Method and system for monitoring technological process of polyester net

Abstract

The invention discloses a technological process monitoring method and system for a polyester net, and belongs to the technical field of intelligent control, and the method comprises the steps: obtaining real-time technological parameter data in the production process of the polyester net; inputting the real-time process parameter data into a digital twin model, predicting the mesh forming quality of the next period, and generating predicted quality parameters; collecting a polyester mesh surface image, inputting the polyester mesh surface image into the semantic segmentation neural network, identifying a mesh abnormal region, and generating an actual mesh quality score; on the basis of a deviation value between the predicted quality parameter and the actual mesh quality score and energy consumption per unit yield, constructing a reward signal and inputting the reward signal into a deep reinforcement learning controller; the deep reinforcement learning controller outputs a process parameter adjustment instruction and sends the process parameter adjustment instruction to an execution mechanism; and feeding back the actual mesh quality score to the digital twin model, correcting the error of the prediction quality parameter, and updating the model synchronization precision. The product quality stability is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control, and particularly relates to a process flow monitoring method and system for polyester mesh. BACKGROUND

[0002] As a high-performance industrial fabric, polyester mesh is widely used in papermaking, filtration, textile and other fields, and its product quality directly affects the forming precision and service life. In recent years, with the development of intelligent manufacturing and industrial internet technology, polyester mesh production lines have gradually upgraded towards automation, continuity and high speed. Some enterprises have introduced online detection systems and PLC control platforms to realize the acquisition of basic parameters and threshold alarm in key process links such as extrusion, drafting and setting, and have initially possessed process monitoring capability. At the same time, the exploratory application of advanced technologies such as digital twinning and artificial intelligence in the field of material processing also provides a new technical path for the intelligent control of the polyester mesh production process.

[0003] However, the existing monitoring method still has significant technical bottlenecks, which are difficult to meet the actual needs of high-quality and high-stability production, mainly in the following three aspects: monitoring lag, lack of prediction ability: the current system is mostly based on real-time parameter over-limit triggering alarm, which belongs to the "after-response" mode, and cannot predict the mesh forming quality trend in advance, resulting in intervention after defects occur, and adjustment lag; quality feedback and process control are disconnected: although some production lines are equipped with machine vision detection systems, the detection results are only used for manual rejudgment or statistical analysis, and do not form a closed loop with the process control system, so the actual mesh quality score cannot be used to drive parameter optimization, resulting in "separation of inspection and control"; control strategy has poor adaptability: traditional PID or empirical rule control cannot cope with dynamic disturbances such as raw material batch fluctuation and equipment aging, and lacks self-learning and self-optimization ability, resulting in blind adjustment of process parameters and large fluctuation of product yield. The above problems jointly restrict the improvement of polyester mesh product consistency and production intelligence level.

[0004] In view of the above problems, the present application proposes a process flow monitoring method for polyester mesh, which simulates and predicts the mesh forming quality of the next process cycle by building a digital twinning model, generates predicted quality parameters, realizes the feedforward perception of process trend, accurately identifies the abnormal area of mesh in the surface image by combining the semantic segmentation neural network, generates the actual mesh quality score, realizes the quantitative feedback of quality state, builds a comprehensive reward signal based on the deviation between prediction and actual value and unit production energy consumption, drives the deep reinforcement learning controller to output the optimal process parameter adjustment instruction, realizes the autonomous optimization of control strategy, and feedbacks the actual quality score to the digital twinning model to dynamically correct the prediction error and continuously improve the model accuracy, forming a whole-process closed-loop monitoring mechanism of "perception-prediction-decision-execution-correction", which effectively solves the core problems of monitoring lag, feedback fragmentation and control rigidity in the prior art. SUMMARY

[0005] In view of the above problems, the present application is proposed.

[0006] To solve the above technical problems, the present application provides the following technical scheme: a process flow monitoring method for polyester mesh, comprising: acquiring real-time process parameter data in the production process of the polyester mesh, wherein the real-time process parameter data comprises extrusion temperature, drawing speed, mesh forming pressure and environmental temperature and humidity; inputting the real-time process parameter data into a digital twin model, predicting mesh forming quality in the next cycle, and generating predicted quality parameters; collecting a surface image of the polyester mesh and inputting it into a semantic segmentation neural network to identify abnormal mesh areas and generate an actual mesh quality score; based on the deviation value of the predicted quality parameters and the actual mesh quality score and the energy consumption per unit output, constructing a reward signal and inputting it into a deep reinforcement learning controller; outputting a process parameter adjustment instruction from the deep reinforcement learning controller and sending it to an actuator; feeding back the actual mesh quality score to the digital twin model to correct the error of the predicted quality parameters and update the model synchronization accuracy.

[0007] As a preferred scheme of the process flow monitoring method for the polyester mesh, the acquisition of the real-time process parameter data in the production process of the polyester mesh comprises: temperature sensors, speed encoders, pressure transmitters and temperature and humidity sensors are respectively arranged at the extrusion section, drawing section, shaping section and winding section of the polyester mesh production line to form a distributed sensing network; a data acquisition module is configured to synchronously acquire output signals of each sensor at a preset sampling frequency, and high-frequency noise and power frequency interference are filtered through a signal conditioning circuit; the conditioned signals are converted into digital data packets through analog-to-digital conversion, and after adding time stamps and process section identifiers, they are transmitted to a central data processing unit through an industrial Ethernet; the digital data packets are subjected to format analysis, abnormal value detection and missing interpolation in the central data processing unit to generate structured real-time process parameter data.

[0008] As a preferred scheme of the process flow monitoring method for the polyester mesh, the step of generating predicted quality parameters comprises: time alignment and normalization processing are performed on the real-time process parameter data to generate a standardized input vector; based on the standardized input vector, a process mode category to which the current operating condition belongs is matched in the digital twin model; Call the multi-physics simulation kernel corresponding to the process mode category to perform mesh morphology evolution simulation of the next process cycle; Extract the mesh size mean value, inter-hole distance coefficient of variation and edge distortion rate from the simulation results to generate the predicted quality parameters.

