Surfactant effect evaluation and process optimization method based on data analysis

Abstract

The present invention relates to a surfactant effect evaluation and process optimization method based on data analysis. The method comprises the following steps: data collection and preprocessing are performed on the production parameters of the surfactant to obtain a preprocessing data set; four unstructured models are constructed based on the preprocessing data set, and the target cleaning efficiency of the surfactant is calculated; a physical information composite network is constructed based on the target cleaning efficiency and the training set in the preprocessing data set, and training is performed to obtain a performance evaluation model; based on the performance evaluation model, multi-granularity information fusion and performance evaluation are performed on the test set in the preprocessing data set to obtain a target performance evaluation result; according to the target performance evaluation result, parameter sensitivity analysis and optimization solution are performed on the surfactant formula and process parameters to obtain a target process optimization scheme. The implementation of the present invention improves the performance evaluation accuracy and process optimization efficiency of the surfactant.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a surfactant effect evaluation and process optimization method based on data analysis. Background Art

[0002] Traditional surfactant effect evaluation methods mainly rely on empirical judgment and simple laboratory tests, which are difficult to fully reflect the actual performance in complex industrial environments. At the same time, surfactant formulation optimization and process parameter adjustment are often based on trial and error, which is time-consuming, labor-intensive and inefficient.

[0003] However, how to effectively integrate multi-source heterogeneous data, build an accurate surfactant performance evaluation model, and how to achieve joint optimization of surfactant formulation and process parameters under multi-objective constraints are still challenges facing this field. In addition, the cleaning process of surfactants involves complex physical and chemical mechanisms. How to combine physical information with data-driven models to improve the interpretability and generalization ability of the model is also an urgent problem to be solved. Summary of the invention

[0004] The main purpose of the present invention is to provide a surfactant effect evaluation and process optimization method based on data analysis to improve the accuracy of surfactant performance evaluation and process optimization efficiency.

[0005] To achieve the above object, the present invention provides a surfactant effect evaluation and process optimization method based on data analysis, comprising the following steps:

[0006] Performing data collection and preprocessing on production parameters of surfactants to obtain a preprocessed data set;

[0007] constructing four unstructured models based on the preprocessed data set and calculating the target cleaning efficiency of the surfactant;

[0008] Building a physical information composite network based on the target cleaning efficiency and a training set in the preprocessed data set, and performing training to obtain a performance evaluation model;

[0009] Based on the performance evaluation model, multi-granularity information fusion and performance evaluation are performed on the test set in the preprocessed data set to obtain a target performance evaluation result;

[0010] According to the target performance evaluation results, parameter sensitivity analysis and optimization solution are performed on the surfactant formula and process parameters to obtain the target process optimization plan.

[0011] The present invention also provides a surfactant effect evaluation and process optimization device based on data analysis, comprising:

[0012] A collection module is used to collect and preprocess the production parameters of the surfactant to obtain a preprocessed data set;

[0013] A calculation module, for constructing four unstructured models based on the preprocessed data set and calculating the target cleaning efficiency of the surfactant;

[0014] A construction module is used to construct a physical information composite network based on the target cleaning efficiency and the training set in the preprocessing data set, and perform training to obtain a performance evaluation model;

[0015] An evaluation module, used to perform multi-granularity information fusion and performance evaluation on the test set in the preprocessed data set based on the performance evaluation model to obtain a target performance evaluation result;

[0016] The solution module is used to perform parameter sensitivity analysis and optimization solution on the surfactant formula and process parameters according to the target performance evaluation results to obtain the target process optimization plan.

[0017] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0018] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0019] In summary, the technical solution provided by the present invention improves data quality and availability through multi-source data collection and preprocessing, combined with technologies such as outlier detection, missing value filling and data enhancement, and provides a reliable data basis for subsequent analysis. Four unstructured models (linear model, Langmuir model, Freundlich model and exponential model) are used to model the cleaning efficiency of surfactants, and the target cleaning efficiency is obtained by comprehensive weighted calculation, which improves the accuracy and reliability of the cleaning efficiency prediction. A physical information composite network containing four deep neural networks is constructed, and modeling is performed for surfactant concentration, dirt concentration, pH value and temperature respectively, realizing multi-dimensional and multi-scale surfactant performance evaluation. Through low-frequency information pre-training and two-stage loss balance training strategy, the training effect and generalization ability of the model are improved, making the performance evaluation model more stable and reliable. The test set is comprehensively analyzed using multi-granularity information fusion technology to obtain comprehensive target performance evaluation results. The Sobol global sensitivity analysis method is used to identify the key parameters that significantly affect the performance of surfactants, and the pertinence and efficiency of the optimization process are improved. The NSGA-III algorithm was used for multi-objective optimization, combined with the fuzzy comprehensive evaluation method, to achieve the joint dynamic optimization of surfactant formulation and process parameters, and obtain the optimal solution that takes into account multiple performance indicators, thereby improving the accuracy of surfactant performance evaluation and the efficiency of process optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the steps of a surfactant effect evaluation and process optimization method based on data analysis in one embodiment of the present invention;

[0021] Figure 2 It is a structural block diagram of a surfactant effect evaluation and process optimization device based on data analysis in one embodiment of the present invention;

[0022] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0023] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0025] Reference Figure 1 This embodiment provides a surfactant effect evaluation and process optimization method based on data analysis, comprising the following steps:

[0026] S1, data collection and preprocessing of production parameters of surfactants to obtain a preprocessed data set;

[0027] Among them, data such as surfactant formula composition, production process parameters, surfactant consumption, and cleaning residues are collected from multiple sources to generate an original data set containing multi-dimensional information. Outlier detection is performed on the original data set. The interquartile range method is used to calculate the upper and lower thresholds of the data. The interquartile range (IQR) is obtained by calculating the first quartile (Q1) and the third quartile (Q3) of each variable in the original data set, and then the upper and lower thresholds are determined by setting a range of 1.5 times the interquartile range. All data points outside this range are marked as outliers, forming an outlier marked data set with outlier marks. The marked outliers are replaced by the median. The median of the data in the column where the outlier is located is calculated, and the marked outlier is replaced with the median of the column. In this way, the impact of outliers on the subsequent modeling process is minimized, while the overall distribution trend of the data is retained to generate a data set after outlier processing. The missing values ​​in the outlier processed data set are filled. For time series data, the missing values ​​are filled by linear interpolation, that is, the missing values ​​are estimated according to the data trend of adjacent time points to ensure the continuity and rationality of the time series data. For non-time series data, the K nearest neighbor interpolation method is used for filling. By calculating the K neighbors most similar to the missing value samples and taking their average or weighted average, the missing data is effectively filled, the overall consistency and representativeness of the data are maintained, and a complete data set is obtained. The complete data set is standardized, the data is converted into a form with a mean of 0 and a standard deviation of 1, the dimensional differences between different features are eliminated, and a standardized data set is generated. The standardized data set is arranged in chronological order to ensure the continuity of the time series features, and a fixed-ratio stratified sampling method is used to divide it into a training set, a validation set, and a test set, with a ratio of 70%, 15%, and 15%, respectively. The divided data set is obtained. Data enhancement is performed on the training set in the divided data set. For time series data, the sliding window technology is used to generate multiple subsequences by sliding on the time series with a fixed window length, increase the number of training samples, and capture the characteristics of different time periods. For non-time series data, the SMOTE (Synthetic Minority Oversampling Technology) algorithm is used to generate synthetic samples. New synthetic samples are generated by interpolating between minority class samples to balance the category distribution in the data set and obtain an enhanced training set. The enhanced training set, validation set, and test set are merged, and batch identifiers and timestamp information are added to form the final preprocessed data set. The batch identifier is used to distinguish different data processing batches for subsequent traceability and analysis; the timestamp information is used to record the data collection and processing time to ensure data traceability and time consistency.

