A data-driven based adaptive control system and method for a knitting process

Abstract

The present application relates to the technical field of industrial adaptive control, in particular to a data-driven adaptive control system and method for a knitting process. The specific implementation process includes: gathering knitting field data streams containing yarn tension and loop density through a knitting data acquisition module; constructing a knitting execution prediction model and generating a control response feedforward prediction; in the control law synthesizer, solving and generating an actuator coordinated adjustment instruction sequence in real time with an optimization objective function as a constraint; establishing an outer loop adaptive correction loop to continuously quantify the residual error between the predicted output and the actual response and use it as the identification accuracy; when the identification accuracy exceeds the dynamic tolerance boundary, automatically triggering a correction reconstruction. The present application combines data-driven feedforward prediction, multi-objective optimization control and adaptive online correction of the model, effectively improving the low control accuracy problem caused by strong coupling, nonlinearity and time-varying nature of the working condition in the knitting process, and improving the adaptive control capability of the knitting process.

Description

Technical Field

[0001] This invention relates to the field of industrial adaptive control technology, specifically to a data-driven adaptive control system and method for a knitting process. Background Technology

[0002] In the field of automated manufacturing, production processes typically involve multiple variables. To effectively control such complex processes, existing technologies usually deploy sensors at key monitoring points in the knitting system to implement feedback control on core process variables (e.g., yarn tension). When the control system detects that the actual measured value of a process variable deviates from a preset setpoint, the controller sends instructions to the corresponding actuator to adjust its output for automated control compensation, suppressing the fluctuation of the variable within the allowable error range.

[0003] However, existing technologies have inherent limitations when dealing with complex dynamic operating conditions, often neglecting the strong coupling and nonlinear relationships that objectively exist between process variables. Forced adjustment of a single variable inevitably disturbs other related variables, thus affecting the overall state of the system. This local, non-cooperative control mode cannot achieve optimal global performance and may even compromise the overall quality of the final product by unilaterally pursuing the stability of a single variable. The system cannot proactively adapt to time-varying characteristics caused by batch differences in raw materials, equipment state drift, and environmental changes. Consequently, it lacks the ability for adaptive control and online strategy optimization, resulting in a significant decline in control performance when facing complex and ever-changing real-world operating conditions, making it difficult to fulfill its intended function.

[0004] To address this, a data-driven adaptive control system and method for the knitting process are proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a data-driven adaptive control system and method for the knitting process, thereby achieving adaptive control of the knitting process.

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

[0007] A data-driven adaptive control system for a knitting process includes:

[0008] The knitting data acquisition module collects knitting field data streams including yarn tension and loop density during the knitting process; using a recursive learning framework, it constructs a knitting execution prediction model with neural network parameterized nonlinear functions as the core, representing multivariable state space equations, and generates control response feedforward predictions for control adjustment through multi-step iterative deduction;

[0009] The process-driven constraint module, based on control response feedforward prediction, configures an optimization objective function in the control law synthesizer that includes minimizing the deviation of the production process state and maximizing the final product quality index; under the premise of satisfying the operating constraints, it solves and generates a sequence of actuator coordinated adjustment instructions in real time, driving multiple actuators of the controlled object to adjust the knitting field data flow;

[0010] An adaptive control correction module establishes an outer-loop adaptive correction circuit, continuously calculates the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected in the knitting field data stream, and quantifies it as identification accuracy. When the identification accuracy exceeds the dynamic tolerance boundary, the internal network parameters of the knitting execution prediction model are corrected and reconstructed.

[0011] Preferably, the specific implementation process for collecting knitting field data streams containing yarn tension and loop density during the knitting process includes:

[0012] The real-time acquired electrical signals are edge-processed and quantized into yarn tension; image processing algorithms are used to extract feature parameters representing the geometry of the coil from the acquired image information and convert them into coil density; the yarn tension and coil density are timestamped and integrated into a knitting field data stream with a unified data format.

[0013] Preferably, a knitting execution prediction model is constructed using a recursive learning framework, with a neural network parameterized nonlinear function as the core, representing a multivariable state-space equation. The specific implementation process for generating a control response feedforward prediction for control adjustment through multi-step iterative deduction includes:

[0014] The knitting field data stream is input into a recursive learning framework. A system identification algorithm is used to perform online parameter estimation of the nonlinear and time-varying coupling relationship between the control input and process output in the knitting field data stream, and a knitting execution prediction model is constructed. The current process state reflected by the knitting field data stream is input into the knitting execution prediction model. Through multi-step iterative deduction, the expected state trajectory of the controlled object in multiple future time steps is calculated and encapsulated as a control response feedforward prediction.

[0015] Preferably, the specific implementation process of configuring an optimization objective function in the control law synthesizer, which includes minimizing the production process state deviation and maximizing the final product quality index, based on the control response feedforward prediction, includes:

[0016] The control response feedforward prediction is input into a control law synthesizer, which contains an optimization objective function for evaluating the control response feedforward prediction. The optimization objective function includes a process state deviation term and a product quality evaluation term. The process state deviation term is quantified by calculating the Euclidean distance between the control response feedforward prediction and the ideal process trajectory. The product quality evaluation term uses an offline-trained quality surrogate model to map the control response feedforward prediction into a quality index. The process state deviation term and the product quality evaluation term are used to construct the optimization objective function.

[0017] Preferably, under the premise of satisfying operational constraints, the specific implementation process of generating a sequence of actuator coordinated adjustment instructions in real time to drive multiple actuators of the controlled object and adjust the knitting field data stream includes:

[0018] The optimization objective function is iteratively solved to extract the optimal control quantity for the current control cycle, and encapsulates it into an actuator coordinated adjustment instruction sequence containing the target setpoint and execution timing information. The actuator coordinated adjustment instruction sequence is distributed to the underlying drive unit corresponding to each actuator. The underlying drive unit parses the received instructions, converts them into electrical signals, and adjusts the knitting field data stream.

