System and method for regulating and controlling dispersion process of carbon nanotube and carbon black composite slurry in real time

Abstract

The invention relates to the technical field of intelligent preparation of new materials, and particularly discloses a real-time regulation and control system and method for a dispersion process of carbon nanotube and carbon black composite slurry, and the system comprises a signal acquisition module which obtains slurry particle size distribution and equipment power parameters of each process link in real time; the state quantification module converts original data into a dynamic state vector sequence representing dispersion evolution; the intelligent decision-making module retrieves similar historical process fragments from a case knowledge base through a dynamic time warping algorithm, a Gaussian process regression prediction model is constructed, a slurry dynamic adaptation factor K is introduced, and the factor synthesizes raw material characteristics, real-time states and process constraints to form a multi-target dispersion potential function; and finally, performing rolling optimization by adopting a model predictive control framework and taking K as a core constraint, and generating a collaborative adjustment instruction of the sanding rotating speed, the ultrasonic power and the dispersing agent flow rate. The closed-loop intelligent regulation and control of the dispersion process of the composite slurry are realized, and the dispersion uniformity, the batch stability and the production energy efficiency are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent preparation of new materials, and particularly relates to a real-time regulation system and method for dispersion process of carbon nanotube and carbon black composite slurry. BACKGROUND

[0002] With the rapid development of new energy, electronic devices and other fields, the demand for high-performance conductive slurry is increasingly urgent. Carbon nanotube and carbon black composite slurry has become a key material for lithium ion battery electrodes, conductive coatings and other key materials due to its excellent conductivity, mechanical strength and stability. In the industrial production process, the dispersion quality of the slurry directly determines the performance and consistency of the final product. Traditional dispersion process relies on fixed process parameters and operator experience, and the dispersion effect is evaluated by offline sampling detection (such as laser particle size analyzer). In recent years, with the progress of automation technology, some dispersion control systems based on single-point sensors (such as online viscometers) and simple PID feedback have appeared, which can adjust a single process parameter (such as stirring speed) to a limited extent.

[0003] However, the existing technical solutions still have significant defects. First, the control logic is mostly in the "open loop" or "single-point simple feedback" mode, which can only make rough adjustments according to the lagging feedback of a single macroscopic physical parameter (such as viscosity, final particle size), and cannot realize real-time sensing and collaborative optimization of the dynamic evolution trajectory of the whole process of sanding, ultrasonic, shearing and other multi-link cooperation. Second, the control strategy relies heavily on pre-set fixed procedures or manual experience, lacks self-adaptive ability to fluctuations in raw material characteristics (such as carbon tube aspect ratio, carbon black specific surface area), and cannot ensure dispersion quality while taking into account energy consumption optimization and protection of carbon nanotube structure integrity, resulting in large differences in performance of different batches of slurry, low production efficiency, and serious loss of high-end raw materials. Finally, the existing system lacks intelligent decision-making ability based on historical successful cases, and it is difficult to convert past optimal process experience into reproducible and generalizable standardized production knowledge, which restricts the intelligent upgrading of the production process.

[0004] Therefore, the present application proposes a real-time regulation system and method for dispersion process of carbon nanotube and carbon black composite slurry. SUMMARY

[0005] The present application provides a real-time regulation system and method for dispersion process of carbon nanotube and carbon black composite slurry, which realizes multi-link dynamic collaborative optimization of the dispersion process of carbon nanotube and carbon black composite slurry by constructing a whole-process closed-loop intelligent regulation architecture, effectively overcomes the defects of existing technologies that rely on fixed parameters and lagging feedback and cannot take into account dispersion quality, energy consumption and material integrity, and significantly improves slurry uniformity, batch stability and production efficiency.

[0006] The present application provides a real-time regulation system for dispersion process of carbon nanotube and carbon black composite slurry, comprising: The signal acquisition module is used to acquire slurry status signals and process parameters in real time at multiple process stages during the dispersion process; The state quantization module is used to generate a dynamic state vector sequence that characterizes the real-time evolution of the dispersion process based on the slurry state signal and process parameters. The case knowledge base is used to store historical, distributed cases containing raw material characteristic vectors, process state matrices, process parameter matrices, and performance evaluation vectors. The intelligent decision-making module is used to execute closed-loop control logic: based on the dynamic state vector sequence and raw material characteristic vector of the current batch, it retrieves a set of similar historical process segments from the case knowledge base; based on the retrieved set of similar historical process segments, it constructs a dynamic prediction model to predict the effect of process parameter adjustment; with the goal of minimizing the dispersion potential function, it combines the dynamic prediction model and generates multi-process parameter coordinated adjustment instructions through rolling optimization calculation. The execution module is used to execute instructions for the coordinated adjustment of multiple process parameters and to intervene in the decentralized process in real time.

[0007] Preferably, the signal acquisition module includes: The first laser particle size sensor is deployed at the outlet of the sand milling stage to collect slurry particle size distribution data in the sand milling stage; The second laser particle size sensor is deployed at the outlet of the ultrasonic process to collect slurry particle size distribution data during the ultrasonic process. The third laser particle size sensor is deployed at the outlet of the shearing stage to collect slurry particle size distribution data during the shearing stage. A sand mill power transmitter is connected to the sand mill driver to collect the sand mill spindle speed and torque in real time, or to directly collect the sand mill drive power. An ultrasonic power monitor, connected to an ultrasonic generator, is used to collect the output electrical power of the ultrasonic generator in real time. Among them, the slurry state signals of multiple process stages include slurry particle size distribution data of the sand milling stage, ultrasonic stage, and shearing stage; Process parameters include the spindle speed and torque of the sand mill or the drive power of the sand mill, as well as the output power of the ultrasonic generator.

[0008] Preferably, the method for the state quantization module to generate a dynamic state vector sequence includes: Using a fixed duration T as the time window, feature extraction is performed on the slurry particle size distribution data of each process step within the current time window to obtain the feature vector within the step. The feature vector within the step includes the distribution width feature value, the main peak position feature value, and the distribution symmetry feature value. Calculate the ratio of the difference in the characteristic values ​​of the main peak positions between adjacent process steps to obtain the coupling characteristic values ​​between the steps; The characteristic value of sand mill power can be calculated based on the spindle speed and torque of the sand mill, or the driving power of the sand mill can be directly used as the characteristic value of sand mill power. Obtain the characteristic value of the output electrical power of the ultrasonic generator within the same time window; The characteristic values ​​of sand milling power and ultrasonic power are used as the energy input characteristic values; The intra-process feature vectors, inter-process coupling feature values, and energy input feature values ​​of all process steps within each time window are combined in a preset order to form a state feature vector; The state feature vectors of consecutive time windows are arranged in chronological order to form a dynamic state vector sequence.

[0009] Preferably, the historical scattered cases in the case knowledge base include: Raw material characteristic vector, process state matrix, process parameter matrix, performance evaluation vector; Among them, the raw material characteristic vector includes the aspect ratio characteristic value of carbon nanotubes, the specific surface area characteristic value of carbon black, and the initial solid content characteristic value of slurry. The rows of the process state matrix correspond to the time series, and the columns correspond to the eigenvalues ​​of each dimension of the state feature vector; The rows of the process parameter matrix correspond to the time series synchronized with the process state matrix, and the columns include the milling speed setting, ultrasonic power setting, and dispersant flow rate setting. The performance evaluation vector includes the slurry conductivity score and the slurry static stability score.

[0010] Preferably, the specific steps for the intelligent decision-making module to retrieve a set of similar historical process fragments from the case knowledge base include: The sequence of dynamic state vectors generated in the current batch is recorded as the query sequence; For each historical scattered case in the case knowledge base, extract the submatrix in the process state matrix of the historical scattered case that has the same time length as the query sequence as the candidate sequence; Calculate the dynamic time-normalized distance between the query sequence and each candidate sequence; A comprehensive similarity score is calculated based on the dynamic time warp distance and the Euclidean distance between raw material characteristic vectors. Select the N historical scattered cases with the highest comprehensive similarity scores, and extract the process state subsequence and process parameter subsequence of the N historical scattered cases after the matching period to form a set of similar historical process segments.

[0011] Preferably, the method for constructing a dynamic prediction model by the intelligent decision-making module includes: Using the state feature vector and raw material characteristic vector of each similar historical process segment at time t in the set of similar historical process segments as the model input sample, and the milling speed adjustment, ultrasonic power adjustment, and dispersant flow rate adjustment of the corresponding similar historical process segment at time t+1 as the model output label, a multi-output Gaussian process regression model is trained as a dynamic prediction model. The multi-output Gaussian process regression model uses a combination of radial basis function kernel and white noise kernel kernel.

