Bayesian optimization-based integrated learning reverse process parameter determination method

Abstract

The invention discloses an ensemble learning reverse process parameter determination method based on Bayesian optimization, and belongs to the field of process parameter optimization. The method comprises the steps that historical process data including target pressure, butterfly valve input angles, circulation volume and circulation volume are collected and preprocessed; a target pressure threshold interval is set, target pressure is divided into two sections, and target models are established respectively; constructing a plurality of machine learning models by taking the butterfly valve input angle, the circulation volume and the circulation volume as independent variables and the target pressure as a dependent variable, and combining the machine learning models through an integrated learning method; an input feature vector is generated, and the minimum value and the maximum value of the input angle of the butterfly valve and the corresponding circulation amount and circulation volume are obtained; inputting the feature vector into an integrated learning model to predict target pressure; the optimal hyper-parameter is automatically searched through the Bayesian optimization algorithm, and the model prediction performance is improved. According to the method, the accuracy and the efficiency of determining the process parameters are effectively improved through a method of combining ensemble learning and Bayesian optimization.

Description

Technical Field

[0001] This application relates to the field of industrial automation control technology, and particularly relates to a method for determining reverse process parameters based on Bayesian optimization and ensemble learning. Background Art

[0002] With the continuous development of industrial automation control systems, the butterfly valve, as a key actuator in fluid control systems, its precise control has an important impact on the stability and efficiency of the entire system. In the industrial production process, accurately controlling the fluid pressure is an important link to ensure product quality and production safety, and the relationship modeling between the input angle of the butterfly valve and the target pressure has become the core issue in determining process parameters.

[0003] Traditional butterfly valve control methods mainly rely on empirical formulas or look-up tables. These methods are usually derived based on steady-state operating conditions and are difficult to adapt to dynamic disturbances such as changes in fluid viscosity and pipeline vibration. At the same time, although traditional PID controllers are widely used, their performance highly depends on accurate mathematical models and is prone to overshoot or oscillation when facing the non-linear flow characteristics of butterfly valves. Especially in the reverse control scenario, that is, when the target pressure is known and the butterfly valve angle needs to be determined, traditional methods often need to be implemented through iterative trial-and-error or look-up tables, with low efficiency and difficulty in dealing with multi-variable coupling problems.

[0004] In recent years, with the development of artificial intelligence technology, machine learning methods have shown great potential in industrial process modeling and optimization. CN118839823A discloses a heat load prediction system and method for a heating system based on ensemble learning. This method uses an improved stacking strategy to integrate machine learning and deep learning algorithms to construct an ensemble learning model, which can adapt to the complexity and dynamic changes of different heating systems1. CN119357914A proposes a method for predicting the hydrate deposition thickness based on ensemble learning. This method constructs a stacking model based on XGBoost, random forest, and gradient boosting regression tree, and uses Bayesian optimization to optimize the weight values of each model, effectively solving the problems of traditional numerical simulation methods in terms of prediction accuracy and efficiency2.

[0005] In terms of Bayesian optimization, CN117057201A discloses a method for collecting and predicting turbine blade wax mold data based on Bayesian optimization. This method combines numerical simulation, can quickly and automatically construct a database, and obtains the predictive relationship between process parameters and deformation indicators through a machine learning model3. This method provides a new idea for optimizing complex process parameters.

[0006] In the field of predicting key parameters in industrial processes, CN112749849A proposed an online prediction method for key parameters in the continuous catalytic reforming process. This method effectively improves the prediction accuracy of the model in the process by constructing multiple adaptive regression submodels and integrating them4. CN106897775B discloses a soft-sensor modeling method based on Bayesian ensemble learning. Under the framework of the model cluster analysis algorithm, multiple soft-sensor algorithms suitable for different industrial process characteristics are used as sub-algorithms, and the Bayesian ensemble learning method is used to calculate the weights of each soft-sensor sub-algorithm model, improving the adaptability and prediction accuracy of the soft-sensor model to different industrial process characteristics5.

