A control method for considering the test bench surface flow rate of exhaust air activated carbon filter

Abstract

The application provides a control method considering the test bench surface flow rate of exhaust activated carbon filter, and belongs to the technical field of laboratory.The application establishes a data fusion set through collecting multi-source sensor data, establishes a differential equation for the activated carbon layer and a double-layer filter membrane area, establishes a Navier-Stokes equation for the air outlet and mesh screen through-hole area, solves the full flow field velocity vector by using a spectral element dimension reduction parallel solution algorithm, establishes an adaptive dynamic slip boundary condition at the double-layer filter membrane interface and corrects the source term by using an equivalent air permeability matrix, establishes a refined multi-layer nested space flow field topology model in a cylindrical coordinate system and introduces a wall function correction, uses artificial intelligence to output each air outlet flow rate prediction value and detect mesh screen blockage anomalies, and solves the technical problem that the closed-loop control precision of the exhaust activated carbon filter bench surface flow rate is insufficient due to the flow field calculation distortion of the coupling area of the porous medium and the free fluid.

Description

Technical Field

[0001] This invention belongs to the field of laboratory technology, and more specifically, relates to a method for controlling the flow velocity on a test bench considering an exhaust activated carbon filter. Background Technology

[0002] In the field of test bench exhaust systems, activated carbon filters are widely used for the adsorption and purification of harmful gases. Their core function relies on the uniform and stable distribution of airflow velocity at the outlet. Existing technologies typically use a single-point pressure sensor combined with a simplified Bernoulli equation or empirical formula to estimate the outlet airflow velocity, and use this as feedback to implement proportional-integral-derivative closed-loop control of the inlet fan speed or valve opening. In actual operation, the activated carbon filter contains multiple heterogeneous structures, including a mesh sieve, an activated carbon layer, a first permeable filter membrane, and a second permeable filter membrane. When the airflow passes through the porous media region and the free fluid region, the form of its momentum conservation equation undergoes a fundamental change, resulting in significant abrupt changes in velocity gradient and discontinuous shear stress at the interface. Under the radial non-uniform geometric constraints of concentric circular through-holes, and in conditions of localized blockage of the activated carbon layer or uneven distribution of filter membrane resistance, existing simplified calculation methods cannot accurately capture the momentum transfer law at the interface between the porous medium and the free fluid. This leads to a systematic deviation between the calculated flow field and the actual distribution, causing the closed-loop controller to output incorrect adjustment values ​​based on distorted velocity information, ultimately resulting in the table surface velocity deviating from the target value. In other words, existing technologies suffer from the technical problem of insufficient closed-loop control accuracy due to the distortion of the flow field calculation in the coupling region between the porous medium and the free fluid in the exhaust activated carbon filter table surface velocity. Summary of the Invention

[0003] In view of this, the present invention provides a method for controlling the flow velocity on the test bench of an exhaust activated carbon filter, which can solve the technical problem in the prior art where the flow velocity on the test bench of the exhaust activated carbon filter is not accurate enough due to the distortion of the flow field calculation in the coupling region between the porous medium and the free fluid.

[0004] This invention is implemented as follows: This invention provides a method for controlling the flow velocity on a test bench considering an exhaust activated carbon filter, comprising the following steps:

[0005] Collect inlet pressure, concentric circular through-hole geometric parameters, first permeable filter membrane resistance coefficient, second permeable filter membrane resistance coefficient, and outlet air velocity data to establish a multi-source sensor data fusion set;

[0006] A generalized Darcy-Brinkman topological differential equation was established for the activated carbon layer and the double-layer filter membrane region, and a Navier-Stokes equation was established for the air outlet and the mesh screen through-hole region. The velocity vector of the entire flow field was solved by the spectral element dimension reduction parallel solution algorithm.

[0007] An adaptive dynamic slip boundary condition is established at the interface between the first and second permeable filter membranes. The source term of the momentum conservation equation at the interface is dynamically corrected using the equivalent permeability matrix to eliminate numerical oscillations caused by abrupt changes in the flow field.

[0008] A refined multi-layered nested spatial flow field topology model is established based on cylindrical coordinates. A high-order tensor product is used to locally refine the control of the concentric circular through-hole region. A wall function correction mechanism is introduced at the wall boundary to analyze the nonlinear coupling between radial shear flow and axial mainstream flow.

[0009] The multi-source sensor data fusion set is input into the attention jump mechanism model constrained by physical fluid dynamics, outputting the predicted flow velocity values ​​of each air outlet, and identifying local weak resistance anomalies caused by mesh screen blockage through the flow field residual variational anomaly detector.

[0010] Based on the deviation between the predicted flow velocity output by the attention jump mechanism model constrained by physical fluid dynamics and the flow velocity on the target platform, the adaptive gating adjustment function value is calculated. According to the interval to which the function value belongs, the corresponding adjustment strategy is adopted to perform closed-loop control of the inlet pressure, so that the flow velocity at the outlet converges to the flow velocity on the target platform.

[0011] Specifically, the multi-source sensor data fusion set is formed by merging the time-series signals collected by the inlet pressure sensor, the flow velocity sensors of each outlet, and the differential pressure sensor before and after the mesh screen into a unified data set after time synchronization and alignment. The sampling frequency of the sensors is the minimum value within the frequency range that prevents the flow field time-series signals from aliasing.

[0012] The mathematical form of the generalized Darcy-Brinkman topological differential equation is: ,in For dynamic viscosity, For penetration rate, It is a velocity vector. For effective viscosity, For pressure, For volumetric forces, permeability The steady-state pressure drop was determined by conducting steady-state pressure drop experiments on the activated carbon layer and the double-layer filter membrane, and the Darcy equation was obtained by fitting linear regression.

[0013] Specifically, the spectral element dimensionality reduction parallel solution algorithm uses high-order Chebyshev polynomials as basis functions to transform partial differential equations into a system of algebraic equations, constructs stress jump continuous boundary conditions at the interface between porous media and free fluid, and uses Scher complement matrix domain decomposition technology to reduce the dimensionality of the global large matrix into independent sub-matrices of each element, and performs fully asynchronous parallel solution on a multi-core processor.

[0014] The slip length parameter of the adaptive dynamic slip boundary condition is obtained by performing segmented pressure drop experiments on the double-layer filter membrane under different flow rate conditions and then fitting it using the generalized Darcy-Brinkman topological differential equation; the matrix elements of the equivalent air permeability matrix are obtained by performing local pressure drop tests on different radial positions of the filter membrane, with the test points evenly arranged in the radial direction.

[0015] Specifically, the refined multi-layered nested spatial flow field topology model is implemented in cylindrical coordinates. Below, nested mesh topologies with different resolutions are established for the concentric circular through-hole region, the wall boundary layer region, and the axial region, respectively. The wall function correction mechanism uses a logarithmic wall function to analytically correct the near-wall velocity distribution.

[0016] Among them, the attention jump mechanism model of physical fluid dynamics constraint adopts an improved residual network backbone structure. The backbone is composed of multiple cascaded physical residual blocks. Each physical residual block is equipped with a physical conservation residual judgment node. The residuals of the Navier-Stokes continuity equation and momentum equation are calculated for the current feature map. When the residual exceeds the physical tolerance threshold, the internal iterative gating unit is activated.

[0017] The internal iterative gating unit adopts a gated loop unit structure, which returns the feature data that does not meet the physical constraints to the previous layer of the current block for physical error correction iteration. The upper limit of the number of iterations is determined by experimental comparison of convergence speed and computational cost under different number of iterations.

[0018] Among them, the attention jump mechanism model of physical fluid dynamics constraints introduces a cylindrical coordinate attention mechanism, using radial position encoding as the query vector and the geometric parameters of concentric circular through holes as key-value pairs. It generates radially weighted features through scaling dot product attention operations and performs jump connections and fusion with the backbone feature map.

