Extraction separation process modeling and optimization method and system based on big data analysis

Abstract

The application provides an extraction separation process modeling and optimization method and system based on big data analysis, relates to the technical field of chemical process control, and comprises the following steps: obtaining multi-source heterogeneous process data of an extraction separation process and performing multi-scale time sequence decomposition to obtain a feature set, constructing a causal relationship graph and identifying a direct causal path, determining a causal topological structure and giving a mass transfer mechanism constraint weight; taking the causal edge as a state transmission path and the weight as a transmission coefficient, solving a state evolution trajectory and correcting the transmission coefficient to obtain a dynamic evolution relationship; calculating the global influence degree of each operation parameter on a separation performance index, identifying a key operation parameter, calculating a separation performance time sequence trajectory in a dimension reduction optimization space, and combining a local gradient search algorithm to determine an optimal operation parameter time sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chemical process control, and in particular to an extraction separation process modeling and optimization method and system based on big data analysis. BACKGROUND

[0002] The extraction separation process is an important operation in the chemical, pharmaceutical, metallurgical and other industries, which realizes the separation and purification of target components by using the difference in distribution coefficient between two phases of different substances, involves complex mass transfer, heat transfer and fluid mechanics phenomena, has the characteristics of multi-variable coupling, strong non-linear dynamic characteristics, sensitive operation conditions, etc., with the improvement of product purity requirements of industrial production and the strictness of energy consumption control.

[0003] The traditional extraction separation process modeling method mainly relies on mechanism model, which describes the process behavior by establishing material balance, energy balance and phase equilibrium equation. With the improvement of industrial process automation level, a large amount of process data has been accumulated in the production site, which provides the basis for data-driven modeling. However, the existing data-driven methods mostly use pure black box model, which still has the problems of regarding the extraction separation process as a static optimization problem, being difficult to adapt to the process dynamic characteristics, being difficult to effectively handle the multi-source heterogeneous data fusion problem in the extraction separation process, being unable to fully mine the process characteristic information at different time scales, causing the modeling precision to be limited and the dimension disaster problem when dealing with high-dimensional operation parameter space, being easy to fall into local optimum, and the calculation cost increasing exponentially with the dimension, etc. SUMMARY

[0004] The embodiments of the present application provide an extraction separation process modeling and optimization method and system based on big data analysis, which can at least solve some of the problems in the prior art.

[0005] In a first aspect, the embodiments of the present application provide an extraction separation process modeling and optimization method based on big data analysis, comprising:

[0006] Obtain multi-source heterogeneous process data of the extraction separation process and perform multi-scale time series decomposition to obtain a multi-scale process state feature set, construct a causal relationship graph based on the multi-scale process state feature set and identify a direct causal path through conditional independence verification, determine a causal topological structure based on the direct causal path and assign a mass transfer mechanism constraint weight to each causal edge;

[0007] Take each causal edge in the causal topological structure as a state transfer path and take the mass transfer mechanism constraint weight as a transfer coefficient, determine an initial state based on the multi-scale process state feature set and solve a state evolution trajectory, calculate the deviation between the state evolution trajectory and the multi-source heterogeneous process data and correct the transfer coefficient to obtain a dynamic evolution relationship;

[0008] The global influence degrees of each operation parameter in the dynamic evolution relationship on the separation performance index are calculated, key operation parameters and a redundant operation parameter subset are identified, the redundant operation parameter subset is fixed as a safe operation interval center value to obtain a reduced optimization space, separation performance time trajectories corresponding to different key operation parameters are calculated in the reduced optimization space, multi-objective evaluation indexes are calculated based on the separation performance time trajectories and a preset separation purity target and energy consumption constraint, an optimal operation parameter time sequence is determined by combining a local gradient search algorithm, and an operation trajectory optimization scheme is obtained.

[0009] In an optional implementation,

[0010] Multi-source heterogeneous process data of an extraction separation process are acquired and multi-scale time sequence decomposition is performed to obtain a multi-scale process state feature set, a causal relationship graph is constructed based on the multi-scale process state feature set, and a direct causal path is identified through conditional independence verification, including:

[0011] For the multi-source heterogeneous process data in the extraction separation process, high-frequency components are extracted as fast-changing dynamic features through wavelet decomposition, low-frequency components are extracted as slow-changing trend features, and the multi-scale process state feature set is obtained by combining the fast-changing dynamic features and the slow-changing trend features, wherein the multi-source heterogeneous process data includes material component concentration data, operation parameter time sequence data, and separation efficiency representation data.

[0012] The fast-changing dynamic features and the slow-changing trend features are respectively mapped to time sequence nodes, and mutual information amounts between different time sequence nodes are calculated, an initial causal relationship graph is obtained by establishing an association edge between the time sequence nodes according to the mutual information amounts.

[0013] Conditional independence tests are performed on each association edge in the initial causal relationship graph, time sequence nodes other than two time sequence nodes connected by the current association edge in the multi-scale process state feature set are taken as a conditional variable set, a conditional mutual information amount between the two time sequence nodes connected by the current association edge is calculated under the constraint of the conditional variable set, when the conditional mutual information amount is higher than a preset independence threshold, the current association edge is marked as a direct causal path, and the calculation and marking are repeated until all association edges are tested, and a direct causal path set is obtained.

[0014] In an optional implementation,

[0015] A mass transfer mechanism constraint weight is assigned to each causal edge based on the direct causal path, including:

[0016] topologically sort operation parameter nodes corresponding to operation parameter time series data in each direct causal path in the set of pre-acquired direct causal paths according to a causal transmission order, determine hierarchical positions of the operation parameter nodes in causal transmission links according to a topological sorting result, and construct a causal topology based on the hierarchical positions;

[0017] For each causal edge in the causal topology, calculate a manifold distance between a multi-scale process state feature set corresponding to a starting node and a terminal node connected by the causal edge, and quantify a state transmission difficulty of the causal edge in a nonlinear state space based on the manifold distance;

[0018] According to the state transmission difficulty, calculate an influence weight of the causal edge on the extraction separation process in combination with a topological centrality of the causal edge in the causal topology, perform a physical interpretability correction on the influence weight by a mass transfer mechanism constraint to obtain a mass transfer mechanism constraint weight of the causal edge and label the causal edge.

[0019] In an optional implementation,

[0020] Taking each causal edge in the causal topology as a state transmission path and taking the mass transfer mechanism constraint weight as a transmission coefficient, determining an initial state based on the multi-scale process state feature set and solving a state evolution trajectory include:

[0021] Determining a mass transfer mechanism constraint weight corresponding to each causal edge and organizing the mass transfer mechanism constraint weight into a state propagation matrix, extracting fast-changing dynamic features and slow-changing trend features corresponding to each node in the multi-scale process state feature set and encoding the fast-changing dynamic features and the slow-changing trend features into vectors to perform tensor product operation to obtain a state tensor, performing singular value decomposition on the state tensor to determine a dominant mode and determining an initial state based on the dominant mode;

[0022] Loading the initial state to a node in the causal topology and iteratively propagating the initial state along the state transmission path through the state propagation matrix, calculating a similarity between a current node and corresponding multi-order neighborhood nodes and determining a receptive field range in each iteration, extracting upstream nodes within the receptive field range, calculating a distance weight based on a topological distance from the upstream nodes to the current node, calculating a modulation coefficient based on the distance weight and the transmission coefficient, and weighting state values of the upstream nodes, combining a preset nonlinear activation function to transform and accumulate to the current node after the weighting, recording state value numbers of each node in each iteration and arranging the state value numbers in time sequence to form a state evolution trajectory.

[0023] In an optional implementation,

[0024] Calculating a deviation between the state evolution trajectory and the multi-source heterogeneous process data and correcting the transmission coefficient to obtain a dynamic evolution relationship include:

[0025] extracting measured data corresponding to nodes in the state evolution trajectory from the multi-source heterogeneous process data, time-aligning the measured data and the state evolution trajectory to obtain an observation sequence, calculating point-by-point difference values of the state evolution trajectory and the observation sequence at each node in each time step and performing square summation to obtain a global deviation;

[0026] calculating partial derivatives of the global deviation with respect to the transfer coefficients, determining key causal edges according to absolute values of the partial derivatives and constructing a key path set, calculating gradients of the global deviation with respect to each transfer coefficient in the key path set and determining an adjustment direction and an adjustment step size of the transfer coefficients, calculating a correction amount based on the adjustment direction and the adjustment step size and applying the correction amount to the transfer coefficients to obtain updated transfer coefficients;

[0027] substituting the updated transfer coefficients into a state propagation matrix obtained in advance and performing the solving process of the state evolution trajectory again to obtain an updated state evolution trajectory and calculate an updated deviation from the observation sequence, determining whether a decrease amplitude of the updated deviation with respect to the global deviation satisfies a preset convergence condition, if yes, taking the updated transfer coefficients as corrected transfer coefficients, if no, replacing the global deviation with the updated deviation and repeating the calculation of the gradients and the corrected transfer coefficients until the preset convergence condition is satisfied, extracting the corrected transfer coefficients and associating and labeling them with corresponding causal edges in the causal topological structure to obtain a dynamic evolution relationship.