[0028] S2, constructs four unstructured models based on the preprocessed dataset and calculates the target cleaning efficiency of surfactants;

[0029] Specifically, feature selection is performed on the preprocessed data set to screen out features that are highly correlated with the cleaning efficiency and utilization rate of surfactants from the preprocessed data set, and features that are irrelevant to the target or have weak correlation are removed to improve the training efficiency and prediction performance of the model. Through feature selection, a streamlined feature selection data set is obtained. Based on the feature selection data set, a linear model is constructed to describe the relationship between surfactant concentration and cleaning efficiency. In this linear model, cleaning efficiency is represented as ,in represents the surfactant concentration, represents the slope, is the intercept. In order to determine the model parameters and , use the least squares method to fit the data and estimate the initial and The goal of the least squares method is to minimize the sum of squared errors between the observed value and the model predicted value, so as to find the best fitting straight line to describe the linear relationship between cleaning efficiency and concentration. After the linear model is constructed, the Langmuir model is constructed based on the same feature selection data set. The Langmuir model is a nonlinear model used to describe the adsorption process of surfactants, which can reflect the gradual change of cleaning efficiency with concentration. In this model, the cleaning efficiency Expressed as ,in Indicates the maximum cleaning efficiency, is the Langmuir constant, which reflects the binding strength between the surfactant and the dirt. and The nonlinear least squares method is used for fitting. This method effectively handles nonlinear relationships and minimizes errors by iteratively adjusting parameters to find the best parameter estimates. At the same time, the Freundlich model is constructed based on the data set after feature selection. The Freundlich model is an empirical formula used to describe the adsorption behavior on multilayer adsorption or heterogeneous surfaces. In this model, the cleaning efficiency Expressed as ,in is the Freundlich constant, is the heterogeneity factor, reflecting the heterogeneity of the adsorption process. and The value of is used, and the original formula is converted into a linear form by logarithmic linear regression, and then these parameters are estimated by linear regression. In this way, the Freundlich model is fitted and the preliminary model parameters are obtained. An exponential model is constructed to describe the relationship between cleaning efficiency and surfactant concentration. In the exponential model, the cleaning efficiency is represented as ,in Indicates the maximum cleaning efficiency, is the rate constant, describing the rate at which concentration affects cleaning efficiency. In order to estimate the initial parameters in the model and , using the Levenberg-Marquardt algorithm. The Levenberg-Marquardt algorithm is an optimization algorithm that combines the advantages of gradient descent and Newton's method and is suitable for nonlinear least squares problems. Through this algorithm, the parameter value that minimizes the error is effectively found and a reasonable exponential model is constructed. The optimal model parameters of the linear model, Langmuir model, Freundlich model and exponential model are estimated to ensure that each model can accurately describe the relationship between the cleaning efficiency and concentration of the surfactant. By optimizing the model parameters, the fitting accuracy and prediction effect of the model are improved, and four final unstructured models are obtained. Based on the four unstructured models, the initial cleaning efficiency of the surfactant is calculated respectively. The initial cleaning efficiencies of the four models are comprehensively weighted. By introducing appropriate weights, considering the applicability and accuracy of each model in different situations, a weighted average is performed to obtain the target cleaning efficiency of the surfactant.

[0030] S3, constructing a physical information composite network based on the target cleaning efficiency and the training set in the preprocessing data set, and training it to obtain a performance evaluation model;

[0031] It should be noted that feature mapping is performed on the training set in the preprocessing data set to convert the original training data into higher-level and more representative features, so as to facilitate the subsequent training of the deep learning model and extract effective information. After the feature mapping is completed, a feature mapping data set is obtained, and the feature mapping data set is reorganized in time series to divide the feature mapping data set into multiple sequences with physical meanings, including surfactant concentration sequence, dirt concentration sequence, pH value sequence and temperature sequence. Based on the surfactant concentration sequence, the first deep neural network in the physical information composite network is constructed. The network adopts the structure of a one-dimensional convolutional neural network (1D-CNN), which specifically includes three one-dimensional convolutional layers, two maximum pooling layers and two fully connected layers. The role of the convolutional neural network is to extract the local features of the surfactant concentration changing over time to better capture the impact of concentration changes on the cleaning effect. In each convolutional layer, the ReLU activation function is used to increase the nonlinear expression ability of the model, and the batch normalization technology is combined to accelerate the convergence speed of the model and improve the generalization performance. Through this structure, the first output result is generated to characterize the characteristics of the impact of surfactant concentration on cleaning efficiency. The second deep neural network of the physical information composite network is constructed based on the dirt concentration sequence. The network adopts the long short-term memory (LSTM) structure. LSTM is a recurrent neural network (RNN) suitable for processing time series data and is good at capturing long-term dependencies. The deep neural network contains two LSTM layers and one fully connected layer. The number of hidden units in each LSTM layer is 64, which is used to fully capture the dynamic impact of dirt concentration changes over time on the cleaning effect. The fully connected layer uses the tanh activation function to map the features extracted by LSTM to a space suitable for subsequent feature fusion to generate the second output result. The third deep neural network is constructed based on the pH value sequence and adopts a bidirectional gated recurrent unit (bidirectional GRU) structure. The bidirectional GRU processes the sequence data from both the forward and backward directions at the same time, which can better capture the bidirectional impact of pH value on cleaning efficiency. The network contains two bidirectional GRU layers and one fully connected layer. The number of hidden units in the GRU layer is 32, which is used to effectively extract the nonlinear impact characteristics of pH value changes on cleaning efficiency. The fully connected layer uses the Sigmoid activation function to map the extracted features to a reasonable range to generate the third output result. The fourth deep neural network is built based on the temperature series and adopts the temporal convolutional network (TCN) structure. TCN is a convolution-based architecture suitable for time series modeling. The deep neural network contains four causal convolutional layers and one fully connected layer. The causal convolutional layer uses dilated convolution and residual connection to increase the receptive field of the model and reduce the gradient vanishing problem, which can effectively capture the dynamic effect of temperature on cleaning efficiency. The fully connected layer uses the LeakyReLU activation function to retain negative information and improve the expressiveness of the model. Through this network, the fourth output result is generated.The first output result, the second output result, the third output result and the fourth output result are feature fused to form a fused feature vector. The feature fusion process combines the features extracted by each deep neural network to retain the comprehensive effects of surfactant concentration, dirt concentration, pH value and temperature on cleaning efficiency. Low-frequency information is extracted from the fused feature vector, and the low-frequency components in the fused feature vector are extracted using Fourier transform to capture the potential trends and patterns in these features. A variational autoencoder is constructed to pre-train the extracted low-frequency information. Through the variational autoencoder, low-frequency features can be compressed and decompressed to enhance the robustness of the features, and finally the pre-trained features are obtained. After obtaining the pre-trained features, the physical information composite network is trained in two stages of loss balance in combination with the target cleaning efficiency. In the first stage, the model is trained using the mean square error loss function to minimize the error between the cleaning efficiency predicted by the model and the target cleaning efficiency, so that the model can accurately fit the training data. In the second stage, in order to improve the physical interpretability and generalization ability of the model, iterative optimization is performed by combining the mean square error loss and the physical constraint loss based on the optimal model parameters. Physical constraint loss introduces constraints related to physical laws, so that the model not only has a good fitting effect on the data, but also conforms to the constraints and limitations in the actual physical process, thereby improving the physical consistency of the model. After a two-stage training process, a performance evaluation model that can accurately evaluate the effect of surfactants is finally obtained.

[0032] S4, based on the performance evaluation model, multi-granularity information fusion and performance evaluation are performed on the test set in the preprocessed data set to obtain the target performance evaluation result;

[0033] Specifically, the test set in the preprocessed data set is input into the first deep neural network of the performance evaluation model. The network is based on the structure of a one-dimensional convolutional neural network and can extract the characteristics of the change of surfactant concentration over time and process parameters from the input data. After multiple layers of convolution and pooling operations, the model extracts the change trend of the concentration distribution and obtains the distribution prediction result of the surfactant concentration. The test set in the preprocessed data set is input into the second deep neural network of the performance evaluation model. The network is based on the LSTM structure and is used to capture the dynamic change characteristics of the dirt concentration in the time series. Through the recursive structure of the two-layer LSTM network, the model can effectively learn the historical change pattern of the dirt concentration under different process conditions and its potential impact on the cleaning effect. In the process of processing the test data, the LSTM network extracts the time series characteristics of the dirt concentration and obtains the distribution prediction result of the dirt concentration through the full connection layer mapping. Through this result, the concentration change of the dirt during the cleaning process is analyzed, so as to evaluate the performance of the surfactant under the cleaning conditions of different dirt. At the same time, the test set is input into the third deep neural network of the performance evaluation model. The network is based on a bidirectional GRU structure and has the ability to process time series data simultaneously from the forward and backward directions, and can capture the bidirectional effect of pH value on the cleaning effect. Through the recursive structure of two-layer bidirectional GRU and the nonlinear mapping of the fully connected layer, the model outputs the distribution prediction results of pH value, reflecting the change of pH value under different conditions and its impact on cleaning efficiency. The change of pH value is an important factor affecting the performance of surfactants. The bidirectional GRU model can help determine the best use of surfactants under specific pH conditions by accurately predicting this parameter. The test set in the preprocessed data set is input into the fourth deep neural network of the performance evaluation model. The network is based on the temporal convolutional network (TCN) structure, which can process time series data and capture the long-term dependency of temperature changes. The TCN network effectively extracts the impact characteristics of temperature changes on the cleaning process through mechanisms such as causal convolution, dilated convolution and residual connection, and obtains the distribution prediction results of temperature through the fully connected layer. Through the prediction results of temperature distribution, the performance of surfactants under different temperature conditions can be understood. The prediction results of surfactant concentration distribution, dirt concentration distribution, pH distribution and temperature distribution are comprehensively analyzed. By combining the characteristics of each prediction result, a comprehensive feature vector is constructed to represent the comprehensive performance of surfactants under various physical parameters and process conditions.

