragmentation and inconsistent time stamps of the membrane printing production line, parameter selection relying on experience, high quality fluctuation and rework, and the causal relationship of defect features, geometric features, process data, production equipment state and material formula is structured and deposited, parameter solving is constrained, interpretable and traceable; the camera and light source pose enter a unified control domain, and the detection condition is stable; reinspection, calibration and incremental updating ensure that the visual recognition model and the digital twin engine can keep up with the changes of the field data in long-term operation, and the first pass yield, unit energy consumption and scrap rate are comprehensively improved. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings, which are not necessarily drawn to scale, like reference numerals describe similar components throughout the several views. The specific embodiments of the present application will now be described with reference to the drawings. Figure 1 FIG. 1 is a flowchart of a shell membrane digital twin production method based on a knowledge graph driving according to an embodiment of the application. DETAILED DESCRIPTION

[0012] In order to make personnel in the technical field better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor should belong to the scope of protection of the present application.

[0013] Referring to Figure 1 The present embodiment provides a shell membrane digital twin production method based on knowledge graph driving, comprising the following steps: S101, the texture-containing image of the same workpiece is collected by a visual imaging module at the membrane output station of a roll-to-roll transfer printer, a process multi-source data in a production line is collected synchronously by a production line controller, a defect feature set and a geometric feature set are obtained based on a visual recognition model, and observation data sets based on a unified time reference are formed in combination with the process multi-source data and the same workpiece identification; the production line comprises a transfer unit, a printing unit and a curing unit; S102, the observation data sets and the texture type, batch and formula of the workpiece are written into a defect knowledge graph, a set of adjustable process parameters and a parameter constraint interval are output by a knowledge graph reasoning engine, a digital twin engine containing a process mechanism model is called to perform state estimation and parameter calibration, a multi-objective optimization is constructed and optimal process parameters are obtained within the constraint, and the set of adjustable process parameters comprises a camera pose and a light source pose; S103, the optimal process parameters are issued to the transfer unit, the printing unit and the curing unit by an execution controller, production is implemented according to a predetermined quality gate sequence, and online calibration and incremental update of the digital twin and the defect knowledge graph are performed based on the reinspection results.

[0014] In this embodiment, the film output station of the roll-to-roll transfer machine collects texture-containing images of the same workpiece by a visual imaging module, and the process line controller synchronously collects multi-source data of the process line. Based on the visual recognition model, an observation data set based on a unified time reference is formed. The process line includes transfer, printing, and curing units. The observation data, texture type, batch, and formula are written into a defect knowledge graph, and the adjustable process parameters and constraints are inferred. A digital twin engine containing a mechanism model is called to perform state estimation and parameter calibration and construct a multi-objective optimization, and the camera pose and light source pose are included in the adjustable set. The optimal parameters are issued by the execution controller and calibrated and incrementally updated online according to the reinspection results, which can solve the problems of image process misalignment caused by data fragmentation and inconsistent time stamps in the film printing production line, parameter selection relying on experience, quality fluctuation, and high rework rate. The scheme of this embodiment structures the causal relationship between defect features, geometric features, process data, production equipment state, and material formula, and the parameter solving is constrained, interpretable, and traceable. The camera and light source pose enter a unified control domain, and the detection conditions are stable. Reinspection, calibration, and incremental updating ensure that the visual recognition model and the digital twin engine always keep up with the changes in the field data in long-term operation, and comprehensively improve the first-time yield, reduce unit energy consumption, and reduce the waste rate. Among them, the visual recognition model is used to extract a defect feature set and a geometric feature set from a workpiece image. It may drift between batches (different batches of workpieces) and within a batch (the same batch but over time) due to factors such as lighting and material batch differences. The digital twin engine is used for state estimation, parameter calibration, and multi-objective optimization. It may also produce deviations between the model and the real process due to equipment degradation and material property fluctuations in long-term operation.

[0015] Further, the observation data set adds image preprocessing and imaging condition standardization steps before being written into the defect knowledge graph, including using a standard whiteboard and a gray scale card for flat field correction to eliminate uneven lighting, using adaptive histogram equalization and glare suppression filtering to improve texture separability, automatically setting exposure time and gain and white balance parameters based on target signal-to-noise ratio and definition evaluation, combining polarization plate angle adjustment to suppress specular components on high-reflectivity materials, and writing the imaging parameter vector including exposure time, gain, white balance, polarization angle, lens focal length, and aperture state into the observation data set based on a unified time reference as image quality metadata, to stabilize the quantization output of geometric features and defect features when the optical conditions change between batches and within a batch, and improve the robustness of knowledge graph reasoning and digital twin engine parameter calibration.

