Intelligent ink control system and method for an ink

Abstract

The present application relates to the field of ink deployment, and discloses an intelligent ink control system and method for ink, comprising a sensing component, a cloud platform, an edge gateway and a controller. The cloud platform calculates the three-dimensional chroma target data of the target color sample and the formula data, and outputs a basic formula data package containing a color Jacobian matrix. The edge gateway calculates the cumulative weight evaluation value and the fluid execution confidence using the original weight signal, controls the fine adjustment valve to stop acting when the confidence is lower than the safety threshold, and records the quality execution deviation of the current execution ink component. Then, a quadratic programming optimization model is established and solved by combining the color Jacobian matrix and the deviation, and the new target weight of the subsequent non-added component is obtained. The present application converts the single-component weight deviation caused by the kinetic fluctuation into a chroma compensation target, adjusts the proportion of the subsequent component to offset the error in the optical layer, controls the color difference of the final product, and avoids the whole batch of materials from being scrapped.

Description

Technical Field

[0001] This invention relates to the field of ink mixing, specifically to an intelligent ink mixing control system and method. Background Technology

[0002] In existing automated ink mixing processes, the control system typically uses feedback from weight sensors to set a fixed threshold to shut off the valve. During the fine-tuning dripping stage, dynamic fluctuations in fluid pressure occur within the pipeline, and current technology cannot effectively assess the stability of the fluid's execution state. This makes it difficult for the system to cope with pipeline dynamic fluctuations, easily leading to over-dripping of single-component inks, resulting in uncontrollable physical execution errors. Once the dripping of a single component goes out of control, it often directly leads to the scrapping of the entire batch of material.

[0003] Meanwhile, existing ink mixing technologies lack the computational means to convert weight errors into color space errors when a certain ink component experiences a weight deviation. The system continues to add subsequent components according to the original static formula, unable to dynamically adjust the distribution ratio of unadded components based on the weight error of the preceding components. This control method, lacking an online optical compensation mechanism, causes the physical deviations accumulated in the preceding steps to be directly reflected in the final ink product, resulting in excessive color difference indicators.

[0004] Furthermore, optical properties can drift between different batches of ink raw materials. Existing technologies typically rely on fixed formulation data for mixing, lacking a closed-loop mechanism for cross-batch data feedback. Because the system cannot use the actual colorimetric data of the final dried product to deduce and iterate the optical parameters in the original formulation, the property differences between raw material batches cannot be automatically corrected. This makes it difficult to improve the accuracy of the initial formulation calculation when the system performs subsequent ink mixing tasks. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent ink mixing control system and method, which solves the problems of color difference in the final product caused by the easy occurrence of fine-tuning dripping errors due to pipeline pressure fluctuations during the existing ink mixing process, and the lack of a dynamic compensation mechanism for color space errors after a weight deviation of a single component.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides an intelligent ink mixing control system for inks, comprising: The sensing component is used to acquire three-dimensional chromaticity target data, fluid pressure data, and raw weight signals of the target color sample; The cloud platform is used to perform calculations on the three-dimensional colorimetric target data and the formula data, and output a basic formula data package containing the initial target quality and the color Jacobian matrix; The edge gateway is used to calculate the cumulative weight evaluation value using the original weight signal, and to define a coarse-adjustment weight threshold based on the initial target mass of the corresponding ink component being executed. It is also used to calculate the fluid execution confidence level after the cumulative weight evaluation value reaches the coarse-adjustment weight threshold, and to record the mass execution deviation of the currently executed ink component when the fluid execution confidence level is lower than a preset confidence level safety threshold. Furthermore, it combines the color Jacobian matrix and the mass execution deviation to establish and solve a quadratic programming optimization model, thereby deriving a new target weight for controlling the subsequent addition of unadded ink components. The controller is used to control the main valve and the fine-tuning valve to perform ink dispensing actions, and to control the operation switching of the main valve and the fine-tuning valve based on the comparison result of the accumulated weight evaluation value and the coarse adjustment weight threshold. When the fluid execution confidence is lower than the preset confidence safety threshold, the controller controls the fine-tuning valve to stop the currently executed ink component dispensing.

[0007] This invention establishes a mathematical relationship between physical weight and color space by introducing a color Jacobian matrix. During the actual ink dispensing process, the edge gateway calculates the fluid execution confidence in real time. When the fluid state is determined to be unstable, the fine-tuning valve is shut off in advance, and the mass execution deviation at this moment is captured. Subsequently, the system uses the color Jacobian matrix to convert the aforementioned mass execution deviation into a color deviation, and uses this color deviation as a compensation target to recalculate the target weight of subsequent undispensed components using a quadratic programming optimization model. Through the above mechanism, this invention transforms the weight error of the preceding components into the weight adjustment parameter of the subsequent components, completing error cancellation at the optical color development level, thereby controlling the color difference of the final product.

[0008] Preferably, the sensing components include a colorimeter, a temperature sensor, a pressure sensor, a weighing sensor, and an ambient temperature and humidity sensor; the colorimeter is used to collect three-dimensional colorimetric target data and actual colorimetric data; the pressure sensor includes a pre-valve pressure sensor and a post-valve pressure sensor for collecting fluid pressure data; the temperature sensor is used to collect the temperature of the pipeline fluid, and the ambient temperature and humidity sensor is used to collect ambient temperature data and ambient humidity data and provide them to the edge gateway for linear trust region boundary parameter correction; the weighing sensor is used to carry the ink component to be added for weight acquisition.

[0009] Furthermore, the cloud platform is also used to construct a target chromaticity column vector based on the three-dimensional chromaticity target data, retrieve the optical absorption coefficient and scattering coefficient of each ink component in the formula data, calculate the predicted chromaticity vector based on the optical absorption coefficient and scattering coefficient, and construct an evaluation model. The fitness function used in the evaluation model is the theoretical color difference value between the target chromaticity column vector and the predicted chromaticity vector. The particle swarm optimization algorithm is used to perform multi-dimensional matrix optimization in the full color space to calculate the initial target quality of each ink component.

[0010] In a preferred embodiment of the present invention, the cloud platform is further used to sequentially introduce mass perturbation for each ink component in the basic formula data package in a simulation environment, calculate the change in three-dimensional chromaticity coordinates of the new formula in the full color space after adding the mass perturbation, and then calculate the local partial derivatives of the lightness axis index with respect to the mass of each ink component, the local partial derivatives of the red-green axis chromaticity index with respect to the mass of each ink component, and the local partial derivatives of the yellow-blue axis chromaticity index with respect to the mass of each ink component, thereby generating a color Jacobian matrix; Subsequently, the cloud platform gradually increases the mass perturbation of the corresponding ink components and compares the linear color difference prediction value calculated by the color Jacobian matrix with the nonlinear color difference real value calculated by the full color space model. When the calculation error between the linear color difference prediction value and the nonlinear color difference real value reaches the preset tolerance ratio, the cloud platform uses the corresponding mass perturbation as the linear trust region boundary of the corresponding ink component and packages the linear trust region boundary into the basic formula data package for distribution.

[0011] Preferably, the edge gateway is also used to combine the accumulated weight assessment value and the instantaneous flow assessment value to generate a two-dimensional state vector, and use a preset state transition matrix containing the sampling time interval parameter to perform linear time-series derivation on the two-dimensional state vector of the previous time to obtain the predicted state vector of the current time. Then, the observation residual is calculated based on the original weight signal obtained at the current time and the predicted state vector of the current time. The state correction amount is obtained by multiplying the observation residual with the preset Kalman gain matrix and performing a posteriori update on the predicted state vector of the current time to obtain the updated two-dimensional state vector. Furthermore, the edge gateway is equipped with a reinforcement learning agent model for generating the duty cycle of the pulse width modulation signal. The input data dimension of the reinforcement learning agent model includes the updated two-dimensional state vector and the remaining target weight obtained by subtracting the accumulated weight evaluation value from the current target weight of the currently executed ink component, which is used to output candidate duty cycle instructions. The edge gateway is also used to convert the candidate duty cycle instructions into control target instructions and issue them when the accumulated weight evaluation value reaches the coarse adjustment weight threshold. The controller is used to receive the control target instructions and output the pulse width modulation signal to start the fine adjustment valve according to the control target instructions. When the fluid execution confidence is lower than the preset confidence safety threshold, the pulse width modulation signal is cut off and the fine adjustment valve is closed.

[0012] Furthermore, the edge gateway is also used to calculate the instantaneous flow rate assessment value using the original weight signal, and to perform a subtraction operation between the pre-valve fluid pressure data and the post-valve fluid pressure data to obtain the fluid pressure difference data, as well as to obtain the preset time delay parameter; then, the edge gateway calculates the variance of the instantaneous flow rate assessment value and the variance of the fluid pressure difference data within a specified time window, and fuses the preset time delay parameter, the variance of the instantaneous flow rate assessment value, and the variance of the fluid pressure difference data to calculate the fluid execution confidence; the fluid execution confidence is calculated by subtracting the overall penalty term from 1. The overall penalty term maps the variance of the instantaneous flow rate assessment value, the preset time delay parameter, and the variance of the fluid pressure difference data respectively through a range normalization function with a limit, and then multiplies each by a preset feature fusion weight coefficient and adds them together; the feature fusion weight coefficient includes the instantaneous flow rate feature fusion weight coefficient, the delay feature fusion weight coefficient, and the pressure difference feature fusion weight coefficient, and the sum of the three is equal to 1.

[0013] In a preferred embodiment of the present invention, the edge gateway is further configured to record the stop time of triggering the hard interrupt, obtain the target weight of the currently executed ink component in the current control stage, and subtract the accumulated weight evaluation value corresponding to the stop time of triggering the hard interrupt from the target weight to calculate the quality execution deviation; then, the edge gateway retrieves the partial derivative column vector of the corresponding currently executed ink component from the color Jacobian matrix, and calculates the actual chromaticity deviation vector and the corresponding three-dimensional chromaticity compensation target in combination with the quality execution deviation. Furthermore, the edge gateway is used to establish a quadratic programming optimization model that includes a three-dimensional chromaticity compensation target, a regularized error tracking term, and physical boundary constraints. The target adjustment amount of the subsequent unadded ink components is used as the solution variable to minimize the Euclidean distance error between the optical compensation amount generated by the target adjustment amount combined with the color Jacobian matrix and the three-dimensional chromaticity compensation target.

[0014] Preferably, the edge gateway is also used to introduce physical boundary constraints, physical non-negativity constraints in the form of logical text constraints, and mixed total capacity constraints in the quadratic programming optimization model; the physical boundary constraints are that the edge gateway converts the pipeline fluid temperature into equivalent pipeline fluid temperature data, and uses the positive boundary compensation coefficient and negative boundary compensation coefficient calculated based on the equivalent pipeline fluid temperature data and the ambient humidity data to correct the linear trust region boundary parameters of the subsequent unadded ink components, and obtain the boundary upper limit correction column vector and boundary lower limit column vector of the constraint target adjustment amount; The physical non-negativity constraint in the form of logical text limits that the new target weight obtained by adding the initial target mass value of all subsequent unadded ink components to the corresponding target adjustment amount must be greater than or equal to zero; the total mixing capacity constraint limits that the total mass of the sum of the cumulative weight assessment value of all ink components already allocated to the ink mixing container and the new target weight after correction of all subsequent unadded ink components must not exceed the total capacity limit parameter determined by the physical properties of the ink mixing container as indicated on the manufacturer's nameplate. Furthermore, before solving the quadratic programming optimization model, the edge gateway is also used to compare the absolute value of the quality execution deviation with the size of the corresponding linear trust region boundary based on the linear trust region boundary and the preset chromaticity residual threshold. Within the boundary range, the quadratic programming optimization model is solved to obtain the target adjustment amount and optimal calculation residual of the subsequent unadded ink components. When it is determined that the optimal calculation residual is less than or equal to the preset chromaticity residual threshold, the corresponding ink droplet action is controlled based on the new target weight after superimposing the target adjustment amount.