[0019] Among them, the loss function of the attention jump mechanism model with physical fluid dynamics constraints is composed of three parts: the mean square error between the predicted flow velocity and the measured flow velocity at each air outlet, the momentum conservation differential residual term, and the mass conservation differential residual term. The weight coefficients are determined on the training set through grid search experiments.

[0020] The flow field residual variational anomaly detector constructs a variational autoencoder to reconstruct the current flow field residual based on the statistical distribution of the residual signal under normal operating conditions. When the reconstruction error exceeds the anomaly threshold, it is determined that there is blockage or abnormality in the mesh screen. The anomaly threshold is determined by calculating the 95th percentile of the reconstruction error under normal operating conditions.

[0021] Wherein, the adjustment index of the adaptive gating function It is the weighted root mean square of the deviation between the current predicted flow velocity of each air outlet and the target flow velocity on the table. The weighting coefficient is determined by the contribution of the radial position of each air outlet to the uniformity of the flow velocity on the table. The contribution coefficient is obtained by regression fitting of the uniformity experimental data.

[0022] Among them, when the adjustment index Not less than the large deviation threshold When the adjustment index is large, a large step size proportional-integral-derivative adjustment strategy is adopted; when the adjustment index is large, a large step size proportional-integral-derivative adjustment strategy is adopted. Not less than the small deviation threshold and less than When the adjustment index is at its maximum, a medium step size proportional-integral adjustment strategy is adopted; when the adjustment index is at its maximum... Less than At that time, a small-step integral fine-tuning strategy is adopted.

[0023] Among them, the training dataset of the attention jump mechanism model under physical fluid dynamics constraints is collected from time-series data of multi-source sensors under different inlet pressure conditions, different activated carbon layer blockage degrees and different ambient temperatures. The full flow field velocity vector calculation results of the spectral element dimensionality reduction parallel solution algorithm are used as physical labels to construct training sample pairs. After normalizing all samples, they are divided according to the ratio of training set, validation set and test set.

[0024] The sensor's sampling frequency range is 20–200 Hz; permeability The experiment was repeated 5–10 times; the experimental working conditions for the slip length parameter were no less than 10; the equivalent air permeability matrix test points were evenly distributed in the radial direction at 5–15 points; the upper limit of the iteration number of the internal iterative gating unit was 3–8 times; the number of sensor data acquisition sets under normal working conditions for the abnormal threshold was 50–200 sets; the number of experimental data sets for the contribution coefficient was 10–20 sets; and the large deviation threshold was... With small deviation threshold The step response was determined through no fewer than 20 sets of experiments; the inlet pressure of the training dataset covered 50–500 Pa, the activated carbon layer blockage level ranged from 5 to 10, the ambient temperature ranged from 15 to 40℃, and the data acquisition time for each set of conditions was no less than 300 s. The ratio of the training set, validation set, and test set was 8:1:1, and the initial learning rate ranged from [value missing]. ~ The number of input samples in each batch ranges from 16 to 64, and the search range for the physical constraint loss weight coefficient is from 0.01 to 10.

[0025] This invention establishes a partitioned coupling model of the generalized Darcy-Brinkman topological differential equation and the Navier-Stokes equation, and applies adaptive dynamic slip boundary conditions and equivalent permeability matrix corrections at the interface between porous media and free fluid. This eliminates the root cause of computational distortion in the interface momentum conservation equation from a physical modeling perspective, ensuring that the solution of the flow field velocity vector accurately reflects the resistance characteristics and coupling effects of each layer of heterogeneous media. Furthermore, a physical fluid dynamics-constrained attention-jump mechanism model embeds the Navier-Stokes residual into the neural network feature propagation process, ensuring that data-driven predictions are always subject to hard constraints of fluid dynamics laws, guaranteeing the physical self-consistency of the predicted flow velocity values ​​for each outlet. Moreover, the adaptive gating control function automatically switches adjustment strategies based on the magnitude of the flow velocity deviation, enabling the closed-loop control system to converge to the target platform flow velocity with matched step sizes and stability across different deviation ranges. In summary, this invention solves the technical problem mentioned in the background art where the flow velocity at the exhaust activated carbon filter platform is insufficiently accurate due to computational distortion of the flow field in the coupling region between porous media and free fluid, leading to insufficient closed-loop control accuracy. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a graph showing the convergence process of the flow velocity at each air outlet over time under moderate blockage conditions.

[0028] Figure 3 This is a diagram showing the steady-state flow velocity distribution of each air outlet under different operating conditions.

[0029] Figure 4 The graph shows the change of the adjustment index over time.

[0030] Figure 5 This is a schematic diagram of the activated carbon exhaust device in Example 2. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0032] like Figure 1 The diagram shows a flowchart of a method for controlling the flow velocity on a test bench surface considering an exhaust activated carbon filter, provided by the present invention. This method includes the following steps:

[0033] S01. Collect inlet pressure, concentric circular through-hole geometric parameters, first permeable filter membrane resistance coefficient, second permeable filter membrane resistance coefficient, and outlet flow velocity data to establish a multi-source sensor data fusion set.

[0034] S02. Establish a generalized Darcy-Brinkman topological differential equation for the activated carbon layer and the double-layer filter membrane region, and establish a Navier-Stokes equation for the air outlet and the mesh screen through-hole region. Use the spectral element dimensionality reduction parallel solution algorithm to solve the velocity vector of the entire flow field.

[0035] S03. An adaptive dynamic sliding boundary condition is established at the interface between the first and second permeable filter membranes. The source term of the momentum conservation equation at the interface is dynamically corrected using the equivalent permeability matrix to eliminate numerical oscillations caused by sudden changes in the flow field.

[0036] S04. Based on cylindrical coordinates, a refined multi-layer nested spatial flow field topology model is established. The high-order tensor product is used to locally refine the control of the concentric circular through-hole region. A wall function correction mechanism is introduced at the wall boundary to analyze the nonlinear coupling between radial shear flow and axial mainstream flow.

[0037] S05. The multi-source sensor data fusion set is input into the attention jump mechanism model constrained by physical fluid dynamics, outputting high-fidelity flow velocity prediction values ​​for each air outlet, and identifying local weak resistance anomalies caused by mesh screen blockage through the flow field residual variational anomaly detector.

[0038] S06. Based on the deviation between the predicted flow velocity output by the attention jump mechanism model under physical fluid dynamics constraints and the target platform flow velocity, calculate the adaptive gating adjustment function value, and adopt the corresponding adjustment strategy according to the interval to which the function value belongs to perform closed-loop control of the inlet pressure so that the outlet flow velocity converges to the target platform flow velocity.

