Surfactant process parameter analysis method and system based on multilevel factors

Abstract

The invention relates to the technical field of process parameter optimization, and discloses a surfactant process parameter analysis method and system based on multi-level factors, and the method comprises the steps: carrying out multi-dimensional process parameter collection and preprocessing on production equipment of a surfactant, and obtaining multi-dimensional process parameter time sequence data; thermodynamic feature extraction is carried out to obtain an equipment temperature distribution feature vector; calculating an equipment performance degradation index; performing process parameter association analysis to obtain a process parameter association probability matrix; executing a double-delay deterministic strategy gradient algorithm to obtain a parameter hierarchical control strategy; according to the technical scheme of the invention, the comprehensive improvement of the product quality, the production efficiency and the equipment reliability is realized through the integration of the equipment state monitoring, the process parameter optimization and the intelligent control strategy. And powerful support is provided for intelligent and fine management of the surfactant production process.

Description

Technical Field

[0001] The present invention relates to the technical field of process parameter optimization, and in particular to a surfactant process parameter analysis method and system based on multi-level factors. Background Art

[0002] The production process of surfactants involves complex chemical reactions and variable process parameters. Traditional production control methods often find it difficult to cope with this complexity, leading to problems such as low production efficiency and unstable product quality.

[0003] How to effectively integrate multi-source heterogeneous data, accurately capture the complex relationship between process parameters, and how to achieve real-time optimization control in a dynamically changing production environment are still technical challenges that need to be solved. The performance degradation problem of surfactant production equipment is also becoming increasingly prominent. The decline in equipment performance not only affects product quality, but may also lead to energy waste and safety hazards. How to monitor equipment status in real time during the production process, predict performance degradation trends, and incorporate this information into process parameter optimization is the key to improving the overall efficiency of the production system. Summary of the invention

[0004] The present invention provides a surfactant process parameter analysis method and system based on multi-level factors. The present invention achieves comprehensive improvement in product quality, production efficiency and equipment reliability by integrating equipment status monitoring, process parameter optimization and intelligent control strategy, and provides strong support for the intelligent and refined management of the surfactant production process.

[0005] In a first aspect, the present invention provides a surfactant process parameter analysis method based on multi-level factors, the surfactant process parameter analysis method based on multi-level factors comprising: Collect and preprocess multi-dimensional process parameters of surfactant production equipment to obtain multi-dimensional process parameter time series data; Performing thermodynamic feature extraction and dimensionality reduction processing on the multi-dimensional process parameter time series data to obtain a device temperature distribution feature vector; Calculating multi-scale sample entropy based on the multi-dimensional process parameter time series data, and performing nonlinear regression analysis to obtain equipment performance degradation indicators; Inputting the equipment temperature distribution feature vector and the equipment performance degradation index into a Bayesian network model to perform process parameter association analysis to obtain a process parameter association probability matrix; A Markov decision process model is constructed according to the process parameter association probability matrix, and a double-delayed deterministic policy gradient algorithm is executed to obtain a parameter hierarchical control strategy; The process parameters of the production equipment are optimized based on the parameter hierarchical control strategy, and the control parameter instruction sequence of the production equipment is continuously updated through a closed-loop feedback mechanism.

[0006] In a second aspect, the present invention provides a surfactant process parameter analysis system based on multi-level factors, the surfactant process parameter analysis system based on multi-level factors comprising: The acquisition module is used to collect and preprocess multi-dimensional process parameters of the surfactant production equipment to obtain multi-dimensional process parameter time series data; An extraction module, used for performing thermodynamic feature extraction and dimensionality reduction processing on the multi-dimensional process parameter time series data to obtain a characteristic vector of equipment temperature distribution; A calculation module, used to calculate multi-scale sample entropy based on the multi-dimensional process parameter time series data, and perform nonlinear regression analysis to obtain equipment performance degradation indicators; An analysis module, used for inputting the equipment temperature distribution feature vector and the equipment performance degradation index into a Bayesian network model to perform process parameter association analysis to obtain a process parameter association probability matrix; A construction module is used to construct a Markov decision process model according to the process parameter association probability matrix, and execute a double-delayed deterministic policy gradient algorithm to obtain a parameter hierarchical control strategy; An updating module is used to optimize the process parameters of the production equipment based on the parameter hierarchical control strategy, and continuously update the control parameter instruction sequence of the production equipment through a closed-loop feedback mechanism.

[0007] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned surfactant process parameter analysis method based on multi-level factors.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned surfactant process parameter analysis method based on multi-level factors.

[0009] In the technical solution provided by the present invention, the present invention realizes the effective fusion of multi-dimensional process parameters such as temperature, pressure, flow, concentration, etc. by collecting and preprocessing multi-type sensor data, adopts short-time Fourier transform and Welch power spectral density analysis, combined with a three-dimensional finite element heat transfer model, realizes high-precision feature extraction of equipment temperature distribution, and constructs equipment performance degradation indicators through multi-scale sample entropy analysis and nonlinear regression, which can capture equipment state changes from multiple time scales, and improve the accuracy and comprehensiveness of performance evaluation. Using the Bayesian network model, the probabilistic association analysis between environmental factors, process parameters and performance indicators is realized, revealing the complex interactive relationship between parameters, and providing a reliable basis for optimization decision-making. Based on the Markov decision process and the double-delay deterministic policy gradient algorithm, an adaptive parameter hierarchical control strategy is constructed, which can effectively cope with the uncertainty and dynamic changes in the production process. Through the hierarchical optimization controller and the adaptive learning rate update mechanism, the continuous optimization of process parameters and the dynamic adjustment of the control strategy are realized, ensuring the stability and efficiency of the production process. By integrating equipment status monitoring, process parameter optimization and intelligent control strategies, the present invention achieves a comprehensive improvement in product quality, production efficiency and equipment reliability, and provides strong support for the intelligent and refined management of the surfactant production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 Schematic diagram of the steps of the surfactant process parameter analysis method based on multi-level factors in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a surfactant process parameter analysis system based on multi-level factors in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] Embodiments of the present invention provide a surfactant process parameter analysis method and system based on multi-level factors. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the surfactant process parameter analysis method based on multi-level factors in the embodiment of the present invention includes: Step S1, collecting and preprocessing multidimensional process parameters of surfactant production equipment to obtain multidimensional process parameter time series data; It is understandable that the execution subject of the present invention may be a surfactant process parameter analysis system based on multi-level factors, or a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0014] Specifically, multiple types of sensors, including temperature sensors, pressure sensors, flow sensors, and concentration sensors, are installed at key locations of surfactant production equipment. Parameters such as temperature, pressure, flow, and concentration are collected in real time to obtain the original multi-dimensional process parameter data stream. During the collection process, all parameters are aligned and synchronized with the timestamps to ensure that different types of process parameters can be analyzed for correlation on the same time scale, and a time-synchronized multi-dimensional process parameter data set is obtained. The data quality of the time-synchronized multi-dimensional process parameter data set is evaluated to ensure that the collected data can accurately reflect the operating status of the equipment. Based on the preset data quality evaluation indicators, the outliers in the data set are detected, and the outliers in the data are identified using statistical analysis methods or machine learning algorithms. After the outliers are detected, the abnormal data are marked to facilitate subsequent processing. For the data marked as abnormal, they are corrected by interpolation and data smoothing methods. The interpolation method is used to replace lost or invalid data points, while data smoothing is used to reduce random noise in the data, making the data more stable and continuous, reflecting the actual change trend of the process. After interpolation and smoothing, a corrected multi-dimensional process parameter data set is obtained. The corrected multidimensional process parameter data set is subjected to data standardization. By scaling all parameters to the same scale, the influence of each parameter can be evaluated at the same level, and a standardized multidimensional process parameter data set is obtained. The standardized data can better adapt to the subsequent analysis and modeling process. The standardized multidimensional process parameter data set is segmented using a sliding window algorithm. By setting a fixed-length time window, the time series data is divided into continuous segments. The length of the sliding window is selected according to the dynamic characteristics of the process to ensure that important information during the operation of the equipment can be captured. After the sliding window processing, a fixed-length data segment sequence is obtained, which reflects the dynamic changes of the process parameters in different time periods. Fast Fourier transform is performed on the fixed-length data segment sequence to extract the frequency domain characteristics of the data from the time domain. Fast Fourier transform is an efficient signal processing method that converts time domain signals into frequency domain to reveal the periodicity and frequency components in the process parameters. By performing fast Fourier transform on multidimensional process parameters, the frequency spectrum feature matrix of multidimensional process parameters is obtained, which reflects the periodic patterns, oscillation characteristics and other information in the process. The spectrum feature matrix and the standardized multidimensional process parameter data set are spliced ​​in time series to obtain multidimensional process parameter time series data containing rich information.

