A digital twin driven intelligent regulation method and system for carbon fiber spinning oil

Abstract

The application discloses a kind of digital twin driven carbon fiber spinning oil intelligent regulation and control method and system, belong to artificial intelligence and intelligent manufacturing field.The method is by real-time acquisition nozzle multi-source sensing data and high-speed spray image, drive high-fidelity CFD digital twin model simulation internal flow field, and combine image processing extracted external spray feature, utilize pre-trained neural network model to predict the risk index of blockage.System according to risk level automatically triggers different cleaning procedures, including preventive ultrasonic cleaning and enhanced thermal-mechanical synergistic cleaning, and after cleaning, by reacquiring data to verify effect and calibrate model, form the closed loop of perception-simulation-decision-execution-verification.The application realizes early warning, accurate evaluation and active intervention to nozzle blockage, effectively guarantees oil agent spraying uniformity and the continuous stability of carbon fiber production.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and intelligent manufacturing, specifically relating to a digital twin-driven intelligent control method and system for carbon fiber spinning oil. Background Technology

[0002] In the preparation of carbon fiber, the spinning precursor yarn needs to be treated with an oiling agent to improve its bundle properties, antistatic ability, and uniformity of subsequent heat treatment. Usually, a precision nozzle is used to atomize and spray the oiling agent onto the surface of the high-speed running precursor yarn to form a continuous oil film with controllable thickness. However, during long-term continuous operation, the oiling agent is prone to blockage of the microchannels inside the nozzle due to factors such as ambient temperature fluctuations, solvent evaporation, and particle deposition. This leads to unstable spray flow and uneven droplet distribution, which seriously affects the mechanical properties of the fiber and the consistency of the product in the subsequent carbonization process.

[0003] Currently, the monitoring and maintenance of nozzle blockage mainly rely on periodic manual cleaning or delayed alarms based on abnormal pipeline pressure and flow. Although existing technologies can identify some blockage situations, they are difficult to distinguish the type of blockage, and usually only trigger a response when the blockage has significantly affected production, lacking the ability to perceive early signs of blockage. In addition, existing digital twin technology in spinning process monitoring is mostly used for parameter visualization and historical backtracking, failing to be deeply integrated with real-time sensing data, and is also difficult to achieve high-fidelity flow field simulation and closed-loop control, thus failing to meet the needs for real-time early warning and autonomous intervention of microscale flow behavior.

[0004] To address the aforementioned technical problems, this invention proposes a control method that integrates multi-source real-time sensing, high-precision flow field simulation, and intelligent decision-making to achieve early identification, severity assessment, and proactive intervention of nozzle clogging risks, thereby ensuring continuous and stable carbon fiber production and consistent product quality. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a digital twin-driven intelligent control method and system for carbon fiber spinning oil, employing the following technical solution.

[0006] Firstly, a digital twin-driven intelligent control method for carbon fiber spinning oil includes the following steps:

[0007] Acquire multi-source sensor data reflecting the nozzle's operating status and a sequence of spray pattern images at the nozzle outlet;

[0008] Based on the multi-source sensor data and the geometric parameters of the nozzle, the computational fluid dynamics digital twin model is driven to perform real-time simulation and extract feature vectors characterizing the internal flow state.

[0009] The spray morphology image sequence is processed to extract feature vectors that characterize the external morphology and dynamic properties of the spray;

[0010] The internal flow state feature vector and the external spray characterization feature vector are input into a pre-trained clogging risk prediction neural network model, which outputs a clogging risk index.

[0011] Based on the comparison result between the blockage risk index and the preset threshold, the corresponding level of cleaning program is automatically triggered and executed. The cleaning program includes at least preventive cleaning and enhanced cleaning.

[0012] After the cleaning procedure is performed, data is collected again to update the state of the digital twin model, and the congestion risk index is recalculated based on the updated model and data to verify the cleaning effect.

[0013] Preferably, the steps for acquiring multi-source sensor data include:

[0014] Collect the nozzle inlet pressure, outlet pressure, oil temperature, and environmental parameters;

[0015] It also performs time-series synchronization and filtering on multi-source data;

[0016] The spray pattern image sequence is acquired through a high-speed backlight imaging system. The acquisition timing includes periodic monitoring acquisition, event-triggered acquisition based on abnormal sensor data or risk index approaching the threshold, and verification acquisition after cleaning.

[0017] Preferably, the steps for driving the computational fluid dynamics digital twin model to perform real-time simulation include:

[0018] A three-dimensional model of the internal flow channel was established based on the nozzle geometry, and a computational mesh was generated.

[0019] The real-time collected inlet and outlet pressure data are used as the dynamic boundary conditions of the model.

[0020] Flow field simulation was performed using a transient solver.

[0021] The internal flow state feature vector, including the average flow velocity at the throat section, wall shear stress, and pressure fluctuation characteristics, is extracted from the simulation results according to a set period.

[0022] Preferably, the step of processing the spray pattern image sequence includes:

[0023] Background subtraction preprocessing is performed on the image to separate moving droplets; the spray cone angle is calculated based on the processed image;

[0024] Particle image velocimetry was used to obtain the droplet velocity vector field and calculate dynamic parameters.

[0025] Identify droplets and statistically analyze their size distribution;

[0026] And calculate the spatial distribution of spray flux density based on droplet position and velocity information;

[0027] Finally, the obtained quantitative indicators are integrated to form a feature vector representing the external characteristics of the spray.

[0028] Preferably, the congestion risk prediction neural network model is a hybrid architecture model that processes high-dimensional features and captures temporal trends; it is trained using a laboratory calibration dataset, the construction of which includes:

[0029] Different degrees of blockage were simulated in the experimental environment. Multi-source sensor data and spray images were collected simultaneously to generate feature vector samples. A quantitative benchmark risk index was assigned to each sample based on the deviation of pressure drop rate and spray morphology index.

[0030] Preferably, the preventative cleaning procedure includes:

[0031] Inject a special cleaning solution into the nozzle that is compatible with the composition of the working oil and contains chelating agents;

[0032] Ultrasonic waves are applied to the nozzle in pulsed operating mode to loosen deposits using cavitation effect;

[0033] The flow rate of the cleaning fluid was controlled at a level close to the normal operating flow rate.