[0028] In an optional implementation,

[0029] calculating global influence degrees of each operation parameter in the dynamic evolution relationship on the separation performance index and identifying key operation parameters and a redundant operation parameter subset, fixing the redundant operation parameter subset as a safe operation interval center value to obtain a reduced optimization space, comprising:

[0030] identifying all paths between source nodes corresponding to operation parameter time series data and target nodes corresponding to a preset separation performance index in the causal topological structure based on the dynamic evolution relationship, constructing a transfer operator sequence in order for causal edges on each path and performing non-commutative compound operation on adjacent transfer operators to obtain a path cascade operator, calculating an operator norm of the path cascade operator as a path transfer strength, and summarizing all path transfer strengths of the same source node to the same target node to obtain a global influence degree corresponding to the operation parameter time series data;

[0031] construct a disturbance field corresponding to each operation parameter in the safe operation interval based on the dynamic evolution relationship and calculate the influence of different operation parameters on the separation performance index under the disturbance field, perform tensor decomposition on the influence to extract the maximum eigenvalue of the second-order tensor component to obtain a curvature index corresponding to each operation parameter;

[0032] fuse the global influence degree and the curvature index to calculate a comprehensive importance score of each operation parameter and perform cluster analysis to obtain a key operation parameter cluster and a redundant operation parameter cluster, extract operation parameters in the redundant operation parameter cluster to obtain a redundant operation parameter subset and fix the center value of the safe operation interval, take the key operation parameters in the key operation parameter cluster as optimization variables and take the corresponding safe operation interval as value constraints to construct a reduced optimization space.

[0033] In an optional implementation,

[0034] In the reduced optimization space, calculate the separation performance time sequence trajectory corresponding to different key operation parameters, calculate the multi-objective evaluation index based on the separation performance time sequence trajectory and the preset separation purity target and energy consumption constraint, determine the optimal operation parameter time sequence by combining the local gradient search algorithm, and obtain the operation trajectory optimization scheme, including:

[0035] In the reduced optimization space, sample the key operation parameters to obtain multiple groups of sampling points, take each group of sampling points as the initial value of the key operation parameters, and perform forward propagation from the source node corresponding to the key operation parameters to the target node along the causal topology structure and sequentially apply the transfer operator corresponding to each causal edge to the propagation signal to obtain the separation performance time sequence trajectory;

[0036] Calculate the separation purity compliance degree and the energy consumption deviation degree corresponding to the separation performance time sequence trajectory based on the preset separation purity target and the energy consumption constraint, construct a Pareto front set in the target space based on the separation purity compliance degree and the energy consumption deviation degree, and perform convex hull analysis to identify an ideal point, calculate the Chebyshev distance of each separation purity compliance degree and energy consumption deviation degree to the ideal point to obtain a multi-objective evaluation index;

[0037] Select the sampling point with the minimum multi-objective evaluation index value as the initial search point, construct a Riemann metric tensor based on the pre-obtained curvature index and map the reduced optimization space to a Riemann manifold, define geodesics on the Riemann manifold and perform gradient search, adjust the local curvature radius of the geodesics according to the curvature index to determine an adaptive search step, calculate the multi-objective evaluation index of the current search point in each iteration until a preset convergence condition is reached, extract the search point at the termination of iteration as the optimal operation parameter, and generate an operation trajectory optimization scheme in combination with the time sequence change.

[0038] In a second aspect, the embodiment of the present application provides an extraction separation process modeling and optimization system based on big data analysis, comprising:

[0039] A first unit is configured to acquire multi-source heterogeneous process data of an extraction separation process, perform multi-scale time series decomposition to obtain a multi-scale process state feature set, construct a causal relationship graph based on the multi-scale process state feature set, identify a direct causal path through conditional independence verification, determine a causal topological structure based on the direct causal path, and assign a mass transfer mechanism constraint weight to each causal edge;

[0040] A second unit is configured to take each causal edge in the causal topological structure as a state transfer path, take the mass transfer mechanism constraint weight as a transfer coefficient, determine an initial state based on the multi-scale process state feature set, solve a state evolution trajectory, calculate a deviation between the state evolution trajectory and the multi-source heterogeneous process data, and correct the transfer coefficient to obtain a dynamic evolution relationship.

[0041] A third unit is configured to calculate a global influence degree of each operation parameter in the dynamic evolution relationship on a separation performance index, identify a key operation parameter and a redundant operation parameter subset, fix the redundant operation parameter subset as a safe operation interval center value to obtain a reduced optimization space, calculate a separation performance time series trajectory corresponding to different key operation parameters in the reduced optimization space, calculate a multi-objective evaluation index based on the separation performance time series trajectory and a preset separation purity target and energy consumption constraint, determine an optimal operation parameter time series sequence by combining a local gradient search algorithm, and obtain an operation trajectory optimization scheme.

[0042] In a third aspect, the embodiment of the present application provides an electronic device, comprising:

[0043] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0044] In a fourth aspect, the embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions are executed by a processor to implement the method described above.

[0045] This invention, through multi-scale time-series decomposition of multi-source heterogeneous data and construction of causal relationship maps, can deeply explore the intrinsic causal relationships between variables in the extraction and separation process. It overcomes the limitations of traditional modeling methods that rely solely on apparent correlations, achieving accurate characterization of complex mass transfer mechanisms. This significantly improves the accuracy and generalization ability of describing actual industrial processes. By combining causal topology with mass transfer mechanism constraints to identify key operating parameters, it effectively reduces the dimensionality and computational complexity of the optimization problem. It solves the problems of traditional high-dimensional optimization methods easily getting trapped in local optima and having low computational efficiency, thus significantly shortening computation time while ensuring optimization quality and enhancing the method's engineering practicality. By constructing a time-series optimization scheme for operating parameters, it achieves a shift from static operating point optimization to dynamic operating trajectory optimization, effectively improving the overall economic efficiency and energy utilization efficiency of the extraction and separation process. This provides strong technical support for the intelligent control and green, low-carbon development of chemical processes. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the extraction and separation process modeling and optimization method based on big data analysis, as described in an embodiment of the present invention.

[0047] Figure 2 This is a flowchart illustrating the causal topology analysis and parameter optimization process of the extraction and separation process modeling and optimization method based on big data analysis, as described in this embodiment of the invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0050] Figure 1 This is a flowchart illustrating the extraction and separation process modeling and optimization method based on big data analysis, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0051] The multi-source heterogeneous process data of the extraction separation process are acquired and multi-scale time series decomposition is performed to obtain a multi-scale process state feature set, a causal relationship graph is constructed based on the multi-scale process state feature set, and a direct causal path is identified through conditional independence verification, a causal topological structure is determined based on the direct causal path, and a mass transfer mechanism constraint weight is assigned to each causal edge;

[0052] Each causal edge in the causal topological structure is taken as a state transfer path, and the mass transfer mechanism constraint weight is taken as a transfer coefficient, an initial state is determined based on the multi-scale process state feature set, and a state evolution trajectory is solved, a deviation between the state evolution trajectory and the multi-source heterogeneous process data is calculated, and the transfer coefficient is corrected to obtain a dynamic evolution relationship;

[0053] The global influence degree of each operating parameter on the separation performance index in the dynamic evolution relationship is calculated, and a key operating parameter and a redundant operating parameter subset are identified, the redundant operating parameter subset is fixed as a safe operation interval center value to obtain a reduced optimization space, the separation performance time series trajectory corresponding to different key operating parameters is calculated in the reduced optimization space, a multi-objective evaluation index is calculated based on the separation performance time series trajectory and a preset separation purity target and energy consumption constraint, and an optimal operating parameter time series sequence is determined by combining a local gradient search algorithm to obtain an operation trajectory optimization scheme.

[0054] In an optional implementation,

[0055] The multi-source heterogeneous process data of the extraction separation process are acquired and multi-scale time series decomposition is performed to obtain a multi-scale process state feature set, a causal relationship graph is constructed based on the multi-scale process state feature set, and a direct causal path is identified through conditional independence verification, a causal topological structure is determined based on the direct causal path, and a mass transfer mechanism constraint weight is assigned to each causal edge;

[0056] For the multi-source heterogeneous process data in the extraction separation process, a high-frequency component is extracted as a fast-changing dynamic feature through wavelet decomposition, a low-frequency component is extracted as a slow-changing trend feature, and the fast-changing dynamic feature and the slow-changing trend feature are combined to obtain a multi-scale process state feature set, wherein the multi-source heterogeneous process data includes material component concentration data, operating parameter time series data, and separation efficiency representation data;

[0057] The fast-changing dynamic feature and the slow-changing trend feature are respectively mapped to time series nodes, and the mutual information amount between different time series nodes is calculated, and an association edge is established between the time series nodes according to the mutual information amount to obtain an initial causal relationship graph;

[0058] The conditional independence test is performed on each association edge in the initial causal graph, the remaining time series nodes except the two time series nodes connected by the current association edge in the multi-scale process state feature set are taken as the condition variable set, the conditional mutual information between the two time series nodes connected by the current association edge under the constraint of the condition variable set is calculated, when the conditional mutual information is higher than the preset independence threshold, the current association edge is marked as a direct causal path, and the calculation and marking are repeated until all association edges are tested, and a direct causal path set is obtained.

[0059] In the extraction separation process, multi-source heterogeneous process data are collected, including material component concentration data, operation parameter time series data and separation efficiency characterization data. The material component concentration data records the concentration change of each related substance in the extraction separation process, such as the concentration of target metal ions in the aqueous phase, the concentration of extractant in the organic phase, etc.; the operation parameter time series data includes the change records of process parameters such as temperature, PH value, stirring speed and flow rate with time; the separation efficiency characterization data includes separation factor, extraction rate, stripping rate and other index data for evaluating the separation effect.