[0012] Preferably, the dispersion potential function E is a function of the state feature vector s, the process parameter vector p, and the raw material characteristic vector r. Its construction and optimization objective are closely related to a slurry dynamic adaptation factor K, specifically: Real-time calculation of the dynamic adaptation factor K of the slurry: K = α × K0 + β × K s -γ×K c ; Where α, β, γ are the preset first set of weight coefficients; K0 is the raw material baseline term, which is calculated based on the characteristic values ​​of the initial solid content of the slurry, the aspect ratio of carbon nanotubes, and the specific surface area of ​​carbon black in the raw material characteristic vector. K s The real-time state correction term is calculated based on the difference ratio between the distribution width feature value and the main peak position feature value between adjacent process steps in the state feature vector. K c The process constraint penalty is calculated based on the ratio of the cumulative volume of dispersant added to the theoretical maximum required volume calculated based on the initial solid content of the slurry, and the ratio of the current total power to the upper limit of the total power of the equipment. The current total power is the sum of the sand mill drive power collected in real time by the signal acquisition module or the power calculated based on the speed and torque of the sand mill spindle and the output power of the ultrasonic generator. The dispersed potential function E is expressed as: E(s,p,r)=w1×f1(distribution width characteristic value deviation)+w2×f2(cumulative energy consumption)+w3×f3(length-to-diameter ratio change rate)+w4×f4(K); Among them, w1, w2, w3, and w4 are the preset second set of weight coefficients; f1 (distribution width eigenvalue deviation) is a monotonically increasing function of the deviation between the distribution width eigenvalue in the state eigenvector s and the preset target distribution width value; f2 (cumulative energy consumption) is a monotonically increasing function of the total energy consumption from the start of the distributed process to the current moment; f3 (aspect ratio change rate) is a monotonically increasing function of the ratio of the current estimated aspect ratio to the initial aspect ratio of carbon nanotubes; f4(K) is the absolute value of the difference between the dynamic adaptation factor K of the slurry and the value 1.

[0013] Preferably, the intelligent decision-making module generates multi-process parameter collaborative adjustment instructions through rolling optimization calculations, and this is implemented using a model predictive control framework. Specific steps include: At each control moment, starting with the current state feature vector and process parameter settings, a set of different exploratory process parameter adjustment sequences are generated within the search space defined below: Based on the current mill speed setting, ultrasonic power setting, and dispersant flow rate setting, within the three-dimensional space formed by the mill speed setting adjustment, ultrasonic power setting adjustment, and dispersant flow rate setting adjustment, a systematic search is performed within the maximum allowable positive and negative range of each adjustment step size to generate a set of different exploratory process parameter adjustment sequences. Using a dynamic prediction model, the evolution trajectories of the system state sequence, slurry dynamic adaptation factor K sequence, and dispersion potential function value sequence are predicted after applying different experimental process parameter adjustment sequences within a future control time domain of length H. Whether the future predicted value of the slurry dynamic adaptation factor K is within the preset adaptation range [K] min ,K max Within this range, it serves as the primary constraint for screening feasible candidate adjustment sequences; Among the candidate sequences that meet the primary constraints, the optimal process parameter adjustment sequence is obtained by solving an optimization problem that minimizes the sum of the dispersed potential function values ​​at the next H time points. The first adjustment instruction in the optimal process parameter adjustment sequence is output to the execution module as the collaborative adjustment instruction at the current control moment; At the next control time step, repeat the above prediction and optimization steps.

[0014] Preferred, fit range [K] min ,K max The value is [0.8, 1.2]. Furthermore, in the rolling optimization calculation, the following constraints are applied to each tentative process parameter adjustment sequence: Sequence adjustment amount constraint: In the trial process parameter adjustment sequence, the absolute values ​​of the mill speed setting value adjustment, ultrasonic power setting value adjustment, and dispersant flow rate setting value adjustment shall not exceed the preset single-step maximum adjustment amplitude, respectively; Total power constraint: The weighted sum of the mill speed setpoint and the ultrasonic power setpoint predicted based on the trial process parameter adjustment sequence shall not exceed the upper limit of the total power of the equipment; Total dispersant volume constraint: The sum of the cumulative volume of dispersant added from the start of the process to the current moment, plus the future volume of dispersant predicted based on the experimental process parameter adjustment sequence, shall not exceed 110% of the theoretical maximum required volume calculated based on the initial solid content characteristic value of the slurry; Particle size safety constraint: The maximum particle size characteristic value in the state sequence predicted by the trial process parameter adjustment sequence must not exceed the preset safety threshold.

[0015] This invention provides a method for real-time control of the dispersion process of carbon nanotube and carbon black composite slurry, including: Real-time acquisition of slurry status signals and process parameters at multiple stages of the dispersion process; Based on the slurry state signal and process parameters, a dynamic state vector sequence characterizing the real-time evolution of the dispersion process is generated. The storage contains historical distributed cases containing raw material characteristic vectors, process state matrices, process parameter matrices, and performance evaluation vectors; Execute closed-loop control logic: Based on the dynamic state vector sequence and raw material characteristic vector of the current batch, retrieve a set of similar historical process segments from the case knowledge base; based on the retrieved set of similar historical process segments, construct a dynamic prediction model for predicting the effect of process parameter adjustment; with the goal of minimizing the dispersion potential function, combine the dynamic prediction model and generate multi-process parameter coordinated adjustment instructions through rolling optimization calculation; It executes commands to coordinate the adjustment of multiple process parameters and intervenes in the decentralized process in real time.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: By constructing a closed-loop intelligent control architecture encompassing signal acquisition, state quantification, case retrieval, intelligent decision-making, and execution, the deficiencies of the aforementioned existing technologies are effectively overcome. The system achieves real-time quantitative perception and tracking of the dynamic states of multiple stages in the dispersion process and innovatively introduces a similarity matching and prediction model based on a historical dispersion case library, enabling the control strategy to dynamically adapt to raw material characteristics and real-time operating conditions. In particular, by focusing on minimizing a multi-objective dispersion potential function, including the slurry dynamic adaptation factor K, and employing model predictive control for rolling optimization, the system ultimately outputs multi-parameter coordinated adjustment commands for milling speed, ultrasonic power, and dispersant flow rate. This scheme significantly improves the uniformity and batch stability of composite slurry dispersion, achieving efficient dispersion while intelligently balancing energy consumption and material structure protection needs, thus propelling the production of carbon nanotube and carbon black composite slurries towards a high-quality, high-consistency, and low-loss intelligent direction.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall architecture diagram of the real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry in an embodiment of the present invention. Figure 2 This is a flowchart of the rolling optimization control in an embodiment of the present invention; Figure 3 This is a logic diagram for calculating the adaptation factor K in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 As shown, this invention provides an embodiment of a real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry, comprising: The signal acquisition module is used to acquire slurry status signals and process parameters in real time at multiple process stages during the dispersion process; The state quantization module is used to generate a dynamic state vector sequence that characterizes the real-time evolution of the dispersion process based on the slurry state signal and process parameters. The case knowledge base is used to store historical, distributed cases containing raw material characteristic vectors, process state matrices, process parameter matrices, and performance evaluation vectors. The intelligent decision-making module is used to execute closed-loop control logic: based on the dynamic state vector sequence and raw material characteristic vector of the current batch, it retrieves a set of similar historical process segments from the case knowledge base; based on the retrieved set of similar historical process segments, it constructs a dynamic prediction model to predict the effect of process parameter adjustment; with the goal of minimizing the dispersion potential function, it combines the dynamic prediction model and generates multi-process parameter coordinated adjustment instructions through rolling optimization calculation. The execution module is used to execute instructions for the coordinated adjustment of multiple process parameters and to intervene in the decentralized process in real time.

[0022] In this embodiment, the process steps specifically refer to the three core physical dispersion units of sand milling, ultrasonication, and shearing that are carried out sequentially on an industrial continuous production line. The system deploys sensors at the outlet of each unit to collect phased slurry status data.

[0023] In this embodiment, the dynamic prediction model refers to a local model within the intelligent decision-making module, built using machine learning methods, used to predict the effects of process parameter adjustments. The model's construction process is as follows: First, the system identifies several historical data segments from a historical case library that are most similar to the current production process. Then, the data from these historical segments is organized into training samples—each sample uses the system state characteristics (such as particle size and energy input at each stage) and raw material characteristics at a specific historical moment as input, and the actual, verified, and effective process parameter adjustment amount adopted at that historical moment as the output target. Finally, these samples are used to train a Gaussian process regression model. The trained model can predict how the system state might evolve if a certain parameter adjustment is made, based on the current real-time state and raw material information, providing a predictive basis for subsequent optimization decisions.

[0024] In this embodiment, the multi-process parameter coordinated adjustment command refers to a set of control commands generated by the optimizer in the intelligent decision-making module, which simultaneously change three key parameters: mill speed, ultrasonic power, and dispersant flow rate. The generation of this command is not a simple combination of independent decisions for each of the three parameters, but rather a coordinated optimization solution obtained through rolling calculations within a model predictive control optimization framework. This optimization aims to minimize a dispersion potential function that integrates multiple objectives such as dispersion quality, energy consumption, and material protection, while considering various process and equipment safety constraints. For example, a single command might simultaneously require appropriately increasing the mill speed to enhance shear, slightly decreasing the ultrasonic power to save energy, and slightly increasing the dispersant flow rate to maintain system stability. This reflects the linkage and balance between parameters.