[0016] Furthermore, through the data quality scoring and weight propagation mechanism, the data quality score DQI is calculated for the object data of the same workpiece formed under a unified time base, which includes an observation data set based on channel integrity, sensor self-test status, range occupancy and saturation detection, zero drift and temperature drift evaluation, focus clarity index, signal-to-noise ratio and glare ratio, and coordinate mapping residual. The DQI is written into the defect knowledge graph along with the workpiece identification and serves as an adjustment factor for the relationship weight and parameter constraint credibility. At the same time, in the multi-objective optimization, the defect probability, geometric deviation and energy consumption per unit area target items are adaptively weighted according to the DQI. In the online learning stage, the DQI and prediction uncertainty are jointly used to drive sample selection and incremental update priority, thereby reducing the risk of misjudgment and misadjustment under the conditions of data noise and equipment status fluctuations and improving the stability of optimal process parameter solution and execution.

[0017] Preferably, the production line controller sends a trigger signal with a unified time reference to the visual imaging module and the process sensor, and collects the image frames and multi-source data of the process in the same time window and adds the same timestamp and workpiece identification. The multi-source data of the process include at least UV curing time, UV curing energy, lamp distance, tension setting, line speed, spraying amount, ambient temperature and humidity, spectral irradiance and light transmittance detection results; a target plate is used to jointly calibrate the camera coordinate system, transfer coordinate system and printing coordinate system, establish a mapping relationship between each coordinate system, and obtain reference values ​​of the camera posture and light source posture, so that the geometric features of line width, grid spacing, waveform amplitude, alignment error and texture similarity are quantified and output in a unified coordinate system, and written into the observation data set after being bound to the same workpiece identification.

[0018] In this embodiment, the production line controller sends a trigger signal with a unified time reference to the visual imaging module and the process sensor, so that the image frames and multi-source process data in the same time window share the same timestamp and workpiece identification; the multi-source data contains at least UV curing time, energy, lamp distance, tension, line speed, spraying amount, ambient temperature and humidity, spectral irradiance and light transmittance detection results; a target plate is used to complete the joint calibration of the camera, transfer and printing coordinate systems and obtain the reference values ​​of the camera posture and light source posture, so that the line width, grid spacing, waveform amplitude, alignment error and texture similarity are quantified in a unified coordinate system, which can solve the problems of incomparable eigenvalues, difficult tracing and large amount of dirty data in model training samples caused by the lack of synchronization between image and process data and the inconsistent coordinate system, and realize cross-device and cross-process time-space integrated measurement and traceable modeling. Among them, time alignment ensures a close correlation between what is seen and what is produced for each workpiece, and joint calibration allows the geometric quantities to eliminate interference from perspective and distortion and express them in a unified manner with the process posture; the posture reference value can provide a benchmark for the subsequent posture linkage control and consistency of detection conditions, reduce the misjudgment rate, and provide a high-quality observation data base for knowledge graphs and digital twins.

[0019] Preferably, the ontology of the defect knowledge graph includes entities and their relationships of texture types, defect types, process parameters, production equipment states, materials, batches, and quality indicators of workpieces, and the knowledge graph reasoning engine outputs a set of adjustable process parameters, parameter constraint intervals, and target geometric ranges based on rule-based reasoning and probabilistic reasoning, manages rule versions and confidence threshold values, triggers re-inspection when the confidence of the visual recognition model is lower than the threshold, and updates relationship weights and rule priorities.