[0039] The multi-source sensor data fusion set refers to the fusion of time-series signals collected by the inlet pressure sensor, the flow velocity sensors at each outlet, and the differential pressure sensor before and after the mesh screen into a unified data set after time synchronization and alignment. This data set is used for subsequent flow field calculations and model input. The sampling frequency of the sensors was determined through multiple experiments, and the frequency range of 20 to 200 Hz was selected. The minimum value that prevents aliasing of the flow field time series signal. The generalized Darcy-Brinkman topological differential equation refers to the topological form of the differential equation that simultaneously introduces the Darcy drag term and the Brinkman viscous shear term in the porous medium region. Its mathematical form is: ;in For dynamic viscosity, For penetration rate, It is a velocity vector. For effective viscosity, For pressure, The permeability is a volumetric force term. Steady-state pressure drop experiments were conducted on the activated carbon layer and the double-layer filter membrane, respectively. The Darcy equation was obtained by fitting linear regression, and the experiments were repeated 5 to 10 times with the average value taken. The Navier-Stokes equations are the standard partial differential equations describing the conservation of momentum in a free fluid region, and are applicable to the free flow field region between the air outlet and the mesh screen openings. The specific implementation of the spectral element dimensionality reduction parallel solution algorithm is as follows: A generalized Darcy-Brinkman topological differential equation is established in the activated carbon layer and double-layer filter membrane region, and a Navier-Stokes equation is established in the air outlet and mesh screen pore region. Using the spectral element method with high-order Chebyshev polynomials as basis functions, the cylindrical flow domain is divided into several large-size spectral elements. Within each element, the partial differential equations are transformed into a system of algebraic equations. Stress-jumping continuous boundary conditions are constructed at the interface between the porous medium and the free fluid. The Schur complement matrix domain decomposition technique is used to reduce the dimensionality of the global large matrix into independent sub-matrices for each element. A fully asynchronous parallel solution is performed on a multi-core processor, and each element achieves rapid iterative assembly of the high-precision velocity field through a small amount of data exchange of boundary basis coefficients. The technical effects of the spectral element dimensionality reduction parallel solution algorithm are as follows: This algorithm uses the generalized Darcy-Brinkman topological differential equation to uniformly describe the flow characteristics of porous media and free fluid regions, and accurately applies stress jump continuous boundary conditions at the interface, fundamentally solving the calculation distortion problem of abrupt changes in flow velocity between the activated carbon layer, the double-layer filter membrane and the free fluid region; the spectral element method uses high-order Chebyshev polynomials as basis functions, achieving extremely high spatial resolution with a very small number of elements, and has a significant advantage in capturing local strong gradient flow fields in concentric circular through-hole regions; the Schur complement matrix domain decomposition technique decomposes the globally coupled large matrix into independent sub-matrices of each element, and each sub-matrices are solved asynchronously and in parallel on a multi-core processor, with only the boundary basis coefficients being exchanged between elements, greatly reducing communication overhead, and significantly reducing the overall computational resource consumption while ensuring solution accuracy, thus meeting the timeliness requirements of real-time flow field control. The adaptive dynamic slip boundary condition refers to the dynamic boundary condition established at the interface between the first and second permeable membranes based on the slip length parameter updated in real time according to the current flow velocity and pressure difference. This slip length parameter is adjusted under different flow velocity conditions (0.1–2.0). A segmented pressure drop experiment was conducted on the double-layer filter membrane, and the pressure drop was obtained using the inversion fitting of the generalized Darcy-Brinkman topological differential equation. At least 10 experimental operating points were used. The equivalent permeability matrix refers to the parameter matrix that equivalently represents the non-uniform microporous resistance characteristics of the first and second permeable filter membranes in matrix form. The matrix elements were obtained by conducting local pressure drop tests at different radial positions of the filter membrane, with 5–15 test points evenly distributed radially. The refined multi-layer nested spatial flow field topology model refers to the model in cylindrical coordinates. Below, nested mesh topologies with different resolutions are established for the concentric circular through-hole region, the wall boundary layer region, and the axial region to accurately capture the velocity gradient distribution in each region. High-order tensor product local refinement control refers to using high-order interpolation basis functions in tensor product form to locally refine the flow field mesh in the concentric circular through-hole region, thereby improving the accuracy of local flow field calculations without increasing the global mesh size. The wall function correction mechanism introduces a logarithmic wall function into the wall boundary layer region to analytically correct the near-wall velocity distribution, avoiding the excessively dense mesh required for directly analyzing the extremely thin boundary layer. The flow field residual variational anomaly detector is based on the statistical distribution of the residual signal under normal flow conditions. A variational autoencoder is constructed to reconstruct the current flow field residual. When the reconstruction error exceeds an anomaly threshold determined by experimental data under normal conditions, it is determined that there is blockage or abnormality in the mesh screen. The anomaly threshold is determined by collecting 50-200 sets of sensor data under normal conditions and calculating the 95th percentile of the reconstruction error. The design of the adaptive gating control function is as follows: Define the regulation index. This is the weighted root mean square of the deviations between the predicted flow velocity at each air outlet and the flow velocity on the target platform. The attention jump mechanism model constrained by physical fluid dynamics outputs and calculates in real time. The weighting coefficients are determined by the contribution of the radial position of each air outlet to the uniformity of the flow velocity on the platform. The contribution coefficients are obtained by regression fitting of 10-20 sets of uniformity experimental data. At that time, a large-step proportional-integral-derivative adjustment strategy is adopted to perform rapid and large-scale adjustment of the inlet air pressure; when At that time, a medium-step proportional-integral control strategy is adopted to smoothly regulate the inlet air pressure; when At that time, a small-step integral fine-tuning strategy is adopted to perform fine adjustment of the inlet pressure; the aforementioned and By conducting step response experiments near the target platform velocity, the balance point between regulation efficiency and stability under different deviation ranges is determined, with no fewer than 20 experimental groups.

[0040] The attention-jumping mechanism model based on physical fluid dynamics constraints is deep learning. Its specific structure is as follows: the network as a whole adopts an improved residual network backbone structure. The input layer receives a fusion set of multi-source sensor data, including inlet pressure, geometric parameters of concentric circular through-holes, resistance coefficients of the first and second permeable membranes, and real-time flow velocities of each outlet. The input dimension corresponds one-to-one with the number of sensor channels. The backbone consists of multiple cascaded physical residual blocks. Each physical residual block contains a convolutional layer, a batch normalization layer, and a linear rectified function activation layer, and also includes a physical conservation residual judgment node. This node calculates Navier-Stokes equations for the current feature map. If the residuals of the Stokes continuity equation and the momentum equation exceed the experimentally calibrated physical tolerance threshold, the internal iterative gating unit is activated. This internal iterative gating unit employs a gated loop unit structure, forcibly returning feature data that does not meet physical constraints to the previous layer of the current block for physical error correction iteration. The upper limit of iterations is 3–8 times, determined through multiple experimental comparisons of convergence speed and computational cost at different iteration counts. In the radial geometric feature processing module, a cylindrical coordinate attention mechanism is introduced. A separate attention weight matrix is ​​established for the radial geometric parameters of the concentric circular mesh sieve. This weight matrix uses the radial position encoding as the query direction. The flow rate is calculated using the through-hole geometric parameters as key-value pairs. Radially weighted features are generated through scaling dot product attention operations and then fused with the backbone feature map via skip connections. The network loss function consists of a weighted sum of three parts: the first part is the mean square error between the predicted and measured flow velocities of each outlet; the second part is the momentum conservation differential residual term; and the third part is the mass conservation differential residual term. The weight coefficients of these three parts are determined on the training set through grid search experiments. The output layer is a fully connected layer that outputs the predicted flow velocity values ​​for each outlet. Regarding computational resource allocation, the CUDA flow allocation strategy distributes the convolution operations of the physical residual blocks and the loop operations of the internal iterative gating units to... Independent CUDA streams execute in parallel. Matrix operations of the cylindrical coordinate attention mechanism occupy a separate CUDA stream. CUDA streams are coordinated through a CUDA event synchronization mechanism. Memory allocation is dynamically planned based on the number of parameters in each layer and the size of intermediate feature maps. The batch normalization layer parameters of the physical residual block are prioritized for resident memory in the persistent memory area. The hidden state tensor of the internal iterative gating unit dynamically expands memory allocation based on the number of iterations. In terms of hierarchical allocation, the input and output layers are placed in host memory, while the backbone physical residual block and cylindrical coordinate attention mechanism module are placed in video memory, realizing an asynchronous data pipeline between the host and the device. The adjustment index of the adaptive gating function... Establish a functional relationship between the number of iterations of the internal iterative gating unit of the attention jump mechanism model with physical fluid dynamics constraints, when At that time, the iteration upper limit of the internal iterative gating unit is dynamically set to the maximum value to improve the accuracy of physical constraint satisfaction. At this time, the iteration limit of the internal iterative gating unit is dynamically set to the minimum value to reduce inference latency while taking into account control real-time performance.