[0015] Furthermore, the sensing component is also used to collect the actual colorimetric data of the final product after all ink components have been added to complete the current batch preparation and after the coating film has dried; the edge gateway is also used to generate a global log file from the final cumulative weight evaluation value and actual colorimetric data corresponding to each ink component and upload it to the cloud platform; the cloud platform uses the global log file as input, uses the backpropagation algorithm to derive the estimated optical absorption coefficient of each ink component in the current batch, multiplies the estimated optical absorption coefficient by the preset evolutionary learning rate to obtain the new knowledge correction term for the current batch, multiplies the original optical absorption coefficient corresponding to the formula data before the update by the difference ratio coefficient to obtain the memory retention term of the original optical parameters, and finally adds the new knowledge correction term and the memory retention term of the original optical parameters to obtain the latest optical absorption coefficient after model evolution update.

[0016] In a preferred embodiment of the present invention, the edge gateway is further configured to extract three-dimensional chromaticity target data and calculate batch global color difference in combination with actual chromaticity data; the edge gateway calculates the difference between the actual lightness axis index in the actual chromaticity data and the lightness axis target index in the three-dimensional chromaticity target data to obtain the lightness axis deviation, calculates the difference between the actual red-green axis chromaticity index in the actual chromaticity data and the red-green axis chromaticity target index in the three-dimensional chromaticity target data to obtain the red-green axis chromaticity deviation, and calculates the difference between the actual yellow-blue axis chromaticity index in the actual chromaticity data and the yellow-blue axis chromaticity target index in the three-dimensional chromaticity target data to obtain the yellow-blue axis chromaticity deviation; then, the square terms of the lightness axis deviation, the red-green axis chromaticity deviation, and the yellow-blue axis chromaticity deviation are added together and the square root is taken to obtain the batch global color difference.

[0017] A second aspect of the present invention provides an intelligent ink mixing control method, comprising the following steps: Acquire the three-dimensional chromaticity target data, fluid pressure data, and raw weight signal of the target color sample; The three-dimensional chromaticity target data and the formula data are processed to output a basic formula data package containing the initial target quality and color Jacobian matrix; The cumulative weight assessment value is calculated using the original weight signal, and the coarse weight threshold is defined based on the initial target mass of the corresponding ink component. The main valve and the fine valve are controlled to perform ink dispensing action, and the operation switching of the main valve and the fine valve is controlled based on the comparison result of the cumulative weight assessment value and the coarse weight threshold. After the accumulated weight assessment value reaches the coarse adjustment weight threshold, the fluid execution confidence level is calculated. When the fluid execution confidence level is lower than the preset confidence level safety threshold, the fine adjustment valve is controlled to stop the current ink component dripping. At the same time, the mass execution deviation of the current ink component is recorded. By combining the color Jacobian matrix and the quality performance deviation, a quadratic programming optimization model is established and solved to obtain a new target weight for controlling the subsequent addition of unadded ink components.

[0018] The second aspect of this invention provides a method that integrates the execution state of the underlying hardware with the upper-level formulation algorithm. By continuously calculating and interrupting the fluid execution state, mass deviation data during the single-component drop addition process is extracted. Combined with a color Jacobian matrix and a quadratic programming optimization model, the mass deviation of this component is converted into a target weight correction amount for subsequent multi-component applications. This method changes the cutting logic in ink formulation that solely relies on weight thresholds, and utilizes iterative calculations of dynamic parameters from multiple distribution ratios to maintain consistency between physical operations and the final optical color.

[0019] This invention provides an intelligent ink mixing control system and method. It has the following beneficial effects: 1. This invention calculates the fluid execution confidence level through an edge gateway, and controls the fine-tuning valve to stop operating and records the quality execution deviation when the confidence level is lower than the safety threshold. This prevents excessive dripping of single-component ink caused by pipeline dynamic fluctuations, and transforms the uncontrollable error in the physical execution process into a definite weight value, thus avoiding the scrapping of the entire batch of materials due to the loss of control of a single component. 2. This invention introduces a color Jacobian matrix to convert the quality execution deviation caused by the pre-ink components into a color space deviation, and establishes a quadratic programming optimization model to solve for the new target weight of the subsequent unadded components. This allows the weight execution error of the pre-components to be optically offset online by adjusting the distribution ratio of the subsequent components, thereby controlling the color difference index of the final product. 3. This invention collects the actual colorimetric data of the final dried product, uses the backpropagation algorithm to derive the estimated optical absorption coefficient of the current batch, and iteratively updates the basic formula data by combining it with a preset evolutionary learning rate, thus establishing a data feedback closed loop across batches. This can automatically correct the optical property drift between raw material batches and improve the accuracy of the system in calculating the initial formula in subsequent ink mixing tasks. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the intelligent ink adjustment control system architecture of the present invention; Figure 2 This is a flowchart of the ink adjustment closed-loop control method of the present invention; Figure 3 This is a flowchart of the multidimensional heterogeneous data acquisition and state estimation process of the present invention; Figure 4 This is a flowchart of the cloud-based global recipe optimization and local optical manifold mapping process of the present invention; Figure 5 This is a flowchart of the dynamic reconfiguration process for edge-end recipes based on trust region constraints according to the present invention. Figure 6 This is a flowchart of the batch-level closed-loop and model evolution for time-series decoupling in this invention; Figure 7 This is a graph showing the accumulated weight tracking during the control execution phase according to an embodiment of the present invention. Figure 8 This is a real-time evolution curve of the confidence level after feature fusion according to an embodiment of the present invention; Figure 9 This is a color difference stability verification diagram of the control group and the experimental group during a continuous production cycle according to an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions in 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.

[0022] See attached document Figure 1 This invention provides an intelligent ink adjustment control system, comprising: a cloud platform, an edge gateway, an execution terminal, and sensing components. The cloud platform achieves bidirectional data transmission with the edge gateway via a network communication interface. The edge gateway is deployed in the production environment and establishes communication connections with both the execution terminal and the sensing components to enable command issuance and data acquisition.

[0023] The sensing components are responsible for acquiring signals from the external environment and the physical state of the fluid. Specifically, they include a colorimeter, temperature sensor, pressure sensor, weighing sensor, and ambient temperature and humidity sensors located in the production area. The colorimeter acquires the spectral reflectance data and three-dimensional colorimetric target data of the target color sample, as well as the actual colorimetric data and actual spectral reflectance data of the final film sample. The temperature sensor is located on the outer wall of the ink flow pipeline to collect the fluid temperature in real time. The pressure sensors include at least a pre-valve pressure sensor located at the input of the fine-tuning valve and a post-valve pressure sensor located at the output of the fine-tuning valve, used to monitor the pressure difference across the fine-tuning valve in real time. The weighing sensor is located at the bottom of the ink mixing container to collect the raw weight signal.

[0024] The execution terminal directly performs the fluid dispensing action, specifically including a controller, a main valve, and a fine-tuning valve. The controller receives control target commands sent by the edge gateway and converts them into underlying electrical signal outputs. The main valve and the fine-tuning valve are connected in parallel to the end of the ink supply line. The controller controls the opening and closing of the main valve according to built-in logic rules and outputs pulse width modulation signals to control the opening degree of the fine-tuning valve.

[0025] The intelligent ink adjustment control system employs a four-layer decoupling mechanism at the logical level to separate the computation and execution processes. The first-level architecture decoupling deploys global formula optimization computation on the cloud platform and millisecond-level fluid control computation on the edge gateway. The second-level execution decoupling divides the ink droplet addition action into a feedforward high-flow-rate ink addition stage independently controlled by the main valve, and a low-flow-rate correction stage controlled by feedback from the fine-tuning valve. The third-level target decoupling separates the active blocking operation during the execution of a single ink component from the reconstruction and compensation operation of the un-dropped ink component in the color space. The fourth-level timing decoupling separates the real-time control loop of the edge gateway from the model parameter update loop of the cloud platform after the batch task is completed.

[0026] See attached document Figure 2 This invention provides a closed-loop control method for ink adjustment, comprising the following steps: S10. System Initialization and Status Data Acquisition. The colorimeter acquires the spectral reflectance data and three-dimensional colorimetric target data of the target color sample and transmits them to the cloud platform through the edge gateway. During the batching process, the edge gateway acquires the pipeline fluid temperature output by the temperature sensor and the fluid pressure data before and after the valve output by the pressure sensor in real time, calculates the fluid pressure difference data, and reads the raw weight signal from the weighing sensor. It uses a filtering algorithm to suppress mechanical vibration interference, calculates and outputs the cumulative weight assessment value and the instantaneous flow assessment value.

[0027] S20. Formulation Optimization and Local Approximation Matrix Generation. The cloud platform performs calculations based on the 3D chromaticity target data and internal formulation data records, outputting basic formulation data. It then performs neighborhood perturbation operations on the basic formulation data to generate a color Jacobian matrix representing the response of each ink component's quality changes to the 3D chromaticity coordinates. Subsequently, based on the effective range of the local linear approximation of the color Jacobian matrix, the cloud platform calculates the linear trust region boundary and chromaticity residual threshold for each ink component. The basic formulation data, color Jacobian matrix, linear trust region boundary, and chromaticity residual threshold are packaged into a basic formulation data package and uniformly distributed to the edge gateway, thereby decoupling the first-order architecture of global formulation optimization and underlying fluid control.

[0028] S30. Feedforward coarse adjustment control execution. Based on the second-order execution decoupling mechanism, the edge gateway uses the basic formula data as the control reference to start the ink addition process. For the currently executed ink component, when the accumulated weight assessment value does not reach the preset coarse adjustment weight threshold, the edge gateway sends a control target command to the controller. The controller converts this into a low-level electrical signal to open the main valve and keep it fully open. When the accumulated weight assessment value reaches the preset coarse adjustment weight threshold, the controller cuts off the main valve and outputs a pulse width modulation signal to start the fine-tuning valve to take over the fluid dripping work.

[0029] S40. Feature Fusion Confidence Assessment and Safety Control. After the fine-tuning valve is activated, the system enters the pulse width modulation (PWM) control stage. The edge gateway continuously calculates the variance of the instantaneous flow rate assessment value and the variance of the fluid pressure difference data within a specified time window. It obtains the time delay parameters from the output control target command to the fine-tuning valve action, and finally to the sensing and physical response of the weighing and pressure sensors. The time delay parameters, the variance of the instantaneous flow rate assessment value, and the variance of the fluid pressure difference data are fused and calculated to obtain the fluid execution confidence of the currently executed ink component. When the fluid execution confidence reaches or exceeds the preset confidence safety threshold, the controller maintains the output PWM signal to maintain the PWM dripping action of the fine-tuning valve. When the fluid execution confidence is lower than the preset confidence safety threshold, the edge gateway sends a control target command to the controller, which cuts off the PWM signal to directly close the fine-tuning valve, stops the currently executed ink component dripping, and records the quality execution deviation of the currently executed ink component. Subsequently, the control flow enters the reconfiguration step.