[0020] In this embodiment, a defect knowledge graph containing texture types, defect types, process parameters, equipment states, materials, batches, and quality indicator entities and relationships is constructed, rule-based reasoning and probabilistic reasoning are combined to output a set of adjustable process parameters, parameter constraint intervals, and target geometric ranges, and rule versions and confidence threshold values are managed; when the confidence of the visual recognition model is lower than the threshold, re-inspection is triggered and relationship weights and rule priorities are updated, which can solve the problems of single empirical rule rigidity, cross-formulation, poor generalization of new materials, and inconsistent processing of low-confidence samples, and realize an adaptive reasoning framework that combines interpretable rules and statistical learning. The knowledge graph explicitly shows the causal context of historical defects, parameters, equipment degradation, and batches, and the constraint intervals are given by both facts and experience, avoiding unbounded divergence in optimization; the confidence gating and re-inspection closed loop suppresses false triggering; rule versioning and weight updating allow the system to evolve with data, maintaining stability while being dynamic, and reducing the impact of incorrect adjustments on quality fluctuations.

[0021] Preferably, the digital twin engine is a hybrid model of mechanism models and data-driven models, the mechanism models at least include ultraviolet curing dynamics sub-models, transfer and superposition alignment sub-models, and spraying rheology and deposition thickness sub-models, and the data-driven models correct the mechanism models through residual learning or online parameter estimation, and uniformly output prediction results of defect occurrence probability, line width variance, alignment error, and light transmittance.

[0022] In this embodiment, a hybrid digital twin of mechanism models (ultraviolet curing dynamics, transfer and superposition alignment, spraying rheology and deposition thickness sub-models) and data-driven models is used, the mechanism models are corrected by residual learning or online parameter estimation, and prediction results of defect occurrence probability, line width variance, alignment error, and light transmittance are uniformly output, which can solve the problems of poor adaptability of pure mechanism models to complex materials and environments, and lack of interpretability and instability in extrapolation scenarios of pure data models, and realize high-precision prediction with both physical interpretability and data adaptability. The mechanism framework provides structured priors and differentiable parameters, the data branch online absorbs drift and unknown disturbances (such as formula micro-variation, lamp decay, and tension disturbance), the residual closed loop reduces systematic bias, makes the prediction closer to the field, supports subsequent model predictive control methods for forward-looking optimization of quality and energy consumption, and reduces offline samples and trial-and-error downtime.

[0023] Preferably, the objective function of the multi-objective optimization includes defect occurrence probability, energy consumption per unit area, waste cost, parameter variation amplitude, and deviation between digital twin simulation geometry and geometric features, the constraint conditions include device safety threshold, material compatibility, quality threshold, and process window, and the optimization algorithm adopts a model predictive control method for solving.

[0024] In this embodiment, the defect occurrence probability, energy consumption per unit area, waste cost, parameter variation amplitude, and deviation between digital twin simulation geometry and geometric features are taken as multi-objectives, the device safety threshold, material compatibility, quality threshold, and process window are taken as constraints, and the optimization algorithm adopts a model predictive control method, which can solve the problem of unstable manual heuristic weighting in single-objective parameter adjustment (only looking at defects or only looking at energy consumption) and realize comprehensive optimal control for quality, cost, energy consumption, and executability. The parameter variation amplitude is introduced in the objective to suppress execution fatigue and system oscillation caused by frequent large-step parameter adjustment, and the constraints ensure that the system does not exceed the boundary and is not damaged; the model predictive control method uses the predictability of digital twin to select a controllable sequence in a rolling time domain, taking into account the response speed and smoothness, and improving the throughput and stability of the production line.

[0025] Further, the specific steps of the optimization algorithm adopting the model predictive control method for solving include: constructing a disturbed state space prediction model based on the mechanism model and the data-driven model of the digital twin engine and performing online parameter calibration under the unified time reference from the observation data set, setting the prediction time domain H, the control step number M, and the sampling period Ts, and taking the last period optimal solution and the parameter constraint interval output by the knowledge graph as the warm start and feasible region, constructing a weighted objective function including the defect occurrence probability, energy consumption per unit area, waste cost, parameter variation amplitude, and deviation between digital twin simulation geometry and geometric features, and imposing device safety threshold, material compatibility, quality threshold, process window, and parameter speed and amplitude limit constraints, solving a finite time domain optimization problem based on current state estimation and disturbance prediction at each sampling time to obtain a future M-step process parameter increment sequence and performing parameter interlocking and cross-unit conflict checking before issuing, issuing only the first step parameter to the transfer unit, the printing unit, and the curing unit and rolling updating the prediction model, state estimation, and objective weight after execution, triggering feasibility repair and safety rollback to continue production using the stored safe compromise parameters when the feasible region is empty or the constraint activity and risk assessment exceed the threshold, and writing the execution results and re-inspection data into the database for incremental updating of digital twin and defect knowledge graph.

