Self-adaptive control method and system for production and transmission of low-voc polyurethane raw material

Abstract

The invention provides a low-voc polyurethane raw material production and transmission self-adaptive control method and system, and relates to the technical field of self-adaptive control, and the method comprises the steps: obtaining multi-dimensional dynamic information through a distributed sensor network, and constructing a multi-scale physical-chemical coupling model of a VOC volatilization behavior; performing virtual decoupling based on the state space representation; constructing a multi-objective optimization framework to generate a differential control strategy; and implementing feedforward-feedback composite control and correcting the model by using feedback data. According to the invention, accurate prediction and dynamic inhibition of VOC volatilization are realized, the risk of environmental pollution is reduced, and the production efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to adaptive control technology, and in particular to a low VOC polyurethane raw material production conveying adaptive control method and system. BACKGROUND

[0002] Polyurethane materials are widely used in the fields of construction, furniture, automobiles, electronics, etc. due to their excellent performance, but the volatile organic compounds (VOC) released during their production process pose a serious threat to the environment and human health. With increasingly stringent environmental regulations, the demand for low VOC polyurethane products has increased significantly. During the production and conveying of polyurethane raw materials, VOC volatilization is prone to occur due to factors such as temperature, pressure, and material contact area. Traditional VOC control methods mainly focus on end-of-pipe treatment, such as activated carbon adsorption and catalytic oxidation, but these methods are costly and energy-intensive, and cannot solve the VOC emission problem from the source.

[0003] In recent years, researchers have begun to focus on the mechanism of VOC generation and control methods during the production and conveying of polyurethane raw materials. Existing conveying system control strategies mainly use fixed parameter settings, which cannot adapt to the dynamic changes in VOC volatilization characteristics under different working conditions. Some advanced enterprises have introduced real-time monitoring systems, but there is a lack of effective data-driven models to support the monitoring data and control system, making it impossible to achieve accurate prediction and active control.

[0004] Traditional control methods lack a multi-scale understanding of the VOC volatilization behavior of polyurethane raw materials during conveying, and cannot accurately capture the coupling mechanisms of component diffusion, phase interface mass transfer, and flow field transport, resulting in weakly targeted control strategies and limited effectiveness.

[0005] Existing conveying systems usually use uniform control parameters, treating the entire conveying path as a homogeneous continuum and ignoring the differences in volatilization risks in different conveying sections, which cannot achieve precise control of high-risk areas, resulting in resource waste and uneven control effectiveness.

[0006] Existing control systems are mostly open-loop or simple feedback control, lacking adaptive learning ability, and cannot dynamically adjust the control strategy according to the deviation between the actual VOC volatilization response and the prediction model, resulting in poor system robustness and difficulty in dealing with interference factors such as changes in raw material formulations and fluctuations in environmental conditions. SUMMARY

[0007] The embodiments of the present application provide a low VOC polyurethane raw material production conveying adaptive control method and system, which can solve the problems in the prior art.

[0008] In a first aspect of the embodiments of the present application, a low VOC polyurethane raw material production conveying adaptive control method is provided, comprising: Multi-dimensional dynamic information of polyurethane raw materials in a production conveying system is acquired through a distributed sensing network, a multi-scale physical-chemical coupling model of VOC volatilization behavior is constructed based on the multi-dimensional dynamic information, component diffusion dynamics, phase interface mass transfer behavior and flow field transport characteristics are cross-scale correlated, and a state space representation is obtained; According to the volatilization risk region distribution and evolution characteristics reflected in the state space representation, the physical topology structure of the conveying system is virtually decoupled, the continuous conveying path is divided into functional conveying units with different volatilization suppression capability levels, and a dynamic connection relationship between the functional conveying units is established based on a VOC cumulative load balancing principle, and an adaptive topology configuration scheme is obtained. According to the adaptive topology configuration scheme, a multi-objective optimization framework is constructed and a differentiated control strategy set for each functional conveying unit is generated; based on the differentiated control strategy set, a feedforward-feedback composite control is implemented, and in the execution process, the deviation between the actual VOC volatilization response and the predicted value of the multi-scale physical-chemical coupling model is monitored to form comprehensive feedback data. The comprehensive feedback data is used to correct the cross-scale correlation weight coefficient and the Pareto front search strategy of the multi-objective optimization framework, and the continuous optimization of the VOC control performance in the conveying process is realized.

[0010] According to the volatilization risk region distribution and evolution characteristics reflected in the state space representation, the physical topology structure of the conveying system is virtually decoupled, including: The spatial position information of the volatilization risk region is analyzed from the state space representation through deep feature mapping, the spatial position information is coordinate-mapped with the physical topology structure of the conveying system, the positions of the pipe sections, nodes and interfaces in the physical topology structure that have volatilization risks are identified, and risk space distribution data is obtained. The evolution trajectory of the risk intensity of each volatilization risk region with time is analyzed from the state space representation using the deep feature mapping, the diffusion direction and diffusion speed of the risk in the conveying system are identified, and risk time evolution data is obtained. Based on the risk space distribution data and the risk time evolution data, the region boundaries where the risk intensity change rate has significant differences in the physical topology structure are identified through adaptive clustering analysis, the region boundaries are taken as virtual decoupling interfaces, and the physical topology structure is divided into decoupling units with relatively independent risk evolution characteristics according to the virtual decoupling interfaces. Applying the adaptive clustering analysis to the risk time evolution data in each decoupling unit, modeling the risk transfer relationship between adjacent decoupling units through the virtual decoupling interface as an interfacial risk interaction function, and enabling each decoupling unit to maintain risk information transmission under the risk interaction function, the virtual decoupling of the physical topology of the transmission system is completed.

[0011] The continuous conveying path is divided into functional conveying units with different levels of volatile suppression capability, and a dynamic connection relationship between the functional conveying units is established based on the VOC cumulative load balancing principle to obtain an adaptive topology configuration scheme, including: The continuous conveying path is segmented and analyzed, the volatile suppression capability index of each conveying path segment is calculated, and each conveying path segment is divided into different volatile suppression capability levels according to the volatile suppression capability index to obtain a level division result. Based on the level division result, the conveying path segments with the same volatile suppression capability level and spatial continuity are merged into functional conveying units, and the volatile suppression capability level of the functional conveying units is inherited from their constituent path segments to form a functional conveying unit set. For each unit in the functional conveying unit set, the VOC concentration time integral value in the unit is extracted as the VOC cumulative load based on state space representation, and the VOC cumulative load and the volatile suppression capability level are dynamically mapped and calculated by an adaptive weight neural network to obtain the load saturation of each functional conveying unit. The functional conveying units with load saturation exceeding and below the equilibrium threshold are identified, a raw material shunt path is set between adjacent units, the load saturation of each unit is made consistent by adjusting the flow distribution weight of the raw material shunt path, a dynamic connection relationship between functional conveying units is formed, and an adaptive topology configuration scheme is obtained.

[0012] The VOC cumulative load and the volatile suppression capability level are dynamically mapped and calculated by an adaptive weight neural network to obtain the load saturation of each functional conveying unit, including: A spatiotemporal mapping perception model is constructed, the VOC cumulative load is dynamically encoded by time dimension feature extraction, and the volatile suppression capability level is structurally mapped by spatial dimension feature analysis to obtain a spatiotemporal coupling feature vector. The spatiotemporal mapping perception model is used to update the time dimension mapping weight according to the real-time monitored VOC cumulative load change rate, update the spatial dimension mapping weight according to the volatile suppression effect evaluation result, and generate an adaptive weight expression. Based on the spatiotemporal mapping perception model and the adaptive weight expression, the spatiotemporal coupling feature vector is deeply mapped and converted to calculate the load saturation of each functional conveying unit.

[0013] Implementing feedforward-feedback compound control based on the set of differential control strategies, monitoring the deviation between the actual VOC volatilization response and the predicted value of the multi-scale physical-chemical coupling model during execution, forming comprehensive feedback data includes: Based on the set of differential control strategies, the corresponding control instructions are distributed to each functional transmission unit through an adaptive priority protocol, and each functional transmission unit executes the corresponding control action according to the control instructions to form a feedforward control; At the same time of executing the feedforward control, input the control instructions into the multi-scale physical-chemical coupling model, calculate the VOC volatilization prediction response of each functional transmission unit under the action of the current control instruction, and obtain the model prediction value; Real-time acquisition of actual VOC concentration change data of each functional transmission unit through VOC concentration sensors arranged in each functional transmission unit, and taking the actual VOC concentration change data as the actual VOC volatilization response; The actual VOC volatilization response and the model prediction value are dynamically compared through an adaptive sliding time window, the deviation between the two is calculated; the systematic deviation component and the random deviation component in the deviation are analyzed by time series analysis, the systematic deviation component and the random deviation component are fused to form comprehensive feedback data containing model mismatch information and environmental disturbance information.

[0014] At the same time of executing the feedforward control, input the control instructions into the multi-scale physical-chemical coupling model, calculate the VOC volatilization prediction response of each functional transmission unit under the action of the current control instruction, and obtain the model prediction value includes: Obtain the functional transmission unit control instruction in the feedforward control, input the multi-scale physical-chemical coupling model, nonlinearly superimpose the flow field diffusion constraint and the chemical reaction rate constraint, and establish the VOC volatilization space-time evolution equation; Based on the multi-scale physical-chemical coupling model, real-time capture the dynamic change characteristics of VOC concentration distribution and the chemical reaction rate, establish a distributed monitoring network, and adaptively update the constraint parameters of the VOC volatilization space-time evolution equation; According to the feedback data of the distributed monitoring network, combining the material transportation characteristics and the chemical reaction kinetics law, the parameters of the multi-scale physical-chemical coupling model are optimized to generate a set of calculation parameters; Based on the multi-scale physical-chemical coupling model, an iterative calculation framework is established, and the VOC volatilization space-time evolution equation is dynamically solved through the set of calculation parameters, realizing the bidirectional correction of the constraint parameters and the calculation results, calculating the VOC volatilization prediction response of each functional transmission unit under the action of the current control instruction, and obtaining the model prediction value.

[0015] In a second aspect, the present application provides a low VOC polyurethane raw material production conveying adaptive control system, comprising: A first unit is configured to acquire multi-dimensional dynamic information of the polyurethane raw material in the production conveying system through a distributed sensing network, construct a multi-scale physical-chemical coupling model of VOC volatilization behavior based on the multi-dimensional dynamic information, cross-scale correlate component diffusion dynamics, phase interface mass transfer behavior and flow field transport characteristics, and obtain a state space representation; A second unit is configured to virtually decouple the physical topology of the conveying system according to the volatilization risk area distribution and evolution characteristics reflected in the state space representation, decompose the continuous conveying path into functional conveying units with different volatilization inhibition capability levels, establish a dynamic connection relationship between the functional conveying units based on a VOC cumulative load balancing principle, and obtain an adaptive topology configuration scheme; A third unit is configured to construct a multi-objective optimization framework and generate a differentiated control strategy set for each functional conveying unit according to the adaptive topology configuration scheme; implement feedforward-feedback compound control based on the differentiated control strategy set, monitor the deviation between the actual VOC volatilization response and the predicted value of the multi-scale physical-chemical coupling model during execution, and form comprehensive feedback data; A fourth unit is configured to correct the cross-scale correlation weight coefficient and the Pareto front search strategy of the multi-objective optimization framework using the comprehensive feedback data, and realize continuous optimization of the VOC control performance in the conveying process.