[0034] S5, based on the target performance evaluation results, perform parameter sensitivity analysis and optimization on the surfactant formula and process parameters to obtain the target process optimization solution.

[0035] Among them, according to the target performance evaluation results, the Sobol global sensitivity analysis of the surfactant formula and process parameters is performed. Sobol global sensitivity analysis is used to evaluate the relative importance of input parameters to the model output. Based on the target performance evaluation results, the surfactant formula parameters and process parameters are analyzed, and the first-order sensitivity index and total sensitivity index of each parameter are calculated. The first-order sensitivity index measures the direct impact of a single parameter on the target output, while the total sensitivity index measures all impacts including interactions. By calculating the sensitivity index, the sensitivity ranking of each parameter is obtained, thereby identifying the key factors that have the greatest impact on the cleaning effect. Based on the results of the sensitivity ranking, parameters with a sensitivity index greater than 0.05 are selected as key optimization variables to construct the decision space of the optimization problem. The decision variables are discretized, and each continuous variable is divided into 10 equally spaced discrete levels to form a discretized decision variable set. Discretization helps to reduce computational complexity while ensuring the comprehensiveness of the search. Based on the discretized decision variable set, NSGA-III (non-dominated sorting genetic algorithm III) is used for optimization. NSGA-III is a multi-objective optimization algorithm suitable for solving problems with multiple conflicting objective functions. In NSGA-III, a population initialization function is constructed to generate an initial population. In this process, the Latin hypercube sampling algorithm is used to generate the initial population. Latin hypercube sampling is an efficient sampling method that can ensure that the population is evenly distributed in the decision space, thereby improving the search efficiency of the algorithm. The generated initial population is input into the performance evaluation model to calculate the fitness value of each population individual. The calculation of the fitness value involves multiple objective functions, including cleaning efficiency, surfactant utilization, and residue content. These three objective functions reflect the cleaning effect, the economic utilization of materials, and the evaluation of the post-cleaning effect, respectively. After the fitness value calculation is completed, the population individuals are non-dominated sorted and reference point associated to identify non-dominated individuals. The crowding distance calculation is used to select individuals to ensure that a diverse solution set is retained, followed by crossover and mutation operations to generate new individuals, and iterate continuously to obtain the Pareto optimal solution set. The Pareto optimal solution set is fuzzy comprehensively evaluated to construct an evaluation index system, including five important indicators: cleaning efficiency, surfactant utilization, residue content, cost, and environmental friendliness. Among them, the cleaning efficiency and surfactant utilization rate are set to weights of 0.3 and 0.2 respectively, the residue content is set to 0.2, and the cost and environmental friendliness are set to 0.15 respectively. The setting of weights reflects the relative importance of each indicator in process optimization. Based on the evaluation index system, a fuzzy relationship matrix is ​​constructed, and fuzzy comprehensive operations are performed to obtain the comprehensive evaluation value of each solution. Fuzzy comprehensive evaluation is a multi-criteria decision-making method that is suitable for dealing with complex problems with high uncertainty. Through fuzzy comprehensive operations, multiple different evaluation indicators are integrated together to obtain the overall score of each solution.The solution with the highest comprehensive evaluation value is selected as the target process optimization solution.

[0036] The sensitivity evaluation function of Sobol sensitivity analysis is constructed based on the target performance evaluation results. The cleaning efficiency, surfactant utilization rate and residue content are taken as components of the multi-sensitivity evaluation function, and the important process evaluation indicators are integrated into a comprehensive sensitivity evaluation function, so that the influence of multiple factors on the target performance can be measured simultaneously. The parameter space is defined for the surfactant formula and process parameters. A reasonable value range and distribution type are set for each parameter to determine the parameter sampling space. The value range of the parameter is set according to the actual process conditions, such as surfactant concentration, pH value, temperature, etc., and these range settings can cover all possible process conditions. At the same time, the distribution type of each parameter is clarified, such as uniform distribution or normal distribution, so as to reasonably describe the value probability of each parameter. Through the definition, a parameter sampling space containing all parameters is formed. The Saltelli sampling method is used to generate a parameter sample matrix from the parameter sampling space. Saltelli sampling is a global sensitivity analysis sampling method that can effectively reduce the variance of the sensitivity index estimation and improve the sampling efficiency. In the process of generating samples, the sample size is set to (2d+2) times the number of parameters, where d is the number of parameters. In this way, the sampling volume is ensured to be sufficient, so as to accurately estimate the influence of each parameter on the sensitivity evaluation function and obtain the sample set of Sobol analysis. Based on the generated Sobol analysis sample set, the sensitivity evaluation function value corresponding to each group of parameter samples is calculated. Each parameter combination is input into the comprehensive sensitivity evaluation function to calculate the corresponding function value. By calculating all samples one by one, an output matrix of the sensitivity evaluation function is formed, which contains the performance of all samples under different parameter combinations. Based on the output matrix of the sensitivity evaluation function, the first-order sensitivity index of each parameter is calculated. The first-order sensitivity index is used to measure the direct influence of a single parameter on the sensitivity evaluation function, reflecting the contribution of the parameter to the target output without the interaction of other parameters. By calculating the first-order sensitivity index of each parameter separately, a first-order sensitivity index set is formed. The total sensitivity index of each parameter is calculated based on the output matrix of the sensitivity evaluation function. The total sensitivity index takes into account the direct influence of a single parameter on the output, as well as the influence of the interaction between the parameter and other parameters. By calculating the total sensitivity index, the importance of each parameter in the entire system is evaluated, and the total sensitivity index set is obtained, which contains the overall contribution of each parameter under all interaction effects. The first-order sensitivity index set and the total sensitivity index set are normalized, and the sensitivity indexes of different dimensions are converted to the same scale so that they can be compared under the same standard. The parameters are arranged in descending order according to the size of the total sensitivity index to obtain the sensitivity ranking of the parameters.

[0037] In one example, data collection and preprocessing are performed on production parameters of surfactants to obtain a preprocessed data set, including:

[0038] The surfactant formula composition, production process parameters, surfactant consumption and cleaning residue data were collected from multiple sources to obtain the original data set;

[0039] Perform outlier detection on the original data set, use the interquartile range method to calculate the upper and lower thresholds, compare the data points in the original data set with the upper and lower thresholds, mark the data points that exceed the range as outliers, and obtain the outlier marked data set;

[0040] Based on the outlier marked data set, the marked outliers are subjected to median replacement processing, the median of the column where the outliers are located is calculated, and the outliers are replaced with the median of the column where they are located, so as to obtain the outlier processed data set;

[0041] Fill missing values ​​in the data set after outlier processing, perform linear interpolation on time series data, and perform K nearest neighbor interpolation on non-time series data to obtain a complete data set;

[0042] The complete data set is processed for data standardization to obtain a standardized data set, and the standardized data set is arranged in chronological order. A fixed-ratio stratified sampling method is used to divide the standardized data set into a training set, a validation set, and a test set, with the ratios of 70%, 15%, and 15%, respectively, to obtain the divided data set;

[0043] The training set in the divided data set is enhanced by using the sliding window technology for time series data and the SMOTE algorithm for non-time series data to generate synthetic samples to obtain an enhanced training set.

[0044] The augmented training set, validation set, and test set are merged, and batch identifier and timestamp information are added to form a preprocessed dataset.