[0019] Preferably, the specific implementation process of establishing an outer-loop adaptive correction loop, continuously calculating the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected in the knitting field data stream, and quantifying it into identification accuracy includes:

[0020] The residuals of the control response feedforward prediction and the knitting field data stream are calculated. The two data streams are strictly aligned based on a unified timestamp, and the vector difference between the predicted value and the actual measured value is calculated point by point to generate a residual sequence that characterizes the degree of model mismatch. The residual sequence is then imported into the accuracy quantization unit, and the residual sequence is converted into continuously changing identification accuracy by integration within a set time window.

[0021] Preferably, when the identification accuracy exceeds the dynamic tolerance boundary, the specific implementation process of correcting and reconstructing the internal network parameters of the knitting execution prediction model includes:

[0022] The identification accuracy is fed into a threshold comparator and continuously compared with an online adjustable dynamic tolerance boundary. Once the identification accuracy is detected to exceed the dynamic tolerance boundary, a reconstruction trigger signal is immediately generated and sent. A dataset of a preset time length is retrieved from the historical data buffer storing the knitting field data stream and the corresponding actuator coordinated adjustment instruction sequence. Based on the retrieved dataset, the internal parameters of the knitting execution prediction model are corrected and reconstructed using the gradient descent method.

[0023] A data-driven adaptive control method for a knitting process includes:

[0024] The knitting field data stream, including yarn tension and loop density, is collected during the knitting process. Using a recursive learning framework, a knitting execution prediction model is constructed with a neural network parameterized nonlinear function as the core, representing a multivariable state space equation. Through multi-step iterative deduction, a control response feedforward prediction is generated for control regulation.

[0025] Based on the control response feedforward prediction, an optimization objective function is configured in the control law synthesizer, which includes minimizing the deviation of the production process state and maximizing the final product quality index. Under the premise of satisfying the operating constraints, the actuator coordinated adjustment instruction sequence is generated in real time to drive multiple actuators of the controlled object and adjust the knitting field data flow.

[0026] An outer-loop adaptive correction loop is established to continuously calculate the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected by the knitting field data stream, and quantify it as identification accuracy. When the identification accuracy exceeds the dynamic tolerance boundary, the internal network parameters of the knitting execution prediction model are corrected and reconstructed.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This invention constructs a multivariate knitting execution prediction model, which incorporates multiple key indicators such as yarn tension and loop density into a unified data stream for comprehensive analysis and feedforward prediction. This enables global coordinated adjustment of the knitting process, stabilizing individual production parameters and maximizing the overall quality indicators of the final product.

[0029] 2. This invention designs an outer-loop adaptive correction circuit that continuously calculates the residual between the model prediction and the actual production process response, and quantifies it as identification accuracy. When the accuracy exceeds the dynamically set tolerance boundary, the system automatically triggers the correction and reconstruction of model parameters. This online learning and adaptive correction capability enables the control system to cope with complex and ever-changing actual working conditions, significantly improving the robustness and adaptability of the control system, and ensuring the continuous stability and efficiency of the production process.

[0030] 3. This invention employs a data-driven feedforward predictive control strategy. By using a knitting execution prediction model to predict the state trajectory at multiple time steps in the future, the controller can plan and generate actuator commands in advance. This forward-looking control approach, combined with real-time solution of the optimal control sequence, significantly improves the response speed and control accuracy of the knitting process. Attached Figure Description

[0031] Figure 1This is a structural diagram of a data-driven adaptive control system for the knitting process proposed in this invention.

[0032] Figure 2 This is a flowchart of a data-driven adaptive control method for knitting processes proposed in this invention.

[0033] Figure 3 This is a schematic diagram of the adaptive control based on dynamic tolerance boundary proposed in this invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It must be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to constitute any limitation on the scope of protection of this invention. Therefore, all equivalent changes or modifications conceived by those skilled in the art based on the content disclosed in this invention without inventive effort should fall within the scope of protection claimed by this invention.

[0035] Reference Figures 1 to 3 This invention provides a data-driven adaptive control system and method for a knitting process, the technical solution of which is as follows:

[0036] Example 1:

[0037] This embodiment provides a data-driven adaptive control system for knitting processes, applicable to knitting process control, referring to... Figure 1 The system includes a knitting data acquisition module, a process-driven constraint module, and an adaptive control correction module.

[0038] The knitting data acquisition module collects knitting field data streams including yarn tension and loop density during the knitting process; using a recursive learning framework, it constructs a knitting execution prediction model with neural network parameterized nonlinear functions as the core, representing multivariable state space equations, and generates control response feedforward predictions for control adjustment through multi-step iterative deduction;

[0039] The process-driven constraint module, based on control response feedforward prediction, configures an optimization objective function in the control law synthesizer that includes minimizing the deviation of the production process state and maximizing the final product quality index; under the premise of satisfying the operating constraints, it solves and generates a sequence of actuator coordinated adjustment instructions in real time, driving multiple actuators of the controlled object to adjust the knitting field data flow;

[0040] An adaptive control correction module establishes an outer-loop adaptive correction circuit, continuously calculates the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected in the knitting field data stream, and quantifies it as identification accuracy. When the identification accuracy exceeds the dynamic tolerance boundary, the internal network parameters of the knitting execution prediction model are corrected and reconstructed.

[0041] Furthermore, the specific implementation process for collecting knitting field data streams, including yarn tension and loop density, during the knitting process includes:

[0042] The real-time acquired electrical signals are edge-processed and quantized into yarn tension; image processing algorithms are used to extract feature parameters representing the geometry of the coil from the acquired image information and convert them into coil density; the yarn tension and coil density are timestamped and integrated into a knitting field data stream with a unified data format.

[0043] Specifically, a set of high-frequency piezoelectric tension sensors is installed along the yarn feeding path in the weaving area. These sensors convert the minute tension fluctuations generated by the yarn during movement into a continuous 0-5 volt analog voltage signal in real time, which is then transmitted to an edge computing unit deployed on the equipment side. This edge computing unit incorporates a high-speed analog-to-digital converter and preprocessing algorithms, acquiring the voltage signal at a sampling frequency of 1000 Hz and running a Kalman filter algorithm to eliminate electromagnetic noise introduced by equipment vibration. After processing, the clean digital signal is quantized into a standard yarn tension value, measured in centinenewtons (cN), according to a preset calibration curve. For example, a 2.5-volt voltage signal, after processing and quantization, is precisely recorded as a yarn tension of 35.5 cN. Each frame of tension data is appended with a high-precision timestamp, accurate to the millisecond level, providing a benchmark for subsequent data alignment.