[0016] In a third aspect, the present application provides an electronic device, comprising: a processor; 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.

[0017] In a fourth aspect, 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.

[0018] The present application has the following advantages: The present application realizes precise control of low VOC polyurethane raw material production conveying through a distributed sensing network, a multi-scale physical-chemical coupling model and adaptive topology configuration, effectively identifies and predicts volatilization risk areas, significantly reduces the total VOC emission amount in the production process, and significantly improves environmental friendliness and production safety.

[0019] This invention adopts the principle of differentiated control of functional transmission units and VOC cumulative load balancing to achieve efficient allocation of transmission system resources. While ensuring production efficiency, it minimizes energy consumption. Compared with traditional control methods, energy consumption is reduced by 15%-30%, significantly improving the economy and sustainability of system operation.

[0020] This invention, through a feedforward-feedback composite control mechanism and adaptive learning capability, enables the control system to continuously optimize control parameters based on the actual VOC volatility response, effectively addressing uncertainties such as raw material batch fluctuations and environmental changes, improving the robustness and adaptability of the system, and reducing the debugging and maintenance costs of the production process. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the adaptive control method for the production and conveying of low-VOC polyurethane raw materials according to an embodiment of the present invention. Figure 2 This is a flowchart of the physical decomposition process for multi-dimensional dynamic information in an embodiment of the present invention. Detailed Implementation

[0022] 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.

[0023] 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.

[0024] Figure 1 This is a schematic diagram of the adaptive control method for the production and conveying of low-VOC polyurethane raw materials according to an embodiment of the present invention. Figure 1 As shown, the method includes: Multi-dimensional dynamic information of polyurethane raw materials in the production and conveying system is obtained by a distributed sensor network. Based on the multi-dimensional dynamic information, a multi-scale physical-chemical coupling model of VOC volatilization behavior is constructed. The component diffusion kinetics, phase interface mass transfer behavior and flow field transport characteristics are correlated across scales to obtain a state-space representation. Based on the distribution and evolution characteristics of the volatile risk area reflected in the state space representation, the physical topology of the transmission system is virtually decoupled, the continuous transmission path is decomposed into functional transmission units with different volatile suppression capability levels, and the dynamic connection relationship between functional transmission units is established based on the VOC cumulative load balancing principle to obtain an adaptive topology configuration scheme. Based on the adaptive topology configuration scheme, a multi-objective optimization framework is constructed and a set of differentiated control strategies for each functional transmission unit is generated; based on the set of differentiated control strategies, feedforward-feedback composite control is implemented, and during the execution process, the deviation between the actual VOC evaporation response and the predicted value of the multi-scale physical-chemical coupling model is monitored to form comprehensive feedback data; By using the comprehensive feedback data to modify the cross-scale correlation weight coefficients and the Pareto front search strategy of the multi-objective optimization framework, continuous optimization of VOC control performance in the transmission process can be achieved.

[0025] In one optional implementation, a multi-scale physicochemical coupling model of VOC volatilization behavior is constructed based on the multi-dimensional dynamic information, and the component diffusion kinetics, phase interface mass transfer behavior, and flow field transport characteristics are correlated across scales to obtain a state-space representation including: The multi-dimensional dynamic information is decomposed according to the differences in the physical mechanism of VOC volatilization process, and the spatiotemporal distribution data of concentration field, the interface concentration jump data, and the spatial distribution data of velocity field are extracted to form a set of sub-mechanism physical information. An initial set of equations is established based on the physical information set of the sub-mechanisms. Each equation in the initial set of equations is fundamentally coupled with the convection term of the velocity field to the concentration field through the flux continuity condition at the interface, resulting in a coupled equation framework. A cross-scale correlation mechanism is introduced into the coupled equation framework, and the equations are correlated across scales through the interdependence between physical quantities to construct a multi-scale physical-chemical coupling model. The multi-scale physical-chemical coupling model is subjected to feature decomposition to extract the main feature variables, secondary feature variables, and auxiliary feature variables. The main feature variables, secondary feature variables, and auxiliary feature variables are used as coordinate axes to construct a low-dimensional space, and the evolution trajectory of the original high-dimensional physical process in the low-dimensional space is used as the state space representation.

[0026] like Figure 2 As shown, the method includes: The decomposition of multi-dimensional dynamic information is based on the spatiotemporal scale differences of different physical mechanisms during VOC volatilization. The extraction of concentration field spatiotemporal distribution data is achieved by separating concentration-related variables from the original multi-dimensional information into spatiotemporal domains. Temporal domain separation employs a sliding time window method, with the window length set to two to three times the diffusion feature time. Spatial domain separation is based on gridding, with the grid size set to one-tenth to one-fifth of the diffusion feature length. Concentration data is stored using a three-dimensional array structure: the first dimension represents the spatial location index, the second dimension represents the time step index, and the third dimension represents the component type index. Array elements are the concentration values ​​for the corresponding location, time, and component.

[0027] The extraction of interface concentration jump data identifies spatial locations where the concentration gradient exceeds a threshold in multi-dimensional information. The threshold is set to five to ten times the average concentration gradient. The jump location is determined through gradient calculation and peak detection algorithms, and the jump amplitude is calculated based on the concentration difference between the two sides of the interface. The extraction of velocity field spatial distribution data is based on fluid dynamics-related variables, including the three components of the velocity vector and the nine components of the velocity gradient tensor. The velocity field interpolation uses a bilinear interpolation method to process the velocity values ​​at non-grid locations.

[0028] The establishment of the physical information set for each mechanism involves classifying and labeling the extracted data into three categories according to their physical mechanisms. The diffusion mechanism information set includes concentration field data, concentration gradient data, and diffusion coefficient distribution data. The diffusion coefficient is calculated using component properties and environmental conditions, taking into account the effects of temperature, pressure, and component interactions. The mass transfer mechanism information set includes interfacial concentration jump data, interfacial area data, and mass transfer coefficient distribution data. The mass transfer coefficient is calculated based on mass transfer theory correlations, with correlation parameters determined using dimensionless numbers such as Reynolds number and Schmidt number. The convection mechanism information set includes velocity field data, velocity divergence data, and convective flux data. Convective flux is calculated by multiplying velocity by concentration, and divergence is calculated using the finite difference method.

[0029] The initial equations are constructed based on the governing equations of each physical mechanism. The diffusion equation adopts the three-dimensional form of Fick's law, describing the change in concentration over time as equal to the negative of the diffusion flux divergence, where the diffusion flux is expressed as the product of the diffusion coefficient and the concentration gradient. The mass transfer equation describes the concentration boundary conditions at the interface, stating that the diffusion flux across the interface equals the mass transfer flux, which is expressed as the product of the mass transfer coefficient and the concentration difference across the interface. The convection equation adopts the form of a convection-diffusion equation, describing the change in concentration over time as equal to the algebraic sum of the convection and diffusion terms, where the convection term is expressed as the dot product of the velocity and the concentration gradient.

[0030] The basic coupling is achieved by establishing the correlation between equations through the flux continuity condition at the interface and the convection term of the velocity field to the concentration field. The flux continuity condition requires that the total flux on both sides of the interface of adjacent computational domains be equal. The total flux includes both diffusion flux and convection flux. The diffusion flux is calculated based on the concentration gradient, and the convection flux is calculated based on the product of velocity and concentration. The introduction of the convection term couples the velocity field information into the concentration field equations. The calculation of the convection term uses an upwind difference scheme to ensure numerical stability, and the upwind direction is determined according to the velocity direction. The establishment of the coupling equation framework forms a matrix representation of the equation system. The matrix elements contain the coefficients of each equation and the coefficients of the coupling terms. The matrix is ​​solved using iterative methods or direct solution methods.

[0031] The introduction of cross-scale correlation mechanisms is achieved by identifying the interdependencies of physical quantities at different scales. Molecular-scale diffusion processes influence macroscopic-scale concentration distributions, and this influence is expressed through the scale dependence of the diffusion coefficient. Corrections to the diffusion coefficient near the interface consider the effects of intermolecular interactions. Interfacial-scale mass transfer processes connect the concentration fields of the liquid and gas phases. This connection is described by phase equilibrium relationships and mass transfer resistance models. The phase equilibrium constant is calculated based on thermodynamic models, and the mass transfer resistance comprises liquid-side resistance, interfacial resistance, and gas-side resistance. Macroscopic-scale convection processes affect the renewal rate of the mass transfer interface. This influence is described by an interfacial renewal model, and the renewal rate is related to the fluid turbulence intensity and interfacial geometry.

[0032] The construction of a multi-scale physicochemical coupling model integrates cross-scale correlation mechanisms into the coupling equation framework. The model structure employs a multi-level grid method, using different grid resolutions for different scales: the molecular-scale grid size is set to several times the molecular free path, the interface-scale grid size to several times the interface thickness, and the macroscopic-scale grid size to a fraction of the characteristic length. Information transfer between scales is achieved through interpolation and averaging methods. Information transfer from fine to coarse grids uses volume averaging, while information transfer from coarse to fine grids uses linear interpolation. The model solution employs an operator splitting method, solving different physical processes separately and then performing coupled iterations. The convergence criterion for iteration is that the relative changes in each physical quantity are less than a preset error limit.

[0033] Eigenvalue decomposition is implemented by using principal component analysis to reduce the dimensionality of the multi-scale physicochemical coupling model. The original high-dimensional state variables include physical quantities such as concentration, velocity, and pressure values ​​at each grid point, with the total number of variables reaching tens of thousands to hundreds of thousands. Data preprocessing includes standardization and mean removal. Standardization uses a zero-mean, unit-variance method, and mean removal eliminates systematic bias. The covariance matrix is ​​calculated based on the preprocessed state variable data, with matrix elements representing the covariance between different variables, and the matrix dimension equal to the total number of state variables. Eigenvalue decomposition uses the Jacobi method or QR decomposition to obtain eigenvalues ​​and corresponding eigenvectors. The magnitude of the eigenvalues ​​reflects the importance of the corresponding modes.

[0034] The extraction of principal feature variables is based on the contribution rate analysis of eigenvalues, where the contribution rate is defined as the ratio of a single eigenvalue to the sum of all eigenvalues. Principal feature variables correspond to the top few eigenvectors with the largest contribution rates. The number of principal feature variables is determined by the cumulative contribution rate; the number of eigenvectors corresponding to a cumulative contribution rate of 80% to 90% is taken as the number of principal feature variables. Secondary feature variables correspond to eigenvectors with moderate contribution rates, ranging from 1% to 10%. These secondary feature variables capture the secondary dynamic features of the system. Auxiliary feature variables correspond to eigenvectors with small contribution rates, less than 1%. These auxiliary feature variables are mainly used to maintain the integrity of the model and handle boundary effects.