[0045] In this example, the surfactant formula composition, production process parameters, surfactant consumption, and cleaning residue data are collected from multiple sources to construct the original data set. Data is collected from a variety of different data sources and sensors, including real-time data from online production monitoring systems, sample data measured in the laboratory, and historical records related to process conditions. Outlier detection is performed on the original data set. The interquartile range (IQR) method is used to calculate the upper and lower limit thresholds. The interquartile range method calculates the first quartile of the data ( ) and the third quartile ( ), and then by the formula Calculate the interquartile range. Based on the interquartile range, calculate the upper and lower thresholds, specifically:

[0046] Lower limit IQR, upper limit IQR;

[0047] Any data point that exceeds the upper and lower thresholds is marked as an outlier to form an outlier marked data set. The marked outliers are subjected to median replacement. The median of the data in the column where the outlier is located is calculated, and all marked outliers are replaced with the median to obtain the outlier processed data set. The missing values ​​in the outlier processed data set are filled to obtain a complete data set. For time series data, linear interpolation is used to fill in the data. According to the data of two adjacent time points, the linear function is used to estimate to ensure the continuity and consistency of the time series data. For non-time series data, the K nearest neighbor interpolation method is used to fill in the missing values ​​by finding the K neighbors most similar to the missing samples and using the mean of these neighbors to fill in the missing values. The complete data set is standardized to eliminate the dimensional differences between different features so that each feature can be compared on the same scale. The Z-standardization method is used to subtract the mean of each data point from its column and then divide it by the standard deviation to obtain a standardized data set. The standardized data set is arranged in chronological order, and a fixed-ratio stratified sampling method is used to divide the data set into a training set, a validation set, and a test set, with a ratio of 70%, 15%, and 15%, respectively. The purpose of stratified sampling is to ensure that the data distribution in each subset is consistent with the original data set, thereby improving the effectiveness of model training and evaluation. The divided data set is used for subsequent model training, validation, and testing to ensure that the model has good generalization ability. Data enhancement is performed on the training set in the divided data set. For time series data, the sliding window technology is used to slide on the time series by setting a fixed-length window, and each slide generates a new subsequence, thereby increasing the number of training samples and retaining the dynamic change characteristics in the time series. For example, assuming that the window length is 5 and each slide is 1 time step, the original time series generates multiple new samples with the same length, thereby increasing the diversity of training data. For non-time series data, the SMOTE (synthetic minority class oversampling) algorithm is used to generate synthetic samples. SMOTE generates new samples by linear interpolation between minority class samples to balance the proportion of each class sample in the data set, thereby improving the model's recognition ability for minority classes and obtaining an enhanced training set. The enhanced training set, validation set, and test set are merged, and batch identifiers and timestamp information are added to each piece of data to form the final preprocessed data set. The batch identifier is used to identify the result of which data processing the data belongs to, which facilitates the tracing and comparison of the effects of different batches of data; the timestamp information is used to record the data collection time to ensure the consistency and traceability of the data in time.

[0048] In one example, four unstructured models were constructed based on the preprocessed dataset and the target cleaning efficiency of the surfactants was calculated, including:

[0049] Performing feature selection on the preprocessed data set, selecting features associated with the cleaning efficiency and utilization rate of the surfactant, and obtaining a data set after feature selection;

[0050] Based on the data set after feature selection, a linear model was constructed, in which the cleaning efficiency E was expressed as E=k*C+b, C was the surfactant concentration, k was the slope, b was the intercept, and the initial k and b values ​​were estimated using the least squares method;

[0051] Based on the data set after feature selection, the Langmuir model is constructed, where the cleaning efficiency E is expressed as E=(Emax*K*C) / (1+K*C), Emax is the maximum cleaning efficiency, K is the Langmuir constant, and the initial Emax and K values ​​are estimated using the nonlinear least squares method;

[0052] Based on the data set after feature selection, the Freundlich model is constructed, where the cleaning efficiency E is expressed as E=F*C (1 / n) , F is the Freundlich constant, n is the heterogeneity factor, and the initial F and n values ​​are estimated using log-linear regression;

[0053] Based on the data set after feature selection, an exponential model was constructed, in which the cleaning efficiency E was expressed as E=a*(1-exp(-s*C)), a was the maximum cleaning efficiency, s was the rate constant, and the initial a and s values ​​were estimated using the Levenberg-Marquardt algorithm;

[0054] The optimal model parameters of the linear model, Langmuir model, Freundlich model and exponential model were estimated to obtain four unstructured models. The corresponding initial cleaning efficiencies were calculated using the four unstructured models, and the initial cleaning efficiencies were comprehensively weighted to obtain the target cleaning efficiency of the surfactant.

[0055] In this example, feature selection is performed on the preprocessed dataset to select features that are closely related to surfactant cleaning efficiency and utilization. By eliminating redundant or irrelevant features, the complexity of the model is reduced and the performance and interpretability of the model are improved. Use correlation analysis (such as the Pearson correlation coefficient) or other feature selection methods (such as recursive feature elimination) to identify important features. For example, after feature selection, features such as surfactant concentration, temperature, stirring speed, and pH value are selected to form a feature-selected dataset. Based on the feature-selected dataset, a simple linear model is constructed to describe the relationship between surfactant concentration and cleaning efficiency. In the linear model, cleaning efficiency It is expressed as:

[0056] ;

[0057] in, Indicates the cleaning efficiency, represents the surfactant concentration, represents the slope, which represents the influence of concentration on cleaning efficiency, and is the intercept, representing the cleaning efficiency baseline when the concentration is zero. In order to determine the parameters in the model and , the least squares method is used to fit the data, minimizing the sum of squared errors between the predicted value and the actual value, thereby finding the best fitting line. For example, in the experimental data, if the effect of concentration change on cleaning efficiency shows a linear relationship, the least squares method effectively finds the most suitable and The value enables the model to accurately describe the changing trend of cleaning efficiency. After completing the construction of the linear model, the Langmuir model is constructed based on the data set after feature selection. The Langmuir model is used to describe the saturation behavior of the surfactant adsorption process and the cleaning efficiency. is represented as:

[0058] ;

[0059] in, is the maximum cleaning efficiency, indicating the saturated cleaning capacity of the surfactant; is the Langmuir constant, which reflects the binding strength between the surfactant and the dirt; represents the surfactant concentration. and The initial value of is fitted by nonlinear least squares method, and the error is minimized by iteratively adjusting the parameters. At the same time, the Freundlich model is constructed based on the data set after feature selection. The Freundlich model is another empirical formula used to describe the adsorption process and is suitable for describing the multilayer adsorption behavior on heterogeneous surfaces. In the Freundlich model, the cleaning efficiency It is expressed as:

[0060] ;

[0061] in, is the Freundlich constant, representing the capacity of the adsorption process; is the heterogeneity factor, reflecting the heterogeneity of surface adsorption. Through logarithmic linearization, the formula is converted into logarithmic form:

[0062] ;

[0063] Use linear regression to estimate and The value of. By converting the nonlinear relationship into a linear relationship, the parameter estimation process is simplified. Based on the data set after feature selection, an exponential model is constructed to describe the relationship between cleaning efficiency and surfactant concentration. The exponential model is used to describe certain processes that gradually increase and tend to stabilize. In this model, the cleaning efficiency It is expressed as:

[0064] ;

[0065] in, Indicates the maximum cleaning efficiency, representing the final saturated cleaning effect; is the rate constant, reflecting the rate at which the concentration increases the cleaning efficiency. and The initial value of is fitted using the Levenberg-Marquardt algorithm. The Levenberg-Marquardt algorithm is a nonlinear optimization algorithm that combines gradient descent and Newton's method, which is suitable for dealing with this type of nonlinear regression problem. For example, when the concentration is low, the cleaning efficiency increases rapidly with the concentration, but at higher concentrations, the increase effect gradually weakens, and this trend is accurately described by the exponential model. The optimal model parameters of the linear model, Langmuir model, Freundlich model and exponential model are estimated to obtain four unstructured models that describe the cleaning efficiency of surfactants, each of which explains the cleaning process from a different perspective. For example, the linear model is used to preliminarily describe the simple linear effect of concentration on the cleaning efficiency, the Langmuir model and the Freundlich model are suitable for capturing the complex nonlinear behavior in the adsorption process, and the exponential model is used to describe the gradual saturation trend of the cleaning efficiency with the change of concentration. The corresponding initial cleaning efficiencies are calculated by the four unstructured models, and the initial cleaning efficiencies are comprehensively weighted to obtain the target cleaning efficiency of the surfactant.

[0066] In one example, a physical information composite network is constructed based on the target cleaning efficiency and the training set in the preprocessing data set, and is trained to obtain a performance evaluation model, including:

[0067] Performing feature mapping on the training set in the preprocessing data set to obtain a feature mapping data set, and reorganizing the feature mapping data set into a time series to divide the feature mapping data set into a surfactant concentration series, a dirt concentration series, a pH value series, and a temperature series;

[0068] A first deep neural network of a physical information composite network is constructed based on the surfactant concentration sequence, and a first output result is generated. The first deep neural network adopts a one-dimensional convolutional neural network structure, including three one-dimensional convolutional layers, two maximum pooling layers, and two fully connected layers. Each convolutional layer uses a ReLU activation function and batch normalization.

