Ink path control system of full-color printer

Abstract

The invention relates to the technical field of industrial control systems, in particular to an ink path control system of a full-color printer, which comprises the following specific implementation steps of: measuring a plurality of physical quantities in a controlled physical process in real time, and outputting the physical quantities as state variables and external disturbance signals; receiving a state variable and an external disturbance signal, performing future state prediction through a nonlinear dynamic process model, and generating a prediction result; constructing a process performance index function, optimizing the performance index function through a sparrow search optimization algorithm based on a prediction result, and generating an optimal target set value; through a model based on Bayesian reasoning, parameters of a PID control law are updated in a self-adaptive mode according to external disturbance signals; the controller is used for receiving the optimal target set value and the adaptive PID control law parameters, executing the PID control law in combination with a preset undisturbed switching module, and generating a final control instruction; and receiving a final control instruction, and converting the final control instruction into driving signals acting on a plurality of physical actuators so as to realize closed-loop feedback control of the controlled physical process.

Description

Technical Field

[0001] This invention relates to the field of industrial control system technology, specifically to an ink path control system for a full-color printer. Background Technology

[0002] Existing control technologies for complex industrial processes are quite mature. At the control strategy level, by establishing a model of the controlled object and combining it with classic feedback control algorithms such as PID, adaptive adjustment of key process parameters has been achieved. In terms of execution structure, integrated drive mechanisms improve the simplicity and reliability of the system. At the underlying drive level, pulse width modulation technology enables fine-grained control of the final execution link. In addition, mature online status maintenance technology ensures the stability of continuous operation of the controlled object.

[0003] When dealing with dynamic industrial environments, the inherent limitations of existing control systems are also very prominent. Their core control models are mostly based on static assumptions, and their adaptive capabilities are insufficient when faced with unmodeled operating conditions or parameter drift. The response speed of the physical execution links is difficult to match the cycle time requirements of high-speed processes, which can easily lead to transient errors in the system. At the same time, existing control architectures generally ignore the coupling effects of multiple physical fields such as thermodynamics, electricity and fluid dynamics. Any external disturbance may cause fluctuations in the physical properties of the controlled medium and process instability, ultimately affecting the quality of the system output.

[0004] In summary, the static and decoupled control strategies adopted by existing control systems are incompatible with the inherent dynamic and strongly coupled physical characteristics of the controlled object. This gap between control theory and physical reality limits the improvement of system performance. Therefore, this invention proposes an ink path control system for a full-color printer. Summary of the Invention

[0005] The purpose of this invention is to provide an ink path control system for a full-color printer, achieving precise, stable, and globally optimal collaborative control of the controlled physical process under complex dynamic conditions. Specifically, this is achieved by constructing a hierarchical control architecture that includes an upper-level intelligent optimizer and a lower-level feedback controller, and deeply integrating various intelligent algorithms such as state prediction based on deep learning, multi-objective optimization, and Bayesian parameter self-tuning.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An ink path control system for a full-color printer, comprising:

[0008] The state observation unit measures multiple physical quantities in the controlled physical process in real time and outputs them as state variables and external disturbance signals.

[0009] Hierarchical digital controllers, including:

[0010] The upper-level intelligent optimizer receives the state variables and external disturbance signals, predicts the future state through a nonlinear dynamic process model, and generates prediction results; it constructs a process performance index function, optimizes the performance index function through a sparrow search optimization algorithm based on the prediction results, and generates the optimal target setpoint; and it adaptively updates the parameters of the PID control law according to the external disturbance signals through a parameter update model based on Bayesian inference.

[0011] The lower-level feedback controller receives the optimal target setpoint and PID control law parameters, and, in conjunction with the preset bumpless switching module, executes the PID control law to generate the final control command.

[0012] The drive execution unit receives the final control command and converts it into drive signals that act on multiple physical actuators.

[0013] Preferably, the state observation unit includes: a group of sensors distributed along the physical path of the controlled physical process to measure real-time dynamic pressure, temperature and flow rate as variables of the controlled physical process; an environmental sensor to measure the external environmental conditions of the equipment as external disturbances; and a soft measurement module that uses a pre-trained model based on physical information to estimate a process medium state parameter in real time and output it as a state variable.

[0014] Preferably, the upper-layer intelligent optimizer receives the state variables and external disturbance signals, predicts the future state through a nonlinear dynamic process model, and generates a prediction result, including: the upper-layer intelligent optimizer is based on a nonlinear dynamic process model of a spatiotemporal separation Transformer architecture, which abstracts the controlled physical process into a graph structure, where each sensor node is defined as a vertex of the graph, the physical connection path connecting the vertices is defined as an edge of the graph, and the attribute of the edge is a dynamic flow resistance parameter calculated from the real-time state variables; a spatial encoder, composed of a graph neural network, aggregates information between adjacent vertices through a message passing mechanism, learns the physical dependency relationship between each node and its neighboring nodes as a spatial distribution feature; a temporal encoder, composed of a recurrent neural network, extracts temporal dependency features from the historical state variable sequence; and a fusion layer, composed of a multi-head self-attention mechanism, adaptively weights and fuses the spatial distribution features and temporal dependency features, and outputs a prediction of the future state of the controlled physical process.