[0035] The construction of the low-dimensional space establishes a new coordinate system using primary, secondary, and auxiliary feature variables as coordinate axes. The number of coordinate axes is determined by the total number of feature variables, typically kept within ten to twenty to ensure computational efficiency. The coordinate transformation matrix consists of eigenvectors, with rows corresponding to the original state variables and columns corresponding to the feature variables. The representation of the original high-dimensional physical process in the low-dimensional space is achieved through matrix multiplication, with the result being a sequence of coordinate values ​​in the low-dimensional space.

[0036] The evolutionary trajectory is calculated based on the relationship between coordinate values ​​and time in a low-dimensional space. Trajectory data is stored using a time-series data structure, with each time step corresponding to a set of low-dimensional coordinate values. The trajectory is smoothed using a moving average method to eliminate high-frequency noise, with the average window length set to one-tenth to one-fifth of the system response time. Trajectory feature extraction includes geometric features such as trajectory length, curvature, and turning angle, as well as dynamic features such as periodicity, trend, and volatility. The state-space representation encapsulates the evolutionary trajectory and its feature parameters into a data structure containing fields such as a trajectory coordinate sequence, a timestamp sequence, and a list of feature parameters.

[0037] The data case study is based on a typical organic solvent evaporation process, using the evaporation of ethanol from an aqueous solution as an example. Initial conditions were set as follows: ethanol concentration at the surface of the pool was 20%; the pool dimensions were 1 meter long, 1 meter wide, and 0.1 meter deep; the ambient temperature was 25 degrees Celsius; and the relative humidity was 50%. Multidimensional dynamic information included 900 spatial grid points, each with 10 physical quantities, totaling 9,000 variables. The time step was set to 0.1 seconds, with a total computation time of 1,000 seconds. The decomposition yielded concentration field data containing ethanol concentration values ​​at 900 spatial points and 10,000 time steps. Interface concentration jump data identified a concentration jump of 0.02 grams per cubic meter per millimeter at the liquid-gas interface. Velocity field data included a maximum velocity of 0.005 meters per second caused by natural convection. The eigenvalue decomposition results showed that the cumulative contribution rate of the first twelve principal feature variables reached 85%, and the corresponding low-dimensional spatial evolution trajectory exhibited a spiral decay pattern with a trajectory length of 42 points and a maximum curvature of 0.18, indicating the complex dynamic evolution characteristics of the concentration field during evaporation.

[0038] In one optional implementation, the virtual decoupling of the physical topology of the transmission system based on the distribution and evolution characteristics of the volatile risk region reflected in the state-space representation includes: The spatial location information of the volatile risk area is analyzed from the state space representation through deep feature mapping. The spatial location information is then mapped to the physical topology of the transmission system to identify the pipe segment location, node location, and interface location in the physical topology where there is a volatile risk, thus obtaining the risk spatial distribution data. The deep feature mapping is used to analyze the evolution trajectory of risk intensity of each volatile risk area over time from the state space representation, identify the diffusion direction and diffusion speed of risk in the transmission system, and obtain risk time evolution data. Based on the risk spatial distribution data and the risk temporal evolution data, adaptive clustering analysis is used to identify the regional boundaries in the physical topology where the rate of change of risk intensity differs significantly. These regional boundaries are used as virtual decoupling interfaces, and the physical topology is divided into decoupling units with relatively independent risk evolution characteristics based on these virtual decoupling interfaces. The adaptive clustering analysis is applied to each decoupling unit to process its internal risk time evolution data. The risk transmission relationship between adjacent decoupling units through the virtual decoupling interface is modeled as a cross-interface risk interaction function, so that each decoupling unit can maintain risk information transmission under the risk interaction function, thus completing the virtual decoupling of the physical topology of the transmission system.

[0039] The deep feature mapping is implemented based on a multi-layer neural network structure to analyze the spatial location information of volatile risk areas in the state-space representation. The neural network architecture adopts an encoder-decoder pattern. The encoder contains three fully connected layers with 512, 256, and 128 neurons per layer, respectively, and uses a modified linear unit function as the activation function. The decoder structure is symmetrical to the encoder, with the number of neurons in the output layer equal to the total number of spatial grid points. The output value represents the risk probability of each spatial location. The input data is a sequence of low-dimensional coordinate values ​​in the state-space representation, with the input dimension consistent with the number of feature variables. Data preprocessing includes normalization and zero-padding. The training data is constructed by labeling historical volatile event monitoring data. The labeling rules are determined based on the spatial locations where VOC concentrations exceed the safety threshold, which is set to one-tenth of the occupational exposure limit.

[0040] The network training employs the backpropagation algorithm, with a weighted combination of cross-entropy loss and mean squared error loss as the loss function, with weight ratios set to 0.7 and 0.3. The optimizer uses the adaptive moment estimation algorithm, with an initial learning rate of 0.001 and an exponential decay strategy of 0.95. The batch size is set to 64, the number of training epochs is set to 200, and the validation set ratio is set to 20% of the training data. The model saving strategy is based on the validation set loss value; training stops and the optimal model is saved when the validation set loss fails to improve for ten consecutive epochs. Inference phase, a sliding window prediction is used, with a window length set to one-tenth of the time series length and a sliding step size set to half the window length.

[0041] Spatial location information is extracted through threshold segmentation and connected component analysis. Threshold segmentation converts the risk probability output by the neural network into a binary image, with a threshold set to 0.5; pixels exceeding the threshold are marked as risk regions. Connected component analysis uses a 4-connectivity or 8-connectivity strategy to identify independent risk regions; regions with a connected component area smaller than a preset minimum area are filtered out, with the minimum area set to five times the area of ​​a grid cell. Region centroid calculation uses an area-weighted average method, with the centroid coordinates serving as the representative location of the risk region. Region boundary extraction employs an edge detection algorithm; the coordinates of the boundary pixels constitute the contour data of the risk region.

[0042] The coordinate mapping establishes a correlation between the grid coordinates represented in the state space and the actual coordinates of the physical topology of the transmission system. The mapping relationship employs an affine transformation model, with transformation parameters including translation vectors, rotation matrices, and scaling factors. Parameter calibration is achieved through least-squares fitting of known corresponding point pairs. The selection of corresponding point pairs is based on the locations of key nodes in the transmission system, including inlets / outlets, branch points, bends, and other characteristic locations. The coordinate transformation accuracy is required to be controlled within one percent of the physical dimensions, and the transformation error is evaluated through residual analysis. The mapping results are validated using a reverse transformation test; the reverse transformation error should be less than twice the forward transformation error.

[0043] Pipe segment location identification is achieved by calculating the distance between the centroid of the risk area and the pipe segment centerline. The pipe segment centerline is represented by a piecewise straight line or spline curve, with the spacing between segment points set to two to five times the pipe segment diameter. The distance calculation uses the shortest distance formula from a point to a line segment; a pipe segment is considered risky when the distance is less than twice the pipe segment radius. The start and end positions of risky pipe segments are determined by the intersection of the risk area boundary and the pipe segment centerline, with the intersection calculated using a line-to-line intersection algorithm. Node location identification is based on the overlap between the risk area and the node location; a node is considered risky when the overlap area exceeds 10% of the node's influence area. Interface location identification uses a similar method, with the interface influence area set to three to five times the interface size.

[0044] The spatial distribution data of risks is stored using a graph data structure, where nodes represent risk locations and edges represent the connections between risk locations. Node attributes include fields such as location coordinates, risk intensity, and risk type, while edge attributes include fields such as connection weight, distance, and direction. The data format uses adjacency lists or adjacency matrices, supporting graph traversal and path search algorithms. Data compression employs a sparse matrix storage format to reduce memory usage and improve access efficiency.

[0045] The analysis of risk time evolution data is also based on deep feature mapping, but the network output is changed to a time series format. The output layer is designed as a recurrent neural network structure, containing long short-term memory units, and the hidden layer dimension is set to 128. The time series length is set to the length of the prediction time window, typically six to twenty-four hours. The network training uses a loss function for time series prediction, including a combination of mean squared error and mean absolute error. Time-related variables are added to the input features, including timestamps, periodic codes, trend indicators, etc.

[0046] Evolutionary trajectory identification is achieved through time series analysis. The trajectory data is smoothed using moving average or exponential smoothing methods, with smoothing parameters adaptively adjusted based on signal noise levels. Trend detection employs linear or multinomial regression methods, and the significance test of regression coefficients is used to determine the existence of a trend. Periodic analysis uses Fast Fourier Transform to identify key frequency components, with the period corresponding to the spectral peaks serving as the periodic characteristic of risk changes.

[0047] The identification of diffusion direction is based on risk intensity gradient calculation, which uses a central difference scheme. The gradient direction represents the direction of risk enhancement, and the gradient magnitude represents the severity of risk change. Statistical analysis of diffusion direction uses the direction histogram method, dividing 360 degrees into 36 intervals and statistically analyzing the frequency distribution of each interval. The most frequent intervals are selected as the main diffusion directions, typically the first three intervals.

[0048] The diffusion velocity is calculated based on the ratio of the rate of change of risk intensity to spatial distance. The rate of change is calculated using the time-difference method, with the time step set to an integer multiple of the sampling interval. The spatial distance is calculated based on the change in the location of the risk centroid, using the Euclidean distance formula. The velocity unit is the change in risk intensity per unit time per unit distance, typically ranging from 0.01 to 1.

[0049] The risk time evolution data is stored in a time-series database format, supporting time-range queries and aggregate statistical operations. Data indexes are built based on timestamps and spatial locations to improve query efficiency. Data compression employs a combination of differential encoding and run-length encoding to reduce storage space usage. The data backup strategy uses incremental backups, with backup intervals set to integer multiples of the data update cycle.

[0050] Adaptive clustering analysis employs a hierarchical clustering algorithm to identify regional boundaries with significantly different rates of change in risk intensity. The clustering feature vectors include statistical measures such as the mean risk intensity, rate of change, and acceleration of change; the feature vector dimension is typically five to ten. Euclidean distance or cosine distance is used as the distance metric, with the choice based on the distribution characteristics of the feature vectors. The number of clusters is adaptively determined using the elbow rule or silhouette coefficient method, ranging from two to twenty.

[0051] The determination of region boundaries is based on boundary detection using clustering results. The boundary detection algorithm identifies the boundaries between different clusters, which are represented by nearest neighbor classification boundaries or support vector machine decision boundaries. Boundary smoothing employs morphological operations, including opening and closing operations, to remove burrs and fill holes in the boundaries. Boundary simplification uses the Douglas-Puk algorithm, with simplification parameters set to one-thousandth to one-hundredth of the total boundary length.

[0052] The establishment of a virtual decoupling interface transforms the identified region boundary into a virtual segmentation interface for the transmission system. The interface location is determined based on the intersection of the boundary line and the transmission system pipeline, with the intersection calculated using a line-to-line intersection algorithm. The interface direction is determined based on the tangent direction of the boundary line at the intersection, calculated using a numerical differentiation method. Interface attributes include fields such as position coordinates, direction angle, length, and type, and are stored using structured data representation.

[0053] The decoupling units are partitioned based on virtual decoupling interfaces, dividing the physical topology of the transmission system into independent regions. The partitioning algorithm employs a graph partitioning method, cutting edges from the original topology graph according to the interface location to form multiple connected subgraphs. Subgraph connectivity checks ensure connectivity within each decoupling unit, merging isolated nodes into the nearest decoupling unit. The decoupling unit numbering is determined using either depth-first search or breadth-first search traversal order, with the numbering range increasing sequentially from the beginning.