[0034] Preferably, the enhanced cleaning procedure includes:

[0035] First, perform a basic ultrasonic cleaning phase that is consistent with the preventative cleaning procedure described above;

[0036] During the basic cleaning stage, the cleaning solution is simultaneously heated online to reduce its viscosity and enhance its dissolving power.

[0037] In addition, the injection flow rate of the cleaning fluid is increased to enhance the shearing and scouring effect of the fluid on the inner wall of the nozzle.

[0038] Preferably, after the step of verifying the cleaning effect, the online calibration step of the digital twin model is also included. The online calibration step is triggered when the cleaning is confirmed to be effective and the conditions of cumulative running time or continuous increase of model prediction deviation are met. Based on the measured spray cone angle value under stable working conditions after cleaning, the wall characteristic parameters in the model are adjusted in reverse through optimization algorithm so that the model prediction value matches the measured value.

[0039] Preferably, the method further includes the step of:

[0040] If the calculated congestion risk index fails to fall back to the safe range, the enhanced cleaning procedure will be automatically triggered again, and the cleaning and verification steps will be repeated until the risk index is confirmed to have fallen back to the safe range.

[0041] Secondly, a digital twin-driven intelligent humidity control system for carbon fiber spinning oil includes:

[0042] The multi-source sensor data acquisition unit is used to collect the nozzle's inlet pressure, outlet pressure, oil temperature, and environmental parameters in real time.

[0043] A high-speed optical imaging unit is used to acquire a sequence of spray pattern images at the nozzle exit.

[0044] The digital twin simulation computing unit is used to carry and run the computational fluid dynamics digital twin model, and to simulate and extract the internal flow state feature vector based on sensor data in real time.

[0045] A spray feature extraction unit is used to process the image sequence, extract and quantize the external representation feature vector of the spray;

[0046] The congestion risk prediction unit has a built-in pre-trained neural network model for receiving the internal and external feature vectors and calculating the congestion risk index.

[0047] The intelligent cleaning decision and execution unit is used to automatically make decisions and execute graded cleaning procedures based on the risk index. It includes a control core, a flow rate adjustment component, an ultrasonic wave generation and coupling component, and a temperature control module.

[0048] The closed-loop validation unit is used to trigger data re-acquisition after cleaning, drive model state updates, recalculate the risk index to verify the effect, and perform online model calibration when the conditions are met.

[0049] In summary, this application includes at least one of the following beneficial technical effects:

[0050] 1. This invention integrates real-time acquired multi-source sensor data with high-speed spray image sequences to drive a high-fidelity computational fluid dynamics digital twin model, enabling dynamic simulation of the internal flow field state of the nozzle and quantitative analysis of the external spray morphology. Furthermore, based on a neural network model that integrates internal and external features, it achieves early identification and quantification of the degree of microchannel blockage risk, overcoming the shortcomings of existing technologies that rely on delayed alarms and cannot provide early warnings of blockage precursors, thereby improving the timeliness and accuracy of risk perception.

[0051] 2. Based on the predicted blockage risk index, this invention intelligently triggers differentiated graded cleaning strategies, including preventive ultrasonic cleaning and enhanced thermo-mechanical synergistic cleaning. It can adaptively adjust the cleaning intensity and method according to the severity of blockage, ensuring the effectiveness and specificity of cleaning while avoiding excessive maintenance. This achieves a shift from passive response to proactive intervention, effectively ensuring spray uniformity and production continuity, and reducing unplanned downtime.

[0052] 3. This invention constructs a complete closed-loop control process that includes post-execution verification and online model calibration. By re-collecting data after cleaning to evaluate the effect, and using measured data under stable conditions to dynamically optimize the key parameters of the digital twin model, the model can continuously adapt to the long-term physical state changes of the nozzle, maintain simulation accuracy and prediction reliability, form a self-iteratory and continuously optimized intelligent control system, and improve the long-term applicability and stability of the method. Attached Figure Description

[0053] Figure 1 This is a schematic flowchart of a digital twin-driven intelligent control method for carbon fiber spinning oil according to the present invention.

[0054] Figure 2 This is a schematic diagram of the process of fusing a high-fidelity computational fluid dynamics digital twin model with a blockage risk prediction neural network in this invention;

[0055] Figure 3 This is a schematic diagram of the process from multi-source sensing to intelligent cleaning decision-making and closed-loop verification in this invention. Detailed Implementation

[0056] This invention provides a digital twin-driven intelligent control method and system for carbon fiber spinning oil. By constructing a high-fidelity digital twin model of the internal fluid dynamics of the nozzle and integrating real-time spray imaging data with multi-source sensor information, it can achieve advanced prediction of nozzle clogging risk. Based on the prediction results, it can dynamically generate and execute differentiated cleaning strategies, thereby proactively intervening before spray performance deteriorates, ensuring the uniformity of oil film thickness and the continuity of carbon fiber production.

[0057] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on specific implementation methods of the present invention.

[0058] A digital twin-driven intelligent control method for carbon fiber spinning oil includes the following steps:

[0059] S1. Acquire multi-source sensor data on nozzle operating status, enabling comprehensive and real-time collection of physical quantity data related to nozzle operating status and operating environment. This is achieved through the following sub-steps:

[0060] S101. An inlet pressure sensor is installed at the outlet of the liquid supply pump, and an outlet pressure sensor is installed about 50mm upstream of the nozzle to monitor the input and output of the liquid supply pressure, respectively.

[0061] An oil temperature sensor is installed near the nozzle in the liquid supply line, and an ambient temperature and humidity sensor is installed on the side wall of the spinning tunnel. The linear speed of the raw yarn is obtained by a rotary encoder installed coaxially with the traction roller.

[0062] The signals from each sensor are transmitted to the central processing unit via industrial Ethernet.

[0063] S102. To ensure the consistency of time sequence of multi-source data, the system adopts a hardware triggering method combined with network time protocol for synchronization to ensure that all sensor data are stamped with a unified millisecond-level timestamp.