[0025] In this embodiment, executing the multi-process parameter coordinated adjustment command means that the execution module converts the digital command into actual control signals for the sand mill, ultrasonic generator and metering pump, thereby completing the closed-loop real-time control of the dispersion process.

[0026] To achieve accurate and synchronous acquisition of core physical parameters across all stages and multiple modes in a distributed process, a signal acquisition module is proposed, including: The first laser particle size sensor is deployed at the outlet of the sand milling stage to collect slurry particle size distribution data in the sand milling stage; The second laser particle size sensor is deployed at the outlet of the ultrasonic process to collect slurry particle size distribution data during the ultrasonic process. The third laser particle size sensor is deployed at the outlet of the shearing stage to collect slurry particle size distribution data during the shearing stage. A sand mill power transmitter is connected to the sand mill driver to collect the sand mill spindle speed and torque in real time, or to directly collect the sand mill drive power. An ultrasonic power monitor, connected to an ultrasonic generator, is used to collect the output electrical power of the ultrasonic generator in real time. Among them, the slurry state signals of multiple process stages include slurry particle size distribution data of the sand milling stage, ultrasonic stage, and shearing stage; Process parameters include the spindle speed and torque of the sand mill or the drive power of the sand mill, as well as the output power of the ultrasonic generator.

[0027] In this embodiment, the outlet of the sand milling stage refers to the pipe or interface position where the slurry enters the next processing unit after it has been processed by the sand mill. The first laser particle size sensor is installed here to collect real-time slurry data after sand milling.

[0028] In this embodiment, the slurry particle size distribution data in the sand milling stage refers to the data measured in real time by the sensor installed at the outlet of the sand milling stage, which characterizes the statistical characteristics of particle size in the slurry after sand milling, and is used to evaluate the dispersion effect of the sand mill.

[0029] In this embodiment, the ultrasonic stage outlet refers to the outflow position of the slurry after ultrasonic dispersion treatment and before entering the shearing stage. The second laser particle size sensor is installed here to collect the state data of the slurry after ultrasonic treatment.

[0030] In this embodiment, the slurry particle size distribution data in the ultrasonic process refers to the particle size distribution information measured in real time by the sensor installed at the outlet of the ultrasonic process, which is used to evaluate the dissociation effect of ultrasonic cavitation on fine agglomerates and the degree of system stabilization.

[0031] In this embodiment, the shearing stage outlet refers to the physical location of the slurry after it has completed the final high-speed shearing homogenization treatment and before it enters the finished product tank or the next process. The third laser particle size sensor is installed here to collect slurry data representing the final dispersion state.

[0032] In this embodiment, the slurry particle size distribution data in the shearing stage refers to the particle size distribution information obtained by the sensor installed at the outlet of the shearing stage. This data characterizes the final state of the slurry after the entire process of dispersion and is a key indicator for judging batch quality.

[0033] In this embodiment, the sand mill driver refers to the electrical transmission device that drives the main shaft of the sand mill to rotate. It receives speed commands from the control system and provides feedback on the operating status. It is the actuator that adjusts the mechanical energy input of the sand mill.

[0034] In this embodiment, the spindle speed and torque of the sand mill are two key physical quantities that describe the operating status of the sand mill. The speed determines the shearing frequency, and the torque reflects the load size. The product of the two reflects the input mechanical power and is used for system calculation of energy consumption and status.

[0035] In this embodiment, the driving power of the sand mill refers to the electrical power consumed by the sand mill driver, which can be directly measured by the power transmitter and used as the basis for calculating the total energy consumption and monitoring the equipment load of the system.

[0036] In this embodiment, the ultrasonic generator refers to a power electronic device that converts power frequency electrical energy into high frequency electrical energy to drive a transducer to generate ultrasonic waves. Its output power is controllable and it is the energy source for the ultrasonic dispersion process.

[0037] In this embodiment, the output power of the ultrasonic generator refers to the high-frequency power actually output by the ultrasonic generator to the transducer. This data is collected in real time by the power monitor and is used to quantify the energy input of the ultrasonic link and to manage the total power of the system.

[0038] To transform high-dimensional, heterogeneous real-time data into standardized feature sequences capable of characterizing the dynamic evolution of the entire process, a method for generating dynamic state vector sequences using a state quantization module is proposed, including: Using a fixed duration T as the time window, feature extraction is performed on the slurry particle size distribution data of each process step within the current time window to obtain the feature vector within the step. The feature vector within the step includes the distribution width feature value, the main peak position feature value, and the distribution symmetry feature value. Calculate the ratio of the difference in the characteristic values ​​of the main peak positions between adjacent process steps to obtain the coupling characteristic values ​​between the steps; The characteristic value of sand mill power can be calculated based on the spindle speed and torque of the sand mill, or the driving power of the sand mill can be directly used as the characteristic value of sand mill power. Obtain the characteristic value of the output electrical power of the ultrasonic generator within the same time window; The characteristic values ​​of sand milling power and ultrasonic power are used as the energy input characteristic values; The intra-process feature vectors, inter-process coupling feature values, and energy input feature values ​​of all process steps within each time window are combined in a preset order to form a state feature vector; The state feature vectors of consecutive time windows are arranged in chronological order to form a dynamic state vector sequence.

[0039] In this embodiment, the fixed duration T refers to the basic time period used by the system for data acquisition and feature calculation, such as 30 seconds or 1 minute. This period is configurable and is used to divide the continuous real-time data stream into a series of analyzable time periods.

[0040] In this embodiment, feature extraction is performed on the slurry particle size distribution data of each process step within the current time window to obtain the feature vector within the step. This means that within each fixed time period T, all particle size data collected by the outlet sensor of a certain step (such as sand mill) are statistically analyzed to extract several key digital features, and these feature values ​​are arranged in order to form a set of numbers representing the dispersion state of the step in the current time period.

[0041] In this embodiment, the distribution width feature value, the main peak position feature value, and the distribution symmetry feature value are the three core values ​​that constitute the feature vector within the segment.

[0042] The distribution width feature is used to quantify the dispersion of particle size distribution, and is usually obtained by calculating the span. Specifically, the span value is a ratio calculated using data from a specific percentage point in the particle size distribution. The smaller the ratio, the more concentrated the particle size distribution and the better the uniformity of the slurry.

[0043] The dominant peak position characteristic value refers to the particle size value corresponding to the highest peak in the particle size distribution curve, and its unit is usually micrometer or nanometer. It reflects the most dominant size range in the particle system, and its variation trend can be used to indicate the progress of the dispersion process or the occurrence of agglomeration.

[0044] The symmetry eigenvalue of distribution is used to describe the degree to which the shape of the distribution curve deviates from symmetry, and is often measured by the statistical concept of skewness. If the skewness is negative, it means that the distribution curve tails to the left, that is, extends towards smaller particle sizes, which may indicate that there are a large number of fine particles in the slurry or that the material has been sufficiently broken down. If the skewness is positive, it means that the distribution curve tails to the right, that is, extends towards larger particle sizes, which may indicate that there are insufficiently dispersed agglomerates in the slurry.

[0045] In this embodiment, calculating the difference ratio of the main peak position feature values ​​between adjacent process steps to obtain the inter-step coupling feature value means comparing the main peak position feature value of the previous step (such as sand milling) within the current time window with the main peak position feature value of the adjacent subsequent step (such as ultrasonication) within the same time window, and calculating their relative change ratio. This ratio is used to quantify the connection and transition state of the processing effects of the two processes.

[0046] In this embodiment, calculating the sand mill power characteristic value based on the sand mill spindle speed and torque refers to calculating the average mechanical power of the input slurry within each fixed time window T, based on the real-time collected data sequence of the sand mill spindle speed (unit: revolutions per minute) and torque (unit: Newton-meters). The calculation follows the physical principle that rotating mechanical power equals the product of speed and torque, multiplied by a constant factor (2π / 60) determined by unit conversion and geometric relationships. Within the time window T, the instantaneous power values ​​calculated according to this principle are averaged to obtain the sand mill power characteristic value for that window. This characteristic value directly characterizes the intensity of mechanical energy input applied to the slurry during the sand milling process within that time period and is a core component of the system's energy input characteristics.

[0047] In this embodiment, obtaining the characteristic value of the output electrical power of the ultrasonic generator within the same time window means: averaging the output electrical power data of the ultrasonic generator collected within the same fixed time period T to obtain a characteristic value representing the energy input level of the ultrasonic process during that time period.

[0048] In this embodiment, the state feature vector is formed by combining the intra-process feature vectors, inter-process coupling feature values, and energy input feature values ​​of all process steps within each time window in a preset order. This means that all features extracted within a time period T (the feature vectors of each of the three steps, the coupling features between the two steps, the grinding power, and the ultrasonic power features) are concatenated together in a predefined order to form a comprehensive digital vector. This vector fully represents the overall operating state of the system within that time period.