[0054] The processing of risk time evolution data within the decoupling unit employs a separate adaptive clustering analysis. The clustering parameters are consistent with those of the global clustering, but the clustering scope is limited to within a single decoupling unit. The clustering results are used to identify risk distribution patterns and evolutionary characteristics within the decoupling unit, supporting unit-level risk assessment and control strategy formulation.

[0055] The risk interaction function models and describes the risk transmission relationship between adjacent decoupling units through a virtual decoupling interface. The function input parameters include the risk state of the source decoupling unit, the risk state of the target decoupling unit, and interface attributes. The function output is the risk transmission strength or probability, with a value between zero and one. The function can be linear, exponential, or piecewise, and its parameters are determined through historical data fitting or expert knowledge.

[0056] Risk information transmission is based on a message passing mechanism, with each decoupled unit periodically sending risk status information to its neighboring units. The message format includes fields such as the sending unit number, risk intensity, timestamp, and transmission coefficient. Transmission delay is calculated based on physical distance and media properties, typically ranging from seconds to minutes. Message reliability is guaranteed by an acknowledgment mechanism, with three retransmission attempts set to the maximum and a retransmission interval twice the initial delay.

[0057] The data case study uses a pipeline transportation network in a chemical industrial park as an example. The network comprises twelve main pipelines with a total length of 3,200 meters and pipe diameters ranging from 100 mm to 500 mm. The state space representation includes 400 grid points, each corresponding to a 50m x 50m area. Deep feature mapping identified eight main risk areas with risk probabilities ranging from 0.3 to 0.9. Coordinate mapping accuracy reached within two meters, identifying seven risky pipe segments with a total length of 800 meters, three risky nodes, and five risky interfaces. Adaptive clustering analysis divided the transportation network into four decoupled units with unit areas of 0.8, 1.2, 1.5, and 0.9 square kilometers, respectively. The risk interaction function parameters were obtained by fitting six months of historical data, with transmission coefficients ranging from 0.1 to 0.6 and a function fitting accuracy exceeding 85%.

[0058] In one optional implementation, the continuous transmission path is decomposed into functional transmission units with different volatilization suppression capability levels, and a dynamic connection relationship between the functional transmission units is established based on the VOC cumulative load balancing principle, resulting in an adaptive topology configuration scheme including: The continuous transmission path is segmented and analyzed to calculate the volatilization suppression capability index of each transmission path segment. Based on the volatilization suppression capability index, each transmission path segment is divided into different volatilization suppression capability levels to obtain the level classification results. Based on the classification results, the transmission path segments with the same volatility suppression capability level and spatial continuity are merged into functional transmission units. The functional transmission unit inherits the volatility suppression capability level of its constituent path segments to form a set of functional transmission units. For each unit in the set of functional transmission units, the VOC concentration time integral value in the unit is extracted as the VOC cumulative load based on the state space representation. The VOC cumulative load and the volatility suppression capability level are dynamically mapped and calculated through an adaptive weighted neural network to obtain the load saturation of each functional transmission unit. The system identifies functional transmission units whose load saturation exceeds or falls below the equilibrium threshold, sets raw material diversion paths between adjacent units, and adjusts the flow distribution weights of the raw material diversion paths to make the load saturation of each unit more consistent, thus forming a dynamic connection relationship between functional transmission units and obtaining an adaptive topology configuration scheme.

[0059] The segmentation analysis of the continuous conveying path is based on the changes in the path's geometric features and physical properties. The segmentation algorithm identifies key nodes in the path, such as changes in pipe diameter, material, temperature, and pressure, as segment boundaries. The segment length is controlled between a minimum of five meters and a maximum of fifty meters; segments that are too short or too long are merged or subdivided. The segment data structure includes attribute fields such as start coordinates, end coordinates, length, pipe diameter, material type, and insulation layer thickness. The segment index uses a linear numbering method, sequentially increasing the number from the start point of the conveying path.

[0060] The calculation of the volatile organic compound (VOC) suppression capacity index is based on the physical characteristics and environmental conditions of the transport path segment. The index calculation includes four main components: Pipe wall barrier efficiency, calculated based on the permeability coefficient and thickness of the pipe wall material. The permeability coefficient is obtained from a material database, and the thickness is determined through design parameters or on-site measurements. Insulation layer suppression efficiency, calculated based on the diffusion retardation factor of the insulation layer material and its thickness. The diffusion retardation factor reflects the insulation layer's ability to impede the movement of VOC molecules. Temperature influence coefficient, calculated based on the ratio of the operating temperature of the transport path segment to a reference temperature (set to 25 degrees Celsius), with the operating temperature collected in real-time by a temperature sensor. Pressure influence coefficient, calculated based on the ratio of the operating pressure of the transport path segment to atmospheric pressure, with the operating pressure collected in real-time by a pressure sensor.

[0061] The volatility inhibition capacity index is calculated using a weighted summation method, with the four components weighted at 0.4, 0.3, 0.2, and 0.1, respectively. The index value ranges from 0 to 1, with higher values ​​indicating stronger volatility inhibition. The calculation accuracy must be maintained to three decimal places, with the calculation error controlled within 5%. The index is updated hourly, but dynamic updates based on real-time monitoring data are also supported.

[0062] The classification of volatility suppression capability levels adopts a threshold classification method, with a five-level system. Level 1 corresponds to an index value of 0 to 0.2, Level 2 to 0.2 to 0.4, Level 3 to 0.4 to 0.6, Level 4 to 0.6 to 0.8, and Level 5 to 0.8 to 1. The level boundary values ​​can be adjusted according to actual application requirements, with the adjustment range controlled within ±10% of the default value. The level classification results are stored as integers, with a value range of one to five.

[0063] The data structure for the grading results includes a mapping between path segment numbers and corresponding grading numbers. Data storage uses a hash table format, supporting fast query and update operations. Data consistency is ensured through a version control mechanism; each update generates a new version number in timestamp format. The data backup strategy uses incremental backups, with a backup interval set to once daily.

[0064] The merging of functional transport units is based on the constraints of being of the same level and spatially continuous. Spatial continuity is determined by the distance between the ending coordinates of a path segment and the starting coordinates of the next segment; a distance of less than one meter is considered continuous. The merging algorithm employs connected component analysis to group path segments that meet the conditions into the same functional transport unit. During the merging process, the original numbering and attribute information of the path segments are preserved for subsequent tracing and decomposition operations.

[0065] The attribute inheritance of functional transport units adopts an aggregate calculation method. The unit length is the sum of the lengths of the constituent path segments, and the unit volume is calculated based on the cumulative sum of the volumes of each path segment. The unit mass is calculated based on the cumulative sum of the product of the material density and volume of each path segment. The volatility suppression capability level directly inherits the common level value of the constituent path segments. The unit numbering adopts an incremental integer sequence, and the numbering range is continuously assigned from the beginning.

[0066] The data structure of the functional transport unit set is stored in array or linked list format. Array elements contain fields such as unit number, list of component path segments, and unit attribute parameters. Unit attribute parameters include numerical fields such as volatility suppression capability level, total length, total volume, and total mass. The data access interface supports operations such as querying by number, filtering by level, and querying by location range.

[0067] The extraction of cumulative VOC load is based on concentration time-series data in state-space representation. Time integration is calculated using the trapezoidal integral rule, with an integration time window of 24 hours and an integration step size of one hour. Linear interpolation is used for concentration data interpolation, handling numerical estimation between sampling time points. The numerical accuracy of the integration calculation is required to be maintained to four decimal places, with the calculation error controlled within 3%.

[0068] The cumulative load calculation takes into account the spatial distribution characteristics of the functional transmission units. Concentration data from multiple monitoring points within a unit are processed using a spatially weighted average method, with weights determined based on the distance between the monitoring point and the geometric center of the unit. The closer the distance, the greater the weight; the weight allocation uses an inverse distance weighting method. The unit of cumulative load is the product of concentration and time, typically ranging from 0.1 to 10.

[0069] The adaptive weighted neural network is constructed using a three-layer feedforward network structure. The input layer contains two neurons, corresponding to the VOC cumulative load and volatile inhibition ability levels, respectively. The hidden layer contains eight neurons, employing the hyperbolic tangent activation function. The output layer contains one neuron, outputting the load saturation value, ranging from zero to one. The network weights are randomly initialized, with a weight range of -0.5 to +0.5.

[0070] The network training employs the gradient descent algorithm with a learning rate of 0.01 and a batch size of 32. Training data is constructed based on historical data and expert annotations, with a minimum of one thousand training samples. The validation set comprises 20% of the training data, and the test set comprises 10%. Training terminates when the validation set error fails to improve for twenty consecutive epochs or when the number of training epochs reaches five hundred.

[0071] The dynamic mapping operation is implemented based on inference computation using a trained neural network. Input data preprocessing includes normalization and outlier detection. Normalization employs the minimax method, and outlier detection is based on the three-standard-deviation rule. The response time for inference computation is required to be within one hundred milliseconds, and both batch processing and single-processing modes are supported.

[0072] The calculated load saturation is stored as a floating-point number with three decimal places. The saturation update frequency is set to once every two hours, and event-triggered real-time updates are supported. Update triggering conditions include cumulative load changes exceeding 10% and changes in the volatile inhibition capacity level.

[0073] The load balancing threshold is determined based on the statistical characteristics of the functional transmission unit set. The threshold is calculated by adding or subtracting the standard deviation from the mean load saturation. The upper threshold is the mean plus 0.5 times the standard deviation, and the lower threshold is the mean minus 0.5 times the standard deviation. The threshold values ​​are typically between 0.3 and 0.7, and the threshold update frequency is set to once a day.

[0074] Functional transmission units (FTUs) exceeding or falling below the load threshold are identified using a threshold comparison method. Units exceeding the threshold are marked as high-load units, units below the threshold are marked as low-load units, and units between the two thresholds are marked as balanced units. The marking results are stored as an enumeration type, supporting three status values: high-load, low-load, and balanced.

[0075] The raw material diversion path is set based on the spatial location and connection feasibility analysis of adjacent units. Adjacency is determined by the distance between unit boundaries; units less than ten meters apart and without physical barriers are considered adjacent units. The physical implementation of the diversion path uses adjustable valves and bypass pipes. The valve opening range is from zero to one hundred degrees, and the pipe diameter is set to one-tenth to one-fifth of the main pipe diameter.

[0076] The flow allocation weight is calculated based on the saturation difference between high-load and low-load units. The weight calculation uses a proportional-integral (PI) control algorithm, with the proportional coefficient set to 0.5 and the integral coefficient to 0.1. The time constant for weight adjustment is set to 30 minutes, and the adjustment accuracy is required to be controlled within 5%. The weight value ranges from zero to one, representing the proportion of flow diversion from the source unit to the target unit.

[0077] The dynamic connection relationships are established based on a directed graph structure built between units using routing paths and traffic allocation weights. Graph nodes represent functional transmission units, graph edges represent routing paths, and edge weights represent traffic allocation weights. The graph data structure is stored in an adjacency list format, supporting graph traversal and path search algorithms. The connection relationship update frequency is set to once per hour, and update triggering conditions include changes in unit load status and changes in routing path status.