[0026] Preferably, the execution controller implements production in accordance with the process gating sequence, which includes, in sequence, bridging varnish printing after the basic texture inspection is qualified, fine texture transfer re-inspection, cover base color printing, lamination glue printing and light transmission re-inspection, and links the camera posture and posture parameters including the light source posture of position and angle to keep the detection conditions consistent. If any gated node is unqualified, it is prohibited to enter the next process; the re-inspection results, defect feature set, geometric feature set and process deviation are written into a database aligned with a unified time base for online calibration of the digital twin model and incremental update of the defect knowledge graph.

[0027] In this embodiment, the following gated processes are executed sequentially: basic texture inspection, bridging varnish printing, fine texture transfer re-inspection, base color printing, laminating adhesive printing, and light transmission re-inspection. The camera posture and light source posture (position and angle) are linked to maintain consistent inspection conditions. If any gate fails, the product is discontinued. Re-inspection results and deviations are written to a database aligned with a unified time base for online calibration and incremental map updates. This approach addresses issues such as misjudgment, defective product transfer, and traceability difficulties caused by inconsistent inspection conditions at different workstations, achieving a strongly interlocked quality gate for process, inspection, decision-making, and documentation. Posture linkage minimizes measurement uncertainty due to differences in illumination and viewing angle. Strict gating prevents the spread of defects. Data accumulation throughout the entire process feeds back into the visual recognition model and digital twin model, forming a closed loop of continuous improvement across inspection, control, and learning, improving first-time pass rates and re-inspection consistency. Re-inspection and inspection data enhance the visual recognition model's accuracy in identifying defects and geometric features. Parameter and residual updates from full-link observation data ensure the accuracy of the digital twin model's predictions and optimization.

[0028] Preferably, a new set of geometric features is extracted in real time at each gated node and compared with the target geometric range. When the deviation exceeds a threshold, parameter fine-tuning is performed. The fine-tuning is limited by the speed and amplitude limiting strategies of the parameter step size and the parameter change amplitude, and supports small-scale grayscale verification in a preset interval at a local workstation or within a short time window. When the geometric feature or defect risk exceeds the threshold, exception handling is triggered. Exception handling includes a virtual test run in a digital twin environment to evaluate the feasibility and risk of the parameters to be issued. If the verification fails, a safety fallback strategy is executed, and production continues using the stored safety compromise parameters. The optimization strategy and closed-loop learning are restored after new high-quality observation data are formed.

[0029] In this embodiment, new geometric features are extracted in real time at each control node and compared with the target range, and the parameter adjustment is performed when the deviation exceeds the threshold; the adjustment is constrained by the speed limit and amplitude limit of the parameter step and change range, and supports small-range gray-scale verification in a preset interval within a local station or short time window; when the geometric feature or defect risk exceeds the threshold, a virtual test run is performed in the digital twin, and if it fails, the production continues with the stored safe compromise parameters, and the optimization and closed-loop learning are restored after new high-quality observation data are formed, which can solve the problems of easy oscillation of fast parameter adjustment, high risk of one-time large change, and high cost of online verification, and realize safe, gradual, and verifiable online self-optimization. The speed limit and amplitude limit suppress control overshoot; gray-scale verification allows changes to be controlled in a local and short time; virtual test run exposes the risk of the scheme in advance; safe rollback ensures the continuity of the production line and the bottom line of yield, and takes into account learning and steady-state production.

[0030] Preferably, the ultraviolet curing kinetics sub-model is used to characterize the relationship between spectral irradiance, energy density, and curing degree, the transfer and superposition alignment sub-model is used to characterize the influence of tension, line speed synchronization, and phase compensation on alignment error and waveform parameters, and the spraying rheology and deposition thickness sub-model is used to characterize the relationship between viscosity and shear rate, wetting angle, and film thickness distribution.