[0064] For high-frequency noise in pressure signals, such as pump pulsation, a low-pass filter is used for processing after signal acquisition.

[0065] Moving average filtering is applied to signals such as temperature and speed to smooth out random interference.

[0066] S103. During normal production, the system continuously collects and updates sensor data at a fixed cycle, such as 100ms.

[0067] When spray image acquisition or post-cleaning verification is triggered, the system automatically starts a high-frequency, multi-cycle synchronous data acquisition window to ensure that the sensor data and image data are aligned in time, providing a consistent input benchmark for the model.

[0068] S2. Acquire a sequence of spray morphology images at the nozzle exit using a high-speed optical imaging device to obtain direct visual evidence reflecting the nozzle's operating state, i.e., the instantaneous microscopic morphology of the spray. Simultaneously, by capturing the high-speed moving droplet field, raw image data is provided for subsequent quantitative analysis of spray quality. This is achieved through the following sub-steps:

[0069] S201. Build a high-speed backlight imaging system, which mainly includes a high-speed industrial camera, a coaxial ring LED light source and an image buffer unit.

[0070] The camera features a global shutter function to avoid motion distortion, and the light source is preferably in the red band to enhance the contrast of the droplet edges.

[0071] Key parameters are set based on the spray characteristics: for example, the camera frame rate is set at about 2000 frames per second to ensure that the high-speed movement of droplets can be resolved; the spatial resolution needs to meet the requirement of resolving droplets at the tens of micrometer level.

[0072] The image buffer unit employs mechanisms such as double buffering to ensure no frame loss during high-speed continuous shooting.

[0073] S202. Position the field of view of the imaging system at the initial spray zone approximately 10 to 50 millimeters downstream of the nozzle outlet, as this region is most sensitive to changes in the internal flow field of the nozzle.

[0074] The triggering and timing of image acquisition are uniformly scheduled by the system, specifically as follows:

[0075] (1) Periodic monitoring and acquisition: In normal production monitoring mode, the system automatically triggers an image sequence acquisition once according to a preset period, for example, once per minute. Each acquisition lasts for a short time, such as 2 seconds, to cover multiple dynamic cycles of spraying.

[0076] (2) Event-triggered acquisition: When the sensor data in step S1 shows an abnormal trend, or when the risk index predicted in step S5 approaches the threshold, the system immediately triggers a high-priority image acquisition.

[0077] (3) Cleaning verification data collection: After the cleaning procedure in step S6 or S7 is executed, step S8 will trigger the data collection for effect verification.

[0078] The acquisition trigger signal is simultaneously sent to the central processing unit responsible for executing step S1, ensuring that the sensor data and image data at that moment are marked as the same time batch.

[0079] S203. The acquired raw image sequence is stably transmitted through a high-speed data interface, such as Camera Link, to a computing device used to perform the subsequent image processing step S4, forming an image data stream to be processed.

[0080] S3. Based on multi-source sensor data and nozzle geometric parameters, initialize and drive a computational fluid dynamics digital twin model to reflect the internal flow field state that cannot be directly observed. This is achieved through the following sub-steps:

[0081] S301. First, obtain the precise geometric information of the nozzle. Usually, based on its manufacturing drawings or 3D scanning data, a 3D digital model of the internal flow channel, including key features such as the contraction section, throat and expansion section, is established in the CAD environment.

[0082] Subsequently, a high-quality computational mesh is generated for this geometric model. The mesh generation must meet two requirements:

[0083] First, it can accurately analyze the flow details within microchannels, such as feature sizes on the order of tens of micrometers, especially in the near-wall region where densification is required to accurately calculate shear stress.

[0084] Secondly, while ensuring the aforementioned accuracy, the overall mesh size is controlled to ensure that subsequent simulation steps can complete a single iteration update within a set time period, such as the 200 milliseconds described in step S304, thus meeting the real-time requirements of online applications.

[0085] Unstructured meshes can typically be used to accommodate complex shapes, and mesh independence can be verified.

[0086] S302. To drive the above model to perform simulation, parameters that reflect the actual physical conditions need to be set.

[0087] The core fluid properties include density, viscosity, and surface tension, which are assigned values ​​based on the actual formulation or typical property data of the oil used in the current production line.

[0088] The simulation solver is a transient solver based on the finite volume method, and a physical model that can effectively simulate the internal turbulent structure, such as large eddy simulation, is used to solve the transient flow control equations.

[0089] The solution time step setting needs to balance numerical stability and computational efficiency to ensure that key transient features in the flow can be captured.

[0090] S303. Use the multi-source sensor data collected in real time in step S1 as the dynamic boundary conditions to drive the model operation.

[0091] In the established computational fluid dynamics digital twin model, the boundaries corresponding to the physical locations are predefined, specifically as follows:

[0092] Set the model section representing the flow channel from the outlet of the liquid supply pump to the inlet of the nozzle as the pressure inlet boundary, and set the model section representing the upstream proximal end of the nozzle, such as the corresponding sensor installation position, as the pressure outlet boundary.

[0093] The data collected in real time by the inlet pressure sensor and the outlet pressure sensor in step S1 are then assigned to the two boundary conditions respectively.

[0094] All flow channel walls are set to no-slip boundary conditions.

[0095] The computational fluid dynamics digital twin model runs continuously on an edge computing server equipped with a dedicated graphics processor. It uses the aforementioned real-time updated boundary conditions to drive transient simulation calculations, thereby achieving dynamic simulation of the velocity field, pressure field, and wall shear stress distribution inside the nozzle.

[0096] S304. The system extracts a set of key physical quantities that quantitatively characterize the internal flow state from the latest simulation flow field results according to the set update cycle, such as 200 milliseconds, and together they form a 16-dimensional internal flow state feature vector.

[0097] These characteristics reflect the health status of the flow channel and may include indicators such as the average flow velocity at the throat section, the maximum shear stress on the wall, the standard deviation of the outlet pressure fluctuation, and the velocity gradient distribution along the centerline of the flow channel.

