Intelligent control method and system for carburant screening

Abstract

The invention belongs to the field of intelligent control, and particularly relates to a carburant screening intelligent control method and system.The method comprises the steps that firstly, screening efficiency and a dimensionless particle size distribution penalty term are weighted and quantized into a unified comprehensive performance index; then, probability modeling is conducted on the nonlinear relation between the index and control input through a deep Gaussian process model, implicit constraints such as material blocking and overload are recognized through a Gaussian process classifier, and a dynamic feasible region is defined; in the dynamic feasible region, efficient Bayesian optimization is carried out by constructing an asymmetric information gain acquisition function, and the optimal working condition is searched; fusing a particle filtering framework with an acquisition function to realize dynamic tracking of a time-varying target; finally, the system operates according to recommended input and updates all models online by using new data to form a self-adaptive closed loop. According to the method, the optimization problems of dynamic, nonlinear and implicit constraint in the screening process are solved, and the intelligent level, safety and economic benefits of control are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent control, and particularly relates to an intelligent control method and system for carbon additive screening. BACKGROUND

[0002] Carbon additive is a kind of auxiliary material commonly used in the metallurgical industry, which is used to adjust the carbon content in molten steel or molten iron. The uniformity of particle size distribution and the screening efficiency of carbon additive directly affect the carbonation effect, melting speed, yield and production cost in the subsequent smelting process. As a key link in the production process of carbon additive, the performance of screening is usually measured by two indicators of screening efficiency and product particle size distribution, which are affected by the complex coupling of various process parameters such as vibration frequency, screen surface inclination, feeding rate, etc.

[0003] In actual production process, the physical properties of carbon additive raw materials (such as moisture content, particle shape, hardness, etc.) will fluctuate with batches, and the running state of screening equipment (such as screen mesh wear, tension change) will also evolve over time. These uncertain factors often lead to the drift of the optimal working condition of the screening process. Existing industrial practice mostly relies on the experience of operators for manual adjustment, which not only responds slowly, but also is difficult to maintain in the optimal working condition for a long time, often leading to decreased screening efficiency, unstable product particle size and increased energy consumption.

[0004] In view of the above problems, researchers try to use traditional optimization methods such as mechanism modeling and response surface analysis. However, these methods are usually based on static assumptions, and a large number of offline experiments are needed to establish the model, and once the model is determined, it is difficult to adapt to the changes of dynamic working conditions. At the same time, traditional methods do not consider the process uncertainty, especially the implicit constraints such as screen mesh blockage and equipment overload, which are difficult to establish explicit mechanism model, thereby affecting the safety and stability of the screening process.

[0005] Therefore, the existing technology still lacks an intelligent control method that can model the screening performance in real time, dynamically track the optimal working condition and update the model online under the premise of ensuring the safety of the equipment, in order to adapt to the complexity, time-varying and uncertainty of the carbon additive screening process. SUMMARY

[0006] The purpose of the present application is to provide an intelligent control method for carbon additive screening, which solves the problems of complexity, time-varying and uncertainty of the carbon additive screening process, comprising the following steps: Obtain vibration frequency, screen surface inclination and feeding rate as control input, and quantify screening efficiency and product particle size distribution as a unified comprehensive performance index; a deep Gaussian process is used to probabilistically model the relationship between the comprehensive performance index and the control input to obtain a deep Gaussian process model, and the deep Gaussian process model uses a deep kernel function to identify nonlinear characteristics; The Gaussian process classifier is constructed to probabilistically model the implicit constraints including screen clogging and / or equipment overload, and a dynamic feasible region is determined based on the prediction result of the Gaussian process classifier; within the dynamic feasible region, a collection function is constructed based on the posterior predictive distribution output by the deep Gaussian process model using asymmetric information gain; A state space model describing dynamic evolution of the optimal working condition is constructed, and a particle filtering framework is adopted to propagate the posterior distribution of the optimal working condition at the previous time through the state space model to obtain the prior distribution at the current time, and the collection function is used to guide importance sampling and resampling to generate the next set of recommended control input; According to the recommended control input, the screening system is adjusted to run, and the deep Gaussian process model and the Gaussian process classifier are updated online using newly collected performance data and constraint states.

[0007] Preferably, the screening efficiency and the product particle size distribution are quantified into a unified comprehensive performance index, which includes: obtaining the screening efficiency and the product particle size ; The target interval of the product particle size is set as , and d max >d min ; A dimensionless product particle size distribution penalty term is defined as: ; The comprehensive performance index is calculated, wherein and are preset positive weight coefficients, is a screening efficiency weight coefficient, is a product particle size weight coefficient.