[0015] Step S2, performing thermodynamic feature extraction and dimensionality reduction processing on the multi-dimensional process parameter time series data to obtain a characteristic vector of equipment temperature distribution; Specifically, the temperature-related parameters in the multidimensional process parameter time series data are short-time Fourier transformed, and the time series data are converted from the time domain to the time-frequency domain to obtain the spectrum information of the temperature parameters in different time periods. Through the transformation, the time-frequency domain distribution characteristics of the temperature parameters are obtained, providing a dual expression of temperature in time and frequency, revealing the spectrum characteristics of the temperature signal changing over time. Based on the time-frequency domain characteristics, the Welch power spectral density is calculated, and the stability and accuracy of the spectrum estimation are improved by segmenting, windowing and averaging the data, and the frequency characteristic vector of the temperature fluctuation is obtained, which reflects the main frequency components and energy distribution of the temperature signal in the process, and helps to identify the vibration characteristics and thermodynamic behavior of the system. The adaptive sliding window algorithm is applied to the multidimensional process parameter time series data to calculate the temperature gradient and heat flux density. The adaptive sliding window algorithm adaptively adjusts the window size according to the data changes in order to more accurately capture the instantaneous change characteristics of temperature. Through this process, the local heat conduction characteristic matrix is ​​obtained, which describes the spatial distribution of temperature gradients and the resulting heat flux density. The local heat conduction characteristics are of great significance for understanding how temperature is transferred and diffused in production equipment, especially when the heat conduction characteristics of different parts of the equipment are uneven. The local heat conduction characteristic matrix can provide detailed information on heat transfer within the equipment. A three-dimensional finite element heat transfer model is constructed based on the local heat conduction characteristic matrix. The finite element method is a numerical analysis method that can describe the heat transfer process in a physical system. Through this model, the temperature distribution in the three-dimensional space inside the equipment is simulated. In order to solve the steady-state heat conduction equation, the conjugate gradient method is used. This numerical solution algorithm can efficiently handle large-scale sparse matrix problems and obtain high-precision temperature field distribution data. The high-precision temperature field distribution data can reflect the temperature conditions of each position of the equipment under steady-state conditions, which is convenient for subsequent thermodynamic analysis. Perform kernel principal component analysis on the temperature field distribution data. By introducing the kernel function, the original data is mapped to a high-dimensional space. The principal component analysis is performed on the data in the high-dimensional space to obtain the thermodynamic feature matrix after dimensionality reduction. The thermodynamic feature matrix after dimensionality reduction is input into a deep stacked autoencoder neural network. The feature matrix is ​​deep-featured extracted through unsupervised learning. The autoencoder neural network can learn the implicit structure of the input data through training and compress it into a more compact representation. This compressed representation can capture the deep-level features of the data and help the subsequent analysis and modeling process. Perform nonlinear dimensionality reduction on the compressed representation to obtain a two-dimensional temperature distribution feature map. This nonlinear dimensionality reduction method can project high-dimensional features into a two-dimensional space by retaining the nonlinear structural characteristics of the data. The two-dimensional temperature distribution feature map helps to visualize the temperature distribution inside the device.Based on the two-dimensional temperature distribution feature mapping, the adaptive kernel density estimation method is used to extract the probability characteristics of temperature distribution. The adaptive kernel density estimation can apply different smoothing parameters in different data distribution areas, so as to more accurately capture the probability characteristics of temperature distribution. Through this method, the probability characteristics of temperature distribution are obtained, which reflects the distribution of temperature at different locations inside the equipment, including the formation of hot spots, the range of temperature fluctuations, etc., and finally forms the equipment temperature distribution feature vector.

[0016] Step S3, calculating multi-scale sample entropy based on multi-dimensional process parameter time series data, and performing nonlinear regression analysis to obtain equipment performance degradation indicators; Specifically, the multi-dimensional process parameter time series data is coarse-grained. In the coarse-grained processing, the scale factor τ is set to a range of 1 to 20, and a corresponding coarse-grained time series is generated for each scale factor τ to obtain a multi-scale time series set. The original time series data is smoothed according to different scales to reduce the impact of high-frequency noise and better analyze the dynamic characteristics of the data at different time scales. The multi-scale time series set contains time series at different granularities, so that the subsequent sample entropy calculation can more comprehensively reflect the complexity of the system. Based on the multi-scale time series set, the sample entropy calculation parameters are set, where the embedding dimension m is 2 and the similarity tolerance r is 0.15 times the standard deviation of the original sequence. This parameter setting can effectively calculate the sample entropy of the time series to measure the complexity and uncertainty of the sequence. By calculating the sample entropy value at each scale, a 20-dimensional multi-scale sample entropy feature vector is obtained, which reflects the complexity of the dynamic changes of process parameters at different time scales. The higher the sample entropy value, the greater the complexity of the time series. The 20-dimensional multi-scale sample entropy feature vector is normalized, and the maximum and minimum normalization method is used to map the entropy values ​​of each scale to the interval [0,1] to obtain the normalized multi-scale sample entropy feature vector. Based on the normalized multi-scale sample entropy feature vector, the slope of the curve of entropy value changing with scale is calculated, and the relationship between entropy value and scale is linearly fitted by the least square method to obtain the entropy value change rate index. The entropy value change rate reflects the trend of system complexity changing with the increase of time scale. If the entropy value change rate is large, it means that there is a significant change in the operating status of the equipment. The normalized multi-scale sample entropy feature vector and the entropy value change rate index are spliced ​​to construct a comprehensive entropy feature vector. Based on the comprehensive entropy feature vector, a radial basis function (RBF) kernel matrix is ​​constructed, in which the kernel function uses a Gaussian kernel and the kernel parameter γ is set to 0.1. The role of the radial basis function kernel is to map the original feature vector to a high-dimensional space, so that the nonlinear relationship between features can be more easily revealed in the high-dimensional space. Through mapping, the feature space mapping matrix is ​​obtained. The feature space mapping matrix is ​​input into the support vector regression model for performance degradation analysis. The parameter settings of the support vector regression model include a penalty factor C of 10 and an ε-insensitive loss function parameter ε of 0.01. The role of the penalty factor C is to control the tolerance of the model to errors and avoid overfitting, while the ε parameter is used to define the range of the loss function so that errors less than ε do not affect the optimization target of the model. The support vector regression model obtains the initial performance degradation prediction value of the equipment by learning the relationship between the comprehensive entropy feature vector and the equipment performance. In order to improve the reliability of the prediction results, the confidence interval of the initial performance degradation prediction value is estimated to obtain the upper and lower limits of the performance degradation prediction. The confidence interval estimation can provide the credibility of the prediction results, making the judgment of the equipment performance status more robust. Combined with the prediction variance of the model, the weighted average method is used to calculate the final equipment performance degradation index.The weighted average method assigns different weights to different prediction values ​​by considering the variance of the prediction results to ensure that the performance degradation index finally calculated is more accurate and reliable.

[0017] Step S4, inputting the equipment temperature distribution feature vector and the equipment performance degradation index into the Bayesian network model to perform process parameter association analysis to obtain a process parameter association probability matrix; Specifically, the equipment temperature distribution feature vector and the equipment performance degradation index are combined into the input feature vector. The comprehensive feature vector is input into the Bayesian network model for association analysis, wherein the Bayesian network model consists of three layers, namely, the environmental factor layer, the process parameter layer and the performance index layer. The environmental factor layer includes three nodes: temperature, humidity and raw material batch. These factors are external conditions in the production process and will have a direct impact on the process; the process parameter layer includes four nodes: reaction temperature, stirring speed, pH value and reaction time. These parameters are process conditions controlled and adjusted in the production process; the performance index layer includes three nodes: product purity, surface activity and stability. These nodes reflect the quality and characteristics of the final product and are important indicators for evaluating the success of the production process. Based on the Bayesian network model, the conditional probability distribution from the environmental factor layer to the process parameter layer is calculated by using the variational inference algorithm to obtain the environment-process parameter association matrix. Variational inference is a Bayesian inference method that approximates the conditional probability distribution by converting complex probability distribution problems into optimization problems, thereby improving computational efficiency. The environment-process parameter association matrix is ​​sampled by parameter posterior distribution to obtain the marginal posterior distribution of the process parameters. By sampling the posterior distribution, the probability distribution of process parameters under different environmental conditions can be estimated more accurately. Based on the marginal posterior distribution of process parameters, the conditional probability from the process parameter layer to the performance indicator layer is calculated to obtain the probability transfer matrix of process parameters-performance indicators. The probability transfer matrix reflects the degree of influence of process parameters on product performance indicators. The environment-process parameter association matrix is ​​combined with the process parameter-performance indicator probability transfer matrix, and the chain rule is used to calculate the joint probability distribution between the three layers to obtain the global parameter association probability table. The chain rule is a basic method used in probability theory to calculate joint probability. By multiplying the conditional probabilities in sequence, the joint probability distribution between multiple variables is calculated. The global parameter association probability table reflects the overall probability relationship from environmental factors to process parameters to performance indicators. The information gain ratio method is applied to the global parameter association probability table. The information gain ratio is an indicator used for feature selection to measure the degree of influence of each process parameter on the performance indicator. The threshold is set to 0.1, and the process parameters with higher information gain are screened out to obtain the process parameter association network. Based on the process parameter association network, the maximum likelihood estimation method is used to calculate the conditional independence between parameters and construct the optimal dependency structure. Maximum likelihood estimation is a statistical method that estimates the relationship between parameters in the model by maximizing the likelihood function of the observed data. This method determines which process parameters are independent and which are dependent, and constructs the optimal dependency structure. Based on the optimal dependency structure, the process parameter association probability matrix is ​​obtained, which reflects the degree of association between the process parameters.