[0049] To structure and store historical successes in a way that creates a reusable and searchable process knowledge base, a case knowledge base is proposed, in which historical, dispersed cases include: Raw material characteristic vector, process state matrix, process parameter matrix, performance evaluation vector; Among them, the raw material characteristic vector includes the aspect ratio characteristic value of carbon nanotubes, the specific surface area characteristic value of carbon black, and the initial solid content characteristic value of slurry. The rows of the process state matrix correspond to the time series, and the columns correspond to the eigenvalues ​​of each dimension of the state feature vector; The rows of the process parameter matrix correspond to the time series synchronized with the process state matrix, and the columns include the milling speed setting, ultrasonic power setting, and dispersant flow rate setting. The performance evaluation vector includes the slurry conductivity score and the slurry static stability score.

[0050] In this embodiment, the aspect ratio of carbon nanotubes, the specific surface area of ​​carbon black, and the initial solid content of the slurry are the three core components of the raw material characteristic vector. Specifically, the aspect ratio of carbon nanotubes represents the ratio of the average length to the diameter of the carbon nanotubes used; the specific surface area of ​​carbon black represents the total surface area per unit mass of the carbon black used; and the initial solid content of the slurry represents the percentage of the total mass of solid particles in the slurry before dispersion begins. These three values ​​collectively characterize the inherent properties of the batch material and form the basis for personalized control system adjustments.

[0051] In this embodiment, the feature values ​​of each dimension of the state feature vector refer to the specific numerical values ​​corresponding to each position in the state feature vector, which is composed of elements in a preset order, and the physical meaning they represent. For example, the first value of the vector may represent the particle size distribution width of the grinding stage, the second value may represent the main peak position of the grinding stage, and so on, until it includes the features of all stages, the coupling features between stages, and the energy input features. The feature value of each dimension is a processed scalar, which together constitute a complete digital description of the system's operating state at a certain moment.

[0052] In this embodiment, the mill speed setpoint, ultrasonic power setpoint, and dispersant flow rate setpoint refer to three core control command values ​​actively applied to the production equipment at each moment, stored in the process parameter matrix of the historical case library. The mill speed setpoint controls the target speed of the mill; the ultrasonic power setpoint controls the target output power of the ultrasonic generator; and the dispersant flow rate setpoint controls the target flow rate of the dispersant addition pump. The historical sequence of these setpoints records the control strategies in successful production processes.

[0053] In this embodiment, the slurry conductivity score and the slurry settling stability score are two core indicators of the performance evaluation vector, used to quantify the final application performance of a batch of slurry. The slurry conductivity score is a normalized score (e.g., 0-100 points) obtained by measuring the volumetric conductivity of the finished slurry using the four-probe method or AC impedance method, and then converting it according to a preset performance standard. Conductivity directly reflects the formation quality of the carbon nanotube and carbon black conductive network; a higher score indicates more uniform dispersion and a more complete conductive pathway. The slurry settling stability score is a normalized score obtained by placing the finished slurry in standardized test tubes and setting it for a specified time (e.g., 24 hours, 7 days), measuring its sedimentation layer height or the proportion of the supernatant, and then converting it according to a stability standard. This score reflects the dispersion system's resistance to agglomeration and sedimentation; a higher score indicates better dispersion stability and a longer shelf life. These two scores together constitute a quantitative evaluation of the final result of a historical dispersion case, stored in the case knowledge base, used to correlate the process with the final performance, and to guide the optimization direction of future batches.

[0054] like Figure 2As shown, to quickly and accurately locate the process segment most similar to the current operating condition from massive historical data, the intelligent decision-making module is proposed to retrieve a set of similar historical process segments from the case knowledge base, including: The sequence of dynamic state vectors generated in the current batch is recorded as the query sequence; For each historical scattered case in the case knowledge base, extract the submatrix in the process state matrix of the historical scattered case that has the same time length as the query sequence as the candidate sequence; Calculate the dynamic time-normalized distance between the query sequence and each candidate sequence; A comprehensive similarity score is calculated based on the dynamic time warp distance and the Euclidean distance between raw material characteristic vectors. Select the N historical scattered cases with the highest comprehensive similarity scores, and extract the process state subsequence and process parameter subsequence of the N historical scattered cases after the matching period to form a set of similar historical process segments.

[0055] In this embodiment, for each historical scattered case in the case knowledge base, extracting a submatrix of the process state matrix of the historical scattered case with the same time length as the query sequence as a candidate sequence means that when the system needs to evaluate the similarity with the current batch (which has been performed for time t and corresponds to a query sequence length of L), it will traverse every historical case in the knowledge base. For each historical case, the algorithm will start from the first row of its process state matrix, slide a window of length L, and extract all rows within the window (i.e., L consecutive state feature vectors) to form a candidate sequence with the same dimension and time length as the query sequence. This operation essentially cuts the long historical process into segments of the same length as the current query for comparison one by one.

[0056] In this embodiment, calculating the dynamic time warping distance between the query sequence and each candidate sequence refers to using a specialized algorithm for comparing time series similarity (dynamic time warping algorithm) to calculate the overall degree of difference between the current batch's state evolution trajectory (query sequence) and each historical candidate sequence. This algorithm can effectively handle potential velocity differences between two sequences on the time axis, and the calculation result is a distance value. The smaller the value, the more similar the morphology and trend of the two sequences are.

[0057] In this embodiment, the Euclidean distance between raw material characteristic vectors refers to the straight-line distance between the raw material characteristic vector of the current batch (including aspect ratio, specific surface area, and solid content) and the raw material characteristic vector of historical cases. This is a simple mathematical method to measure the difference between the two in terms of their basic raw material properties; the smaller the distance, the closer the raw materials are.

[0058] In this embodiment, calculating the comprehensive similarity score based on the dynamic time warp distance and the Euclidean distance between raw material characteristic vectors involves normalizing the two distance metrics (process evolution similarity and raw material attribute similarity) and combining them according to a preset weight ratio to calculate a single score. This score considers both the dynamic process and the static raw material similarity; a higher score indicates a higher comprehensive similarity between the historical case and the current production scenario.

[0059] In this embodiment, extracting the process state subsequences and process parameter subsequences of N historical dispersed cases after the matching time period to form a set of similar historical process segments means that the system selects the N historical dispersed cases with the highest comprehensive similarity scores. For each selected case, not only is the historical data that matches the current state recorded, but more importantly, a complete record of what happened immediately after the matching time period is extracted, including the state evolution data and the process parameter data used at the same time. These K sets of subsequent evolution records constitute a set of similar historical process segments used to guide subsequent operations of the current batch.

[0060] like Figure 2 As shown, to establish a local dynamic model based on similar historical fragments that can accurately predict the effect of parameter adjustment, a method for constructing a dynamic prediction model using an intelligent decision-making module is proposed, including: Using the state feature vector and raw material characteristic vector of each similar historical process segment at time t in the set of similar historical process segments as the model input sample, and the milling speed adjustment, ultrasonic power adjustment, and dispersant flow rate adjustment of the corresponding similar historical process segment at time t+1 as the model output label, a multi-output Gaussian process regression model is trained as a dynamic prediction model. The multi-output Gaussian process regression model uses a combination of radial basis function kernel and white noise kernel kernel.

[0061] In this embodiment, the mill speed adjustment, ultrasonic power adjustment, and dispersant flow rate adjustment refer to the specific changes in the three key process parameters—mill speed, ultrasonic generator power, and dispersant addition flow rate—that the system needs to calculate and output at each control decision moment. These adjustment amounts are the direct content of the control command; positive values ​​represent increases, and negative values ​​represent decreases.

[0062] In this embodiment, the state feature vector and raw material characteristic vector of each similar historical process segment at time t in the set of similar historical process segments are used as model input samples. The mill speed adjustment, ultrasonic power adjustment, and dispersant flow rate adjustment of the corresponding similar historical process segment at time t+1 are used as model output labels. Training a multi-output Gaussian process regression model as a dynamic prediction model means that this process transforms historical successful experience into quantifiable predictive relationships. First, for each similar historical segment in the set (i.e., a record of a successful process), the algorithm traverses every time step from the beginning to the end of the segment. At each time step t, two key inputs are extracted: 1) the state feature vector recorded at time t (reflecting the system snapshot at that time); 2) the raw material characteristic vector of the case (reflecting the unchanging material basis). These two vectors are concatenated to form an input sample. At the same time, the three process parameter adjustments (ΔR, ΔP, ΔF) actually executed at time t+1 of the segment are extracted to form a three-dimensional output label. In this way, a large number of (input, output) sample pairs can be generated from all similar segments. Then, these samples are used to train a multi-output Gaussian process regression model. Through learning, this model establishes a complex nonlinear mapping from "current state + raw material characteristics" to "optimal parameter adjustment suggestions." Once training is complete, when faced with new, unseen states and raw materials, the model can predict the parameter adjustment combination most likely to achieve good results in the current situation.