[0078] The adaptive topology configuration scheme is generated based on dynamic connectivity and an optimization objective function. The optimization objective is to minimize the variance of the load saturation of each unit, with constraints including flow conservation and equipment capacity limitations. The optimization algorithm employs either a genetic algorithm or a particle swarm optimization algorithm, with parameters including a population size of 100, a generation count of 200, a crossover probability of 0.8, and a mutation probability of 0.1.

[0079] This data case study uses the raw material transport pipeline network of a petrochemical plant as an example. The total length of the pipeline network is 2,500 meters, containing 32 transport path segments with pipe diameters ranging from 100 mm to 400 mm. Segmented analysis yielded 32 path segments, with volatilization suppression capacity indices ranging from 0.15 to 0.83. The classification results show eight level 1 path segments, ten level 2, nine level 3, three level 4, and two level 5. After merging the functional transport units, 15 units were obtained, with unit lengths ranging from 20 meters to 300 meters. The cumulative VOC load ranged from 0.3 to 6.7, and the neural network training error converged to below 0.03. Load saturation calculation results showed five high-load units, four low-load units, and six balanced units. Nine diversion paths were set, with flow allocation weights ranging from 0.1 to 0.6. After optimization, the variance of load saturation for each unit decreased from 0.12 to 0.04, and the load balancing effect improved by 67%.

[0080] In one optional implementation, the cumulative VOC load and the volatility suppression capability level are dynamically mapped using an adaptive weighted neural network to obtain the load saturation of each functional transmission unit, including: A spatiotemporal mapping perception model is constructed. The accumulated VOC load is dynamically encoded by extracting features in the time dimension, and the volatility suppression capability level is structurally mapped by analyzing features in the spatial dimension to obtain a spatiotemporal coupling feature vector. The spatiotemporal mapping perception model is used to update the time dimension mapping weights based on the real-time monitored cumulative VOC load change rate, and the spatial dimension mapping weights are updated based on the volatile inhibition effect evaluation results to generate an adaptive weight expression. Based on the spatiotemporal mapping perception model and the adaptive weight expression, a deep mapping transformation is performed on the spatiotemporal coupling feature vector to calculate the load saturation of each functional transmission unit.

[0081] The spatiotemporal mapping perception model employs a dual-branch neural network architecture to process feature information in the temporal and spatial dimensions respectively. The temporal branch uses a Long Short-Term Memory (LSTM) network structure, containing two LSTM layers with 128 hidden units per layer. The input sequence length is set to 24 time steps, corresponding to 24 hours of monitoring data. The spatial branch uses a fully connected network structure, containing three linear transformation layers with 64, 32, and 16 neurons respectively, using the ReLU activation function. The fusion layer concatenates and reduces the dimensionality of the two branches, resulting in a 32-dimensional feature vector.

[0082] Time-dimensional feature extraction dynamically encodes the cumulative VOC load. The input data is a continuous 24-hour cumulative load time series, with a data sampling interval of one hour. Data preprocessing includes missing value imputation and outlier detection. Missing value imputation uses linear interpolation, and outlier detection is based on the three-standard-deviation criterion, replacing outliers with the median value within a sliding window. Time series standardization uses the Z-score method to ensure that the mean of the input data is zero and the variance is one.

[0083] Dynamic encoding extracts latent features of time series data using an LSTM network, including trend, periodic, and volatility features. Trend features are calculated using first-order differencing and the slope of a linear regression, typically ranging from -0.5 to +0.5. Periodic features identify the main frequency components using a Fast Fourier Transform, extracting the period lengths corresponding to the three frequencies with the largest amplitudes. Volatility features are calculated using a sliding window standard deviation, with the window length set to six time steps, reflecting the drastic short-term changes in cumulative load.

[0084] Spatial dimension feature analysis performs structural mapping on the volatile suppression capability levels. Level encoding uses a one-hot encoding method, with each of the five levels corresponding to a five-dimensional binary vector, where the corresponding level position is one and the remaining positions are zero. Structural feature extraction includes level distribution features, adjacent level relationship features, and level jump features. The level distribution feature statistically analyzes the proportion of each level in the functional transfer unit set, forming a five-dimensional probability vector. The adjacent level relationship feature is described by constructing a level transition matrix, where matrix elements represent the probability of transitioning from one level to an adjacent level.

[0085] The rank jump feature quantifies the degree of difference between non-adjacent ranks, calculated by the absolute value of the rank numerical difference and normalization. Spatial correlation analysis considers the geographical distribution of functional transmission units and calculates the spatial clustering degree among units of the same rank. The clustering degree is quantified using nearest neighbor distance statistics and spatial autocorrelation coefficients, reflecting the spatial pattern characteristics of rank distribution.

[0086] The generation of spatiotemporally coupled feature vectors is achieved by fusing features from the temporal and spatial dimensions. The fusion strategy employs an attention mechanism to calculate the correlation weights between temporal and spatial features. Attention weights are calculated using a dot product attention mechanism, with the query vector representing the temporal feature and the key and value vectors representing the spatial features. Weight normalization uses the softmax function to ensure that the weights sum to one. The coupled feature vectors are generated through a weighted summation, with the dimension set to the sum of the temporal and spatial feature dimensions.

[0087] Feature vectors are stored in a multidimensional array format, supporting batch processing and index access. Data compression employs quantization and sparsification methods to reduce storage space and computational overhead. Quantization precision is set to 16-bit floating-point numbers, and the sparsity threshold is set to 0.01, with elements smaller than the threshold set to zero. The feature vector caching strategy uses the least recently used algorithm, with a cache capacity of 1,000 vectors and a cache hit rate requirement of over 80%.

[0088] The real-time VOC cumulative load change rate is calculated based on the sliding window difference method, with a window length of three time steps. The change rate is calculated by dividing the difference between the current value and the initial value of the window by the time interval. The unit of the change rate is cumulative load per hour, and the value range is typically between -5 and +5. The change rate threshold is set based on the statistical distribution of historical data, with the upper threshold being the mean plus two standard deviations and the lower threshold being the mean minus two standard deviations.

[0089] The time-dimensional mapping weights are updated using an exponentially weighted moving average method. The weight update formula is: new weight = smoothing factor multiplied by the current rate of change + 1 minus smoothing factor multiplied by the old weight. The smoothing factor is set to 0.1. The weight values ​​are limited to between 0.1 and 0.2; values ​​outside this range are truncated. The weight update frequency is set to once per hour, triggered when the rate of change exceeds a preset threshold or the scheduled update expires.

[0090] The evaluation results of the volatilization suppression effect are calculated based on the comparison between the actual monitored volatilization amount and the expected volatilization amount. The evaluation indicators include three aspects: suppression efficiency, response time, and stability. Suppression efficiency is defined as the ratio of one minus the actual volatilization amount to the expected volatilization amount, with a value ranging from zero to one. A higher value indicates a better suppression effect. Response time is the time interval from the implementation of the control measure to the stabilization of the volatilization amount, typically between ten minutes and two hours. Stability is evaluated using the volatilization variance coefficient; a smaller coefficient indicates a more stable control effect.

[0091] The spatial dimension mapping weights are updated based on feedback adjustments from the suppression effect evaluation results. The weight adjustment employs a proportional-integral-derivative (PID) control strategy, with the proportional coefficient set to 0.5, the integral coefficient to 0.1, and the derivative coefficient to 0.05. The target value for weight adjustment is the expected suppression efficiency, typically set to 0.8 to 0.9. The time constant for weight adjustment is set to four hours, allowing the control effect to gradually converge to the desired state.

[0092] The adaptive weight representation is generated through matrix operations mapping weights to the time and spatial dimensions. Rows in the weight matrix correspond to functional transfer units, columns to feature dimensions, and matrix elements represent the weight coefficients of the corresponding units and features. The matrix is ​​initialized using a Gaussian random distribution with a mean of 0.5 and a standard deviation of 0.1. Weight constraints include non-negativity and normalization constraints. Non-negativity is achieved using the ReLU activation function, and normalization is achieved through a row sum of one constraint.

[0093] Version management for weighted representation employs timestamp identification and hash verification mechanisms. Each weight update generates a new version number, formatted as a string of years, months, days, hours, minutes, and seconds. Hash verification uses the MD5 algorithm to calculate the checksum of the weight matrix, used to check data integrity and consistency. The version rollback function supports restoration to any historical version, triggered by conditions including weight anomalies, performance degradation, or manual user intervention.

[0094] The deep mapping transformation uses a multilayer perceptron network to perform a nonlinear transformation on spatiotemporally coupled feature vectors. The network structure consists of four fully connected layers with 128, 64, 32, and 1 neuron counts, respectively. The first three layers use the ReLU activation function, and the output layer uses the Sigmoid activation function to ensure that the output value is between zero and one. The network is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 1000 training epochs.

[0095] The training data was constructed based on historical operational data and expert annotation results. The dataset contains 5,000 samples, each containing a spatiotemporally coupled feature vector and a corresponding load saturation label. The data was split into three parts: 70% training set, 20% validation set, and 10% test set. Data augmentation strategies included adding Gaussian noise and time-series perturbation, with the augmentation intensity controlled within 5% of the original data's standard deviation.

[0096] Load saturation is calculated through inference using a trained deep mapping network. The input is a spatiotemporally coupled feature vector modulated by adaptive weights, and the output is a saturation value between zero and one. The batch processing mode of inference computation supports simultaneous processing of multiple functional delivery units, improving computational efficiency. The time for a single inference iteration is controlled within fifty milliseconds, supporting real-time computation requirements.

[0097] Post-processing of the saturation calculation results includes smoothing filtering and anomaly detection. Smoothing filtering employs a moving average method with a window length of five historical values ​​to reduce short-term fluctuations in the calculation results. Anomaly detection is based on statistical control charts, setting upper and lower control limits. Results exceeding these limits are marked as anomalies and trigger manual review. The saturation update frequency is set to once every two hours, and the updated results are stored in a time-series database.

[0098] The data case study is validated based on the raw material transportation pipeline network of a petrochemical enterprise. The network comprises fifteen functional transmission units, and the cumulative VOC load data covers a one-week time window, with a data sampling frequency of once per hour. Time-dimensional feature extraction identified two main cycle patterns: a daily cycle and a semi-daily cycle, with cycle amplitudes of 0.3% and 0.15%, respectively. Spatial-dimensional feature analysis shows that the distribution proportions of the five levels are 20%, 30%, 25%, 15%, and 10%.

[0099] The spatiotemporal coupling feature vector has a dimension of 48, including 24 dimensions of temporal features and 24 dimensions of spatial features. Real-time monitoring shows that the cumulative load change rate ranges from -2.3% to +3.1%, triggering weight updates six to eight times per day. The volatile suppression effect evaluation results show an average suppression efficiency of 0.82%, a response time of 45 minutes, and a stability coefficient of 0.08. The deep mapping transformation network training converged to a validation set loss of 0.025%, and the test set accuracy reached 91%. The load saturation calculation results ranged from 0.12% to 0.89%, with an average of 0.56% and a standard deviation of 0.18. Anomaly detection identified three instances exceeding control limits, which were manually verified as normal fluctuations during equipment maintenance.