[0015] Preferably, a process performance index function is constructed. Based on the prediction results, the performance index function is optimized using a sparrow search optimization algorithm to generate the optimal target setpoint. This includes: constructing a process performance index function composed of three weighted sub-objectives: a tracking error representing a quantitative indicator of material conveying volume, a positioning error representing a spatial accuracy indicator of the material's action point, and the total energy consumption required to control the actuator; in the offline initialization phase before the start of the operation cycle, a sparrow search optimization algorithm is used to achieve a global search by introducing Lévy flight random perturbations; in the offline phase, the process performance index function is iteratively solved using an improved sparrow search optimization algorithm to generate a control strategy lookup table, which stores a solution set consisting of a large number of Pareto optimal solutions corresponding to different operating conditions; in the online real-time control phase of the operation cycle, the optimal target setpoint is obtained by quickly retrieving and interpolating from the control strategy lookup table based on the current working mode and the real-time state variables obtained from the state observation unit.

[0016] Preferably, the Bayesian inference parameter update model includes: treating the three PID parameters of the lower-level feedback controller as a parameter vector to be estimated; in the offline training phase, running a complete Bayesian inference process to train a parameter update model, the Bayesian inference process including: constructing a prior probability distribution of the parameter vector in the form of a Gaussian mixture model; and calculating the likelihood probability and posterior probability distribution based on the external environmental disturbance signal and the corresponding optimal PID control law parameters in historical data; the parameter update model is a lightweight neural network that represents the mapping relationship from the external environmental disturbance signal to the optimal PID control law parameter vector; in the online real-time control phase, the real-time acquired external environmental disturbance signal is input into the parameter update model, and through one forward propagation calculation, the estimated value of the optimal PID control law parameter is directly obtained as the adaptive PID control law parameter under the current operating condition.

[0017] Preferably, the lower-level feedback controller includes: receiving the optimal target setpoint from the upper-level intelligent optimizer as the setpoint of the lower-level feedback controller control loop; receiving adaptive PID control law parameters from the upper-level intelligent optimizer and updating the internal PID control law in real time, the update process being executed by a bumpless switching module, the bumpless switching module achieving smooth transition of PID control law parameters through a combination of reverse calculation and setpoint filtering; using the real-time state variable output by the state observation unit as the process variable, calculating the error between the process variable and the setpoint, and performing proportional, integral, and derivative operations according to the adaptive PID control law parameters to generate the final control command.

[0018] Preferably, the hierarchical digital controller further includes a fault-tolerant control module, which determines that the control system has entered an abnormal operating condition when any of the following conditions are met:

[0019] The posterior probability distribution confidence of the parameter update model output is consistently below a preset threshold; a residual-based anomaly detector detects that the deviation between the actual measured value of the state observation unit and the predicted value of the nonlinear dynamic process model consistently exceeds a preset range.

[0020] Upon determining that an abnormal operating condition has been entered, the control system automatically suspends parameter adaptive updates, switches to conservative PID control law parameters, and sends an alarm message to a human-machine interface. The conservative PID control law parameters are a set of fixed parameters with low gain and long integral time characteristics that are pre-stored in the controller's memory.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. This invention utilizes a nonlinear dynamic process model based on a spatiotemporally separated Transformer architecture in the upper-level intelligent optimizer. This model autonomously learns the complex spatiotemporal coupling relationship between pressure, temperature, flow rate, and ink viscosity within the ink path system from multi-dimensional, multi-channel sensor data. This overcomes the fundamental deficiency of existing technologies that rely on static, decoupled models, enabling the system to accurately predict the ink path state under high-speed jetting or environmental changes. Furthermore, through a parameter update model based on Bayesian inference, the system can dynamically and online adjust the parameters of the lower-level PID controller according to real-time environmental disturbances, achieving real-time adaptation to external changes. This fundamentally solves the problem of drastic performance degradation of existing fixed-parameter controllers when facing operating condition drift, significantly improving control accuracy and environmental adaptability.

[0023] 2. This invention uses an improved sparrow search optimization algorithm in the upper-level intelligent optimizer to solve a process performance index function that includes droplet volume, positioning accuracy, and energy consumption. The algorithm enhances global search capability by introducing the Lévy flight mechanism and outputs a Pareto optimal solution set. It can find an optimal balance point between conflicting performance objectives, overcoming the shortcomings of existing technologies that can only optimize for a single objective, resulting in the overall system performance not reaching the optimal level. This allows the system to operate in a globally optimal state under different working modes.

[0024] 3. This invention, through its hierarchical digital controller architecture, decouples complex intelligent computing from high-speed feedback execution, ensuring the real-time performance and determinism of the underlying control loop. Furthermore, by integrating anti-integral saturation and disturbance-free switching functions into the lower-level feedback controller, it effectively suppresses system overshoot and oscillation when the actuator reaches its physical limits or control parameters are updated. Through a fault-tolerant control module, it can automatically switch to a safe, conservative control strategy when the model is uncertain about the current operating conditions, overcoming the risk of system collapse that may occur in existing intelligent control methods when model mismatch or sensor anomalies. This makes the stability and robustness of the system in this invention far exceed existing technologies. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the ink path control system architecture of a full-color printer according to an embodiment of the present invention;

[0026] Figure 2 This is a flowchart of the upper-layer intelligent optimizer according to an embodiment of the present invention;

[0027] Figure 3 This is a flowchart of the lower-level feedback controller according to an embodiment of the present invention. Detailed Implementation

[0028] Next, the technical solution of the present invention will be explained in detail and clearly with reference to specific examples and related drawings. It should be understood that the examples described below represent only a part of the present invention, not all of it. Based on these examples, those skilled in the art can deduce other possible embodiments of the present invention without creative effort, and these embodiments are also protected by the present invention.