[0007] However, the existing technologies still have the following deficiencies: First, the existing valve angle calibration methods are mainly based on fitting laboratory flow curves, which are difficult to reproduce the complexity of real working conditions, have long testing times, and the fitting curves are static mapping relationships, unable to compensate for the characteristic offsets caused by valve mechanical wear in real time. Second, when existing machine learning is applied to valve characteristic modeling, the modeling ability of a single model for high-dimensional parameter spaces is limited and it is prone to falling into local optima. In addition, traditional optimization algorithms have too high computational costs during hyperparameter tuning and are difficult to meet the requirements of real-time control. Especially in the case where the target pressure has different working condition intervals, a single model is difficult to simultaneously take into account the prediction accuracy in different pressure intervals.

[0008] Finally, although there are applications of ensemble learning and Bayesian optimization in the existing technologies, there is a lack of a method to organically combine the two and apply them to the determination of the reverse process parameters of butterfly valves. Especially, there are technical gaps in dealing with segmented modeling of target pressure, generation of feature vectors, and optimization of model hyperparameters, unable to meet the industrial control requirements of high precision and high efficiency. Summary of the Invention

[0009] To solve the technical problems faced by traditional butterfly valve control methods in industrial automation control systems and achieve high-precision and strong self-adaptive butterfly valve angle prediction, the present invention provides a method and system for determining reverse process parameters based on Bayesian optimization and ensemble learning.

[0010] The present application specifically adopts the following technical solutions to achieve the above objectives: Provide a method and system for determining reverse process parameters based on Bayesian optimization and ensemble learning, including the following steps: First, collect historical process data, including flow rate, volume, target pressure, and the corresponding butterfly valve angles, and perform preprocessing on the data, including operations such as noise removal and normalization, to ensure the accuracy of the data and enable the model to learn on the same scale, avoiding learning biases caused by variable scale differences.

[0011] Further, the pressure is divided into two segments. A first target model is separately established for a segment of data with a pressure range of 12000 - 94000 Pa, and a second target model is established for pressure values not within the range of 12000 - 94000 Pa, so as to improve the prediction accuracy of the model in different pressure ranges.

[0012] Preferably, an ensemble learning model is constructed: taking the input angle, flow rate, and volume of the butterfly valve as independent variables and the pressure as the dependent variable, multiple machine learning models are constructed, including an SVN pressure prediction model, an xgboost pressure prediction model, and an elm extreme learning pressure prediction model, and these models are combined through a weighted average ensemble learning method to establish a forward pressure prediction model.

[0013] Further, the minimum and maximum values of the angle, as well as the flow rate and volume, are obtained. A feature vector is generated for every 0.1 angle, and then the above-mentioned ensemble learning model is called to predict the pressure.

[0014] Preferably, the hyperparameters of the ensemble learning model are optimized using the Bayesian optimization algorithm. The specific process includes: constructing a surrogate model (such as a Gaussian process) to fit the objective function; calculating the Expected Improvement (EI) value; selecting the position corresponding to the maximum EI value in the calculated EI for evaluation; adding the new observed value to the model; and obtaining the input angle of the butterfly valve corresponding to all parameters by continuously generating the parameter space through the above process.

[0015] Finally, the angle corresponding to the minimum pressure error is obtained as the final valve angle.

[0016] The beneficial effects of this application include: By integrating multiple machine learning models and combining the Bayesian optimization algorithm for hyperparameter tuning, the accuracy of butterfly valve angle prediction is significantly improved; this method can automatically adjust the butterfly valve angle according to different process conditions and target pressures, adapt to complex industrial environments, and has strong self - adaptability; the Bayesian optimization algorithm can quickly find the optimal combination of hyperparameters, reduce the time and computational resource consumption of model training, and improve efficiency; the ensemble learning method can effectively reduce the overfitting risk of a single model, improve the robustness and generalization ability of the model; in the dynamic pressure compensation scenario, the angle prediction error ≤ 0.8°, and the response delay < 200 ms, which is two orders of magnitude higher than traditional methods, solving the limitations of traditional butterfly valve control methods in the face of non - linear flow characteristics, multivariable coupling problems, and characteristic offsets caused by valve mechanical wear.