[0060] For the collected multi-source heterogeneous process data, multi-scale features are extracted by using wavelet decomposition method. Specifically, the original time series data is decomposed by using db4 wavelet basis function for 5 layers. In the decomposition process, the original signal is decomposed into approximation coefficients and detail coefficients. The high-frequency component, i.e. the detail coefficient, is a fast-changing dynamic feature, which can reflect the sudden change, disturbance and noise in the process; the low-frequency component, i.e. the approximation coefficient, is a slow-changing trend feature, which can reflect the long-term change trend and steady-state behavior of the process. For example, for the measured pH time series data in an extraction process, the high-frequency features reflecting the sudden change of pH and the low-frequency features reflecting the slow change trend of pH can be obtained by wavelet decomposition. By combining the extracted fast-changing dynamic features and slow-changing trend features, a complete multi-scale process state feature set is formed, which can comprehensively describe the dynamic behavior of the extraction separation process.

[0061] To construct the initial causal graph, the extracted fast-varying dynamic features and slow-varying trend features are mapped into time series nodes respectively. Specifically, each time series of features is regarded as a node in the graph, such as the temperature high-frequency component node, the pH low-frequency component node, and so on. The correlation between nodes is established by calculating the mutual information between different time series nodes. The mutual information is calculated by using the kernel density estimation method, which estimates the joint probability distribution and the marginal probability distribution by using the Gaussian kernel function, and then calculates the mutual information value. For two time series nodes, when the mutual information between them exceeds a preset threshold, such as 0.2, an association edge is established between the two nodes, indicating that there is a potential causal relationship. For example, if the mutual information between the high-frequency component of the extractant concentration and the low-frequency component of the extraction rate is 0.35, which exceeds the threshold of 0.2, an association edge is established between the two nodes. By traversing all node pairs and calculating the mutual information, the construction of the initial causal graph is completed.

[0062] To eliminate indirect associations in the initial causal graph, conditional independence tests are performed on each association edge in the graph. In specific implementation, for the current association edge to be tested, all other nodes in the multi-scale process state feature set except the two time series nodes connected by the current association edge are defined as the conditional variable set. The conditional mutual information between the two time series nodes connected by the current association edge is calculated under the constraint of the conditional variable set. The conditional mutual information is calculated by using the nearest neighbor estimation method, which estimates the conditional mutual information by investigating the local distribution characteristics of data points in high-dimensional space. For example, for time series data with a data volume of 1000 sampling points, 5 to 10 nearest neighbor points can be selected for estimation. If the calculated conditional mutual information is higher than a preset independence threshold, such as 0.15, it indicates that there is still significant association between the two nodes even under the condition of all other variables, and the current association edge is marked as a direct causal path. For example, in a certain extraction process, the conditional mutual information between the temperature high-frequency component node and the separation factor low-frequency component node under the condition of all other nodes is 0.22, which is higher than the threshold of 0.15, so it is confirmed that there is a direct causal relationship between them. The above calculation and marking process is repeated until all association edges in the graph are tested, and finally a set of direct causal paths is obtained.

[0063] In this embodiment, the fast-varying dynamic features and slow-varying trend features are extracted by wavelet decomposition, which can distinguish between rapid disturbances and slow evolutionary trends, can completely capture the multi-scale behavior features of the extraction separation process, avoid the problem of insufficient state description caused by ignoring high-frequency or low-frequency information in the prior art, construct the initial causal graph by mutual information, which can identify potential associations in nonlinear and multi-variable coupled scenarios, has stronger applicability, eliminate indirect associations by conditional independence test, and only retain the real direct causal paths, effectively improve the reliability of causal structure identification, and overcome the defect that the causal chain is confused due to insufficient consideration of conditional dependence.

[0064] In an alternative embodiment,

[0065] Determining a causal topology based on the direct causal paths and assigning a mass transfer mechanism constraint weight to each causal edge comprises:

[0066] Topologically sorting the operation parameter nodes corresponding to the operation parameter time series data in each direct causal path in the pre-obtained direct causal path set according to the causal transmission order, determining the hierarchical positions of the operation parameter nodes in the causal transmission link according to the topological sorting results, and constructing the causal topology based on the hierarchical positions.

[0067] For each causal edge in the causal topology, calculating the manifold distance between the multi-scale process state feature sets corresponding to the start node and the end node connected by the causal edge, and quantifying the state transmission difficulty of the causal edge in the nonlinear state space based on the manifold distance.

[0068] According to the state transmission difficulty, the influence weight of the causal edge on the extraction separation process is calculated in combination with the topological centrality of the causal edge in the causal topology, the influence weight is physically interpretable by mass transfer mechanism constraint correction, the mass transfer mechanism constraint weight of the causal edge is obtained and labeled.

[0069] Traverse the pre-obtained direct causal path set, extract the nodes related to operation parameter time series data and their causal relationship. Operation parameter nodes include extraction temperature nodes, pH nodes, flow nodes, stirring speed nodes, etc. Topological sorting of causal transmission order is realized by using depth-first search algorithm. Specifically, starting from any unvisited node, recursively visit all its successor nodes, and when all successor nodes of a node have been visited, add the node to the sorting result. In the process of topological sorting, the loop problem may exist, and the loop is detected and broken by the causal edge with the smallest influence weight in the loop to ensure the effectiveness of the sorting. According to the topological sorting results, the hierarchical positions of the operation parameter nodes in the causal transmission link are determined, and the smaller the hierarchical position value is, the more forward the node is in the causal transmission link, and the more extensive the influence on other nodes is. For example, in a certain rare earth extraction separation process, the pH node level is 1, the temperature node level is 2, the flow node level is 3, and the stirring speed node level is 2 after topological sorting. Based on the hierarchical position information, the causal topology is constructed to form a directed acyclic graph, which clearly represents the causal transmission relationship between operation parameters.

[0070] For each causal edge in the causal topology, the manifold distance between the sets of multi-scale process state features corresponding to the start node and the end node of the edge is calculated to quantify the state transfer difficulty in the nonlinear state space. The manifold distance calculation adopts the geodesic distance estimation method, which considers the nonlinear distribution characteristics of data points in high-dimensional feature space, constructs a k-nearest neighbor graph in the feature space, and each data point is connected to its k nearest neighbors, with k typically set to 5-10. The Euclidean distance between the nearest neighbors is calculated as the edge weight, and the shortest path length between any two points is calculated as the geodesic distance using Dijkstra's algorithm. For a feature set with 500 sampling points, a 10-nearest neighbor graph is constructed for calculation. For example, in a certain extraction separation process, the manifold distance from the temperature node to the separation factor node is 2.37, indicating that the state transfer difficulty from temperature change to separation factor change is relatively large; while the manifold distance from the pH node to the extraction rate node is 1.28, indicating that the state transfer difficulty is relatively small.

[0071] According to the calculated state transfer difficulty, combined with the topological centrality of the causal edge in the causal topology, the influence weight of each causal edge on the extraction separation process is calculated. The topological centrality is measured by the degree centrality index, which is calculated as the sum of the in-degree and out-degree of the node divided by the total number of edges in the graph. For each causal edge, the influence weight is calculated as the product of the degree centrality of the start node and the end node of the edge divided by the manifold distance of the edge. For example, in a certain extraction separation process, the influence weight of the causal edge from the temperature node to the extraction rate node is 0.42, and the influence weight of the causal edge from the pH node to the separation factor node is 0.63, indicating that the latter has a greater impact on the extraction separation process.

[0072] To make the influence weight physically interpretable, a mass transfer mechanism constraint is introduced for modification. The mass transfer mechanism constraint is based on the basic theory of the extraction separation process, including the two-phase distribution coefficient, the interface mass transfer rate, and the phase equilibrium relationship, etc. Specifically, the theoretical influence degree of different operating parameter changes on the mass transfer performance is calculated to form a mass transfer sensitivity matrix. The element value of the mass transfer sensitivity matrix represents the percentage change of the mass transfer performance caused by a unit change in the operating parameter. The data-driven influence weight is weighted and fused with the corresponding element value in the mass transfer sensitivity matrix, with the fusion weight generally set to 0.7 and 0.3, i.e. the data-driven result accounts for 70% and the theoretical model result accounts for 30%, to obtain the weight with mass transfer mechanism constraint, and mark it on the corresponding edge of the causal topology. For example, in a certain rare earth extraction process, the mass transfer mechanism constraint weight of temperature on extraction rate is 0.39, and the mass transfer mechanism constraint weight of pH on separation factor is 0.58.

[0073] In this embodiment, by topologically sorting the operation parameter nodes and determining their hierarchical positions, the order and structural role of the operation parameters in the causal transmission link can be clearly revealed, overcoming the problem of unclear causal hierarchy due to multi-variable cross-influence. The manifold distance is used to quantify the state transmission difficulty in a nonlinear state space, which can identify the real influence path under complex coupling conditions and more accurately reflect the nonlinear dynamic characteristics in the extraction and separation process. By combining the topological centrality of the causal edge with the state transmission difficulty to calculate the influence weight and introducing the mass transfer mechanism constraint for correction, the weight not only has quantitative differentiation ability, but also can conform to the actual physical mass transfer law, solving the short board of lacking mechanism interpretation in the existing data-driven causal analysis method.