[0069] A second deep neural network of the physical information composite network is constructed based on the dirt concentration sequence, and a second output result is generated. The second deep neural network adopts an LSTM structure, including 2 LSTM layers and 1 fully connected layer. The number of hidden units in the LSTM layer is 64, and the fully connected layer uses a tanh activation function.

[0070] A third deep neural network of a physical information composite network is constructed based on the pH value sequence, and a third output result is generated. The third deep neural network adopts a bidirectional GRU structure, including 2 bidirectional GRU layers and 1 fully connected layer. The number of hidden units in the GRU layer is 32, and the fully connected layer uses a Sigmoid activation function.

[0071] A fourth deep neural network of a physical information composite network is constructed based on the temperature sequence, and a fourth output result is generated. The fourth deep neural network adopts a TCN structure, including 4 causal convolutional layers and 1 fully connected layer. The causal convolutional layer uses dilated convolution and residual connection, and the fully connected layer uses a LeakyReLU activation function.

[0072] The first output result, the second output result, the third output result and the fourth output result are subjected to feature fusion to obtain a fused feature vector, low-frequency information is extracted from the fused feature vector, low-frequency components are extracted using Fourier transform, and a variational autoencoder is constructed to perform low-frequency information pre-training to obtain pre-trained features;

[0073] Based on the pre-training features and target cleaning efficiency, the physical information composite network is trained in two stages of loss balance. The first stage uses the mean square error loss function, and the second stage combines the mean square error loss and the physical constraint loss based on the optimal model parameters for iterative optimization to obtain a performance evaluation model.

[0074] In this example, feature mapping is performed on the training set in the preprocessed dataset. By performing nonlinear transformation on the original features, a set of new features with higher representation power is generated, thereby improving the prediction accuracy of the model. After feature mapping, a feature mapping dataset is obtained. The feature mapping dataset is reorganized in time series and divided into four different sequences, namely surfactant concentration sequence, dirt concentration sequence, pH value sequence and temperature sequence. The various physical quantities are arranged according to the time dimension to capture their dynamic change characteristics over time. Based on the surfactant concentration sequence, the first deep neural network in the physical information composite network is constructed. The network adopts the structure of a one-dimensional convolutional neural network (1D-CNN), which includes three one-dimensional convolutional layers, two maximum pooling layers and two fully connected layers. Each convolutional layer extracts local features in the surfactant concentration sequence, processes the sequence by sliding the convolution kernel, and uses the ReLU activation function in the convolution layer:

[0075] ;

[0076] in, is the output after activation, is the input of the convolutional layer. The nonlinear activation function makes the network more expressive and learns the complex patterns in the data. Batch normalization is used to speed up the convergence of the model and improve the generalization ability. Through three layers of convolution and two layers of pooling operations, the network can effectively extract the local features of the surfactant concentration changing over time, map these features to higher dimensions through the fully connected layer, and generate the first output result. The second deep neural network of the physical information composite network is constructed based on the dirt concentration sequence. The network adopts the structure of the long short-term memory network (LSTM). LSTM is a recurrent neural network (RNN) suitable for time series data. It can capture long-term dependencies in the sequence and is suitable for processing tasks with memory requirements. The network contains two layers of LSTM layers, each with 64 hidden units, which means that the input of each time step will be mapped to a 64-dimensional hidden state, thereby capturing the dynamic change characteristics of the dirt concentration in the time series. The LSTM network effectively controls the transmission and retention of information in the sequence through its gating structure, including the input gate, forget gate and output gate, avoiding the problem of gradient disappearance. Finally, through a fully connected layer, the tanh activation function is used:

[0077] ;

[0078] The features extracted by LSTM are nonlinearly mapped to generate the second output result. Based on the pH value sequence, the third deep neural network of the physical information composite network is constructed. The network adopts the structure of bidirectional gated recurrent unit (bidirectional GRU). GRU is a variant of RNN. It has fewer parameters than LSTM and is suitable for relatively simple sequence modeling tasks. The characteristic of bidirectional GRU is that it processes sequence data from both forward and backward directions at the same time, thereby capturing the bidirectional effect of pH value on cleaning efficiency. The network contains two layers of bidirectional GRU, and the number of hidden units in each layer is 32, which is used to effectively extract the temporal dynamic characteristics of pH value, and uses Sigmoid activation function through a fully connected layer:

[0079] ;

[0080] The extracted features are mapped to the range of 0 to 1 to generate the third output result. The fourth deep neural network of the physical information composite network is constructed based on the temperature series. The network adopts the structure of the temporal convolutional network (TCN). TCN is a convolutional architecture suitable for time series modeling. It contains four causal convolutional layers and one fully connected layer. The causal convolution ensures that the output only depends on the current and previous time steps, and does not use future information. Dilated convolution and residual connection are used in the causal convolution layer to increase the receptive field and enhance the transmission ability of information flow. Through the fully connected layer, the LeakyReLU activation function is used:

[0081] ;

[0082] in is a constant less than 1, which is used to retain negative information. Finally, the fourth output result is generated, which represents the characteristics of the impact of the temperature sequence on the cleaning efficiency. The first, second, third and fourth output results are feature fused to obtain a fused feature vector. In order to extract the low-frequency information in the fused feature vector, Fourier transform is used to separate the low-frequency components. The low-frequency components usually represent the overall dynamic trend of the system, while ignoring the high-frequency noise. The formula for Fourier transform is:

[0083] ;

[0084] in, represents the signal in the frequency domain, is a signal in the time domain, Indicates the length of the signal, is the frequency index. Through Fourier transform, the low-frequency components in the fused feature vector are extracted to capture the macroscopic behavior of the system. A variational autoencoder (VAE) is constructed to pre-train the low-frequency features. VAE is a generative model that maps the input features to the latent space through the encoder, reconstructs the input through the decoder, and learns the potential representation of the input data. Through pre-training of low-frequency features, VAE can effectively compress and represent important patterns in the data, improve the training effect of subsequent deep learning models, and obtain pre-trained features. Based on the pre-trained features and target cleaning efficiency, the physical information composite network is trained in two stages of loss balance. In the first stage, the mean square error (MSE) loss function is used for training. The formula of MSE is:

[0085] ;

[0086] in, is the actual target cleaning efficiency, is the cleaning efficiency predicted by the model, is the number of samples. By minimizing the MSE, the prediction of the model is made as close to the actual value as possible, thus improving the accuracy of the model. In the second stage, the mean square error loss and the physical constraint loss based on the optimal model parameters are combined for iterative optimization. The physical constraint loss introduces constraints related to physical laws, so that the model can not only have a good fit effect on the data, but also meet the limitations of the actual physical process, thereby improving the physical consistency and generalization ability of the model. Through two-stage loss balance training, a performance evaluation model that can accurately evaluate the effect of surfactants is obtained.

[0087] In one example, based on the performance evaluation model, multi-granularity information fusion and performance evaluation are performed on the test set in the preprocessed data set to obtain the target performance evaluation results, including:

[0088] Inputting the test set in the preprocessed data set into the first deep neural network of the performance evaluation model to obtain the surfactant concentration distribution prediction result;

[0089] Input the test set in the preprocessed data set into the second deep neural network of the performance evaluation model to obtain the prediction result of the dirt concentration distribution;

[0090] The test set in the preprocessed data set is input into the third deep neural network of the performance evaluation model to obtain the pH value distribution prediction result;

[0091] Input the test set in the preprocessed data set into the fourth deep neural network of the performance evaluation model to obtain the temperature distribution prediction result;

[0092] The prediction results of surfactant concentration distribution, fouling concentration distribution, pH value distribution and temperature distribution are comprehensively analyzed to obtain the target performance evaluation results.