[0030] S50. Target Dynamic Reconstruction under Trust Region Constraints. Based on the third-order target decoupling mechanism, the edge gateway compares the absolute value of the quality execution deviation with the corresponding linear trust region boundary. If the absolute value of the quality execution deviation is within the boundary range, the edge gateway extracts the matrix parameters corresponding to the subsequent unadded ink components in the color Jacobian matrix, establishes a quadratic programming optimization model including a three-dimensional chromaticity compensation target, a regularized error tracking term, and physical boundary constraints, and then solves the quadratic programming optimization model to obtain the target adjustment amount and optimal calculation residual for the unadded ink components. If the optimal calculation residual is less than or equal to the chromaticity residual threshold, the edge gateway controls the subsequent ink component addition action according to the new target weight after superimposing the target adjustment amount.

[0031] S60. Batch Feedback Data Integration and Model Update. Based on the fourth-order temporal decoupling mechanism, after all ink components have been added to complete the current batch mixing, the prepared ink is coated onto the surface of a standard substrate according to the preset wet film thickness. After film formation and drying under preset drying temperature, preset drying time, and preset ambient humidity conditions, the colorimeter collects the actual spectral reflectance data and actual colorimetric data of the final product. The preset wet film thickness, preset drying temperature, preset drying time, and preset ambient humidity are all pre-configured by technicians based on the curing characteristics of the prepared ink's base material and industry standard testing specifications to ensure the repeatability of the colorimetric data of the film sample.

[0032] The edge gateway asynchronously uploads a global log file containing the final cumulative weight assessment of each ink component, pipeline fluid temperature, ink component interruption records, and actual spectral reflectance and chromaticity data to the cloud platform. The cloud platform reads the data from the global log file, performs convergence calculations on the internal formula data records, and updates the basic optical parameters in the internal formula data records for regenerating the partial derivative values ​​of the color Jacobian matrix in the next batch.

[0033] See attached document Figure 3 In step S10, namely the system initialization and status data acquisition step, the edge gateway and sensing components work together to convert continuous multidimensional heterogeneous signals into discrete digital signals after noise reduction. This specifically includes the following sub-steps: S11. Colorimetric Data Acquisition and Upload. The colorimeter acquires the spectral reflectance data and three-dimensional colorimetric target data of the target color sample and transmits them to the cloud platform via an edge gateway. Specifically, the colorimeter captures the reflected light from the target color sample and performs photoelectric conversion to generate spectral reflectance data within the wavelength distribution range. Then, the colorimeter's digital signal processor converts the spectral reflectance data into three-dimensional colorimetric target data based on the CIELAB color space. The three-dimensional colorimetric target data specifically includes lightness axis indices, red-green axis colorimetric indices, and yellow-blue axis colorimetric indices. The edge gateway reads the spectral reflectance data and the three-dimensional colorimetric target data and encapsulates them into an initialization task frame, which is then uploaded to the cloud platform. The photoelectric conversion circuit design and color space conversion process of the colorimeter are well-known technologies in this field.

[0034] S12. Multidimensional Environmental and Pressure Data Acquisition. The edge gateway acquires in real time the pipeline fluid temperature output by the temperature sensor, the ambient temperature and humidity data output by the ambient temperature and humidity sensors deployed in the production site, and the fluid pressure data before and after the valve output by the pressure sensor, and calculates the fluid pressure difference data. The temperature sensor uses a patch-type RTD probe, which is directly attached and fixed to the outer wall of the ink flow pipeline at the front end of the main valve and the fine-tuning valve. It acquires continuous pipeline outer wall temperature data according to a preset polling cycle. The preset polling cycle is based on the temperature sensor hardware sampling rate set to a fixed range of 10ms to 50ms, and converts it into equivalent pipeline fluid temperature data based on a preset thermal conductivity correction coefficient between the pipeline wall temperature and the ink fluid temperature. The ambient temperature and humidity sensors also send the workshop's ambient temperature and humidity data to the edge gateway according to a preset polling cycle.

[0035] Because the rheological properties of high-viscosity inks exhibit a significant environmental correlation with temperature and humidity changes, pipeline fluid temperature data, ambient temperature data, and ambient humidity data will be used together in the boundary correction and anomaly determination of the dynamic proportion calculation of the remaining ink components in subsequent steps.

[0036] In terms of fluid pressure detection, the pressure sensors include a pre-valve pressure sensor located at the input end of the fine-tuning valve and a post-valve pressure sensor located at the output end of the fine-tuning valve. The pre-valve and post-valve pressure sensors output fluid pressure data at the inlet and outlet of the valve, respectively. The edge gateway synchronously receives the pre-valve and post-valve fluid pressure data and directly performs a subtraction operation between the two to obtain the fluid differential pressure data, which characterizes the real-time fluid flow resistance inside the fine-tuning valve. By continuously monitoring the fluid differential pressure data, the edge gateway provides physical feedback indicators for fluid performance confidence calculations in subsequent safety control stages to identify abnormal fluid overflow phenomena in the fine-tuning valve.

[0037] S13. Weight Signal Filtering and State Estimation. The edge gateway reads the raw weight signal from the weighing sensor, uses a filtering algorithm to suppress mechanical vibration interference, and calculates and outputs the accumulated weight assessment value and the instantaneous flow assessment value. To eliminate mechanical vibration interference in the raw weight signal and extract the true physical accumulation, the edge gateway uses a discrete Kalman filter algorithm to dynamically track the raw weight signal over time.

[0038] The edge gateway constructs a two-dimensional state vector in memory. .in, Indicates the current time The cumulative weight assessment value, Indicates the current time The instantaneous flow rate assessment value; This indicates the transpose operation.

[0039] The edge gateway uses a state transition matrix containing a sampling time interval parameter to perform linear temporal derivation on the two-dimensional state vector from the previous time step, thus obtaining the predicted state vector for the current time step. The state prediction process satisfies the following formula: ; in, Indicates based on the previous moment The calculated current time The predicted state vector; This indicates that the matrix contains a sampling time interval parameter and the matrix elements are configured as follows: The state transition matrix, where This represents the discrete time interval corresponding to the current control clock tick; Indicates the previous moment A two-dimensional state vector containing the cumulative weight assessment value and the instantaneous flow assessment value; the state prediction process integrates the instantaneous flow assessment value of the previous moment into the cumulative weight prediction component of the current moment through the state transition matrix, and uses the instantaneous flow assessment value as an approximately constant state component in the short period to participate in the prediction.

[0040] Subsequently, the edge gateway calculates the observation residual based on the original weight signal acquired at the current sampling time and the predicted state vector at the current time, and then performs a posterior update on the predicted state vector at the current time using the Kalman gain matrix. The posterior update process satisfies the following formula: ; in, This indicates the current time after posterior update. A two-dimensional state vector containing the accumulated weight assessment value and the instantaneous flow rate assessment value; Indicates based on the previous moment The calculated current time The predicted state vector; Indicates the current time The Kalman gain matrix obtained through iterative calculation; Indicates the current time The raw weight signal actually collected by the physical output terminal of the weighing sensor; Indicates that the matrix elements are configured as The observation matrix; the posterior update process updates the observation matrix by updating it with the current time step. The predicted weight component is obtained by multiplying the predicted state vector with the original weight signal, and the observation residual is obtained by subtracting it from the original weight signal. The state correction is then obtained by multiplying the Kalman gain matrix with the observation residual, and this correction is used to adjust the current time step. The predicted state vector.

[0041] Kalman gain matrix Its function is to dynamically adjust the trust weights of the predicted state vector and the original weight signal at the current moment based on the relative magnitudes of the system process noise covariance and the sensor measurement noise covariance. The system process noise covariance is preset based on the theoretical disturbance range of the fluid system, and the sensor measurement noise covariance is determined by the factory calibration error parameters of the weighing sensor.

[0042] The edge gateway sets the control clock cycle based on the clock frequency performance of its internal microprocessor. The control clock cycle is configured as a high-frequency discrete time interval between 1ms and 5ms. The edge gateway cyclically executes control decisions and state estimations according to the control clock cycle, and asynchronously refreshes the observation data based on the hardware sampling periods of the weighing sensor, temperature sensor, pressure sensor, and ambient temperature and humidity sensor. When the weighing sensor does not generate a new original weight signal within the current control clock cycle, the edge gateway skips the matrix operation of the posterior update process within the current clock cycle and directly updates the current predicted state vector as the two-dimensional state vector after the posterior update to the system memory in real time. When the temperature sensor, pressure sensor, or ambient temperature and humidity sensor does not generate a new sample value within the current control clock cycle, the edge gateway automatically retains and uses its most recent valid sample value to participate in the current control judgment. Among them, the accumulated weight evaluation value is used independently to determine the critical point of the main valve shut-off operation, while the instantaneous flow evaluation value is used to determine the start and stop status of the fine-tuning valve and to calculate the variance of the instantaneous flow evaluation value.

[0043] To eliminate the interference of the cumulative physical weight at the bottom of the container during continuous feeding of multiple ink components on the closed-loop control of the current individual ink component, at the initial timestamp of the start of dripping for each ink component group in the basic formula data package, the edge gateway reads and records the current filtered output reading of the weighing sensor as the basic bias of the ink component. Throughout the entire dripping cycle of the current ink component, the edge gateway also uses the raw weight signal acquired in real time to... The base bias of the ink component is deducted from the total weight to ensure the original weight signal during the current execution phase. And the calculated cumulative weight evaluation value. All are converted to the net cumulative weight assessment value of the single component for the current ink composition.

[0044] See attached document Figure 4 Based on the system initialization and status data acquisition process in step S10, the cloud platform receives relevant data and performs dimensionality reduction and mapping operations to complete the basic data construction before allocation and execution. Step S20 specifically includes the following sub-steps: S21. Basic Formula Data Optimization Calculation. The cloud platform performs calculations based on the three-dimensional chromaticity target data combined with internal formula data records, outputting basic formula data. Specifically, the three-dimensional chromaticity target data constitutes a target chromaticity column vector. The cloud platform retrieves the optical absorption coefficient and scattering coefficient of each ink component from the internal formula data records. In the general principle of ink optical matching, the spectral reflectance characteristics of the mixed multi-ink components have a complex nonlinear mapping relationship with the absorption and scattering characteristics of each ink component. The cloud platform constructs an evaluation model with the theoretical color difference calculated between the target chromaticity column vector and the predicted chromaticity vector as the fitness function, and uses the particle swarm optimization algorithm to perform multi-dimensional matrix optimization in the full color space. By continuously iterating to find the quality combination that minimizes the fitness function value, the initial target quality of each ink component corresponding to the target chromaticity column vector is calculated.

[0045] The basic formulation data is formed by the initial target quality combination of each ink component. The basic formulation data includes... Each ink component, of which This represents the total number of ink components included in the basic formulation data, from the initial target mass of the first ink component in the basic formulation data to the mass of the second ink component in the basic formulation data. The initial target mass of each ink component is labeled as follows: to The construction and optimization iteration process of the fitness function based on optical coefficients and particle swarm optimization algorithm is a well-known technique in this field and will not be elaborated here.