[0017] Other beneficial effects or advantages of this application will be described in detail in combination with the specific structure in the specific implementation manner. Brief Description of the Drawings

[0018] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those skilled in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings. In addition, it should be understood that the proportional relationships of the various components in the accompanying drawings of this specification do not represent the proportional relationships in actual material selection and design. It is only a schematic diagram of the structure or position, where: Figure 1 Timing diagram representing the target pressure value Figure 2 Normalized processing diagram of the timing of the target pressure value Figure 3 Fitting model diagram of the target pressure value based on the SVM flow valve angle Figure 4 Fitting model diagram of the target pressure value based on the XGBoost flow valve angle Figure 5 Fitting model diagram of the target pressure value based on the ELM flow valve angle Figure 6 Fitting model diagram of the target pressure value based on the integrated model flow valve angle Figure 7 Reverse angle prediction fitting model diagram of the integrated model based on Bayes Specific implementation manners

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Usually, the components of the embodiments of the present application described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application to be protected, but only represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application. Embodiment

[0021] An integrated learning reverse process parameter determination method based on Bayesian optimization, comprising the following steps: Data collection and preprocessing: Collect historical process data including target pressure, butterfly valve input angle, flow rate, and flow volume, and preprocess the data. Specifically, extract production data from the industrial production system for the past three months, including target pressure records, butterfly valve input angle setting values, flow rate measurement values, and flow volume calculation values for each production batch. These data are usually stored in the production database in the form of a time series, and each record contains a timestamp, target pressure value (in Pa), butterfly valve input angle value (in degrees), flow rate value (in m³ / h), and flow volume value (in m³). Specifically, as Figure 1 shown, it presents a sample of historical process data with different pressure values.

[0022] Preprocess the collected data, including noise removal and normalization. Noise removal is to ensure the accuracy of the data and remove possible outliers or noise. The specific operation is to identify outliers through statistical methods, such as using the 3σ principle, that is, considering data points deviating from the mean by more than 3 standard deviations as outliers and removing them. Normalization is to enable the model to learn on the same scale and avoid learning biases caused by variable scale differences. The Min-Max normalization method is adopted to map each eigenvalue to the interval [0,1]. The conversion formula is: X_norm = (X - X_min) / (X_max - X_min), where X is the original data, X_min and X_max are the minimum and maximum values of this feature respectively, and X_norm is the normalized data. Specifically, as Figure 2 shown, it presents the normal standardization after preprocessing a sample of historical process data with different pressure values.

[0023] Target pressure segmentation: Set the target pressure threshold interval for the target pressure. The part overlapping with the target pressure threshold interval value is the first segment of target pressure values, and a first target model is established based on the first segment of target pressure values. The part not overlapping with the target pressure threshold interval value is the second segment of target pressure values, and a second target model is established based on the second segment of target pressure values.

[0024] Specifically, set the target pressure threshold interval as 12000Pa - 94000Pa. Classify the preprocessed dataset according to the target pressure value: Data points with target pressure in the range of 12000Pa - 94000Pa are classified as the first segment of target pressure data for establishing the first target model; Data points with target pressure below 12000Pa or above 94000Pa are classified as the second segment of target pressure data for establishing the second target model.

[0025] The establishment process of the first target model is as follows: Using the first segment of target pressure data, with the input angle of the butterfly valve, the flow rate, and the flow volume as input features, and the target pressure as the output label, a regression model is trained. This model uses the random forest algorithm, sets the number of trees to 100, the maximum depth to 10, and the minimum number of samples in a leaf node to 5, and determines the optimal parameter configuration through cross-validation.

[0026] The establishment process of the second target model is as follows: Using the second segment of target pressure data, also with the input angle of the butterfly valve, the flow rate, and the flow volume as input features, and the target pressure as the output label, another regression model is trained. This model uses the gradient boosting decision tree algorithm, sets the number of trees to 150, the learning rate to 0.05, and the maximum depth to 8, and determines the optimal parameter configuration through grid search.

[0027] Construction of the ensemble learning model: Let the input angle of the butterfly valve, the flow rate, and the flow volume be independent variables, and the target pressure be the dependent variable. Construct multiple machine learning models, and combine these models through the method of ensemble learning to obtain an ensemble learning model that can predict the target pressure.

[0028] Specifically, construct three different types of machine learning models: the SVN pressure prediction model, the xgboost pressure prediction model, and the elm extreme learning pressure prediction model.