[0093] In this example, the test set in the preprocessed dataset is input into the first deep neural network of the performance evaluation model to obtain the predicted result of the surfactant concentration distribution. The first deep neural network adopts a one-dimensional convolutional neural network (1D-CNN) structure, which can effectively capture the local characteristics of the surfactant concentration in the time series. By inputting the concentration sequence, after a series of convolutional layers, pooling layers, and fully connected layers, the predicted value of the surfactant concentration changing over time is obtained. The test set in the preprocessed dataset is input into the second deep neural network of the performance evaluation model, which is built based on the long short-term memory network (LSTM) structure and is used to predict the dirt concentration distribution. LSTM is a recurrent neural network that is good at processing sequence data with time-dependent characteristics. For the dirt concentration sequence, by inputting the test set, LSTM can remember the changes in concentration in the past time steps and capture the dynamic relationship between dirt concentration and time. The test set is input into the third deep neural network of the performance evaluation model, which is built based on the bidirectional gated recurrent unit (GRU) structure and is used to predict the distribution of pH value. The bidirectional GRU network can simultaneously process information from the past to the present and from the future to the present, and is suitable for modeling the bidirectional influence of pH value in the cleaning process. For the input pH value sequence, the bidirectional GRU generates pH value predictions for each time step. The test set in the preprocessed dataset is input into the fourth deep neural network of the performance evaluation model, which is based on the temporal convolutional network (TCN) structure and is responsible for predicting the distribution of temperature. TCN uses dilated convolution and residual connection to enable the network to model the trend of temperature change over a longer time range and predict the temperature value at each time point. Through this network structure, the temporal variation characteristics of temperature and its impact on the cleaning efficiency of surfactants are captured. The prediction results of surfactant concentration distribution, dirt concentration distribution, pH value distribution and temperature distribution are comprehensively analyzed to obtain the target performance evaluation results. The comprehensive analysis process includes feature fusion and comprehensive weighted calculation. The features of each prediction result are fused, and different features are combined together by splicing, weighted averaging or attention-based methods to form a fused feature vector. The fused feature vector is analyzed in the frequency domain. Fourier transform is used to extract low-frequency information in the fused feature. The Fourier transform formula is:

[0094] ;

[0095] in, is the frequency domain signal, is the time domain fusion feature, is the sample length, is the frequency index. Through Fourier transform, the distribution of fused features in the frequency domain is obtained. The low-frequency component usually represents the overall trend of change, while the high-frequency component is mostly noise. The low-frequency information is processed and the variational autoencoder (VAE) is used to pre-train the low-frequency information. By learning the potential representation of low-frequency features, the variational autoencoder can effectively extract important patterns in the data, thereby improving the generalization ability of the model. The low-frequency features pre-trained by VAE are used to guide the evaluation of target performance to ensure that the model can better comprehensively analyze the overall effect of the surfactant. The final target performance evaluation result is obtained by performing two-stage loss balance training on the fused features and target cleaning efficiency. In the first stage, the model is trained using the mean square error (MSE) loss function to ensure that the model's prediction results are as close to the actual values ​​as possible. The calculation formula for MSE is:

[0096] ;

[0097] in, For the actual target cleaning efficiency, is the cleaning efficiency predicted by the model, is the number of samples. By minimizing the MSE, the prediction accuracy of the model is improved so that it can accurately reflect the changing trend of the cleaning efficiency. In the second stage of training, the MSE loss and the physical constraint loss based on the optimal model parameters are combined for iterative optimization. The physical constraint loss introduces constraints related to physical laws into the loss function, so that the model not only has good prediction performance, but also conforms to the physical characteristics of the cleaning process. For example, there is a certain physical law between the change of temperature and the cleaning efficiency. Through physical constraints, the model conforms to this physical law when predicting the effect of temperature on the cleaning efficiency, thereby improving the physical consistency of the model.

[0098] In one example, based on the target performance evaluation results, parameter sensitivity analysis and optimization solutions are performed on the surfactant formulation and process parameters to obtain the target process optimization solution, including:

[0099] According to the target performance evaluation results, Sobol global sensitivity analysis was performed on the surfactant formula and process parameters, and the first-order sensitivity index and total sensitivity index of each parameter were calculated to obtain the parameter sensitivity ranking;

[0100] Based on the parameter sensitivity ranking, the parameters with sensitivity index greater than 0.05 are selected as key optimization variables, the decision space of the optimization problem is constructed, and the decision space is discretized, the continuous variables are divided into 10 equally spaced discrete levels, and the discretized decision variable set is obtained;

[0101] The population initialization function of the NSGA-III algorithm is constructed based on the discretized decision variable set, and the Latin hypercube sampling algorithm is used to generate the initial population;

[0102] The initial population is input into the performance evaluation model to calculate the fitness values ​​of the individuals in the population, including three objective functions: cleaning efficiency, surfactant utilization, and residue content. The individuals in the population are sorted non-dominatedly and associated with reference points. The crowding distance calculation is used to perform selection, crossover, and mutation operations to obtain the Pareto optimal solution set.

[0103] The Pareto optimal solution set was subjected to fuzzy comprehensive evaluation, and an evaluation index system was constructed, including five indicators: cleaning efficiency, surfactant utilization, residue content, cost, and environmental friendliness, and the weights were set to 0.3, 0.2, 0.2, 0.15, and 0.15 respectively;

[0104] A fuzzy relationship matrix is ​​constructed based on the evaluation index system, and fuzzy comprehensive operation is performed to obtain the comprehensive evaluation value of each solution. The solution with the highest comprehensive evaluation value is selected as the target process optimization solution.

[0105] In this example, a Sobol global sensitivity analysis is performed on the surfactant formulation and process parameters based on the target performance evaluation results. Sobol global sensitivity analysis is used to evaluate the impact of input parameters on output variables. In this analysis, the first-order sensitivity index and the total sensitivity index are calculated for each parameter to determine which parameters have a significant impact on cleaning efficiency and process performance. First-order sensitivity index Representation parameters direct contribution to the output, while the overall sensitivity index Consider the parameters The overall impact of its interaction with other parameters on the output. By calculating the sensitivity index, the parameter sensitivity ranking is obtained, and the variables that have a key impact on system performance are identified. Based on the parameter sensitivity ranking, the parameters with a sensitivity index greater than 0.05 are selected as key optimization variables to construct the decision space of the optimization problem. The decision space contains the value range of all key parameters. Discretize the decision space. Divide each continuous variable into 10 equally spaced discrete levels. For example, suppose a parameter The value range of is [0,1], then it is divided into 10 discrete values , the spacing between each discrete value is 0.1. This reduces the computational complexity while ensuring comprehensive coverage of the parameters, and obtains a discretized set of decision variables. Based on the discretized set of decision variables, the population initialization function of the NSGA-III (Non-dominated Sorting Genetic Algorithm III) algorithm is constructed, and the Latin hypercube sampling (LHS) algorithm is used to generate the initial population. LHS is an efficient sampling method that can ensure that the population is evenly distributed throughout the decision space and maximize the diversity of the initial population. Through Latin hypercube sampling, a set of samples is generated, each of which represents a combination of a set of different parameters, thus forming the initial population. The generated initial population is input into the performance evaluation model to calculate the fitness value of each population individual. The fitness value is calculated based on three objective functions, namely cleaning efficiency, surfactant utilization, and residue content. Assume represent cleaning efficiency, surfactant utilization and residue content respectively, where is the input parameter vector, and the fitness of each individual is the value of these three objective functions. For example, for a particular individual , and its fitness is . These objective functions often conflict with each other. For example, improving the cleaning efficiency will increase the consumption of surfactants, so multi-objective optimization is used to weigh the various objectives. The individuals in the population are non-dominated sorted and associated with reference points to identify which individuals are candidate solutions for the Pareto front. NSGA-III introduces reference points for non-dominated sorting and combines crowding distance calculation to maintain the diversity of the population. In the selection, crossover and mutation operations, individuals with higher fitness and diversity are selected for reproduction, and new individuals are generated through crossover and mutation, so as to conduct a wider search in the solution space. After multiple iterations, a set of Pareto optimal solutions is obtained. The Pareto optimal solution set is fuzzy comprehensively evaluated. An evaluation index system containing five indicators is constructed, including cleaning efficiency, surfactant utilization, residue content, cost and environmental friendliness. In order to reflect the importance of each indicator in process optimization, a corresponding weight is set for each indicator, where the weight of cleaning efficiency is 0.3, the surfactant utilization and residue content are 0.2 respectively, and the cost and environmental friendliness are 0.15 respectively. Assume that the cleaning efficiency is The surfactant utilization rate is , the residue content is , the cost is , environmental friendliness is , then the comprehensive evaluation value of each solution is Calculated by the following weighted formula:

[0106] ;

[0107] in, are the values ​​of each evaluation index respectively. Based on the evaluation index system, a fuzzy relationship matrix is ​​constructed, and fuzzy comprehensive operations are performed to obtain the comprehensive evaluation value of each solution. In the fuzzy comprehensive evaluation process, the index values ​​of each solution are fuzzified, that is, they are converted into a fuzzy membership degree, which indicates the performance of each solution on each index. By constructing a fuzzy relationship matrix, the index values ​​of all solutions are combined together, and then each solution is scored through fuzzy comprehensive operations. For example, let the fuzzy relationship matrix be , whose elements Representation solution In the indicator The membership degree on Perform weighted operations to obtain the comprehensive evaluation value of each solution. By comparing the comprehensive evaluation values ​​of each solution, the solution with the highest comprehensive evaluation value is selected as the target process optimization solution.