[0046] S22. Construction of the Color Jacobian Matrix. The cloud platform performs neighborhood perturbation operations on the basic formula data to generate a color Jacobian matrix characterizing the response of each ink component's mass change to the three-dimensional chromaticity coordinates. To quantify the color shift caused by the loss of mass in a single ink component, the cloud platform sequentially introduces mass perturbation for each ink component in the basic formula data within its internal simulation environment. The operation of introducing mass perturbation aims to approximate the nonlinear color space into a locally linear space near the formula working point. The value of the mass perturbation is a fixed step value within the range of 0.1% to 1% of the initial target mass of the corresponding ink component. The value of the fixed step value is preset and extracted based on the rheological sensitivity of the corresponding ink type in historical formula data records.

[0047] The cloud platform calculates the change in three-dimensional chromaticity coordinates of the new formula in the CIELAB color space after adding mass perturbation. Based on the mass perturbation and the corresponding change in three-dimensional chromaticity coordinates, the cloud platform calculates the local partial derivatives of the lightness axis, red-green axis, and yellow-blue axis chromaticity indices with respect to the mass of each ink component, thereby constructing a dimensionless chromaticity coordinate system. The color Jacobian matrix is ​​calculated using the following formula: ; in, Represents the color Jacobian matrix; Indicates the brightness axis index; This represents the chromaticity index of the red and green axes; This represents the yellow-blue axis chromaticity index; This indicates the mass of the first ink component in the basic formulation data; This indicates the mass of the second ink component in the basic formulation data; This indicates the first in the basic formula data. The mass of each ink component; This represents the local partial derivative of the lightness axis index with respect to the quality of the first ink component; Indicates the value of the lightness axis index relative to the first Local partial derivatives of the mass of each ink component; This represents the local partial derivative of the red-green axis chromaticity index with respect to the quality of the first ink component; Indicates the red-green axis chromaticity index for the first Local partial derivatives of the mass of each ink component; This represents the local partial derivative of the yellow-blue axis chromaticity index with respect to the quality of the first ink component; Indicates the yellow-blue axis chromaticity index for the first The local partial derivative of the mass of each ink component.

[0048] S23. Linear Trust Region Boundary and Basic Recipe Data Package Generation. Based on the effective range of the local linear approximation of the color Jacobian matrix, the cloud platform calculates the linear trust region boundary and chromaticity residual threshold for each ink component. Since the color Jacobian matrix is ​​essentially a local linear dimensionality reduction approximation of the nonlinear color space, it only has guiding significance within a specific quality variation range. The cloud platform gradually increases the quality perturbation of the corresponding ink component in the simulation environment, comparing in real time the linear color difference prediction value calculated using the color Jacobian matrix with the nonlinear color difference true value calculated using the full color space model. When the calculation error between the linear color difference prediction value and the nonlinear color difference true value reaches a preset tolerance ratio, the cloud platform uses the corresponding quality perturbation as the linear trust region boundary of the corresponding ink component.

[0049] In the parameter configuration, the preset tolerance ratio is calibrated according to 5% to 10% of the actual target color difference tolerance value. The specific tolerance ratio value is bound and mapped according to the visual tolerance level of the target color sample application scenario. If the calculation error exceeds the preset tolerance ratio under the initial amplified quality perturbation, the cloud platform forcibly sets the linear trust region boundary of that ink component according to the preset minimum safety step size. The preset minimum safety step size is determined based on the physical accuracy of the single minimum pulse drop of the fine-tuning valve. The linear trust region boundary corresponding to each ink component is determined through error comparison and judgment logic. This is applied to the first ink component in the basic formula data. Each ink component has a corresponding linear trust region boundary parameter denoted as . ,in This indicates the serial number of the ink component in the basic formulation data. The value range is 1 to Positive integers.

[0050] The linear trust region boundary sets the maximum upper and lower limits of quality that can be adjusted in subsequent dynamic reconstruction steps for each ink component. Simultaneously, the cloud platform sets a batch-permissible chromaticity residual threshold based on the industry tolerance standards of the target color sample. For example, the chromaticity residual threshold is set to an evaluation standard of 1.0 in the CIELAB color difference formula, and the chromaticity residual threshold is denoted as... Subsequently, the cloud platform binary-packages the basic formulation data, including the initial target quality of each ink component, the color Jacobian matrix, the linear trust region boundary of each ink component, and the chromaticity residual threshold, into a basic formulation data package. The basic formulation data package contains the initial target quality corresponding to... parameter, Parameters, corresponding to each ink component Parameters and A composite data set of parameters. The specific calculation process of the CIELAB color difference formula is well-known in this field and will not be elaborated here.

[0051] S24. Data Packet Distribution, Parsing, and Storage. The cloud platform distributes basic recipe data packets to the edge gateway. The edge gateway receives and parses the basic recipe data packets, storing them in its local memory. This decouples the global recipe optimization from the first-order architecture of the underlying fluid control. The edge gateway retrieves the data baseline and boundary limit parameters required for closed-loop control by accessing the basic recipe data packets stored in its local memory.

[0052] The second-order execution decoupling between rapid high-flow approximation and precise low-flow compensation in the control process is mainly achieved through the combined action of feedforward coarse-tuning control execution in step S30 and feature fusion confidence assessment and safety control in step S40. Step S30 can specifically include the following sub-steps: S31. Coarse Adjustment Weight Threshold Calculation. The edge gateway extracts the initial target mass corresponding to the currently executed ink component from the basic formula data package and calculates the preset coarse adjustment weight threshold for the currently executed ink component by combining it with a preset coarse adjustment ratio coefficient. In the actual ink addition process, the edge gateway performs dropwise control of each ink component sequentially according to the arrangement order of the ink components in the basic formula data. For the currently executed ink component, the edge gateway multiplies the initial target mass corresponding to the currently executed ink component by the coarse adjustment ratio coefficient.

[0053] Generally, the higher the fluid viscosity, the lower the fluid inertia of the intelligent ink mixing control system when the dripping stops. Based on the ink component viscosity parameters preset by a physical rheometer and combined with the statistical distribution of a large amount of historical fluid dripping experimental data, the intelligent ink mixing control system sets the coarse adjustment ratio coefficient to a fixed value within the range of 85% to 95%. The logic is that the higher the viscosity parameter, the closer the coarse adjustment ratio coefficient is to 95%. The edge gateway uses this to calculate the coarse adjustment weight threshold used to distinguish the physical boundary between high-flow-rate control and low-flow-rate control.

[0054] S32. Main Valve Control and Fine-Tuning Valve Switching. The edge gateway sends a control target command to the controller to open the main valve and compares the accumulated weight assessment value with the preset coarse-adjustment weight threshold in real time. When the preset coarse-adjustment weight threshold is reached, the main valve is shut off and the fine-tuning valve is activated. In the initial stage of the ink filling process, the edge gateway sends a control target command to the controller. The controller converts the control target command into a low-level electrical signal to open the main valve and keep it fully open, utilizing the large diameter of the main valve to achieve rapid fluid filling.

[0055] The edge gateway cyclically reads the current accumulated weight assessment value according to the control clock cycle. When the accumulated weight assessment value is less than the coarse adjustment weight threshold, the main valve remains open; when the accumulated weight assessment value reaches the coarse adjustment weight threshold, the edge gateway sends a shutdown command, and the controller outputs a low-level signal to control the main valve to close and simultaneously outputs a pulse width modulation signal to start the fine-tuning valve to take over the fluid dripping work, thereby overcoming the overshoot phenomenon caused by the inertia of the fluid inside the pipeline. The fine-tuning valve uses the high-frequency opening and closing action of the internal solenoid valve core to take over the dripping task of the remaining mass, and achieves small flow feedback compensation by adjusting the duty cycle of the pulse width modulation signal.

[0056] After the main valve is shut off and the system switches to the fine-tuning valve, the system enters the feedback control process of step S40. Step S40 specifically includes the following sub-steps: S41. Physical State Parameter Extraction and Feature Calculation. After the fine-tuning valve is activated, the edge gateway extracts the physical state parameters within a specified time window, calculates the variance of the instantaneous flow rate assessment value and the variance of the fluid pressure difference data, and obtains the time delay parameters. The edge gateway allocates a sliding data queue in memory. The length of the sliding data queue is configured to correspond to a specified time window of 10 to 50 consecutive control clock cycles. The specific number of clock cycles included in the specified time window is dynamically configured by the edge gateway based on the fine-tuning valve's calibrated mechanical response frequency and the fluid rheological cycle, and the length of the specified time window covers a complete cycle of fluid fluctuation.

[0057] The edge gateway performs statistical calculations on the instantaneous flow rate assessment values ​​under multiple consecutive control clock cycles within a specified time window to obtain the variance of the instantaneous flow rate assessment values. Similarly, it performs statistical calculations on the fluid pressure difference data within a specified time window to obtain the variance of the fluid pressure difference data. The numerical changes in the variances of the instantaneous flow rate assessment values ​​and the fluid pressure difference data directly reflect non-ideal physical phenomena such as fluid interruption, stringing, or minor nozzle blockage at the ink outlet of the fine-tuning valve.

[0058] Simultaneously, the edge gateway continuously monitors the time span from outputting the target control command to the fine-tuning valve's action, and finally to the weighing and pressure sensors sensing and generating a physical response. Specifically, the edge gateway confirms that the fine-tuning valve has generated a physical response when at least one of the following response determination conditions is met:

[0059] The first response determination condition is that the rate of change of the cumulative weight assessment value output by the weighing sensor exceeds a preset weight change rate threshold. The second response determination condition is that the rate of change of the fluid differential pressure data output by the pressure sensor exceeds a preset differential pressure change rate threshold. Furthermore, each response determination condition must continuously reach a preset number of sampling points.

[0060] The weight change rate threshold is calibrated at the factory using the no-load background noise amplitude of the weighing sensor. The preset differential pressure change rate threshold is calibrated based on the extreme value of static pressure fluctuations in the fluid pipeline. The preset number of sampling points is fixed at 3 to 5 continuous control clock cycles to filter out high-frequency burst pulse interference. The edge gateway subtracts the system timestamp of the control target command from the system timestamp corresponding to the physical response to obtain the time delay parameter characterizing the hysteresis characteristics of the fluid system.

[0061] S42. Fluid Execution Confidence Assessment. The edge gateway fuses the variance of the instantaneous flow rate assessment value, the variance of the fluid pressure difference data, and the time delay parameter to derive the fluid execution confidence of the current ink component. Because the rheological properties of high-viscosity fluids during micro-droplet addition can easily cause nonlinear fluctuations, the edge gateway introduces a penalty mechanism for real-time safety assessment. The fusion calculation process satisfies the following formula: ; in, Indicates the current time The corresponding fluid execution confidence level; This represents the weighting coefficient for the fusion of instantaneous flow characteristics; This represents a range normalization function with a limit. Its operation logic is to subtract the minimum baseline value from the system's historical operation records from the input variance or physical quantity value, and then divide by the difference between the maximum and minimum values ​​and a preset minimum positive number. The larger of the two values ​​is used, and the calculation result is constrained to the standard dimensionless interval of 0 to 1 by a limiting function, where a minimum positive number is preset. Usually 10 -6 A floating-point constant of magnitude used to avoid division by zero errors when the maximum value in the system's historical running records is equal to the minimum base value; This represents the variance of the instantaneous flow rate assessment value calculated within a specified time window; Indicates the weight coefficients for delayed feature fusion; Indicates the time delay parameter; This represents the weighting coefficient for pressure difference feature fusion; This represents the variance of the fluid pressure difference data calculated within a specified time window. The fluid execution confidence is calculated by subtracting the overall penalty term from 1. The overall penalty term is obtained by mapping the variance of the instantaneous flow rate assessment value, the time delay parameter, and the variance of the fluid pressure difference data through the range normalization function, multiplying each by the corresponding feature fusion weight coefficient, and then adding them together.