[0029] Construction of the SVN pressure prediction model: Use the support vector regression (SVR) algorithm, select the RBF (radial basis function) as the kernel function, set the penalty parameter C to 10, the epsilon parameter to 0.1, and the gamma parameter to 0.01. The input features are the input angle of the butterfly valve, the flow rate, and the flow volume, and the output is the predicted target pressure value. Specifically, as Figure 3 shown, when the correlation coefficient r between the target pressure value and the predicted value is 0.96296 / / mae = 3478.433, it shows the fitting correlation coefficient and the fitting mean absolute error of this model, which is convenient for further evaluating the goodness of model fitting; Construction of the xgboost pressure prediction model: Use the XGBoost algorithm, set the number of trees to 200, the learning rate to 0.1, the maximum depth to 6, the subsampling rate to 0.8, and the column sampling rate to 0.8. The input features are also the input angle of the butterfly valve, the flow rate, and the flow volume, and the output is the predicted target pressure value. Specifically, as Figure 4 shown, when the correlation coefficient r between the target pressure value and the predicted value is 0.95716 / / mae = 4332.6201, it shows the fitting correlation coefficient and the fitting mean absolute error of this model, which is convenient for further evaluating the goodness of model fitting; Construction of the elm extreme learning pressure prediction model: Using the extreme learning machine (ELM) algorithm, the number of hidden layer neurons is set to 100, and the sigmoid function is selected as the activation function. The input features and output are the same as those of the previous two models. Specifically, as Figure 5 shown, when the correlation coefficient r between the target pressure value and the predicted value is 0.95655 / / mae = 4899.9888, the fitting correlation coefficient and the fitting mean absolute error of the model are shown, which is convenient for further evaluating the goodness of model fitting.

[0030] As Figure 6 shown, when the correlation coefficient r between the target pressure value and the predicted value is 0.97331 / / mae = 3333.2029, the fitting correlation coefficient and the fitting mean absolute error of the model are shown, which is convenient for further evaluating the goodness of model fitting; The above three models are combined by the weighted average method to obtain the target pressure value predicted by the ensemble learning model. Specifically, weights are assigned to each model, with the weight of the SVN model being 0.3, the weight of the xgboost model being 0.5, and the weight of the elm model being 0.2. The final prediction result is the weighted average of the predicted values of the three models: Predicted pressure = 0.3 × SVN predicted value + 0.5 × xgboost predicted value + 0.2 × elm predicted value Optimize the hyperparameters of the ensemble learning model: Generate input feature vectors, obtain the minimum and maximum values of the input angle of the butterfly valve, as well as the flow rate and flow volume at each angle. A feature vector is generated every 0.1-degree change. Specifically, the minimum value of the input angle of the butterfly valve extracted from the preprocessed data is 0 degrees, and the maximum value is 90 degrees. In this range, an angle sequence is generated at a step of 0.1 degree: 0 degrees, 0.1 degrees, 0.2 degrees,..., 89.9 degrees, 90 degrees. For each angle value, the corresponding average flow rate and flow volume are calculated according to the historical data. If a certain angle value has no corresponding record in the historical data, its flow rate and flow volume are estimated by linear interpolation. Finally, 900 feature vectors are generated, and each feature vector contains three elements: the input angle of the butterfly valve, the corresponding flow rate, and the flow volume.

[0031] Use the ensemble learning model to predict the pressure, input all the feature vectors into the trained ensemble learning model for prediction, and obtain the target pressure corresponding to each input. Specifically, the generated 900 feature vectors are sequentially input into the ensemble learning model to obtain 900 predicted target pressure values. These predicted values constitute the mapping relationship between the input angle of the butterfly valve and the target pressure, which can be used for subsequent determination of reverse parameters.

[0032] Optimize the model hyperparameters. Through the Bayesian optimization algorithm, automatically search for the optimal hyperparameters to improve the prediction performance of the ensemble learning model. The specific process of the Bayesian optimization algorithm is as follows: Construct a surrogate model: Select the Gaussian process as the surrogate model to fit the objective function. The Gaussian process is a probability model that can model functions and provide an estimate of the uncertainty of the predicted values. In this embodiment, the objective function is the root mean square error (RMSE) of the ensemble learning model on the validation set, and this error needs to be minimized. The input of the surrogate model is the hyperparameter combination, and the output is the corresponding RMSE estimate value and its uncertainty.

[0033] Calculate the EI value: Calculate the Expected Improvement (EI) at each trial point. EI is a commonly used acquisition function in Bayesian optimization, aiming to calculate which point to choose next to maximize the expected value of the optimization result by analyzing the current optimization status. The goal of EI is to select a point for sampling and evaluate its function value in each iteration to find the global optimal solution. The calculation formula for the EI value is: EI(x) = E[max(f_best - f(x), 0)] where f_best is the currently known best function value, and f(x) is the function value at point x.