[0108] In one example, based on the target performance evaluation results, a Sobol global sensitivity analysis was performed on the surfactant formulation and process parameters, and the first-order sensitivity index and total sensitivity index of each parameter were calculated to obtain the parameter sensitivity ranking, including:

[0109] The sensitivity evaluation function of Sobol sensitivity analysis was constructed based on the target performance evaluation results, and the cleaning efficiency, surfactant utilization rate and residue content were taken as components of the multi-sensitivity evaluation function to obtain a comprehensive sensitivity evaluation function.

[0110] Define the parameter space for surfactant formula and process parameters, set the value range and distribution type of each parameter, and obtain the parameter sampling space;

[0111] Based on the parameter sampling space, the Saltelli sampling method is used to generate the parameter sample matrix, and the sample size is set to (2d+2) times the number of parameters, where d is the number of parameters, to obtain the Sobol analysis sample set;

[0112] According to the Sobol analysis sample set, the sensitivity evaluation function value corresponding to each group of parameter samples is calculated to obtain the sensitivity evaluation function output matrix;

[0113] Based on the output matrix of the sensitivity evaluation function, the first-order sensitivity index of each parameter is calculated. The first-order sensitivity index reflects the direct impact of a single parameter on the sensitivity evaluation function, and a first-order sensitivity index set is obtained.

[0114] Based on the output matrix of the sensitivity evaluation function, the total sensitivity index of each parameter is calculated. The total sensitivity index includes the direct influence and interactive influence of the parameter, and the total sensitivity index set is obtained;

[0115] The first-order sensitivity index set and the total sensitivity index set are normalized and arranged in descending order according to the size of the total sensitivity index to obtain the parameter sensitivity ranking.

[0116] In this example, a sensitivity evaluation function of Sobol sensitivity analysis is constructed based on the target performance evaluation results. This function aims to measure the comprehensive impact of different parameters on the cleaning performance of surfactants. Cleaning efficiency, surfactant utilization and residue content are taken as components of the multi-sensitivity evaluation function. Define a comprehensive sensitivity evaluation function , which combines three important indicators:

[0117] ;

[0118] in, Indicates the cleaning efficiency, represents the surfactant utilization rate, Indicates the residue content, are the weights of each indicator, which are used to measure their relative importance in the overall performance. The comprehensive sensitivity evaluation function can reflect the multi-faceted performance of surfactants in the cleaning process. Define the parameter space for the formulation and process parameters of the surfactant. Set the value range and distribution type of each parameter to construct the sampling space of the parameter. For example, for the surfactant concentration , set its value range to [0.1,1.0]mol / L, and the distribution type to uniform distribution; for temperature , set its value range to [25,75]°C, and also assume a uniform distribution. For other parameters such as stirring speed and pH value, specify their range and distribution type to ensure that the parameter space can fully cover possible process conditions. The definition of parameter space is the basis for Sobol sensitivity analysis, which determines the range of each parameter to analyze the impact of the objective function. Based on the parameter sampling space, the Saltelli sampling method is used to generate a parameter sample matrix. Saltelli sampling is an improved sampling method commonly used in Sobol sensitivity analysis, which can reduce the variance when estimating the sensitivity index and improve sampling efficiency. In this step, assume is the number of parameters, the sampling amount is set to the number of parameters times to ensure that the direct effect of each parameter and its interactive effects with other parameters are fully captured. For example, if there are 4 parameters, the resulting sample size is sample points. Through these samples, a matrix containing all parameter combinations is formed, namely the Sobol analysis sample set. According to the Sobol analysis sample set, the sensitivity evaluation function value corresponding to each group of parameter samples is calculated. Each sample in the sample set is input into the comprehensive sensitivity evaluation function By traversing all samples, we get a sensitivity evaluation function output matrix, which is recorded as ,in , indicating the sample The corresponding evaluation function value. Output matrix It contains the comprehensive performance of the system under different parameter combinations. Based on the sensitivity evaluation function output matrix , then calculate the first-order sensitivity index of each parameter. First-order sensitivity index Represents the direct effect of a single parameter on the output, which is defined as:

[0119] ;

[0120] in, Indicates that given parameters In the case of The variance of the conditional expectation of Represents the total variance of the evaluation function. Calculates the total sensitivity index of each parameter based on the output matrix. Total sensitivity index It includes the direct effect of a parameter on the output and the interactive effect with other parameters, which is defined as:

[0121] ;

[0122] in, Indicates to remove parameters The variance of the conditional expectation of other parameters is Indicates that all parameters except Total sensitivity index Higher means that the parameter The overall impact on the output is also great when interacting with other parameters. The first-order sensitivity index set and the total sensitivity index set are normalized so that the values ​​of the sensitivity indexes can be compared on the same scale to facilitate understanding of their relative importance in the overall system. For example, each sensitivity index is divided by the sum of all sensitivity indexes so that the normalized value is between [0,1]. The parameters are sorted in descending order according to the size of the total sensitivity index to obtain the sensitivity ranking of the parameters.

[0123] Reference Figure 2 This embodiment provides a surfactant effect evaluation and process optimization device based on data analysis, including:

[0124] The acquisition module 1 is used to collect and preprocess the production parameters of the surfactant to obtain a preprocessed data set;

[0125] A calculation module 2 is used to construct four unstructured models based on the preprocessed data set and calculate the target cleaning efficiency of the surfactant;

[0126] Building module 3, used to build a physical information composite network based on the target cleaning efficiency and the training set in the preprocessing data set, and perform training to obtain a performance evaluation model;

[0127] Evaluation module 4 is used to perform multi-granularity information fusion and performance evaluation on the test set in the preprocessed data set based on the performance evaluation model to obtain the target performance evaluation result;

[0128] The solution module 5 is used to perform parameter sensitivity analysis and optimization solution on the surfactant formula and process parameters according to the target performance evaluation results to obtain the target process optimization solution.

[0129] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0130] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0131] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0132] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0133] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0134] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0135] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A surfactant effect evaluation and process optimization method based on data analysis, characterized in that: The following steps are involved: Performing data collection and preprocessing on production parameters of surfactants to obtain a preprocessed data set; Four unstructured models are constructed based on the preprocessed data set, and the target cleaning efficiency of the surfactant is calculated; specifically, the following steps are performed: feature selection is performed on the preprocessed data set, and features associated with the cleaning efficiency and utilization rate of the surfactant are selected to obtain a data set after feature selection; a linear model is constructed based on the data set after feature selection, wherein the cleaning efficiency E is expressed as E=k*C+b, C is the surfactant concentration, k is the slope, b is the intercept, and the initial k and b values ​​are estimated using the least squares method; a Langmuir model is constructed based on the data set after feature selection, wherein the cleaning efficiency E is expressed as E=(Emax*K*C) / (1+K*C), Emax is the maximum cleaning efficiency, K is the Langmuir constant, and the initial Emax and K values ​​are estimated using the nonlinear least squares method; a Freundlich model is constructed based on the data set after feature selection, wherein the cleaning efficiency E is expressed as E=F*C (1 / n) , F is the Freundlich constant, n is the heterogeneity factor, and the initial F and n values ​​are estimated using log-linear regression; based on the data set after the feature selection, an exponential model is constructed, wherein the cleaning efficiency E is expressed as E=a*(1-exp(-s*C)), a is the maximum cleaning efficiency, s is the rate constant, and the initial a and s values ​​are estimated using the Levenberg-Marquardt algorithm; the optimal model parameters of the linear model, the Langmuir model, the Freundlich model and the exponential model are estimated to obtain four unstructured models, and the corresponding initial cleaning efficiencies are calculated respectively through the four unstructured models, and the initial cleaning efficiencies are comprehensively weighted calculated to obtain the target cleaning efficiency of the surfactant; Building a physical information composite network based on the target cleaning efficiency and a training set in the preprocessed data set, and performing training to obtain a performance evaluation model; Based on the performance evaluation model, multi-granularity information fusion and performance evaluation are performed on the test set in the preprocessed data set to obtain a target performance evaluation result; According to the target performance evaluation results, parameter sensitivity analysis and optimization solution are performed on the surfactant formula and process parameters to obtain the target process optimization plan.