[0009] As a preferred scheme of the polyester mesh process flow monitoring method, the step of generating the actual mesh quality score comprises: Acquire high-resolution surface images of the running polyester mesh under constant illumination conditions through a line scan camera to generate original image data; Perform denoising, grayscale correction and geometric distortion correction on the original image data to output a standardized image; Input the standardized image into the trained semantic segmentation neural network to output a pixel-level classification mask of the mesh abnormal area; Based on the pixel-level classification mask, the area proportion, density distribution and connected domain number of the abnormal area are counted to generate the actual mesh quality score.

[0010] As a preferred scheme of the polyester mesh process flow monitoring method, the step of constructing the reward signal and inputting it into the deep reinforcement learning controller comprises: Calculate the deviation of each index in the predicted quality parameters from the preset target value to generate a quality deviation vector; Map the quality deviation vector to an equivalent quality loss value, and combine it with the actual mesh quality score to form a quality reward component; Perform sliding average processing on the real-time collected unit production energy consumption to generate a smoothed energy consumption value, and determine an energy consumption penalty component based on the ratio of the smoothed energy consumption value to the baseline energy consumption; Weighted sum the quality reward component and the energy consumption penalty component to generate a comprehensive reward signal, and input the comprehensive reward signal into the deep reinforcement learning controller.

[0011] As a preferred scheme of the polyester mesh process flow monitoring method, the step of outputting the process parameter adjustment instruction comprises: Input the comprehensive reward signal into the trained deep reinforcement learning policy network to generate an original control action vector; Decode the original control action vector in the action space to obtain the extrusion temperature gradient adjustment amount and the draw speed set value adjustment amount; Apply upper and lower limit constraints and change rate limits to the adjustment amounts to generate a safe control instruction; Package the safe control instruction into a standard communication protocol data packet and send it to the corresponding actuator to adjust the process parameters.

[0012] As a preferred scheme of the polyester mesh process monitoring method, wherein: the step of correcting the error of the predicted quality parameter comprises: Calculate the deviation value between the actual mesh quality score and the predicted score generated based on the predicted quality parameter to obtain a quality prediction error; Determine the model key parameters to be adjusted through a preset error-parameter sensitivity matrix; Update the model key parameters using an exponential moving average method to generate corrected model key parameters; Write the corrected model key parameters into a digital twin model configuration file to complete state synchronization accuracy update.

[0013] The application provides a polyester mesh process monitoring system.

[0014] To solve the above technical problems, the application further provides the following technical scheme: a polyester mesh process monitoring system, comprising: a data acquisition module that acquires real-time process parameter data in a polyester mesh production process, wherein the real-time process parameter data includes extrusion temperature, drawing speed, mesh forming pressure and environmental temperature and humidity; A prediction quality module inputs the real-time process parameter data into a digital twin model to predict the next cycle mesh forming quality and generate a predicted quality parameter; A quality score module acquires a polyester mesh surface image and inputs it into a semantic segmentation neural network to identify mesh abnormal areas and generate an actual mesh quality score; A reward generation block constructs a reward signal based on the deviation value between the predicted quality parameter and the actual mesh quality score and unit production energy consumption, and inputs the reward signal into a deep reinforcement learning controller; An instruction generation module outputs process parameter adjustment instructions from the deep reinforcement learning controller and sends them to an execution mechanism; A model correction module is used to feed back the actual mesh quality score to the digital twin model to correct the error of the predicted quality parameter and update the model synchronization accuracy.

[0015] The application provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the polyester mesh process monitoring method when executing the computer program.

[0016] The application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the polyester mesh process monitoring method.

[0017] The application has the beneficial effects that the application realizes intelligentization and self-adaptive control of the polyester mesh production process by constructing a closed-loop monitoring mechanism of "perception-prediction-decision-execution-correction", and significantly improves the accuracy and response ability of process monitoring. Through the feedforward prediction of the mesh forming quality by the digital twin model, the actual mesh quality score is generated by combining the semantic segmentation neural network, realizing the change from "after-alarm" to "trend prediction"; the reward signal is constructed based on the deviation between the prediction and the actual value and the energy consumption index, driving the deep reinforcement learning controller to autonomously output the optimal process parameter adjustment instruction, enhancing the adaptability of the system to dynamic working conditions; and the prediction error of the digital twin model is corrected through the actual quality score feedback, realizing the continuous optimization of the model precision. The overall scheme effectively solves the problems of monitoring lag, disconnection between detection and control, and strategy rigidity in the prior art, improves the product quality stability, production efficiency and yield, reduces the dependence on artificial experience, and has significant technical progress and industrial application value. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 A general flowchart of a polyester mesh process flow monitoring method provided by an embodiment of the application; Figure 2 A quality prediction error correction flowchart of a polyester mesh process flow monitoring method provided by an embodiment of the application; Figure 3 A predicted quality parameter generation flowchart of a polyester mesh process flow monitoring method provided by an embodiment of the application. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor should be within the protection scope of the application.

[0021] Embodiment 1, refer to Figure 1 For the first embodiment of the application, the embodiment provides a polyester mesh process flow monitoring method, which comprises: S1: Obtain real-time process parameter data in the production process of the polyester mesh, the real-time process parameter data including extrusion temperature, drawing speed, mesh forming pressure, and environmental temperature and humidity; S2: Input the real-time process parameter data into a digital twin model to predict mesh forming quality in the next cycle and generate predicted quality parameters; S3: Collect a surface image of the polyester mesh and input it into a semantic segmentation neural network to identify abnormal mesh regions and generate an actual mesh quality score; S4: Based on the deviation value of the predicted quality parameters and the actual mesh quality score and the energy consumption per unit output, construct a reward signal and input it into a deep reinforcement learning controller; S5: Output a process parameter adjustment instruction from the deep reinforcement learning controller and send it to an actuator; S6: Feedback the actual mesh quality score to the digital twin model to correct the error of the predicted quality parameters and update the model synchronization accuracy.