[0062] The sum of the instantaneous flow rate feature fusion weighting coefficient, the delay feature fusion weighting coefficient, and the differential pressure feature fusion weighting coefficient equals 1. The instantaneous flow rate feature fusion weighting coefficient is set to a value between 0.4 and 0.5, the delay feature fusion weighting coefficient to a value between 0.2 and 0.3, and the differential pressure feature fusion weighting coefficient to a value between 0.2 and 0.3. These specific values ​​are fixedly assigned based on a preset ink component rheological sensitivity level, which is generated by the cloud platform based on historical formula execution logs. After calculating the fluid execution confidence level, the edge gateway further constrains the fluid execution confidence level to the range of 0 to 1 using a limiting function.

[0063] S43. Dynamic Dropping Control and Hard Interruption Execution. The edge gateway compares the fluid execution confidence level with a preset confidence level safety threshold, executes a sustained dropping action or triggers a hard interrupt action, and records the quality execution deviation. A reinforcement learning agent model for generating the duty cycle of the pulse width modulation signal is deployed within the edge gateway. In a specific embodiment, the reinforcement learning agent model adopts an Actor-Critic network structure based on a deep deterministic policy gradient algorithm. The input data dimension of the reinforcement learning agent model includes a two-dimensional state vector at the current moment, and the remaining target weight obtained by subtracting the accumulated weight evaluation value from the current target weight of the currently executed ink component. The Actor network includes an input layer, two hidden layers respectively configured with ReLU activation functions, and an output layer configured with a Sigmoid activation function. The output layer maps the calculation result to the 0% to 100% interval as a candidate duty cycle command for the fine-tuning valve.

[0064] In the offline phase, the reinforcement learning agent model is pre-trained using an experience replay pool containing historical fluid dripping trajectory data and state transition records. The reward function during training is designed as follows: a positive reward is given when the accumulated weight evaluation value smoothly and rapidly approaches the initial target mass; a penalty is given when dripping overshoot occurs or the variance of the instantaneous flow rate evaluation value changes abruptly. By continuously updating the internal layer weights of the Actor and Critic networks, the reinforcement learning agent model is able to output the optimal duty cycle in complex fluid environments. The underlying network operation mechanisms of the reinforcement learning agent model, such as network weight updates, loss function calculations, and gradient backpropagation, are well-known technologies in this field and will not be elaborated upon here.

[0065] During the online control phase, the reinforcement learning agent model calculates and outputs candidate duty cycle commands through forward propagation based on the real-time input two-dimensional state vector and the remaining target weight. Before issuing the candidate duty cycle commands, the edge gateway performs a security interception judgment:

[0066] When the fluid execution confidence level reaches or exceeds the preset confidence level safety threshold, the edge gateway determines that the current fluid state is in a stable range, converts the candidate duty cycle command into a control target command, and issues it. The controller maintains the output of the corresponding duty cycle pulse width modulation signal to control the fine-tuning valve to continue dripping. Simultaneously, the edge gateway compares the cumulative weight assessment value of the current ink component with the current target weight in real time. When the cumulative weight assessment value reaches the current target weight, or the difference between the cumulative weight assessment value and the current target weight is less than the minimum single pulse dripping mass of the fine-tuning valve, the edge gateway sends a stop command to the controller. The controller then cuts off the pulse width modulation signal and closes the fine-tuning valve, completing the normal dripping of the current ink component. The preset confidence level safety threshold is typically set between 0.6 and 0.8 based on the production fault tolerance rate pre-issued by the manufacturing execution system. The lower the production fault tolerance rate issued by the manufacturing execution system, the higher the confidence level safety threshold should be.

[0067] When the fluid execution confidence level is lower than the preset confidence level safety threshold, the edge gateway determines that there is a risk of fluid overflow or blockage and nonlinear runaway. It then deprives the reinforcement learning agent model of its control authority and triggers a hard interrupt action. Simultaneously, it sends a control target instruction to the controller, which cuts off the pulse width modulation signal to directly close the fine-tuning valve and forcibly stops the physical dripping process of the current ink component.

[0068] After a hard interrupt occurs, the edge gateway records the stopping time that triggered the hard interrupt and calculates the current ink component's quality execution deviation. The calculation process satisfies the following formula: ; in, This indicates the sequence number of the ink component currently being processed in the basic formulation data. The value range is 1 to Positive integers; This indicates the first in the basic formula data. The quality execution deviation generated by each ink component when a hard interruption occurs; the positive, negative, and zero values ​​of the quality execution deviation correspond to the physical states of underfilling, overfilling, and reaching the current target weight, respectively; when When it indicates a shortfall, When it indicates over-betting, This indicates that the current target weight has been reached; Indicates the first The target weight used for each ink component in the current control phase, specifically the initial target mass or the new target weight after reconstruction and updating; Indicates the stopping time The corresponding cumulative weight assessment value.

[0069] After recording the quality deviation, the edge gateway no longer forcibly opens the fine-tuning valve to make up for the missing quality of the current ink component, but directly transfers the system state to the subsequent reconstruction steps.

[0070] See attached document Figure 5 After triggering a hard interrupt and recording the quality deviation of the currently executed ink component, the intelligent ink adjustment control system enters the third-order target decoupling reconstruction compensation process. Step S50 specifically includes the following sub-steps:

[0071] S51. Trust domain boundary verification and Jacobian submatrix extraction. The absolute value of the quality execution deviation extracted by the edge gateway is compared with the first value in the basic recipe data package. The linear trust region boundary parameters corresponding to each ink component. The linear trust region boundary parameters are pre-calculated by the cloud platform based on the Taylor expansion truncation error tolerance limit of the color mapping model, or set by technical personnel with reference to historical distribution experience, and are usually 5% to 10% of the initial target quality of the corresponding ink component.

[0072] In the judgment logic, if the absolute value of the quality execution deviation is greater than the first... The linear trust region boundary parameters corresponding to each ink component indicate that the deviation of a single interruption, underfilling, or overfilling is too large, exceeding the compensation limit of the local linear model. The edge gateway directly suspends the dripping control process and triggers an alarm signal to stop subsequent actions.

[0073] If the absolute value of the quality performance deviation is less than or equal to the first The linear trust region boundary parameters corresponding to each ink component indicate that the optical offset caused by the quality execution deviation is within the effective range of local linear approximation. The intelligent ink adjustment control system meets the condition of using a linear matrix for reconstruction compensation calculation at the edge computing end.

[0074] Subsequently, the edge gateway extracts the color Jacobian matrix stored in memory and retrieves the corresponding color Jacobian matrix. The partial derivative sequence vector of each ink component is used, and combined with the quality execution deviation, to calculate the actual color deviation vector and the corresponding compensation target vector. The actual color deviation vector represents the... The optical shift caused by the deviation of the actual weight of each ink component from the current target weight is calculated using the following formula: ; The compensation target vector is the inverse vector of the actual chromaticity deviation vector, and its calculation formula is as follows: ; in, Indicates that due to the first The actual colorimetric deviation vector resulting from the quality deviation of each ink component; This represents the target compensation vector that needs to be generated by the subsequent unadded ink components; This represents the corresponding color extracted from the color Jacobian matrix. A vector of partial derivatives of each ink component, with dimensions of Column matrix; Indicates the first Quality deviation of each ink component.

[0075] Simultaneously, the edge gateway extracts the matrix columns corresponding to the subsequent unadded ink components from the color Jacobian matrix. The sequence range of the subsequent unadded ink components is as follows: to .like If the current hard interruption occurs, it indicates that the ink component is the last ink component in the basic formula data. The edge gateway determines that there are no subsequent unadded ink components available for color reconstruction compensation, and directly suspends the dripping control process, triggers an alarm signal, and records the current batch as pending manual review. The edge gateway combines the extracted matrix columns to generate a dimension of The Jacobian submatrix of the subsequent unadded ink components.

[0076] S52. Construct a quadratic programming optimization model with boundary constraints. The edge gateway establishes a quadratic programming optimization model that includes a three-dimensional chromaticity compensation objective, a regularized error tracking term, and physical boundary constraints. The established objective function aims to minimize the Euclidean distance error between the optical compensation amount caused by subsequent unadded ink component adjustments and the compensation objective vector, and simultaneously introduces a norm penalty term to prevent over-adjustment of subsequent unadded ink components. Its formula is: ; in, A mathematical operator for finding the minimum value of an objective function; Represents an unknown column vector The objective function of a quadratic programming problem with independent variables; This represents the Jacobian submatrix of the subsequent unadded ink components generated by the combination; This represents an unknown column vector consisting of the target adjustment amounts of all subsequent unadded ink components; This represents the target compensation vector that needs to be generated by the subsequent unadded ink components; This represents the pre-configured penalty regularization coefficient, which is empirically derived based on the ratio between the maximum allowable chromaticity error of the edge gateway and the flow control accuracy of the underlying fine-tuning valve, and its value ranges from 0.01 to 0.1. This represents a mathematical operation that calculates the modulus of a vector by performing a 2-norm operation. This represents the optical compensation column vector generated by local linear mapping after subsequent adjustments to the ink components not added. This represents the compensation error vector between the optical compensation column vector and the compensation target vector; This represents one-half of the L2 norm square operation of the compensation error vector, and is used as the basic error tracking term of the optimization model; This represents a penalty regularization control term that imposes a size limit on the target adjustment amount (i.e., the unknown column vector).

[0077] Simultaneously, the edge gateway sets physical boundary constraints to limit the magnitude of the target adjustment. To ensure the intelligent ink adjustment control system operates within its linear trust region while also considering rheological changes, the physical boundary constraints are transformed into inequality constraint formulas: ; in, This represents an unknown column vector consisting of the target adjustment amounts of all subsequent unadded ink components; This represents the upper bound column vector of the boundary, which is formed by sequentially arranging the linear trust region boundary parameters corresponding to the subsequent unadded ink components. This represents the positive boundary compensation coefficient dynamically calculated based on equivalent pipeline fluid temperature data and ambient humidity data, where... This represents the equivalent fluid temperature variable in the pipeline. Indicates the ambient humidity variable; This represents the negative boundary compensation coefficient dynamically calculated based on equivalent pipeline fluid temperature data and ambient humidity data. This represents the lower boundary column vector obtained by combining the negative boundary compensation coefficient; This represents the boundary upper limit correction column vector obtained by combining the positive boundary compensation coefficient.