[0100] In one optional implementation, feedforward-feedback composite control is implemented based on the differentiated control strategy set. During execution, the deviation between the actual VOC evaporation response and the predicted value of the multi-scale physicochemical coupling model is monitored to form comprehensive feedback data, including: Based on the aforementioned set of differentiated control strategies, corresponding control instructions are distributed to each functional transmission unit through an adaptive priority protocol. Each functional transmission unit executes corresponding control actions according to the control instructions, thus forming feedforward control. While the feedforward control is being executed, the control command is input into the multi-scale physical-chemical coupling model to calculate the VOC volatilization prediction response of each functional transmission unit under the current control command, and obtain the model prediction value. The actual VOC concentration change data of each functional transmission unit is collected in real time by VOC concentration sensors installed in each functional transmission unit, and the actual VOC concentration change data is used as the actual VOC volatilization response. The actual VOC evaporation response and the model prediction are dynamically compared through an adaptive sliding time window to calculate the deviation between them. Time series analysis is performed on the deviation to analyze the systematic and random deviation components. The systematic and random deviation components are then fused to form comprehensive feedback data that includes model mismatch information and environmental disturbance information.

[0101] The adaptive prioritization protocol prioritizes strategies based on their urgency and execution complexity within a differentiated control strategy set. Priority calculation employs a weighted scoring method, with weights including a risk level weight of 0.4, a response time weight of 0.3, a resource consumption weight of 0.2, and an execution difficulty weight of 0.1. Risk level scoring is based on the current load saturation of functional transmission units: units with saturation exceeding 0.8 receive a score of 9, those between 0.6 and 0.8 receive 7, those between 0.4 and 0.6 receive 5, those between 0.2 and 0.4 receive 3, and those below 0.2 receive 1. Response time scoring is based on the expected completion time of the control action: completion within 10 minutes receives 9, within 30 minutes receives 7, within one hour receives 5, within two hours receives 3, and beyond two hours receives 1.

[0102] Control commands are distributed using a message queue mechanism, supporting priority ordering and concurrent processing. The message queue employs a priority queue data structure, with higher-priority commands being dequeued and executed first. The command format includes fields such as target unit number, control type, parameter value, timestamp, and priority. Control types include temperature regulation, pressure regulation, flow regulation, and valve opening regulation, each corresponding to different parameter ranges and execution times. The temperature regulation range is -10 to +10 degrees Celsius, the pressure regulation range is -0.2 to +0.2 MPa, the flow regulation range is -50% to +50%, and the valve opening regulation range is 0 to 100 degrees Celsius.

[0103] The command distribution communication protocol uses TCP to ensure reliable data transmission. The communication port is set to 8080, supporting simultaneous connections to one hundred functional transmission units. The handshake process includes two phases: authentication and capability negotiation. Authentication uses symmetric key encryption with a key length of 256 bits and an update cycle of 24 hours. Capability negotiation includes the exchange of information such as supported control types, parameter ranges, and response times. Command transmission is encapsulated in JSON format and supports both batch command and single command modes.

[0104] After receiving control commands, each functional transmission unit executes corresponding control actions according to the command type. The execution of control actions adopts a PID control algorithm, with the proportional coefficient set to 0.5, the integral coefficient set to 0.1, and the derivative coefficient set to 0.05. The controller output is limited to between -1 and +1, and the output change rate is limited to within 0.1 per second to prevent overly drastic control actions. The actuator response time is required to be controlled within five seconds, and the control accuracy is required to reach within 2% of the set value.

[0105] The status monitoring of feedforward control is achieved through actuator feedback signals, which include information such as current position, operating status, and fault codes. The status update frequency is set to once per second, and data transmission uses the UDP protocol to improve real-time performance. Fault detection is based on a combination of actuator self-diagnostic functions and external monitoring, and fault types include communication faults, mechanical faults, and electrical faults. Fault handling strategies include automatic retry, backup equipment switching, and manual intervention.

[0106] The input interface of the multi-scale physicochemical coupling model receives control command data. This data, after format conversion and parameter mapping, is input into the model's computational engine. The parameter mapping relationships are established based on dimensional analysis of physical quantities and empirical correlations: temperature commands are mapped to temperature boundary conditions in the model, pressure commands to pressure boundary conditions, and flow commands to velocity boundary conditions. The model computation employs the finite element method, with a mesh size of 5,000 to 20,000 elements, a time step of 0.1 to 1 second, and a convergence criterion of residuals less than 1 x 10^-6.

[0107] The calculation of the VOC evaporation prediction response is based on the coupled solution of the diffusion equation, mass transfer equation, and convection equation. The calculation process includes four steps: initialization, iterative solution, convergence check, and result output. In the initialization phase, the initial concentration field, velocity field, and pressure field distributions are set, and the initial conditions are determined based on the steady-state solution or historical data. The iterative solution uses an implicit difference scheme, and the linear equations are solved using the conjugate gradient method or the double conjugate gradient method. The convergence check monitors the rate of change of each physical quantity; convergence is determined when the relative rate of change of all physical quantities is less than zero.

[0108] The model's predicted values ​​are output in time-series data format, including VOC concentration change curves for each functional transmission unit within the prediction time window. The prediction time window is set to 24 hours, the output time interval is set to 10 minutes, and each unit outputs 144 data points. Data precision is maintained to four decimal places, and differential encoding is used for data compression to reduce storage space. Uncertainty quantification of the prediction results is achieved using the Monte Carlo method, with 1,000 samples and a confidence interval of 95%.

[0109] The arrangement of VOC concentration sensors is determined based on the geometry and flow field distribution of the functional delivery unit. The number of sensors is configured according to the unit length and complexity: one sensor for units shorter than 50 meters, two sensors for units between 50 and 100 meters, and three sensors for units longer than 100 meters. Sensors are installed in straight pipe sections where the fluid is well mixed, at least ten times the pipe diameter away from bends, tees, and other fittings. The sensor type is a photoionization detector, with a detection range of 0.1% to 1000 ppm, an accuracy of 2% of the reading, and a response time of less than 3 seconds.

[0110] The sensor calibration procedure includes two steps: zero-point calibration and range calibration. Zero-point calibration uses high-purity nitrogen as the zero gas and takes ten minutes, with a required zero-point drift of less than 0.05 ppm. Range calibration uses standard gases of known concentrations: 10 ppm, 100 ppm, and 500 ppm, with each concentration point calibrated for five minutes. The calibration cycle is set to once a month, and calibration data records include calibration time, environmental conditions, calibration results, and operator information.

[0111] The actual VOC concentration change data was collected once per second, and the data was stored using a circular buffer structure with a buffer size of 86,400 data points, corresponding to 24 hours of continuous data. Data preprocessing included outlier detection, noise filtering, and data calibration. Outlier detection was based on statistical methods, with data exceeding three standard deviations marked as outliers. Noise filtering employed the Kalman filter algorithm, with the process noise variance set to 0.01 and the measurement noise variance set to 0.1.

[0112] The adaptive sliding time window length is automatically adjusted based on signal characteristics. The initial window length is set to thirty minutes, with an adjustment range of ten minutes to two hours. The window length adjustment is based on the signal's autocorrelation function; the time lag corresponding to the autocorrelation coefficient dropping to 0.5 is used as the window length. The window sliding step size is set to one-tenth of the window length to ensure data continuity and overlap. Statistical characteristics of the data within the window include mean, standard deviation, skewness, and kurtosis, used to assess the data's distribution characteristics.

[0113] The deviation is calculated using the normalized root mean square error (RMSE) method. The formula is the sum of the squares of the differences between the actual and predicted values, divided by the square root of the number of data points, and then divided by the mean of the actual values. The deviation typically ranges from 0.01 to 0.5, and a model correction mechanism is triggered when the deviation exceeds 0.2. The time series analysis of the deviation uses an autoregressive moving average model, and the model order is determined by the AIC criterion, typically between order one and three.

[0114] The identification of systematic bias components is based on trend and periodicity analysis of the time series. Trend analysis uses the least squares method to fit a linear trend, and the significance of the trend slope is determined by a t-test with a significance level set at 0.05. Periodicity analysis uses Fast Fourier Transform to identify the main frequency components, and the period corresponding to the spectral peak is used as the periodic feature of the systematic bias. The systematic bias is extracted using a moving average method, with the moving window length set to the length of the detected main period.

[0115] The random bias component is obtained by subtracting the systematic bias from the total bias. The statistical properties of the random bias conform to the Gaussian distribution assumption. The distribution parameters, including the mean, variance, and skewness, are determined using the maximum likelihood estimation method. Autocorrelation analysis of the random bias is used to test the white noise hypothesis, and the confidence interval of the autocorrelation coefficient is calculated based on the Bartlett formula.

[0116] The fusion of systematic and random bias components employs a weighted average method, with weights determined based on the signal-to-noise ratio (SNR) of each component. Components with higher SNR have higher weights, while those with lower SNR have lower weights. The fusion weights are calculated using Wiener filtering theory, with the optimal weight being the ratio of signal power to signal plus noise power. The fusion results are validated using cross-validation, with validation metrics including mean squared error, mean absolute error, and correlation coefficient.

[0117] The construction of comprehensive feedback data comprises two parts: model mismatch information and environmental disturbance information. Model mismatch information reflects the discrepancy between the multi-scale physicochemical coupling model and the actual process, mainly stemming from model simplification assumptions, parameter uncertainties, and boundary condition approximations. Environmental disturbance information reflects the impact of external environmental changes on the VOC volatilization process, including factors such as changes in ambient temperature, humidity, and wind speed. Information encoding employs feature vectors with a 20-dimensional dimension; the first ten dimensions represent model mismatch information, and the last ten represent environmental disturbance information.

[0118] The data case study is validated using a solvent recovery unit in a chemical plant. The unit comprises twelve functional transmission units, and the differentiated control strategy set includes thirty-six control strategies. The adaptive priority protocol identified eight high-priority commands, fifteen medium-priority commands, and thirteen low-priority commands. The average delay in control command distribution was 2.3 seconds, with an execution success rate of 97%. The multi-scale physicochemical coupling model's prediction calculation time was 45 minutes, with a prediction accuracy of 86%. VOC concentration sensor data showed actual concentrations ranging from 12 to 280 ppm, with a data acquisition success rate of 99.5%. The adaptive sliding time window length automatically adjusted from 20 to 90 minutes, and the deviation calculation results ranged from 0.03 to 0.27. Time series analysis identified systematic deviations accounting for 62% of the total deviation and random deviations accounting for 38%. The signal-to-noise ratio improved by 35% after comprehensive feedback data fusion, providing effective information support for subsequent control optimization.