[0018] Step S5, constructing a Markov decision process model according to the process parameter association probability matrix, and executing a double-delayed deterministic policy gradient algorithm to obtain a parameter hierarchical control strategy; Specifically, the state space of the Markov decision process is constructed based on the process parameter association probability matrix. Each process parameter is discretized into 5 levels to form a multidimensional discrete state space. The discretization process helps to convert continuous process parameters into a finite state set that is easy to analyze by the Markov decision process. At the same time, the action space of the Markov decision process is constructed, and the adjustment action of each process parameter is defined as three options: increase, decrease and maintain, to obtain a multidimensional discrete action space. The discrete action space provides a clear adjustment path for each process parameter, making the transition between states more operational. The state transition probability matrix of the Markov decision process is calculated based on historical production data. The maximum likelihood estimation method is used to count the frequency of state transitions. The maximum likelihood estimation can effectively estimate the probability of transitioning from one state to another and obtain the state transition function. The state transition function describes the probability that the system transitions from one state to another after performing a certain action. The reward function of the Markov decision process is designed to evaluate the pros and cons of each state and action. When designing the reward function, the product quality index, energy consumption index and equipment performance degradation index are comprehensively considered, and different indicators are integrated into a scalar reward signal through the weight coefficient method. The current state of the process parameters is input into the double-delay deterministic policy gradient algorithm. The double-delay deterministic policy gradient algorithm is an effective method for deep reinforcement learning, which uses two Q networks and a deterministic policy network. Both Q networks are four-layer fully connected neural network structures, each hidden layer contains 256 neurons, and the output layer is 1, which is used to evaluate the value of the current state-action pair. The deterministic policy network uses a three-layer fully connected neural network structure, each hidden layer contains 128 neurons, which is used to generate deterministic actions. The output of the deterministic policy network is used to guide the adjustment of process parameters to ensure that the optimal control strategy can be found in a complex state space. The deterministic actions generated by the deterministic policy network need to be properly explored, so Gaussian noise is added to the generated action a. By adding Gaussian noise, the action space is explored to prevent falling into the local optimum and improve the generalization ability of the strategy. The actual execution action generated is used to adjust the process parameters, and the feedback of the environment is observed according to the adjusted process parameters, including the new system state and immediate rewards. These feedback information are used for updates in the reinforcement learning process. After each step is executed, the state transition sample consisting of the new state, the executed action, the immediate reward, and the new state is input into the experience replay buffer, and the priority of the sample is updated. The use of the experience replay buffer effectively breaks the time correlation of the sample and improves the efficiency and stability of learning. Samples are extracted from the experience replay buffer, and two Q networks are used to calculate the Q value estimate of the current state-action pair. The Q value is used to evaluate the long-term reward of performing an action in the current state.Based on the current Q-value estimate, the target Q-value is generated. The calculation of the target Q-value is a key step in reinforcement learning, which is used to guide the learning direction of the Q network and ensure that the model can gradually approach the optimal strategy. The two Q networks are updated using the time difference error. By minimizing the mean square error between the predicted value of the Q network and the target Q value, the Q network can better fit the value of the state-action pair. At the same time, the policy network is updated by the deterministic policy gradient. The goal of the deterministic policy gradient is to maximize the evaluation value of the current policy action by the Q1 network. The policy network continuously adjusts its output according to the evaluation results of the Q1 network, so that the final generated strategy can maximize the long-term benefits. According to the evaluation results of the Q1 network, the optimal strategy is selected as the parameter hierarchical control strategy to obtain a hierarchical control scheme that effectively optimizes the process parameters.

[0019] Step S6: Optimize the process parameters of the production equipment based on the parameter hierarchical control strategy, and continuously update the control parameter instruction sequence of the production equipment through a closed-loop feedback mechanism.

[0020] Specifically, the parameter hierarchical control strategy is input into the hierarchical optimization controller. The role of the hierarchical optimization controller is to reasonably hierarchically divide the process parameters of the production equipment to achieve efficient control and coordination of parameters at different levels. In the hierarchical control strategy, the process parameters are divided into a high-level control parameter set and a low-level control parameter set, where the high-level control parameter set includes reaction temperature and pH value, while the low-level control parameter set includes stirring speed and reaction time. The process parameters are optimized independently at different time scales and control priorities to improve the control effect of the entire process. Based on the high-level control parameter set, a model predictive controller is constructed to optimize the high-level parameters. In this process, the nonlinear autoregressive exogenous (NARX) model is used as a prediction model to describe the dynamic characteristics of the process. The NARX model can capture the nonlinear behavior of the system and provide accurate prediction capabilities for the model predictive controller. The prediction period of the model predictive controller is set to 10 and the control period is set to 2. Through this setting, the model predictive controller can make real-time adjustments based on a longer prediction time and a shorter control time to obtain the optimization instruction sequence of the high-level parameters. The high-level parameter optimization instruction sequence reflects the ideal adjustment method of reaction temperature and pH value in the future time period in order to achieve the process goals required in the production process. Based on the optimization instruction sequence of high-level parameters, the low-level control parameter set is constrained and optimized. The optimization of the low-level control parameter set needs to take into account the influence of high-level parameters, and the sequential quadratic programming (SQP) algorithm is used to solve the constrained optimization problem. The sequential quadratic programming algorithm is an effective nonlinear optimization method that can gradually approach the optimal solution by solving quadratic approximation problems in each iteration to obtain the optimization instruction sequence of low-level parameters. The high-level parameter optimization instruction sequence is combined with the low-level parameter optimization instruction sequence to construct a complete process parameter optimization instruction sequence. The instruction sequence contains the optimal adjustment scheme of all process parameters in different time periods, and the optimization instruction sequence is sent to the actuator of the production equipment. After receiving the optimization instruction, the actuator adjusts the process parameters of the production equipment accordingly according to the instruction to achieve the expected production goals. During the execution process, the performance deviation of the actuator is calculated. The performance deviation refers to the difference between the equipment state and the target state in the actual production process. The effectiveness of the current control strategy is evaluated by calculating the performance deviation. In order to optimize the control strategy and improve the stability and accuracy of production, an adaptive learning rate update mechanism is constructed based on performance deviation, through which the parameter hierarchical control strategy is fine-tuned. The purpose of the adaptive learning rate update mechanism is to adjust the learning rate according to the current performance feedback, so that a larger adjustment is made when the deviation is large, and a fine adjustment is made when the deviation is small, so as to gradually improve the control effect. After fine-tuning, the updated control strategy is obtained. The updated control strategy is input into the hierarchical optimization controller, and the control parameter instruction sequence of the production equipment is continuously updated through the closed-loop feedback optimization mechanism.In the closed-loop feedback process, the hierarchical optimization controller continuously adjusts according to the equipment status and feedback signals until the system performance deviation is reduced to meet the preset optimization termination conditions. The application of the closed-loop feedback mechanism can ensure that the control strategy is continuously optimized and adjusted according to the actual situation during the production process, and ultimately the production process tends to be stable and optimal.

[0021] In the embodiment of the present invention, the present invention realizes the effective fusion of multi-dimensional process parameters such as temperature, pressure, flow rate, concentration, etc. by collecting and preprocessing multi-type sensor data, adopts short-time Fourier transform and Welch power spectral density analysis, and combines the three-dimensional finite element heat transfer model to realize high-precision feature extraction of equipment temperature distribution. Through multi-scale sample entropy analysis and nonlinear regression, the equipment performance degradation index is constructed, which can capture the equipment state change from multiple time scales, and improve the accuracy and comprehensiveness of performance evaluation. Using the Bayesian network model, the probabilistic association analysis between environmental factors, process parameters and performance indicators is realized, and the complex interactive relationship between parameters is revealed, which provides a reliable basis for optimization decision-making. Based on the Markov decision process and the double-delay deterministic policy gradient algorithm, an adaptive parameter hierarchical control strategy is constructed, which can effectively cope with the uncertainty and dynamic changes in the production process. Through the hierarchical optimization controller and the adaptive learning rate update mechanism, the continuous optimization of process parameters and the dynamic adjustment of the control strategy are realized, and the stability and efficiency of the production process are guaranteed. By integrating equipment status monitoring, process parameter optimization and intelligent control strategies, the present invention achieves a comprehensive improvement in product quality, production efficiency and equipment reliability, and provides strong support for the intelligent and refined management of the surfactant production process.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: (1) Install multiple types of sensors at key locations of surfactant production equipment to collect temperature, pressure, flow, and concentration parameters in real time to obtain the original multi-dimensional process parameter data stream, and perform time stamp alignment and synchronization on the original multi-dimensional process parameter data stream to obtain a time-synchronized multi-dimensional process parameter data set; (2) performing outlier detection on the time-synchronized multidimensional process parameter data set based on preset data quality evaluation indicators to obtain an abnormal data marking result, and performing interpolation and smoothing processing on the time-synchronized multidimensional process parameter data set according to the abnormal data marking result to obtain a corrected multidimensional process parameter data set; (3) performing data standardization on the corrected multidimensional process parameter data set, unifying parameters of different dimensions to the same scale, obtaining a standardized multidimensional process parameter data set, and performing segmentation processing on the standardized multidimensional process parameter data set based on a sliding window algorithm to obtain a fixed-length data segment sequence; (4) Perform fast Fourier transform on a fixed-length data segment sequence to extract frequency domain features and obtain a multidimensional process parameter spectrum feature matrix. The multidimensional process parameter spectrum feature matrix is ​​then time-series spliced ​​with the standardized multidimensional process parameter data set to obtain multidimensional process parameter time series data.