[0063] In this embodiment, the multi-output Gaussian process regression model employs a combination of radial basis function kernel and white noise kernel. The kernel function refers to the core mathematical structure and characteristic definition of the selected prediction model. Multi-output Gaussian process regression is a probabilistic nonparametric model that provides not only the predicted mean (the most likely adjustment) but also the predicted variance (an estimate of uncertainty). The kernel function defines how the model measures the similarity between two input samples, thereby determining the output prediction value. In this embodiment, a combination of two kernel functions is used: 1) Radial basis function kernel, also known as the quadratic exponential kernel, is the most commonly used kernel function in Gaussian processes. This kernel function assumes that the closer two points are in the input space (state and raw material feature space), the more similar their output values ​​(adjustments) should be. It can smoothly capture the complex nonlinear relationship between input and output. 2) White noise kernel, which is superimposed on the radial basis function kernel, is used to represent the inherent noise level in the data, such as measurement errors or random fluctuations in the process itself. This is achieved by adding a diagonal term to the predicted covariance matrix, ensuring that the model does not overfit to small noises in the training data, thereby improving the model's robustness and generalization ability. This combined kernel function allows the model to learn complex trends while also reasonably representing uncertainty.

[0064] likeFigure 3 As shown, to construct a unified optimization objective function that comprehensively and quantitatively evaluates dispersion quality, energy efficiency, material damage, and process adaptability, a dispersion potential function E is proposed. This function is a function of the state characteristic vector s, the process parameter vector p, and the raw material characteristic vector r. Its construction and optimization objective are closely related to a slurry dynamic adaptability factor K, specifically: Real-time calculation of the dynamic adaptation factor K of the slurry: K = α × K0 + β × K s -γ×K c ; Where α, β, γ are the preset first set of weight coefficients; K0 is the raw material baseline term, which is calculated based on the characteristic values ​​of the initial solid content of the slurry, the aspect ratio of carbon nanotubes, and the specific surface area of ​​carbon black in the raw material characteristic vector. K s The real-time state correction term is calculated based on the difference ratio between the distribution width feature value and the main peak position feature value between adjacent process steps in the state feature vector. K c The process constraint penalty is calculated based on the ratio of the cumulative volume of dispersant added to the theoretical maximum required volume calculated based on the initial solid content of the slurry, and the ratio of the current total power to the upper limit of the total power of the equipment. The current total power is the sum of the sand mill drive power collected in real time by the signal acquisition module or the power calculated based on the speed and torque of the sand mill spindle and the output power of the ultrasonic generator. The dispersed potential function E is expressed as: E(s,p,r)=w1×f1(distribution width characteristic value deviation)+w2×f2(cumulative energy consumption)+w3×f3(length-to-diameter ratio change rate)+w4×f4(K); Among them, w1, w2, w3, and w4 are the preset second set of weight coefficients; f1 (distribution width eigenvalue deviation) is a monotonically increasing function of the deviation between the distribution width eigenvalue in the state eigenvector s and the preset target distribution width value; f2 (cumulative energy consumption) is a monotonically increasing function of the total energy consumption from the start of the distributed process to the current moment; f3 (aspect ratio change rate) is a monotonically increasing function of the ratio of the current estimated aspect ratio to the initial aspect ratio of carbon nanotubes; f4(K) is the absolute value of the difference between the dynamic adaptation factor K of the slurry and the value 1.

[0065] In this embodiment, the preset first set of weighting coefficients refers to three preset proportional coefficients used to balance the contributions of the raw material baseline item, the real-time status correction item, and the process constraint penalty item when calculating the adaptation factor used for comprehensive evaluation of process adaptability. These weighting coefficients are fixed values ​​predetermined through statistical analysis of a large amount of historical successful production data or through experimental design optimization, and are not dynamically adjusted in real-time control. For example, a possible set of values ​​is α=0.4, β=0.5, γ=0.3, which means that when evaluating the current process adaptability, the real-time status correction item is given the highest weight, followed by the raw material baseline item, and a certain degree of adjustment is given to the constraint penalty item. The specific values ​​of the weights reflect the degree of importance attached to different influencing factors, and their sum α+β is usually greater than γ to ensure that the main component of the adaptation factor K is contributed by the first two items.

[0066] In this embodiment, the calculation method for the raw material baseline term is based on the characteristic values ​​of the initial solid content of the slurry, the aspect ratio of carbon nanotubes, and the specific surface area of ​​carbon black in the raw material characteristic vector. This baseline term is a process of mapping the above three key raw material attributes to a scalar value using an empirical formula. This scalar value is used to characterize the basic dispersion difficulty or ideal process adaptation benchmark of the batch of raw materials. In a specific example, this empirical formula can be expressed as: calculating the raw material baseline term equals a first constant term, adding the initial solid content characteristic value of the slurry multiplied by a second coefficient, subtracting the aspect ratio characteristic value of carbon nanotubes multiplied by a third coefficient, and then subtracting the specific surface area characteristic value of carbon black multiplied by a fourth coefficient. The first constant term can be 1.0, the second coefficient can be 0.5, the third coefficient can be 0.3, and the fourth coefficient can be 0.2. All the above characteristic values ​​have been normalized to the range of 0 to 1 before being substituted into the calculation. The specific coefficient values ​​in the formula can be obtained by analyzing historical production data from a specific production line. This calculation method results in lower baseline values ​​for raw materials with higher solid content, larger carbon nanotube aspect ratios, and larger carbon black specific surface areas, indicating greater difficulty in basic dispersion. This baseline provides a static benchmark point for the entire adaptability factor that is closely related to the specific raw material properties.

[0067] In this embodiment, the calculation method for the real-time state correction term is based on the ratio of the difference between the distribution width feature value and the main peak position feature value between adjacent process steps in the state feature vector. This correction term reflects the degree of deviation of the current real-time state of the dispersion process from the ideal state. Its calculation is usually designed as a function with relevant state indicators as input. In a specific example, this function can be expressed as: the real-time state correction term is equal to the value 1 minus a comprehensive deviation; the comprehensive deviation is equal to the difference between the distribution width feature value of the current main step (such as the shear outlet) and the preset target value (e.g., 0.2) multiplied by a first weighting coefficient (e.g., 0.6), plus the average of the ratio of the main peak position difference between adjacent steps multiplied by a second weighting coefficient (e.g., 0.4). The design intent of this function is that when the particle size distribution width increases (i.e., uniformity deteriorates) or the connection between steps is not smooth (i.e., the main peak position change rate is large), the value of this correction term will decrease, thereby conveying a signal that the process needs to be adjusted through the adaptation factor.

[0068] In this embodiment, the calculation method for the process constraint penalty term is based on the ratio of the cumulative volume of dispersant added to the theoretical maximum required volume calculated based on the initial solid content of the slurry, and the ratio of the current total power to the upper limit of the total power of the equipment. This term aims to impose a negative penalty on the adaptation factor when the production process approaches or violates critical safety and resource constraints. Its calculation consists of two parts: the first part is the dispersant constraint, which calculates the ratio of the current total volume of dispersant added to the theoretical maximum allowable volume; the second part is the power constraint, which calculates the ratio of the current total input power of the equipment (i.e., the sum of the sand mill power and the ultrasonic power) to the upper limit of the system's safe power (e.g., 55 kW). This penalty term is usually designed as a function, and its specific form can be: the dispersant penalty component is calculated as the dispersant ratio minus a first trigger threshold (e.g., 0.9), and if the result is less than 0, it is taken as 0; the power penalty component is calculated as the power ratio minus a second trigger threshold (e.g., 0.85), and if the result is less than 0, it is taken as 0; the value of the process constraint penalty term is the sum of the above dispersant penalty component and the power penalty component. When any ratio exceeds its corresponding threshold, the penalty term becomes positive, thereby reducing the value of the fit factor.

[0069] In this embodiment, the current total power is the sum of the sand mill drive power acquired in real time by the signal acquisition module, or the power calculated based on the sand mill spindle speed and torque, and the output electrical power of the ultrasonic generator. This clarifies the physical definition and data source of the "current total power" used in calculating constraint penalty terms and managing total power. This total power is obtained by directly adding two parts: one part is the sand mill power, which can be obtained in two equivalent ways depending on the system configuration: either by directly reading the electrical power measured by the power transmitter, or by calculating the mechanical power based on the physical relationship of "the product of rotational speed (unit: revolutions per minute) and torque (unit: Newton-meters), multiplied by a constant determined by unit conversion and pi (this constant is approximately 2 / 60 π)"; the other part is the output electrical power of the ultrasonic generator measured by the ultrasonic power monitor. This explicit definition ensures the consistency of energy consumption calculation and power constraint verification throughout the system.

[0070] In this embodiment, the preset second set of weighting coefficients refers to four preset proportional coefficients used to balance the importance of the four sub-objectives—particle size uniformity deviation, cumulative production energy consumption, carbon nanotube structural damage, and process adaptation factor deviation—when constructing the overall optimization objective function, the dispersion potential function. By adjusting these coefficients, the priority given to different performance indicators in production can be reflected.