[0029] See Figures 1 to 3 This invention provides an ink path control system for a full-color printer, the technical solution of which is as follows:

[0030] Example 1:

[0031] Reference Figure 1 An ink path control system for a full-color printer includes:

[0032] The state observation unit measures multiple physical quantities in the controlled physical process in real time and outputs them as state variables and external disturbance signals.

[0033] Hierarchical digital controllers, including:

[0034] The upper-level intelligent optimizer receives the state variables and external disturbance signals, predicts the future state through a nonlinear dynamic process model, and generates prediction results; it constructs a process performance index function, optimizes the performance index function through a sparrow search optimization algorithm based on the prediction results, and generates the optimal target setpoint; and it adaptively updates the parameters of the PID control law according to the external disturbance signals through a parameter update model based on Bayesian inference.

[0035] The lower-level feedback controller receives the optimal target setpoint and PID control law parameters, and, in conjunction with the preset bumpless switching module, executes the PID control law to generate the final control command.

[0036] The drive execution unit receives the final control command and converts it into drive signals that act on multiple physical actuators.

[0037] Furthermore, the state observation unit includes:

[0038] Multiple sensor arrays distributed along the physical path of the controlled physical process measure real-time dynamic pressure, temperature, and flow rate as variables of the controlled physical process;

[0039] An environmental sensor measures external environmental conditions, such as temperature and humidity, as external disturbances to the device;

[0040] And a soft measurement module, which is a software functional module embedded in the controller, utilizes a pre-trained model based on physical information. The pre-trained model learns the physical properties of the controlled medium (such as ink) offline (such as the relationship between viscosity and temperature and pressure). When running online, the soft measurement module receives partial measurement values ​​from the physical sensor group, calculates through the model, and estimates in real time a key process medium state parameter (such as dynamic viscosity) that is difficult to measure directly but is crucial to control, and outputs this parameter as part of the state variable.

[0041] The pre-trained model based on physical information is a fully connected neural network containing an input layer, two hidden layers, and an output layer. The number of nodes in the input layer corresponds to the number of physical sensors selected; for example, if three temperature sensors and two pressure sensors are selected, the input layer has five nodes. The two hidden layers contain sixty-four and thirty-two neurons respectively, and use exponential linear units (ELUs) as non-linear activation functions. The output layer has one node, outputting a scalar value representing dynamic viscosity in centipoise (cP).

[0042] The training data for the model includes calibration data for inks with multiple formulations in different temperature and pressure ranges. For example, it includes 1,200 data points for 20 formulations in a temperature range of 15 to 50 degrees Celsius and a pressure range of 0.1 MPa to 0.5 MPa. The training process uses the Adam optimizer with a learning rate set to 0.1%.

[0043] The state observation unit also includes an anomaly detection module that integrates multimodal sensor fusion, which integrates acoustic and image sensors.

[0044] The acoustic sensor is used to capture the acoustic signal spectrum generated in real time as ink flows in the pipe; the image sensor is used to monitor the morphological changes on the outside of the ink pipe.

[0045] The anomaly detection module also includes a multimodal deep fusion model, which is composed of a convolutional neural network and a recurrent neural network in parallel. It receives acoustic signal spectrum and image data from an image sensor as input; it extracts visual spatial features through the convolutional neural network and acoustic temporal features through the recurrent neural network; and it integrates the multimodal features through a fusion layer to output a real-time probability value that characterizes the physical anomaly of the ink system.

[0046] The anomaly detection module, which integrates data from multiple sensors, improves the accuracy and robustness of anomaly detection, especially in printing tasks under high load or complex environments.

[0047] This application embodiment, through a combination of "hard measurement + soft estimation" and virtual and real methods, can comprehensively and accurately grasp every subtle state change inside the controlled process, providing high-dimensional real-time data input for upper-level intelligent prediction and optimization, and solving the problem of insufficient control accuracy caused by incomplete state perception in existing methods.

[0048] Furthermore, the upper-level intelligent optimizer receives the state variables and external disturbance signals, predicts future states through a nonlinear dynamic process model, and generates prediction results, including:

[0049] The upper-level intelligent optimizer, based on a nonlinear dynamic process model of a spatiotemporally separated Transformer architecture, abstracts the controlled physical process (such as the ink system) and its sensor network into a graph structure containing multiple vertices and edges. Each vertex represents a sensor node, and its attribute is the real-time reading of the sensor. Each edge represents a physical path connecting two nodes, and its attribute is a dynamic flow resistance parameter. The dynamic flow resistance parameter is calculated in each control cycle according to the Darcy-Weisbach equation. Specifically, its value is directly proportional to the fluid density and pipe length, inversely proportional to the square of the pipe diameter and cross-sectional area, and is affected by a key friction coefficient. The friction coefficient itself is obtained through an iteratively solved functional relationship (such as the Colebrook-White formula) related to pipe wall roughness and Reynolds number.

[0050] A spatial encoder, composed of a graph neural network, aggregates information between adjacent vertices of a graph through its unique message passing mechanism, learning the dependencies and crosstalk effects between nodes caused by physical connections as spatial distribution features.

[0051] The message passing mechanism adopts the GraphSAGE framework. The message function concatenates the state vectors of two adjacent nodes with the attributes of the edges between them (e.g., pipe diameter and length), and performs a linear transformation through a weight matrix with a non-linear activation function (such as ReLU) to generate a message. The aggregation function performs max pooling on all messages received by a node from its neighboring nodes, concatenates the aggregated message with the node's current state vector, and then updates the node's state vector through another weight matrix and activation function.