[0041] The technical effects of the attention jump mechanism model with physical fluid dynamics constraints are as follows: This model embeds the residuals of the Navier-Stokes continuity equation and the momentum equation into the physical conservation residual judgment node of the residual block, so that the feature propagation process of the neural network is always subject to the hard constraints of fluid dynamics laws, rather than relying solely on data fitting; the internal iterative gating unit forcibly triggers feedback correction when the features do not meet the physical constraints, which is equivalent to embedding a physically interpretable iterative solution process in the forward propagation of the network; the cylindrical coordinate attention mechanism explicitly encodes the radial geometric non-uniformity of the concentric circular mesh screen into attention weights, enabling the network to have directional perception of radial position-related flow field distortions; the joint training of the three weighted loss functions enables the network to output physically consistent flow velocity prediction results even with very few training samples, effectively avoiding the overfitting risk of pure data-driven models under small sample conditions, while ensuring the spatial consistency of flow velocity predictions for each air outlet.

[0042] The specific steps for establishing the training dataset for the attention jump mechanism model under physical fluid dynamics constraints include: arranging inlet pressure sensors, outlet velocity sensors, and differential pressure sensors before and after the mesh screen on the exhaust device of the test bench; and testing under different inlet pressure conditions (covering the actual operating range of 50–500). Under different activated carbon layer blockage levels (constructing 5-10 blockage levels by gradually adding simulated microparticles) and different ambient temperatures (15-40℃, at 5℃ intervals), time-series data from multiple sensors were collected, with each set of operating conditions requiring a collection time of no less than 300 seconds. The full-field velocity vector calculation results from the spectral element dimensionality reduction parallel solution algorithm are used as physical labels and matched with corresponding sensor data to construct training sample pairs; all samples are normalized, with a normalization range of [missing information]. The normalization coefficient is determined by the total sample statistic.

[0043] The specific steps for training the attention jump mechanism model under the physical fluid dynamics constraints include: dividing the training dataset into a training set, a validation set, and a test set in an 8:1:1 ratio; updating the parameters using an adaptive moment estimation optimizer, with the initial learning rate ranging from [value missing]. ~ The learning rate is dynamically adjusted using a cosine annealing strategy for the validation set loss; the number of input samples per batch ranges from 16 to 64, and the optimal batch size is determined by comparing the convergence speed of the validation set loss experimentally; the weight coefficients of the physical constraint loss are optimized on the validation set using a grid search, with a search range of [missing information]. ~ The step size is 0.1; when the loss on the validation set no longer decreases after 20 to 50 consecutive training rounds, training is stopped; after training is completed, the root mean square error of the velocity prediction of each air outlet is evaluated on the test set. The root mean square error must meet the upper limit threshold determined by the actual engineering accuracy requirements. The upper limit threshold is determined by comparing the velocity vector of the entire flow field calculated using the parallel solution algorithm of spectral element dimensionality reduction.

[0044] The activated carbon exhaust device has the following structure: It is a cylindrical structure, with an air inlet, a mesh screen, an activated carbon layer, a first breathable filter membrane, a second breathable filter membrane, and an air outlet arranged sequentially along the airflow direction. The air inlet is located at one end of the cylindrical structure, and airflow enters the device through it. The mesh screen is positioned adjacent to the inner side of the air inlet, and has multiple circular through-holes arranged concentrically. The size of these through-holes is smaller than the size of the activated carbon particles in the activated carbon layer, allowing airflow to pass through the mesh screen before entering the device. An activated carbon layer is disposed between a mesh sieve and a first breathable filter membrane to adsorb harmful gas molecules in the airflow. The first breathable filter membrane is disposed between the activated carbon layer and a second breathable filter membrane, and the first breathable filter membrane is made of polytetrafluoroethylene. The second breathable filter membrane is disposed between the first breathable filter membrane and the air outlet, and the second breathable filter membrane is made of polypropylene non-woven fabric. The air outlet is disposed at the other end of the cylindrical structure, and the air outlet includes multiple air outlet holes. The airflow filtered by the double-layer filter membrane is discharged to the outside of the device through the multiple air outlet holes.

[0045] Optionally, the present invention also provides a computer-implemented system form, forming a test bench flow velocity control system considering an exhaust activated carbon filter. The system is connected to an inlet pressure sensor, flow velocity sensors at each outlet, and a differential pressure sensor before and after the mesh screen. The system is equipped with a readable storage medium storing program instructions. When the program instructions are executed in the computer, they drive the computer to call the attention jump mechanism model of physical fluid dynamics constraints, the spectral element dimensionality reduction parallel solution algorithm, and the adaptive gating adjustment function to execute the above control method, thereby realizing closed-loop control of the inlet pressure and causing the outlet flow velocity to converge to the target test bench flow velocity.

[0046] The specific implementation of step S01 is as follows: a pressure sensor is arranged at the air inlet of the activated carbon exhaust device, a flow velocity sensor is arranged at each air outlet, and a differential pressure sensor is arranged before and after the mesh screen. The sampling frequency is determined based on the Nyquist sampling theorem. Within the frequency range of 20–200 Hz, the sampling frequency is gradually increased through multiple experiments until the spectrum of the flow field time-series signal no longer exhibits aliasing. The minimum sampling frequency at this point is taken as the unified operating frequency of each sensor to reduce unnecessary computational resource consumption. The time-series signals collected by each sensor are aligned by hardware triggering or software timestamp to eliminate phase deviations between channels. They are then fused into a multi-source sensor data fusion set containing the air inlet pressure sequence, the flow velocity sequence of each air outlet, and the differential pressure sequence before and after the mesh screen, for subsequent flow field calculations and model input.

[0047] The specific implementation of step S02 is as follows: For the internal structure of the activated carbon exhaust device, two types of equations are divided according to the fluid domain type. In the porous media region composed of the activated carbon layer and the double-layer filter membrane, a generalized Darcy-Brinkman topological differential equation is established, simultaneously introducing the Darcy resistance term and the Brinkman viscous shear term, and the required permeability... Five to ten steady-state pressure drop experiments were conducted on the activated carbon layer and the double-layer filter membrane, respectively. The mean value was obtained by fitting the slope of the Darcy equation using linear regression. The standard Navier-Stokes equations were established in the free fluid region formed by the air outlet and the mesh screen openings. A parallel solution algorithm for spectral element dimensionality reduction was used, with higher-order Chebyshev polynomials as basis functions, to divide the cylindrical flow domain into several large-size spectral elements. Within each element, the partial differential equations were transformed into a system of algebraic equations. Stress-jumping continuous boundary conditions were applied at the interface between the porous medium and the free fluid to ensure the conservation of momentum flux on both sides of the interface. The Scher complement matrix domain decomposition technique was used to reduce the dimensionality of the globally coupled large matrix into independent sub-matrices for each element. Each sub-matrices were solved asynchronously and in parallel on a multi-core processor, with only the boundary basis coefficients being exchanged between elements. Finally, the velocity vector of the entire flow field was obtained through iterative concatenation.

[0048] The specific implementation of step S03 is as follows: At the interface between the first and second permeable filter membranes, the slip length parameter of the adaptive dynamic slip boundary condition is obtained by performing segmented pressure drop experiments on the double-layer filter membrane at no less than 10 flow rate points within the range of 0.1 to 2.0 m / s. The parameter is obtained through inversion fitting using the generalized Darcy-Brinkman topological differential equation as the forward model, and is dynamically updated during operation based on real-time flow velocity and pressure difference. The matrix elements of the equivalent permeability matrix are obtained by measuring the local pressure drop at 5 to 15 test points uniformly arranged radially on the first and second permeable filter membranes, representing the spatial distribution of the non-uniform microporous resistance of the filter membrane in matrix form. The equivalent permeability matrix is ​​used to dynamically correct the source term of the momentum conservation equation at the interface, ensuring that the momentum transfer on both sides of the interface remains numerically continuous, thereby avoiding numerical oscillations caused by abrupt changes in filter membrane resistance.