[0022] It should be noted that the polyester mesh involves multiple continuous process links such as melt extrusion, multi-stage drawing, heat setting, and winding in the production process. The process parameters (such as extrusion temperature, drawing speed, and forming pressure) of each link are coupled with each other and are affected by environmental temperature and humidity fluctuations, which easily lead to quality defects such as uneven mesh structure, fracture, or blockage. Due to the fast running speed of the production line and the short process cycle, the traditional monitoring method based on threshold alarm can only achieve post-detection and is difficult to intervene in process deviation in advance. At the same time, the polyester mesh is a continuous flexible material, and its quality state cannot be fully evaluated by offline sampling, and machine vision detection can only provide current quality results, lacking the ability to predict process trends, resulting in control lag and blind adjustment. In addition, the uncertainty introduced by different batches of raw materials, equipment aging, and other factors makes the fixed control strategy less adaptable, and the finished product rate fluctuates greatly. Therefore, how to realize dynamic prediction, intelligent decision-making, and closed-loop optimization of the polyester mesh process has become a key problem to improve product quality stability and production intelligence level.

[0023] Therefore, in order to solve the problems of process monitoring lag, single quality feedback and rigid control strategy, the digital twin model is constructed through steps S1-S6 to simulate and predict the mesh forming quality of the next process cycle, generate predicted quality parameters, and realize the feedforward perception of process trend; the semantic segmentation neural network is used to accurately identify the abnormal area of the mesh in the surface image, generate the actual mesh quality score, and realize the quantitative feedback of the quality state; the reward signal is constructed based on the deviation between the prediction and the actual value and the energy consumption index, and the optimal adjustment instruction is output by the deep reinforcement learning controller to realize the adaptive adjustment of the process parameters; and the actual quality score is fed back to the digital twin model to dynamically correct the prediction error and continuously improve the model accuracy, forming a whole-process closed-loop monitoring mechanism of "perception-prediction-decision-execution-correction"; and then the real-time monitoring of the polyester mesh process state, the early warning and control of the quality abnormality, and the autonomous optimization of the control strategy are realized, and the stability and intelligent level of the production process are significantly improved.

[0024] Embodiment 2, refer to Figures 1-3 The second embodiment of the present application provides a process flow monitoring method for polyester mesh.

[0025] S1: Obtain real-time process parameter data of each process section of the polyester mesh production line, wherein the real-time process parameter data includes extrusion temperature, drawing speed, mesh forming pressure and environmental temperature and humidity; It should be noted that the polyester mesh production line is a continuous, multi-section coupled polymer material forming system, and its process stability is affected by the dynamic parameters of extrusion, drawing, shaping and other links and the external environment. Due to the wide distribution of equipment in each process section, the variety of signal types (analog, digital, high-frequency pulse, etc.), and the fact that some sensors are easily disturbed by high temperature and vibration, there is a risk of delayed, distorted or asynchronous data collection. In addition, key parameters such as extrusion temperature gradient, drawing speed matching, and forming zone pressure uniformity have strong nonlinearity and time-varying characteristics. If only local or discrete sampling is performed, it is difficult to truly reflect the whole-process process state, which will affect the subsequent prediction and control accuracy. Therefore, a high-precision, high-synchronization multi-source data acquisition mechanism needs to be constructed to ensure the integrity, timeliness and traceability of the process parameters, and to provide a reliable input basis for digital twin modeling and intelligent control.

[0026] S11: Deploy temperature sensors, speed encoders, pressure transmitters and temperature and humidity sensors in the extrusion section, drawing section, shaping section and winding section of the polyester mesh production line to form a distributed sensor network; S12: Configure a data acquisition module to synchronously collect the output signals of each sensor at a preset sampling frequency, and filter out high-frequency noise and power interference through a signal conditioning circuit; S13: The conditioned signal is converted into a digital data packet by analog-to-digital conversion, and then transmitted to the central data processing unit through industrial Ethernet after adding a time stamp and process section identifier; S14: The digital data packet is parsed, abnormal value detected and missing interpolated in the central data processing unit to generate structured real-time process parameter data.

[0027] Specifically, to realize the process state perception of the whole polyester mesh production process, multiple types of sensors need to be deployed at key process sections to form a distributed sensor network. In the present embodiment, a platinum resistance temperature sensor is installed at the extrusion section to monitor the melt temperature, a rotary encoder is installed at the drawing section to measure the speed of the traction roller, a piezoresistive pressure transmitter is installed at the setting zone to detect the mesh forming pressure, and a temperature and humidity integrated sensor is arranged at a key position in the workshop to collect environmental parameters. The analog or digital signals output by each sensor are connected to the local data acquisition module through shielded cables to ensure the anti-interference ability of signal transmission. The acquisition module reads sensor data periodically and uploads them in packets to form a structured real-time process parameter data stream, which serves as the basic input for subsequent monitoring and control.

[0028] Specifically, the data acquisition module uses an industrial-grade data acquisition card with time synchronization function, and the sampling frequency is set to 100 Hz to meet the dynamic response requirements of the continuous polyester mesh production process. In the present embodiment, the temperature sensor outputs a 4-20 mA analog signal, which is impedance matched and filtered by the built-in conditioning circuit of the acquisition card to filter out 50 Hz power frequency interference and high-frequency electromagnetic noise; the speed encoder outputs A / B phase pulse signals, which are converted from frequency to speed by the high-speed counting channel of the acquisition card; the pressure transmitter and the temperature and humidity sensor both output digital signals using Modbus RTU protocol, which are connected to the acquisition module through RS-485 bus. The hardware trigger synchronization mechanism is enabled for all acquisition channels to ensure that the data of each process section is collected under the same time reference, avoiding input distortion of the model caused by time sequence misalignment.

[0029] Specifically, the timestamp is generated by the IEEE 1588 Precision Time Protocol (PTP) time signal provided by the industrial Ethernet switch, with an accuracy of ±1 μs. In this embodiment, each frame of digital data packet contains fields: process section identification (such as "extrusion section" "drafting section"), parameter type identification (such as "extrusion temperature" "drafting speed"), numerical value (calibrated engineering unit value), and acquisition timestamp. The data packet is encapsulated into a standard TCP / IP protocol frame and transmitted to the central data processing unit through the redundant ring network structure of the industrial Ethernet. The central data processing unit runs a data receiving service program, parses the received data packet, and stores it in a real-time database in chronological order to form the real-time process parameter data. For missing or abnormal data points (such as numerical value jump exceeding ±3σ), linear interpolation or Kalman filtering is used for repair to ensure data continuity and reliability.