[0052] A time encoder, composed of a recurrent neural network, analyzes the trend of the evolution of each state variable over time and extracts time-dependent features from the historical state variable sequence;

[0053] And a fusion layer consisting of a multi-head self-attention mechanism, which simultaneously receives spatial distribution features output by the spatial encoder and temporal dependency features output by the temporal encoder, adaptively weights and fuses the spatial distribution features and temporal dependency features, and outputs a prediction of the future state of the controlled physical process.

[0054] This application adopts a nonlinear dynamic process model based on a spatiotemporal separation Transformer architecture. By using a graph neural network to model the controlled process in a graph structure, it deeply understands and quantifies the crosstalk effect between multiple loops. This prediction based on global physical dependencies enables the controller to upgrade from passive adjustment to active "predictive" collaborative control, solving the problem of system performance degradation caused by inter-channel coupling.

[0055] Furthermore, a process performance index function is constructed. Based on the prediction results, the performance index function is optimized using a sparrow search optimization algorithm to generate the optimal target setting value, with reference to... Figure 2 ,include:

[0056] In the software architecture of the upper-level intelligent optimizer, a process performance index function for evaluating the control effect is predefined. The process performance index function is composed of three sub-objective functions combined by a linear weighting method. The first sub-objective function is used to calculate the root mean square error between the actual value and the expected value of the quantitative index representing the material conveying volume. The second sub-objective function is used to calculate the Euclidean distance between the actual coordinates and the expected coordinates of the spatial accuracy index representing the material action point. The third sub-objective function is used to calculate the total power integral of all actuators in the drive execution unit per unit time. Each of the three sub-objective functions has an adjustable weight coefficient. The adjustable range of the weight coefficient representing the material conveying volume tracking error is set between 0.4 and 0.8, and the sum of all weight coefficients is 1.

[0057] Depending on the working mode, these weighting coefficients are assigned different default values. For example, in "high-quality mode", the weighting is as follows: the weight of droplet volume tracking error is 0.6, the weight of positioning accuracy error is 0.3, and the weight of total energy consumption is 0.1, in order to prioritize printing quality. In "economy mode", the weighting is adjusted to: the weight of total energy consumption is 0.6, and the weights of volume error and positioning error are each 0.2.

[0058] The process performance index function is solved in the offline initialization phase before the start of an operation cycle, such as during device power-on self-test or standby. The controller calls an offline optimization program to execute this program, which performs an improved Sparrow Search Optimization Algorithm (SSA). In this algorithm population, individuals designated as Producers have a random step size that follows a Lévy distribution when updating their positions, enabling them to make long-distance jumps with a small probability and perform a global search.

[0059] The step size coefficient beta of the Lévy distribution is set to 1.5, the size of the Pareto optimal solution set generated by the iterative solution process is fixed at 500 sets, and the final control strategy lookup table is indexed by a KD tree. The index dimension of the KD tree is six-dimensional, corresponding to an encoding of the working mode and five real-time state variables: pressure, temperature, flow rate, dynamic viscosity, and ambient humidity.

[0060] The offline optimization program takes a preset simulation environment model covering all possible operating conditions as input, iteratively solves the aforementioned process performance index function, calculates and outputs a set of solutions constituting the Pareto front. The iterative solution process continues to run until a preset number of iterations or a convergence threshold is reached. The maximum number of iterations is set to one thousand, and the convergence threshold is defined as follows: when the improvement rate of the Pareto front obtained from 20 consecutive iterations (i.e., the proportion of newly generated non-dominated solutions to the total number of solutions) is less than one-thousandth, the algorithm is judged to have converged and the iteration is terminated. Finally, all the calculated Pareto optimal solutions, together with the corresponding operating condition inputs and performance weights, are organized into a multi-dimensional control strategy lookup table with a KD-tree index structure and stored in the controller's non-volatile memory.

[0061] When entering the online real-time control phase of the operation cycle, the upper-level intelligent optimizer enters a high-speed call mode. Within each control cycle, the upper-level intelligent optimizer first combines the current operating mode and the real-time state variables obtained from the state observation unit into a query vector. Then, using the KD-tree index, it performs a high-speed nearest neighbor or range query in the control strategy lookup table generated in the previous step to find several pre-stored optimal strategy points that are closest to the current query vector. If the current query vector happens to fall between two or more pre-stored strategy points, a fast multilinear interpolation algorithm is started to calculate an accurate optimal target setpoint that is most suitable for the current specific instantaneous operating condition based on the distance between the current query vector and the neighboring strategy points. Finally, the optimal target setpoint calculated by interpolation is output to the lower-level feedback controller.

[0062] The upper-layer intelligent optimizer also employs a joint optimization mechanism based on causal representation decoupling, including:

[0063] A shared representation encoder is used to capture common features that are useful for both future state prediction tasks and optimal goal decision tasks; a prediction task-specific encoder is used to extract task-specific information that is only relevant to future state prediction; and a decision task-specific encoder is used to extract task-specific information that is only relevant to optimal goal decision.

[0064] During the model training phase, an orthogonality loss function is introduced. This loss function forces the separation of spurious associations between tasks by minimizing the vector cosine similarity between the output of the prediction task-specific encoder and the output of the decision task-specific encoder.

[0065] The joint optimization mechanism based on causal representation decoupling explicitly separates shared information from task-specific information, cutting off spurious correlation paths between prediction and decision-making. This enables the optimizer to learn more fundamental causal relationships, thereby improving the robustness of prediction and decision-making.