[0008] By directly obtaining the two most critical physical quantities, the screening efficiency and the product particle size , raw and accurate data inputs are provided for performance evaluation; by setting the target interval of the product particle size as This study clarifies the "acceptable" range for product quality, providing a benchmark for subsequent calculations of particle size distribution deviations, thus making quality control objectives more specific and quantifiable. By defining a dimensionless product particle size distribution penalty term P(d), a penalty value is only generated when the particle size d exceeds the target range, and the magnitude of the penalty value is proportional to the degree of deviation. Normalization of the denominator makes it dimensionless, facilitating combination with other indicators of different units, such as screening efficiency, greatly simplifying the complexity of multi-objective trade-offs. Finally, a unified performance index J is constructed through linear weighted summation (efficiency as a positive incentive, penalty term as a negative incentive). The introduction of weighting coefficients w1 and w2 allows for flexible adjustment of the relative importance of the two objectives based on actual production needs (e.g., prioritizing quality or quantity), providing an effective tool for achieving customized, economically efficient optimization strategies.

[0009] Preferably, the step of using a deep Gaussian process to probabilistically model the relationship between the comprehensive performance index and the control input to obtain a deep Gaussian process model includes: The control input is used as the input to the first-level Gaussian process to generate latent variables; The latent variables are used as input to the second-level Gaussian process to output the posterior prediction distribution of the comprehensive performance index.

[0010] The deep Gaussian process model with the aforementioned two-layer structure can simulate the intrinsic logical chain from "control action" to "intermediate physical state" and then to "final performance." The first layer maps the control input to a set of abstract latent variables, and the second layer links these latent variables to the final performance index. This structure not only endows the model with powerful nonlinear fitting capabilities to improve prediction accuracy, but its complete probability distribution (rather than a single predicted value) provides crucial uncertainty information for subsequent Bayesian optimization, risk assessment, and exploratory decision-making, thereby significantly enhancing the robustness and intelligence of the control strategy.

[0011] Preferably, the prediction results based on the Gaussian process classifier determine the dynamic feasible region, including: For any candidate point in the control input space, the probability that the candidate point will not experience screen clogging or equipment overload is predicted using the Gaussian process classifier. ; All satisfied The set of control input points that are greater than or equal to the safety probability threshold is defined as the dynamic feasible region at the current moment.

[0012] The safety probability P(safe|x) of any candidate control point is predicted by using a Gaussian process classifier, and all points with safety probability greater than or equal to a preset threshold are combined to form a dynamic feasible region. This design realizes active risk avoidance from "post-alarm" to "pre-prediction", and more importantly, the safety boundary can be updated online according to newly collected data, achieving intelligent dynamic balance between safety and performance pursuit.

[0013] Preferably, the particle filtering framework is used to propagate the posterior distribution of the optimal operating condition at the previous time through the state space model to obtain the prior distribution at the current time, and the acquisition function is used to guide importance sampling and resampling to generate the next set of recommended control input quantities, including: a) Propagating a set of particles representing the posterior distribution of the optimal operating condition at the previous time through the state space model to obtain a prior particle set at the current time; b) Calculating the acquisition function value of each particle in the prior particle set corresponding to the control input quantity, and taking the acquisition function value as the importance weight of the particle; c) Resampling from the prior particle set according to the importance weight calculated in step b) to form a posterior particle set at the current time; d) Selecting the control input quantity corresponding to the particle with the highest comprehensive performance index prediction mean from the posterior particle set as the next set of recommended control input quantities.

[0014] The particle filtering framework uses the acquisition function value to guide the weight distribution and resampling of particles, so that the particle cloud can intelligently focus on the current most valuable exploration area, rather than blindly tracking. Finally, the optimal decision is made from the "elite" particle set selected by survival of the fittest, ensuring that the system can quickly and accurately track the time-varying optimal target, and each recommended operation takes into account the wisdom of exploration and the efficiency of production. That is, by deeply integrating the "efficient exploration" capability of Bayesian optimization with the "dynamic tracking" capability of particle filtering, the problem of dynamic drift of the optimal operating condition point caused by changes in raw material characteristics or equipment state is solved.

[0015] Preferably, the deep Gaussian process model and the Gaussian process classifier are updated online using newly collected performance data and constraint states, including: The newly collected control input quantity, corresponding comprehensive performance index, and constraint state are supplemented to the historical database; A gradient optimization-based method is used to update the hyperparameters of the deep Gaussian process model and the Gaussian process classifier.

[0016] By supplementing the valuable data generated by each online operation to the database and updating the model hyperparameters online using gradient optimization algorithms, it is ensured that the two core models (deep Gaussian process model and Gaussian process classifier) inside the system can always keep up with the slow changes of the process characteristics. This online updating mechanism guarantees the long-term effectiveness and advancement of the control strategy, enabling the model to continuously approximate the true production process characteristics, thereby maximizing the value of data and continuously optimizing the performance of the system, overcoming the limitations of offline models that gradually lose effectiveness due to "model-reality mismatch" in long-term operation, and endowing the entire control system with the ability of continuous learning and adaptation.