[0098] By optimizing the model and computing platform, we can ensure that the entire process from reading real-time data, driving simulation, to outputting the feature vector can be completed within the update cycle, such as within 200 milliseconds, thereby meeting the timeliness requirements of online monitoring and risk prediction.

[0099] S4. Perform image processing on the spray morphology image sequence to form the external representation feature vector of the spray. This is achieved through the following sub-steps:

[0100] S401. First, the acquired raw image sequence is preprocessed to eliminate fixed background interference and separate the moving droplets.

[0101] Specifically, the background subtraction method is adopted: using the image acquired when the nozzle is unloaded as the background template, the template is subtracted from each frame of spray image, thereby effectively removing the image of ambient light and fixed structure of the equipment, and obtaining a foreground image sequence mainly composed of moving droplets, which lays the foundation for subsequent quantitative analysis.

[0102] S402, the spray cone angle is a key macroscopic indicator characterizing the uniformity and directional stability of spray diffusion.

[0103] To calculate this angle, representative frames or average images from the aforementioned foreground image sequence are processed.

[0104] By employing edge detection algorithms such as Hough transform, the left and right outer boundaries of the spray field are automatically identified, and the angle between these two boundary lines at the nozzle exit plane is calculated to obtain the spray cone angle value.

[0105] S403. In order to quantify the dynamic characteristics of the spray, it is necessary to obtain the velocity information of the droplets.

[0106] The particle image velocimetry technique is used for processing. Based on the foreground image sequence, image pairs with known time intervals are selected. For example, the time interval can be taken as two adjacent frames with a time interval of 0.5 milliseconds according to the acquisition frequency of 2000 frames / second. Cross-correlation analysis is performed on local regions in the image to calculate the displacement of the droplet swarm and thus obtain the local velocity vector.

[0107] A two-dimensional droplet velocity vector field is formed by synthesizing local vectors throughout the entire field of view.

[0108] Based on the two-dimensional droplet velocity vector field, dynamic parameters reflecting flow stability, such as average velocity and turbulence intensity, can be further calculated.

[0109] S404. Droplet size distribution is a key microscopic indicator for evaluating atomization uniformity.

[0110] Based on the foreground image sequence obtained from S401, adaptive threshold segmentation is performed on the image to completely separate the droplets from the background.

[0111] Subsequently, each individual droplet region was identified through connected component analysis, and its equivalent circle diameter was calculated.

[0112] For a complete batch of image acquisition, i.e., an observation window, such as the 4000 frames of images described in S203, the diameters of all identified droplets are counted, and a droplet size distribution histogram is plotted.

[0113] Statistical features that represent the degree of concentration and dispersion of droplet size distribution can be extracted from the droplet size distribution histogram, such as the mean and variance of the droplet size.

[0114] S405, spray flux density reflects the intensity of the oil's projection in space and is directly related to the uniformity of spraying.

[0115] Its calculation integrates the position and velocity information of the droplets, specifically as follows:

[0116] Based on the two-dimensional droplet velocity vector field generated by S403 and the droplet position identified by S404, the image is divided into multiple small unit regions. The number of droplets in each unit is counted, and the average velocity amplitude of the droplets in that unit is used for weighting to obtain a two-dimensional spatial distribution map of the spray flux density.

[0117] Analyzing the two-dimensional spatial distribution map can extract features such as peak position and distribution uniformity index, which can reflect whether the spray is eccentric or has local weak areas.

[0118] S406. Integrate and encode all the quantitative indicators obtained in steps S402 to S405 above to form a fixed-dimensional, machine-readable spray external characterization feature vector.

[0119] The external characterization vector of the spray integrates macroscopic morphology, microscopic statistics and dynamic characteristics. For example, it may include the spray cone angle, the mean and variance of droplet size, the average velocity and turbulence intensity of the velocity field, the geometric characteristics of the spray flux density distribution such as peak coordinates and statistical characteristics such as uniformity index, as well as other representative statistics extracted from the particle size distribution histogram and flux density map such as higher-order moments and spatial moments.

[0120] All feature values ​​were normalized before integration to ensure their comparability in the model.

[0121] The external spray characterization feature vector serves as a comprehensive digital description of the external spray state, which is then used by subsequent risk prediction models.

[0122] S5. Input the internal flow state feature vector and the spray external characterization feature vector into the pre-trained clogging risk prediction neural network model, and output the clogging risk index. This is achieved through the following sub-steps:

[0123] S501. To achieve the fusion analysis of internal and external multi-dimensional features, a dedicated neural network prediction model is constructed and adopted. The design of this model must meet two key requirements:

[0124] First, it can effectively process and mine complex patterns in high-dimensional feature vectors, such as the 48-dimensional vector in this embodiment, which is formed by splicing 16-dimensional internal features and 32-dimensional external features.

[0125] Secondly, it can capture the evolution trend of features in time series or continuous assessment, which is crucial for predicting the development of congestion.

[0126] Therefore, neural network prediction models preferably employ a hybrid architecture that combines local feature extraction with sequence dependency learning. For example, convolutional layers can be included to automatically learn local correlations between features, and recurrent neural network layers can be included to understand trends in state changes.

[0127] The final output layer of a neural network prediction model maps it to a single predicted value.

[0128] S502. The output of the neural network prediction model is defined as the blockage risk index, which is a continuous value used to quantify the risk state of the nozzle from fully unobstructed to fully blocked.

[0129] The congestion risk index is usually normalized in the range of 0 to 1, where 0 represents an ideal unobstructed state and 1 represents a completely blocked failure state.

[0130] The actual physical meaning of the blockage risk index is established through laboratory calibration. During the model training phase, different degrees of blockage are simulated and the corresponding physical effects, such as pressure drop changes and spray pattern deterioration, are observed. Each training sample is then assigned a real risk index between 0 and 1.

[0131] The neural network prediction model learns and eventually establishes a mapping relationship between the complex 48-dimensional mixed features and this physically calibrated single-value risk index.