[0044] An industrial-grade high-resolution CCD camera, equipped with a ring-shaped shadowless LED light source, is fixedly mounted directly above the needle bed, enabling it to clearly and uniformly capture the newly woven fabric area. The camera continuously captures images of the fabric at a rate of 50 frames per second, with an image resolution of 1920*1080 pixels. Each captured frame is immediately sent to the image processing module. This module converts the original color image to grayscale, then uses a median filtering algorithm to eliminate random noise points such as yarn fuzz; it employs the Canny edge detection operator to sharpen and identify the loop outlines formed by the yarn in the image. Based on the outline information, the Hough transform is used to identify and count the number of rows and columns within a predefined 1 square centimeter virtual measurement window. For example, at a specific sampling moment, the algorithm identifies 45 rows and 52 columns. Based on this, the system calculates the current loop density to be 2340 loops per square centimeter. Similar to the yarn tension data, each calculated loop density value is also precisely timestamped to the millisecond level.

[0045] A key step in the data acquisition process is the effective integration of two data streams from different sources and with varying frequencies. Yarn tension data is generated at a frequency of 1000 Hz, while loop density data is generated at a frequency of 50 Hz. To construct a dataset that accurately reflects the dynamic causal relationship between the two, the system's data fusion module performs a timestamp-based alignment operation. Using the timestamp of the lower-frequency loop density data as a reference, this module calculates, for each loop density sampling point, a yarn tension value that is completely synchronized with the sampling time of that loop density by using a linear interpolation algorithm among the nearest few high-frequency yarn tension sampling points before and after it. Through this process, the system integrates and serializes these data pairs containing timestamps, yarn tension values, and loop density values, forming a continuous knitting field data stream with a unified data format.

[0046] This embodiment generates a knitting field data stream that accurately reflects the inherent physical laws of the knitting process through high-precision timestamp alignment and data integration. This enables subsequent predictive models to more accurately learn and identify the dynamic characteristics of the system, thereby allowing the entire adaptive control system to make more precise and forward-looking adjustments.

[0047] Furthermore, utilizing a recursive learning framework, a knitting execution prediction model is constructed, with a neural network parameterized nonlinear function as its core, representing the multivariable state-space equation. Through multi-step iterative deduction, the specific implementation process for generating feedforward prediction of the control response for control regulation includes:

[0048] The knitting field data stream is input into a recursive learning framework. A system identification algorithm is used to perform online parameter estimation of the nonlinear and time-varying coupling relationship between the control input and process output in the knitting field data stream, and a knitting execution prediction model is constructed. The current process state reflected by the knitting field data stream is input into the knitting execution prediction model. Through multi-step iterative deduction, the expected state trajectory of the controlled object in multiple future time steps is calculated and encapsulated as a control response feedforward prediction.

[0049] Specifically, to address the nonlinear and time-varying coupling relationships prevalent in knitting processes, such as the elasticity differences caused by yarn batch changes or the impact of environmental temperature and humidity variations on the yarn friction coefficient, this embodiment employs a neural state-space model. This model utilizes a neural network to parameterize the nonlinear state transition function and observation function in the multivariate state-space equation. Both the state transition function and the observation function are constructed from a multilayer perceptron neural network with two hidden layers. Each hidden layer contains 64 neurons and uses a modified linear unit as the activation function. The neural network input to the state transition function is the current state vector (containing yarn tension and coil density) and the current control input (e.g., yarn feeder speed and take-up roller torque); its output is the predicted value of the state vector at the next moment. The neural network of the observation function maps the predicted state vector back to the observable measurement value. The recursive learning framework is an online update mechanism integrating an extended Kalman filter algorithm. It treats the model parameters as augmented states. When a continuous knitting data stream is input, the system state is estimated in real time through a prediction-update loop, while the neural network parameters are also estimated online. At the core of this framework is a multivariate state-space equation. Its state vector contains not only directly observable yarn tension and loop density, but also several hidden state variables characterizing the system's intrinsic properties. When a continuous stream of knitting data, such as a set of data points including a feeder speed of 30.5 revolutions per minute, a yarn tension of 40.2 centinewtons, and a loop density of 2510 loops per square centimeter, is input into this recursive learning framework, the system identification algorithm is activated. The extended Kalman filter algorithm estimates and corrects the parameters in the state-space model online through a prediction-update loop. It uses the model parameters from the previous time step to predict the system output at the current time step and compares this prediction with the actual measurements collected by the sensors. The residual between the two is used as a feedback signal to adjust and update the state transition matrix and observation matrix parameters within the model through Kalman gain. This process is performed continuously at a frequency of 10 times per second, ensuring that the model can track and fit the dynamic characteristic drift of the controlled object (i.e., the knitting process) caused by fluctuations in raw material properties or equipment wear in real time.

[0050] When control adjustments are required, the system first acquires the current process state reflected in the knitting data stream, for example, the current yarn tension is 41.0 centinenewtons and the loop density is 2505 loops / cm². This current state is used as the initial condition input into the knitting execution prediction model that has just been updated online. The control system enters a multi-step iterative deduction phase, generating a control input sequence for a future time period (e.g., the next 2 seconds, corresponding to 100 control steps) based on a preliminary control strategy. The model will then iteratively calculate step by step based on this input sequence, starting from the current state, to solve for the expected state trajectory of the controlled object within these 100 time steps. The deduction result is a detailed multi-dimensional time series containing the predicted yarn tension and predicted loop density values ​​for each future moment. For example, the model might predict that under the current control strategy, the yarn tension will rise to 43.5 centinenewtons after 0.5 seconds and cause the loop density to drop to 2490 loops / cm² after 1.2 seconds, thus deviating from the process target. This complete expected state trajectory is eventually encapsulated into a standardized data structure, namely "control response feedforward prediction", and sent to the downstream control law synthesizer.