[0074] In an alternative embodiment,

[0075] Each causal edge in the causal topology structure is taken as a state transmission path, and the mass transfer mechanism constraint weight is taken as a transmission coefficient. Based on the multi-scale process state feature set, an initial state is determined and a state evolution trajectory is solved, including:

[0076] The mass transfer mechanism constraint weight corresponding to each causal edge is determined and organized into a state propagation matrix. The fast-changing dynamic features and slow-changing trend features of each node in the multi-scale process state feature set are extracted and encoded into a vector for tensor product operation to obtain a state tensor. Singular value decomposition is performed on the state tensor to determine a dominant mode, and an initial state is determined based on the dominant mode.

[0077] The initial state is loaded into the nodes in the causal topology structure and iteratively propagated along the state transmission path through the state propagation matrix. In each iteration, the similarity between the current node and the corresponding multi-order neighborhood nodes is calculated and the receptive field range is determined. The upstream nodes within the receptive field range are extracted. The distance weight is calculated according to the topological distance from the upstream nodes to the current node. The modulation coefficient is calculated based on the distance weight and the transmission coefficient, and the state values of the upstream nodes are weighted. After transformation by a pre-set nonlinear activation function, the state values are accumulated to the current node. In each iteration, the state value of each node is recorded and arranged in time sequence to form a state evolution trajectory.

[0078] The mass transfer mechanism constraint weight corresponding to each causal edge obtained in the foregoing is organized into a state propagation matrix, which represents the state transmission coefficient between nodes in the causal topology. Specifically, for a causal topology containing n nodes, an n x n state propagation matrix is constructed, in which the matrix element value corresponds to the mass transfer mechanism constraint weight between nodes. If there is no causal edge between two nodes, the corresponding matrix element value is set to 0. For example, in a certain rare earth extraction separation process, the matrix element value from the temperature node to the extraction rate node in the state propagation matrix containing 6 nodes is 0.39, the matrix element value from the acid-base node to the separation factor node is 0.58, and the matrix element value corresponding to the node pair without direct causal relationship is 0. After the state propagation matrix is constructed, the matrix is subjected to row normalization processing to ensure that the matrix has good numerical stability.

[0079] The fast-changing dynamic features and slow-changing trend features corresponding to each node are extracted from the multi-scale process state feature set and encoded into vector form. For each node in the extraction separation process, its state is represented as a two-dimensional vector containing fast-changing dynamic features and slow-changing trend features. For example, the state of the temperature node can be represented as a vector containing the temperature fast-changing dynamic feature value 0.45 and the temperature slow-changing trend feature value 0.68. The state vectors of all nodes are subjected to tensor product operation to construct a state tensor. For a causal topology containing m nodes, the dimension of the state tensor is 2 m . To reduce the computational complexity, the state tensor is subjected to singular value decomposition to extract the dominant mode. Singular value decomposition is realized through an iterative algorithm, and the principal component with a singular value contribution rate of more than 85% is extracted as the dominant mode. For example, the state tensor of a certain extraction separation process is decomposed to obtain 3 dominant modes, and the singular value contribution rates thereof are 65%, 15% and 10% respectively. Based on the dominant mode, the initial state value of each node is determined, and the initial state reflects the starting condition of the extraction separation process.

[0080] The determined initial state is loaded into each node in the causal topology, preparing for state propagation simulation. The simulation is performed in an iterative manner, and in each iteration, the state value propagates between nodes along the state transmission path. For each node, the similarity between the multi-order neighborhood nodes is calculated to determine the receptive field range. The similarity calculation uses the cosine similarity method, with a value range of 0 to 1, and the similarity threshold is set to 0.6, that is, nodes with a similarity greater than 0.6 are included in the receptive field range. For example, in a certain iteration step, the similarity of the temperature node with the pH node, flow node, and stirring speed node is 0.78, 0.52, and 0.65, respectively, so the pH node and the stirring speed node are included in the receptive field range of the temperature node. The upstream nodes are extracted from the receptive field range, that is, the nodes pointing to the current node in the causal topology. For each upstream node, the distance weight is calculated according to its topological distance to the current node. The topological distance is defined as the length of the shortest path between two nodes, and the distance weight is calculated as the inverse of the topological distance. The smaller the topological distance, the greater the distance weight, indicating a more significant impact.

[0081] Based on the calculated distance weight and the transmission coefficient in the state propagation matrix, the modulation coefficient is calculated and the state value of the upstream node is weighted. The modulation coefficient is calculated as the product of the distance weight and the transmission coefficient, representing the contribution of the upstream node to the state update of the current node. The weighted state value of the upstream node is transformed through a predetermined nonlinear activation function, and the activation function uses the hyperbolic tangent function. The transformed state value is accumulated to the current node to complete a state update. For example, for the extraction rate node, the modulation coefficients of its upstream nodes temperature and pH are 0.35 and 0.62, respectively. The weighted state value is transformed and accumulated through the activation function to obtain the updated state value of the extraction rate node.

[0082] In each iteration, the state value of each node is recorded, and these state values are arranged in chronological order to form a state evolution trajectory. The state evolution trajectory reflects the trend of the state of each node changing with time in the extraction separation process, and can be used for process behavior analysis and anomaly detection. By setting a reasonable number of iterations, such as 100 times, the state evolution process in a long enough time period can be simulated.

[0083] In this embodiment, the mass transfer mechanism constraint weight is organized into a state propagation matrix, and the dominant mode is extracted using singular value decomposition of the state tensor. Based on multi-scale feature fusion, the core features that best represent the key trends of the process can be selected, effectively avoiding the accumulation of redundant interference and noise caused by direct use of original multi-variable features, improving the representativeness and stability of the initial state. By loading the initial state into the causal topology structure and iteratively propagating along the causal path, the state evolution can follow the real causal chain corrected by the mechanism constraint. By weighting the upstream node state with the modulation coefficient and introducing the nonlinear activation, the expression ability for nonlinear coupling relationship and threshold effect can be effectively enhanced, overcoming the limitation of not reflecting the dynamic of complex process, and significantly improving the modeling accuracy, stability and physical credibility of the dynamic behavior of the extraction separation process. More reliable basis is provided for process optimization control and early abnormal identification.

[0084] In an optional embodiment,

[0085] The deviation between the state evolution trajectory and the multi-source heterogeneous process data is calculated, and the transfer coefficient is corrected to obtain a dynamic evolution relationship.

[0086] The measured data corresponding to the nodes in the state evolution trajectory in the multi-source heterogeneous process data is extracted, the measured data and the state evolution trajectory are time-aligned to obtain an observation sequence, the point-by-point difference of the state evolution trajectory and the observation sequence at each node at each time step is calculated, and the global deviation is obtained by summing the squares.

[0087] The partial derivative of the global deviation with respect to the transfer coefficient is calculated, the key causal edges are determined according to the absolute value of the partial derivative, and the key path set is constructed. The gradient of the global deviation with respect to each transfer coefficient in the key path set is calculated, and the adjustment direction and adjustment step of the transfer coefficient are determined. The correction amount is calculated based on the adjustment direction and the adjustment step, and the correction amount is applied to the transfer coefficient to obtain an updated transfer coefficient.

[0088] The updated transfer coefficient is substituted into the state propagation matrix obtained in advance, and the state evolution trajectory solving process is performed again to obtain an updated state evolution trajectory and calculate an updated deviation with the observation sequence. It is determined whether the decrease amplitude of the updated deviation with respect to the global deviation satisfies a preset convergence condition. If it satisfies, the updated transfer coefficient is taken as a corrected transfer coefficient. If it does not satisfy, the updated deviation is used to replace the global deviation, and the gradient and the corrected transfer coefficient are repeatedly calculated until the preset convergence condition is satisfied. The corrected transfer coefficient is extracted and associated with the corresponding causal edge in the causal topology structure to obtain a dynamic evolution relationship.

[0089] The measured data corresponding to the nodes in the state evolution trajectory is extracted from the multi-source heterogeneous process data, including time series data such as temperature, pH, flow rate, stirring speed, extraction rate, etc. The measured data and the state evolution trajectory are time-aligned to generate an observation sequence. Time alignment uses the dynamic time warping algorithm to find the best matching path on the time axis, solving the inconsistency of two time series data in sampling frequency, starting time, etc. After time alignment, the point-by-point difference of the state evolution trajectory and the observation sequence at each node in each time step is calculated, and the global deviation is obtained by summing the squares. For example, in a certain rare earth extraction separation process, the state evolution trajectory and the observation sequence of 8 nodes are compared, the point-by-point difference sum of 100 time steps is calculated, and the global deviation value is 3.27.

[0090] The partial derivative of the global deviation with respect to the transfer coefficient, i.e. the sensitivity, is calculated. The sensitivity calculation uses the numerical difference method, which observes the change rate of the global deviation by applying a small perturbation to the transfer coefficient. Exemplarily, an incremental perturbation of 0.01 is applied to each transfer coefficient, the state evolution trajectory and the global deviation are recalculated, and the approximate partial derivative is obtained by dividing the deviation change by the perturbation. According to the absolute value of the partial derivative, the key causal edge is determined, and the larger the absolute value, the more significant the influence of the causal edge on the global deviation. The top 30% of the causal edges with the largest absolute value of the partial derivative are selected to construct the critical path set. For example, in a certain extraction separation process, there are 15 causal edges, and after calculating the partial derivative, the partial derivative of temperature to extraction rate is -2.35, the partial derivative of pH to separation factor is -1.87, and the partial derivative of flow rate to stripping rate is -0.56. The 5 causal edges with the largest absolute value of the partial derivative are selected to form the critical path set.