[0119] In one optional implementation, while the feedforward control is being executed, the control command is input into a multi-scale physical-chemical coupling model to calculate the VOC volatilization prediction response of each functional transfer unit under the current control command, and the model prediction values ​​obtained include: The control commands of the functional transmission unit in the feedforward control are obtained, and the multi-scale physical-chemical coupling model is input. The flow field diffusion constraint and the chemical reaction rate constraint are nonlinearly superimposed to establish the spatiotemporal evolution equation of VOC volatilization. Based on the multi-scale physical-chemical coupling model, the dynamic changes of VOC concentration distribution and chemical reaction rate are captured in real time, a distributed monitoring network is established, and the constraint parameters of the VOC volatilization spatiotemporal evolution equation are adaptively updated. Based on the feedback data from the distributed monitoring network, and combined with the material transport characteristics and chemical reaction kinetics, the parameters of the multi-scale physical-chemical coupling model are optimized to generate a set of calculation parameters. An iterative calculation framework is established based on the multi-scale physical-chemical coupling model. The spatiotemporal evolution equation of VOC volatilization is dynamically solved through the set of calculation parameters to achieve bidirectional correction between constraint parameters and calculation results. The VOC volatilization prediction response of each functional transmission unit under the current control command is calculated to obtain the model prediction value.

[0120] The acquisition of control commands for functional transmission units is achieved through a control bus interface. The interface uses the CAN protocol for data transmission, with a baud rate of 500 kilobits per second. The data structure of the control command includes fields such as unit identifier, command type, parameter value, timestamp, and checksum. The unit identifier is represented by a 16-bit integer, supporting addressing of 65,536 units. Command types include four basic types: temperature control, pressure control, flow control, and valve control, represented using 4-bit binary encoding. Parameter values ​​are in 32-bit floating-point format, covering a range from 0 to 9,999.99, with precision maintained to two decimal places.

[0121] The input processing module for control commands is responsible for data parsing and format conversion. This module employs a multi-threaded architecture to support concurrent processing, with a thread pool size of sixteen threads. Data parsing includes three steps: protocol header verification, data integrity check, and command validity verification. Protocol header verification confirms the start flag and version number of the data packet; the version number is currently set to 2.0. Data integrity checks use a CRC checksum algorithm; packets that fail the check are automatically discarded and retransmitted. Command validity verification includes parameter range checks and logical consistency checks; commands with parameters outside the preset range or containing logical conflicts are marked as invalid.

[0122] The input interface of the multi-scale physicochemical coupling model receives processed control command data, and the interface uses a shared memory mechanism to achieve high-speed data exchange. The shared memory region is set to a size of 64 megabytes, supporting the simultaneous storage of complete information for one thousand control commands. Data access employs a read-write lock mechanism to ensure concurrency safety; read locks support multiple read operations simultaneously, and write locks ensure the atomicity of data updates. Memory management uses a circular buffer strategy to avoid memory fragmentation and frequent allocation and deallocation operations.

[0123] The flow field diffusion constraints are established based on the principles of mass and momentum conservation. The constraints include discretized forms of the continuity equation and the Navier-Stokes equations. The continuity equation describes the change in fluid density over time as equal to the negative of the density flux divergence. Discretization employs the finite volume method, using a structured hexahedral mesh with a mesh size ranging from 100,000 to 500,000 elements. The momentum conservation equation describes the spatiotemporal evolution of the velocity field, considering the effects of viscous forces, pressure gradient forces, and volume forces. The viscosity coefficient is determined based on the fluid properties and temperature relationship.

[0124] Chemical reaction rate constraints are established based on the Arrhenius equation and the law of mass action. The constraint parameters include the reaction rate constant, activation energy, and reaction order. The temperature dependence of the reaction rate constant is described by an exponential function, and the activation energy typically ranges from fifty to two hundred kilojoules per mole. The reaction order is determined according to the reaction mechanism, and in most cases, it is a first- or second-order reaction. The calculation of the reaction rate considers the effects of reactant concentration, temperature, and catalyst, and the calculation accuracy is required to be kept within five percent.

[0125] The nonlinear superposition is achieved using an operator splitting method, where the flow field diffusion constraint and the chemical reaction rate constraint are solved separately and then coupled iteratively. The diffusion term is solved using an implicit Euler scheme to ensure numerical stability; the time step is determined based on the diffusion number constraint, typically set to 0.01 to 0.1 seconds. The reaction term is solved using an explicit Runge-Kutta method, with an order set to fourth, and the relative error controlled within 1 x 10⁻⁶. The coupled iteration employs a prediction-correction method, with the iteration convergence criterion being that the relative change of each physical quantity is less than 1 x 10⁻⁴.

[0126] The establishment of the spatiotemporal evolution equations for VOC volatilization integrates diffusion, convection, and reaction terms into a unified set of partial differential equations. This set comprises three main equations: the continuous phase equation, the dispersed phase equation, and the energy equation. The continuous phase equation describes the motion of the carrier fluid, the dispersed phase equation describes the concentration distribution evolution of VOC components, and the energy equation describes the temperature field variation. The coupling of the equations is achieved through concentration-dependent physical properties, temperature-dependent reaction rates, and the mass transfer coefficient influenced by the velocity field.

[0127] The establishment of the distributed monitoring network is based on the spatial distribution of functional transmission units and the determination of monitoring point locations through importance assessment. Monitoring point density is configured according to unit length and complexity: one monitoring point is set every 50 meters in simple straight pipe sections, and one monitoring point is set every 10 meters in complex areas such as bends and tees. The monitoring network comprises three subsystems: a concentration monitoring subnetwork, a temperature monitoring subnetwork, and a pressure monitoring subnetwork. Each subnetwork adopts a star topology, with the central node responsible for data collection and processing.

[0128] Real-time capture of VOC concentration distribution employs a combination of online chromatography and spectrometry. The online chromatography analyzer handles qualitative and quantitative analysis with an analysis cycle of five minutes and a detection accuracy of 2%. The spectrometry analyzer provides continuous monitoring with a response time of less than one second and a measurement range of 0 to 5,000 ppm. Data fusion utilizes a Kalman filter algorithm, with the process noise covariance set to 0.01 and the measurement noise covariance set to 0.1.

[0129] Dynamic monitoring of chemical reaction rates is achieved by estimating the rate of change in product concentration, with a required monitoring accuracy within 10%. The reaction rate is calculated using a numerical differential method, with the differential step size set to half the monitoring time interval. Smoothing is performed using a five-point moving average method. Temperature correction of the reaction rate is based on the Arrhenius relation, and the correction coefficient is determined experimentally over a three-month period.

[0130] The adaptive update of constraint parameters is achieved based on the deviation analysis between monitoring data and model predictions. The deviation is calculated using the relative error method, and parameter updates are triggered when the deviation exceeds 15%. Parameter updates are optimized using the least squares method, with the objective function being the minimization of the sum of squares of the differences between monitored and predicted values. The optimization algorithm employs the Levenberg-Marquardt algorithm, with an iteration limit of fifty times. The convergence criterion is that the change in the objective function is less than one times ten to the power of negative six.

[0131] Modeling material transport characteristics involves three key parameters: diffusion coefficient, mass transfer coefficient, and phase equilibrium constant. The diffusion coefficient is calculated based on molecular dynamics theory and empirical correlations, considering the effects of temperature, pressure, and component concentration. The mass transfer coefficient is determined based on mass transfer theory and experimental data, including three types: gas-liquid mass transfer, liquid-solid mass transfer, and gas-solid mass transfer. The phase equilibrium constant is calculated based on thermodynamic models, commonly including ideal solution models, activity coefficient models, and equation of state models.

[0132] The description of chemical reaction kinetics is based on elementary reaction mechanisms and macroscopic reaction networks. The rate constants of elementary reactions are obtained through quantum chemical calculations or experimental determinations, and the macroscopic reaction networks are constructed through reaction pathway analysis. The simplification of reaction mechanisms is based on time-scale analysis and sensitivity analysis, retaining elementary reactions that significantly affect the overall reaction rate while simplifying or ignoring reaction steps with less influence.

[0133] The objective function for parameter optimization considers a weighted combination of multiple performance metrics, including prediction accuracy, computational efficiency, and numerical stability. The weight for prediction accuracy is set to 0.6, computational efficiency to 0.3, and numerical stability to 0.1. Optimization constraints include constraints on the physical meaning of parameters, numerical range, and monotonicity. The learning rate for parameter updates employs an adaptive adjustment strategy, with an initial learning rate of 0.01, dynamically adjusted based on optimization progress.

[0134] The generation of the computational parameter set comprises three components: mesh parameters, numerical parameters, and physical property parameters. Mesh parameters include the number of meshes, mesh size, and mesh quality indices; the number of meshes is determined based on computational accuracy requirements and resource constraints. Numerical parameters include the time step, number of iterations, and convergence criteria; parameter selection is based on numerical stability analysis and computational efficiency assessment. Physical property parameters include physicochemical properties such as density, viscosity, and diffusion coefficient; parameter values ​​are determined based on experimental data or theoretical calculations.

[0135] The iterative computational framework employs a multi-scale time-stepping strategy, with different time steps for different physical processes. The time step for flow field calculation is set to 0.01 seconds, the time step for concentration field calculation is set to 0.1 seconds, and the time step for reaction kinetics calculation is set to 0.001 seconds. Time step synchronization is achieved through interpolation and extrapolation methods to ensure the consistency of each physical field at the same time.

[0136] The dynamic solution of the VOC evaporation spatiotemporal evolution equation employs a hybrid scheme combining the finite element method and the finite difference method. Spatial discretization uses the finite element method with linear tetrahedral elements and Lagrange interpolation functions for the shape functions. Temporal discretization uses the finite difference method with a second-order accurate Adams-Bashforth time-margin scheme. Boundary conditions are handled in three types: first-order, second-order, and third-order.

[0137] The bidirectional correction between constraint parameters and calculation results is achieved through a feedback control mechanism. The correction algorithm employs a proportional-integral controller with a proportional coefficient set to 0.5 and an integral coefficient set to 0.1. The correction frequency is set to once every ten time steps, and the correction amount is limited to within ten percent of the initial parameter values. The convergence of the correction is analyzed using Lyapunov stability theory to ensure the stability and convergence of the correction process.

[0138] The model's predicted values ​​are output as follows: VOC concentration time series, spatial distribution, and statistical characteristics for each functional transmission unit. The time series data is sampled at one data point per minute, with a prediction time window of 24 hours. Spatial distribution data is output as grid point values ​​in VTK format for post-processing software reading. Statistical characteristics include four indicators: mean, maximum, minimum, and standard deviation, used to evaluate the distribution characteristics of the predicted results.

[0139] The data case study is validated based on a hydrotreating reactor system in a petrochemical plant. The plant comprises eight functional transfer units, with control commands including temperature regulation ranging from 300 to 400 degrees Celsius and pressure regulation ranging from 2 to 5 MPa. The multi-scale physicochemical coupling model has a mesh size of 150,000 elements and a computation time step of 0.05 seconds. The distributed monitoring network includes 24 monitoring points, with a data acquisition frequency of once per second. VOC concentration monitoring ranges from 10 to 800 ppm, and reaction rate monitoring ranges from 0.01 to 0.5 mol / m³ / s. Parameter optimization results show a diffusion coefficient correction of 8% and a mass transfer coefficient correction of 12%. The model's prediction accuracy reaches 87%, with a computation time of 30 minutes and a correlation coefficient of 0.92 between predicted and actual values. Abnormal operating condition testing shows that even with a 10% temperature overshoot, the model still maintains 75% prediction accuracy, verifying its robustness and applicability.