[0132] Therefore, the risk index output during online prediction is actually a quantitative assessment of the current nozzle condition relative to the calibration benchmark, and its value directly reflects the severity of the blockage.

[0133] S503. To train the risk prediction model, a calibration dataset that can correlate nozzle physical state with multi-source feature vectors is constructed. The construction method is as follows:

[0134] In the laboratory, using the same nozzles as the production line, different degrees of clogging, from slight to severe, are simulated by controllably introducing standard particles of known size and concentration gradients, such as silica particles, into their supply lines.

[0135] For each set simulated blockage condition, steps S1 to S4 are started and executed synchronously to fully collect multi-source sensor data and spray image sequences under that condition, and finally generate the corresponding internal flow state feature vector and spray external characterization feature vector to form a data sample.

[0136] The key point is to assign an objective benchmark risk index to each sample. The benchmark risk index does not rely purely on subjective judgment, but is determined comprehensively based on observable and quantifiable physical indicators under the simulated working conditions.

[0137] For example, based on the rate of change of the pressure drop measured under the operating condition relative to the unobstructed baseline, and the degree of deviation of image analysis indicators such as spray cone angle or droplet uniformity, a value between 0 and 1 can be calculated according to a preset conversion rule or reference table, which can be used as the benchmark risk index for the sample.

[0138] By repeating the above process extensively, a dataset covering various congestion conditions is constructed, with each sample having an objective quantitative label.

[0139] S504. Using the dataset constructed above, supervise the training of the neural network prediction model in step S501.

[0140] Before training, the dataset is usually divided into non-overlapping training, validation, and test sets, for example, randomly divided in a 7:2:1 ratio, to ensure the independence of the final evaluation.

[0141] The training process aims to minimize the gap between the model's predicted values ​​and the sample's baseline risk index. Commonly used loss functions such as mean squared error are suitable for this regression task.

[0142] During training, a gradient descent-type optimization algorithm is used to update parameters, and the training process is monitored using a validation set to prevent overfitting.

[0143] After the model is trained, its performance is evaluated on an independent test set. The error between the predicted risk index and the benchmark index is required to be controlled within an acceptable range for engineering applications, such as an accuracy of less than 0.05 in this embodiment, to ensure that the model has reliable generalization ability and prediction accuracy.

[0144] Model parameters that have been trained and validated to meet the requirements will be fixed and deployed to the edge computing units of the production system for online real-time prediction.

[0145] S6. When the blockage risk index exceeds the preset first threshold, start the preventive cleaning procedure.

[0146] When the predictive model identifies early risk of clogging, the system triggers a first-level automated response strategy designed to remove initial deposits through gentle, proactive physicochemical processes to prevent further clogging.

[0147] Specifically, this is achieved through the following sub-steps:

[0148] S601. The system monitors the congestion risk index output from step S5 in real time.

[0149] When the index continuously exceeds a preset first threshold (set to 0.4 in this embodiment) and reaches a predetermined stabilization time, such as three consecutive prediction cycles, the nozzle is determined to have entered an early risk state.

[0150] The first threshold setting and stabilization time requirement here are based on the analysis of a large amount of historical data or laboratory calibration data, aiming to ensure that the system responds to real and continuous early clogging tendencies, while avoiding false triggering due to instantaneous fluctuations.

[0151] Once the determination is confirmed, the system's central controller automatically generates instructions to trigger and initiate the preventative cleaning process.

[0152] S602. The system uses a dedicated pulsed ultrasonic cleaning fluid. The basic components of this cleaning fluid are exactly the same as the working oil used in the current production line to ensure compatibility and avoid introducing new contaminants.

[0153] To improve the cleaning effect, especially for inorganic salt deposits that may form in the early stages, add an appropriate amount of chelating agent to the base solution, such as sodium citrate.

[0154] Chelating agents can effectively complex metal ions, such as calcium and magnesium ions, that may precipitate from oils, preventing them from forming hard soap scum or salt deposits, thereby enhancing the cleaning solution's ability to dissolve and disperse early inorganic deposits.

[0155] S603. The physical method of cleaning is to utilize the cavitation effect generated by ultrasound in the liquid, and to generate high-frequency mechanical vibration through an ultrasonic generator, for example, the frequency can be selected at around 40kHz.

[0156] Mechanical vibration is effectively transmitted to the inside of the nozzle through a mechanical device tightly coupled to the nozzle metal body, such as a special clamp or a transmission rod, causing the cleaning fluid inside and on the nozzle microchannel wall to vibrate at high frequency. The microjet and shock wave generated by the cavitation effect help to loosen and peel off the initial deposits attached to the wall.

[0157] S604. To avoid the possibility of local overheating or cumulative damage to the precision nozzle structure caused by continuous ultrasonic action, the cleaning process adopts a pulse working mode.

[0158] In this mode, the ultrasound operates in a work-pause cycle. For example, it can be set to work for a few seconds and then pause for a longer period of time, such as working for 5 seconds and then pausing for 10 seconds. A complete preventative cleaning cycle lasts about 2 minutes.

[0159] By operating intermittently, the effective time of cavitation effect is ensured to loosen the deposits, while providing sufficient quiet time for the cleaning fluid to remove the stripped particles and allow the system to cool down, thus achieving gentle yet effective cleaning.

[0160] S605. During the cleaning process, the injection flow rate of the special cleaning solution shall be controlled at a level that is basically consistent with the oil supply flow rate during normal spinning of the nozzle, for example, at about 50 ml / min.

[0161] Maintaining the flow rate is based on two main considerations:

[0162] Firstly, it can ensure that the cleaning fluid fills the entire flow channel smoothly and fully, forming a stable flow environment, which is conducive to the uniform effect of ultrasonic cavitation and the transport of the stripped material.

[0163] Secondly, it avoids unnecessary hydraulic shocks or disturbances to the upstream liquid supply pipeline and pump valve system caused by drastic changes in flow rate, reflecting the friendliness of the cleaning process to the production system.

[0164] S7. When the blockage risk index exceeds the preset second threshold, the enhanced cleaning program is started.