[0078] The logic for dynamically calculating the boundary compensation coefficient of the edge gateway is as follows: It uses the equivalent pipeline fluid temperature data acquired in real time to calculate the boundary compensation coefficient. The variable value is subtracted from the factory-set standard reference temperature (usually set to 25℃) to obtain the temperature difference term. This temperature difference term is then multiplied by a weighting coefficient pre-set based on the fluid viscosity-temperature change rate measured by the rheometer to obtain the trust region correction term, which is then combined with the corresponding ambient humidity data. The empirical correction terms generated from the variable values ​​are combined to obtain the initial compensation coefficients. The edge gateway further uses a limiting function to constrain the initial compensation coefficients within a preset compensation range, for example, between 0.8 and 1.2, thereby calculating the positive boundary compensation coefficients separately. With negative boundary compensation coefficient This is to avoid physical distortions such as negative values, excessive relaxation, or excessive contraction of boundary constraints under low temperature or high humidity conditions.

[0079] To ensure the feasibility of the target adjustment at the physical execution level, the edge gateway simultaneously introduces physical non-negativity constraints and hybrid total capacity constraints in the form of logical text constraints into the quadratic programming optimization model.

[0080] The physical non-negativity constraint requires that the sum of the initial target mass values ​​of all subsequent unadded ink components and their corresponding target adjustments must be greater than or equal to zero to avoid execution dead zones where the equipment cannot reverse fluid extraction. The total mixing capacity constraint requires that the total mass, summed from the cumulative weight assessment of all ink components already dispensed into the ink mixing container and the corrected new target weights of all subsequent unadded ink components, must not exceed the maximum total capacity parameter specified on the ink mixing container's nameplate to prevent ink overflow. The corresponding formula is as follows: ; ; in, This represents the initial target mass column vector, which is composed of the initial target mass values ​​of each subsequent unadded ink component arranged sequentially. Represents an unknown column vector; Represents a zero column vector with the same dimensions as the unknown column vector; This represents the target state column vector consisting of the new target weights after correction of each subsequent unadded ink component; This represents the cumulative weight assessment value of all ink components currently dispensed into the ink mixing container; Indicates the first The initial target mass value of each subsequent ink component that was not added; Indicates the first The target adjustment amount corresponding to each subsequent ink component that was not added; This indicates the maximum total capacity parameter determined by the physical properties specified on the ink mixing container's factory nameplate. This indicates that all subsequent ink components not added are summed. This represents the total predicted compensation mass calculated by summing the new target weights after corrections for all subsequent unadded ink components.

[0081] S53. Solving the optimization problem and issuing the reconstruction target. The edge gateway uses the embedded interior-point method convex optimization solver to iteratively solve the quadratic programming optimization model constrained by inequality, and obtains the minimum value of the quadratic programming objective function. The optimal solution vector, because the objective function contains The quadratic programming optimization model is a convex optimization model because it has a 2-norm regularization term, and the physical boundary constraints, physical non-negativity constraints, and mixed total capacity constraints are all linear constraints. The edge gateway is configured with a maximum number of iterations and a maximum solution time for its embedded interior-point convex optimization solver. If a solution that meets the convergence condition is obtained within the maximum solution time, the optimal solution vector is output. If convergence is not achieved after the maximum solution time, the dripping control process is suspended and an alarm signal is triggered.

[0082] The edge gateway decomposes the optimal solution vector, separating the target adjustment amount corresponding to each subsequent unadded ink component. Then, it substitutes the optimal solution vector into the first term of the quadratic programming objective function, i.e., calculates... The magnitude of the L2 norm of this compensation error vector is used to calculate the optimal computational residual. For the specific numerical iterative solution mechanism of the interior point method, those skilled in the art can refer to the implementation of conventional convex optimization algorithm architectures, which are well-known techniques in the field and will not be elaborated upon here.

[0083] After calculating the optimal residual, the edge gateway compares the optimal residual with the chromaticity residual threshold parsed from the basic formula data package. The chromaticity residual threshold is pre-entered by technicians into the human-machine interface and sent to the basic formula data package based on the customer's color difference acceptance standard for the final product. To ensure the accuracy requirements of local linear reconstruction at the edge, the value of the chromaticity residual threshold is usually set to 0.5 to 1.0.

[0084] If the optimal calculation residual is less than or equal to the chromaticity residual threshold, the edge gateway determines that the error tolerance is met. The edge gateway adds the target adjustment amount of each subsequent unadded ink component to its corresponding initial target quality value, calculating the corrected new target weight for each subsequent unadded ink component. After completing the update of the new target weight, the edge gateway reactivates the ink adding process and calls the execution logic of steps S30 and S40. It then controls the main valve and the fine-tuning valve sequentially according to the corrected new target weights of each subsequent unadded ink component until all ink components have been distributed.

[0085] If the optimal calculated residual is greater than the chromaticity residual threshold, it indicates that the color difference caused by the quality execution deviation cannot be effectively neutralized by the subsequent unadded ink components. At this time, the edge gateway suspends the dripping control process and triggers an alarm signal, stopping subsequent actions.

[0086] By isolating the active blocking operation during the execution of a single ink component from the optical reconstruction compensation operation of the un-dropped ink component in the color space, the intelligent ink adjustment control system successfully achieves the third-order objective decoupling of correcting ink configuration errors within boundary condition constraints, reducing the scrap rate of the entire batch of materials caused by occasional fluid control failures.

[0087] See attached document Figure 6 Based on the physical mixing state after the dynamic reconstruction in step S50 is completed and steps S30 and S40 are called back until all ink components have been distributed, the intelligent ink mixing control system performs batch-level data evaluation and long-term evolution of the underlying model. Step S60 specifically includes the following sub-steps:

[0088] S61. Batch Mixing Verification and Global Color Difference Calculation. After the drop-addition control process of all ink components is completed, the external stirring mechanism physically mixes the fluid in the ink mixing container and applies the prepared ink to the surface of a standard substrate according to the preset wet film thickness. The ink is then dried under preset drying temperature, preset drying time, and preset ambient humidity conditions. Subsequently, the edge gateway sends a trigger command to the colorimeter, instructing it to perform spectral scanning and photoelectric conversion on the final film sample to obtain actual spectral reflectance data and actual chromaticity data. The actual chromaticity data specifically includes actual lightness axis indices, actual red-green axis chromaticity indices, and actual yellow-blue axis chromaticity indices.

[0089] Subsequently, the edge gateway extracts the 3D chromaticity target data obtained in step S10 and calculates the batch global color difference by combining it with the actual chromaticity data. The calculation of the batch global color difference satisfies the following formula: ; in, This represents the lightness axis target index in three-dimensional chromaticity target data; This represents the red-green axis chromaticity target index in the three-dimensional chromaticity target data; This represents the yellow-blue axis chromaticity target index in the three-dimensional chromaticity target data; This represents the actual lightness axis index in the actual chromaticity data; This represents the actual red-green axis chromaticity index in the actual chromaticity data; This represents the actual yellow-blue axis chromaticity index in the actual chromaticity data; This indicates the overall color difference of the batch after the current batch has been completed; This represents the squared term of the lightness axis deviation, which consists of the difference between the actual lightness axis index and the target lightness axis index. This represents the squared term of the red-green axis chromaticity deviation, which consists of the difference between the actual red-green axis chromaticity index and the target red-green axis chromaticity index. This represents the squared term of the yellow-blue axis chromaticity deviation, which is the difference between the actual yellow-blue axis chromaticity index and the target yellow-blue axis chromaticity index.

[0090] S62. Asynchronous Data Encapsulation and Cloud Backhaul. The edge gateway summarizes the final cumulative weight assessment value of each ink component, pipeline fluid temperature, ambient temperature and humidity data collected by ambient temperature and humidity sensors, ink component interruption records, and actual spectral reflectance and actual colorimetric data. This data, along with the calculated batch global color difference and the underlying execution logs recorded during execution, is then binary-encoded and packaged to generate a global log file. The data included in the underlying execution logs specifically includes the variance of instantaneous flow rate assessment values, the variance of fluid pressure difference data, time delay parameters, and quality execution deviation.

[0091] The edge gateway is configured with an independent communication scheduling thread, which monitors the bandwidth utilization of the underlying execution bus. When the underlying execution bus is idle, the communication scheduling thread is responsible for asynchronously uploading the global log file to the cloud platform. Asynchronous data encapsulation and cloud backhaul operations achieve fourth-order temporal decoupling between the millisecond-level high-frequency real-time control layer and the minute-level or hour-level long-term model evolution layer, ensuring the skewed allocation of underlying hardware resources at the edge computing end and avoiding network communication delays interfering with the physical disconnection response of the fluid.

[0092] S63. Long-term evolution of cloud-based basic optical parameters. The cloud platform reads data from the global log file, performs convergence calculations on the internal formula data records, and updates the basic optical parameters in the internal formula data records for regenerating the partial derivative values ​​of the color Jacobian matrix for the next batch. The specific convergence calculation process is as follows: the cloud platform receives the global log file and unpacks it. When the batch's global color difference exceeds the preset evolution trigger threshold, the system indicates that the current formula derivation mathematical model has experienced a systematic deviation due to changes in raw material batches or long-term environmental drift. The preset evolution trigger threshold is pre-entered and persistently stored by technicians in the human-machine interface of the intelligent ink adjustment control system based on the lower limit of human visual tolerance. The preset evolution trigger threshold is typically configured to range from 1.5 to 2.0.

[0093] After triggering the evolution mechanism, the cloud platform retrieves the full color space model, using the actual chromaticity data and the final cumulative weight assessment value corresponding to each ink component as input. It then uses a backpropagation algorithm to derive the estimated optical absorption coefficient of each ink component in the current batch. In one embodiment, the cloud platform uses the optical absorption coefficient as the primary evolution target. In another embodiment, the cloud platform can also use the same exponential moving average logic to synchronously update the optical scattering coefficient, or keep it unchanged when the change in the optical scattering coefficient is below a preset threshold. Subsequently, combined with the derived estimated optical absorption coefficient, the exponential moving average algorithm is used to smoothly update the original basic optical parameters in the internal formula data record. The smooth update process satisfies the following formula: ; in, This indicates the sequence number of the ink components in the basic formulation data; This indicates the first internal recipe data record before the update. The original optical absorption coefficients corresponding to each ink component; This indicates the first [number]th [item] derived from the underlying execution log of the current batch. Estimated optical absorption coefficients of each ink component; This represents the preset evolutionary learning rate, which is pre-configured in the cloud platform as a factory calibration parameter of the intelligent ink adjustment control system. The preset evolutionary learning rate is configured to range from 0.01 to 0.05. This indicates the th iteration after model evolution and update. The latest optical absorption coefficients corresponding to each ink component; This represents the product of the original optical absorption coefficient and the difference proportionality coefficient. The memory retention terms of the original optical parameters obtained; This represents the new knowledge correction term for the current batch feedback, derived by multiplying the estimated optical absorption coefficient by a preset evolutionary learning rate.

[0094] After calculating the latest optical absorption coefficients for all ink components, the cloud platform substitutes the latest optical absorption coefficients into the color mapping model to recalculate the neighborhood perturbation partial derivatives, thereby completing the convergence operation within the current formulation point domain and writing the updated basic optical parameters into the internal formulation data record. When executing steps S21 to S23 in the next batch, the cloud platform re-executes the basic formulation optimization, neighborhood perturbation operation, and color Jacobian matrix construction based on the updated basic optical parameters to generate color Jacobian matrix partial derivative values ​​that are suitable for the target color samples in the next batch.