[0023] Specifically, multiple types of sensors are installed at key locations of the surfactant production equipment to collect key process parameters such as temperature, pressure, flow rate and concentration in real time. Sensors are installed in key locations such as reactors, delivery pipelines, and stirring equipment to capture the changing trends of various important variables in the production process. The data collected by the sensors in real time form an original multi-dimensional process parameter data stream, where each data stream contains time series information, which is used to record the dynamic changes of temperature, pressure, flow rate and concentration in the production process. The original multi-dimensional process parameter data stream is timestamp aligned and synchronized. Through the method of linear interpolation, the acquisition time series of each sensor is adjusted to the same time node to obtain a time-synchronized multi-dimensional process parameter data set. Assume that the collected temperature data is , the pressure data is , the traffic data is , the concentration data is After aligning the timestamps of each data stream, a synchronized multidimensional time series dataset is obtained. Based on the preset data quality evaluation indicators, outlier detection is performed on the time-synchronized multi-dimensional process parameter data set to ensure that the collected data can accurately reflect the operating status of the equipment and is not interfered by sensor noise, sampling errors, etc. Outlier detection determines whether a data point is abnormal by setting a threshold range. For example, if the normal range of temperature is , any temperature value outside this range is considered to be abnormal data. By performing outlier detection on the time-synchronized data, the marking result of the abnormal data is obtained. According to the abnormal data marking result, the time-synchronized multi-dimensional process parameter data set is interpolated and smoothed to keep the data continuous and reasonable. For interpolation processing, Lagrange interpolation or linear interpolation methods are used to fill in the missing or abnormal data points with adjacent valid data points. The smoothing process uses the moving average or exponentially weighted average method to reduce the random noise in the data and obtain a smooth and continuous multi-dimensional process parameter data set. After interpolation and smoothing, the corrected multi-dimensional process parameter data set Reduce unreasonable fluctuations and get closer to the actual production process. The corrected multidimensional process parameter data set contains parameters of different dimensions such as temperature, pressure, flow, concentration, etc. In order to make parameters of different dimensions comparable, the data is standardized. The maximum and minimum value normalization method is used to scale each parameter to the same range (usually [0,1]). Standardized multidimensional process parameter data set Eliminate the influence between different units and orders of magnitude, which is convenient for subsequent data analysis and feature extraction. Based on the sliding window algorithm, the standardized multi-dimensional process parameter data set is segmented, and the time series is divided into several fixed-length data segments by setting the length and step size of the sliding window. For example, if the sliding window length is , the step length is , the time series Divide into several pieces of length Sequence of data fragments ,in is the total number of segments after segmentation. These segments contain the change patterns of process parameters in different time periods. Fast Fourier transform is performed on the fixed-length data segment sequence to extract the frequency domain characteristics of the time series. Fast Fourier transform is an effective tool to transform the time series from the time domain to the frequency domain, revealing the periodicity and frequency components of the data. For example, suppose the data segment is , its frequency domain representation By formula: ; in, For frequency The corresponding spectrum amplitude is is the length of the data segment, is an imaginary unit. The energy distribution of process parameters at different frequencies is obtained by fast Fourier transform. After fast Fourier transform, the spectrum characteristic matrix of multi-dimensional process parameters is obtained. , which contains the amplitude information of each process parameter in the frequency domain. The multidimensional process parameter spectrum feature matrix is ​​spliced ​​with the standardized multidimensional process parameter data set in time series to integrate the information in the time domain and frequency domain, and obtain a multidimensional process parameter time series data set. The spliced ​​time series data set contains the changes of process parameters in the time domain and their frequency domain characteristics, describing the dynamic changes in the production process. The spliced ​​data set is expressed as ,in Represents the time domain characteristics, Represents frequency domain features.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: (1) Perform short-time Fourier transform on the temperature-related parameters in the multidimensional process parameter time series data to obtain the temperature distribution characteristics in the time-frequency domain, and calculate the Welch power spectrum density based on the temperature distribution characteristics in the time-frequency domain to obtain the temperature fluctuation frequency feature vector; (2) Applying the adaptive sliding window algorithm to the multi-dimensional process parameter time series data, the temperature gradient and heat flux density are calculated to obtain the local heat conduction characteristic matrix; (3) A three-dimensional finite element heat transfer model is constructed based on the local heat conduction characteristic matrix. The steady-state heat conduction equation is solved using the conjugate gradient method to obtain high-precision temperature field distribution data. Kernel principal component analysis is performed on the high-precision temperature field distribution data to obtain the thermodynamic characteristic matrix after dimensionality reduction. (4) Input the thermodynamic feature matrix into a deep stacked autoencoder neural network, extract deep features through unsupervised learning, obtain a compressed representation, and perform nonlinear dimensionality reduction on the compressed representation to obtain a two-dimensional temperature distribution feature map; (5) Based on the two-dimensional temperature distribution feature map, the adaptive kernel density estimation is calculated to extract the probability distribution features and obtain the device temperature distribution feature vector.

[0025] Specifically, the temperature-related parameters in the multidimensional process parameter time series data are short-time Fourier transformed to obtain the time-frequency domain distribution characteristics of the temperature parameters. Short-time Fourier transform is an effective time-frequency analysis method. By dividing the temperature data into multiple time windows and performing Fourier transform on each window, the spectrum information of the temperature signal at different time points is obtained. Assume that the temperature data is , using the window function The data is windowed and the short-time Fourier transform is expressed as: ; in, Indicates the time Frequency The spectrum distribution of is the window function, usually a Gaussian window or a Hanning window. is an imaginary unit. Through this process, the time-frequency domain distribution characteristics of temperature are obtained, so as to analyze the frequency change of temperature signal. Based on the time-frequency domain distribution characteristics, the Welch power spectral density (PSD) is calculated. This method can effectively estimate the frequency characteristics of temperature fluctuations by segmenting, windowing and averaging the temperature signal. The calculation formula of Welch power spectral density is: ; in, Indicates frequency The power spectral density at is the number of segments, For the The Fourier transform result of the segment. The frequency characteristic vector of temperature fluctuation is obtained by Welch method. The adaptive sliding window algorithm is applied to the multi-dimensional process parameter time series data to calculate the temperature gradient and heat flux density and obtain the local heat conduction characteristic matrix. The adaptive sliding window algorithm adaptively adjusts the window size according to the dynamic characteristics of temperature change in order to more accurately capture the local temperature change trend. Within the sliding window, the temperature gradient is expressed as: ; in, represents the temperature gradient, and are the temperatures at adjacent locations, is the distance between the two points. According to the temperature gradient, the heat flux density is calculated. The relationship between the heat flux density and the temperature gradient is given by Fourier's law: ; in, represents the heat flux density, is the thermal conductivity of the material, and the negative sign indicates that the direction of heat flow is opposite to the temperature gradient. Through the above calculations, the local heat conduction characteristic matrix is ​​obtained, which describes the temperature changes at different positions in the equipment and the resulting heat flux density. Based on the local heat conduction characteristic matrix, a three-dimensional finite element heat transfer model is constructed to simulate the heat transfer process in the production equipment. The finite element method is an effective numerical calculation method for fine simulation of complex physical systems. By constructing a three-dimensional finite element model, the temperature distribution in different areas inside the equipment is obtained. In order to solve the steady-state heat conduction equation, the conjugate gradient method is used for calculation. The conjugate gradient method is an efficient method for solving large sparse linear systems and quickly obtaining high-precision temperature field distribution data. Assume that the steady-state heat conduction equation is: ; The above equations are solved by the conjugate gradient method to obtain high-precision temperature field distribution data. These data provide detailed information on how heat is transferred inside the device and are used to evaluate the thermal equilibrium state of the device. Kernel principal component analysis is performed on the high-precision temperature field distribution data for dimensionality reduction. Kernel principal component analysis uses kernel functions to map data to high-dimensional space, performs principal component analysis in high-dimensional space, effectively extracts the main features in the temperature field, and reduces the dimension of the data to obtain a thermodynamic feature matrix after dimensionality reduction. The thermodynamic feature matrix after dimensionality reduction is input into a deep stacked autoencoder neural network, and deep features are extracted from the feature matrix through unsupervised learning. The deep stacked autoencoder neural network is composed of multiple stacked autoencoders. Each autoencoder compresses the input data into a low-dimensional representation through learning, and extracts the deep-level features of the data layer by layer. Through this process, a compressed representation of the thermodynamic features is obtained, which contains the main information of the temperature distribution and reduces redundancy. Nonlinear dimensionality reduction is performed on the compressed representation to obtain a two-dimensional temperature distribution feature map. The nonlinear dimensionality reduction method can retain important nonlinear features in the data and project high-dimensional features into two-dimensional space for visualization and subsequent analysis. Based on the two-dimensional temperature distribution feature map, the adaptive kernel density estimation method is used to extract the probability characteristics of the temperature distribution. Adaptive kernel density estimation is a non-parametric method that estimates the probability density function by applying the kernel function to each point in the data. In this process, the bandwidth of the kernel function is adaptively adjusted according to the distribution of the data, thereby more accurately describing the distribution characteristics of the data. Assume that the two-dimensional feature map is , then the kernel density estimate is: ; in, For at point The estimated density at is the sample size, is the kernel function, is the bandwidth parameter. Through adaptive kernel density estimation, the probability characteristics of temperature distribution are obtained. These characteristics describe the probability distribution of temperature at different locations inside the device, and finally the temperature distribution feature vector of the device is obtained.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: (1) Coarse-graining the multidimensional process parameter time series data, setting the scale factor τ to a range of 1 to 20, generating a corresponding coarse-grained time series for each scale factor τ, and obtaining a multi-scale time series set; (2) Based on the multi-scale time series set, the sample entropy calculation parameters are set, the embedding dimension m is 2, the similarity tolerance r is 0.15 times the standard deviation of the original sequence, and the sample entropy value at each scale is calculated to obtain a 20-dimensional multi-scale sample entropy feature vector; (3) The 20-dimensional multi-scale sample entropy feature vector is normalized by using the maximum and minimum value normalization method to map the entropy values ​​of each scale to the interval [0, 1] to obtain the normalized multi-scale sample entropy feature vector; (4) Based on the normalized multi-scale sample entropy feature vector, the slope of the curve of entropy value changing with scale is calculated, and the least squares method is used for linear fitting to obtain the entropy value change rate index. The normalized multi-scale sample entropy feature vector and the entropy value change rate index are then concatenated to construct a comprehensive entropy feature vector as a representation of the equipment performance status. (5) Based on the comprehensive entropy feature vector, a radial basis function kernel matrix is ​​constructed. The kernel function uses a Gaussian kernel and the kernel parameter γ is set to 0.1 to obtain the feature space mapping matrix. (6) Input the feature space mapping matrix into the support vector regression model. The model parameters include the penalty factor C of 10 and the ε-insensitive loss function parameter ε of 0.01 to obtain the initial performance degradation prediction value. (7) Confidence intervals are estimated for the initial performance degradation prediction values ​​to obtain the upper and lower limits of the performance degradation prediction. Based on the upper and lower limits of the performance degradation prediction and combined with the model prediction variance, the weighted average method is used to calculate the equipment performance degradation index.