[0165] When the predictive model identifies a moderate or higher risk of blockage, the system triggers a second, more powerful automated cleaning strategy. This strategy, based on preventative cleaning, incorporates thermal and mechanical energy to effectively remove pre-formed deposits with stronger adhesion.

[0166] Specifically, this is achieved through the following sub-steps:

[0167] S701, The system continuously monitors the congestion risk index.

[0168] If the index continues to exceed the preset second threshold, for example, 0.7 in this embodiment, and reaches the predetermined stabilization time, such as for two consecutive prediction cycles, the nozzle is determined to be in a medium-risk or medium-term blockage state.

[0169] The second threshold is higher than the first threshold. It is set based on more significant physical performance degradation, such as further increase in pressure drop and obvious deterioration of spray pattern. It aims to distinguish between moderate and severe blockage conditions that require stronger intervention and early risks.

[0170] Once the determination is confirmed, the system's central controller immediately generates an upgrade command, triggering and initiating the enhanced cleaning process.

[0171] S702. The enhanced cleaning procedure first performs the basic cleaning stage, which is consistent with the preventive cleaning procedure defined in step S6. That is, a special cleaning fluid is injected into the nozzle and pulsed ultrasonic waves are applied to loosen the deposited layer in advance by utilizing the cavitation effect.

[0172] The cleaning fluid formulation and ultrasonic parameters used in the basic cleaning stage, such as frequency, pulse timing, and flow rate, all follow the principles and example parameters set for preventative cleaning to ensure consistency from basic loosening to subsequent strengthening steps.

[0173] S703. While the basic ultrasonic cleaning is being performed in step S702, the system starts online heating of the cleaning fluid.

[0174] The electric heating module raises the temperature of the cleaning fluid flowing through the nozzle from the ambient temperature and stabilizes it within a suitable cleaning temperature range, such as around 60°C in this embodiment.

[0175] The purpose of raising the temperature is mainly threefold:

[0176] First, it significantly reduces the viscosity of the cleaning fluid, improving its fluidity and penetration.

[0177] Secondly, it enhances the cleaning solution's ability to dissolve organic components in oil-based agents;

[0178] Third, heat energy is used to soften or loosen certain organic or composite deposits, reducing their adhesion to the wall surface.

[0179] Precise temperature control helps maintain consistency and safety in the cleaning process.

[0180] S704. Simultaneously, the system adjusts the flow control valve in the pipeline to appropriately increase the injection flow rate of the cleaning fluid from the basic level of the preventive cleaning stage to a higher set value, such as about 1.5 times the normal operating flow rate.

[0181] The purpose of increasing the flow rate is to further increase the shear stress on the wall surface of the cleaning fluid flowing through the microchannel inside the nozzle, based on the fact that the fluid viscosity has decreased due to heating.

[0182] This enhanced mechanical scouring action, combined with the ultrasonic cavitation effect and thermal softening effect, can more effectively peel off the deposit particles that have been loosened by the aforementioned steps from the wall and carry them out with the fluid, thereby improving the cleaning efficiency for moderate blockages. In addition, the increase in flow rate needs to take into account factors such as cleaning effect, system pressure capacity and fluid stability.

[0183] S8. After the cleaning procedure is executed, multi-source sensor data and spray pattern image sequences are re-acquired, the state of the computational fluid dynamics digital twin model is updated, and the blockage risk index is verified to have fallen back to the safe range.

[0184] Specifically, this is achieved through the following sub-steps:

[0185] S801. After the preventive or enhanced cleaning procedure is completed, the controller will not collect data immediately, but will wait for a preset stabilization time, such as 30 seconds, to allow the cleaning fluid in the pipeline to be completely replaced by the working oil and to restore the flow inside the nozzle and the external spray field to a stable production condition.

[0186] Once the stabilization period ends, the controller automatically triggers a dedicated closed-loop verification data acquisition process. This process re-executes steps S1 and S2, but the acquisition strategy focuses on obtaining stable state values. Multi-source sensor data will be continuously acquired within a short time window, and their average value will be calculated to represent steady-state conditions. At the same time, a spray image sequence acquisition with the same periodic monitoring acquisition parameters as in step S2 is triggered to obtain baseline data on the health status of the nozzle under stable working conditions after cleaning.

[0187] S802. Using the newly acquired real-time sensor data from S801, especially the average pressure and temperature values, as new boundary conditions, drive the computational fluid dynamics digital twin model in step S3 to perform a transient simulation from the initial state until the flow field reaches stability, and extract the corresponding new internal flow state feature vector. This process can be understood as refreshing the digital twin with verification data.

[0188] Simultaneously, the complete processing flow of step S4 is performed on the newly acquired image sequence to generate a new spray external characterization feature vector.

[0189] Subsequently, this new set of internal and external feature vectors is input into the congestion risk prediction neural network model deployed in step S5 to calculate the real-time congestion risk index under stable operating conditions after cleaning.

[0190] S803. The system compares the new congestion risk index calculated in S802 with a preset safe range. In this embodiment, the safe range is defined as the risk index continuously being lower than a first threshold, such as 0.4.

[0191] It should be noted that confirming a decline usually requires the risk index to remain consistently below 0.4 in several consecutive verification assessments, such as twice, in order to avoid misjudgments caused by fluctuations in a single data point.

[0192] If the verification results show that the risk index has met the confirmation and decline conditions, the cleaning is considered successful, the nozzle is restored to a safe state, and the system returns to normal monitoring mode.

[0193] If the risk index is still higher than or equal to 0.4, or if the continuous and stable confirmation condition is not met, the cleaning is deemed incomplete. In this case, the system will automatically trigger and execute the enhanced cleaning procedure defined in step S7 again, and enter a new round of cleaning-verification cycle.

[0194] This iterative process will continue until the risk index is confirmed to have stabilized and fallen back into a safe range, thereby ensuring that the congestion problem is completely resolved.

[0195] S804. To maintain the high fidelity of the computational fluid dynamics digital twin model throughout the entire equipment lifecycle, the system is designed with an online calibration mechanism to compensate for the gradual drift of the model caused by long-term physical wear of the nozzle, such as changes in wall roughness.