[0034] Optimize EI: Select the position corresponding to the maximum EI value in the calculated EI for evaluation. Specifically, use the L-BFGS-B optimization algorithm to find the point with the maximum EI value in the hyperparameter space. This point represents the hyperparameter combination that is most likely to improve the current optimal solution.

[0035] Update the model: Add the new observed value to the model. Specifically, use the selected hyperparameter combination to train the ensemble learning model, calculate the RMSE on the validation set, and then add this hyperparameter combination and its corresponding RMSE to the observed dataset to update the Gaussian process model.

[0036] By continuously generating the parameter space as described above, obtain the butterfly valve input angles corresponding to all parameters. Specifically, repeat the above steps 50 times, or until the improvement amplitude of the RMSE is less than 0.001 for 10 consecutive iterations. Finally, select the hyperparameter combination with the minimum RMSE on the validation set as the optimal hyperparameters for training the final ensemble learning model.

[0037] The optimal hyperparameter combination includes: the C parameter of the SVN model is 15.7, the epsilon parameter is 0.08, and the gamma parameter is 0.015; the number of trees of the xgboost model is 180, the learning rate is 0.08, and the maximum depth is 7; the number of neurons in the hidden layer of the elm model is 120; the weights of the three models are 0.25, 0.55, and 0.2 respectively.

[0038] Retrain the ensemble learning model using the optimal hyperparameters and use it to predict the target pressure corresponding to different input angles of the butterfly valve. By establishing the mapping relationship between the input angle of the butterfly valve and the target pressure, the reverse process parameter determination can be realized: given the target pressure value, find the corresponding input angle of the butterfly valve. Embodiment

[0039] An ensemble learning reverse process parameter determination method based on Bayesian optimization includes the following steps: Data collection and preprocessing: Collect historical process data including target pressure, butterfly valve input angle, flow rate, and flow volume, and preprocess the data. Specifically, extract the production data of the past six months from the industrial production system, including the target pressure records, butterfly valve input angle setting values, flow rate measurement values, and flow volume calculation values of each production batch.

[0040] Preprocess the collected data, including noise removal and normalization. Noise removal is to ensure the accuracy of the data and remove possible outliers or noise. In this embodiment, the box plot method is used to identify outliers, and the data points that exceed the upper quartile plus 1.5 times the interquartile range or are lower than the lower quartile minus 1.5 times the interquartile range are regarded as outliers and removed. Normalization is to enable the model to learn on the same scale and avoid learning biases caused by variable scale differences. In this embodiment, the Z-score standardization method is used to convert each feature into a distribution with a mean of 0 and a standard deviation of 1. The conversion formula is: X_std = (X - μ) / σ, where X is the original data, μ is the mean of this feature, σ is the standard deviation of this feature, and X_std is the standardized data.

[0041] Target pressure segmentation: Set the target pressure threshold interval of the target pressure. The part overlapping with the target pressure threshold interval value is the first segment of target pressure values, and the first target model is established through the first segment of target pressure values. The part not overlapping with the target pressure threshold interval value is the second segment of target pressure values, and the second target model is established through the second segment of target pressure values.

[0042] Specifically, the target pressure threshold range is set to 12000Pa - 94000Pa. The preprocessed dataset is classified according to the target pressure value: data points with target pressure in the range of 12000Pa - 94000Pa are classified as the first-stage target pressure data for establishing the first target model; data points with target pressure below 12000Pa or above 94000Pa are classified as the second-stage target pressure data for establishing the second target model.

[0043] The establishment process of the first target model is as follows: Using the first-stage target pressure data, with the input angle of the butterfly valve, the flow rate, and the flow volume as input features, and the target pressure as the output label, a regression model is trained. This model uses the support vector regression algorithm, the kernel function is selected as the polynomial kernel, the polynomial degree is 3, and the penalty parameter C is set to 5.

[0044] The establishment process of the second target model is as follows: Using the second-stage target pressure data, also with the input angle of the butterfly valve, the flow rate, and the flow volume as input features, and the target pressure as the output label, another regression model is trained. This model uses the neural network algorithm, the hidden layer structure is [64, 32], the activation function is ReLU, the optimizer is Adam, the learning rate is 0.001, and the number of training epochs is 200.