[0017] In addition, the present application also provides an intelligent control system for carbon additive screening, comprising the following units: A control input quantity acquisition unit acquires the vibration frequency, screen surface inclination angle and feeding rate as control input quantities, quantifies the screening efficiency and product particle size distribution as a unified comprehensive performance index; a deep Gaussian process is used to probabilistically model the relationship between the comprehensive performance index and the control input quantities to obtain a deep Gaussian process model, and the deep Gaussian process model uses a deep kernel function to identify nonlinear characteristics; A probabilistic modeling unit constructs a Gaussian process classifier to probabilistically model implicit constraints including screen clogging and / or equipment overload, and determines a dynamic feasible region based on the prediction results of the Gaussian process classifier; within the dynamic feasible region, an acquisition function is constructed using asymmetric information gain based on the posterior predictive distribution output by the deep Gaussian process model; A model construction unit constructs a state space model describing the dynamic evolution of the optimal working condition, uses a particle filtering framework to propagate the posterior distribution of the optimal working condition at the previous time through the state space model to obtain the prior distribution at the current time, and uses the acquisition function to guide importance sampling and resampling to generate the next set of recommended control input quantities; A control and updating unit adjusts the operation of the screening system according to the recommended control input quantities and updates the deep Gaussian process model and the Gaussian process classifier online using newly collected performance data and constraint states.

[0018] Preferably, the quantification of the screening efficiency and the product particle size distribution as a unified comprehensive performance index comprises: acquiring the screening efficiency and the product particle size ; setting the target interval of the product particle size as , and d max >d min ; defining a dimensionless product particle size distribution penalty term : ; Calculate comprehensive performance indicators ,in and These are preset positive weighting coefficients. This is the weighting coefficient for screening efficiency. This is the product granularity weighting coefficient.

[0019] Preferably, the step of using a deep Gaussian process to probabilistically model the relationship between the comprehensive performance index and the control input to obtain a deep Gaussian process model includes: The control input is used as the input to the first-level Gaussian process to generate latent variables; The latent variables are used as input to the second-level Gaussian process to output the posterior prediction distribution of the comprehensive performance index.

[0020] Preferably, the prediction results based on the Gaussian process classifier determine the dynamic feasible region, including: For any candidate point in the control input space, the probability that the candidate point will not experience screen clogging or equipment overload is predicted using the Gaussian process classifier. ; All satisfied The set of control input points that are greater than or equal to the safety probability threshold is defined as the dynamic feasible region at the current moment.

[0021] Preferably, the particle filter framework is used to propagate the posterior distribution of the optimal operating condition at the previous time step through the state-space model to obtain the prior distribution at the current time step. The acquisition function is then used to guide importance sampling and resampling to generate the next set of recommended control inputs, including: a) The set of particles representing the posterior distribution of the optimal working condition in the previous moment is propagated through the state space model to obtain the prior particle set at the current moment. b) Calculate the acquisition function value of the control input for each particle in the prior particle set, and use the acquisition function value as the importance weight of the particle; c) Based on the importance weights calculated in step b), resample from the prior particle set to form the posterior particle set at the current moment; d) From the posterior particle set, select the control input corresponding to the particle with the highest predicted mean of comprehensive performance index as the next set of recommended control inputs.

[0022] Preferably, the step of updating the deep Gaussian process model and the Gaussian process classifier online using newly acquired performance data and constraint states includes: The newly collected control inputs, corresponding comprehensive performance indicators, and constraint states are added to the historical database. The hyperparameters of the deep Gaussian process model and the Gaussian process classifier are updated by using a gradient optimization based method.

[0023] Compared with the prior art, the present application has at least the following advantages: By constructing a deep Gaussian process model, the present application can capture the complex nonlinear relationship between the control input and the comprehensive performance index in the carbon additive screening process, quantify the uncertainty of the model, and improve the reliability of prediction and control. Moreover, the present application uses a Gaussian process classifier to probabilistically identify implicit constraints such as screen clogging and equipment overload, which are difficult to model, and defines a dynamic feasible region for safe operation, thereby reducing potential production risks and ensuring the stability and safety of equipment operation. On this basis, by constructing an asymmetric information gain acquisition function, efficient search and convergence of the optimal working condition are achieved. The present application can effectively track the changes in the optimal working condition point caused by fluctuations in raw material characteristics or evolution of equipment state, so that the screening process can continuously approach and maintain at the best operating point, thereby improving the screening efficiency and uniformity of product particle size distribution without relying on human experience. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 A method flowchart of steps S1-S4 in the carbon additive screening intelligent control method of the embodiment of the present application; Figure 2 A deep Gaussian process model structure diagram in the embodiment of the present application; Figure 3 A dynamic feasible region diagram at a certain moment in the embodiment of the present application; Figure 4 An asymmetric information gain acquisition function diagram in the embodiment of the present application; Figure 5 A particle filter tracking optimal working condition diagram in the embodiment of the present application; Figure 6 A structure block diagram of the carbon additive screening intelligent control system of the embodiment of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application.

[0026] The control method of the application is implemented on a typical carbon additive screening production line, which mainly includes a parameter-adjustable vibrating screening device, a frequency converter for controlling the vibration frequency, a servo motor and an angle sensor for controlling the screen surface inclination angle, a material conveyor for controlling the feeding rate, an online particle size detector for measuring the product particle size distribution, a weighing sensor for calculating the screening efficiency, a motor current sensor and a material blocking sensor for monitoring the constraint state, and an industrial computer; wherein the industrial computer communicates with the above-mentioned actuators and sensing systems through a bus, and the feeding speed of the material conveyor is controlled by an independent frequency converter.