[0051] This embodiment empowers the control system with the ability to predict the future by constructing a dynamic predictive model capable of online adaptive learning. Through the extrapolation of future state trajectories, the controller can anticipate the trend and magnitude of quality deviations before they actually occur, thereby enabling proactive intervention and ensuring the high stability of knitted product quality.

[0052] Furthermore, based on control response feedforward prediction, the specific implementation process of configuring an optimization objective function in the control law synthesizer that includes minimizing production process state deviations and maximizing final product quality indicators includes:

[0053] The control response feedforward prediction is input into a control law synthesizer, which contains an optimization objective function for evaluating the control response feedforward prediction. The optimization objective function includes a process state deviation term and a product quality evaluation term. The process state deviation term is quantified by calculating the Euclidean distance between the control response feedforward prediction and the ideal process trajectory. The product quality evaluation term uses an offline-trained quality surrogate model to map the control response feedforward prediction into a quality index. The process state deviation term and the product quality evaluation term are used to construct the optimization objective function.

[0054] Specifically, the control law synthesizer is internally configured with a composite optimization objective function, which consists of two core parts: a process state deviation term and a product quality evaluation term. To calculate the process state deviation term, the system loads an ideal process trajectory. In a preferred embodiment, for automotive interior fabrics, the ideal process trajectory is set with a stable yarn tension of 25.0 centinewtons and a constant loop density of 1850 loops / cm². The quantification of the process state deviation term is accomplished by calculating the Euclidean distance between the predicted trajectory and this ideal trajectory point by point. For example, at the 10th time step in the future, if the predicted yarn tension is 26.2 centinewtons and the loop density is 1845 loops / cm², then the state vector at that point is (26.2, 1845), and the Euclidean distance between it and the ideal state vector (25.0, 1850) is calculated. The system accumulates all distance values ​​over the next 150 time steps to obtain a total deviation value (e.g., 128.4), which intuitively reflects the overall fluctuation of the future production process.

[0055] To calculate product quality assessment items, the system uses an offline-trained deep neural network as a quality proxy model. The model's training process utilizes months of accumulated historical production data. The input is a complete trajectory of process parameters, while the output label is a comprehensive quality score from 0 to 100, derived from laboratory tests of the finished fabric's abrasion resistance, colorfastness grading, and image analysis of texture clarity. Specifically, the predicted quality score is calculated as follows: when the control law synthesizer inputs the current complete trajectory containing 150 predicted state points into the quality proxy model (e.g., a gated recurrent unit network), it processes the sequence step-by-step, and the hidden state vector of the last time step is input into one or more fully connected layers. These fully connected layers ultimately output a single predicted quality score. For example, for the current predicted trajectory, the proxy model predicts that it can produce fabric with a comprehensive quality score of 95.2.

[0056] After obtaining the two heterogeneous evaluation values, the process state deviation item (128.4) and the product quality evaluation item (95.2), the system must unify them onto a common evaluation scale. To this end, the system first normalizes these two values, mapping them to the interval between 0 and 1. Based on the specific production goals, the two normalized values ​​are weighted and fused. In this embodiment, due to the stringent requirements for the final quality of automotive interior fabrics, the product quality evaluation item is assigned a weight of 0.7, while the process state deviation item is assigned a weight of 0.3. These two weighted values ​​are linearly combined to construct a final optimization objective function. The goal of the optimization algorithm is to find a set of future actuator operation sequences such that the predicted trajectory generated by this operation sequence scores optimally (i.e., has the lowest cost) under this final cost function.

[0057] This embodiment introduces an optimization objective function, enabling the control system to quantify and predict the direct impact of its decisions on the final product quality. This forward-looking decision-making mechanism based on multi-objective optimization ensures the stability of the control system's regulation, thereby effectively improving the quality and efficiency of high-end knitwear production.

[0058] Furthermore, under the premise of satisfying operational constraints, the specific implementation process of generating a sequence of actuator coordinated adjustment instructions in real time to drive multiple actuators of the controlled object and adjust the knitting field data flow includes:

[0059] The optimization objective function is iteratively solved to extract the optimal control quantity for the current control cycle, and encapsulates it into an actuator coordinated adjustment instruction sequence containing the target setpoint and execution timing information. The actuator coordinated adjustment instruction sequence is distributed to the underlying drive unit corresponding to each actuator. The underlying drive unit parses the received instructions, converts them into electrical signals, and adjusts the knitting field data stream.

[0060] Specifically, at the start of each 100-millisecond control cycle, the model predictive control solver within the control law synthesizer initiates iterative calculations. This solver, constrained by a composite optimization objective function that minimizes process state deviations (e.g., deviation from the target yarn tension of 30.0 centinewtons) and maximizes predicted product quality (e.g., a target quality score of 98 points), explores possible combinations of actuator actions within the prediction time domain of the next 3 seconds (i.e., 30 control steps) using a sequential quadratic programming algorithm. After intensive computation of no more than 80 milliseconds, the solver derives a complete set of optimal control sequences. The system extracts the optimal control quantity for the first time step, i.e., the current control cycle (within the next 100 milliseconds), from this sequence. For example, the solver might calculate the optimal control quantity as follows: adjusting the yarn feeding speed of yarn feeder #3 from the current 12.5 meters per minute to 12.8 meters per minute, simultaneously fine-tuning the position of the sinker triangle #5 from 4.2 millimeters to 4.15 millimeters, and increasing the rotation speed of the take-up roller to 1.8 revolutions per minute.

[0061] These discrete values ​​are then encapsulated into a sequence of actuator coordination control instructions with execution timing information. Specifically, this sequence consists of a data frame containing fields such as timestamp, target actuator ID, target setpoint, unit, and ramp time. For example, the instruction explicitly requires that the adjustment of the yarn feeder speed be completed within 20 milliseconds after the instruction is issued, while the adjustment of the sinker cam position must begin at 40 milliseconds and be completed within 30 milliseconds, thus ensuring the coordination of actions. This instruction sequence is precisely distributed to the underlying drive units corresponding to each actuator via an industrial Ethernet bus. For example, the instruction for the yarn feeder is sent to its servo driver, and the instruction for the sinker cam is sent to its stepper motor controller. Upon receiving the instruction, each underlying drive unit immediately parses it, converting physical quantity instructions such as "speed 12.8 m / min" into the duty cycle or bus current value of the pulse width modulation signal required to drive the motor. These electrical signals directly drive the motor to change its operating state, thereby applying precise physical adjustments to the yarn tension and the coil formation process. The adjustment effect will be reflected in the knitting field data stream collected in the next cycle, forming a complete and high-precision closed-loop control.