[0045] Construction of the ensemble learning model: Let the input angle of the butterfly valve, the flow rate, and the flow volume be independent variables, and the target pressure be the dependent variable. Multiple machine learning models are constructed, and these models are combined through the method of ensemble learning to obtain an ensemble learning model that can predict the target pressure.

[0046] Specifically, three different types of machine learning models are constructed: the SVN pressure prediction model, the xgboost pressure prediction model, and the elm extreme learning pressure prediction model.

[0047] Construction of the SVN pressure prediction model: Use the support vector regression (SVR) algorithm, select the linear kernel as the kernel function, set the penalty parameter C to 1, and the epsilon parameter to 0.2. The input features are the input angle of the butterfly valve, the flow rate, and the flow volume, and the output is the predicted target pressure value.

[0048] Construction of the xgboost pressure prediction model: Use the XGBoost algorithm, set the number of trees to 150, the learning rate to 0.05, the maximum depth to 5, the subsampling rate to 0.7, and the column sampling rate to 0.7. The input features are also the input angle of the butterfly valve, the flow rate, and the flow volume, and the output is the predicted target pressure value.

[0049] Construction of the elm extreme learning pressure prediction model: Using the extreme learning machine (ELM) algorithm, the number of hidden layer neurons is set to 80, and the tanh function is selected as the activation function. The input features and output are the same as those of the previous two models.

[0050] The above three models are combined by the weighted average method to obtain an ensemble learning model. Specifically, weights are assigned to each model, with the weight of the SVN model being 0.2, the weight of the xgboost model being 0.6, and the weight of the elm model being 0.2. The final prediction result is the weighted average of the predicted values of the three models: Predicted pressure = 0.2 × SVN predicted value + 0.6 × xgboost predicted value + 0.2 × elm predicted value Optimize the hyperparameters of the ensemble learning model: Generate input feature vectors, obtain the minimum and maximum values of the input angle of the butterfly valve, as well as the flow rate and flow volume at each angle. A feature vector is generated every 0.1-degree change. Specifically, the minimum value of the input angle of the butterfly valve extracted from the preprocessed data is 5 degrees, and the maximum value is 85 degrees. In this range, an angle sequence is generated at a step size of 0.1 degree: 5 degrees, 5.1 degrees, 5.2 degrees,..., 84.9 degrees, 85 degrees. For each angle value, the corresponding average flow rate and flow volume are calculated based on historical data. If a certain angle value has no corresponding record in the historical data, its flow rate and flow volume are estimated by polynomial interpolation. Finally, 800 feature vectors are generated, and each feature vector contains three elements: the input angle of the butterfly valve, the corresponding flow rate, and the flow volume.

[0051] Use the ensemble learning model to predict pressure. Input all the feature vectors into the trained ensemble learning model for prediction to obtain the target pressure corresponding to each input. Specifically, the 800 generated feature vectors are sequentially input into the ensemble learning model to obtain 800 predicted target pressure values.

[0052] Optimize the model hyperparameters. Through the Bayesian optimization algorithm, automatically find the optimal hyperparameters to improve the prediction performance of the ensemble learning model. The specific process of the Bayesian optimization algorithm is as follows: Construct a surrogate model: Select the Gaussian process as the surrogate model to fit the objective function. In this embodiment, the Matérn kernel function is used, with the smoothing parameter ν set to 2.5 and the length scale parameter set to 1.0.

[0053] Calculate the EI value: Calculate the Expected Improvement (EI) at each trial point. In this embodiment, the EI formula with the exploration parameter k is used: EI(x) = E[max(fbest - f(x) - k, 0)] where k is set to 0.01, which is used to control the balance between exploration and exploitation.

[0054] Optimize EI: Select the position corresponding to the maximum EI value in the calculated EI for evaluation. Specifically, use the particle swarm optimization algorithm to find the point with the maximum EI value in the hyperparameter space. The number of particles is set to 20, and the number of iterations is set to 100.

[0055] Update the model: Add the new observation value to the model. After each evaluation, update the hyperparameters of the Gaussian process model, including the length scale and signal variance of the kernel function.

[0056] By continuously generating the parameter space as described above, the input angles of the butterfly valve corresponding to all parameters are obtained. Specifically, repeat the above steps 30 times, or until the improvement amplitude of the RMSE in 5 consecutive iterations is less than 0.005. Finally, select the combination of hyperparameters with the minimum RMSE of the validation set as the optimal hyperparameters.