[0091] The gradient of the global deviation with respect to each transfer coefficient in the critical path set is calculated to determine the adjustment direction and adjustment step of the transfer coefficient. The gradient calculation also uses the numerical difference method, but with higher precision, and the perturbation amount is set to 0.001. The adjustment direction is opposite to the gradient direction, i.e. the transfer coefficient is decreased when the gradient is positive, and the transfer coefficient is increased when the gradient is negative. The adjustment step uses an adaptive strategy to dynamically adjust based on the gradient size. Specifically, the adjustment step is calculated as the base step multiplied by the gradient normalization factor, the base step is set to 0.05, and the gradient normalization factor is the ratio of the current gradient absolute value to the maximum gradient absolute value. Based on the adjustment direction and the adjustment step, the correction amount is calculated, and the updated transfer coefficient is obtained by applying the correction amount to the transfer coefficient. For example, the original value of the transfer coefficient from temperature to extraction rate is 0.39, the gradient is -2.35, the adjustment step is 0.05, and the calculated correction amount is 0.05x2.35 / 2.35=0.05, and the updated transfer coefficient is 0.39+0.05=0.44.

[0092] The updated transfer coefficient is substituted into the pre-obtained state propagation matrix, and the solving process of the state evolution trajectory is performed again to obtain an updated state evolution trajectory. Compared with the observation sequence, an updated deviation is calculated to determine the reduction amplitude of the updated deviation relative to the global deviation. The reduction amplitude is calculated as the ratio of the deviation reduction amount to the original deviation, expressed as a percentage. It is determined whether the reduction amplitude meets a preset convergence condition, and the convergence condition is set as the reduction amplitude being less than 1% or the number of iterations exceeding 50. If the convergence condition is met, the updated transfer coefficient is used as the corrected transfer coefficient; if not, the updated deviation is used to replace the global deviation, and the gradient and the corrected transfer coefficient are repeatedly calculated until the preset convergence condition is met. Exemplarily, the initial global deviation of a certain extraction separation process is 3.27, the updated deviation after the first iteration is 2.85, and the reduction amplitude is 12.8%, which does not meet the convergence condition, and the iteration continues. After 15 iterations, the updated deviation is 2.52, which is 0.8% lower than the previous 2.54, and the reduction amplitude meets the convergence condition, the iteration is terminated, and the final corrected transfer coefficient is obtained.

[0093] The corrected transfer coefficient is extracted and associatedly labeled with the corresponding causal edges in the causal topological structure to obtain a dynamic evolution relationship.

[0094] Exemplarily, taking a certain rare earth element extraction separation process as an example, a causal topological structure containing 8 nodes is selected, and the initial transfer coefficient is set based on the mass transfer mechanism constraint weight. The measured data is extracted from the process data of 7 consecutive days, time-aligned with the state evolution trajectory, and the global deviation is calculated to be 3.27. After calculating the partial derivative, 5 key causal edges are determined: temperature to extraction rate, pH to separation factor, extractant concentration to extraction rate, flow rate to stripping rate, and stirring speed to separation factor. The gradient is calculated and the transfer coefficient is updated for the 5 key causal edges. The transfer coefficient of temperature to extraction rate is updated from 0.39 to 0.44, the transfer coefficient of pH to separation factor is updated from 0.58 to 0.63, the transfer coefficient of extractant concentration to extraction rate is updated from 0.47 to 0.51, the transfer coefficient of flow rate to stripping rate is updated from 0.32 to 0.35, and the transfer coefficient of stirring speed to separation factor remains unchanged. After 15 iterations of optimization, the global deviation is reduced from 3.27 to 2.52, with a reduction amplitude of 23.0%, and the corrected transfer coefficient is obtained. The corrected transfer coefficient is associatedly labeled with the corresponding causal edges in the causal topological structure to obtain a dynamic evolution relationship.

[0095] In this embodiment, by time-aligning the state evolution trajectory with the actual process data and constructing the observation sequence, the problem of deviation of causal strength from the actual situation caused by relying only on theoretical assumptions or single modeling can be avoided. By point-by-point difference calculation and summarization, the global deviation can be obtained, the adaptability of the current causal parameter can be quantified from the degree of coincidence of the overall dynamic behavior, the comprehensiveness and reliability of the deviation measurement can be improved, by calculating the partial derivative of the global deviation to each transfer coefficient, the key causal edge can be screened, focusing on the most significant path affecting the deviation for key modification, avoiding the dimension disaster and invalid update caused by full parameter update, improving the calculation efficiency and the sensitivity of path identification, by determining the correction amount of the transfer coefficient through the gradient direction and the step length, the update process has a clear optimization target and stable iteration direction, enhancing the controllability and stability of parameter convergence.

[0096] In an alternative embodiment,

[0097] Calculating the global influence degree of each operation parameter on the separation performance index in the dynamic evolution relationship and identifying the key operation parameter and the redundant operation parameter subset, fixing the redundant operation parameter subset as the center value of the safe operation interval to obtain the reduced optimization space includes:

[0098] Based on the dynamic evolution relationship, all paths between the source node corresponding to the operation parameter time series data and the target node corresponding to the preset separation performance index in the causal topology structure are identified, the transfer operator sequence is constructed in order for each causal edge on each path, and the non-commutative composite operation is performed on adjacent transfer operators to obtain a path cascade operator. The operator norm of the path cascade operator is calculated as the path transfer strength, and the global influence degree corresponding to the operation parameter time series data is obtained by summarizing the path transfer strengths of all paths from the same source node to the same target node.

[0099] Based on the dynamic evolution relationship, a perturbation field corresponding to each operation parameter within the safe operation interval is constructed, and the influence of different operation parameters on the separation performance index under the perturbation field is calculated. The maximum eigenvalue of the second-order tensor component is extracted by tensor decomposition to obtain the curvature index corresponding to each operation parameter.

[0100] The global influence degree and the curvature index are fused to calculate the comprehensive importance score of each operation parameter, and clustering analysis is performed to obtain the key operation parameter cluster and the redundant operation parameter cluster. The operation parameters in the redundant operation parameter cluster are extracted to obtain the redundant operation parameter subset, which is fixed as the center value of the safe operation interval. The key operation parameters in the key operation parameter cluster are taken as optimization variables, and the corresponding safe operation interval is taken as the value constraint to construct a reduced optimization space.

[0101] All paths from the source node corresponding to the operation parameter time series data to the target node corresponding to the preset separation performance index are identified based on the dynamic evolution relationship. The identification process uses a depth-first search algorithm to start from the operation parameter node and traverse all possible paths to the separation performance index node. In the extraction and separation process, common operation parameters include temperature, pH, flow rate, stirring speed, and extractant concentration, and separation performance indicators include extraction rate, separation factor, and stripping rate. For example, for the temperature node to the extraction rate node, there may be a direct path "temperature → extraction rate", or an indirect path "temperature → pH → separation factor → extraction rate". For each causal edge on the path, a transfer operator sequence is constructed in order. The transfer operator is defined as the transfer function corresponding to the causal edge, which describes the influence of the source node change on the target node. The non-commutative composition operation is performed on adjacent transfer operators to obtain the path cascade operator. Non-commutative composition operation means applying transfer operators in order according to path order. For example, the transfer operator sequence on the path "temperature → pH → separation factor → extraction rate" is "temperature to pH transfer operator", "pH to separation factor transfer operator", and "separation factor to extraction rate transfer operator". The path cascade operator obtained by composition operation represents the comprehensive influence of temperature change on extraction rate through the path.

[0102] The operator norm of the path cascade operator is calculated as the path transfer strength. The operator norm uses the two norm, which is specifically calculated as the square root of the sum of the squares of the output change produced by applying the cascade operator to the unit input. The path transfer strength reflects the efficiency of transferring information along the path, and the greater the strength, the more significant the influence. For example, in a certain rare earth extraction and separation process, the direct path transfer strength of temperature to extraction rate is 0.44, and the indirect path transfer strength of temperature to extraction rate through pH is 0.22. The global influence degree corresponding to the operation parameter time series data is obtained by summarizing all path transfer strengths from the same source node to the same target node. The global influence degree is calculated as the weighted sum of the path transfer strengths, and the weight is inversely proportional to the path length, reflecting the comprehensive influence ability of the operation parameter on the separation performance index. For example, the global influence degree of temperature on extraction rate is 0.78, the global influence degree of pH on extraction rate is 0.65, and the global influence degree of flow rate on extraction rate is 0.32.

[0103] A disturbance field corresponding to each operating parameter in the safe operation interval is constructed based on the dynamic evolution relationship. The disturbance field describes the influence distribution of operating parameter changes on system state. The construction method is to uniformly select multiple sampling points in the safe operation interval of the operating parameter, apply a small disturbance to each sampling point, and record the system response change. For example, the safe operation interval of the temperature operating parameter is 35°C to 60°C, 9 sampling points are selected in the interval: 35°C, 38°C, 42°C, 46°C, 50°C, 54°C, 57°C and 60°C, ±1°C disturbance is applied to each sampling point, and the change of extraction rate is recorded. The influence of different operating parameters on the separation performance index under the disturbance field is calculated, and the influence is defined as the ratio of the change of the separation performance index to the disturbance of the operating parameter. For example, when 50°C is applied +1°C disturbance, the extraction rate increases by 0.025, and the influence of the point is 0.025. The maximum eigenvalue of the second-order tensor component is obtained by tensor decomposition of the influence to extract the curvature index corresponding to each operating parameter. Tensor decomposition uses high-order singular value decomposition algorithm to express the influence as a linear combination of tensors of different orders. The curvature index reflects the sensitivity change rate of the separation performance index to the change of the operating parameter, that is, the instability of the sensitivity. The greater the curvature index, the greater the difference in the influence of the operating parameter on the separation performance index under different operating conditions, and the more difficult the control. For example, the curvature index of temperature is 0.15, the curvature index of acid-base degree is 0.08, and the curvature index of flow is 0.21.