[0027] Referring to Figure 1 , an intelligent control method for carbon additive screening, specifically comprising the following steps: S1, obtaining the vibration frequency, the screen surface inclination angle and the feeding rate as control input quantities, and quantifying the screening efficiency and the product particle size distribution into a unified comprehensive performance index; a deep Gaussian process is used to probabilistically model the relationship between the comprehensive performance index and the control input quantities to obtain a deep Gaussian process model, and the deep Gaussian process model uses a deep kernel function to identify nonlinear characteristics; The vibration frequency is set and read by the frequency converter, the screen surface inclination angle is controlled and obtained by the angle sensor and the servo motor, and the feeding rate is controlled by adjusting the speed or amplitude of the belt conveyor or the electromagnetic vibrating feeder; at the system output end, the product particle size distribution is measured by image analysis or an online particle size detector, the particle proportion within the preset target particle size range or the Wasserstein distance from the target distribution is calculated as the particle size distribution index, and the screening efficiency is calculated by weighing the mass of the material under the screen through the weighing sensor; a weighted summation method is used, for example, the screening efficiency and the negative particle size distribution deviation are multiplied by their respective economic weight coefficients and then added to form a comprehensive performance index in the form of a scalar .

[0028] A deep Gaussian process model with a multi-layer structure is constructed, wherein the first layer maps the three-dimensional input vector composed of the vibration frequency, the screen surface inclination angle and the feeding rate to a high-dimensional hidden space, the input of each subsequent Gaussian process is the output of the previous layer, and the output of the last Gaussian process is the predicted mean and variance of the comprehensive performance index ; the structure diagram of the deep Gaussian process model is shown in Figure 2 ; the hierarchical structure is equivalent to a deep kernel function automatically learned from data, which can obtain the highly nonlinear and non-stationary relationship between the input and output.

[0029] S2, constructing a Gaussian process classifier to probabilistically model the implicit constraints including screen clogging and / or equipment overload, determining a dynamic feasible region based on the prediction result of the Gaussian process classifier; within the dynamic feasible region, constructing an acquisition function using asymmetric information gain based on the posterior predictive distribution output by the deep Gaussian process model; During operation, whether screen clogging or equipment overload occurs is determined through a clogging sensor, motor current monitoring or acoustic signal analysis, and is recorded as class 1, and normal is recorded as class 0; the control input and the corresponding class label are combined to form a training data set, which is used to train the Gaussian process classifier; the Gaussian process classifier can predict the probability of occurrence of constraint violation events for any new control input; by setting a safety probability threshold, for example, 95%, all control input combinations that make the predicted constraint violation probability lower than the threshold form the dynamic feasible region at the current time, as shown in Figure 3 .

[0030] Within the dynamic feasible region defined by the Gaussian process classifier, a series of candidate control input points are evaluated; for each candidate point, the posterior mean and variance of the comprehensive performance index of the point are predicted using the deep Gaussian process model; the acquisition function is preferably in the form of asymmetric information gain, as shown in Figure 4 , the value function of the acquisition function not only considers the amount of uncertainty reduction about the position of the optimal value, i.e., entropy reduction, brought by sampling the candidate point, but also additionally gives more weight to the regions whose posterior mean is higher than the current optimal value, i.e., giving higher weight to exploring high potential regions, and giving lower weight to exploring low performance regions or purely reducing uncertainty; the candidate point that maximizes the asymmetric information gain is selected as the next test working condition.

[0031] In a more specific embodiment, within the dynamic feasible region, an acquisition function is constructed using asymmetric information gain based on the posterior predictive distribution output by the deep Gaussian process model, specifically: a safety probability weighted expected improvement function is constructed. This function first uses the deep Gaussian process model to calculate the expected improvement amount that the system performance at a certain candidate control input point can obtain compared to the current optimal value; then, the safety probability that the point will not clog or overload is predicted using the Gaussian process classifier or a new model. The value of the acquisition function is the product of the two, i.e., safety probability x expected improvement amount. Only those points that are predicted to be both safe and have high potential for improvement can get high scores, thereby guiding the system to find the optimal working condition most efficiently while avoiding risks.

[0032] S3, constructing a state space model describing the dynamic evolution of the optimal working condition, using a particle filtering framework to propagate the posterior distribution of the optimal working condition at the last time through the state space model to obtain the prior distribution at the current time, using the acquisition function to guide importance sampling and resampling to generate the next set of recommended control input; Suppose that the optimal working condition point evolves over time according to a Gaussian random walk model, that is, the optimal working condition at the current time is based on the optimal working condition at the last time plus a small random disturbance; using particle filtering, a set of particles representing the probability distribution of the optimal working condition at the last time is propagated through the random walk model to obtain a prior particle set of the optimal working condition at the current time; using the asymmetric information gain acquisition function constructed in the last step as the importance density function, the propagated particles are weighted, that is, the particles with larger acquisition function values are given higher weights; resampling according to the weights to generate a new particle set, from which the particle with the maximum acquisition function value is selected as the recommended control input at the current time.