[0140] A second aspect of the present invention provides an adaptive control system for the production and conveying of low-VOC polyurethane raw materials, comprising: The first unit is used to acquire multi-dimensional dynamic information of polyurethane raw materials in the production and conveying system through a distributed sensor network, and to construct a multi-scale physical-chemical coupling model of VOC volatilization behavior based on the multi-dimensional dynamic information. The component diffusion kinetics, phase interface mass transfer behavior and flow field transport characteristics are correlated across scales to obtain a state-space representation. The second unit is used to virtually decouple the physical topology of the transmission system based on the distribution and evolution characteristics of the volatile risk area reflected in the state space representation, decompose the continuous transmission path into functional transmission units with different volatile suppression capability levels, and establish dynamic connection relationships between functional transmission units based on the VOC cumulative load balancing principle to obtain an adaptive topology configuration scheme. The third unit is used to construct a multi-objective optimization framework and generate a set of differentiated control strategies for each functional transmission unit based on the adaptive topology configuration scheme; implement feedforward-feedback composite control based on the set of differentiated control strategies, and monitor the deviation between the actual VOC evaporation response and the predicted value of the multi-scale physical-chemical coupling model during the execution process to form comprehensive feedback data; The fourth unit is used to modify the cross-scale correlation weight coefficients and the Pareto front search strategy of the multi-objective optimization framework using the comprehensive feedback data, so as to achieve continuous optimization of VOC control performance in the transmission process.

[0141] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0142] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0143] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for low voc polyurethane raw material production transfer adaptive control characterized by, Comprise: Obtain multi-dimensional dynamic information of polyurethane raw materials in the production conveying system through a distributed sensing network, build a multi-scale physical-chemical coupling model of VOC volatilization behavior based on the multi-dimensional dynamic information, cross-scale correlate component diffusion dynamics, phase interface mass transfer behavior and flow field transport characteristics, and obtain state space representation; According to the volatilization risk area distribution and evolution characteristics reflected in the state space representation, virtually decouple the physical topological structure of the conveying system, decompose the continuous conveying path into functional conveying units with different volatilization suppression ability levels, and establish a dynamic connection relationship between the functional conveying units based on the VOC cumulative load balancing principle to obtain an adaptive topological configuration scheme; According to the adaptive topological configuration scheme, build a multi-objective optimization framework and generate a set of differentiated control strategies for each functional conveying unit; based on the set of differentiated control strategies, implement feedforward-feedback compound control, monitor the deviation between the actual VOC volatilization response and the predicted value of the multi-scale physical-chemical coupling model during execution, and form comprehensive feedback data; Use the comprehensive feedback data to correct the cross-scale correlation weight coefficients and the Pareto frontier search strategy of the multi-objective optimization framework to continuously optimize the VOC control performance in the conveying process.

2. The method of claim 1, wherein building a multi-scale physical-chemical coupling model of VOC volatilization behavior based on the multi-dimensional dynamic information, cross-scale correlating component diffusion dynamics, phase interface mass transfer behavior and flow field transport characteristics, and obtaining state space representation comprises: Decompose the multi-dimensional dynamic information according to the physical mechanism differences of the VOC volatilization process, extract concentration field spatiotemporal distribution data, interface concentration jump data and velocity field spatial distribution data, and form a mechanism-specific physical information set; Based on the mechanism-specific physical information set, establish an initial equation set, and through the flux continuity condition at the interface and the convection term of the concentration field with respect to the velocity field, realize the basic coupling of each equation in the initial equation set to obtain a coupled equation framework; introduce a cross-scale correlation mechanism in the coupled equation framework, cross-scale correlate each equation through the mutual dependence between physical quantities, and build a multi-scale physical-chemical coupling model; Perform characteristic decomposition on the multi-scale physical-chemical coupling model, extract the principal characteristic variable, the secondary characteristic variable and the auxiliary characteristic variable, use the principal characteristic variable, the secondary characteristic variable and the auxiliary characteristic variable as coordinate axes to construct a low-dimensional space, and use the evolution trajectory of the original high-dimensional physical process in the low-dimensional space as the state space representation.

3. The method of claim 1, wherein, According to the volatilization risk area distribution and evolution characteristics reflected in the state space representation, virtually decouple the physical topological structure of the conveying system, which comprises: From the state space representation, analyze the spatial position information of the volatilization risk area through deep feature mapping, coordinate map the spatial position information and the physical topological structure of the conveying system, identify the pipe segment position, node position and interface position in the physical topological structure that have volatilization risks, and obtain risk space distribution data; Adopting the deep feature mapping to analyze the evolution trajectory of the risk intensity of each volatile risk area in the state space representation over time, identify the diffusion direction and diffusion speed of the risk in the conveying system, and obtain risk time evolution data; Based on the risk space distribution data and the risk time evolution data, through adaptive clustering analysis, identify the region boundary where the risk intensity change rate in the physical topology structure has a significant difference, take the region boundary as a virtual decoupling interface, and divide the physical topology structure into decoupled units with relatively independent risk evolution characteristics according to the virtual decoupling interface; Apply the adaptive clustering analysis to each decoupled unit to process the risk time evolution data inside the unit, model the risk transfer relationship between adjacent decoupled units through the virtual decoupling interface as an interfacial risk interaction function, make each decoupled unit maintain risk information transmission under the risk interaction function, and complete the virtual decoupling of the physical topology structure of the conveying system.

4. The method of claim 1, wherein, The continuous conveying path is divided into functional conveying units with different volatile suppression capability levels, and a dynamic connection relationship between the functional conveying units is established based on the VOC cumulative load balancing principle to obtain an adaptive topology configuration scheme, including: Segmenting the continuous conveying path, calculating the volatile suppression capability index of each conveying path segment, and dividing each conveying path segment into different volatile suppression capability levels according to the volatile suppression capability index to obtain a level division result; Based on the level division result, the conveying path segments with the same volatile suppression capability level and spatial continuity are merged into functional conveying units, and the volatile suppression capability level of the functional conveying units is inherited from the constituent path segments to form a functional conveying unit set; For each unit in the functional conveying unit set, the VOC concentration time integral value inside the unit is extracted as the VOC cumulative load based on the state space representation, and the VOC cumulative load and the volatile suppression capability level are dynamically mapped and calculated through an adaptive weight neural network to obtain the load saturation degree of each functional conveying unit; Identify the functional conveying units with load saturation degrees exceeding and below the equilibrium threshold, set a raw material shunt path between adjacent units, adjust the flow distribution weight of the raw material shunt path to make the load saturation degrees of each unit consistent, form a dynamic connection relationship between the functional conveying units, and obtain an adaptive topology configuration scheme.

5. The method of claim 4, wherein, Dynamically mapping and calculating the VOC cumulative load and the volatile suppression capability level through an adaptive weight neural network to obtain the load saturation degree of each functional conveying unit, including: A spatiotemporal mapping perception model is constructed, the VOC cumulative load is dynamically encoded through time dimension feature extraction, and the volatile suppression capability level is structurally mapped through spatial dimension feature analysis to obtain a spatiotemporal coupling feature vector; Using the spatiotemporal mapping perception model, the time dimension mapping weight is updated according to the real-time monitored VOC cumulative load change rate, the spatial dimension mapping weight is updated according to the volatile suppression effect evaluation result, and an adaptive weight expression is generated. Based on the spatio-temporal mapping perception model and the adaptive weight expression, a deep mapping conversion is performed on the spatio-temporal coupling feature vector to calculate the load saturation of each functional conveying unit.

6. The method of claim 1, wherein, Based on the set of differential control strategies, a front-end feedback composite control is implemented, and a comprehensive feedback data is formed by monitoring the deviation between the actual VOC volatile response and the multi-scale physical-chemical coupling model prediction value during the execution process. Based on the set of differential control strategies, an adaptive priority protocol is used to distribute corresponding control instructions to each functional conveying unit, and each functional conveying unit performs corresponding control actions according to the control instructions to form a front-end control. During the execution of the front-end control, the control instructions are input into the multi-scale physical-chemical coupling model to calculate the VOC volatile prediction response of each functional conveying unit under the current control instruction, and the model prediction value is obtained. The actual VOC concentration change data of each functional conveying unit is collected in real time by the VOC concentration sensor arranged on each functional conveying unit, and the actual VOC concentration change data is used as the actual VOC volatile response. The actual VOC volatile response and the model prediction value are dynamically compared through an adaptive sliding time window to calculate the deviation between them. The systematic deviation component and the random deviation component in the deviation are analyzed by time series analysis, and the systematic deviation component and the random deviation component are fused to form comprehensive feedback data containing model mismatch information and environmental disturbance information.

7. The method of claim 6, wherein, During the execution of the front-end control, the control instructions are input into the multi-scale physical-chemical coupling model to calculate the VOC volatile prediction response of each functional conveying unit under the current control instruction, and the model prediction value is obtained. The functional conveying unit control instructions in the front-end control are obtained, input into the multi-scale physical-chemical coupling model, and the flow field diffusion constraint and the chemical reaction rate constraint are nonlinearly superimposed to establish the VOC volatile spatio-temporal evolution equation. Based on the multi-scale physical-chemical coupling model, the dynamic change characteristics of the VOC concentration distribution and the chemical reaction rate are captured in real time, a distributed monitoring network is established, and the constraint parameters of the VOC volatile spatio-temporal evolution equation are adaptively updated. According to the feedback data of the distributed monitoring network, the material transportation characteristics and the chemical reaction kinetics law are combined to optimize the parameters of the multi-scale physical-chemical coupling model, and a calculation parameter set is generated. Based on the multi-scale physical-chemical coupling model, an iterative calculation framework is established, the VOC volatile spatio-temporal evolution equation is dynamically solved through the calculation parameter set, the constraint parameters and the calculation results are bidirectionally corrected, the VOC volatile prediction response of each functional conveying unit under the current control instruction is calculated, and the model prediction value is obtained.

8. A low VOC polyurethane raw material production transfer adaptive control system for implementing the method of any of the preceding claims 1-7, characterized in that, It includes: The first unit is configured to acquire multi-dimensional dynamic information of polyurethane raw materials in a production conveying system through a distributed sensing network, construct a multi-scale physical-chemical coupling model of VOC volatilization behavior based on the multi-dimensional dynamic information, correlate component diffusion dynamics, phase interface mass transfer behavior and flow field transport characteristics across scales, and obtain state space representation; The second unit is configured to decouple the physical topology of the conveying system according to the volatilization risk area distribution and evolution characteristics reflected in the state space representation, decompose the continuous conveying path into functional conveying units with different volatilization inhibition capability levels, establish a dynamic connection relationship between the functional conveying units based on a VOC cumulative load balancing principle, and obtain an adaptive topology configuration scheme; The third unit is configured to construct a multi-objective optimization framework and generate a set of differentiated control strategies for each functional conveying unit according to the adaptive topology configuration scheme; implement feedforward-feedback compound control based on the set of differentiated control strategies, monitor the deviation between the actual VOC volatilization response and the predicted value of the multi-scale physical-chemical coupling model during the execution process, and form comprehensive feedback data; The fourth unit is configured to correct the cross-scale correlation weight coefficient and the Pareto front search strategy of the multi-objective optimization framework using the comprehensive feedback data, and realize continuous optimization of the VOC control performance in the conveying process.

9. An electronic device, comprising: It comprises: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to realize the method of any one of claims 1 to 7.

Images