[0062] This embodiment, through iterative solving based on model prediction, can proactively calculate the action commands to achieve the globally optimal goal within a future timeframe. The timing information in the command sequence ensures that the actions of different actuators are performed in the optimal order and rhythm, avoiding new process disturbances caused by improper action timing, thereby improving control accuracy and process stability.

[0063] Furthermore, the specific implementation process of establishing an outer-loop adaptive correction circuit, continuously calculating the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected in the knitting field data stream, and quantifying it into identification accuracy includes:

[0064] The residuals of the control response feedforward prediction and the knitting field data stream are calculated. The two data streams are strictly aligned based on a unified timestamp, and the vector difference between the predicted value and the actual measured value is calculated point by point to generate a residual sequence that characterizes the degree of model mismatch. The residual sequence is then imported into the accuracy quantization unit, and the residual sequence is converted into continuously changing identification accuracy by integration within a set time window.

[0065] Specifically, the system constructs a mechanism that can monitor and quantify the accuracy of the knitting execution prediction model in real time. This mechanism runs continuously in a parallel computing thread, and its input consists of two key data streams: one is the "control response feedforward prediction" generated by the system at a certain point in the past, and the other is the "knitting field data stream" collected in real time by sensors at the current moment.

[0066] The first step in residual calculation is to achieve strict alignment of the two data streams in the time dimension. Since all data in the system, whether predicted or measured, is tagged with a unified timestamp accurate to the millisecond, the alignment process is precisely achieved. For example, at system timestamp T, the adaptive correction module retrieves the state prediction value generated at time Tk (where k is the prediction time domain length) from the historical database for time T. At T-2000 milliseconds, the prediction model outputs a state prediction value of yarn tension of 32.5 centinewtons and a loop density of 2150 loops / cm². Simultaneously, the correction module obtains the actual knitting field data measured and processed by the sensor at time T: yarn tension of 33.1 centinewtons and a loop density of 2147 loops / cm². At this point, the system calculates the difference between these two state vectors point by point, i.e., (33.1-32.5, 2147-2150), thus obtaining a two-dimensional residual vector (+0.6, -3.0). This residual vector quantifies the prediction error of the model at the current moment in two key dimensions. The calculation process is repeated once per control cycle (100 milliseconds) to generate a continuous stream of residual sequences that characterize the degree of model mismatch.

[0067] To transform this potentially noisy residual sequence into a smooth metric with clear physical meaning, the sequence is fed into a precision quantization unit, the core of which is a sliding time window integrator. In this embodiment, a sliding time window of 10 seconds is set, meaning that at any given time, the unit caches all residual vectors from the past 10 seconds (i.e., 100 control cycles). The unit then calculates the Euclidean norm (i.e., the magnitude of the vector) for each residual vector within the window to obtain a scalar error magnitude. These 100 error magnitudes are then summed to obtain the error integral within the time window. This continuously varying, smoothed integral error value is defined as the quantization metric for "identification accuracy." The lower this metric value, the closer the overall predicted output of the prediction model matches the actual response of the physical process over a given period, and the higher the model's identification accuracy.

[0068] This embodiment transforms the abstract concept of model accuracy into a continuous index that can be monitored and quantitatively compared in real time by establishing an outer-loop adaptive correction circuit. This enables the system to proactively perceive the "outdatedness" of the prediction model. This ability to learn and adaptively correct online ensures that the control system can continuously cope with complex and ever-changing actual operating conditions, significantly improving the robustness and adaptability of the control system.

[0069] Furthermore, when the identification accuracy exceeds the dynamic tolerance boundary, the specific implementation process of correcting and reconstructing the internal network parameters of the knitting execution prediction model includes:

[0070] The identification accuracy is fed into a threshold comparator and continuously compared with an online adjustable dynamic tolerance boundary. Once the identification accuracy is detected to exceed the dynamic tolerance boundary, a reconstruction trigger signal is immediately generated and sent. A dataset of a preset time length is retrieved from the historical data buffer storing the knitting field data stream and the corresponding actuator coordinated adjustment instruction sequence. Based on the retrieved dataset, the internal parameters of the knitting execution prediction model are corrected and reconstructed using the gradient descent method.

[0071] Specifically, the system calculates an "identification accuracy" index characterizing the degree of model mismatch, which is then sent to the threshold comparator module in real time. This comparator incorporates a "dynamic tolerance boundary," whose value is not fixed but can be adjusted online by process engineers based on the criticality of the production task. For example, in routine production, this boundary is set to an integral error value of 4.5. However, when producing a critical precast of a main load-bearing structural component, engineers tighten this boundary to 3.0 via the human-machine interface to ensure the highest control accuracy. The threshold comparator continuously compares the real-time identification accuracy value with the current dynamic tolerance boundary once per control cycle. During a continuous production process, due to the replacement of a carbon fiber tow with slightly drifting tension characteristics, the system's identification accuracy value gradually climbed from a stable 2.5 within minutes. When it reached 3.01, it immediately exceeded the set dynamic tolerance boundary of 3.0. Upon detecting this boundary violation, the threshold comparator immediately generates a high-level reconstruction trigger signal and broadcasts it through the system's internal event bus.

[0072] A dedicated data management module is activated immediately upon detecting the reconstruction trigger signal. It accesses a historical data buffer that continuously stores all high-frequency operation data from the past hour. The module retrieves a dataset from the buffer covering the period 15 minutes prior to the trigger signal. This dataset includes not only all knitting data streams collected by sensors during those 15 minutes (i.e., yarn tension and loop density sequences), but also a sequence of precisely timestamped coordinated control commands issued by the system to each actuator during the same time period. This constitutes an input-output pair sample set containing tens of thousands of data points, recording the system's dynamic response characteristics under the latest operating conditions.