[0033] In one embodiment, the use of the acquisition function to guide importance sampling and resampling to generate the next set of recommended control input is as follows: obtain a prior particle set representing the possible position of the optimal working condition, where each particle is a specific control input set, such as vibration frequency, screen inclination angle, and feed rate. Calculate the acquisition function value for each particle, which is the product of the safety probability and the expected improvement, and use this value directly as the importance weight of the particle. Particles in the more promising (safe and high potential for performance improvement) region will have higher weights. Resampling according to these weights: high-weight particles are copied multiple times, while low-weight particles are eliminated. To give a definite operation instruction, the particle with the highest comprehensive performance index mean predicted by the deep Gaussian process model is selected from the refined posterior particle set, and the corresponding control input combination is used as the next set of optimal working conditions recommended for the screening system to execute.

[0034] S4, adjusting the operation of the screening system according to the recommended control input, and updating the deep Gaussian process model and the Gaussian process classifier online using newly collected performance data and constraint states.

[0035] The particle filtering recommended vibration frequency, screen inclination angle, and feed rate values are sent to the underlying PID controller or device driver unit for execution; after stable operation for a period, the corresponding comprehensive performance index measured values and the labels indicating whether the constraints are triggered are collected; this new set of data points, i.e., control input, performance index, and constraint label, are appended to the respective historical databases, and the hyperparameters of the deep Gaussian process model and the Gaussian process classifier are retrained or updated online, so that both models can absorb the latest information in real time and adjust their predictions accordingly to reflect changes in the process.

[0036] In one embodiment, the quantification of the screening efficiency and the product size distribution into a unified comprehensive performance index comprises: obtaining the screening efficiency and the product size ; setting a target interval of the product size as , and d max > d min ; defining a dimensionless product size distribution penalty term : ; calculating the comprehensive performance index , wherein and are preset positive weight coefficients, is a screening efficiency weight coefficient, is a product size weight coefficient.

[0037] In screening, the screening efficiency is the ratio of the undersize mass to the feed mass, and the value range is [0%, 100%]. The product size is in millimeters, d min and d max are the lower limit and the upper limit of the product size, respectively. Assuming that the target product size interval is set as 0.1 millimeter to 0.5 millimeter. If the screening efficiency is 92% at a certain time, and the product size is 0.6 millimeter, the size exceeds the target upper limit at this time. According to the formula, the penalty term is 0.25. Assuming that the weight is set as 1, is set as 2 to emphasize the importance of size control, then the comprehensive performance index is calculated as 0.42.

[0038] In another working condition, if the screening efficiency is 90%, and the product size is 0.4 millimeter, the size value falls within the ideal interval, and the penalty term is calculated as 0. At this time, the comprehensive performance index is 0.90. By comparing the values of of the two working conditions, 0.90 is much higher than 0.42, indicating that even if the screening efficiency is slightly low, the latter working condition performs better because the product size meets the requirements. This enables the optimization algorithm to clearly distinguish the pros and cons of different control strategies and find the balance point that optimizes the overall production target.

[0039] In one embodiment, the deep Gaussian process is employed to probabilistically model the relationship between the comprehensive performance indicator and the control input, resulting in a deep Gaussian process model, comprising: inputting the control input as an input of a first layer Gaussian process, to generate latent variables; inputting the latent variables as an input of a second layer Gaussian process, to output a posterior predictive distribution of the comprehensive performance indicator.

[0040] A complex production process is understood through a hierarchical structure. For example, the vibration frequency and amplitude of a vibrating screen are inputted into a first layer Gaussian process as control inputs. The first layer does not directly predict the final performance, but rather transforms these physical inputs into a set of abstract latent variables, which can correspond to some intermediate process states that are not directly measurable, such as the layering effect or throwing strength of the material on the screen surface.

[0041] The latent variables representing the intermediate states are inputted into a second layer Gaussian process, which is responsible for modeling the mapping relationship between the intermediate states and the final comprehensive performance indicator . Its output is not a deterministic value, but a probability distribution, such as a predicted mean of 0.85 and a variance of 0.05. The variance represents the degree of uncertainty of the current prediction of the model.

[0042] In one embodiment, the prediction result of the Gaussian process classifier is used to determine the dynamic feasible region, comprising: for any candidate point in the control input space, the Gaussian process classifier is used to predict the probability of the candidate point not causing screen clogging or equipment overload ; a set of control input points that satisfy greater than or equal to a safety probability threshold is defined as the dynamic feasible region at the current time.

[0043] To ensure the stability and safety of the production process, a dynamic definition mechanism of a safety boundary is introduced. A Gaussian process classifier trained based on historical data is used to evaluate the risk of any set of potential control inputs, such as the combination of a feed rate of fifty tons per hour and a screen surface inclination of fifteen degrees, causing equipment failure. The Gaussian process classifier will output a safety probability, such as a probability of 98% for the above combination to predict safe operation.