[0030] Specifically, the platinum resistance temperature sensor is Pt100, installed near the extruder die, with a measurement range of 0-400℃ and an accuracy level of A class (±(0.15+0.002|t|)℃); the rotary encoder is an incremental optical encoder with a resolution of 1024 PPR, installed at the end of the three-stage drafting roller shaft, and the actual linear speed is obtained through gear transmission ratio conversion; the piezoresistive pressure transmitter has a range of 0-10 MPa and an output signal of 4-20 mA, installed in the hydraulic circuit of the heat setting zone pressure roller; the temperature and humidity sensor uses a capacitive sensing element, with a temperature measurement range of -10-50℃ and a humidity range of 0-100% RH, installed at a ventilation position 1.5 m above the production line. All sensors are calibrated by the national legal metrology institution, with a calibration period of 6 months.

[0031] S2: inputting the real-time process parameter data into a pre-constructed digital twin model, predicting the mesh forming quality of the next process cycle through the digital twin model, and generating a predicted quality parameter; It should be noted that the mesh forming quality of the polyester mesh is affected by the coupling of multiple physical fields, involving complex nonlinear processes such as melt flowability, cooling rate, tension distribution, and thermal shrinkage behavior. Traditional empirical models are difficult to accurately describe the dynamic evolution law. Although the digital twin model has simulation capability, if the original process parameters are directly input, it is easy to cause prediction deviation due to data noise, time delay or model mismatch. In addition, different production batches of raw materials have different characteristics (such as molecular weight distribution, additive content), and if the model is not adapted to the working condition, the prediction result will deviate from the actual value. Therefore, the real-time process parameter data needs to be preprocessed and matched with the working condition at the input stage, and the digital twin model is driven by a dynamic simulation engine to output the mesh forming quality prediction result with spatiotemporal continuity, ensuring that the predicted quality parameter has engineering usability.

[0032] S21: time alignment and normalization processing is performed on the real-time process parameter data to generate a standardized input vector; S22: based on the standardized input vector, a process mode category to which a current running condition belongs is matched in a digital twin model; S23: a multi-physical field simulation kernel corresponding to the process mode category is called to perform mesh morphology evolution simulation of a next process cycle; S24: a mesh size mean value, a hole spacing variation coefficient and an edge distortion rate are extracted from a simulation result to generate a predicted quality parameter.

[0033] Specifically, to ensure that the input data of the digital twin model has consistency and timeliness, the real-time process parameter data needs to be preprocessed. In this embodiment, a time synchronization interpolation algorithm is used to resample the data collected by each process section, unified to a 10ms time step, and an input sequence containing the last 1 second of historical data is constructed in a sliding window manner; then the extrusion temperature, the drawing speed, the mesh forming pressure and the environmental temperature and humidity are respectively subjected to Min-Max normalization processing to eliminate the dimension difference and generate a standardized input vector. The vector is used as the driving input of the digital twin model, ensuring that the simulation starting point is consistent with the actual working condition and improving the accuracy of the prediction initial state.

[0034] Specifically, the digital twin model is built-in with a process mode classifier for identifying the process category to which the current running state belongs. In this embodiment, the process mode category includes "normal production mode", "high-speed running mode" and "material changing transition mode", and the classification basis is the combined features of the drawing speed and the extrusion temperature in the standardized input vector. The classifier uses a trained support vector machine (SVM) model, and the training samples come from the stable running section labeled in the historical production data. The classification result is used to select the corresponding multi-physical field simulation kernel parameter set, ensuring that the model simulation logic matches the current actual process and avoiding prediction distortion caused by mode mismatch.

[0035] Specifically, the multi-physical field simulation kernel is built based on the finite element method and includes three coupled physical fields of heat conduction, viscoelastic fluid flow and structural stress. In this embodiment, the simulation time step is set to 1ms, and the total simulation time covers a 2-second interval of the next process cycle; the initial boundary conditions are determined by the standardized input vector, in which the extrusion temperature is set as the initial temperature field of the die area, the drawing speed is set as the boundary displacement loading rate, and the mesh forming pressure is set as the contact pressure load of the pressure roller. The material constitutive model uses a generalized Maxwell model to describe the viscoelastic behavior of the polyester melt, and the convective heat transfer coefficient is set by the environmental temperature and humidity during the cooling process. The strain rate and the hole evolution path of the grid nodes are calculated in real time during the simulation process, which are used for subsequent mesh morphology extraction.

[0036] Specifically, the mesh morphology data extracted from the simulation results is analyzed by the post-processing module to generate structured prediction quality parameters. In this embodiment, the system binarizes and skeletonizes the mesh structure image of the simulation final frame, calculates the mean of the equivalent diameters of all meshes in the region as the "mesh size average", calculates the ratio of the standard deviation of the inter-pore distance to the average distance as the "inter-pore distance coefficient of variation", and identifies the local stretching distortion area through edge curvature analysis, and calculates the area ratio as the "edge distortion rate". The three indicators together constitute the prediction quality parameters, which are packaged as numerical vectors in a fixed field order and output to the subsequent control module for bias calculation and reward signal construction.

[0037] S3: Collecting the polyester mesh surface image and inputting it into the trained semantic segmentation neural network to identify the mesh abnormal area and generate the actual mesh quality score; It should be noted that the polyester mesh is a continuous fabric structure with high light transmittance and strong texture repeatability. Its surface image is easily affected by factors such as uneven light, motion blur, and background interference, making it difficult for traditional image processing methods (such as threshold segmentation and edge detection) to accurately identify the mesh abnormal area. In addition, the mesh abnormal types are diverse, including breakage, blockage, fusion, deformation, etc., with large morphological differences and blurred boundaries. If only relying on artificial design features, the generalization ability is poor. Therefore, a semantic segmentation model based on deep learning is needed to accurately identify the abnormal area through pixel-level classification and further quantify it into an "actual mesh quality score" that can participate in control decisions, ensuring that the detection results are comparable and engineering usable.