[0104] The global influence and the curvature index are fused to calculate the comprehensive importance score of each operating parameter. Fusion uses weighted average method, in which the weight of global influence is 0.7 and the weight of curvature index is 0.3. The comprehensive importance score takes into account the influence capacity and control difficulty of the operating parameter, and the higher the score, the more important the operating parameter. For example, the comprehensive importance score of temperature is 0.78x0.7+0.15x0.3=0.59, the comprehensive importance score of acid-base degree is 0.65x0.7+0.08x0.3=0.48, and the comprehensive importance score of flow is 0.32x0.7+0.21x0.3=0.29. Cluster analysis is performed on the comprehensive importance score to obtain key operating parameter clusters and redundant operating parameter clusters. Clustering uses K-means algorithm, and the number of clusters is set to 2, and the initial center point is the highest point and the lowest point. For example, in a certain extraction separation process, the comprehensive importance scores of 5 operating parameters, temperature, acid-base degree, flow, stirring speed and extractant concentration, are 0.59, 0.48, 0.29, 0.25 and 0.43, respectively. Cluster analysis is performed to obtain key operating parameter cluster {temperature, acid-base degree, extractant concentration} and redundant operating parameter cluster {flow, stirring speed}.

[0105] The operation parameters in the redundant operation parameter cluster are extracted to obtain a redundant operation parameter subset and fixed as the center value of the safe operation interval. For example, the flow is fixed as 150 mL / min, which is the center value of the safe operation interval of 100-200 mL / min, and the stirring speed is fixed as 250 rpm, which is the center value of the safe operation interval of 200-300 rpm. The key operation parameters in the key operation parameter cluster are taken as optimization variables and the corresponding safe operation interval is taken as the value constraint to construct a reduced dimension optimization space. For example, the value constraint of temperature is 35-60℃, the value constraint of pH is 1.5-4.5, and the value constraint of extractant concentration is 0.05-0.20 mol / L, and a three-dimensional optimization space is constructed.

[0106] In the embodiment, the full path from the source node to the target node is identified based on the dynamic evolution relationship, and the non-exchange composition of the propagation operators on the path is performed to obtain the path cascade operator. The multi-path and multi-level influence of the operation parameters on the performance indicators in the actual extraction separation process can be completely captured. The path propagation strength is quantified by the operator norm, which effectively reflects the nonlinear coupling characteristics and strong and weak differences between different paths, so as to obtain a more comprehensive and accurate global influence degree. The multi-dimensional index system constructed by the nonlinear influence path and the local curvature change makes the parameter clustering more stable and more accurate, which can significantly reduce the misjudgment probability. The redundant parameters are fixed as the center value of the safe interval, and only the key operation parameters are taken as optimization variables to construct a reduced dimension optimization space, which reduces the optimization dimension, improves the convergence efficiency, and also improves the robustness of the optimization result.

[0107] Figure 2 The causal topology analysis and parameter optimization flowchart of the extraction separation process modeling and optimization method based on big data analysis of the embodiment of the application.

[0108] In an optional implementation,

[0109] The separation performance time sequence trajectory corresponding to different key operation parameters is calculated in the reduced dimension optimization space. The multi-objective evaluation index is calculated based on the separation performance time sequence trajectory and the preset separation purity target and energy consumption constraint. The optimal operation parameter time sequence is determined by combining the local gradient search algorithm to obtain an operation trajectory optimization scheme, which includes:

[0110] A plurality of groups of sampling points are obtained by sampling the key operation parameters in the reduced dimension optimization space. Each group of sampling points is taken as an initial value of the key operation parameters. The separation performance time sequence trajectory is obtained by performing forward propagation from the source node corresponding to the key operation parameters to the target node along the causal topology structure and sequentially applying the propagation operator corresponding to each causal edge to the propagation signal.

[0111] The separation purity target and energy consumption deviation degree corresponding to the separation performance time sequence trajectory are calculated based on the preset separation purity target and energy consumption constraint, the separation purity target and energy consumption deviation degree are used to construct a Pareto front set in a target space and perform convex hull analysis to identify an ideal point, and the Chebyshev distance of each separation purity target and energy consumption deviation degree to the ideal point is calculated to obtain a multi-objective evaluation index;

[0112] The sampling point with the minimum multi-objective evaluation index value is selected as the initial search point, a Riemann metric tensor is constructed based on the pre-obtained curvature index, and the reduced dimension optimization space is mapped to a Riemann manifold, a geodesic line on the Riemann manifold is defined and gradient search is performed, the local curvature radius of the geodesic line is adjusted according to the curvature index to determine an adaptive search step, the multi-objective evaluation index of the current search point is calculated in each iteration until a preset convergence condition is reached, and the search point at the end of iteration is extracted as the optimal operation parameter and an operation trajectory optimization scheme is generated in combination with the time sequence change.

[0113] A plurality of groups of sampling points are obtained by sampling the key operation parameters in the reduced dimension optimization space. The sampling adopts Latin hypercube sampling technology to generate uniformly distributed sampling points in a multi-dimensional space. In a three-dimensional optimization space, the number of sampling points is set to 27, which uniformly covers the entire optimization space. For example, in a three-dimensional optimization space composed of temperature, pH and extractant concentration, the temperature range is 35-60℃, the pH range is 1.5-4.5, and the extractant concentration range is 0.05-0.20 mol / L. The generated sampling points include temperature 38℃, pH 2.0, extractant concentration 0.07 mol / L, temperature 42℃, pH 2.5, extractant concentration 0.10 mol / L, temperature 47℃, pH 3.0, extractant concentration 0.12 mol / L, etc. 27 groups of parameter values. Each group of sampling points is used as the initial value of the key operation parameters, and the forward propagation is performed from the source node corresponding to the key operation parameters to the target node along the causal topology. In the propagation process, the transfer operator corresponding to each causal edge is sequentially applied to the propagation signal. The transfer operator multiplies the input signal by the transfer coefficient and adds dynamic characteristics such as time delay and damping effect. For example, for the case of temperature changing from 38℃ to 42℃, the pH change amount is calculated to be 0.15 by the temperature-to-pH transfer operator, and the separation factor change amount is calculated to be 0.22 by the pH-to-separation factor transfer operator, and a series of separation performance parameters varying with time, i.e. a separation performance time sequence trajectory, is obtained.

[0114] The separation purity target and energy consumption constraint are preset. The separation performance time sequence trajectory is calculated. The separation purity target and energy consumption deviation are calculated. The separation purity target is defined as the ratio of the actual separation purity to the target purity. The larger the value is, the closer to or exceeding the target. The energy consumption deviation is defined as the difference between the actual energy consumption and the upper limit of the allowed energy consumption. The smaller the value is, the lower the energy consumption. For example, the target purity of a rare earth extraction separation process is 99.5%, and the upper limit of the allowed energy consumption is 250 kW·h / ton. For a sampling point temperature of 42°C, an acid-base degree of 2.5, and an extractant concentration of 0.10 mol / L, the final separation purity calculated by forward propagation is 99.2%, and the separation purity target is 99.2% divided by 99.5%, which is 0.997. The calculated energy consumption is 230 kW·h / ton, and the energy consumption deviation is 230 minus 250, which is -0.08, and the absolute value is 0.08. The Pareto front set is constructed in the target space based on the separation purity target and the energy consumption deviation. The Pareto front represents a set of solutions that cannot be further improved without compromising one target. The construction method is to plot the separation purity target and energy consumption deviation of all sampling points on a two-dimensional plane, and select non-dominated solutions to form the Pareto front. Convex hull analysis is performed to identify ideal points, which are hypothetical points with the maximum separation purity target and the minimum energy consumption deviation. For example, by analyzing the evaluation results of 27 sampling points, the Pareto front contains 8 non-dominated solutions, and the ideal point coordinates are separation purity target 1.02 and energy consumption deviation 0.05, indicating that the separation purity target is 1.02, i.e. the purity exceeds the target by 2%, and the energy consumption deviation is 0.05, i.e. the energy consumption is 5% lower than the upper limit.

[0115] The Chebyshev distance of each group of separation purity target and energy consumption deviation to the ideal point is calculated to obtain a multi-objective evaluation index. The Chebyshev distance is defined as the maximum value of the difference in each dimension, which can balance the deviation of multiple targets. For example, the evaluation result of the sampling point temperature 42°C, acid-base degree 2.5, and extractant concentration 0.10 mol / L is separation purity target 0.997 and energy consumption deviation 0.08, and the Chebyshev distance to the ideal point separation purity target 1.02 and energy consumption deviation 0.05 is calculated as the maximum value of the absolute value of the difference between 0.997 and 1.02 and the absolute value of the difference between 0.08 and 0.05, i.e. the maximum value of 0.023 and 0.03 is 0.03, i.e. the multi-objective evaluation index value is 0.03. The sampling point with the minimum multi-objective evaluation index value is selected as the initial search point. After calculating the multi-objective evaluation index of 27 sampling points, it is found that the index value of the sampling point temperature 47°C, acid-base degree 3.0, and extractant concentration 0.12 mol / L is 0.018, which is the smallest, so this point is selected as the initial search point.