[0031] In this embodiment, ultraviolet curing kinetics is used to describe the relationship between spectral irradiance, energy density, and curing degree; the transfer and superposition alignment sub-model is used to describe the influence of tension, line speed synchronization, and phase compensation on alignment error and waveform parameters; and the spraying rheology and deposition thickness sub-model is used to describe the relationship between viscosity, shear rate, wetting angle, and film thickness distribution, which can solve the problem of unclear key mechanism influence path and difficulty in positioning sensitive parameters, and realize sensitivity list and quantitative coupling for adjustable parameter channels. The three types of sub-models integrate energy, reaction, and structure, motion, synchronization, and geometry, and rheology, spreading, and thickness, guiding the selection of the most effective parameter adjustment lever (such as lamp distance, energy, tension, phase compensation, spraying amount, and line speed) under different defect spectra, improving the directionality of diagnosis and control, and reducing invalid trial and error.

[0032] Preferably, the set of process parameters includes at least ultraviolet curing time, ultraviolet curing energy, lamp distance, tension setting, line speed, spraying amount, camera pose, and light source pose, the amplitude limit interval is determined jointly by the parameter constraint interval output by the knowledge graph and the device safety threshold, the speed limit is defined by the parameter increment generated by the model predictive control method, and the interlock and conflict verification is performed before the parameter is issued to ensure consistency with the process control sequence and avoid strategy conflicts across units, while the drift monitoring is performed on the historical parameter trajectory to prevent cumulative errors.

[0033] In this embodiment, the process parameter set is explicitly defined as UV curing time, energy, lamp distance, tension setting, line speed, spray volume, camera pose, and light source pose; the parameter constraint interval output by the knowledge graph is jointly determined with the equipment safety threshold to determine the limiting interval; the parameter increment limited by the model predictive control method (MPC) generates a speed limit; interlocking and conflict checking are performed before parameter issuance to avoid cross-unit strategy conflicts, while drift monitoring is performed on historical parameter trajectories to solve the problem of mutual interference conflicts, out-of-range risks, and chronic drifts that are not discovered in time, and to achieve controlled, coordinated, and auditable parameter management and issuance. The limiting amplitude and speed limit ensure the feasibility and smoothness of the execution level; interlocking verification avoids mutual contradictions such as increasing line speed and increasing exposure time; trajectory drift monitoring can early warn of light intensity attenuation or tension system deviation, and correct deviation in advance to stabilize process capability.

[0034] Preferably, when printing the base color of the cover, a secondary or tertiary printing strategy is used for the base color covering layer, and light transmission detection is performed after each printing. When the light transmission rate is higher than the threshold, the knowledge graph determines the reason to be insufficient ink layer thickness or uneven coating and gives the adjustment direction of the spray volume and line speed. The digital twin engine selects the combination of parameters that meet the light blocking and surface quality after simulating and screening different layer orders and single ink amounts.

[0035] In this embodiment, when printing the base color of the cover, a secondary or tertiary printing strategy is used for the base color covering layer, and light transmission detection is performed after each printing. When the light transmission rate is higher than the threshold, the knowledge graph determines the reason to be insufficient ink layer thickness or uneven coating and gives the adjustment direction of the spray volume and line speed. The digital twin engine selects the combination of parameters that meet the light blocking and surface quality after simulating and screening different layer orders and single ink amounts.

[0036] Preferably, when the knowledge graph determines that there is at least one of the following degradation signs: lamp group light intensity attenuation, roller system eccentricity, bearing gap increase, or lens contamination, based on the historical association of defect spectrum and production equipment state, the digital twin engine performs parameter inversion and sensitivity evaluation to estimate the degree of degradation and its influence on defect probability and energy consumption. When the maintenance threshold is exceeded, planned maintenance, lamp group calibration, or lens cleaning is triggered, and after maintenance is completed, baseline re-calibration is performed, and the new calibration parameters and rules are updated to the knowledge graph.

[0037] In this embodiment, when the knowledge graph determines the existence of lamp group light intensity attenuation, roller system eccentricity, bearing gap increase or lens pollution and other degradation signs according to the association between defect spectrum and device state history, the digital twin performs parameter inversion and sensitivity evaluation to estimate the degradation degree and its influence on defect probability and energy consumption; exceeding the maintenance threshold triggers planned maintenance, lamp group calibration or lens cleaning, and after maintenance, the baseline is recalibrated, and the new calibration and rules are updated to the knowledge graph, which can solve the problems of passive maintenance lag, quality fluctuation and timely attribution, and realize the transition from after-maintenance to predictive maintenance. Inversion and sensitivity correspond the defect increase to the degradation of a certain component; maintenance threshold triggers reduce unnecessary downtime; recalibration refreshes the model baseline to avoid old parameter pollution in the new stage; knowledge writing allows faster identification of similar degradation next time, stabilizing energy consumption and yield.