[0038] S31: Collecting high-resolution surface images of the running polyester mesh under constant light conditions through a line scan camera to generate raw image data; S32: Denoising, grayscale correction, and geometric distortion correction are performed on the raw image data to output a standardized image; S33: Inputting the standardized image into the trained semantic segmentation neural network to output a pixel-level classification mask of the mesh abnormal area; S34: Based on the pixel-level classification mask, the area ratio, density distribution, and connected component number of the abnormal area are calculated to generate the actual mesh quality score.

[0039] Specifically, to adapt to the characteristics of high-speed continuous operation of the polyester mesh, a line scan camera is used in combination with an encoder triggering mode for image acquisition. In this embodiment, the line scan camera is installed above the stable transmission zone before the winding section, and the line scan frequency is synchronized with the running speed of the mesh belt through the rotary encoder signal to ensure that the image is not stretched and distorted. The light source is a high-uniformity LED line light source arranged on the opposite side of the camera to form a transmission illumination mode, enhancing the contrast of the mesh hole edges. The acquired image is stored in grayscale format with a resolution of 4096 pixels / line and a line frequency of 2000 Hz, covering the full-width image data to form the original image data as the basis for subsequent analysis.

[0040] Specifically, after the original image data is transmitted to the image processing unit, it is sequentially subjected to non-local mean denoising, illumination field grayscale correction, and lens geometric distortion correction. In this embodiment, the denoising process uses the Non-Local Means algorithm with a search window set to 21x21 to filter out random noise while preserving edge details. The grayscale correction builds an uneven illumination model using a pre-acquired blank mesh area reference image and performs pixel-by-pixel compensation. The geometric distortion correction uses camera calibration parameters based on a calibration board and uses a bilinear interpolation method to correct the image in reverse mapping. After processing, the standardized image with consistent spatial resolution and uniform brightness is generated, meeting the input quality requirements of the semantic segmentation model.

[0041] Specifically, the semantic segmentation neural network is a U-Net model based on an encoder-decoder structure, with the encoder using a ResNet-34 backbone network and the decoder recovering spatial details through upsampling and jump connection. In this embodiment, the model input size is 512x512 pixels, and the output is a pixel-level classification mask with the same resolution as the input, where 0 represents a normal mesh area and 1 represents a mesh anomaly area (including breakage, blockage, fusion, etc.). The model is trained on a training set containing 10,000 labeled images, and the labeling is verified by multiple experts to ensure accuracy. In the inference stage, TensorRT is used for acceleration, and the single-frame processing time on the edge computing device is less than 50 ms, meeting the real-time requirements.

[0042] Specifically, the pixel-level classification mask is analyzed by post-processing to generate a structured quality index. In this embodiment, the system calculates the ratio of the total number of abnormal area pixels to the total number of detected area pixels as the "abnormal area ratio"; counts the number of independent abnormal connected domains within a unit area as the "abnormal density"; and extracts the equivalent diameter of the largest connected abnormal area through morphological analysis as the "maximum defect size". The three indicators are normalized by weighting (weights are 0.5, 0.3, and 0.2, respectively), combined into a single numerical value, and the range is 0-100 points. The lower the value, the worse the quality. This value is the actual mesh quality score. The score result is stored in the real-time database after adding a timestamp, and is simultaneously output to the control module for subsequent deviation analysis and feedback adjustment.

[0043] S4: Based on the deviation value of the predicted quality parameter and the actual mesh quality score and the energy consumption per unit output, an reward signal is constructed and input into a deep reinforcement learning controller; It should be noted that the decision quality of the deep reinforcement learning controller is highly dependent on the rationality of the reward signal design. If the reward signal is based on a single indicator (such as quality deviation or energy consumption), it is easy to cause the control strategy to deviate, resulting in non-optimal behavior such as "emphasizing energy saving and ignoring quality" or "emphasizing quality and being uneconomical". In addition, the predicted quality parameter is a multi-dimensional vector (such as the average mesh size, the coefficient of variation of the hole pitch, and the edge distortion rate), while the actual mesh quality score is a scalar. There is a dimension mismatch problem between the two, and direct calculation of the deviation may lose information. At the same time, the energy consumption per unit output is affected by the noise of instantaneous load fluctuations, and if it is directly involved in the reward calculation, it may cause control oscillation. Therefore, it is necessary to normalize, weight, and dynamically adjust the multi-source indicators to construct a stable, balanced, and drivable reward signal that can optimize the strategy, and ensure that the process parameter adjustment instructions output by the deep reinforcement learning controller have both quality stability and operational economy.

[0044] To address the rationality and robustness of the reward signal, S4 specifically includes the following sub-steps: S41: Calculate the deviation of each indicator in the predicted quality parameter from the preset target value to generate a quality deviation vector; S42: Map the quality deviation vector to an equivalent quality loss value, and together with the actual mesh quality score to form a quality reward component; S43: Perform a sliding average on the real-time collected energy consumption per unit output to generate a smoothed energy consumption value, and determine an energy consumption penalty component based on the ratio of the smoothed energy consumption value to the baseline energy consumption; S44: Weight and sum the quality reward component and the energy consumption penalty component to generate a comprehensive reward signal, and input the comprehensive reward signal into the deep reinforcement learning controller.

[0045] Specifically, the prediction quality parameter includes three indicators: mesh size average, pore spacing coefficient of variation, and edge distortion rate, and the preset target values thereof are 1.2 mm, 8%, and 3%, respectively. In this embodiment, the normalized Euclidean distance method is used to calculate the deviations in S41: the size deviation is |actual average -1.2, the coefficient of variation deviation is |actual coefficient of variation -8%, and the distortion rate deviation is |actual distortion rate -3%. The three deviations constitute a three-dimensional quality deviation vector, which is used as the basis input for subsequent mapping to ensure that the contributions of each indicator are interpretable and adjustable.