[0044] A pre-set safety probability threshold, such as 95%, is established. This threshold represents the acceptable risk level for production management, and is typically set between 90% and 99%. All control input points with a safety probability of at least 95% after evaluation by the Gaussian process classifier collectively constitute a safe operating region, i.e., the dynamic feasible region. If another set of candidate control inputs, such as a feed rate of 80 tons per hour and a screen inclination angle of 12 degrees, has a predicted safety probability of only 60%, then this input point will be excluded from the dynamic feasible region. As new data is continuously collected, such as when the motor current approaches the alarm value during an operation, the Gaussian process classifier will be updated online, and the boundary of the dynamic feasible region will be adjusted accordingly, ensuring that the optimization algorithm always seeks the optimal operating condition within a dynamically updated safety range.

[0045] The acquisition function quantifies the expected reduction in posterior distribution entropy at the optimal operating position due to sampling at candidate points. Asymmetric weights are introduced to distinguish the information value of performance-enhancing and degrading regions. The acquisition function is the core of the Bayesian optimization algorithm; its main function is to guide the algorithm in each iteration to select the next most valuable sample point, i.e., to control the input for evaluation, i.e., to run on the actual system and observe the results. The information gain acquisition function selects a point x, and after evaluation, obtains y, which allows for a more accurate assessment of the optimal solution position. posterior probability distribution The entropy decreases the most. Asymmetric information gain assumes that different types of information have different values. The process of constructing the asymmetric information gain acquisition function is as follows: Define the value function The value function measures the value or utility derived from observing a certain outcome y, i.e., the value of the overall performance index J. It is asymmetric. For example, when the y-value is higher than the current optimal value... At that time, its value Rapid growth; while when the y-value is lower than At this point, the value is very small or zero. In constrained problems, if a point x leads to failure such as blockage, the resulting value y is negative or extremely low. This invention does not limit the specific form of the value function.

[0046] Calculating the expected information value: Acquisition function The value of is the value V(y) for all possible observations y, given its predicted probability. The expectations below. It is given by the deep Gaussian process model, representing the posterior predicted distribution of y observed at point x. Mathematically, it can be expressed as:

[0047] Where x is the candidate control input vector (e.g., vibration frequency, screen inclination angle, feed rate, etc.); D is the current historical dataset; For the posterior predictive distribution of the DGP at point x, random variable y represents the possible observation of the performance indicator J. The integral is the expected value if an experiment is conducted at point x.

[0048] More complex asymmetric information gain can modify the way of calculating information entropy. For example, in the calculation of KL divergence, different weights are given to different result regions, such as "performance greatly improved region", "performance slightly improved region", "performance decreased region", and "system failure region".

[0049] The acquisition function is used to guide the exploration direction of the next step to find the optimal working condition most efficiently. Not all unknown regions are worth exploring equally. Suppose the current known optimal performance indicator J is 0.8. Now evaluate two candidate sampling points A and B. Candidate point A is located in a region where the model predicts that the performance can reach 0.9 but has high uncertainty, while candidate point B is located in a region where the model predicts that the performance is about 0.6 and also has high uncertainty.

[0050] The acquisition function may give similar scores to both points because both can significantly reduce uncertainty. However, with asymmetric information gain, more reward is given to exploring regions that may bring performance improvement. For candidate point A, because it is expected to find a better working condition than the current optimal value 0.8, the information value is amplified by a larger weight coefficient. For candidate point B, although information can be obtained, this information is only about a known suboptimal region, and the value is weakened by a smaller weight coefficient. Therefore, a higher acquisition function score will be calculated for point A, and the next experiment will be preferentially conducted at point A, accelerating the convergence speed to the higher performance region and avoiding wasting valuable experimental opportunities in inefficient regions.

[0051] In one embodiment, the particle filtering framework is used to propagate the posterior distribution of the optimal working condition at the last time through the state space model to obtain the prior distribution at the current time, use the acquisition function to guide importance sampling and resampling, and generate the next set of recommended control input quantities, including: a) Propagate a set of particles representing the posterior distribution of the optimal working condition at the last time through the state space model to obtain a prior particle set at the current time; b) Calculate the acquisition function value of the control input quantity corresponding to each particle in the prior particle set, and use the acquisition function value as the importance weight of the particle; c) Resample from the prior particle set according to the importance weight calculated in step b) to form a posterior particle set at the current time; ​d) From the posterior particle set, select the control input corresponding to the particle with the highest predicted mean of the comprehensive performance index as the next set of recommended control input.

[0052] A set of particles, such as one thousand, are used to track the optimal operating condition over time. In the first propagation stage, the one thousand particles representing the optimal operating condition at the previous time step are moved according to a state space model describing the drift of the system to predict the possible new location of the optimal operating condition at the current time step, forming a prior particle cloud as shown in Figure 5

[0053] In the second weight calculation and resampling stage, each particle in the prior particle cloud is evaluated using the aforementioned asymmetric information gain acquisition function. Particles falling in areas of high information value are assigned high weights, and vice versa. Resampling is performed according to the weights, with high-weight particles having a higher probability of being copied and retained, and low-weight particles being more likely to be eliminated. This causes the distribution of the particle cloud to gradually concentrate in the current region of highest exploration value, forming a posterior particle set.