[0049] The specific implementation of step S04 is as follows: in cylindrical coordinate system The cylindrical flow domain is divided into three regions based on regional characteristics: a concentric circular through-hole region, a wall boundary layer region, and an axial region. Nested mesh topologies with three levels of resolution (coarse, medium, and fine) are established for each region, forming a refined multi-layered nested spatial flow field topology model. In the concentric circular through-hole region, higher-order interpolation basis functions in the form of tensor products are used to locally refine the mesh, improving the accuracy of local flow field calculations without increasing the global mesh size and effectively capturing the strong radial shear gradient at the through-hole edge. In the wall boundary layer region, a logarithmic law wall function correction mechanism is introduced. The near-wall velocity distribution is analyzed using the logarithmic law formula, avoiding the excessively dense mesh required for directly analyzing the extremely thin boundary layer, while ensuring the accuracy of near-wall shear stress calculations. Through the synergistic effect of the nested mesh and wall functions, the nonlinear coupling relationship between radial shear flow and axial mainstream flow is accurately analyzed.

[0050] The specific implementation of step S05 is as follows: The attention skipping mechanism model with physical fluid dynamics constraints takes a multi-source sensor data fusion set as input, and the dimension of the input layer corresponds one-to-one with the number of sensor channels. The backbone consists of multiple cascaded physical residual blocks. The physical conservation residual judgment node in each physical residual block calculates the Navier-Stokes continuity equation residual and momentum equation residual for the current feature map. When the residual exceeds the physical tolerance threshold calibrated by experiments, the internal iterative gating unit is activated. The internal iterative gating unit adopts a gated loop unit structure, returning the feature data to the previous layer for physical error correction iteration, with an upper limit of 3 to 8 iterations. The cylindrical coordinate attention mechanism uses radial position encoding as the query vector and the geometric parameters of concentric circular through-holes as key-value pairs. It generates radially weighted features through scaling dot product attention operations and fuses them with the backbone feature map skipping connections, enabling the network to have directional perception capabilities for radial position-related flow field distortions. The output layer is a fully connected layer, outputting the predicted flow velocity values ​​for each air outlet. The variational anomaly detector for flow field residuals trains a variational autoencoder by collecting 50 to 200 sets of sensor data under normal operating conditions. The 95th percentile of the reconstruction error is used as the anomaly threshold. When the reconstruction error of the current flow field residual exceeds this threshold, it is determined that there is blockage in the mesh screen and an alarm is triggered.

[0051] The specific implementation of step S06 is as follows: the adaptive gating control function first calculates the control index. , which is defined as the weighted root mean square of the deviation between the predicted flow velocity of each air outlet and the target flow velocity on the platform. The weighting coefficient is determined by the contribution of the radial position of each air outlet to the uniformity of the flow velocity on the platform. The contribution coefficient is obtained by regression fitting of 10 to 20 sets of uniformity experimental data. Within the specified range, the control system automatically switches the adjustment strategy: when Not less than the large deviation threshold At that time, a large-step proportional-integral-derivative adjustment strategy is adopted to perform rapid and large-scale adjustment of the inlet air pressure; when Within the small deviation threshold and When the interval is between, a moderate step size proportional-integral control strategy is adopted to perform smooth control; when Less than At that time, a small-step integral fine-tuning strategy is adopted to perform fine adjustment. and Through at least 20 sets of step response experiments near the target platform velocity, the balance point between regulation efficiency and stability was determined under different deviation ranges. The above three-stage regulation strategy forms a complete closed-loop control loop, causing the outlet velocity to gradually converge to the target platform velocity.

[0052] It should be noted that the key technologies of this invention include: a partitioned coupling modeling strategy of generalized Darcy-Brinkman topological differential equations and Navier-Stokes equations, which unifies the momentum transfer descriptions of porous media and free fluids at the equation level, eliminating computational distortion at the interface during the physical modeling stage, rather than relying on post-processing compensation; a spectral element dimensionality reduction parallel solution algorithm that combines high-order Chebyshev basis functions with Scher complement matrix domain decomposition to achieve high spatial resolution in strong gradient regions with a very small number of units, while satisfying the time constraints of real-time control through fully asynchronous parallel computation, thus resolving the contradiction between accuracy and speed; a physical fluid dynamics constraint attention jump mechanism model that embeds the fluid dynamics equation residuals into the feature propagation process of the neural network, subjecting data-driven prediction to hard constraints of physical laws, and explicitly encoding radial geometric non-uniformity through a cylindrical coordinate attention mechanism, giving the network the ability to directionally perceive local flow field distortions; and an adaptive gating adjustment function that automatically switches between three levels of deviation intervals, balancing rapid response under large deviation conditions with fine stability under small deviation conditions. The four key technologies work together to form a complete technology chain from physical modeling, high-precision solving, intelligent prediction to adaptive control, making the overall control accuracy significantly better than the effect of any single technology applied alone.

[0053] It should be noted that under conditions of non-uniform localized blockage in the activated carbon layer, some pores of the mesh screen are covered by particles, resulting in a strong local gradient in the flow field within the pore area. This leads to significant differences in flow velocity at the outlets at different radial positions, severely deteriorating the uniformity of the flow velocity distribution on the platform. The cause of this technical problem is that localized blockage alters the effective flow cross-section of each pore in the mesh screen, thereby creating a non-uniform pressure distribution at the inlet side of the activated carbon layer. This non-uniform pressure distribution is transmitted along the flow direction to the double-layer filter membrane and the outlets, causing spatial differences in flow velocity at each outlet. Since the location and degree of blockage change dynamically with operating time, its impact on the flow field distribution is non-linear and time-varying, making it difficult to compensate for with a fixed geometric correction factor. A common solution to this technical problem is to add flow velocity sensors at multiple outlets and adjust the total inlet pressure through multi-point feedback proportional-integral-derivative control. However, total pressure regulation is a global control method, which can only change the overall flow rate and cannot specifically correct the velocity differences between individual air outlets. Simultaneously, the fixed-gain proportional-integral-derivative control cannot adaptively match the current flow field state under dynamically changing blockage conditions, leading to overshoot or oscillation in the control system when blockage worsens, making it difficult to maintain the spatial uniformity of the flow velocity on the platform. This invention effectively solves this technical problem. The flow field residual variational anomaly detector reconstructs the current flow field residual through a variational autoencoder. Utilizing the deviation of the reconstruction error from the statistical distribution under normal operating conditions, it accurately identifies subtle resistance anomalies caused by local blockage of the mesh screen, triggering an alarm in the early stages of blockage and transmitting the anomaly information to the control system. The cylindrical coordinate attention mechanism of the physical fluid dynamics-constrained attention jump mechanism model explicitly encodes the radial geometric parameters of each through-hole as attention weights, enabling the model to perceive the impact of local geometric changes caused by blockage on the flow field distribution, thereby outputting high-fidelity predicted flow velocity values ​​for each air outlet, reflecting the true spatial non-uniformity. The adaptive gating regulation function, upon identifying an increase in the spatial difference in flow velocity, adjusts the regulation index accordingly. The value is automatically switched to the adjustment strategy matching the step size, and the pressure at the air inlet is precisely closed-loop adjusted so that the flow velocity of each air outlet converges in space to the target flow velocity on the table, thereby solving the technical problem that it is difficult to maintain the uniformity of the flow velocity distribution on the table under non-uniform local blockage conditions.

[0054] Specifically, the principle of this invention is:

[0055] The fundamental reason why this invention can solve the above-mentioned technical problems is that it has built a complete technical chain from four levels: flow field physical modeling, high-precision numerical solution, physical constraint neural network prediction and adaptive closed-loop control. The logical relationship between each level is rigorous and mutually supportive.

[0056] First, at the physical modeling level, the generalized Darcy-Brinkman topological differential equations simultaneously introduce Darcy drag terms and Brinkman viscous shear terms, enabling a unified description of the macroscopic seepage characteristics and near-interface viscous shear effects in porous media regions within the same equation framework. Meanwhile, the Navier-Stokes equations precisely describe the momentum conservation in the free fluid region. The two types of equations are seamlessly connected at the interface through adaptive dynamic slip boundary conditions. The slip length parameter is dynamically updated based on real-time flow velocity and pressure difference, ensuring the physical consistency of momentum transfer at the interface. The equivalent permeability matrix further quantifies the non-uniform microporous resistance characteristics of the filter membrane in matrix form, correcting the source term of the momentum conservation equation at the interface and thus eliminating numerical oscillations caused by uneven resistance distribution. This modeling strategy fundamentally eliminates the root cause of distortion in flow field calculations at the equation level.