[0071] In this embodiment, the monotonically increasing function relating the deviation between the distribution width feature value and the preset target distribution width value in the state feature vector s refers to a sub-term in the optimization objective function whose value monotonically increases as the current slurry particle size distribution width deviates from its ideal target value. Specifically, it can be a square function of the deviation value. For example, squaring the difference between the current distribution width feature value and the target value yields the value of this sub-term. This means that the more uneven the distribution, the greater the penalty of this term, thereby driving the system to adjust the process to make the distribution more concentrated.

[0072] In this embodiment, the monotonically increasing function of total energy consumption from the start of the decentralized process to the current moment refers to a sub-term in the optimization objective function that monotonically increases with the increase of the total electrical energy consumed by all process equipment (mainly sand mills and ultrasonic generators) in the production process. Specifically, it can be a linear function of the cumulative energy consumption. For example, multiplying the total energy consumption by a preset cost coefficient yields the value of this sub-term. This aims to incorporate energy-saving objectives into the optimization, prompting the system to find more energy-efficient process paths while ensuring quality.

[0073] In this embodiment, the monotonically increasing function relating the current estimated aspect ratio to the initial aspect ratio of the carbon nanotubes refers to an objective function containing a sub-term whose value monotonically increases as the degree of breakage of the carbon nanotubes due to processing increases (manifested as a decrease in the current estimated aspect ratio relative to the initial value). Specifically, it can be a square function of the deviation in aspect ratio retention rate. For example, first calculate "one minus the ratio of the current aspect ratio to the initial aspect ratio," then square the result; the resulting value is the value of this sub-term. This aims to prioritize the protection of the expensive carbon nanotube structure and the prevention of excessive damage as a key optimization objective.

[0074] like Figure 2 As shown, to calculate the optimal collaborative adjustment strategy that takes multiple objectives into account in the future time period through forward-looking rolling optimization, an intelligent decision-making module is proposed to generate collaborative adjustment instructions for multiple process parameters through rolling optimization calculation, which is implemented using a model predictive control framework. The specific steps include: At each control moment, starting with the current state feature vector and process parameter settings, a set of different exploratory process parameter adjustment sequences are generated within the search space defined below: Based on the current mill speed setting, ultrasonic power setting, and dispersant flow rate setting, within the three-dimensional space formed by the mill speed setting adjustment, ultrasonic power setting adjustment, and dispersant flow rate setting adjustment, a systematic search is performed within the maximum allowable positive and negative range of each adjustment step size to generate a set of different exploratory process parameter adjustment sequences. Using a dynamic prediction model, the evolution trajectories of the system state sequence, slurry dynamic adaptation factor K sequence, and dispersion potential function value sequence are predicted after applying different experimental process parameter adjustment sequences within a future control time domain of length H. Whether the future predicted value of the slurry dynamic adaptation factor K is within the preset adaptation range [K] min ,K max Within this range, it serves as the primary constraint for screening feasible candidate adjustment sequences; Among the candidate sequences that meet the primary constraints, the optimal process parameter adjustment sequence is obtained by solving an optimization problem that minimizes the sum of the dispersed potential function values ​​at the next H time points. The first adjustment instruction in the optimal process parameter adjustment sequence is output to the execution module as the collaborative adjustment instruction at the current control moment; At the next control time step, repeat the above prediction and optimization steps.

[0075] In this embodiment, the control moment refers to the point in time when the system automatically triggers and executes a complete "acquisition-decision-output" cycle according to a fixed period (e.g., every 10 seconds or 30 seconds). Each control moment is a decision-making rhythm for the system to intervene in real time.

[0076] In this embodiment, based on the current mill speed setting, ultrasonic power setting, and dispersant flow rate setting, a systematic search is performed within the three-dimensional space formed by the adjustment amounts of the mill speed setting, ultrasonic power setting, and dispersant flow rate setting, using a preset adjustment step size, within the maximum allowable positive and negative range of each adjustment amount. This generates a set of different exploratory process parameter adjustment sequences. Specifically, at each control moment, the optimizer, starting from the current equipment settings, systematically enumerates multiple possible parameter adjustment scheme combinations for future times within the allowed safe adjustment range and according to fixed change intervals. These combinations constitute the set of candidate control strategies to be evaluated. The operation steps are as follows: First, the three settings actually executed by the equipment at the current moment are used as reference points. Then, for the three decision variables—mill speed adjustment, ultrasonic power adjustment, and dispersant flow rate adjustment—a series of equally spaced possible values ​​are generated within the interval defined by the maximum allowable positive and negative change values ​​of each variable, according to a preset fixed change unit. For example, if the maximum allowable adjustment of the mill speed is 100 rpm in the positive direction and 100 rpm in the negative direction, and the preset adjustment step size is 20 rpm, then the possible values ​​for the mill speed adjustment include a series of discrete values ​​from -100, -80, down to zero, and then to +80, +100. Next, all possible discrete values ​​of the three decision variables are combined to form a three-dimensional grid set. Each grid point corresponds to a specific combination of single-step control actions that may be applied at the next control time. Finally, by combining the current baseline setpoint with these combinations of single-step control actions, and assuming that a certain strategy is adopted for these control actions in a subsequent prediction time domain (e.g., keeping the combination unchanged, or changing it according to a certain rule), a set of different tentative process parameter adjustment sequences spanning multiple future control times can be generated as input for subsequent model predictions.

[0077] In this embodiment, the preset adjustment step size refers to the interval unit set for the adjustment of each process parameter in the above-mentioned systematic search (for example, the rotation speed changes by 50 revolutions per minute each time, and the power changes by 10 watts each time). The step size determines the fineness of the search.

[0078] In this embodiment, a dynamic prediction model is used to predict the evolution trajectory of the system state sequence, the slurry dynamic adaptation factor K sequence, and the dispersion potential function value sequence after applying different experimental process parameter adjustment sequences within a future control time domain of length H. This means that for each candidate parameter adjustment scheme, the system uses a pre-trained dynamic prediction model to simulate and calculate how the state of the entire dispersion process, the comprehensive adaptation index, and the optimization objective function value will change step by step within a predetermined future time period if the scheme is executed. This allows for the evaluation of the expected effect of each scheme.

[0079] In this embodiment, the future predicted value of the slurry dynamic adaptation factor K is checked against a preset adaptation range [K]. min ,K max Within this framework, the primary constraint for screening feasible candidate adjustment sequences is that the system sets a acceptable range for process adaptability. When evaluating candidate solutions, the first step is to check whether the predicted future process ensures that key adaptability indicators consistently fall within this acceptable range. If the predicted value exceeds this range at any given time, the solution is deemed infeasible. This is the first hurdle to ensuring the process rationality of the control scheme.

[0080] In this embodiment, the candidate sequences that satisfy the primary constraints refer to those parameter adjustment scheme sequences that have passed the above-mentioned adaptability check, which constitute the set of feasible solutions for subsequent selection of the best among the best.

[0081] In this embodiment, "obtaining the optimal process parameter adjustment sequence by solving an optimization problem that minimizes the sum of the dispersed potential function values ​​over the next H time steps" is the core mathematical description of the "optimization" stage in rolling optimization. The objective function J of the optimization problem is defined as follows: for each candidate adjustment sequence U that satisfies the primary constraint, the predicted dispersed potential function values ​​E corresponding to it over the next H control time steps are summed, and the sum is the objective value J(U) of that sequence. Solving this optimization problem means finding a specific optimal adjustment sequence U among all feasible candidate sequences that minimizes its objective function value J(U). This sequence U... * This refers to the optimal process parameter adjustment sequence that, based on the prediction of future H steps, can most effectively optimize multiple comprehensive objectives such as particle size uniformity, energy consumption, material protection, and process adaptability at the current control moment.

[0082] In this embodiment, outputting the first adjustment instruction in the optimal process parameter adjustment sequence as the collaborative adjustment instruction for the current control moment to the execution module means that although the optimal strategy is a complete plan covering multiple future moments, the system adopts a rolling execution strategy, that is, only sending the first step operation instruction to be executed immediately in the optimal plan to the production line equipment. This ensures the real-time performance of the control.

[0083] In this embodiment, repeating the above prediction and optimization steps at the next control moment means that when the next decision cycle is reached, the system will perform a complete prediction and optimization calculation again based on the latest production site data and output new real-time instructions. This process is repeated continuously to achieve dynamic closed-loop optimization control of the entire decentralized process.

[0084] To ensure that the optimization process always operates within safe, feasible, and process-adaptable boundaries, and to guarantee the robustness of control, a fit interval [K] is proposed. min ,K max The value is [0.8, 1.2]. Furthermore, in the rolling optimization calculation, the following constraints are applied to each tentative process parameter adjustment sequence: Sequence adjustment amount constraint: In the trial process parameter adjustment sequence, the absolute values ​​of the mill speed setting value adjustment, ultrasonic power setting value adjustment, and dispersant flow rate setting value adjustment shall not exceed the preset single-step maximum adjustment amplitude, respectively; Total power constraint: The weighted sum of the mill speed setpoint and the ultrasonic power setpoint predicted based on the trial process parameter adjustment sequence shall not exceed the upper limit of the total power of the equipment; Total dispersant volume constraint: The sum of the cumulative volume of dispersant added from the start of the process to the current moment, plus the future volume of dispersant predicted based on the experimental process parameter adjustment sequence, shall not exceed 110% of the theoretical maximum required volume calculated based on the initial solid content characteristic value of the slurry; Particle size safety constraint: The maximum particle size characteristic value in the state sequence predicted by the trial process parameter adjustment sequence must not exceed the preset safety threshold.