[0027] Specifically, the multi-dimensional process parameter time series data is coarse-grained and the scale factor is set The range of is 1 to 20, so as to generate the corresponding coarse-grained time series and obtain a multi-scale time series set. Coarse-grained processing is a data preprocessing method that analyzes the dynamic behavior of the system at different time scales by adjusting the scale factor of the time series. For example, assuming that the original time series data is , then for each scale factor , generating a coarse-grained time series

[0028] ; in, For the elements in the coarse-grained time series, is the length of the original sequence, is the scale factor. By coarse-graining the original sequence at different scales, a multi-scale time series set at different time scales is obtained. Based on the multi-scale time series set, the sample entropy value at each scale is calculated to measure the complexity of the system. Set the sample entropy calculation parameters, where the embedding dimension 2, similarity tolerance It is 0.15 times the standard deviation of the original sequence. Sample entropy is an indicator used to measure the complexity of time series. Lower sample entropy means that the time series has stronger predictability, while higher sample entropy means that the system is more complex and difficult to predict. The calculation formula of sample entropy is: ; in, Indicates that the matching embedding dimension is The mode logarithm of Indicates that the matching embedding dimension is The mode logarithm of is the length of the time series. Sample entropy is calculated at each scale to obtain a 20-dimensional multi-scale sample entropy feature vector. These entropy values ​​reflect the complexity characteristics of the system at different time scales. The 20-dimensional multi-scale sample entropy feature vector is normalized. The maximum and minimum value normalization method is used to map the entropy values ​​at each scale to the [0,1] interval to eliminate the differences in the numerical ranges between different scales. The normalized sample entropy feature vector is expressed as: ; in, For the The sample entropy value of the scale is and are the minimum and maximum values ​​of the sample entropy, respectively. After normalization, the entropy values ​​at all scales are scaled to the same range, reflecting the relative complexity between scales. Based on the normalized multi-scale sample entropy feature vector, the slope of the curve of entropy value changing with scale is calculated to analyze the changing trend of system complexity at different scales. The least squares method is used to perform linear fitting on the change of entropy value with scale to obtain the entropy value change rate index. The least squares method is a method for fitting data. By minimizing the sum of squares of errors, the slope and intercept of the best fitting line are obtained. Let the fitting line be: ; in, represents the normalized sample entropy value, is the scale factor, is the slope, i.e. the entropy value change rate indicator, is the intercept. The entropy change rate reflects the changing trend of system complexity over time. The normalized multi-scale sample entropy feature vector is concatenated with the entropy change rate index to construct a comprehensive entropy feature vector. Based on the comprehensive entropy feature vector, a radial basis function (RBF) kernel matrix is ​​constructed to map the original feature vector to a high-dimensional feature space so as to perform linear analysis in the high-dimensional space. The kernel function uses a Gaussian kernel with a kernel parameter of Set to 0.1, the Gaussian kernel is defined as: ; in, and are two input feature vectors, is the Euclidean distance between the two, is the kernel parameter, which controls the width of the Gaussian function. Through the Gaussian kernel function, the comprehensive entropy feature vector is mapped to a high-dimensional space to obtain the feature space mapping matrix. The feature space mapping matrix is ​​input into the support vector regression model to predict the performance degradation of the device. In the SVR model, the penalty factor is set is 10, -Insensitive loss function parameters is 0.01. Penalty factor Used to control the tolerance to errors, the larger Will make the model pay more attention to the fitting of training data, while smaller It emphasizes the generalization ability of the model. The parameter is used to define an allowable error range so that the error within this range is not included in the loss function, thereby improving the robustness of the model. The initial performance degradation prediction value of the equipment is obtained through the SVR model. In order to improve the reliability of the prediction results, the confidence interval of the initial performance degradation prediction value is estimated to obtain the upper and lower limits of the performance degradation prediction. The confidence interval estimation provides a credible interval range for the prediction result, making the assessment of the equipment status more robust. Combined with the prediction variance of the model, the performance degradation index of the equipment is calculated by the weighted average method. The weight of the weighted average is adjusted according to the size of the prediction variance to ensure that the final performance degradation index can more accurately reflect the true status of the equipment.

[0029] In a specific embodiment, the process of executing step S4 may specifically include the following steps: (1) The equipment temperature distribution feature vector and the equipment performance degradation index are combined into an input feature vector, and the input feature vector is input into the Bayesian network model. The Bayesian network model includes: an environmental factor layer, a process parameter layer, and a performance index layer. The environmental factor layer includes three nodes: temperature, humidity, and raw material batch; the process parameter layer includes four nodes: reaction temperature, stirring speed, pH value, and reaction time; and the performance index layer includes three nodes: product purity, surface activity, and stability. (2) Based on the Bayesian network model, the variational inference algorithm is used to calculate the conditional probability distribution from the environmental factor layer to the process parameter layer to obtain the environment-process parameter association matrix. The posterior distribution of the parameters of the environment-process parameter association matrix is ​​sampled to obtain the marginal posterior distribution of the process parameters. (3) Based on the marginal posterior distribution of process parameters, the conditional probability from the process parameter layer to the performance indicator layer is calculated to obtain the process parameter-performance indicator probability transfer matrix; (4) Combine the environment-process parameter association matrix with the process parameter-performance index probability transfer matrix, use the chain rule to calculate the joint probability distribution between the three layers, and obtain the global parameter association probability table; (5) Apply the information gain ratio method to the global parameter association probability table, set the threshold to 0.1, screen out the key process parameter set, and obtain the process parameter association network; (6) Based on the process parameter association network, the maximum likelihood estimation method is used to calculate the conditional independence between parameters, and the optimal dependency structure is constructed to obtain the process parameter association probability matrix.

[0030] Specifically, the temperature distribution feature vector and the performance degradation index are combined to form a comprehensive input feature vector. Assume that the temperature distribution feature vector is ,in The temperature characteristic components, and the equipment performance degradation index is , then the input feature vector Expressed as The input feature vector is input into the Bayesian network model. The Bayesian network model contains three levels, namely, the environmental factor layer, the process parameter layer, and the performance index layer. The environmental factor layer contains three nodes: temperature, humidity, and raw material batch, which are denoted as ,in , and Represent temperature, humidity and raw material batch respectively. The process parameter layer contains four nodes: reaction temperature, stirring speed, pH value and reaction time, which are recorded as ,in , , and Represent the reaction temperature, stirring speed, pH value and reaction time respectively. The performance index layer contains three nodes: product purity, surface activity and stability, which are recorded as ,in , and Respectively represent the purity, surface activity and stability of the product. Based on the Bayesian network model, the variational inference algorithm is used to calculate the conditional probability distribution from the environmental factor layer to the process parameter layer, and the environment-process parameter association matrix is ​​obtained. Variational inference is an effective approximate inference method that transforms the complex Bayesian inference problem into an optimization problem and obtains the conditional probability distribution by minimizing the free energy or log-likelihood approximation. Assume that the conditional probability between environmental factors and process parameters is , through the variational inference algorithm, this conditional probability distribution is approximately calculated and expressed as the environment-process parameter association matrix ,in Represents an environmental factor node Process parameter node The degree of influence of the environment-process parameter association matrix is ​​sampled by the parameter posterior distribution to obtain the marginal posterior distribution of the process parameters. The posterior distribution represents the model's estimate of the state of the process parameters after the observation data is given. By sampling the posterior distribution, the probability distribution of the process parameters under different environmental conditions is obtained. Based on the marginal posterior distribution of the process parameters, the conditional probability from the process parameter layer to the performance indicator layer is calculated to obtain the probability transfer matrix of the process parameter-performance indicator. Assume that the conditional probability between the process parameters and the performance indicators is By inferring the marginal posterior distribution of the process parameters, the influence of each process parameter on the performance index is calculated, and the process parameter-performance index probability transfer matrix is ​​obtained. ,in Indicates the process parameter node For performance indicator nodes The environment-process parameter correlation matrix and process parameter-performance index probability transfer matrix Combined, the chain rule is used to calculate the joint probability distribution between the three layers. According to the chain rule of probability theory, the joint probability of the three layers is expressed as: ; in, is the prior probability of the environmental factor, is the conditional probability of environmental factors on process parameters, is the conditional probability of the process parameters to the performance index. Through the joint probability calculation, the global parameter association probability table is obtained , which describes the global correlation between environmental factors, process parameters and performance indicators. In order to screen out the key process parameters that have a greater impact on the system, the information gain ratio method is applied to the global parameter association probability table. The information gain ratio is an indicator that measures the influence of a feature on a target variable. By calculating the information gain ratio of each process parameter to the performance indicator and setting the threshold to 0.1, the key process parameter set is screened out and the process parameter association network is obtained. The calculation formula for the information gain ratio is: ; in, is the process parameter The information gain of For parameters The entropy of is calculated, and the contribution of each process parameter to the overall performance of the system is measured by calculating the information gain ratio. Based on the selected process parameter association network, the maximum likelihood estimation method is used to calculate the conditional independence between the parameters and construct the optimal dependency structure. Maximum likelihood estimation is a statistical inference method that estimates the optimal dependency relationship between the nodes in the model by maximizing the likelihood function of the observed data. Assume that the observed data is , then the goal of maximum likelihood estimation is to find the parameter , so that the likelihood function By this method, it is determined which process parameters are conditionally independent and which are dependent, so as to construct the optimal dependency structure. Based on the optimal dependency structure, the process parameter association probability matrix is ​​obtained. ,in Indicates the strength of association between process parameters. This matrix can fully reflect the degree of association between various process parameters.