[0057] Secondly, at the numerical solution level, the spectral element dimensionality reduction parallel solution algorithm uses high-order Chebyshev polynomials as basis functions, achieving high spatial resolution in the local strong gradient region of the concentric circular aperture with a very small number of elements. This overcomes the computational cost problem of traditional finite difference or finite volume methods requiring extremely dense meshes in strong gradient regions. The Scher complement matrix domain decomposition technique decomposes the globally coupled large matrix into independent sub-matrices for each element, enabling fully asynchronous parallel solution on multi-core processors, meeting the timeliness requirements of real-time control. The refined multi-layer nested spatial flow field topology model establishes nested meshes of different resolutions for the concentric circular aperture region, the wall boundary layer, and the axial region in cylindrical coordinates. The wall function correction mechanism analyzes the near-wall velocity distribution through logarithmic law, avoiding the computational redundancy of directly analyzing the extremely thin boundary layer. The above solution strategies ensure the high fidelity of the velocity vector results for the entire flow field, providing physically reliable label data for subsequent neural network training.

[0058] Third, at the velocity prediction level, the physical fluid dynamics constraint-based attention jump mechanism model embeds the Navier-Stokes continuity equation residuals and momentum equation residuals into the conservation judgment nodes of each physical residual block. When a feature fails to meet the physical constraints, the internal iterative gating unit forcibly triggers feedback correction, equivalent to embedding a physically interpretable iterative solution process into the forward propagation of the neural network. The cylindrical coordinate attention mechanism explicitly encodes the radial geometric non-uniformity of the concentric circular mesh sieve into attention weights, enabling the network to have directional perception of radial position-related flow field distortions, effectively avoiding the overfitting risk of pure data-driven models under small sample conditions.

[0059] Fourth, at the closed-loop control level, the adaptive gating adjustment function automatically switches between three strategies—proportional-integral-derivative, proportional-integral, or integral fine-tuning—based on the range of the adjustment index. This balances rapid response under large deviation conditions with fine stability under small deviation conditions, forming a complete closed loop that deeply integrates physical modeling and intelligent control, thereby ultimately converging the air outlet velocity to the target table velocity.

[0060] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0061] The specific implementation method of step S01 is as follows:

[0062] Sensor sampling frequency Between 20 and 200 The minimum value within the range that prevents aliasing of the flow field time series signals is selected, determined by the Nyquist theorem. Multi-source sensor data fusion set. The time-series signals collected by the inlet pressure sensor, the flow velocity sensors at each outlet, and the differential pressure sensor before and after the mesh screen are time-synchronized and fused together to form:

[0063] ;

[0064] In the formula, for Inlet pressure at all times, in units of For the first The geometric parameters of a concentric circular through hole include geometric quantities such as the through hole radius and the distance between the centers. The first air-permeable filter membrane resistance coefficient, in units of The second air-permeable filter membrane resistance coefficient, in units of For the first One air vent Real-time flow rate at any given moment, in units of For time, the unit is .

[0065] The specific implementation method of step S02 is as follows:

[0066] A generalized Darcy-Brinkman topological differential equation is established in the porous media region between the activated carbon layer and the double-layer filter membrane:

[0067] ;

[0068] In the formula, This refers to the gas dynamic viscosity, in units of... Permeability of porous media, in units of It is a velocity vector, with units of . Effective viscosity, in units of Pressure, unit: This is a volumetric force term, with units of . This refers to the volume force per unit volume. The first term on the left-hand side of the above equation... The term is Darcy's resistance, with dimensions of The first term describes the overall resistance of porous media to fluids; the second term... This is the Brinkman viscous shear term, with dimensions of... The third term describes the viscous shear effect of fluids in porous media. The pressure gradient term has dimensions of , which is the driving force for fluid flow; right side This is a volume force term with dimensions of . The dimensions are consistent with those of the quantities on the left. The steady-state pressure drop was obtained by conducting experiments on the activated carbon layer and the double-layer filter membrane, with the experiments repeated 5-10 times and the average value taken. The Darcy equation was then fitted using linear regression. Navier-Stokes equations were established in the air outlet and mesh sieve pore regions to describe the momentum conservation in the free fluid region. A spectral element dimensionality reduction parallel solution algorithm was used to divide the cylindrical flow domain into several spectral elements using high-order Chebyshev polynomials as basis functions. Stress-jumping continuous boundary conditions were applied at the interface between the porous medium and the free fluid. The global large matrix was reduced in dimensionality and decomposed into independent sub-matrices for each element using the Schur complement matrix domain decomposition technique. The solution was performed asynchronously and in parallel on a multi-core processor. High-precision velocity field rapid iterative assembly was achieved through a small amount of data exchange between boundary basis coefficients for each element.

[0069] The specific implementation method of step S03 is as follows:

[0070] An adaptive dynamic slip boundary condition is established at the interface between the first and second permeable filter membranes, with the slip length parameter... (Unit is) It updates in real time based on the current flow rate and pressure difference, within the range of 0.1 to 2.0. Segmented pressure drop experiments were conducted on the double-layer filter membrane under flow rate conditions. The pressure drop was obtained using inversion fitting of the generalized Darcy-Brinkman topological differential equation, with no fewer than 10 experimental operating points. The equivalent permeability matrix was then calculated. The non-uniform micropore resistance characteristics of the first and second air-permeable filter membranes are equivalently characterized as follows:

[0071] ;

[0072] In the formula, For the filter membrane Layer Local air permeability at radial locations, in units of The number of filter membrane layers is 2. The radial test points are evenly distributed, ranging from 5 to 15. Utilizing... The source term of the momentum conservation equation at the interface is dynamically corrected to eliminate numerical oscillations caused by abrupt changes in the flow field.

[0073] The specific implementation method of step S04 is as follows:

[0074] In cylindrical coordinate system A refined, multi-layered nested spatial flow field topology model is established below, in which... Radial coordinates (unit: ), Circumferential angular coordinates (unit: ), Axial coordinates (unit: Nested mesh topologies with different resolutions were established for the concentric circular through-hole region, the wall boundary layer region, and the axial region to capture the velocity gradient distribution in each region. In the concentric circular through-hole region, high-order tensor product interpolation basis functions were used for local refinement control. In the wall boundary layer region, a logarithmic law wall function was introduced to analytically correct the near-wall velocity distribution, thus analyzing the nonlinear coupling between radial shear flow and axial mainstream flow.

[0075] The specific implementation method of step S05 is as follows:

[0076] The attention jump mechanism model under physical fluid dynamics constraints employs an improved residual network backbone structure, with the input layer receiving a fusion set of data from multiple sensor sources. The input dimension corresponds one-to-one with the number of sensor channels. The backbone consists of multiple cascaded physical residual blocks, each containing a convolutional layer, a batch normalization layer, and a linear rectified function activation layer. The physical conservation residual judgment node calculates the Navier-Stokes continuity equation residual for the current feature map. With momentum equation residual :

[0077] ;

[0078] ;

[0079] In the formula, This is the density of the gas, in units of... The dimensions are The divergence residual characterizing the continuity equation The dimensions are The residual, representing the sum of the terms in the momentum equation, reflects the degree of deviation from the conservation of momentum. If... or If the physical tolerance threshold is exceeded, the internal iterative gating unit is activated. This unit employs a gated loop unit structure, forcibly returning feature data that does not meet the physical constraints to the previous layer of the current block for physical error correction iteration. The upper limit of iterations is 3–8 times, determined through multiple experimental comparisons of convergence speed and computational cost at different iteration numbers. The cylindrical coordinate attention mechanism uses radial position encoding as the query vector. Using the through-hole geometry parameters as the bond matrix AND-value matrix Radially weighted features are generated through scaling dot product attention operations:

[0080] ;

[0081] In the formula, For querying the matrix, the dimension is Obtained by radial position encoding linear transformation The key matrix has dimensions of . Obtained by linear transformation of the through hole geometric parameters It is a value matrix with dimension . Obtained by linear transformation of the through hole geometric parameters The dimension of the key vector Number of query vectors Number of key-value pairs Dimension of value vector The attention score is normalized using a normalized exponential function. The dimension is , is a radially weighted eigenvalue matrix, with dimensions of and . Consistent. After being fused with the main feature map via skip connections, the predicted flow velocity values ​​for each air outlet are fed into the output fully connected layer. (Unit is) Network loss function It consists of a weighted sum of three parts:

[0082] ;

[0083] In the formula, The data error weighting coefficient is dimensionless. These are the residual weighting coefficients for the momentum equation, in units of... ,make Dimensions are These are the residual weighting coefficients for the continuity equation, in units of... ,make Dimensions are Total number of air outlets For the first Predicted flow velocity values ​​for each air outlet, in units of For the first The measured flow velocity at each air outlet, in units of Three-part weighting coefficients , , Through grid search experiments on the training set, the search range was determined to be 0.01–10, with a step size of 0.1, so that the weighted sum of the three parts has a unified dimension. The variational anomaly detector for flow field residuals reconstructs the current flow field residuals based on a variational autoencoder. When the reconstruction error exceeds an anomaly threshold... The screen was determined to be clogged or malfunctioning at that time. The 95th percentile of the reconstruction error is determined by collecting 50 to 200 sets of sensor data under normal operating conditions.

[0084] The specific implementation method of step S06 is as follows:

[0085] Define the adjustment index The predicted flow velocity of each air outlet and the target table surface flow velocity are given. Weighted root mean square of the deviation:

[0086] ;

[0087] In the formula, The target surface velocity, in units of For the first The contribution coefficient of the radial position of each air outlet to the uniformity of the flow velocity on the table surface is dimensionless and obtained by regression fitting of 10-20 sets of uniformity experimental data. The unit is This reflects the weighted deviation between the flow velocity at each outlet and the target flow velocity. Based on... The corresponding adjustment strategy is adopted for the corresponding interval: when At that time, a large-step proportional-integral-derivative adjustment strategy is adopted to perform rapid and large-scale adjustment of the inlet pressure, while the iteration upper limit of the internal iterative gating unit is dynamically set to the maximum value to improve the accuracy of physical constraint satisfaction; when At that time, a medium-step proportional-integral control strategy is used to smoothly regulate the inlet pressure; when At the same time, a small-step integral fine-tuning strategy is used to perform fine adjustment of the inlet air pressure, while the iteration upper limit of the internal iterative gating unit is dynamically set to the minimum value to reduce inference latency. and By conducting step response experiments near the target platform velocity, the balance point between regulation efficiency and stability was determined under different deviation ranges. At least 20 experimental groups were conducted. Closed-loop control was implemented on the inlet pressure to converge the outlet velocity to the target platform velocity. .

[0088] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0089] This embodiment uses a cylindrical activated carbon exhaust filtration device as the object to verify the flow rate control effect of the control method described in this invention under actual operating conditions. The device has a cylindrical diameter of 0.3 m and is arranged sequentially along the airflow direction as an air inlet, a mesh screen, an activated carbon layer, a first permeable filter membrane, a second permeable filter membrane, and an air outlet. The air outlet has a total of 12 air outlet holes, divided into two groups radially: an inner ring of 4 holes and an outer ring of 8 holes. The target flow rate on the platform is set to 0.5 m / s, and the experimental ambient temperature is 25℃.

[0090] The sensor arrangement is as follows: a pressure sensor with a range of 0–600 Pa is installed at the air inlet; a hot-wire velocity sensor is installed at each of the 12 air outlets; and a differential pressure sensor is installed before and after the mesh screen. After multiple experiments, the sampling frequency was determined to be 50 Hz to ensure that the flow field time-series signals do not aliasing. The sensor signals are then time-synchronized and fused into a multi-source sensor data fusion set, serving as the unified data source for subsequent flow field calculations and model input.

[0091] The permeability calibration experiment was conducted on both the activated carbon layer and the double-layer filter membrane. Each operating condition was repeated 7 times, and the average value was taken as the permeability of each layer. The calibration results. Among them, the permeability of the activated carbon layer. The calibration result is The permeability of the first breathable filter membrane (polytetrafluoroethylene) for The permeability of the second breathable filter membrane (polypropylene nonwoven fabric) for The above parameters serve as inputs to the generalized Darcy-Brinkman topological differential equations, which, together with the Navier-Stokes equations, constitute the partitioned coupled flow field equation set. The spectral element dimensionality reduction parallel solution algorithm uses high-order Chebyshev polynomials as basis functions to divide the cylindrical flow domain into 18 spectral elements. The algorithm is solved in parallel on an 8-core processor using Scher's complement matrix domain decomposition technology. The time for a single full-field velocity vector solution is approximately 0.12 s, meeting the time requirements for real-time control.

[0092] The equivalent permeability matrix was calibrated by uniformly arranging 10 test points along the radial direction of the double-layer filter membrane, measuring the local pressure drop at each point, and constructing a 10×10 equivalent permeability matrix. The slip length parameter of the adaptive dynamic slip boundary condition was obtained through piecewise pressure drop experiments at 12 flow velocity points within the range of 0.1–2.0 m / s, and obtained via inversion fitting. The refined multi-layer nested spatial flow field topology model used the finest mesh in cylindrical coordinates for the concentric circular through-hole region, applied logarithmic law wall function correction for the wall boundary layer region, and used a coarser mesh for the axial region, forming a three-level nested structure.

[0093] The training dataset for the attention-jumping mechanism model constrained by physical fluid dynamics was collected under a full combination of nine inlet pressure conditions (50 Pa, 100 Pa, 150 Pa, 200 Pa, 250 Pa, 300 Pa, 350 Pa, 400 Pa, 500 Pa), seven activated carbon layer blockage levels, and five ambient temperatures (15℃, 20℃, 25℃, 30℃, 40℃). Each condition was collected for 300 seconds, resulting in 315 sets of sensor data. The velocity vector of the entire flow field output by the spectral reduction parallel solution algorithm was used as the physical label to construct training sample pairs. The training, validation, and test sets were divided in an 8:1:1 ratio. An adaptive moment estimation optimizer was used, with an initial learning rate of... The cosine annealing strategy was used to adjust the number of input samples per batch to 32. The weight coefficient of the physical constraint loss was determined to be 0.5 in the range of 0.01 to 10 through grid search. Training was stopped when the loss on the validation set no longer decreased after 30 consecutive training rounds.

[0094] Table 1 shows the typical steady-state acquisition data of each sensor under different operating conditions:

[0095] Table 1. Sensor Steady-State Acquisition Data under Typical Operating Conditions

[0096]

[0097] Condition 5 corresponds to a moderately clogged activated carbon layer, with a consistently high mesh screen pressure difference and a significantly increased difference in flow velocity between the inner and outer rings. The reconstruction error of the flow field residual variational anomaly detector exceeds the anomaly threshold, triggering a clog anomaly alarm. Under this condition, the model uses a cylindrical coordinate attention mechanism to weightedly perceive radial geometric parameters, outputting corrected flow velocity predictions for each outlet and adjusting the index. The calculated result is 0.087 m / s, which exceeds the large deviation threshold. (Value taken as 0.06 m / s), the system automatically switches to a large-step proportional-integral-derivative adjustment strategy, rapidly and significantly adjusting the inlet pressure, causing the outlet velocity to converge to approximately 0.5 m / s on the target platform within about 8 seconds, reducing the velocity deviation of each orifice to within 0.01 m / s. Figure 2 As shown.