[0095] S64. Experience replay and policy update of the reinforcement learning agent model. For the underlying intelligent control of the fine-tuning valve, the edge gateway extracts the underlying execution logs from the global log file and converts them into a time sequence containing the current moment, according to the control clock tick. Two-dimensional state vector Predict duty cycle actions, instant reward value, and the next moment. Two-dimensional state vector The edge gateway generates Markov decision tuples and appends them to the experience replay pool in memory. While the system is in a standby state between batches, the edge gateway performs the operation of randomly selecting multiple Markov decision tuples from the experience replay pool to form training batches, thereby updating the parameters of the reinforcement learning agent model deployed at the network layer.

[0096] Regarding the supplementary model structure, the Critic network in the reinforcement learning agent model is designed to include an input layer, two hidden layers configured with ReLU activation functions, and a linear output layer without an activation function. The input layer of the Critic network receives the current two-dimensional state vector and the corresponding predicted duty cycle action, and concatenates them. The output layer outputs the expected evaluation value of the corresponding state-action pair.

[0097] The edge gateway computes the temporal difference objective value for the Critic network in the reinforcement learning agent model, which guides the evolution of the value assessment system. The process of computing the temporal difference objective value satisfies the following formula: ; in, Indicates the current time The immediate reward value fed back from the underlying environment; the sign of the immediate reward value depends on the current moment. Does the calculated fluid performance confidence level meet the stability requirements? This represents the set future reward discount factor, which is statically initialized by the underlying configuration file of the intelligent ink adjustment control system, and its value ranges from 0.9 to 0.99. Indicates the next moment after the posterior update. A two-dimensional state vector containing the accumulated weight assessment value and the instantaneous flow rate assessment value; This represents the policy output function of the target Actor network in a reinforcement learning agent model; This represents the value evaluation function of the target Critic network in a reinforcement learning agent model. Indicates the current time The corresponding time difference target value; This indicates that the target Actor network updates the data based on the posterior time step. The predicted duty cycle action is output by the two-dimensional state vector operation. This indicates that the target Critic network is updated based on the posterior time step. The expected future value is calculated by evaluating the two-dimensional state vector and the predicted duty cycle action. This means multiplying the expected future value by a set future reward discount factor to obtain the discounted future value.

[0098] After calculating the time difference target value, the edge gateway uses the mean squared error function to calculate the network loss between the expected evaluation value of the current output of the Critic network and the time difference target value, and updates the weights of the Critic network using the gradient descent method. At the same time, it calculates the policy gradient with respect to the action by combining the updated Critic network, and updates the internal layer weights of the Actor network using the gradient ascent method, so as to encourage the Actor network to output action policies with higher expected evaluation value.

[0099] For the underlying computational operator implementation mechanism of network weight update and gradient backpropagation, those skilled in the art can refer to the conventional deep reinforcement learning architecture, which is a well-known technology in this field and will not be elaborated here. Through long-term evolution and strategy update, the intelligent ink adjustment control system has the closed-loop iterative capability to adapt to ink components with different viscosity characteristics.

[0100] See attached document Figure 7 To be continued Figure 9 To further verify the operating mechanism and technical effect of the intelligent ink mixing control system in a real industrial production environment, this embodiment sets the target batching task as mixing a specific batch of industrial blue phase ink. The three-dimensional chromaticity target data of the target color sample in the CIELAB color space is defined as follows: , , The maximum total capacity parameter of the ink mixing container. The target total weight for the current batch is set at 500g, with a target weight of 1000g.

[0101] The cloud platform receives 3D chromaticity target data uploaded by the edge gateway. The cloud platform then accesses its internal formula data records, using the theoretical color difference between the target chromaticity column vector and the predicted chromaticity vector as the fitness function, and employs a particle swarm optimization algorithm to calculate the basic formula data across the full color space. The calculated basic formula data includes the initial target quality corresponding to four basic ink components, in the following order: cyan component... g, magenta component g, Black component g and white base components g.

[0102] The cloud platform applies mass perturbations to the four basic ink components in a simulation environment, calculates the changes in three-dimensional chromaticity coordinates using a full color space model, and thus generates a dimensionless model. Color Jacobian Matrix By increasing the quality perturbation and comparing the predicted linear color difference with the actual nonlinear color difference, the cloud platform sets the preset tolerance ratio of the linear trust region to 7%, and calculates the boundary parameters of the linear trust region corresponding to the cyan component. It weighs 6.5g. The cloud platform will include the initial target mass of the four components and the color Jacobian matrix. The linear trust region boundary parameters for each component and the chromaticity residual threshold set to 0.8. The basic recipe data package is packaged and sent to the edge gateway.

[0103] The edge gateway parses the basic formula data packet and starts the cyan component (the currently executing ink component) according to the component number. The ink addition process is described. Based on the rheological viscosity parameters of the cyan component, the edge gateway sets the coarse adjustment ratio coefficient to 90%, and calculates the preset coarse adjustment weight threshold of the cyan component to be 108.0g.

[0104] The edge gateway sends a control target command to the controller to open the main valve, allowing a large flow of cyan ink to be injected into the ink mixing container. The edge gateway runs a discrete Kalman filter algorithm according to a 2ms control clock cycle to track the original weight signal. When the accumulated weight reaches 108.0g, the edge gateway sends a shut-off command to the controller to cut off the main valve. Simultaneously, the controller outputs a pulse width modulation signal to activate the fine-tuning valve to take over the fluid dripping operation. During this process, combined with the attached... Figure 7 It can be seen that the system calibrated the initial target mass (120g) and the preset coarse adjustment weight threshold (108g) for the corresponding level, and recorded the physical upward trajectory.

[0105] After entering the fine-tuning phase, the reinforcement learning agent model inside the edge gateway (based on a deep deterministic policy gradient algorithm) outputs candidate duty cycle commands based on the remaining target weight (obtained by subtracting the current accumulated weight evaluation value from the initial target mass of 120.0g). The fine-tuning valve executes a high-frequency pulse dripping operation. When the accumulated weight evaluation value reaches 115.0g, a physical anomaly of momentary micro-clogging of the filter screen at the end of the ink supply pipeline occurs in the simulated production environment. This physical anomaly causes irregular fluctuations in the fluid differential pressure data at the output of the fine-tuning valve, and a sharp drop in the instantaneous flow rate evaluation value.

[0106] The edge gateway calculates the variance of the instantaneous traffic assessment value within a specified time window (set to 40 consecutive control clock cycles). Significantly reduced, while the variance of fluid pressure difference data... With time delay parameter A surge. The edge gateway substitutes the above three characteristic parameters into the fluid execution confidence calculation formula: After processing with the range normalization function and multiplying by their respective feature fusion weight coefficients (set), The calculated fluid performance confidence level It dropped from a stable 0.92 to 0.48 in a relatively short period of time.

[0107] At this time, combined with the appendix Figure 8 As can be seen, because the confidence level fell below the safety threshold (0.70) indicated in the figure, the edge gateway determined that the fluid state had deviated from a steady state and immediately revoked the control authority of the reinforcement learning agent model. (See attached diagram.) Figure 7 As shown in the figure, the system stopped at 115.5g in response to the abnormally triggered hardware interrupt, and recorded the accumulated weight evaluation value at the time of stopping. The mass deviation of the cyan component is calculated to be 115.5g. g (under-injection status), which corresponds to a quality execution deviation of 4.5g in the figure.

[0108] The system transitions to the third-order target decoupling reconstruction compensation process. The edge gateway compares the absolute value of the quality execution deviation (4.5g) with the linear trust region boundary parameter (6.5g) corresponding to the cyan component, confirming that it is within the effective range of local linear approximation. The edge gateway then... Extracting the partial derivative column vector corresponding to the cyan component from the matrix Calculate the target compensation vector that needs to be borne by the subsequent unadded ink components (magenta, black, and white base materials). The edge gateway extracts the Jacobian submatrix corresponding to the ink components that were not subsequently added. (dimension is) A quadratic programming optimization model is established, which includes a three-dimensional chromaticity compensation objective, a regularized error tracking term, and physical boundary constraints.

[0109] Under the constraints of physical nonnegativity and total mixed capacity, the embedded interior-point convex optimization solver completes the iterative solution within 150ms, yielding the optimal solution vector. After decomposing the optimal solution vector into the target adjustment amounts for each subsequent ink component not added, the new target weights for the magenta component (corrected) are calculated as 15.4g, the black component as 8.1g, and the white base material as 356.2g. The calculated optimal residual is 0.62, which is less than the chromaticity residual threshold of 0.8, satisfying the reconstruction requirements.

[0110] The edge gateway sequentially controls the addition of magenta, black, and white base components according to the revised target weight. An external stirring mechanism completes the physical mixing and prepares a film sample. A colorimeter acquires the actual colorimetric data, and the edge gateway calculates the global color difference for the current batch. The value is 0.75. This color difference value meets the visual acceptance standards for industrial applications (typically requiring...). Subsequently, the edge gateway asynchronously uploads the global log file. The cloud platform utilizes an exponential moving average algorithm (with a set evolutionary learning rate). The basic optical parameters in the internal formula data record are updated based on the data of this batch, providing a corrected benchmark for the next batch of calculations.

[0111] Combining the above embodiments with the attached process Figure 9 This allows us to draw conclusions regarding the full lifecycle performance verification of the control system. Figure 9 This study demonstrates the color difference stability verification between the control and experimental groups over a continuous production cycle. A total of 50 batches of continuous ink adjustment were conducted. The control group used a traditional fixed formula and a traditional weight-tracking closed-loop control system, while the experimental group used the dynamic reconstruction and cloud-based long-term evolution of this invention. Three abnormal injection points were artificially introduced into the 12th, 28th, and 41st batches.

[0112] In the regular batch without significant disturbance, the color difference performance of the control group gradually showed an upward trend due to the slow drift caused by environmental factors, while the experimental group relied on the cloud platform to update the basic optical parameters under the fourth-order temporal decoupling mechanism, so that the underlying control benchmark of the system continuously approached the real rheological state, showing a slight downward convergence trend.

[0113] The most crucial difference lies in the system's behavior when faced with the aforementioned abnormal injection points. For example... Figure 9 As shown, the control group, lacking a low-level perception of the physical state and a target reconstruction mechanism, can only forcibly increase the control output in an attempt to approach the 120g target, which easily leads to severe overfilling after the filter breaks through, causing the batch-wide color difference it generates to directly exceed the evolution trigger threshold marked in the figure. This resulted in a batch of waste; however, the intelligent ink adjustment control system of this invention, through the coordinated linkage of the edge gateway and the cloud platform, instantly identifies underlying nonlinear risks and triggers a hard interrupt action. In subsequent execution cycles, it accurately projects the damaged chromaticity vector into the subsequent quadratic programming constraint model without ink components. Through reconstruction compensation calculation, the experimental group, even when dealing with severe pipeline anomalies, still steadily suppressed the batch global color difference within the evolution trigger threshold ( )under.

[0114] This solution transforms the originally rigid fluid flow control into an optical error approximation process based on manifold mapping and online trust region constraints. This not only frees formulation optimization from the inherent defects of unidirectional execution, but also effectively eliminates the occasional errors of single components within the system at the physical execution level. This fundamentally reduces the dependence of fluid mixing on the absolute accuracy of mechanical actuators, thereby improving product yield under varying industrial conditions.