[0054] In the decision stage, for each particle in the posterior particle set, the mean of the comprehensive performance index predicted by the deep Gaussian process model is calculated. For example, among the one thousand posterior particles after resampling, the particle numbered 372, with control input amplitude of 5 mm and frequency of 900 rpm, has the highest predicted performance mean of 0.92. This set of control input is then selected as the recommended operating instruction for the next period and issued to the control system for execution. In an alternative embodiment, the point that maximizes the acquisition function is found in the entire dynamic feasible region, which is the combination of all possible control input within the safety constraints, and is selected as the next recommended control input.

[0055] In one embodiment, the updating of the deep Gaussian process model and the Gaussian process classifier using newly acquired performance data and constraint states online comprises: Supplementing the newly acquired control input, corresponding comprehensive performance index, and constraint state to the historical database; Using a gradient-based optimization method to update the hyperparameters of the deep Gaussian process model and the Gaussian process classifier.

[0056] After the recommended control instruction of the previous round is executed, such as running for ten minutes at an amplitude of 5 mm and a frequency of 900 rpm, it will collect the actual running data of this period. These data constitute a complete data point, including the input control, the actually measured comprehensive performance index ​For example, 0.91, and recorded constraint states, such as normal motor current, no material jamming, i.e. safe state.

[0057] The new data points are immediately added to the system's history database. The model is retrained or fine-tuned with the updated dataset. Through gradient-based optimization algorithms, the internal hyperparameters of the deep Gaussian process model and the Gaussian process classifier, such as the length scale of the kernel function, can be adjusted. For example, new data may reveal that the sensitivity of performance to amplitude is higher than the model originally thought at the current humidity, and the update of the hyperparameters will reflect this new information. This enables the model to continuously approximate the true production process characteristics, adapt to changes in raw materials, equipment wear and other slow changes, and ensure the long-term effectiveness and accuracy of optimization recommendations.

[0058] Reference Figure 6 The present application also provides an intelligent control system for carbon additive screening, which is used to execute any of the above-mentioned embodiments of the intelligent control method for carbon additive screening; the system is specifically an industrial computer configured with a special software or an integrated control device; in the embodiment, the control system comprises the following units: A control input quantity acquisition unit acquires the vibration frequency, the screen surface inclination angle and the feeding rate as control input quantities, quantizes the screening efficiency and the product particle size distribution into a unified comprehensive performance index; a deep Gaussian process is used to probabilistically model the relationship between the comprehensive performance index and the control input quantities to obtain a deep Gaussian process model, and the deep Gaussian process model uses a deep kernel function to identify nonlinear characteristics; A probabilistic modeling unit constructs a Gaussian process classifier to probabilistically model implicit constraints including screen mesh jamming and / or equipment overload, and determines a dynamic feasible region based on the prediction results of the Gaussian process classifier; within the dynamic feasible region, an acquisition function is constructed using asymmetric information gain based on the posterior predictive distribution output by the deep Gaussian process model; A model construction unit constructs a state space model describing the dynamic evolution of the optimal working condition, uses a particle filtering framework to propagate the posterior distribution of the optimal working condition at the previous time through the state space model to obtain the prior distribution at the current time, and uses the acquisition function to guide importance sampling and resampling to generate the next set of recommended control input quantities; A control and update unit adjusts the operation of the screening system according to the recommended control input quantities, and updates the deep Gaussian process model and the Gaussian process classifier online using newly collected performance data and constraint states.

[0059] The above describes in detail the carbon additive screening intelligent control method and system provided by the present application, and the principles and implementation manners of the present application are described by using specific examples. The above description of the embodiments is only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges will be changed. In conclusion, the content of the present description should not be understood as a limitation of the present application.

Claims

1. A recarburizer screening intelligent control method, characterized in that, The method comprises the following steps: obtaining vibration frequency, screen surface inclination and feeding rate as control input, quantifying screening efficiency and product particle size distribution as a unified comprehensive performance index; a deep Gaussian process is used to probabilistically model the relationship between the comprehensive performance index and the control input to obtain a deep Gaussian process model, and the deep Gaussian process model uses a deep kernel function to identify nonlinear characteristics; a Gaussian process classifier is constructed to probabilistically model implicit constraints including screen mesh clogging and / or equipment overload, and a dynamic feasible region is determined based on the prediction result of the Gaussian process classifier; within the dynamic feasible region, an acquisition function is constructed using asymmetric information gain based on the posterior predictive distribution output by the deep Gaussian process model; a state space model describing the dynamic evolution of the optimal working condition is constructed, and a particle filtering framework is used to propagate the posterior distribution of the optimal working condition at the previous time through the state space model to obtain the prior distribution at the current time, and the acquisition function is used to guide importance sampling and resampling to generate the next set of recommended control input; the screening system is adjusted according to the recommended control input, and the deep Gaussian process model and the Gaussian process classifier are updated online using newly collected performance data and constraint states.