[0116] The Riemannian metric tensor is constructed based on the pre-obtained curvature indicators, and the reduced dimension optimization space is mapped to a Riemannian manifold. The Riemannian metric tensor is a mathematical tool for describing the local geometric characteristics of the space. The diagonal elements are proportional to the curvature indicators of each dimension. For example, the curvature indicators of temperature, pH and extractant concentration are 0.15, 0.08 and 0.12 respectively, and the main diagonal elements of the constructed Riemannian metric tensor are in the ratio of 0.15:0.08:0.12, and the non-diagonal elements are zero, indicating that the three parameters are mutually orthogonal. The Riemannian manifold is a surface with local Euclidean structure, which can more accurately describe the distance and direction relationship in the optimization space. The geodesic line on the Riemannian manifold is defined and gradient search is performed. The geodesic line is the shortest path between two points on the Riemannian manifold, and the gradient search is performed in the direction of the fastest performance indicator decline. The local curvature radius of the geodesic line is adjusted according to the curvature indicator to determine the adaptive search step. The curvature radius is inversely proportional to the curvature indicator. The larger the curvature indicator, the smaller the curvature radius, and the smaller the search step, to ensure the stability of the search.

[0117] The multi-objective evaluation index of the current search point is calculated in each iteration until the preset convergence condition is reached. The specific search process is as follows: starting from the initial point of temperature 47℃, pH 3.0, and extractant concentration 0.12 mol / L, the gradient direction is calculated and the search step is determined. The curvature indicator of temperature is 0.15, and the corresponding basic step is 0.5℃; the curvature indicator of pH is 0.08, and the corresponding basic step is 0.1; the curvature indicator of extractant concentration is 0.12, and the corresponding basic step is 0.01 mol / L. The first iteration moves to the point of temperature 47.5℃, pH 3.1, and extractant concentration 0.13 mol / L, and the multi-objective evaluation index is calculated to be 0.015; the second iteration moves to the point of temperature 48.0℃, pH 3.2, and extractant concentration 0.14 mol / L, and the multi-objective evaluation index is calculated to be 0.012; the third iteration moves to the point of temperature 48.5℃, pH 3.2, and extractant concentration 0.15 mol / L, and the multi-objective evaluation index is calculated to be 0.011; the fourth iteration moves to the point of temperature 49.0℃, pH 3.3, and extractant concentration 0.15 mol / L, and the multi-objective evaluation index is calculated to be 0.010. When the multi-objective evaluation index changes by less than 0.002 for three consecutive iterations or the number of iterations reaches 20, the search is terminated. After the twelfth iteration, the search converges to the point of temperature 52.5℃, pH 3.4, and extractant concentration 0.16 mol / L, with a corresponding multi-objective evaluation index of 0.007, a separation purity standard of 1.015, and an energy consumption deviation of 0.07.

[0118] The search point at the end of the extraction termination iteration is taken as the optimal operation parameter, and an operation trajectory optimization scheme is generated in combination with the timing change. The optimal operation parameters, temperature 52.5 DEG C, pH 3.4, and extractant concentration 0.16 mol / L, are substituted into the causal topological structure for forward propagation to obtain the change curve of the separation performance over time. Further considering the characteristics of the process start, stable operation and shutdown and the like, a complete operation trajectory is designed. For example, the operation trajectory of the temperature parameter is: the temperature is gradually increased from the ambient temperature 25 DEG C to 52.5 DEG C at a temperature increasing rate controlled at 5 DEG C / min in the start-up stage; the temperature is maintained at 52.5 DEG C ± 0.5 DEG C in the stable operation stage; and the temperature is naturally cooled to the ambient temperature in the shutdown stage. The operation trajectory of the pH parameter is: the pH is gradually adjusted from the initial value 2.0 to 3.4 at an adjustment rate controlled at 0.2 / min in the start-up stage; the pH is maintained at 3.4 ± 0.1 in the stable operation stage; and the pH is maintained unchanged in the shutdown stage. The operation trajectory of the extractant concentration parameter is: the extractant concentration is adjusted to 0.16 mol / L in the preparation stage before the start; and the extractant concentration is maintained unchanged throughout the whole process.

[0119] In the embodiment, in the reduced dimension optimization space, multiple groups of sampling are performed based on the key operation parameters, and forward propagation is performed by using the causal topological structure, so that the influence of each sampling point can be transmitted to the separation performance index along the real causal chain, the separation performance timing trajectory which is physically consistent and reflects the non-linear dynamics is obtained, the credibility of the performance evaluation is significantly improved, the target space of the separation purity standard and the energy consumption deviation degree is constructed, and the Pareto front is calculated, so that the multi-objective compromise relationship can be systematically identified on the premise of meeting the target conflict, the problem that the target weighting needs artificial experience setting and is difficult to reflect the real trade-off of multiple targets is overcome, the ideal point is determined by convex hull analysis, and a multi-objective evaluation index is constructed by using Chebyshev distance, so that the deviation amount between different targets can be uniformly quantified, the objectivity and sensitivity of the sampling point quality evaluation are improved, the geodesic line is used for guiding the search, the step length is automatically shortened in the high-curvature area, and the step length is appropriately enlarged in the flat area, so that a high-efficiency search mechanism with reasonable direction and adaptive step length is realized, and the optimization convergence speed and global search ability are significantly improved.

[0120] In a second aspect of the embodiment of the present application, an extraction separation process modeling and optimization system based on big data analysis is provided, comprising:

[0121] A first unit is configured to acquire multi-source heterogeneous process data of an extraction separation process, perform multi-scale timing decomposition to obtain a multi-scale process state feature set, construct a causal relationship graph based on the multi-scale process state feature set, identify a direct causal path through conditional independence verification, determine a causal topological structure based on the direct causal path, and assign a mass transfer mechanism constraint weight to each causal edge.

[0122] The second unit is configured to determine an initial state based on the multi-scale process state feature set and solve a state evolution trajectory by taking each causal edge in the causal topology as a state transfer path and taking the mass transfer mechanism constraint weight as a transfer coefficient, calculate a deviation between the state evolution trajectory and the multi-source heterogeneous process data, and correct the transfer coefficient to obtain a dynamic evolution relationship.

[0123] The third unit is configured to calculate a global influence degree of each operation parameter in the dynamic evolution relationship on a separation performance index, identify a key operation parameter and a redundant operation parameter subset, fix the redundant operation parameter subset as a safe operation interval center value to obtain a reduced optimization space, calculate a separation performance time sequence trajectory corresponding to different key operation parameters in the reduced optimization space, calculate a multi-objective evaluation index based on the separation performance time sequence trajectory and a preset separation purity target and energy consumption constraint, determine an optimal operation parameter time sequence by combining a local gradient search algorithm, and obtain an operation trajectory optimization scheme.

[0124] In a third aspect, an electronic device is provided, including:

[0125] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0126] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0127] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0128] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for modeling and optimizing extraction and separation processes based on big data analysis, characterized in that, include: Multi-source heterogeneous process data of extraction and separation process are obtained and multi-scale time series decomposition is performed to obtain multi-scale process state feature set. Based on the multi-scale process state feature set, a causal relationship map is constructed and direct causal path is identified through conditional independence verification. Based on the direct causal path, the causal topology is determined and mass transfer mechanism constraint weight is assigned to each causal edge. Each causal edge in the causal topology is taken as a state transfer path, the mass transfer mechanism constraint weight is taken as a transfer coefficient, the initial state is determined based on the multi-scale process state feature set and the state evolution trajectory is solved, the deviation between the state evolution trajectory and the multi-source heterogeneous process data is calculated and the transfer coefficient is corrected to obtain the dynamic evolution relationship. The global influence of each operational parameter on the separation performance index in the dynamic evolution relationship is calculated, and key operational parameters and redundant operational parameter subsets are identified. The redundant operational parameter subset is fixed as the center value of the safe operation interval to obtain the dimensionality reduction optimization space. Within the dimensionality reduction optimization space, the separation performance time-series trajectory corresponding to different key operational parameters is calculated. Based on the separation performance time-series trajectory and the preset separation purity target and energy consumption constraints, multi-objective evaluation index is calculated. The optimal operational parameter time-series sequence is determined by combining the local gradient search algorithm to obtain the operational trajectory optimization scheme.