[0098] Under normal operating conditions, the adjustment index Stabilized at small deviation threshold (Values ​​below 0.02 m / s) maintain a small-step integral fine-tuning strategy, resulting in minimal adjustment of the inlet pressure and highly stable flow velocity on the platform. The steady-state distribution of flow velocity at each outlet under different operating conditions is shown below. Figure 3 As shown, the air velocity uniformity of the inner and outer ring outlet holes is good, and the maximum velocity difference between holes does not exceed 0.015 m / s.

[0099] The root mean square error of the model's velocity prediction for each air outlet on the test set is summarized in Table 2:

[0100] Table 2 Summary of Root Mean Square Errors for Predicting Flow Velocity at Each Air Outlet

[0101]

[0102] The root mean square errors mentioned above all meet the upper limit threshold (0.02 m / s) determined by the engineering accuracy requirements, verifying the physical self-consistent prediction capability of the attention jump mechanism model with physical fluid dynamics constraints under small sample conditions.

[0103] The switching logic of the three-level regulation strategy of the adaptive gating function was determined through 20 sets of step response experiments. The value is 0.02 m / s. With a value of 0.06 m / s, each strategy exhibits good convergence and stability within the corresponding deviation interval. Adjustment index The curve of change over time is as follows Figure 4 As shown, the system's adjustment process and convergence characteristics during operating condition switching are clearly illustrated. It should be noted that the activated carbon exhaust device involved in this embodiment is as follows: Figure 5 As shown.

[0104] The advancements of this invention over traditional methods are reflected in the following aspects. Traditional methods estimate the outlet velocity using a simplified Bernoulli equation, neglecting the fundamental shift in momentum transfer equations at the interface between porous media and free fluid, leading to systematic biases in flow field calculations at the interface. This invention eliminates the physical root of this bias at the equation level by partitioning the generalized Darcy-Brinkman topological differential equation with the Navier-Stokes equation, ensuring that the flow field calculation results accurately reflect the resistance characteristics of each medium layer. Traditional neural network models rely solely on data fitting, easily leading to overfitting or physically inconsistent predictions under small sample conditions or unconventional clogging conditions. This invention embeds the residuals of the fluid dynamics equations into the neural network feature propagation process, subjecting the prediction results to hard constraints of physical laws under any condition, fundamentally guaranteeing the physical reliability of the predictions. Traditional proportional-integral-derivative closed-loop control uses a fixed gain, which cannot adaptively match the adjustment step size under different deviation levels, and is prone to problems such as slow response under large deviations or overshoot oscillation under small deviations. The adaptive gating control function of this invention enables the control system to converge with the optimal strategy in the entire deviation range through automatic switching of three deviation intervals, thus solving the problem of insufficient adaptability of fixed gain control under dynamic conditions.

[0105] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0106] Table 3. Variable Explanation Table

[0107]

[0108] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for controlling the flow velocity on a test bench considering an exhaust activated carbon filter, characterized in that, Includes the following steps: Collect inlet pressure, concentric circular through-hole geometric parameters, first permeable filter membrane resistance coefficient, second permeable filter membrane resistance coefficient, and outlet air velocity data to establish a multi-source sensor data fusion set; A generalized Darcy-Brinkman topological differential equation was established for the activated carbon layer and the double-layer filter membrane region, and a Navier-Stokes equation was established for the air outlet and the mesh screen through-hole region. The velocity vector of the entire flow field was solved by the spectral element dimension reduction parallel solution algorithm. An adaptive dynamic slip boundary condition is established at the interface between the first and second permeable filter membranes. The source term of the momentum conservation equation at the interface is dynamically corrected using the equivalent permeability matrix to eliminate numerical oscillations caused by abrupt changes in the flow field. A refined multi-layered nested spatial flow field topology model is established based on cylindrical coordinates. A high-order tensor product is used to locally refine the control of the concentric circular through-hole region. A wall function correction mechanism is introduced at the wall boundary to analyze the nonlinear coupling between radial shear flow and axial mainstream flow. The multi-source sensor data fusion set is input into the attention jump mechanism model constrained by physical fluid dynamics, outputting the predicted flow velocity values ​​of each air outlet, and identifying local weak resistance anomalies caused by mesh screen blockage through the flow field residual variational anomaly detector. Based on the deviation between the predicted flow velocity output by the attention jump mechanism model constrained by physical fluid dynamics and the flow velocity on the target platform, the adaptive gating adjustment function value is calculated. According to the interval to which the function value belongs, the corresponding adjustment strategy is adopted to perform closed-loop control of the inlet pressure, so that the flow velocity at the outlet converges to the flow velocity on the target platform.

2. The control method according to claim 1, characterized in that, The multi-source sensor data fusion set specifically involves merging the time-series signals collected by the inlet pressure sensor, the flow velocity sensors of each outlet, and the differential pressure sensor before and after the mesh screen into a unified data set after time synchronization and alignment. The sampling frequency of the sensors is taken as the minimum value within the frequency range that prevents the flow field time-series signals from aliasing.

3. The control method according to claim 2, characterized in that, The variational anomaly detector for flow field residuals is based on the statistical distribution of residual signals under normal operating conditions. It constructs a variational autoencoder to reconstruct the current flow field residuals. When the reconstruction error exceeds the anomaly threshold, it is determined that there is blockage or abnormality in the mesh screen. The anomaly threshold is determined by calculating the 95th percentile of the reconstruction error under normal operating conditions.

4. The control method according to claim 3, characterized in that, The spectral element dimensionality reduction parallel solution algorithm specifically uses high-order Chebyshev polynomials as basis functions to transform partial differential equations into a system of algebraic equations, constructs stress jump continuous boundary conditions at the interface between porous media and free fluid, and uses Scher complement matrix domain decomposition technology to reduce the dimensionality of the global large matrix into independent sub-matrices of each element, and performs fully asynchronous parallel solution on a multi-core processor.

5. The control method according to claim 4, characterized in that, The slip length parameter of the adaptive dynamic slip boundary condition is obtained by performing segmented pressure drop experiments on the double-layer filter membrane under different flow rate conditions and then fitting it using the generalized Darcy-Brinkman topological differential equation; the matrix elements of the equivalent air permeability matrix are obtained by performing local pressure drop tests on different radial positions of the filter membrane, with the test points evenly distributed in the radial direction.

6. The control method according to claim 5, characterized in that, The refined multi-layer nested spatial flow field topology model is specifically designed in cylindrical coordinates, where nested mesh topologies of different resolutions are established for the concentric circular through-hole region, the wall boundary layer region, and the axis region. The wall function correction mechanism uses a logarithmic wall function to analytically correct the near-wall velocity distribution.

7. The control method according to claim 6, characterized in that, The attention jump mechanism model with physical fluid dynamics constraints adopts an improved residual network backbone structure. The backbone is composed of multiple cascaded physical residual blocks. Each physical residual block is equipped with a physical conservation residual judgment node. The residuals of the Navier-Stokes continuity equation and momentum equation are calculated for the current feature map. When the residual exceeds the physical tolerance threshold, the internal iterative gating unit is activated.

8. The control method according to claim 7, characterized in that, The internal iterative gating unit adopts a gated loop unit structure, which returns feature data that does not meet the physical constraints to the previous layer of the current block for physical error correction iteration. The upper limit of the number of iterations is determined by experimental comparison of convergence speed and computational cost under different number of iterations.

9. The control method according to claim 8, characterized in that, The attention skipping mechanism model of physical fluid dynamics constraints introduces a cylindrical coordinate attention mechanism, using radial position encoding as the query vector and the geometric parameters of concentric circular through holes as key-value pairs. It generates radially weighted features through scaling dot product attention operations and then performs skip connections and fusion with the backbone feature map.

10. The control method according to claim 9, characterized in that, The loss function of the attention jump mechanism model under physical fluid dynamics constraints is composed of three weighted sums: the mean square error between the predicted and measured flow velocities at each outlet, the momentum conservation differential residual term, and the mass conservation differential residual term. The weight coefficients are determined on the training set through grid search experiments.

Images