2. The surfactant effect evaluation and process optimization method based on data analysis according to claim 1, characterized in that: The data collection and preprocessing of the production parameters of the surfactant to obtain a preprocessed data set includes: The surfactant formula composition, production process parameters, surfactant consumption and cleaning residue data were collected from multiple sources to obtain the original data set; Perform outlier detection on the original data set, calculate upper and lower thresholds using the interquartile range method, compare data points in the original data set with the upper and lower thresholds, mark data points that exceed the range as outliers, and obtain an outlier marked data set; Based on the outlier labeled data set, perform median replacement processing on the labeled outliers, calculate the median of the column where the outliers are located, replace the outliers with the median of the column where the outliers are located, and obtain a data set after outlier processing; Filling the missing values ​​in the data set after outlier processing, performing linear interpolation on the time series data, and performing K-nearest neighbor interpolation on the non-time series data to obtain a complete data set; Performing data standardization processing on the complete data set to obtain a standardized data set, and arranging the standardized data set in chronological order, and using a fixed-ratio stratified sampling method to divide the standardized data set into a training set, a validation set, and a test set, with the ratios of 70%, 15%, and 15%, respectively, to obtain a divided data set; Performing data enhancement on the training set in the divided data set, using sliding window technology for time series data, and using SMOTE algorithm for non-time series data to generate synthetic samples, to obtain an enhanced training set; The enhanced training set, the validation set and the test set are merged, and batch identifier and timestamp information are added to form a preprocessed data set.

3. The surfactant effect evaluation and process optimization method based on data analysis according to claim 1, characterized in that: The physical information composite network is constructed based on the target cleaning efficiency and the training set in the preprocessed data set, and is trained to obtain a performance evaluation model, including: Performing feature mapping on the training set in the preprocessed data set to obtain a feature mapping data set, and performing time series reorganization on the feature mapping data set to divide the feature mapping data set into a surfactant concentration sequence, a dirt concentration sequence, a pH value sequence, and a temperature sequence; Constructing a first deep neural network of a physical information composite network based on the surfactant concentration sequence, and generating a first output result, wherein the first deep neural network adopts a one-dimensional convolutional neural network structure, including 3 one-dimensional convolutional layers, 2 maximum pooling layers and 2 fully connected layers, and each convolutional layer uses a ReLU activation function and batch normalization; Constructing a second deep neural network of the physical-information composite network based on the dirt concentration sequence, and generating a second output result, wherein the second deep neural network adopts an LSTM structure, including 2 LSTM layers and 1 fully connected layer, the number of hidden units in the LSTM layer is 64, and the fully connected layer uses a tanh activation function; Constructing a third deep neural network of the physical-information composite network based on the pH value sequence, and generating a third output result, wherein the third deep neural network adopts a bidirectional GRU structure, including 2 bidirectional GRU layers and 1 fully connected layer, the number of hidden units in the GRU layer is 32, and the fully connected layer uses a Sigmoid activation function; A fourth deep neural network of the physical information composite network is constructed based on the temperature sequence, and a fourth output result is generated, wherein the fourth deep neural network adopts a TCN structure, includes four causal convolutional layers and one fully connected layer, the causal convolutional layer uses dilated convolution and residual connection, and the fully connected layer uses a LeakyReLU activation function; Performing feature fusion on the first output result, the second output result, the third output result, and the fourth output result to obtain a fused feature vector, extracting low-frequency information from the fused feature vector, extracting low-frequency components using Fourier transform, constructing a variational autoencoder to perform low-frequency information pre-training, and obtaining pre-trained features; Based on the pre-training features and the target cleaning efficiency, the physical information composite network is subjected to two-stage loss balance training. The first stage uses the mean square error loss function, and the second stage combines the mean square error loss and the physical constraint loss based on the optimal model parameters for iterative optimization to obtain a performance evaluation model.

4. The surfactant effect evaluation and process optimization method based on data analysis according to claim 3, characterized in that: Based on the performance evaluation model, multi-granularity information fusion and performance evaluation are performed on the test set in the preprocessed data set to obtain a target performance evaluation result, including: Inputting the test set in the preprocessed data set into the first deep neural network of the performance evaluation model to obtain a surfactant concentration distribution prediction result; Inputting the test set in the preprocessed data set into the second deep neural network of the performance evaluation model to obtain a prediction result of the dirt concentration distribution; Inputting the test set in the preprocessed data set into the third deep neural network of the performance evaluation model to obtain a pH value distribution prediction result; Inputting the test set in the preprocessed data set into the fourth deep neural network of the performance evaluation model to obtain a temperature distribution prediction result; The surfactant concentration distribution prediction result, the scale concentration distribution prediction result, the pH value distribution prediction result and the temperature distribution prediction result are comprehensively analyzed to obtain a target performance evaluation result.

5. The surfactant effect evaluation and process optimization method based on data analysis according to claim 4, characterized in that: According to the target performance evaluation results, the surfactant formula and process parameters are subjected to parameter sensitivity analysis and optimization solution to obtain the target process optimization solution, including: According to the target performance evaluation results, a Sobol global sensitivity analysis is performed on the surfactant formula and process parameters, and the first-order sensitivity index and the total sensitivity index of each parameter are calculated to obtain the parameter sensitivity ranking; Based on the parameter sensitivity ranking, parameters with a sensitivity index greater than 0.05 are selected as key optimization variables, a decision space for the optimization problem is constructed, and the decision space is discretized to divide the continuous variables into 10 equally spaced discrete levels to obtain a discretized decision variable set; Based on the discretized decision variable set, a population initialization function of the NSGA-III algorithm is constructed, and an initial population is generated using a Latin hypercube sampling algorithm; Inputting the initial population into the performance evaluation model, calculating the fitness values ​​of the individuals in the population, including three objective functions of cleaning efficiency, surfactant utilization rate and residue content, performing non-dominated sorting and reference point association on the individuals in the population, and performing selection, crossover and mutation operations using crowding distance calculation to obtain a Pareto optimal solution set; A fuzzy comprehensive evaluation is performed on the Pareto optimal solution set to construct an evaluation index system, including five indicators: cleaning efficiency, surfactant utilization, residue content, cost, and environmental friendliness, and the weights are set to 0.3, 0.2, 0.2, 0.15, and 0.15 respectively; A fuzzy relationship matrix is ​​constructed based on the evaluation index system, and fuzzy comprehensive operation is performed to obtain a comprehensive evaluation value of each solution, and the solution with the highest comprehensive evaluation value is selected as the target process optimization solution.

6. The surfactant effect evaluation and process optimization method based on data analysis according to claim 5, characterized in that: According to the target performance evaluation results, the Sobol global sensitivity analysis is performed on the surfactant formula and process parameters, and the first-order sensitivity index and the total sensitivity index of each parameter are calculated to obtain the parameter sensitivity ranking, including: A sensitivity evaluation function of Sobol sensitivity analysis is constructed according to the target performance evaluation result, and cleaning efficiency, surfactant utilization rate and residue content are used as components of the multi-sensitivity evaluation function to obtain a comprehensive sensitivity evaluation function; Define the parameter space for surfactant formula and process parameters, set the value range and distribution type of each parameter, and obtain the parameter sampling space; Based on the parameter sampling space, the parameter sample matrix is ​​generated using the Saltelli sampling method, and the sample size is set to (2d+2) times the number of parameters, where d is the number of parameters, to obtain a Sobol analysis sample set; According to the Sobol analysis sample set, the sensitivity evaluation function value corresponding to each group of parameter samples is calculated to obtain a sensitivity evaluation function output matrix; Based on the sensitivity evaluation function output matrix, the first-order sensitivity index of each parameter is calculated, and the first-order sensitivity index reflects the direct influence of a single parameter on the sensitivity evaluation function, so as to obtain a first-order sensitivity index set; Based on the sensitivity evaluation function output matrix, the total sensitivity index of each parameter is calculated, and the total sensitivity index includes the direct influence and interactive influence of the parameter to obtain a total sensitivity index set; The first-order sensitivity index set and the total sensitivity index set are normalized, and arranged in descending order according to the size of the total sensitivity index to obtain a parameter sensitivity ranking.

7. A surfactant effect evaluation and process optimization device based on data analysis, characterized in that: For implementing the steps of the method according to any one of claims 1 to 6, the device comprises: A collection module is used to collect and preprocess the production parameters of the surfactant to obtain a preprocessed data set; A calculation module, for constructing four unstructured models based on the preprocessed data set and calculating the target cleaning efficiency of the surfactant; A construction module is used to construct a physical information composite network based on the target cleaning efficiency and the training set in the preprocessing data set, and perform training to obtain a performance evaluation model; An evaluation module, used to perform multi-granularity information fusion and performance evaluation on the test set in the preprocessed data set based on the performance evaluation model to obtain a target performance evaluation result; The solution module is used to perform parameter sensitivity analysis and optimization solution on the surfactant formula and process parameters according to the target performance evaluation results to obtain the target process optimization plan.

8. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images