[0057] The optimal hyperparameter combination includes: the C parameter of the SVN model is 8.3, and the epsilon parameter is 0.15; the number of trees of the xgboost model is 165, the learning rate is 0.06, and the maximum depth is 6; the number of neurons in the hidden layer of the elm model is 95; the weights of the three models are 0.15, 0.65, and 0.2 respectively.

[0058] As Figure 7 shown, when the correlation coefficient between the input angle value of the butterfly valve and the predicted value is r = 0.98421 / / mae = 0.39821, the fitting correlation coefficient and the fitting mean absolute error of the model are shown, which is convenient for further evaluating the goodness of the model fitting; Retrain the ensemble learning model using the optimal hyperparameters for reverse prediction and use it to predict the target pressure corresponding to different input angles of the butterfly valve. By establishing the mapping relationship between the input angle of the butterfly valve and the target pressure, the reverse process parameter determination can be realized: given the target pressure value, find the corresponding input angle of the butterfly valve.

[0059] It should be noted that both Example 1 and Example 2 are one kind of the method for determining the reverse process parameters of ensemble learning based on Bayesian optimization.

[0060] Among them: To objectively evaluate the established model, this scheme uses the correlation coefficient r and the mean absolute error mae as indicators for evaluation. The calculation formula of the coefficient of determination r is the prior art and will not be elaborated here.

[0061] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, by way of example and not limitation, alternative configurations of the embodiments of this specification may be regarded as consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly presented and described in this specification.

Claims

1. An integrated learning reverse process parameter determination method based on Bayesian optimization, characterized in that, Including: Data acquisition and preprocessing: Collect historical process data including target pressure, butterfly valve input angle, flow rate, and flow volume, and preprocess the data; Target pressure segmentation: Set the target pressure threshold range of the target pressure. The part overlapping with the target pressure threshold range value is the first segment of target pressure values, and the first target model is established through the first segment of target pressure values. The part not overlapping with the target pressure threshold range value is the second segment of target pressure values, and the second target model is established through the second segment of target pressure values; Integrated learning model construction: Set the butterfly valve input angle, flow rate, and flow volume as independent variables and the target pressure as the dependent variable, construct multiple machine learning models, and combine these models through the method of integrated learning to obtain an integrated learning model that can predict the target pressure; Optimize the hyperparameters of the integrated learning model: Generate input feature vectors, obtain the minimum and maximum values of the butterfly valve input angle, as well as the flow rate and flow volume at each angle, and generate a feature vector every 0.1 degree change; Use the integrated learning model to predict pressure, input all the feature vectors into the trained integrated learning model for prediction, and obtain the target pressure corresponding to each input; Optimize the model hyperparameters, and automatically find the optimal hyperparameters through the Bayesian optimization algorithm to improve the prediction performance of the integrated learning model.

2. The method for determining reverse process parameters of an ensemble learning based on Bayesian optimization according to claim 1, wherein The process of preprocessing the data includes noise removal and normalization processing.

3. The method for determining reverse process parameters of ensemble learning based on Bayesian optimization according to claim 1, wherein, The target pressure threshold range of the target pressure is 12000Pa - 94000Pa.

4. The method for determining reverse process parameters of ensemble learning based on Bayesian optimization according to claim 1, wherein The construction of multiple machine learning models specifically includes: SVN pressure prediction model, xgboost pressure prediction model, and elm extreme learning pressure prediction model.

5. The method for determining reverse process parameters of ensemble learning based on Bayesian optimization according to claim 4, characterized in that The SVN pressure prediction model, xgboost pressure prediction model, and elm extreme learning pressure prediction model are combined through the weighted average method to obtain an integrated learning model that can predict the target pressure.

6. The method for determining reverse process parameters of ensemble learning based on Bayesian optimization according to claim 1, wherein The specific process of the Bayesian optimization algorithm is: Construct a surrogate model: Select a model to fit the objective function; Calculate the EI value: Calculate the expected improvement at each trial point; Optimize EI: Select the position corresponding to the maximum EI value in the calculated EI for evaluation; Update the model: Add the new observation value to the model; By continuously generating the parameter space as described above, obtain the butterfly valve input angles corresponding to all parameters.

Images