[0196] (1) Calibration triggering conditions: Online calibration is not performed after every cleaning, but is automatically triggered when the following two conditions are met simultaneously:

[0197] a) Step S803 confirms that the cleaning is effective and the nozzle has been restored to a safe state;

[0198] b) Since the last successful calibration, the cumulative production run time of the nozzle has exceeded the preset threshold, such as 200 hours, or the system has detected that the deviation between the spray pattern characteristics predicted by the model, such as the cone angle and the measured value, shows a continuous increasing trend.

[0199] This design ensures that calibration is performed only in a stable and healthy state where the model may have drifted significantly, avoiding incorrect calibration in faulty or incompletely cleaned conditions.

[0200] (2) Calibration method: Calibration is to use the measured data under stable working conditions after cleaning to perform reverse optimization of the key physical parameters of the model.

[0201] Since the spray cone angle is highly sensitive to the internal flow channel condition of the nozzle, which includes its geometry and wall roughness, and is a stable macroscopic indicator that can be directly obtained in step S4, this embodiment preferably uses the measured spray cone angle as the target benchmark for calibration. The specific process is as follows:

[0202] The system calls the measured stable spray cone angle value obtained in step S801;

[0203] Under the same set of boundary conditions used for verification in step S802, the current computational fluid dynamics digital twin model is run to obtain its predicted spray cone angle value;

[0204] With the goal of minimizing the error between measured and predicted values, optimization algorithms, such as least squares method and genetic algorithm, are used to adjust the parameters characterizing wall properties, such as wall roughness coefficient, in the computational fluid dynamics digital twin model in reverse.

[0205] Through iterative optimization, the model's predicted output under the given boundary conditions achieves the best match with the measured data, thereby completing the dynamic correction of the model's physical parameters.

[0206] (3) The new parameters obtained from the optimization will be updated and saved in the computational fluid dynamics digital twin model configuration corresponding to the nozzle, and used for all subsequent online simulation predictions until the next calibration is triggered.

[0207] To achieve the aforementioned method, this invention also provides a digital twin-driven intelligent humidity control system for carbon fiber spinning oil. This system, through hardware and software collaboration, constructs a complete closed loop of "perception-simulation-decision-execution-verification." Specifically, it includes the following units:

[0208] The multi-source sensor data acquisition unit is responsible for real-time acquisition of nozzle operating status and environmental parameters. It integrates various types of sensors mentioned in step S1, specifically including inlet and outlet pressure sensors for monitoring hydraulic pressure, an immersion platinum resistance thermometer for measuring oil temperature, an integrated sensor for acquiring ambient temperature and humidity, and a rotary encoder for detecting the linear velocity of the precursor fiber. All sensors communicate with the system's central processing unit via a fieldbus to achieve real-time data upload.

[0209] The high-speed optical imaging unit, used to acquire microscopic morphological images of the spray, corresponds to step S2 of the method. Its core is a high-speed backlit imaging system, mainly composed of a high-speed industrial camera, a coaxial ring LED light source, and an image buffer module. This unit is independently mounted on a shockproof platform to isolate the vibration of the spinning equipment and ensure stable imaging quality; the camera has a global shutter and is strictly synchronized with the light source pulse to freeze high-speed droplets.

[0210] The digital twin simulation computing unit carries and runs the high-fidelity computational fluid dynamics digital twin model from step S3. Deployed on an edge computing server and equipped with a dedicated graphics processor, it accelerates simulation calculations. Through optimization, this unit supports a complete iterative update of the internal flow field simulation of a single nozzle within 200 milliseconds, thus meeting online real-time requirements.

[0211] The spray feature extraction unit, corresponding to step S4, is responsible for processing and analyzing the original image sequence acquired by the high-speed optical imaging unit. This unit runs on a dedicated image processing workstation. By calling image processing algorithm libraries, such as background subtraction, Hough transform, and PIV cross-correlation analysis, and using parallel computing technology to accelerate the processing flow, it finally extracts and quantifies features such as spray cone angle, droplet size distribution, and velocity field from the image, forming a 32-dimensional spray external characterization feature vector.

[0212] The congestion risk prediction unit, corresponding to step S5, is a pre-trained neural network model at its core. This unit loads the pre-trained model parameters and receives in real time a 16-dimensional internal feature vector from the digital twin simulation calculation unit and a 32-dimensional external feature vector from the spray feature extraction unit. Through model calculation, it outputs a congestion risk index between 0 and 1.

[0213] The intelligent cleaning decision-making and execution unit, based on the output of the risk prediction unit, executes the differentiated cleaning strategies defined in steps S6 and S7. Essentially, it is an integrated electromechanical control unit, mainly comprising:

[0214] Programmable Logic Controller (PLC): As the control core, it receives the risk index, makes cleaning decisions based on preset threshold logic, and outputs control signals.

[0215] Proportional control valve: Receives analog signals from the PLC to precisely adjust the injection flow rate of the cleaning fluid, such as maintaining 50 mL / min during preventative cleaning and increasing it to 75 mL / min during intensive cleaning.

[0216] Ultrasonic generator and coupling device: generates ultrasonic waves of a specific frequency, such as 40kHz, and transmits the vibration energy to the nozzle body through mechanical coupling, such as clamps.

[0217] Temperature control module: Includes electric heating element and temperature sensor, used to heat the cleaning fluid in the enhanced cleaning process and control it in a closed loop at a set temperature, such as 60°C±1°C.

[0218] The closed-loop verification unit, corresponding to step S8, is responsible for coordinating the system to verify the effectiveness and calibrate the model after the cleaning process is completed. It is not an independent physical device, but rather implemented by the control logic and software modules within the central processing unit. Its main functions are to trigger a new round of data acquisition after the cleaning procedure is finished, update the digital twin model status, recalculate the risk index to verify the cleaning effect, and initiate the online parameter calibration process of the digital twin model when conditions are met.