Claims

1. An intelligent ink mixing control system for ink, characterized in that, include: The sensing component is used to acquire three-dimensional chromaticity target data, fluid pressure data, and raw weight signals of the target color sample; The cloud platform is used to perform calculations on the three-dimensional colorimetric target data and the formula data, and output a basic formula data package containing the initial target quality and color Jacobian matrix; The edge gateway is used to calculate the cumulative weight assessment value and the instantaneous flow rate assessment value using the original weight signal, define a coarse weight threshold based on the initial target mass corresponding to the currently executed ink component, and after the cumulative weight assessment value reaches the coarse weight threshold, perform a subtraction operation between the fluid pressure data before and after the valve to obtain the fluid pressure difference data, and obtain a preset time delay parameter, calculate the variance of the instantaneous flow rate assessment value and the variance of the fluid pressure difference data within a specified time window, and fuse the preset time delay parameter, the variance of the instantaneous flow rate assessment value and the variance of the fluid pressure difference data to calculate the fluid execution confidence. The fluid execution confidence is calculated by subtracting the overall penalty term from 1. The overall penalty term is obtained by mapping the variance of the instantaneous flow rate evaluation value, the preset time delay parameter, and the variance of the fluid pressure difference data through a range normalization function with a limit, multiplying each by a preset feature fusion weight coefficient, and adding them together. The feature fusion weighting coefficient includes the instantaneous flow feature fusion weighting coefficient, the delay feature fusion weighting coefficient, and the pressure difference feature fusion weighting coefficient, and the sum of the three is equal to 1; When the fluid execution confidence level is lower than the preset confidence level safety threshold, the mass execution deviation of the currently executed ink component is recorded. A quadratic programming optimization model is established and solved by combining the color Jacobian matrix and the mass execution deviation, thereby obtaining a new target weight for controlling the subsequent addition of ink components. The controller is used to control the main valve and the fine-tuning valve to perform ink dripping action, control the operation switching of the main valve and the fine-tuning valve based on the comparison result of the accumulated weight evaluation value and the coarse adjustment weight threshold, and control the fine-tuning valve to stop the currently executed ink component dripping when the fluid execution confidence is lower than the preset confidence safety threshold.

2. The intelligent ink mixing control system for ink according to claim 1, characterized in that, The sensing components include a colorimeter, a temperature sensor, a pressure sensor, a weighing sensor, and an ambient temperature and humidity sensor. The colorimeter is used to collect the three-dimensional colorimetric target data and the actual colorimetric data. The pressure sensor includes a pre-valve pressure sensor and a post-valve pressure sensor for collecting the fluid pressure data. The temperature sensor is used to collect the temperature of the fluid in the pipeline, the ambient temperature and humidity sensor is used to collect ambient temperature data and ambient humidity data and provide them to the edge gateway for linear trust domain boundary parameter correction, and the weighing sensor is used to carry the ink component to be added for weight acquisition.

3. The intelligent ink mixing control system for ink according to claim 1, characterized in that, The cloud platform is also used to construct a target chromaticity column vector based on the three-dimensional chromaticity target data, retrieve the optical absorption coefficient and scattering coefficient of each ink component in the formula data, calculate the predicted chromaticity vector based on the optical absorption coefficient and the scattering coefficient, and construct an evaluation model. The evaluation model uses the theoretical color difference between the target chromaticity column vector and the predicted chromaticity vector as its fitness function, and employs a particle swarm optimization algorithm to perform multi-dimensional matrix optimization in the full color space to calculate the initial target quality of each ink component.

4. The intelligent ink mixing control system for ink according to claim 1, characterized in that, The cloud platform is also used to sequentially introduce mass perturbation for each ink component in the basic formula data package in a simulation environment, calculate the change in three-dimensional chromaticity coordinates of the new formula in the full color space after adding the mass perturbation, and then calculate the local partial derivatives of the lightness axis index with respect to the mass of each ink component, the local partial derivatives of the red-green axis chromaticity index with respect to the mass of each ink component, and the local partial derivatives of the yellow-blue axis chromaticity index with respect to the mass of each ink component, thereby generating the color Jacobian matrix. Furthermore, the cloud platform gradually increases the mass perturbation amount of the corresponding ink components and compares the linear color difference prediction value calculated by the color Jacobian matrix with the nonlinear color difference real value calculated by the full color space model. When the calculation error between the linear color difference prediction value and the nonlinear color difference true value reaches a preset tolerance ratio, the cloud platform uses the corresponding quality perturbation amount as the linear trust region boundary of the corresponding ink component, and packages the linear trust region boundary into the basic formula data package for distribution.

5. The intelligent ink mixing control system for ink according to claim 2, characterized in that, The edge gateway is also used to combine the accumulated weight assessment value and the instantaneous traffic assessment value to generate a two-dimensional state vector, and use a preset state transition matrix containing a sampling time interval parameter to perform linear time-series derivation on the two-dimensional state vector of the previous moment to obtain the predicted state vector of the current moment. Then, it calculates the observation residual based on the original weight signal obtained at the current moment and the predicted state vector of the current moment, and then combines the preset Kalman gain matrix with the observation residual to obtain the state correction amount to perform a posteriori update on the predicted state vector of the current moment to obtain the updated two-dimensional state vector. Furthermore, the edge gateway is equipped with a reinforcement learning agent model for generating the duty cycle of the pulse width modulation signal. The input data dimension of the reinforcement learning agent model includes the updated two-dimensional state vector and the remaining target weight obtained by subtracting the accumulated weight evaluation value from the current target weight of the currently executed ink component, which is used to output candidate duty cycle instructions. The edge gateway is also used to convert the candidate duty cycle instruction into a control target instruction and issue it when the accumulated weight evaluation value reaches the coarse weight threshold. The controller is used to receive the control target instruction, output a pulse width modulation signal to start the fine-tuning valve according to the control target instruction, and cut off the pulse width modulation signal to close the fine-tuning valve when the fluid execution confidence is lower than the preset confidence safety threshold.

6. The intelligent ink mixing control system for ink according to claim 1, characterized in that, The edge gateway is also used to record the stop time of triggering a hard interrupt, obtain the target weight of the currently executed ink component in the current control stage, and subtract the accumulated weight evaluation value corresponding to the stop time of triggering the hard interrupt from the target weight to calculate the quality execution deviation. Then, the edge gateway retrieves the partial derivative sequence vector of the ink component currently being executed from the color Jacobian matrix, and calculates the actual chromaticity deviation vector and the corresponding three-dimensional chromaticity compensation target by combining the quality execution deviation amount. Furthermore, the edge gateway is used to establish the quadratic programming optimization model that includes the three-dimensional chromaticity compensation target, regularized error tracking term and physical boundary constraints, and uses the target adjustment amount of the subsequent unadded ink components as the solution variable to minimize the Euclidean distance error between the optical compensation amount generated by the target adjustment amount combined with the color Jacobian matrix and the three-dimensional chromaticity compensation target. The physical boundary constraint is that the edge gateway converts the pipeline fluid temperature into equivalent pipeline fluid temperature data, and uses the positive boundary compensation coefficient and negative boundary compensation coefficient calculated based on the equivalent pipeline fluid temperature data and the ambient humidity data to correct the linear trust region boundary parameters of the subsequent unadded ink components, thereby obtaining the upper boundary correction column vector and the lower boundary column vector that constrain the size of the target adjustment amount.

7. The intelligent ink mixing control system for ink according to claim 6, characterized in that, The edge gateway is also used to introduce the physical boundary constraints, the physical non-negativity constraints in the form of logical text constraints, and the hybrid total capacity constraints into the quadratic programming optimization model. The physical non-negativity constraint of the logical text constraint form restricts the new target weight obtained by adding the initial target mass value of all subsequent unadded ink components to the corresponding target adjustment amount to be greater than or equal to zero; The total mass of the total mixing capacity constraint limit, which is the sum of the cumulative weight assessment of all ink components assigned to the ink mixing container and the new target weight after correction of all subsequent unadded ink components, shall not exceed the total capacity limit parameter determined by the physical properties specified on the ink mixing container's nameplate. Furthermore, the edge gateway is also used to, before solving the quadratic programming optimization model, compare the absolute value of the quality execution deviation with the size of the corresponding linear trust region boundary based on the linear trust region boundary and a preset chromaticity residual threshold, solve the quadratic programming optimization model within the boundary range to obtain the target adjustment amount and optimal calculation residual of the subsequently unadded ink component, and when it is determined that the optimal calculation residual is less than or equal to the preset chromaticity residual threshold, control the corresponding ink droplet action based on the new target weight after superimposing the target adjustment amount.

8. The intelligent ink mixing control system for ink according to claim 2, characterized in that, The sensing component is also used to collect the actual colorimetric data of the final product after all ink components have been added to complete the current batch preparation and after the coating film has dried. The edge gateway is also used to generate a global log file from the final cumulative weight evaluation value corresponding to each ink component and the actual colorimetric data, and upload it to the cloud platform; The cloud platform takes the global log file as input, uses the backpropagation algorithm to derive the estimated optical absorption coefficient of each ink component in the current batch, multiplies the estimated optical absorption coefficient by a preset evolutionary learning rate to obtain the new knowledge correction term for the current batch, multiplies the original optical absorption coefficient corresponding to the formula data before the update by the difference ratio coefficient to obtain the memory retention term of the original optical parameters, and finally adds the new knowledge correction term and the memory retention term of the original optical parameters to obtain the latest optical absorption coefficient after model evolution update.

9. A method for intelligent ink mixing control, characterized in that, An intelligent ink mixing control system for an ink as described in any one of claims 1-8 includes the following steps: Acquire the three-dimensional chromaticity target data, fluid pressure data, and raw weight signal of the target color sample; The three-dimensional chromaticity target data and the formula data are processed to output a basic formula data package containing the initial target quality and the color Jacobian matrix; The cumulative weight assessment value and instantaneous flow rate assessment value are calculated using the original weight signal, and a coarse weight threshold is defined based on the initial target mass of the ink component currently being executed; the main valve and the fine-tuning valve are controlled to perform ink dripping action, and the operation switching of the main valve and the fine-tuning valve is controlled based on the comparison result of the cumulative weight assessment value and the coarse weight threshold. After the accumulated weight assessment value reaches the coarse adjustment weight threshold, the fluid pressure data before the valve and the fluid pressure data after the valve are subtracted to obtain the fluid pressure difference data, and the time delay parameter is obtained. The variance of the instantaneous flow assessment value and the variance of the fluid pressure difference data within the specified time window are calculated, and the time delay parameter, the variance of the instantaneous flow assessment value and the variance of the fluid pressure difference data are fused together to obtain the fluid execution confidence. The fluid execution confidence level is calculated by subtracting the overall penalty term from 1. The overall penalty term is obtained by mapping the variance of the instantaneous flow rate assessment value, the time delay parameter, and the variance of the fluid pressure difference data through a range normalization function with a limit, multiplying each by a preset feature fusion weight coefficient, and then adding them together. The feature fusion weight coefficient includes the instantaneous flow rate feature fusion weight coefficient, the time delay feature fusion weight coefficient, and the pressure difference feature fusion weight coefficient, and the sum of the three is equal to 1. And when the confidence level of the fluid execution is lower than the preset confidence level safety threshold, the fine-tuning valve is controlled to stop the current ink component dripping, and the mass execution deviation of the current ink component is recorded. By combining the color Jacobian matrix and the quality execution deviation, a quadratic programming optimization model is established and solved to obtain a new target weight for controlling the subsequent addition of ink components.

Images