[0073] Once the dataset is ready, the system automatically starts a background training task. The correction and reconstruction process specifically involves using the retrieved dataset as training samples and employing a momentum-based mini-batch gradient descent algorithm to correct and reconstruct the internal parameters of the currently running online knitting execution prediction model. The loss function between the model's predicted output and the actual output on the training samples is calculated, the error is backpropagated, and the weights and bias parameters in the network are iteratively updated. The entire training process is performed on dedicated computing hardware without interrupting normal production and converges within approximately 45 seconds. When the training loss falls below a threshold, it indicates that the model has successfully learned and adapted to the new operating conditions. The system then replaces the original model parameters with this optimized set of new parameters, thus completing a full automated model correction and reconstruction cycle.

[0074] This embodiment establishes a fully automated closed loop of "monitoring-triggering-data retrieval-reconstruction," enabling the system to proactively and promptly respond to model mismatches caused by unavoidable factors such as raw material fluctuations, equipment aging, and environmental changes, thus achieving "online evolution" of the model. This self-correcting capability ensures that the control system maintains near-optimal control performance throughout its entire lifecycle, significantly improving the robustness and yield of the knitting process.

[0075] Example 2:

[0076] Reference Figure 2 This embodiment proposes a data-driven adaptive control method for the knitting process, including:

[0077] The knitting field data stream, including yarn tension and loop density, is collected during the knitting process. Using a recursive learning framework, a knitting execution prediction model is constructed with a neural network parameterized nonlinear function as the core, representing a multivariable state space equation. Through multi-step iterative deduction, a control response feedforward prediction is generated for control regulation.

[0078] Based on the control response feedforward prediction, an optimization objective function is configured in the control law synthesizer, which includes minimizing the deviation of the production process state and maximizing the final product quality index. Under the premise of satisfying the operating constraints, the actuator coordinated adjustment instruction sequence is generated in real time to drive multiple actuators of the controlled object and adjust the knitting field data flow.

[0079] An outer-loop adaptive correction loop is established to continuously calculate the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected by the knitting field data stream, and quantify it as identification accuracy. When the identification accuracy exceeds the dynamic tolerance boundary, the internal network parameters of the knitting execution prediction model are corrected and reconstructed.

[0080] Furthermore, in the data acquisition and feedforward prediction stage, yarn tension and loop density are acquired at frequencies of 1000 Hz and 50 Hz using piezoelectric sensors and high-speed industrial cameras, respectively. After timestamp alignment and interpolation fusion, a unified format knitting field data stream is generated. For example, at time T, the acquired current process state is a yarn tension of 28.5 centinenewtons and a loop density of 3210 loops per square centimeter. This state is immediately input into the knitting execution prediction model. A "control response feedforward prediction" is then generated, which is a detailed multidimensional time series describing the expected evolution trajectory of yarn tension and loop density over the next 2.5 seconds (25 control steps) if the current control strategy is maintained.

[0081] Furthermore, the control response feedforward prediction is fed into the control law synthesizer as a key input to the model predictive control solver. The solver iteratively solves a composite optimization objective function, which comprises two parts: a process state deviation term, which quantifies process stability by calculating the cumulative Euclidean distance between the predicted trajectory and the ideal process trajectory (e.g., target yarn tension constant at 28.0 centinewtons, coil density constant at 3250 coils / cm²); and a product quality assessment term, which uses an offline-trained quality surrogate model that maps the process parameter trajectory to a comprehensive score of the final fabric elastic recovery rate and breathability, to provide an expected quality score for the predicted trajectory, e.g., 97.2 points. Depending on the different emphases of stability and final quality in the current production order, these two terms are weighted and summed with weights of 0.4 and 0.6, respectively. Within 80 milliseconds, the solver calculates the optimal actuator coordinated adjustment command sequence for the next 25 steps that minimizes this weighted composite cost function. The first instruction in the sequence is extracted, for example, setting the speed of yarn feeder #2 to 15.4 meters per minute and the torque of the main take-up roller to 3.5 Nm. This instruction is encapsulated into a data frame containing precise execution timing and distributed to the corresponding servo drive unit via the EtherCAT industrial bus. The drive unit parses it into corresponding current and voltage signals to apply precise adjustments to the knitting data stream.

[0082] Furthermore, an outer-loop adaptive correction loop is run to ensure high accuracy of the model during long-term operation. This loop continuously calculates the residual between the predicted output of the prediction model and the actual response of the physical process. For example, the actual measured value at time T (e.g., yarn tension 28.5 cmN, coil density 3210 coils per square centimeter) is compared with the predicted value for time T made by the model at time T-2500 milliseconds (e.g., yarn tension 28.1 cmN, coil density 3225 coils per square centimeter), resulting in a residual vector (+0.4, -15). These continuously generated residuals are fed into an integrator with a 30-second sliding time window, quantifying them into a continuously changing "identification accuracy" index. After a production material change, due to a slight change of about 3% in the elastic modulus of the new batch of yarn, the model's predictions began to deviate from reality, and the identification accuracy index gradually climbed from below 2.5 within 5 minutes. When the index value reaches and exceeds the dynamically adjustable tolerance boundary 4.0 set online by the process engineer, the threshold comparator immediately generates and sends a reconstruction trigger signal. This signal activates the data management module, which retrieves a high-fidelity dataset (containing actuator instruction sequences and corresponding knitting field data streams) from the historical data buffer, covering the most recent 20 minutes after material change and fully reflecting the characteristics of the new working conditions. Using this dataset, the Adam optimizer is used to train the internal network parameters of the knitting execution prediction model using mini-batch gradient descent. The entire reconstruction process is completed automatically in the background, taking approximately 70 seconds. After training convergence, this updated set of model parameters, adapted to the characteristics of the new raw materials, is seamlessly deployed online, adaptively updating the old parameters.

[0083] This embodiment constructs a complete adaptive control closed loop with capabilities from perception, prediction, decision-making, execution, and self-reflection and evolution. It not only achieves high-precision and forward-looking control of complex knitting processes but also, through an outer-loop adaptive correction circuit, endows the model with the ability to cope with changes in operating conditions and achieve autonomous iterative optimization. This ensures the continuous and stable production of knitted products that meet quality standards, greatly improving the intelligence level and economic efficiency of the automated production line.