[0031] In a specific embodiment, the process of executing step S5 may specifically include the following steps: (1) Based on the process parameter association probability matrix, the Markov decision process state space is constructed. Each process parameter is discretized into five levels to obtain a multidimensional discrete state space. The Markov decision process action space is also constructed. Three adjustment actions, namely, increase, decrease, and maintain, are defined for each process parameter to obtain a multidimensional discrete action space. (2) Based on historical production data, the state transition probability matrix of the Markov decision process is calculated, and the state transition frequency is statistically estimated using the maximum likelihood estimation method to obtain the state transition function. The Markov decision process reward function is designed, and the product quality index, energy consumption index, and equipment performance degradation index are comprehensively considered. The scalar reward signal is obtained through the weight coefficient method; (3) The current process parameter state is input into the double-delay deterministic policy gradient algorithm. The two Q networks of the double-delay deterministic policy gradient algorithm adopt a four-layer fully connected neural network structure, with 256 neurons in each hidden layer and 1 in the output layer; the deterministic policy network adopts a three-layer fully connected neural network structure with 128 neurons in the hidden layer; (4) Generate deterministic actions through the deterministic policy network, and add Gaussian noise to the generated deterministic action a to explore the action space and obtain the actual execution action; (5) Adjust process parameters according to the actual execution action, observe environmental feedback, record new states and immediate rewards, input state transition samples into the experience replay buffer, and update sample priorities; (6) Using two Q networks, calculate the Q-value estimate of the current state-action pair from the experience replay buffer, and generate the target Q-value based on the calculated Q-value estimate; (7) Use the temporal difference error to update the two Q networks, minimize the mean square error between the predicted value of the Q network and the target Q value, and update the policy network through deterministic policy gradient to maximize the evaluation value of the Q1 network for the current policy action. Select the optimal strategy as the parameter hierarchical control strategy based on the evaluation value.

[0032] Specifically, each process parameter is discretized. Assuming there is a multidimensional set of process parameters, such as reaction temperature, stirring speed, pH value, and reaction time, each parameter is discretized into 5 levels, such as "low, medium-low, medium, medium-high, high". By discretizing each parameter, a multidimensional discrete state space is obtained, in which each state represents a specific operating condition under a different combination of process parameters. For example, for reaction temperature, stirring speed, and pH value, each parameter has 5 possible levels, then the state space of the entire system is the Cartesian product of these three parameters, with possible states. At the same time, in order to adjust the process parameters, the action space of the Markov decision process is constructed. Three adjustment actions are defined for each process parameter, namely "increase", "decrease" and "maintain", forming a multi-dimensional discrete action space. In this discrete action space, any combination of actions represents a way to adjust a certain process parameter. For example, for the reaction temperature, choose to increase or decrease the temperature, or maintain the current state. Based on historical production data, the state transition probability matrix of the Markov decision process is calculated. By analyzing the changes in process parameters and the corresponding state transition frequencies in historical data, the maximum likelihood estimation method is used to count the transition frequencies of each state to the next state to obtain the state transition probability. Assume that the state changes from Transfer to state The number of times is ,state The total number of transfers is , then the state transition probability is expressed as: ; in, From the state Transfer to state The conditional probability of . In this way, a state transition probability matrix describing the state transition of the entire system is obtained. This matrix is ​​used in the Markov decision process to describe the possibility of the system changing from one state to another. In order to make the entire system evolve in the optimal direction during the decision-making process, a reward function of the Markov decision process is designed. The design of the reward function takes into account a variety of factors, including product quality indicators, energy consumption indicators, and equipment performance degradation indicators. In order to integrate indicators of different natures, the weight coefficient method is used to assign weights to each indicator. For example, assuming that the product quality indicator is , energy consumption index is , equipment degradation index is , the weights are , , , then the scalar reward signal It is expressed as: ; in, , , is the weight coefficient, which is used to measure the degree of influence of each indicator on the overall system. By setting these weights reasonably, the system can take into account production quality, energy consumption and equipment stability during the optimization process. The state of the current process parameters is input into the double-delay deterministic policy gradient algorithm (TD3). The TD3 algorithm jointly realizes the learning of the optimal strategy through two Q networks and a deterministic policy network. Both Q networks adopt a four-layer fully connected neural network structure. Each hidden layer contains 256 neurons and the output layer is 1, which is used to evaluate the Q value of the current state-action pair. The deterministic policy network adopts a three-layer fully connected neural network structure. Each hidden layer contains 128 neurons to generate deterministic actions. These actions are used to guide the specific adjustment of process parameters. In order to increase the exploratory nature and prevent falling into the local optimum when generating deterministic actions in the deterministic policy network, the generated deterministic actions are Adding Gaussian noise increases a certain degree of randomness in the action space, enables exploration of unknown areas, and obtains actual execution actions ,in, The mean is 0 and the variance is Gaussian noise. According to the actual execution action generated, adjust the process parameters of the production equipment, observe the feedback from the environment, record the new state of the equipment and the immediate reward. Store the state transition samples in the experience replay buffer and update the priority of the samples. The role of the experience replay buffer is to break the time correlation of the samples and improve the stability and efficiency of training. Randomly extract samples from the experience replay buffer and use two Q networks to calculate the Q value estimate of the current state-action pair. The Q value is used to measure the expected future benefits of performing an action in a given state. The dual Q network is used in the TD3 algorithm to reduce estimation bias and improve learning stability. Assume that the current Q value estimate is , the target Q value is , expressed by the Bellman equation: ; in, For instant rewards, is the discount factor, and are the estimated values ​​of the two Q networks, and Represent the next state and the next action respectively. By calculating the target Q value, the two Q networks are updated using the time difference error to minimize the mean square error between the Q network prediction value and the target Q value, ensuring that the Q network can accurately evaluate the value of the state-action pair. While updating the Q network, the policy network is updated through the deterministic policy gradient, with the goal of maximizing the evaluation value of the Q1 network for the current policy action, that is: ; in, The actions generated by the policy network, is the parameter of the policy network. By maximizing the evaluation value of the Q1 network, the policy network can gradually generate the optimal control strategy. Based on the training results of the Q network and the policy network, the optimal control strategy is selected as the parameter hierarchical control strategy and used to guide the process parameter adjustment of the production equipment.

[0033] In a specific embodiment, the process of executing step S6 may specifically include the following steps: (1) Inputting the parameter hierarchical control strategy into the hierarchical optimization controller, hierarchically dividing the process parameters of the production equipment, and obtaining a high-level control parameter set and a low-level control parameter set, wherein the high-level control parameter set includes reaction temperature and pH value, and the low-level control parameter set includes stirring speed and reaction time; (2) A model predictive controller is constructed based on the high-level control parameter set. A nonlinear autoregressive exogenous model is used as the prediction model. The prediction period is set to 10 and the control period is set to 2. The high-level parameter optimization instruction sequence is obtained. (3) According to the high-level parameter optimization instruction sequence, the low-level control parameter set is constrained optimized, and the sequential quadratic programming algorithm is used to solve the constrained optimization problem to obtain the low-level parameter optimization instruction sequence; (4) Merging the high-level parameter optimization instruction sequence and the low-level parameter optimization instruction sequence to construct a process parameter optimization instruction sequence, and sending the process parameter optimization instruction sequence to the actuator of the production equipment, and calculating the performance deviation of the actuator; (5) Based on the performance deviation, an adaptive learning rate update mechanism is constructed to fine-tune the parameter hierarchical control strategy to obtain the updated control strategy; (6) The updated control strategy is input into the hierarchical optimization controller to perform a closed-loop feedback optimization mechanism to continuously update the control parameter instruction sequence of the production equipment until the preset optimization termination condition is met.