[0046] Further, the equivalent quality loss value is calculated by a weighted sum of squares. In this embodiment, each deviation in the quality deviation vector obtained in S41 is squared and multiplied by the corresponding weight (size deviation weight 0.4, coefficient of variation weight 0.35, and distortion rate weight 0.25), and the sum is taken as the quality loss term. At the same time, the actual mesh quality score is normalized to the interval [0, 1] as the measured quality feedback term. The two terms are weighted by a ratio of 0.6:0.4 to generate a quality reward component, which ranges from , and the closer to 0 indicates better quality.

[0047] Further, the unit production energy consumption is calculated by combining an electric energy meter with a production encoder, and is defined as the electric energy (kWh / m) consumed per unit length of polyester mesh produced. In this embodiment, the length-weighted moving average method is used to process the energy consumption data, with a window length of 100 meters to avoid the influence of transient fluctuations. The reference energy consumption is set as the historical average value under the current process mode; and the energy consumption penalty component is calculated by the formula: the ratio of the smoothed energy consumption to the reference energy consumption is used as the comparison reference, and the result obtained by subtracting the ratio from 1 is the energy consumption penalty component. Specifically, when the smoothed energy consumption is lower than the reference energy consumption, the calculation result is positive, indicating energy saving, which is used as a positive reward; and when the smoothed energy consumption is higher than the reference energy consumption, the calculation result is negative, indicating energy consumption exceeding the standard, which is used as a negative punishment. Through this calculation method, energy-saving behavior can be positively encouraged, and high-energy-consumption behavior can be inhibited, thereby guiding the control system to optimize energy consumption under the premise of ensuring quality.

[0048] Specifically, the comprehensive reward signal is generated by linear weighting. In this embodiment, the quality reward component is multiplied by a weight of 0.7, and the energy consumption penalty component is multiplied by a weight of 0.3, and the two are added to obtain the comprehensive reward signal. The value range of the quality reward component is negative one to zero, indicating that the greater the quality deviation, the lower the reward value; the value range of the energy consumption penalty component is negative zero point five to positive zero point five, which is positive reward when saving energy and negative penalty when over-consumption. The comprehensive reward signal is standardized after generation to ensure stable numerical range, and then transmitted to the deep reinforcement learning controller deployed on the edge server through the MQTT communication protocol. The generation cycle of the signal is synchronized with the process control cycle, and the signal is updated and sent once every 2 seconds to ensure that the controller can learn and optimize the strategy based on stable and periodic input reward signals.

[0049] S5: outputting a process parameter adjustment instruction by the deep reinforcement learning controller according to the reward signal, and sending the process parameter adjustment instruction to an actuator; It should be noted that the polyester mesh production process has the characteristics of strong nonlinearity, multivariable coupling and response delay, and the traditional PID control is difficult to adapt to the dynamic working condition change. Although the deep reinforcement learning controller has autonomous decision-making ability, if the strategy output lacks safety constraints or the action space design is unreasonable, it may lead to dramatic fluctuations in control instructions, causing equipment overload or quality out of control. In addition, the extrusion temperature gradient and the drawing speed set value are key process parameters, and their adjustment needs to consider the response capability of the actuator (such as the heating zone temperature rise rate, motor acceleration and deceleration time), to avoid the instruction exceeding the physical limit. Therefore, action boundary restrictions, smooth transition mechanisms and safety verification processes need to be introduced in the strategy reasoning process to ensure that the process parameter adjustment instruction meets the optimization goal and conforms to the actual operation constraints of the production line.

[0050] To solve the above control safety and engineering adaptability problems, S5 specifically includes the following sub-steps: S51: inputting the reward signal into the trained deep reinforcement learning strategy network to generate an original control action vector; S52: decoding the original control action vector to obtain an extrusion temperature gradient adjustment amount and a drawing speed set value adjustment amount; S53: applying upper and lower limit constraints and change rate restrictions to the adjustment amounts to generate a safe control instruction; S54: packaging the safe control instruction into a standard communication protocol data packet and sending it to the corresponding actuator to adjust the process parameters.

[0051] Specifically, the deep reinforcement learning policy network adopts a TD3 (Twin Delayed Deep Deterministic Policy Gradient) algorithm architecture, the input of which is a spliced vector containing the reward signal sequence of the last 5 cycles and the process state features, and the output of which is a two-dimensional original control action vector corresponding to the "extrusion temperature gradient adjustment direction and amplitude" and the "draft speed set value adjustment direction and amplitude". In this embodiment, the network inference runs on an edge computing device equipped with an NVIDIA Jetson AGX Xavier, and the time consumption of a single inference is less than 50 ms, meeting the control cycle requirement of once every 2 seconds.

[0052] Further, the action space decoding process is implemented based on a preset mapping relationship table. In this embodiment, the value range of each component of the original control action vector is , which is linearly mapped to the extrusion temperature gradient adjustment amount , and the draft speed set value adjustment amount , respectively. For example, when the model output is , the decoding obtains an extrusion temperature gradient increase of 3.0 °C / m and a draft speed decrease of 0.4 m / min. The mapping relationship is stored in the controller configuration file, supporting flexible switching under different product specifications.

[0053] Further, the upper and lower limit constraints include static boundaries and dynamic change rate limits. In this embodiment, the extrusion temperature gradient adjustment amount is limited within ±5 °C / m, and the adjustment amount between adjacent two control cycles does not change by more than ±1 °C / m; the draft speed set value adjustment amount is limited within ±1.0 m / min, and the change rate does not exceed ±0.5 m / min / s. If the decoded adjustment amount exceeds the limit value, it is truncated to the boundary value; if the change rate is out of limit, it is executed step by step by interpolation. This process is completed by the safety logic module built-in the controller, ensuring smooth transition of the instructions.