[0066] This application adopts a two-stage strategy of "offline long-term global optimization and online millisecond-level fast invocation". In the offline stage, an improved sparrow search optimization algorithm is used to solve the multi-objective performance index function and generate a multi-dimensional control strategy lookup table. In the online stage, the lookup table is searched and interpolated at high speed according to the real-time operating conditions. This realizes the successful application of advanced global optimization algorithms in real-time control systems, solves the fundamental contradiction that complex optimization algorithms cannot meet real-time requirements due to huge computational load, and enables the system to achieve the best balance among multiple conflicting performance objectives.

[0067] Furthermore, the parameter update model based on Bayesian inference includes:

[0068] In the offline training phase, a parameter update model is trained. The training process first treats the three PID control law parameters (proportional gain Kp, integral gain Ki, and derivative gain Kd) of the lower-level feedback controller as a three-dimensional parameter vector to be estimated.

[0069] Using a large amount of historical operating data collected under different operating conditions, a complete Bayesian inference process is run. The historical data includes the pairing of the optimal PID control law parameter vectors under various external environmental disturbance signals.

[0070] A prior probability distribution of the PID control law parameter vector is constructed using a Gaussian Mixture Model (GMM). The GMM represents a complex probability distribution with multiple peaks by weighted sum of multiple Gaussian components. The posterior probability distribution is calculated based on historical data. A numerical approximation method using variational inference can be used. Variational inference finds an approximate distribution that is closest to the true posterior distribution but has a simpler form by optimizing an objective function called the lower bound of evidence.

[0071] The variational inference uses a Gaussian distribution family of mean fields as an approximate distribution. The parameter update model is a multilayer perceptron (MLP) with four layers. The network structure is as follows: the input layer contains three nodes, corresponding to external disturbance signals such as ambient temperature and humidity; the two hidden layers contain sixty-four and thirty-two neurons, respectively, and use exponential linear units (ELU) as nonlinear activation functions; the output layer contains three nodes, corresponding to the estimated values ​​of the proportional, integral, and derivative parameters of the PID controller. The Adam optimizer is used when training the parameter update model, and the learning rate is set to one-thousandth.

[0072] By performing a complete Bayesian inference process on all historical data, a large number of "input (environmental disturbance signal) - output (mean and variance of the posterior distribution of the optimal PID control law parameter vector)" data pairs are obtained. Using these data pairs, a lightweight neural network is trained as the parameter update model. The parameter update model is a multilayer perceptron (MLP) containing multiple hidden layers and employing a nonlinear activation function. This trained parameter update model establishes a complex and nonlinear mapping relationship from "external environmental disturbance signal" to "optimal PID control law parameter vector".

[0073] During the online real-time control phase, when the external environmental disturbance signal acquired in real time is input into this pre-trained parameter update model, the model directly outputs a set of estimated values ​​for a new PID control law parameter through a forward propagation calculation, and uses this set of estimated values ​​as the adaptive PID control law parameter under the current operating condition, and sends it down to the lower-level feedback controller.

[0074] This application's embodiments employ an "offline learning, online inference" model. In the offline stage, a lightweight neural network parameter update model is trained using a complete Bayesian inference process. In the online stage, adaptive parameters are quickly obtained through a single forward propagation calculation. This achieves rapid, real-time, and intelligent adaptation to changes in the external environment, solving the problem of drastic performance degradation of existing fixed-parameter PID controllers when facing operating condition drift. It ensures that the system maintains high-performance operation under different operating conditions.

[0075] Furthermore, refer to Figure 3 The lower-level feedback controller includes:

[0076] It receives the optimal target setpoint from the upper-level intelligent optimizer and uses it as the setpoint for the control loop of the lower-level feedback controller.

[0077] After receiving the optimal target setpoint and the real-time updated adaptive PID control law parameters from the upper-level intelligent optimizer, the lower-level feedback controller executes its core control law. The feedback controller integrates two key functional modules: a bumpless switching module for gradual transition of the integral term and an anti-integral saturation module.

[0078] The disturbanceless switching module is used to handle the updating of PID control law parameters (especially integral time parameters or integral gain). After receiving a new set of adaptive PID control law parameters, the disturbanceless switching module does not immediately update the integral term in the PID control law to the target value determined by the new parameters. Instead, it adopts an integral term reset strategy based on reverse calculation. Specifically, the module first calculates a new initial value of the integral term based on the newly received PID control law parameters (especially the new proportional gain Kp and the new derivative gain Kd), as well as the current error value and error change rate. The mathematical relationship of the initial value is set as follows: the new initial value of the integral term is equal to the total output value of the controller at the instant before the switch, minus the value obtained by multiplying the new proportional gain by the current error, minus the value obtained by multiplying the new derivative gain by the current error change rate, and finally the result is divided by the new integral gain Ki. This ensures that the total output value of the controller remains unchanged at the instant of parameter switching, thereby avoiding output jumps. At the same time, in order to cope with the step change of the target setpoint, the module also performs first-order or second-order low-pass filtering on the setpoint from the upper layer, converting it into a smoothly changing internal setpoint.