[0085] In this embodiment, the absolute values ​​of the mill speed setting adjustment, ultrasonic power setting adjustment, and dispersant flow rate setting adjustment not exceeding the preset single-step maximum adjustment amplitude means that, each time a control command is output, the maximum change in the adjustment range of these three parameters—mill speed, ultrasonic power, and dispersant flow rate—must not exceed their respective preset safety limits. This constraint aims to prevent overly drastic control actions and ensure smooth equipment operation and stable process flow.

[0086] In this embodiment, the preset maximum single-step adjustment amplitude refers to the maximum value that is allowed to be changed in a single control command for each of the three adjustable parameters mentioned above. This value is a safety parameter of the system, preset according to equipment performance and process sensitivity, for example, the rotational speed is adjusted no more than 100 revolutions per minute in a single adjustment.

[0087] In this embodiment, the weighted sum of the mill speed setpoint and ultrasonic power setpoint predicted based on the trial process parameter adjustment sequence, not exceeding the upper limit of the total equipment power, means that when evaluating each candidate future parameter adjustment scheme, the system needs to predict the combined load level of the mill and ultrasonic generator after implementing the scheme. This load level (calculated by weighting the speed and power) must not exceed the upper limit of the total power supply or design power of the equipment system at any future time to ensure electrical safety.

[0088] In this embodiment, the sum of the cumulative volume of dispersant added from the start of the process to the current moment, plus the future volume predicted based on the experimental process parameter adjustment sequence, must not exceed 110% of the theoretical maximum required volume calculated based on the initial solids content characteristic value of the slurry. This is an economic and safety constraint regarding the total amount of chemicals consumed. This constraint limits the total amount of dispersant used from the perspective of the entire batch. It consists of three parts: 1) Historical cumulative amount: The total volume of dispersant actually added from the start of the batch to the current control moment, denoted as V. used hist 2) Future Predicted Volume: For each trial adjustment sequence, based on the predicted dispersant flow rate at each future control time, calculate the total additional volume to be added in the prediction time domain, denoted as V. future pred The calculation method is as follows: multiply the predicted dispersant flow rate at each control time point within the prediction time domain by the control cycle duration to obtain the predicted addition volume at that time point. Then, sum the predicted addition volumes at all control times within the prediction time domain, and the sum is V. future pred 3) Theoretical maximum required volume: The maximum volume of dispersant theoretically required to complete the dispersion of this batch, calculated in advance based on the initial solid content of the slurry, the formulation ratio, and process experience. This is denoted as V. max theory The specific requirement of this constraint is: historical cumulative amount V used hist With future forecast V future pred The sum obtained by adding them together must be less than or equal to the theoretical maximum required volume V. max theory 1.1 times (i.e., 110%). The 1.1 times factor serves as a safety factor.

[0089] In this embodiment, the cumulative dispersant addition volume from the start of the process to the current moment refers to the total fluid volume corresponding to all executed dispersant addition commands from the start of the current batch dispersion process until the current control moment. This data is accumulated and recorded by the system in real time and forms the basis for executing the aforementioned total quantity constraint. The accumulation method is as follows: within each past control cycle, the dispersant flow rate setpoint output by the system in that cycle is multiplied by the control cycle duration to obtain the addition volume for that cycle; then, the addition volumes of all control cycles from the start of the process to the current moment are added sequentially, and the sum is the cumulative addition volume. It reflects the amount of dispersant resources consumed up to the present.

[0090] In this embodiment, the predicted future addition volume based on the trial process parameter adjustment sequence is an estimate of the total amount of dispersant to be added within the prediction time domain (i.e., several consecutive control cycles in the future) for a specific trial adjustment sequence. The calculation method is as follows: First, the predicted dispersant flow rate value for each control moment within the future prediction time domain is extracted from the trial sequence, forming a prediction value sequence. Then, for each flow rate prediction value in the sequence, it is multiplied by the control cycle duration to obtain the predicted addition volume for the corresponding moment. Finally, the predicted addition volumes for all control moments within the prediction time domain are summed, and the total sum is the predicted future addition volume. This predicted value is used to assess the incremental impact on overall dispersant consumption if the trial sequence is executed.

[0091] In this embodiment, the theoretical maximum required volume, calculated based on the initial solids content characteristic value of the slurry, is an estimate of the maximum reference amount of dispersant required to achieve effective dispersion, based on the inherent properties (initial solids content) of this batch of slurry and process knowledge. A common calculation method is as follows: the theoretical maximum required volume equals an empirical coefficient K, multiplied by the initial solids content of the slurry (expressed as a mass fraction), multiplied by the total mass of the batch of slurry, and finally divided by the density of the dispersant. Here, the empirical coefficient K represents the mass or volume ratio of dispersant typically required per unit mass of solid particles. This coefficient is obtained through regression analysis of a large amount of historically successful formulation production data. The theoretical maximum required volume provides a dosage benchmark based on material balance and long-term process experience to constrain total consumption and prevent cost increases and potential negative impacts on slurry performance (such as conductivity) due to excessive addition.

[0092] In this embodiment, the maximum particle size characteristic value in the state sequence predicted based on the trial process parameter adjustment sequence refers to: for each candidate future parameter adjustment scheme, the system uses a dynamic prediction model to calculate the particle state of the slurry over a future period of time. From these predicted states, the maximum possible particle size value (e.g., the predicted maximum D90 value) is identified.

[0093] In this embodiment, the preset safety threshold refers to an absolute upper limit set for the predicted maximum particle size characteristic value. This constraint requires that the maximum particle size predicted by any candidate scheme must not exceed this threshold, which is the bottom line requirement to ensure that the final slurry product does not experience unacceptable agglomeration or quality degradation. For example, for a slurry with a target D90 of 500 nm, the safety threshold may be set at 800 nm.

[0094] This invention provides an embodiment of a method for real-time control of the dispersion process of carbon nanotube and carbon black composite slurry, comprising: Real-time acquisition of slurry status signals and process parameters at multiple stages of the dispersion process; Based on the slurry state signal and process parameters, a dynamic state vector sequence characterizing the real-time evolution of the dispersion process is generated. The storage contains historical distributed cases containing raw material characteristic vectors, process state matrices, process parameter matrices, and performance evaluation vectors; Execute closed-loop control logic: Based on the dynamic state vector sequence and raw material characteristic vector of the current batch, retrieve a set of similar historical process segments from the case knowledge base; based on the retrieved set of similar historical process segments, construct a dynamic prediction model for predicting the effect of process parameter adjustment; with the goal of minimizing the dispersion potential function, combine the dynamic prediction model and generate multi-process parameter coordinated adjustment instructions through rolling optimization calculation; It executes commands to coordinate the adjustment of multiple process parameters and intervenes in the decentralized process in real time.

[0095] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry, characterized in that, include: The signal acquisition module is used to acquire slurry status signals and process parameters in real time at multiple process stages during the dispersion process; The state quantization module is used to generate a dynamic state vector sequence that characterizes the real-time evolution of the dispersion process based on the slurry state signal and process parameters. The case knowledge base is used to store historical, distributed cases containing raw material characteristic vectors, process state matrices, process parameter matrices, and performance evaluation vectors. The intelligent decision-making module is used to execute closed-loop control logic: based on the dynamic state vector sequence and raw material characteristic vector of the current batch, it retrieves a set of similar historical process segments from the case knowledge base; based on the retrieved set of similar historical process segments, it constructs a dynamic prediction model to predict the effect of process parameter adjustment; with the goal of minimizing the dispersion potential function, it combines the dynamic prediction model and generates multi-process parameter coordinated adjustment instructions through rolling optimization calculation. The execution module is used to execute instructions for the coordinated adjustment of multiple process parameters and to intervene in the decentralized process in real time.

2. The real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry according to claim 1, characterized in that, The signal acquisition module includes: The first laser particle size sensor is deployed at the outlet of the sand milling stage to collect slurry particle size distribution data in the sand milling stage; The second laser particle size sensor is deployed at the outlet of the ultrasonic process to collect slurry particle size distribution data during the ultrasonic process. The third laser particle size sensor is deployed at the outlet of the shearing stage to collect slurry particle size distribution data during the shearing stage. A sand mill power transmitter is connected to the sand mill driver to collect the sand mill spindle speed and torque in real time, or to directly collect the sand mill drive power. An ultrasonic power monitor, connected to an ultrasonic generator, is used to collect the output electrical power of the ultrasonic generator in real time. Among them, the slurry state signals of multiple process stages include slurry particle size distribution data of the sand milling stage, ultrasonic stage, and shearing stage; Process parameters include the spindle speed and torque of the sand mill or the drive power of the sand mill, as well as the output power of the ultrasonic generator.