[0054] Specifically, the safety control instruction is packaged as a Modbus TCP protocol data packet, including a function code, a register address, a numerical value, and a CRC check field. In this embodiment, the extrusion temperature gradient adjustment instruction is written to the D4010 register of the PLC, and the draft speed set value adjustment instruction is written to the D4020 register, which is further distributed to the heating zone temperature control module and the frequency converter driving unit by the PLC. The communication period is 2 seconds, synchronized with the control period; if communication fails for three consecutive times, the system automatically switches to the backup communication link and triggers an alarm signal to prevent control interruption. After receiving the instruction, the actuator gradually adjusts the heating power or motor speed to realize closed-loop adjustment of the process parameters.

[0055] S6: feeding the actual mesh quality score to the digital twin model, correcting the error of the predicted quality parameter by using the actual mesh quality score, and updating the state synchronization accuracy of the digital twin model.

[0056] It should be noted that the digital twin model will gradually deviate from the actual state due to factors such as material property drift, equipment aging or environmental disturbance during long-term operation. If dynamic correction is not performed, the prediction ability will degrade, affecting the reliability of the entire monitoring system. However, directly using the actual mesh quality score as the only basis for model parameter inversion is easily disturbed by single measurement noise or abnormal working conditions, leading to model mis-correction. In addition, the predicted quality parameter is a multi-dimensional index, while the actual mesh quality score is a scalar, and there is a dimension mismatch problem between the two, which requires a mapping mechanism from scalar feedback to multi-dimensional parameter correction. Therefore, an online correction strategy based on error memory and weighted update needs to be constructed to gradually improve the model accuracy and avoid oscillation or overfitting.

[0057] To solve the above problems of model degradation and correction stability, S6 specifically includes the following sub-steps: S61: calculating the deviation value between the actual mesh quality score and the predicted score generated based on the predicted quality parameter to obtain a quality prediction error; S62: inputting the quality prediction error into a model correction module to determine the key model parameters that need to be adjusted through a pre-set error-parameter sensitivity matrix; S63: updating the key model parameters using an exponential moving average method to generate corrected model parameters; S64: writing the corrected model parameters into a digital twin model configuration file to complete dynamic updating of the state synchronization accuracy.

[0058] Specifically, the predicted score is an equivalent score value calculated by the same weighting formula in S3 according to the mesh size average, the inter-pore distance coefficient of variation and the edge distortion rate in the predicted quality parameter. In this embodiment, the predicted score calculation formula is as follows:

[0059] wherein, is the predicted score generated by the predicted quality parameter, is the normalized deviation of the mesh size average relative to the target value, is the normalized deviation of the inter-pore distance coefficient of variation relative to the target value, is the normalized deviation of the edge distortion rate relative to the target value.

[0060] The difference between the predicted score and the actual mesh quality score is the quality prediction error, which is used as the input signal for model correction, and the formula is as follows:

[0061] wherein, is the mass prediction error, is the actual mesh quality score, used to drive the subsequent model correction process after S6. If <0 indicates that the model overestimates the quality (prediction is too optimistic), and the model performance expectation needs to be adjusted or the physical parameters need to be corrected.

[0062] Specifically, the error-parameter sensitivity matrix is obtained by historical data regression analysis. In this embodiment, the matrix contains the response coefficients of three key model parameters: when the mass prediction error is consistently negative (prediction is too high), it indicates that the model underestimates the melt cooling rate, and the thermal conductivity coefficient should be adjusted higher; when the error fluctuates greatly, it indicates that the viscoelasticity parameters are not matched, and the Maxwell relaxation time constant needs to be adjusted; when the error shows a downward trend, the die pressure loss coefficient should be corrected. The matrix is stored in the model correction module in the form of a lookup table, supporting automatic matching of parameter adjustment direction according to the error characteristics.

[0063] Specifically, the update of the key parameters of the model uses the exponential moving average algorithm, and the update formula is:

[0064] wherein, is the new model parameter value after correction, is the original parameter value stored in the current digital twin model, is the target correction value obtained by inverting the current mass prediction error, is a smoothing coefficient. In this embodiment, parameter update is triggered only when the mass prediction error exceeds ±5 points for three consecutive periods, avoiding frequent disturbances; the update amplitude is limited to within ±10% each time, ensuring smooth evolution of the model.

[0065] Specifically, the corrected model parameters are written into the runtime configuration area of the digital twin model through the OPC UA protocol. In this embodiment, the digital twin model is deployed on an industrial edge server, and its core simulation engine is realized by secondary development of COMSOL Multiphysics, and the parameter interface is packaged as a standard UA variable node. After completing the parameter calculation, the model correction module writes new values to the corresponding node through a secure session connection and triggers a model reinitialization instruction. After the update is completed, the system records the correction time, parameter change amount, and operation log for subsequent tracing and auditing, ensuring that the model update process is monitorable and verifiable.

[0066] Embodiment 3, as a third embodiment of the present application, provides a process flow monitoring system of a polyester mesh, comprising: a data acquisition module, acquiring real-time process parameter data in a polyester mesh production process, wherein the real-time process parameter data comprises extrusion temperature, drawing speed, mesh forming pressure and environmental temperature and humidity; a prediction quality module, inputting the real-time process parameter data into a digital twin model, predicting mesh forming quality in a next cycle, and generating a predicted quality parameter; a quality score module, acquiring a surface image of the polyester mesh and inputting the surface image into a semantic segmentation neural network, identifying an abnormal area of the mesh, and generating an actual mesh quality score; a reward generation block, constructing a reward signal based on a deviation value of the predicted quality parameter and the actual mesh quality score and unit production energy consumption, and inputting the reward signal into a deep reinforcement learning controller; an instruction generation module, outputting process parameter adjustment instructions from the deep reinforcement learning controller and sending the process parameter adjustment instructions to an actuator; a model correction module, used for feeding back the actual mesh quality score to the digital twin model, correcting an error of the predicted quality parameter, and updating model synchronization accuracy Embodiment 4, as a fourth embodiment of the present application, is different from the first three embodiments in that: if the function is realized in the form of a software function unit and sold or used as an independent product, the function can be stored in a computer readable storage medium. Based on this understanding, the technical solution of the present application or the part that contributes to the prior art or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0067] The logic and / or steps represented in the flow diagrams or otherwise described herein, for example, can be considered as a sequence of executable instructions, and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. In the context of this specification, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, device, or propagation medium.