[0084] Example 3:

[0085] This embodiment fully deploys the aforementioned data-driven adaptive control system and method for the knitting process in the knitting workshop of Factory X, referring to... Figure 3 This enables adaptive control of the knitting process.

[0086] High-frequency piezoelectric tension sensors are installed on each yarn path of the computerized flat knitting machine in the workshop, monitoring the tension fluctuations of the yarn in real time before it enters the needle feeding area at a frequency of 1000 Hz. A high-resolution industrial camera is fixed directly above the needle bed, using a strobe light source to capture images of the fabric surface just formed at a rate of 50 frames per second. The image processing unit runs edge detection and image recognition algorithms in real time to calculate the number of horizontal and vertical loops within the preset measurement area and converts them into a loop density value in "loops per square centimeter". The system aligns and synchronizes the tension and density data, which are from these two different types of sensors and have millisecond-level high-precision timestamps, and generates a continuous knitting field data stream in a uniform format that is completely synchronized with the high-frequency tension data stream by performing linear interpolation between the low-frequency loop density data points.

[0087] The predictive model constructed in this embodiment is based on a multivariate state-space representation. This representation's state vector includes not only the directly measurable physical quantities of yarn tension and loop density, but also two key hidden state variables to more profoundly capture the system's intrinsic dynamics: one variable characterizes the dynamic frictional characteristics between the yarn and metal components such as yarn guides and sinkers, which change due to subtle variations in the yarn surface sizing agent or increases in equipment temperature; the other hidden state variable characterizes the minute dimensional drift caused by thermal expansion of critical components (e.g., the needle bed) due to prolonged equipment operation. The system utilizes an online system identification algorithm based on extended Kalman filtering to continuously estimate and update the parameters within this state-space model based on real-time input knitting field data streams and actuator action commands. At the beginning of each control cycle, for example, at time T, the system collects the current yarn tension as 45.5 centinewtons and the loop density as 1850 loops / cm². The system then uses this current state as the initial condition and inputs it into the newly updated prediction model. Through multi-step iterative deduction, it generates a control response feedforward prediction that describes the expected evolution trajectory of yarn tension and loop density within the next 2.5 seconds (corresponding to 25 control steps).

[0088] In the control law synthesizer, optimization decisions and instruction generation are performed based on the feedforward prediction of the control response. The optimization objective is a composite objective function that simultaneously minimizes the process state deviation and maximizes the final product quality. The process state deviation term is calculated by comparing the cumulative Euclidean distance between the feedforward predicted trajectory and an ideal process trajectory (e.g., setting the yarn tension to a stable 45.0 centinewtons and the coil density to a constant 1860 coils / cm²). The product quality evaluation term utilizes an offline-trained gated recurrent unit neural network as a quality surrogate model. The input to this surrogate model is the predicted 2.5-second state trajectory and the actuator action sequence driving that trajectory, and its output is a comprehensive product quality score from 0 to 100. The training labels for this score are obtained by weighted fusion of the 3D scanning dimensional accuracy of the finished prefabricated part (weight 0.5), the fiber volume content uniformity analysis results (weight 0.3), and the machine vision score for the absence of surface wrinkles and defects (weight 0.2). In this embodiment, since the final product quality is crucial, the quality assessment item is assigned a weight of 0.7, while the process state deviation item is assigned a weight of 0.3. The control law synthesizer, using a sequential quadratic programming algorithm, solves for the optimal actuator coordinated adjustment instruction sequence for the next 25 steps within a computation time of no more than 80 milliseconds. The system extracts the first instruction from this sequence, for example: fine-tuning the yarn feeding speed of yarn feeder No. 4 to 18.5 meters per minute, while simultaneously setting the torque of the main take-up roller to 8.2 Nm, and sends this instruction to the corresponding servo drive unit via the industrial Ethernet bus, driving the motor to adjust the knitting field data stream.

[0089] Simultaneously, the system runs an outer-loop adaptive correction loop in parallel to ensure the accuracy of the prediction model during continuous operation. This loop continuously calculates the residual between the model's predicted value for the current state at a past time (e.g., T-2500 milliseconds) and the actual value measured by the sensor at the current time T. For example, if the predicted value is a tension of 45.1 centinewtons and a density of 1855 particles / cm², while the actual value is 45.5 centinewtons and 1850 particles / cm², a residual vector is generated. Within a sliding time window of twenty seconds, the system sums these continuously generated residual vectors to obtain the error integral within that time window, generating a continuously changing "identification accuracy" index. The tolerance boundary of this index is dynamic, not a fixed value, but is related to the part number of the currently produced prefabricated component and its criticality level in the final product. For the column structure reinforcement in this embodiment, which belongs to the highest criticality level, its identification accuracy tolerance boundary is automatically retrieved from the process database and set to an integral error value of 2.5. When the production task switches to non-load-bearing interior trim, this boundary automatically widens to 4.0. During one production run, due to the replacement of a roll of carbon fiber yarn with slightly different surface properties, the system's recognition accuracy gradually increased over several minutes. When it reached 2.51, it exceeded the dynamic tolerance boundary of 2.5. At this point, the system automatically triggered a correction and reconstruction of the internal parameters of the knitting execution prediction model. The system retrieved high-fidelity knitting data streams and corresponding actuator instruction sequences from the historical data buffer, containing the most recent twenty minutes of data after the material change, as a new training sample set. Without interrupting normal production, the system retrained the prediction model's internal network parameters in the background using a momentum-based mini-batch gradient descent algorithm. The entire reconstruction process took approximately 90 seconds. After training, the new model parameters, successfully adapted to the new working conditions, were seamlessly deployed online, replacing the old parameters for adaptive model evolution, thus achieving adaptive control of the knitting process within the factory workshop.