[0038] Preferably, the visual recognition model and the digital twin parameters are incrementally updated through an online learning mechanism and support transfer learning, and small sample data is used for self-adaptation when a new substrate or new formula is online to shorten the cold start time, and the update record is written synchronously to the knowledge graph and parameter library to keep the version traceable.

[0039] In this embodiment, the online learning mechanism is used to incrementally update the visual recognition model and the digital twin parameters and support transfer learning, and small sample self-adaptation is used to shorten the cold start time when a new substrate or new formula is online, and the update record is written synchronously to the knowledge graph and parameter library to keep the version traceable, which can solve the problems of few samples, large model bias and slow online climbing during the introduction period of new products, and realize fast adaptation and compliance traceability. Incremental learning fine-tunes the boundary with the latest batch data; transfer learning reuses historical domain knowledge to reduce labeling and trial production costs; versioned records support auditing and rollback to avoid long-term quality decline caused by false updates; comprehensive improvement of introduction efficiency and early yield makes the production line more flexible.

[0040] Preferably, the quality gate monitors the upper and lower control limits of line width variance, grid spacing deviation, waveform amplitude unevenness and transmittance based on statistical process control, dynamically adjusts the weight of multi-objective optimization and narrows the process window when the first yield or unit area energy consumption deviates from the target interval, and triggers rule relearning and mechanism parameter reestimation when a continuous multiple batches still do not meet the standard.

[0041] In this embodiment, statistical process control (SPC) is used to monitor upper and lower control limits for line width variance, grid spacing deviation, waveform amplitude unevenness, and transmittance. When the first-pass yield or energy consumption per unit area deviates from the target range, multi-objective optimization weights are dynamically adjusted and the process window is narrowed. When multiple consecutive batches fail to meet the target, rule relearning and mechanism parameter reassessment are triggered. This addresses issues such as process capability drift over time, rigid single-threshold control, and long-term weight mismatch, enabling adaptive process control and strategy upgrades in both quality and energy consumption. SPC provides real-time trends and anomaly signals; automatic weight adjustment dynamically shifts the optimization focus between quality and cost; window narrowing strengthens robustness during anomaly periods; and relearning and reassessment during persistent non-compliance completes strategy evolution, preventing symptomatic treatment without addressing the root cause, ultimately improving process stability and long-term first-pass yield.

[0042] It should be pointed out that the above embodiments are only preferred specific implementation methods of the present invention, and the protection scope of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. The protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A knowledge graph-driven housing diaphragm digital twin production method, characterized in that: include: At the film output station of the roll-to-roll transfer machine, a visual imaging module captures a textured image of the same workpiece. The production line controller simultaneously collects multi-source data of the production line process and obtains a defect feature set and a geometric feature set based on a visual recognition model. These are combined with the multi-source data of the process and the identification of the same workpiece to form an observation dataset based on a unified time base. The production line includes a transfer unit, a printing unit, and a curing unit. The observation data set and the texture type, batch and recipe of the workpiece are written into the defect knowledge graph, and the knowledge graph reasoning engine outputs the set of adjustable process parameters and parameter constraint intervals; Calling the digital twin engine containing the process mechanism model to perform state estimation and parameter calibration, constructing a multi-objective optimization and finding the optimal process parameters within the constraints. The adjustable process parameter set includes the camera posture and light source posture; The execution controller sends the optimal process parameters to the transfer unit, the printing unit and the curing unit, implements production according to the predetermined quality gating sequence, and performs online calibration and incremental update of the digital twin and defect knowledge graph based on the re-inspection results.

2. The method according to claim 1, characterized in that The production line controller sends a trigger signal with a unified time base to the visual imaging module and process sensor, and collects the image frames and multi-source data of the process in the same time window and adds the same timestamp and workpiece identification. The multi-source data of the process include at least UV curing time, UV curing energy, lamp distance, tension setting, line speed, spraying amount, ambient temperature and humidity, spectral irradiance and light transmittance detection results; a target plate is used to jointly calibrate the camera coordinate system, transfer coordinate system and printing coordinate system, establish a mapping relationship between each coordinate system, and obtain reference values ​​of the camera posture and light source posture, so that the geometric features such as line width, grid spacing, waveform amplitude, alignment error and texture similarity are quantified and output in a unified coordinate system, and written into the observation data set after being bound to the same workpiece identification.