[0034] Specifically, the hierarchical control strategy is input into the hierarchical optimization controller to hierarchically divide the process parameters of the production equipment. The parameter set is divided into a high-level control parameter set and a low-level control parameter set, where the high-level control parameter set includes reaction temperature and pH value, while the low-level control parameter set includes stirring speed and reaction time. Through the hierarchical division method, the high-level control parameters are the macroscopic conditions that affect the entire production process and determine the chemical kinetics of the reaction, while the low-level control parameters are more fine-tuned for production operations to ensure the expected reaction effect. After the parameters are stratified, a model predictive controller (MPC) is constructed based on the high-level control parameter set. The model predictive controller is a control strategy that predicts the future system state and makes decisions based on optimization criteria by establishing a process model. The nonlinear autoregressive exogenous (NARX) model is used as the prediction model. The NARX model can well describe the nonlinear characteristics of the system, and its form is: ; in, represents the output of the system (e.g. reaction temperature or pH), Indicates the control input (such as the adjusted heating power or the amount of acid or alkali added), is a nonlinear function, is the noise term of the system. In the model predictive controller, the prediction period is set to 10, which means that the model can predict the system state within the next 10 time steps, and the control period is set to 2, which means that the parameters are adjusted every 2 time steps. Through the setting, the optimization instruction sequence of high-level parameters is obtained. The optimization instruction sequence contains the adjustment plan of the reaction temperature and pH value in the future period, so as to ensure that the production process always proceeds in the predetermined direction. According to the optimization instruction sequence of high-level parameters, the low-level control parameter set is constrained and optimized. The optimization of the low-level control parameter set is subject to the constraints of the high-level parameters to ensure the coordination and consistency of the overall process. The sequential quadratic programming algorithm is used to solve the constrained optimization problem. Sequential quadratic programming is an effective algorithm suitable for nonlinear constrained optimization problems. It gradually approaches the optimal solution by iteratively solving the quadratic programming problem. Let the objective function be , represents the total energy consumption or total error of the system. The constraints are the instruction sequence of high-level parameter optimization. The optimization problem is expressed as: ; in, For low-level control parameters (such as stirring speed and reaction time), and Represent inequality constraints and equality constraints respectively. The constraints ensure that the low-level control is optimally adjusted while meeting the high-level requirements. By solving the above optimization problem, the optimization instruction sequence of the low-level parameters is obtained. These instructions are used to fine-tune the operating state of the production equipment to achieve more stable reaction conditions. The high-level parameter optimization instruction sequence and the low-level parameter optimization instruction sequence are combined to construct a complete process parameter optimization instruction sequence. The optimization instruction sequence is sent to the actuator of the production equipment, and the actuator adjusts the various process parameters accordingly. In order to ensure that the equipment operates as expected, the performance deviation of the actuator is calculated, that is, the difference between the actual operation results and the expected instructions. The performance deviation is expressed as: ; in, For the Performance deviation at any time, is the reference instruction value (i.e. the set value in the optimized instruction sequence), is the actual output of the actuator. By calculating the performance deviation, the control effect of the equipment is evaluated and a reference is provided for subsequent strategy adjustments. In order to improve the adaptability of the control strategy, an adaptive learning rate update mechanism is constructed based on the performance deviation to fine-tune the parameter hierarchical control strategy. The goal of the adaptive learning rate update mechanism is to dynamically adjust the learning rate according to the size of the performance deviation, ensuring rapid adjustment when the deviation is large and fine adjustment when the deviation is small. Suppose the learning rate is , the update formula of adaptive learning rate is: ; in, is the learning rate at the next moment, is the learning rate adjustment coefficient, is the absolute value of the deviation at the current moment. In this way, the control strategy is fine-tuned according to the actual feedback of the system, and the control effect is gradually improved. The updated control strategy is input into the hierarchical optimization controller again, and the control parameter instruction sequence of the production equipment is continuously updated using the closed-loop feedback optimization mechanism. The role of the closed-loop feedback optimization mechanism is to continuously monitor various parameters in the production process, and continuously optimize the control strategy according to the difference between the actual production results and the expected goals until the preset optimization termination conditions are met. The optimization termination condition is set as the performance deviation is less than a certain threshold, or the fluctuation of the process parameters remains stable within the set range. Through closed-loop feedback optimization, the system can continuously adapt to changes in the production environment and realize dynamic optimization control of process parameters.

[0035] The above describes the surfactant process parameter analysis method based on multi-level factors in the embodiment of the present invention. The following describes the surfactant process parameter analysis system based on multi-level factors in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a surfactant process parameter analysis system based on multi-level factors includes: The acquisition module is used to collect and preprocess multi-dimensional process parameters of the surfactant production equipment to obtain multi-dimensional process parameter time series data; The extraction module is used to extract thermodynamic features and reduce the dimension of multi-dimensional process parameter time series data to obtain the equipment temperature distribution feature vector; A calculation module is used to calculate multi-scale sample entropy based on multi-dimensional process parameter time series data and perform nonlinear regression analysis to obtain equipment performance degradation indicators; An analysis module is used to input the equipment temperature distribution feature vector and the equipment performance degradation index into the Bayesian network model to perform process parameter association analysis and obtain a process parameter association probability matrix; A construction module is used to construct a Markov decision process model according to the process parameter association probability matrix and execute a double-delayed deterministic policy gradient algorithm to obtain a parameter hierarchical control strategy; The update module is used to optimize the process parameters of the production equipment based on the parameter hierarchical control strategy, and continuously update the control parameter instruction sequence of the production equipment through a closed-loop feedback mechanism.

[0036] Through the synergy of the above-mentioned components, the present invention realizes the effective fusion of multi-dimensional process parameters such as temperature, pressure, flow, concentration, etc. by collecting and preprocessing multi-type sensor data. The short-time Fourier transform and Welch power spectral density analysis are combined with a three-dimensional finite element heat transfer model to realize high-precision feature extraction of equipment temperature distribution. Through multi-scale sample entropy analysis and nonlinear regression, the equipment performance degradation index is constructed, which can capture equipment state changes from multiple time scales and improve the accuracy and comprehensiveness of performance evaluation. Using the Bayesian network model, the probabilistic association analysis between environmental factors, process parameters and performance indicators is realized, revealing the complex interactive relationship between parameters, and providing a reliable basis for optimization decision-making. Based on the Markov decision process and the double-delay deterministic policy gradient algorithm, an adaptive parameter hierarchical control strategy is constructed, which can effectively cope with the uncertainty and dynamic changes in the production process. Through the hierarchical optimization controller and the adaptive learning rate update mechanism, the continuous optimization of process parameters and the dynamic adjustment of the control strategy are realized, ensuring the stability and efficiency of the production process. By integrating equipment status monitoring, process parameter optimization and intelligent control strategies, the present invention achieves a comprehensive improvement in product quality, production efficiency and equipment reliability, and provides strong support for the intelligent and refined management of the surfactant production process.

[0037] Reference Figure 3In 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.

[0038] 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.

[0039] 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.

[0040] 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.

[0041] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0042] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0043] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A surfactant process parameter analysis method based on multi-level factors, characterized in that: The method comprises: Collect and preprocess multi-dimensional process parameters of surfactant production equipment to obtain multi-dimensional process parameter time series data; Performing thermodynamic feature extraction and dimensionality reduction processing on the multi-dimensional process parameter time series data to obtain a device temperature distribution feature vector; Calculating multi-scale sample entropy based on the multi-dimensional process parameter time series data, and performing nonlinear regression analysis to obtain equipment performance degradation indicators; Inputting the equipment temperature distribution feature vector and the equipment performance degradation index into a Bayesian network model to perform process parameter association analysis to obtain a process parameter association probability matrix; A Markov decision process model is constructed according to the process parameter association probability matrix, and a double-delayed deterministic policy gradient algorithm is executed to obtain a parameter hierarchical control strategy; The process parameters of the production equipment are optimized based on the parameter hierarchical control strategy, and the control parameter instruction sequence of the production equipment is continuously updated through a closed-loop feedback mechanism.

2. The surfactant process parameter analysis method based on multi-level factors according to claim 1, characterized in that: The multi-dimensional process parameter collection and preprocessing of the surfactant production equipment to obtain multi-dimensional process parameter time series data includes: Install multiple types of sensors at key locations of surfactant production equipment to collect temperature, pressure, flow rate, and concentration parameters in real time to obtain an original multi-dimensional process parameter data stream, and perform time stamp alignment and synchronization processing on the original multi-dimensional process parameter data stream to obtain a time-synchronized multi-dimensional process parameter data set; Performing outlier detection on the time-synchronized multidimensional process parameter data set based on a preset data quality assessment index to obtain an abnormal data marking result, and performing interpolation and smoothing processing on the time-synchronized multidimensional process parameter data set according to the abnormal data marking result to obtain a corrected multidimensional process parameter data set; Performing data standardization processing on the modified multidimensional process parameter data set to unify parameters of different dimensions to the same scale to obtain a standardized multidimensional process parameter data set, and performing segmentation processing on the standardized multidimensional process parameter data set based on a sliding window algorithm to obtain a data segment sequence of fixed length; A fast Fourier transform is performed on the fixed-length data segment sequence to extract frequency domain features to obtain a multidimensional process parameter spectrum feature matrix, and the multidimensional process parameter spectrum feature matrix is ​​time-series spliced ​​with the standardized multidimensional process parameter data set to obtain multidimensional process parameter time series data.