[0079] The anti-integral saturation module is used to handle situations where the drive actuator reaches its physical limit (e.g., the pump speed has reached its maximum). When the module detects that the actual output value of the actuator is inconsistent with the expected output value calculated by the controller (i.e., saturation occurs), it uses a reverse calculation method. Based on this deviation, the reverse calculation method corrects the accumulation of the integral term through an additional feedback loop. Specifically, the reverse calculation method corrects the accumulated value of the integral term based on the deviation between the actual output value of the actuator and the expected output value calculated by the controller through an additional feedback loop. The feedback loop multiplies the output deviation by an anti-saturation gain and subtracts the result from the accumulated value of the integral term, thereby keeping it within a reasonable, unsaturated range. The value of the anti-saturation gain is set to be proportional to the proportional gain Kp in the current PID control law, and the adjustment range of the proportional coefficient is limited to between 0.3 and 0.7, for example, set to 0.5 times the proportional gain.

[0080] With the combined effect of bumpless switching and anti-integral saturation function, the lower-level feedback controller finally performs standard discrete-time PID calculation based on the real-time updated setpoints and parameters, combined with the process variables obtained from the state observation unit, to generate a stable and smooth final control command.

[0081] This application embodiment integrates a smooth transition module for gradual integral term switching and an anti-integral saturation module into the lower-level feedback controller, thereby achieving smooth transition during online updates of PID control parameters and stable adjustment under actuator saturation conditions. This reduces system shocks and oscillations caused by improper handling of integral terms during parameter switching or actuator saturation.

[0082] Furthermore, the hierarchical digital controller includes:

[0083] The hierarchical digital controller also includes a fault-tolerant control module. This module employs a multi-source information fusion strategy to continuously and in parallel monitor two completely different but mutually corroborating conditions. The fault-tolerant control module determines that the control system has entered an abnormal operating condition when any of the following conditions are met:

[0084] Firstly, there is the confidence level of the posterior probability distribution, which is updated by the Bayesian parameters in the upper-level intelligent optimizer. It reflects the model's confidence in the currently calculated PID control law parameters. In a specific implementation, this confidence level is calculated from the variance of the posterior probability distribution. A smaller variance indicates a concentrated distribution, meaning the model's parameter estimation is highly certain, resulting in high confidence; conversely, a larger variance indicates a divergent distribution, meaning the model's parameter estimation is highly uncertain, resulting in low confidence.

[0085] Secondly, there is the process residual, calculated by an anomaly detector based on the residual. This is the normalized error between the actual measured values ​​of the state observation unit and the predicted values ​​of the nonlinear dynamic process model. The magnitude of the process residual directly reflects the degree of deviation between physical reality and model expectations. A consistently small residual indicates that the system is operating normally, while a consistently large residual indicates that the physical system may have experienced unexpected changes or that sensors may be malfunctioning.

[0086] The threshold settings for the two judgment conditions of the fault-tolerant control module are as follows:

[0087] First, the confidence threshold of the posterior probability distribution is set to a variance greater than 0.25.

[0088] Secondly, the preset range threshold of the process residual is set to the absolute value of the normalized error being greater than twice the standard deviation of the historical residual. The normalized error is calculated by dividing the difference between the actual measured value and the model predicted value by the difference between the upper limit and the lower limit of the sensor's range, thereby unifying the sensor deviations of different dimensions into a comparable range. The statistical window of the standard deviation of the historical residual is set to the most recent ten minutes.

[0089] Meanwhile, the time window for the "continuous" judgment is set to five consecutive control cycles; in a typical implementation, this time window is fifty milliseconds.

[0090] The system immediately determines that it has entered an abnormal operating condition if and only if any one of the above conditions is triggered and continues to exceed its respective preset threshold. For example, even if the physical residual is still within the normal range, the system will trigger the safety mode as long as the model itself is "not confident" (low confidence) in the parameters of its output PID control law. Conversely, even if the model is "very confident", the system will also trigger the safety mode as long as the result it predicts is seriously inconsistent with the physical reality (large residual).

[0091] Once an abnormal operating condition is detected, the system immediately executes a safety mode switch: automatically suspending all adaptive update functions of parameters and instructing the lower-level feedback controller to switch to a set of conservative PID control law parameters with low proportional gain and long integral time characteristics pre-stored in the controller's memory; the low gain and long integral time characteristics make the controller's response slow but extremely stable. At the same time, an 8-bit status code containing the fault source and fault type is sent to a human-machine interface (HMI) to guide the operator in troubleshooting.

[0092] The hierarchical digital controller also executes a closed-loop control method that integrates decision interpretability and human-machine collaborative evolution, including:

[0093] The decision interpretability step employs a model post-interpretation technique based on SHAP values. After the upper-level intelligent optimizer outputs key decisions, it automatically analyzes and calculates the input features that contribute the most. The analysis results are then combined with preset natural language templates to generate and present explanatory text to the operator.

[0094] Based on the presented explanatory text, a human-machine collaborative evolution process is executed, recording the operator's manual intervention actions on the system decision after understanding the explanatory text. When an intervention is verified by the system to produce a better control effect than the original system decision, the complete event of pre-intervention state - intervention action - better post-intervention result is automatically used as a high-quality expert demonstration sample. Using the expert demonstration sample, the strategy model of the upper-level intelligent optimizer is fine-tuned online through reinforcement learning from human feedback.

[0095] Control methods that possess decision interpretability and human-machine co-evolution build a bridge of trust between humans and machines by giving AI the ability to "explain its actions"; and create a closed loop of continuous learning by giving AI the ability to "learn from human strengths". This enables the control system to integrate the valuable experience and intuition of human experts into its mathematical model, achieving the integration and co-evolution of human and machine intelligence, and making the system increasingly intelligent and efficient in long-term use.