2. The method according to claim 1, characterized in that, The process involves acquiring multi-source heterogeneous process data from the extraction and separation process and performing multi-scale time-series decomposition to obtain a multi-scale process state feature set. Based on this multi-scale process state feature set, a causal relationship graph is constructed, and direct causal paths are identified through conditional independence verification, including: For the multi-source heterogeneous process data in the extraction and separation process, high-frequency components are extracted as fast-changing dynamic features and low-frequency components are extracted as slow-changing trend features by wavelet decomposition. The fast-changing dynamic features and the slow-changing trend features are combined to obtain a multi-scale process state feature set. The multi-source heterogeneous process data includes material component concentration data, time series data of operating parameters, and separation efficiency characterization data. The fast-changing dynamic features and slow-changing trend features are mapped to time-series nodes respectively, and the mutual information between different time-series nodes is calculated. Based on the mutual information, the association edges between the time-series nodes are established to obtain the initial causal relationship graph. For each associated edge in the initial causal relationship graph, a conditional independence test is performed. The remaining time-series nodes in the multi-scale process state feature set, excluding the two time-series nodes connected by the current associated edge, are taken as the conditional variable set. The conditional mutual information between the two time-series nodes connected by the current associated edge is calculated under the constraints of the conditional variable set. When the conditional mutual information is higher than a preset independence threshold, the current associated edge is marked as a direct causal path. The calculation and marking are repeated until all associated edges are tested, and a set of direct causal paths is obtained.

3. The method according to claim 1, characterized in that, Determining the causal topology based on the direct causal path and assigning mass transfer mechanism constraint weights to each causal edge includes: Traverse the pre-acquired set of direct causal paths, sort the operation parameter nodes corresponding to the operation parameter time series data in each direct causal path according to the causal transmission order, determine the hierarchical position of each operation parameter node in the causal transmission link based on the topology sorting result, and construct a causal topology structure based on the hierarchical position. For each causal edge in the causal topology, calculate the manifold distance between the multi-scale process state feature sets corresponding to the starting node and the ending node connected by the causal edge, and quantify the state propagation difficulty of the causal edge in the nonlinear state space based on the manifold distance. Based on the difficulty of state transmission, the influence weight of the causal edge on the extraction and separation process is calculated in combination with the topological centrality of the causal edge in the causal topology. The influence weight is then corrected for physical interpretability through mass transfer mechanism constraints to obtain and label the mass transfer mechanism constraint weight of the causal edge.

4. The method according to claim 1, characterized in that, Using each causal edge in the causal topology as a state propagation path and the mass transfer mechanism constraint weights as propagation coefficients, the initial state is determined and the state evolution trajectory is solved based on the multi-scale process state feature set, including: The mass transfer mechanism constraint weights corresponding to each causal edge are determined and organized into a state propagation matrix. The fast dynamic features and slow trend features corresponding to each node in the multi-scale process state feature set are extracted and encoded into vectors. Tensor product operation is performed to obtain the state tensor. Singular value decomposition is performed on the state tensor to determine the dominant mode and the initial state is determined based on the dominant mode. The initial state is loaded onto the nodes in the causal topology and iteratively propagated along the state propagation path using the state propagation matrix. In each iteration, the similarity between the current node and its corresponding multi-order neighboring nodes is calculated, and the receptive field range is determined. Upstream nodes within the receptive field range are extracted, and distance weights are calculated based on the topological distance from the upstream nodes to the current node. Modulation coefficients are calculated based on the distance weights and the propagation coefficients, and the state values ​​of the upstream nodes are weighted and then accumulated to the current node after transformation using a preset nonlinear activation function. In each iteration, the state value of each node is recorded, and the state value values ​​are arranged in chronological order to form a state evolution trajectory.

5. The method according to claim 1, characterized in that, Calculating the deviation between the state evolution trajectory and the multi-source heterogeneous process data, and correcting the transfer coefficient to obtain the dynamic evolution relationship includes: Extract the measured data corresponding to the nodes in the state evolution trajectory from the multi-source heterogeneous process data, align the measured data with the state evolution trajectory in time to obtain the observation sequence, calculate the point-by-point difference between the state evolution trajectory and the observation sequence at each time step and sum the squares to obtain the global bias; Calculate the partial derivative of the global deviation with respect to the transmission coefficient, determine the critical causal edge based on the absolute value of the partial derivative and construct the critical path set, calculate the gradient of the global deviation with respect to each transmission coefficient in the critical path set and determine the adjustment direction and adjustment step size of the transmission coefficient, calculate the correction amount based on the adjustment direction and the adjustment step size and apply the correction amount to the transmission coefficient to obtain the updated transmission coefficient; The updated transfer coefficients are substituted into the pre-acquired state propagation matrix, and the process of solving the state evolution trajectory is executed again to obtain the updated state evolution trajectory. The update deviation between the updated trajectory and the observed sequence is calculated. It is determined whether the decrease of the updated deviation relative to the global deviation meets the preset convergence condition. If it does, the updated transfer coefficients are used as the corrected transfer coefficients. If it does not, the updated deviation is used to replace the global deviation, and the gradient and corrected transfer coefficients are repeatedly calculated until the preset convergence condition is met. The corrected transfer coefficients are extracted and associated with the corresponding causal edges in the causal topology to obtain the dynamic evolution relationship.

6. The method according to claim 1, characterized in that, Calculate the global impact of each operational parameter on the separation performance index in the dynamic evolution relationship, identify key operational parameters and subsets of redundant operational parameters, and fix the subset of redundant operational parameters to the center value of the safe operating interval to obtain the dimensionality reduction optimization space, including: Based on the dynamic evolution relationship, identify all paths in the causal topology between the source node corresponding to the time series data of the operation parameters and the target node corresponding to the preset separation performance index. Construct a sequence of transit operators for the causal edges on each path in sequence and perform non-commutative composition operations on adjacent transit operators to obtain path concatenation operators. Calculate the operator norm of the path concatenation operators as the path transit strength. Summarize the path transit strengths of all paths from the same source node to the same target node to obtain the global influence degree corresponding to the time series data of the operation parameters. Based on the dynamic evolution relationship, a disturbance field corresponding to each operating parameter within the safe operating range is constructed, and the influence of different operating parameters on the separation performance index under the disturbance field is calculated. The influence is then decomposed into tensors to extract the maximum eigenvalue of the second-order tensor components to obtain the curvature index corresponding to each operating parameter. The comprehensive importance score of each operation parameter is calculated by integrating the global influence degree and curvature index, and cluster analysis is performed to obtain the key operation parameter cluster and the redundant operation parameter cluster. The operation parameters in the redundant operation parameter cluster are extracted to obtain the redundant operation parameter subset and fixed as the center value of the safe operation interval. The key operation parameters in the key operation parameter cluster are used as optimization variables and the corresponding safe operation interval is used as the value constraint to construct the dimensionality reduction optimization space.

7. The method according to claim 1, characterized in that, Within the dimensionality reduction optimization space, the separation performance time-series trajectory corresponding to different key operating parameters is calculated. Based on the separation performance time-series trajectory and the preset separation purity target and energy consumption constraints, a multi-objective evaluation index is calculated. The optimal operating parameter time-series sequence is determined using a local gradient search algorithm, resulting in an optimized operating trajectory scheme including: Multiple sets of sampling points are obtained by sampling the key operation parameters in the dimensionality reduction optimization space. Each set of sampling points is used as the initial value of the key operation parameter. The signal is propagated forward from the source node corresponding to the key operation parameter along the causal topology and the propagation operator corresponding to each causal edge is applied to the propagation signal in sequence to obtain the separation performance time-series trajectory. Based on the preset separation purity target and energy consumption constraint, calculate the separation purity attainment and energy consumption deviation corresponding to the separation performance time-series trajectory. Based on the separation purity attainment and energy consumption deviation, construct a Pareto front set in the target space and perform convex hull analysis to identify ideal points. Calculate the Chebyshev distance from each set of separation purity attainment and energy consumption deviation to the ideal point to obtain multi-objective evaluation indicators. The sampling point with the smallest multi-objective evaluation index value is selected as the initial search point. A Riemann metric tensor is constructed based on the pre-acquired curvature index, and the dimensionality reduction optimization space is mapped to a Riemann manifold. Geodesics on the Riemann manifold are defined and gradient search is performed. The local curvature radius of the geodesics is adjusted according to the curvature index to determine the adaptive search step size. In each iteration, the multi-objective evaluation index of the current search point is calculated until the preset convergence condition is reached. The search point at the end of the iteration is extracted as the optimal operation parameter, and an operation trajectory optimization scheme is generated by combining the temporal changes.

8. A big data analysis-based extraction and separation process modeling and optimization system for implementing the method described in any one of claims 1-7, characterized in that, include: The first unit is used to acquire multi-source heterogeneous process data of the extraction and separation process and perform multi-scale time-series decomposition to obtain a multi-scale process state feature set. Based on the multi-scale process state feature set, a causal relationship graph is constructed and direct causal paths are identified through conditional independence verification. Based on the direct causal paths, a causal topology is determined and mass transfer mechanism constraint weights are assigned to each causal edge. The second unit is used to take each causal edge in the causal topology as a state transmission path, take the mass transfer mechanism constraint weight as a transmission coefficient, determine the initial state based on the multi-scale process state feature set and solve the state evolution trajectory, calculate the deviation between the state evolution trajectory and the multi-source heterogeneous process data and correct the transmission coefficient to obtain the dynamic evolution relationship. The third unit is used to calculate the global influence of each operation parameter on the separation performance index in the dynamic evolution relationship and identify key operation parameters and redundant operation parameter subsets. The redundant operation parameter subsets are fixed as the center value of the safe operation interval to obtain the dimensionality reduction optimization space. Within the dimensionality reduction optimization space, the separation performance time-series trajectory corresponding to different key operation parameters is calculated. Based on the separation performance time-series trajectory and the preset separation purity target and energy consumption constraints, multi-objective evaluation indexes are calculated. The optimal operation parameter time-series sequence is determined by combining the local gradient search algorithm to obtain the operation trajectory optimization scheme.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images