3. The method according to claim 1, characterized in that The ontology of the defect knowledge graph includes entities and their relationships of the workpiece's texture type, defect type, process parameters, production equipment status, materials, batches and quality indicators. The knowledge graph reasoning engine outputs a set of adjustable process parameters, parameter constraint intervals and target geometric ranges based on rule reasoning and probabilistic reasoning, and manages rule versions and confidence thresholds. When the confidence of the visual recognition model is lower than the threshold, a re-inspection is triggered and the relationship weights and rule priorities are updated.

4. The method according to claim 2, characterized in that The digital twin engine is a hybrid model of a mechanism model and a data-driven model. The mechanism model includes at least a UV curing kinetics sub-model, a transfer and overlay alignment sub-model, and a spray rheology and deposition thickness sub-model. The data-driven model corrects the parameters of the mechanism model through residual learning or online parameter estimation, and uniformly outputs the prediction results of defect occurrence probability, line width variance, alignment error and transmittance.

5. The method according to claim 1, wherein The objective functions of multi-objective optimization include the probability of defect occurrence, energy consumption per unit area, scrap cost, parameter variation range, and the deviation between the digital twin simulation geometric quantity and geometric characteristics. The constraints include equipment safety threshold, material compatibility, quality threshold and process window. The optimization algorithm uses model predictive control method for solution.

6. The method according to claim 5, characterized in that The execution controller implements production in accordance with the process gating sequence, which includes bridging varnish printing after the basic texture inspection is qualified, fine texture transfer re-inspection, cover base color printing, lamination glue printing and light transmission re-inspection, and links the camera posture and posture parameters including the light source posture of position and angle to keep the detection conditions consistent. If any gated node fails, it is prohibited to enter the next process; the re-inspection results, defect feature set, geometric feature set and process deviation are written into a database aligned with a unified time base for online calibration of the digital twin model and incremental update of the defect knowledge graph.

7. The method according to claim 6, characterized in that At each gated node, a new set of geometric features is extracted in real time and compared with the target geometric range. When the deviation exceeds the threshold, parameter fine-tuning is performed. The fine-tuning is limited by the speed and amplitude limiting strategies of the parameter step size and the parameter change amplitude, and supports small-scale grayscale verification in a preset interval at a local workstation or within a short time window. When the geometric feature or defect risk exceeds the threshold, exception handling is triggered. Exception handling includes a virtual test run in a digital twin environment to evaluate the feasibility and risk of the parameters to be issued. If the verification fails, a safe fallback strategy is executed, and production continues using the stored safety compromise parameters. After new high-quality observation data is formed, the optimization strategy and closed-loop learning are restored.

8. The method according to claim 4, characterized in that The UV curing kinetics sub-model is used to characterize the relationship between spectral irradiance, energy density and curing degree; the transfer and superposition alignment sub-model is used to characterize the influence of tension, linear speed synchronization and phase compensation on alignment error and waveform parameters; the spray rheology and deposition thickness sub-model is used to characterize the relationship between viscosity and shear rate, wetting angle and film thickness distribution.

9. The method according to claim 7, characterized in that The process parameter set includes at least UV curing time, UV curing energy, lamp distance, tension setting, line speed, spraying amount, camera posture and light source posture. The limit interval is determined jointly by the parameter constraint interval output by the knowledge graph and the equipment safety threshold. The speed limit is limited by the parameter increment generated by the model predictive control method. Interlocking and conflict checks are performed before the parameters are issued to ensure consistency with the process gating sequence and avoid cross-unit policy conflicts. At the same time, the historical parameter trajectory is monitored for drift to prevent cumulative errors.

10. The method according to claim 6, characterized in that When printing the base color, the base color covering layer adopts a two- or three-printing strategy and a light transmittance test is performed after each printing. When the transmittance is higher than the threshold, the knowledge graph determines that the root cause is insufficient ink layer thickness or uneven coating and gives the adjustment direction of spraying amount and line speed. The digital twin engine simulates and screens different stacking sequences and single ink amounts, and then selects the combination parameters that meet the requirements of both light shading and surface quality for execution.

Images