[0096] This application embodiment employs a fault-tolerant control module that combines "model confidence" and "physical residual" dual verification logic. Specifically, the system is only considered to be in an abnormal operating condition when the confidence level of the model's output is consistently below a threshold, and the deviation between the model's predicted value and physical reality also consistently exceeds the acceptable range. This achieves reliable diagnosis of abnormal operating conditions such as model mismatch or sensor anomalies, mitigating the risk of catastrophic errors in existing intelligent control methods due to single, unreliable judgment criteria. By automatically switching to a safe, conservative strategy, the reliability and safety of the entire control system are improved.

[0097] This application embodiment constructs a hierarchical control architecture that includes an upper-level intelligent optimizer and a lower-level feedback controller. A state observation unit collects real-time data; the upper-level intelligent optimizer performs state prediction, multi-objective optimization, and adaptive calculation of PID control law parameters based on the real-time data, and outputs the optimal target setpoint and adaptive PID control law parameters to the lower level; the lower-level feedback controller executes high-speed PID calculations according to the received instructions and transmits the final control instructions to the drive execution unit. By organically integrating and cooperating with three advanced control strategies—state prediction, optimization algorithms, and adaptive PID control—within a hierarchical architecture, synergistic effects of multiple algorithms are achieved, thus solving the problem that existing single control strategies cannot simultaneously achieve predictability, optimality, and adaptability.

[0098] Example 2:

[0099] This embodiment fully deploys the ink path control system of the aforementioned full-color printer in the actual production environment of A Paper Company, which consists of printers and toner shakers, to achieve intelligent collaborative control of the entire process from printing to finished product curing.

[0100] Once a printing job (e.g., a colorful T-shirt design with a large white background) is submitted, the control system begins to execute a complete workflow.

[0101] As the core of system decision-making, the upper-level intelligent optimizer refers to... Figure 2It utilizes a built-in nonlinear dynamic process model based on a spatiotemporally separated Transformer architecture to pre-simulate and calculate the entire printing process. This process model is a digital twin that includes all key physical processes. The input comes from a comprehensive network of state-monitoring sensors deployed throughout the device: in the printer's white ink filtration circulation system, miniature pressure and turbidity sensors report ink pressure and uniformity data in real time; in the heating section between the seven-color ink cartridges and the printhead, thin-film temperature sensors ensure that the ink viscosity is at its optimal state; in the powder shaker, array-type infrared sensors monitor the uniformity of curing temperature, while ultrasonic level sensors monitor the remaining powder.

[0102] A soft measurement module uses measured temperature and pressure data to estimate in real time the viscosity and surface tension of ink droplets at the moment of ejection—key but difficult-to-measure state parameters. Based on this comprehensive real-time and predictive data, the upper-level optimizer uses a sparrow search optimization algorithm to solve a multi-objective performance index function that aims to balance print quality, production efficiency, and energy consumption costs. Ultimately, it generates a set of globally optimal target settings, such as setting the printer preheating plate to 115°C, the curing lamp temperature of the toner shaker to 130°C, and planning an optimal printing speed curve.

[0103] These optimal target settings, along with a set of adaptive PID control law parameters calculated by the Bayesian inference module for the current operating conditions, are sent to the lower-level feedback controller. Upon receiving the instructions, the lower-level feedback controller executes its core PID control law, precisely drives the heater through a high-frequency PWM signal, and controls the electric pressure roller and tension system through a pulse sequence to ensure that the DTF film moves smoothly at the optimal speed and tension.

[0104] During the printing process, the system continuously performs closed-loop feedback and adaptive adjustments. For example, if the temperature and humidity of the workshop environment fluctuate, the state observation unit will immediately capture this external disturbance. The Bayesian parameter update model in the upper optimizer will immediately identify the impact of this external disturbance on ink drying and powder adsorption, and recalculate the PID control law parameters in real time. It will then instruct the lower controller to fine-tune the heating power and powder shaking amplitude, thereby actively offsetting the negative impact of environmental changes and ensuring the continuous stability of print quality.

[0105] The system's built-in fault-tolerant control module safeguards production safety throughout the process. If improper loading of the DTF film roll causes continuous abnormal fluctuations in the tension sensor readings, significantly deviating from the predicted values ​​of the dynamic model, and the posterior probability distribution confidence of the parameter update model output remains below the preset threshold, the fault-tolerant module will immediately determine that the system has entered an abnormal operating condition of model mismatch and trigger a safety mode: automatically suspending the parameter adaptive function, switching to a set of preset, conservative PID control law parameters (such as reducing printing speed and tension), and simultaneously displaying specific alarm information on the human-machine interface to guide the operator in checking and correcting the film path.

[0106] This invention achieves a deep deconstruction and reshaping of the printing process by constructing a complete technical system from intelligent decision-making and adaptive execution to closed-loop fault tolerance. It is no longer an automated device that passively executes preset instructions, but an intelligent production platform with autonomous perception, forward prediction, multi-objective collaborative optimization and dynamic environmental adaptability, providing a brand-new solution for achieving high-quality, high-efficiency and high-stability printing production.

[0107] It should be emphasized that the specific embodiments described herein are merely examples illustrating the core ideas of the present invention, and not intended to limit its scope. For those skilled in the art, any modifications, combinations, or equivalent substitutions made to the above embodiments without departing from the core principles and spirit disclosed in this invention should fall within the protection scope claimed by this invention. The final protection scope of this invention should be determined by the contents of the appended claims.