3. The real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry according to claim 1, characterized in that, The methods for generating dynamic state vector sequences by the state quantization module include: Using a fixed duration T as the time window, feature extraction is performed on the slurry particle size distribution data of each process step within the current time window to obtain the feature vector within the step. The feature vector within the step includes the distribution width feature value, the main peak position feature value, and the distribution symmetry feature value. Calculate the ratio of the difference in the characteristic values ​​of the main peak positions between adjacent process steps to obtain the coupling characteristic values ​​between the steps; The characteristic value of sand mill power can be calculated based on the spindle speed and torque of the sand mill, or the driving power of the sand mill can be directly used as the characteristic value of sand mill power. Obtain the characteristic value of the output electrical power of the ultrasonic generator within the same time window; The characteristic values ​​of sand milling power and ultrasonic power are used as the energy input characteristic values; The intra-process feature vectors, inter-process coupling feature values, and energy input feature values ​​of all process steps within each time window are combined in a preset order to form a state feature vector; The state feature vectors of consecutive time windows are arranged in chronological order to form a dynamic state vector sequence.

4. The real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry according to claim 1, characterized in that, The historical scattered cases in the case knowledge base include: Raw material characteristic vector, process state matrix, process parameter matrix, performance evaluation vector; Among them, the raw material characteristic vector includes the aspect ratio characteristic value of carbon nanotubes, the specific surface area characteristic value of carbon black, and the initial solid content characteristic value of slurry. The rows of the process state matrix correspond to the time series, and the columns correspond to the eigenvalues ​​of each dimension of the state feature vector; The rows of the process parameter matrix correspond to the time series synchronized with the process state matrix, and the columns include the milling speed setting, ultrasonic power setting, and dispersant flow rate setting. The performance evaluation vector includes the slurry conductivity score and the slurry static stability score.

5. The real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry according to claim 1, characterized in that, The intelligent decision-making module retrieves a set of similar historical process fragments from the case knowledge base through the following specific steps: The sequence of dynamic state vectors generated in the current batch is recorded as the query sequence; For each historical scattered case in the case knowledge base, extract the submatrix in the process state matrix of the historical scattered case that has the same time length as the query sequence as the candidate sequence; Calculate the dynamic time-normalized distance between the query sequence and each candidate sequence; A comprehensive similarity score is calculated based on the dynamic time warp distance and the Euclidean distance between raw material characteristic vectors. Select the N historical scattered cases with the highest comprehensive similarity scores, and extract the process state subsequence and process parameter subsequence of the N historical scattered cases after the matching period to form a set of similar historical process segments.

6. The real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry according to claim 5, characterized in that, The methods for constructing dynamic prediction models in intelligent decision-making modules include: Using the state feature vector and raw material characteristic vector of each similar historical process segment at time t in the set of similar historical process segments as the model input sample, and the milling speed adjustment, ultrasonic power adjustment, and dispersant flow rate adjustment of the corresponding similar historical process segment at time t+1 as the model output label, a multi-output Gaussian process regression model is trained as a dynamic prediction model. The multi-output Gaussian process regression model uses a combination of radial basis function kernel and white noise kernel kernel.

7. The real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry according to claim 1, characterized in that, The dispersion potential function E is a function of the state characteristic vector s, the process parameter vector p, and the raw material characteristic vector r. Its construction and optimization objective are closely related to a slurry dynamic adaptation factor K, specifically: Real-time calculation of the dynamic adaptation factor K of the slurry: K = α × K0 + β × K s -γ×K c ; Where α, β, γ are the preset first set of weight coefficients; K0 is the raw material baseline term, which is calculated based on the characteristic values ​​of the initial solid content of the slurry, the aspect ratio of carbon nanotubes, and the specific surface area of ​​carbon black in the raw material characteristic vector. K s The real-time state correction term is calculated based on the difference ratio between the distribution width feature value and the main peak position feature value between adjacent process steps in the state feature vector. K c The process constraint penalty is calculated based on the ratio of the cumulative volume of dispersant added to the theoretical maximum required volume calculated based on the initial solid content of the slurry, and the ratio of the current total power to the upper limit of the total power of the equipment. The current total power is the sum of the sand mill drive power collected in real time by the signal acquisition module or the power calculated based on the speed and torque of the sand mill spindle and the output power of the ultrasonic generator. The dispersed potential function E is expressed as: E(s,p,r)=w1×f1(distribution width characteristic value deviation)+w2×f2(cumulative energy consumption)+w3×f3(length-to-diameter ratio change rate)+w4×f4(K); Among them, w1, w2, w3, and w4 are the preset second set of weight coefficients; f1 (distribution width eigenvalue deviation) is a monotonically increasing function of the deviation between the distribution width eigenvalue in the state eigenvector s and the preset target distribution width value; f2 (cumulative energy consumption) is a monotonically increasing function of the total energy consumption from the start of the distributed process to the current moment; f3 (aspect ratio change rate) is a monotonically increasing function of the ratio of the current estimated aspect ratio to the initial aspect ratio of carbon nanotubes; f4(K) is the absolute value of the difference between the dynamic adaptation factor K of the slurry and the value 1.

8. The real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry according to claim 7, characterized in that, The intelligent decision-making module generates multi-process parameter coordinated adjustment instructions through rolling optimization calculations, and adopts a model predictive control framework for implementation. Specific steps include: At each control moment, starting with the current state feature vector and process parameter settings, a set of different exploratory process parameter adjustment sequences are generated within the search space defined below: Based on the current mill speed setting, ultrasonic power setting, and dispersant flow rate setting, within the three-dimensional space formed by the mill speed setting adjustment, ultrasonic power setting adjustment, and dispersant flow rate setting adjustment, a systematic search is performed within the maximum allowable positive and negative range of each adjustment step size to generate a set of different exploratory process parameter adjustment sequences. Using a dynamic prediction model, the evolution trajectories of the system state sequence, slurry dynamic adaptation factor K sequence, and dispersion potential function value sequence are predicted after applying different experimental process parameter adjustment sequences within a future control time domain of length H. Whether the future predicted value of the slurry dynamic adaptation factor K is within the preset adaptation range [K] min ,K max Within this range, it serves as the primary constraint for screening feasible candidate adjustment sequences; Among the candidate sequences that meet the primary constraints, the optimal process parameter adjustment sequence is obtained by solving an optimization problem that minimizes the sum of the dispersed potential function values ​​at the next H time points. The first adjustment instruction in the optimal process parameter adjustment sequence is output to the execution module as the collaborative adjustment instruction at the current control moment; At the next control time step, repeat the above prediction and optimization steps.

9. The real-time control system for the dispersion process of carbon nanotube and carbon black composite slurry according to claim 8, characterized in that, Fitting range [K] min ,K max The value is [0.8, 1.2]. Furthermore, in the rolling optimization calculation, the following constraints are applied to each tentative process parameter adjustment sequence: Sequence adjustment amount constraint: In the trial process parameter adjustment sequence, the absolute values ​​of the mill speed setting value adjustment, ultrasonic power setting value adjustment, and dispersant flow rate setting value adjustment shall not exceed the preset single-step maximum adjustment amplitude, respectively; Total power constraint: The weighted sum of the mill speed setpoint and the ultrasonic power setpoint predicted based on the trial process parameter adjustment sequence shall not exceed the upper limit of the total power of the equipment; Total dispersant volume constraint: The sum of the cumulative volume of dispersant added from the start of the process to the current moment, plus the future volume of dispersant predicted based on the experimental process parameter adjustment sequence, shall not exceed 110% of the theoretical maximum required volume calculated based on the initial solid content characteristic value of the slurry; Particle size safety constraint: The maximum particle size characteristic value in the state sequence predicted by the trial process parameter adjustment sequence must not exceed the preset safety threshold.

10. A method for real-time control of the dispersion process of carbon nanotube and carbon black composite slurry, characterized in that, include: Real-time acquisition of slurry status signals and process parameters at multiple stages of the dispersion process; Based on the slurry state signal and process parameters, a dynamic state vector sequence characterizing the real-time evolution of the dispersion process is generated. The storage contains historical distributed cases containing raw material characteristic vectors, process state matrices, process parameter matrices, and performance evaluation vectors; Execute closed-loop control logic: Based on the dynamic state vector sequence and raw material characteristic vector of the current batch, retrieve a set of similar historical process segments from the case knowledge base; based on the retrieved set of similar historical process segments, construct a dynamic prediction model for predicting the effect of process parameter adjustment; with the goal of minimizing the dispersion potential function, combine the dynamic prediction model and generate multi-process parameter coordinated adjustment instructions through rolling optimization calculation; It executes commands to coordinate the adjustment of multiple process parameters and intervenes in the decentralized process in real time.

Images