2. The intelligent control method of recarburizer screening according to claim 1, characterized in that, The comprehensive performance index is quantified by the screening efficiency and the product particle size distribution, which comprises: Acquisition of screening efficiency and product particle size ; The target interval of the product granularity is set as , and d max > d min ; Defining a dimensionless product particle size distribution penalty term : ； Computing an overall performance indicator wherein and is a predetermined positive weight coefficient, is a weight coefficient for the screening efficiency, is a weight coefficient for the product grain size.

3. The intelligent control method of recarburizer screening according to claim 1, characterized in that, The deep Gaussian process is used to probabilistically model the relationship between the comprehensive performance index and the control input to obtain a deep Gaussian process model, which comprises: The control input is used as the input of the first layer Gaussian process to generate hidden variables; The hidden variables are used as the input of the second layer Gaussian process to output the posterior predictive distribution of the comprehensive performance index.

4. The intelligent control method of recarburizer screening according to claim 1, characterized in that, The dynamic feasible region is determined based on the prediction result of the Gaussian process classifier, which comprises: for any candidate point in the control input space, predicting a probability that the candidate point does not result in screen blinding or equipment overload using the Gaussian process classifier ; All the requirements The set of control input points greater than or equal to the safety probability threshold is defined as the dynamic feasible region at the current time.

5. The intelligent control method of recarburizer screening according to claim 1, characterized in that, The particle filtering framework is used to propagate the posterior distribution of the optimal working condition at the previous time through the state space model to obtain the prior distribution at the current time, and the acquisition function is used to guide importance sampling and resampling to generate the next set of recommended control input, which comprises: a) A set of particles representing the posterior distribution of the optimal working condition at the previous time are propagated through the state space model to obtain a prior particle set at the current time; b) The acquisition function value of the control input corresponding to each particle in the prior particle set is calculated, and the acquisition function value is used as the importance weight of the particle; c) According to the importance weight calculated in step b), the posterior particle set at the current time is formed by resampling from the prior particle set; d) The control input corresponding to the particle with the highest comprehensive performance index prediction mean is selected from the posterior particle set as the next set of recommended control input.

6. The intelligent control method of recarburizer screening according to claim 1, characterized in that, The deep Gaussian process model and the Gaussian process classifier are updated online using newly collected performance data and constraint states, which comprises: The newly collected control input, corresponding comprehensive performance index and constraint state are supplemented into the historical database; The hyperparameters of the deep Gaussian process model and the Gaussian process classifier are updated using a gradient optimization-based method.

7. A recarburizer screening intelligent control system, characterized in that, The method comprises the following units: The control input quantity acquisition unit acquires the vibration frequency, the screen surface inclination angle and the feeding rate as the control input quantity, quantifies the screening efficiency and the product particle size distribution into a unified comprehensive performance index; a deep Gaussian process is used to probabilistically model the relationship between the comprehensive performance index and the control input quantity to obtain a deep Gaussian process model, and the deep Gaussian process model uses a deep kernel function to identify nonlinear characteristics; The probability modeling unit constructs a Gaussian process classifier to probabilistically model implicit constraints including screen clogging and / or equipment overload, determines a dynamic feasible region based on the prediction result of the Gaussian process classifier; within the dynamic feasible region, an acquisition function is constructed using the asymmetric information gain based on the posterior predictive distribution output by the deep Gaussian process model; The model construction unit constructs a state space model describing the dynamic evolution of the optimal working condition, uses a particle filtering framework to propagate the posterior distribution of the optimal working condition at the previous time through the state space model to obtain the prior distribution at the current time, uses the acquisition function to guide importance sampling and resampling, and generates the next set of recommended control input quantities; The control and update unit adjusts the operation of the screening system according to the recommended control input quantities, and updates the deep Gaussian process model and the Gaussian process classifier online using newly collected performance data and constraint states.

8. The recarburizer screening intelligent control system of claim 7, wherein, The unified comprehensive performance index includes: Acquisition of screening efficiency and product particle size ; The target interval of the product granularity is set as , and d max > d min ; Defining a dimensionless product particle size distribution penalty term : ； Computing an overall performance indicator wherein and is a predetermined positive weight coefficient, is a weight coefficient for the screening efficiency, is a weight coefficient for the product size.

9. The recarburizer screening intelligent control system of claim 7, wherein, The deep Gaussian process model includes: The control input quantity is used as the input of the first layer Gaussian process to generate hidden variables; The hidden variables are used as the input of the second layer Gaussian process to output the posterior predictive distribution of the comprehensive performance index.

10. The recarburizer screening intelligent control system of claim 7, wherein, The dynamic feasible region includes: for any candidate point in the control input space, predicting a probability that the candidate point does not result in screen blinding or equipment overload using the Gaussian process classifier ; All the requirements The set of control input points greater than or equal to the safety probability threshold value is defined as the dynamic feasible region at the current time.

Images