[0090] It should be clarified that the embodiments described above are merely exemplary and are intended to aid in understanding the present invention, not to limit it. Those skilled in the art can make various changes and modifications after grasping the core ideas of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data-driven adaptive control system for a knitting process, characterized in that, include: The knitting data acquisition module collects knitting field data streams, including yarn tension and loop density, during the knitting process. Using a recursive learning framework, a knitting execution prediction model is constructed with a neural network parameterized nonlinear function as the core, representing a multivariable state space equation. Through multi-step iterative deduction, a control response feedforward prediction is generated for control regulation. The recursive learning framework is an online update mechanism that integrates the extended Kalman filter algorithm. It treats the knitting execution prediction model parameters as augmented states. When a continuous knitting field data stream is input, the state is estimated in real time through a prediction-update loop, and the neural network parameters are estimated online. The multivariate state space equation also includes hidden state variables to characterize intrinsic properties. One hidden state variable is used to characterize the dynamic friction characteristics between the yarn and the yarn guide and sinker metal parts, while the other hidden state variable is used to characterize the dimensional drift caused by the thermal expansion of the parts due to long-term operation of the equipment. The process-driven constraint module, based on the control response feedforward prediction, configures an optimization objective function in the control law synthesizer that includes minimizing the production process state deviation and maximizing the final product quality index. The specific process includes: inputting the control response feedforward prediction into the control law synthesizer, which internally contains an optimization objective function for evaluating the control response feedforward prediction; the optimization objective function includes a process state deviation term and a product quality evaluation term; the process state deviation term is quantified by calculating the Euclidean distance between the control response feedforward prediction and the ideal process trajectory; the product quality evaluation term uses an offline-trained quality surrogate model to map the control response feedforward prediction into a quality index; and, under the premise of satisfying operational constraints, generating a sequence of actuator coordinated adjustment instructions in real time to drive multiple actuators of the controlled object and adjust the knitting field data stream. An adaptive control correction module establishes an outer-loop adaptive correction circuit, continuously calculates the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected in the knitting field data stream, and quantifies it as identification accuracy. When the identification accuracy exceeds the dynamic tolerance boundary, the internal network parameters of the knitting execution prediction model are corrected and reconstructed.

2. The data-driven adaptive control system for a knitting process according to claim 1, characterized in that, The specific implementation process of collecting knitting field data streams, including yarn tension and loop density, during the knitting process includes: The real-time acquired electrical signals are edge-processed and quantized into yarn tension; image processing algorithms are used to extract feature parameters representing the geometry of the coil from the acquired image information and convert them into coil density; the yarn tension and coil density are timestamped and integrated into a knitting field data stream with a unified data format.

3. The data-driven adaptive control system for a knitting process according to claim 1, characterized in that, Using a recursive learning framework, a knitting execution prediction model is constructed, with a neural network parameterized nonlinear function as its core, representing a multivariable state-space equation. The specific implementation process for generating a control response feedforward prediction for control regulation through multi-step iterative deduction includes: The knitting field data stream is input into a recursive learning framework. A system identification algorithm is used to perform online parameter estimation of the nonlinear and time-varying coupling relationship between the control input and process output in the knitting field data stream, and a knitting execution prediction model is constructed. The current process state reflected by the knitting field data stream is input into the knitting execution prediction model. Through multi-step iterative deduction, the expected state trajectory of the controlled object in multiple future time steps is calculated and encapsulated as a control response feedforward prediction.

4. The data-driven adaptive control system for a knitting process according to claim 1, characterized in that, Under the premise of satisfying operational constraints, the specific implementation process of generating a sequence of actuator coordinated adjustment instructions in real time to drive multiple actuators of the controlled object and adjust the data flow of the knitting field includes: The optimization objective function is iteratively solved to extract the optimal control quantity for the current control cycle, and encapsulates it into an actuator coordinated adjustment instruction sequence containing the target setpoint and execution timing information. The actuator coordinated adjustment instruction sequence is distributed to the underlying drive unit corresponding to each actuator. The underlying drive unit parses the received instructions, converts them into electrical signals, and adjusts the knitting field data stream.

5. The data-driven adaptive control system for a knitting process according to claim 1, characterized in that, The specific implementation process of establishing an outer-loop adaptive correction loop, continuously calculating the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected in the knitting field data stream, and quantifying it into identification accuracy includes: The residuals of the control response feedforward prediction and the knitting field data stream are calculated. The two data streams are strictly aligned based on a unified timestamp, and the vector difference between the predicted value and the actual measured value is calculated point by point to generate a residual sequence that characterizes the degree of model mismatch. The residual sequence is then imported into the accuracy quantization unit, and the residual sequence is converted into continuously changing identification accuracy by integration within a set time window.

6. The data-driven adaptive control system for a knitting process according to claim 1, characterized in that, When the identification accuracy exceeds the dynamic tolerance boundary, the specific implementation process of correcting and reconstructing the internal network parameters of the knitting execution prediction model includes: The identification accuracy is fed into a threshold comparator and continuously compared with an online adjustable dynamic tolerance boundary. Once the identification accuracy is detected to exceed the dynamic tolerance boundary, a reconstruction trigger signal is immediately generated and sent. A dataset of a preset time length is retrieved from the historical data buffer storing the knitting field data stream and the corresponding actuator coordinated adjustment instruction sequence. Based on the retrieved dataset, the internal parameters of the knitting execution prediction model are corrected and reconstructed using the gradient descent method.

7. A data-driven adaptive control method for a knitting process, said method being applied to a data-driven adaptive control system for a knitting process as described in any one of claims 1 to 6, characterized in that, include: Collect knitting field data streams containing yarn tension and loop density during the knitting process; Using a recursive learning framework, a knitting execution prediction model is constructed with a neural network parameterized nonlinear function as the core, representing a multivariable state space equation. Through multi-step iterative deduction, a control response feedforward prediction is generated for control regulation. Based on the control response feedforward prediction, an optimization objective function containing minimizing the production process state deviation and maximizing the final product quality index is configured in the control law synthesizer. Under the premise of satisfying the operating constraints, the actuator coordinated adjustment instruction sequence is generated in real time to drive multiple actuators of the controlled object and adjust the knitting field data flow. An outer-loop adaptive correction loop is established to continuously calculate the residual between the predicted output of the knitting execution prediction model and the actual physical process response reflected by the knitting field data stream, and quantify it as identification accuracy. When the identification accuracy exceeds the dynamic tolerance boundary, the internal network parameters of the knitting execution prediction model are corrected and reconstructed.

Images