[0068] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example via an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in a form that can be later executed by a computer.

[0069] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, known in the art, or combinations thereof, can be used: a discrete logic circuit having logic gates for implementing logic functions upon data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), or the like.

[0070] It should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application but not limit the present application, and although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalent replaced without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A method for monitoring the process flow of polyester web, characterized in that: include, Acquire real-time process parameter data during the production of polyester mesh, including extrusion temperature, drawing speed, mesh forming pressure, and ambient temperature and humidity; The real-time process parameter data is input into the digital twin model to predict the mesh forming quality of the next cycle and generate predicted quality parameters. Acquire images of the polyester mesh surface and input them into a semantic segmentation neural network to identify abnormal mesh areas and generate actual mesh quality scores; Based on the deviation between the predicted quality parameters and the actual mesh quality score, and the energy consumption per unit output, a reward signal is constructed and input into the deep reinforcement learning controller. The deep reinforcement learning controller outputs process parameter adjustment instructions and sends them to the actuator; The actual mesh quality score is fed back to the digital twin model to correct the error in the predicted quality parameters and update the model synchronization accuracy.

2. The process flow monitoring method for polyester web as described in claim 1, characterized in that: The acquisition of real-time process parameter data during the polyester web production process includes: Temperature sensors, speed encoders, pressure transmitters, and temperature and humidity sensors are deployed in the extrusion section, drawing section, shaping section, and winding section of the polyester web production line to form a distributed sensor network. The data acquisition module is configured to synchronously acquire the output signals of each sensor at a preset sampling frequency, and the signal conditioning circuit filters out high-frequency noise and power frequency interference. The conditioned signal is converted from analog to digital into digital data packets, and after being appended with a timestamp and process segment identifier, it is transmitted to the central data processing unit via industrial Ethernet. The central data processing unit performs format parsing, outlier detection, and missing value imputation on the digital data packets to generate structured real-time process parameter data.

3. The process flow monitoring method for polyester web as described in claim 2, characterized in that: The steps for generating the predicted quality parameters include: The real-time process parameter data is time-aligned and normalized to generate a standardized input vector; Based on the standardized input vector, the process mode category to which the current operating condition belongs is matched in the digital twin model; Invoke the multiphysics simulation kernel corresponding to the process mode category to perform the mesh morphology evolution simulation for the next process cycle; The mean mesh size, the coefficient of variation of the mesh spacing, and the edge distortion rate are extracted from the simulation results to generate predicted quality parameters.

4. The process flow monitoring method for polyester web as described in claim 3, characterized in that: The steps for generating actual mesh quality scores include: High-resolution surface images of the polyester web in operation are acquired using a line scan camera under constant illumination conditions to generate raw image data. The original image data is denoised, grayscale corrected, and geometric distortion corrected to output a standardized image; The standardized image is input into the trained semantic segmentation neural network, which outputs a pixel-level classification mask for mesh anomaly regions. Based on the pixel-level classification mask, the area ratio, density distribution, and number of connected components of abnormal regions are statistically analyzed to generate an actual mesh quality score.

5. The process flow monitoring method for polyester web as described in claim 4, characterized in that: The steps of constructing the reward signal and inputting it into the deep reinforcement learning controller include: Calculate the degree of deviation between each indicator in the predicted quality parameters and the preset target value, and generate a quality deviation vector; The quality deviation vector is mapped to an equivalent quality loss value, and together with the actual mesh quality score, it constitutes a quality reward component. The energy consumption per unit output collected in real time is processed by moving average to generate a smoothed energy consumption value, and the energy consumption penalty component is determined based on the ratio of the smoothed energy consumption to the baseline energy consumption. The quality reward component and the energy consumption penalty component are weighted and summed to generate a comprehensive reward signal, which is then input into the deep reinforcement learning controller.

6. The process flow monitoring method for polyester web as described in claim 5, characterized in that: The steps for issuing the process parameter adjustment command include: The comprehensive reward signal is input into the trained deep reinforcement learning policy network to generate the original control action vector; The original control action vector is decoded in action space to obtain the extrusion temperature gradient adjustment amount and the stretching speed setpoint adjustment amount; Apply upper and lower limit constraints and a rate of change limit to the adjustment amount to generate a safety control command; The safety control commands are encapsulated into standard communication protocol data packets and sent to the corresponding actuators to adjust process parameters.

7. The process flow monitoring method for polyester web as described in claim 6, characterized in that: The steps for correcting the error in the predicted quality parameters include: The deviation between the actual mesh quality score and the predicted score generated based on the predicted quality parameters is calculated to obtain the quality prediction error. The key parameters of the model that need to be adjusted are determined by using a preset error-parameter sensitivity matrix; The key parameters of the model are updated using an exponential moving average method to generate corrected key parameters of the model; The corrected key parameters of the model are written into the digital twin model configuration file to complete the state synchronization accuracy update.

8. A process flow monitoring system for polyester web, employing the process flow monitoring method for polyester web as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module acquires real-time process parameter data during the polyester mesh production process. The real-time process parameter data includes extrusion temperature, drawing speed, mesh forming pressure, and ambient temperature and humidity. The prediction quality module inputs the real-time process parameter data into the digital twin model to predict the mesh forming quality of the next cycle and generate prediction quality parameters. The quality scoring module collects images of the polyester mesh surface and inputs them into a semantic segmentation neural network to identify abnormal areas of the mesh and generate actual mesh quality scores. The reward generation block constructs a reward signal and inputs it into the deep reinforcement learning controller based on the deviation between the predicted quality parameters and the actual mesh quality score and the energy consumption per unit output. The instruction generation module outputs process parameter adjustment instructions from the deep reinforcement learning controller and sends them to the actuator; The model calibration module is used to feed back the actual mesh quality score to the digital twin model, correct the error of the predicted quality parameters, and update the model synchronization accuracy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the process flow monitoring method for polyester web according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the process flow monitoring method for polyester web according to any one of claims 1 to 7.

Images