Claims

1. An ink path control system for a full-color printer, characterized in that, include: The state observation unit measures multiple physical quantities in the controlled physical process in real time and outputs them as state variables and external disturbance signals. Hierarchical digital controllers, including: The upper-level intelligent optimizer receives the state variables and external disturbance signals, predicts the future state through a nonlinear dynamic process model, and generates prediction results; it constructs a process performance index function, optimizes the performance index function through a sparrow search optimization algorithm based on the prediction results, and generates the optimal target setpoint; and it adaptively updates the parameters of the PID control law according to the external disturbance signals through a parameter update model based on Bayesian inference. The lower-level feedback controller receives the optimal target setpoint and PID control law parameters, and, in conjunction with the preset bumpless switching module, executes the PID control law to generate the final control command. The drive execution unit receives the final control command and converts it into drive signals that act on multiple physical actuators.

2. The ink path control system for a full-color printer according to claim 1, characterized in that, The state observation unit includes: a group of sensors distributed along the physical path of the controlled physical process to measure real-time dynamic pressure, temperature and flow rate as variables of the controlled physical process; an environmental sensor to measure the external environmental conditions of the equipment as external disturbances; and a soft measurement module that uses a pre-trained model based on physical information to estimate a process medium state parameter in real time and output it as a state variable.

3. The ink path control system for a full-color printer according to claim 1, characterized in that, The upper-layer intelligent optimizer receives the state variables and external disturbance signals, predicts the future state through a nonlinear dynamic process model, and generates a prediction result, including: the upper-layer intelligent optimizer is based on a nonlinear dynamic process model of a spatiotemporal separation Transformer architecture, which abstracts the controlled physical process into a graph structure, where each sensor node is defined as a vertex of the graph, the physical connection path connecting the vertices is defined as an edge of the graph, and the attribute of the edge is a dynamic flow resistance parameter calculated from the real-time state variables; a spatial encoder, composed of a graph neural network, aggregates information between adjacent vertices through a message passing mechanism, learns the physical dependencies between each node and its neighbors as spatial distribution features; a temporal encoder, composed of a recurrent neural network, extracts temporal dependency features from the historical state variable sequence; and a fusion layer, composed of a multi-head self-attention mechanism, adaptively weights and fuses the spatial distribution features and temporal dependency features, and outputs a prediction of the future state of the controlled physical process.

4. The ink path control system for a full-color printer according to claim 1, characterized in that, A process performance index function is constructed. Based on the prediction results, the performance index function is optimized using a sparrow search optimization algorithm to generate the optimal target setpoint. This includes: constructing a process performance index function composed of three weighted sub-objectives: a tracking error representing a quantitative indicator of material conveying volume, a positioning error representing a spatial accuracy indicator of the material's action point, and the total energy consumption required to control the actuator; in the offline initialization phase before the start of the operation cycle, a sparrow search optimization algorithm is used to achieve a global search by introducing Lévy flight random perturbations; in the offline phase, the process performance index function is iteratively solved using an improved sparrow search optimization algorithm to generate a control strategy lookup table, which stores a solution set consisting of a large number of Pareto optimal solutions corresponding to different operating conditions; in the online real-time control phase of the operation cycle, the optimal target setpoint is obtained by quickly retrieving and interpolating from the control strategy lookup table based on the current operating mode and the real-time state variables obtained from the state observation unit.

5. The ink path control system for a full-color printer according to claim 1, characterized in that, The parameter update model based on Bayesian inference includes: treating the three PID parameters of the lower-level feedback controller as a parameter vector to be estimated; in the offline training phase, running a complete Bayesian inference process to train a parameter update model, the Bayesian inference process including: constructing a prior probability distribution of the parameter vector in the form of a Gaussian mixture model; and calculating the likelihood probability and posterior probability distribution based on the external environmental disturbance signal and the corresponding optimal PID control law parameters in historical data; the parameter update model is a lightweight neural network that represents the mapping relationship from the external environmental disturbance signal to the optimal PID control law parameter vector; in the online real-time control phase, the real-time acquired external environmental disturbance signal is input into the parameter update model, and through one forward propagation calculation, the estimated value of the optimal PID control law parameter is directly obtained as the adaptive PID control law parameter under the current operating condition.

6. The ink path control system for a full-color printer according to claim 1, characterized in that, The lower-level feedback controller includes: receiving the optimal target setpoint from the upper-level intelligent optimizer as the setpoint of the lower-level feedback controller control loop; receiving adaptive PID control law parameters from the upper-level intelligent optimizer and updating the internal PID control law in real time, the update process being executed by a bumpless switching module, the bumpless switching module achieving smooth transition of PID control law parameters through a combination of reverse calculation and setpoint filtering; using the real-time state variable output by the state observation unit as the process variable, calculating the error between the process variable and the setpoint, and performing proportional, integral, and derivative operations according to the adaptive PID control law parameters to generate the final control command.

7. The ink path control system for a full-color printer according to claim 1, characterized in that, The hierarchical digital controller includes: the hierarchical digital controller further includes a fault-tolerant control module, which determines that the control system has entered an abnormal operating condition when any of the following conditions are met: The posterior probability distribution confidence of the parameter update model output is consistently below a preset threshold; a residual-based anomaly detector detects that the deviation between the actual measured value of the state observation unit and the predicted value of the nonlinear dynamic process model consistently exceeds a preset range. Upon determining that an abnormal operating condition has been entered, the control system automatically suspends parameter adaptive updates, switches to conservative PID control law parameters, and sends an alarm message to a human-machine interface. The conservative PID control law parameters are a set of fixed parameters with low gain and long integral time characteristics that are pre-stored in the controller's memory.

Images