3. The surfactant process parameter analysis method based on multi-level factors according to claim 2 is characterized in that: The step of performing thermodynamic feature extraction and dimensionality reduction processing on the multi-dimensional process parameter time series data to obtain a device temperature distribution feature vector includes: Performing short-time Fourier transform on the temperature-related parameters in the multidimensional process parameter time series data to obtain the temperature distribution characteristics in the time-frequency domain, and calculating the Welch power spectrum density based on the temperature distribution characteristics in the time-frequency domain to obtain the temperature fluctuation frequency characteristic vector; Applying an adaptive sliding window algorithm to the multi-dimensional process parameter time series data, calculating the temperature gradient and heat flux density, and obtaining a local heat conduction characteristic matrix; A three-dimensional finite element heat transfer model is constructed based on the local heat conduction characteristic matrix, a conjugate gradient method is used to solve the steady-state heat conduction equation to obtain high-precision temperature field distribution data, and a kernel principal component analysis is performed on the high-precision temperature field distribution data to obtain a thermodynamic characteristic matrix after dimensionality reduction; Inputting the thermodynamic feature matrix into a deep stacked autoencoder neural network, extracting deep features through unsupervised learning to obtain a compressed representation, and performing nonlinear dimensionality reduction on the compressed representation to obtain a two-dimensional temperature distribution feature map; Based on the two-dimensional temperature distribution feature map, an adaptive kernel density estimation is calculated, probability distribution features are extracted, and a device temperature distribution feature vector is obtained.

4. The surfactant process parameter analysis method based on multi-level factors according to claim 3 is characterized in that: The multi-scale sample entropy is calculated based on the multi-dimensional process parameter time series data, and nonlinear regression analysis is performed to obtain the equipment performance degradation index, including: Performing coarse-graining processing on the multi-dimensional process parameter time series data, setting the scale factor τ to a range of 1 to 20, generating a corresponding coarse-grained time series for each scale factor τ, and obtaining a multi-scale time series set; Based on the multi-scale time series set, set the sample entropy calculation parameters, the embedding dimension m is 2, the similarity tolerance r is 0.15 times the standard deviation of the original sequence, calculate the sample entropy value at each scale, and obtain a 20-dimensional multi-scale sample entropy feature vector; The 20-dimensional multi-scale sample entropy feature vector is normalized by using a maximum and minimum value normalization method to map the entropy value of each scale to the interval [0, 1] to obtain a normalized multi-scale sample entropy feature vector; Based on the normalized multi-scale sample entropy feature vector, the slope of the curve of entropy value changing with scale is calculated, and the least square method is used for linear fitting to obtain the entropy value change rate index, and the normalized multi-scale sample entropy feature vector is spliced ​​with the entropy value change rate index to construct a comprehensive entropy feature vector as a representation of the performance status of the device; Based on the comprehensive entropy feature vector, a radial basis function kernel matrix is ​​constructed, the kernel function adopts a Gaussian kernel, and the kernel parameter γ is set to 0.1, to obtain a feature space mapping matrix; Inputting the feature space mapping matrix into a support vector regression model, wherein the model parameters include a penalty factor C of 10 and an ε-insensitive loss function parameter ε of 0.01, to obtain an initial performance degradation prediction value; A confidence interval is estimated for the initial performance degradation prediction value to obtain the upper and lower limits of the performance degradation prediction, and based on the upper and lower limits of the performance degradation prediction and in combination with the model prediction variance, a weighted average method is used to calculate the equipment performance degradation index.

5. The surfactant process parameter analysis method based on multi-level factors according to claim 4 is characterized in that: The device temperature distribution feature vector and the device performance degradation index are input into the Bayesian network model to perform process parameter association analysis to obtain a process parameter association probability matrix, including: The device temperature distribution feature vector and the device performance degradation index are combined into an input feature vector, and the input feature vector is input into a Bayesian network model, wherein the Bayesian network model includes: an environmental factor layer, a process parameter layer, and a performance index layer, wherein the environmental factor layer includes three nodes: temperature, humidity, and raw material batch; the process parameter layer includes four nodes: reaction temperature, stirring speed, pH value, and reaction time; and the performance index layer includes three nodes: product purity, surface activity, and stability; Based on the Bayesian network model, a variational inference algorithm is used to calculate the conditional probability distribution from the environmental factor layer to the process parameter layer to obtain an environment-process parameter association matrix, and the environment-process parameter association matrix is ​​sampled for parameter posterior distribution to obtain a marginal posterior distribution of process parameters; Based on the marginal posterior distribution of the process parameters, the conditional probability from the process parameter layer to the performance indicator layer is calculated to obtain a process parameter-performance indicator probability transfer matrix; The environment-process parameter association matrix is ​​combined with the process parameter-performance index probability transfer matrix, and the joint probability distribution among the three layers is calculated by using the chain rule to obtain a global parameter association probability table; Applying the information gain ratio method to the global parameter association probability table, setting the threshold to 0.1, screening out a set of key process parameters, and obtaining a process parameter association network; Based on the process parameter association network, the conditional independence between parameters is calculated using the maximum likelihood estimation method, and the optimal dependency structure is constructed to obtain the process parameter association probability matrix.

6. The surfactant process parameter analysis method based on multi-level factors according to claim 5, characterized in that: The Markov decision process model is constructed according to the process parameter association probability matrix, and a double-delayed deterministic policy gradient algorithm is executed to obtain a parameter hierarchical control strategy, including: A Markov decision process state space is constructed based on the process parameter association probability matrix, each process parameter is discretized into 5 levels to obtain a multidimensional discrete state space, and a Markov decision process action space is constructed, and three adjustment actions of increase, decrease and maintenance are defined for each process parameter to obtain a multidimensional discrete action space; The Markov decision process state transition probability matrix is ​​calculated based on historical production data, and the state transition frequency is statistically analyzed using the maximum likelihood estimation method to obtain the state transition function. The Markov decision process reward function is designed, and the product quality index, energy consumption index and equipment performance degradation index are comprehensively considered. The scalar reward signal is obtained through the weight coefficient method. The current process parameter state is input into the double-delay deterministic policy gradient algorithm. The two Q networks of the double-delay deterministic policy gradient algorithm adopt a four-layer fully connected neural network structure, each hidden layer has 256 neurons, and the output layer is 1; the deterministic policy network adopts a three-layer fully connected neural network structure, and the hidden layer has 128 neurons; Generate a deterministic action through the deterministic policy network, and add Gaussian noise to the generated deterministic action a to achieve action space exploration and obtain the actual execution action; Adjust process parameters according to actual execution actions, observe environmental feedback, record new states and immediate rewards, input state transition samples into the experience replay buffer, and update sample priorities; Calculate a Q-value estimate for the current state-action pair from the experience replay buffer using the two Q-networks, and generate a target Q-value based on the calculated Q-value estimate; The two Q networks are updated using the temporal difference error to minimize the mean square error between the predicted value of the Q network and the target Q value, and the policy network is updated using the deterministic policy gradient to maximize the evaluation value of the Q1 network for the current policy action. The optimal strategy is selected as the parameter hierarchical control strategy based on the evaluation value.

7. The surfactant process parameter analysis method based on multi-level factors according to claim 6 is characterized in that: The step of optimizing the process parameters of the production equipment based on the parameter hierarchical control strategy and continuously updating the control parameter instruction sequence of the production equipment through a closed-loop feedback mechanism includes: Inputting the parameter hierarchical control strategy into a hierarchical optimization controller, hierarchically dividing the process parameters of the production equipment to obtain a high-level control parameter set and a low-level control parameter set, wherein the high-level control parameter set includes reaction temperature and pH value, and the low-level control parameter set includes stirring speed and reaction time; Based on the high-level control parameter set, a model predictive controller is constructed, a nonlinear autoregressive exogenous model is used as a prediction model, a prediction period is set to 10, a control period is set to 2, and a high-level parameter optimization instruction sequence is obtained; Performing constrained optimization on the low-level control parameter set according to the high-level parameter optimization instruction sequence, solving the constrained optimization problem using a sequential quadratic programming algorithm, and obtaining a low-level parameter optimization instruction sequence; Merging the high-level parameter optimization instruction sequence and the low-level parameter optimization instruction sequence to construct a process parameter optimization instruction sequence, sending the process parameter optimization instruction sequence to an actuator of a production device, and calculating a performance deviation of the actuator; Building an adaptive learning rate update mechanism based on the performance deviation, fine-tuning the parameter hierarchical control strategy, and obtaining an updated control strategy; The updated control strategy is input into the hierarchical optimization controller to perform a closed-loop feedback optimization mechanism, and the control parameter instruction sequence of the production equipment is continuously updated until a preset optimization termination condition is met.

8. A surfactant process parameter analysis system based on multi-level factors, characterized in that: For executing the surfactant process parameter analysis method based on multi-level factors according to any one of claims 1 to 7, the system comprises: The acquisition module is used to collect and preprocess multi-dimensional process parameters of the surfactant production equipment to obtain multi-dimensional process parameter time series data; An extraction module, used for performing thermodynamic feature extraction and dimensionality reduction processing on the multi-dimensional process parameter time series data to obtain a characteristic vector of equipment temperature distribution; A calculation module, used to calculate multi-scale sample entropy based on the multi-dimensional process parameter time series data, and perform nonlinear regression analysis to obtain equipment performance degradation indicators; An analysis module, used for inputting the equipment temperature distribution feature vector and the equipment performance degradation index into a Bayesian network model to perform process parameter association analysis to obtain a process parameter association probability matrix; A construction module is used to construct a Markov decision process model according to the process parameter association probability matrix, and execute a double-delayed deterministic policy gradient algorithm to obtain a parameter hierarchical control strategy; An updating module is used to optimize the process parameters of the production equipment based on the parameter hierarchical control strategy, and continuously update the control parameter instruction sequence of the production equipment through a closed-loop feedback mechanism.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, the surfactant process parameter analysis method based on multi-level factors described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the surfactant process parameter analysis method based on multi-level factors according to any one of claims 1 to 7.

Images