[0219] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0220] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A digital twin-driven intelligent control method for carbon fiber spinning oil, characterized in that, Includes the following steps: Acquire multi-source sensor data reflecting the nozzle's operating status and a sequence of spray pattern images at the nozzle outlet; Based on the multi-source sensor data and the geometric parameters of the nozzle, the computational fluid dynamics digital twin model is driven to perform real-time simulation and extract feature vectors characterizing the internal flow state. The spray morphology image sequence is processed to extract feature vectors that characterize the external morphology and dynamic properties of the spray; The internal flow state feature vector and the external spray characterization feature vector are input into a pre-trained clogging risk prediction neural network model, which outputs a clogging risk index. Based on the comparison result between the blockage risk index and the preset threshold, the corresponding level of cleaning program is automatically triggered and executed. The cleaning program includes at least preventive cleaning and enhanced cleaning. After the cleaning procedure is performed, data is collected again to update the state of the digital twin model, and the congestion risk index is recalculated based on the updated model and data to verify the cleaning effect.

2. The digital twin-driven intelligent humidity control method for carbon fiber spinning oil according to claim 1, characterized in that, The steps for acquiring multi-source sensor data include: Collect the nozzle inlet pressure, outlet pressure, oil temperature, and environmental parameters; It also performs time-series synchronization and filtering on multi-source data; The spray pattern image sequence is acquired through a high-speed backlight imaging system. The acquisition timing includes periodic monitoring acquisition, event-triggered acquisition based on abnormal sensor data or risk index approaching the threshold, and verification acquisition after cleaning.

3. The digital twin-driven intelligent humidity control method for carbon fiber spinning oil according to claim 1, characterized in that, The steps for driving real-time simulation of computational fluid dynamics digital twin models include: A three-dimensional model of the internal flow channel was established based on the nozzle geometry, and a computational mesh was generated. The real-time collected inlet and outlet pressure data are used as the dynamic boundary conditions of the model. Flow field simulation was performed using a transient solver. The internal flow state feature vector, including the average flow velocity at the throat section, wall shear stress, and pressure fluctuation characteristics, is extracted from the simulation results according to a set period.

4. The digital twin-driven intelligent humidity control method for carbon fiber spinning oil according to claim 1, characterized in that, The steps for processing the spray pattern image sequence include: Background subtraction preprocessing is performed on the image to separate moving droplets; the spray cone angle is calculated based on the processed image; Particle image velocimetry was used to obtain the droplet velocity vector field and calculate dynamic parameters. Identify droplets and statistically analyze their size distribution; And calculate the spatial distribution of spray flux density based on droplet position and velocity information; Finally, the obtained quantitative indicators are integrated to form a feature vector representing the external characteristics of the spray.

5. The digital twin-driven intelligent humidity control method for carbon fiber spinning oil according to claim 1, characterized in that, The congestion risk prediction neural network model is a hybrid architecture model that processes high-dimensional features and captures temporal trends; It is trained using a laboratory-calibrated dataset, the construction of which includes: Different degrees of blockage were simulated in the experimental environment. Multi-source sensor data and spray images were collected simultaneously to generate feature vector samples. A quantitative benchmark risk index was assigned to each sample based on the deviation of pressure drop rate and spray morphology index.

6. The digital twin-driven intelligent humidity control method for carbon fiber spinning oil according to claim 1, characterized in that, The preventative cleaning procedure includes: Inject a special cleaning solution into the nozzle that is compatible with the composition of the working oil and contains chelating agents; Ultrasonic waves are applied to the nozzle in pulsed operating mode to loosen deposits using cavitation effect; The flow rate of the cleaning fluid was controlled at a level close to the normal operating flow rate.

7. The digital twin-driven intelligent humidity control method for carbon fiber spinning oil according to claim 1, characterized in that, The enhanced cleaning procedure includes: First, perform a basic ultrasonic cleaning phase that is consistent with the preventative cleaning procedure described above; During the basic cleaning stage, the cleaning solution is simultaneously heated online to reduce its viscosity and enhance its dissolving power. In addition, the injection flow rate of the cleaning fluid is increased to enhance the shearing and scouring effect of the fluid on the inner wall of the nozzle.

8. The digital twin-driven intelligent humidity control method for carbon fiber spinning oil according to claim 1, characterized in that, After verifying the cleaning effect, the online calibration step of the digital twin model is also included. The online calibration step is triggered when the cleaning is confirmed to be effective and the conditions of cumulative running time or continuous increase of model prediction deviation are met. Based on the measured spray cone angle value under stable working conditions after cleaning, the wall characteristic parameters in the model are adjusted in reverse through optimization algorithm so that the model prediction value matches the measured value.

9. The digital twin-driven intelligent humidity control method for carbon fiber spinning oil according to claim 1, characterized in that, The method further includes the following steps: If the calculated congestion risk index fails to fall back to the safe range, the enhanced cleaning procedure will be automatically triggered again, and the cleaning and verification steps will be repeated until the risk index is confirmed to have fallen back to the safe range.

10. A digital twin-driven intelligent humidity control system for carbon fiber spinning oil, used to implement the digital twin-driven intelligent humidity control method for carbon fiber spinning oil as described in any one of claims 1 to 9, characterized in that, include: The multi-source sensor data acquisition unit is used to collect the nozzle's inlet pressure, outlet pressure, oil temperature, and environmental parameters in real time. A high-speed optical imaging unit is used to acquire a sequence of spray pattern images at the nozzle exit. The digital twin simulation computing unit is used to carry and run the computational fluid dynamics digital twin model, and to simulate and extract the internal flow state feature vector based on sensor data in real time. A spray feature extraction unit is used to process the image sequence, extract and quantize the external representation feature vector of the spray; The congestion risk prediction unit has a built-in pre-trained neural network model for receiving the internal and external feature vectors and calculating the congestion risk index. The intelligent cleaning decision and execution unit is used to automatically make decisions and execute graded cleaning procedures based on the risk index. It includes a control core, a flow rate adjustment component, an ultrasonic wave generation and coupling component, and a temperature control module. The closed-loop validation unit is used to trigger data re-acquisition after cleaning, drive model state updates, recalculate the risk index to verify the effect, and perform online model calibration when the conditions are met.

Images