Multi-stage filtration method and system for wastewater oil separation based on coalescing materials

Abstract

The application provides a wastewater oil liquid separation multi-stage filtering treatment method and system based on coalescence material, relates to the technical field of wastewater treatment, and comprises the following steps: tracking the oil droplet motion trajectory through deep reinforcement learning, and constructing a filtering layer space fiber network structure; analyzing the oil droplet coalescence network topological structure by using a graph neural network, determining the coalescence position and expansion trend; accordingly, the filtering state is evaluated, an evolution potential field is constructed to optimize the filtering layer parameters, and the iterative optimization of the filtering device is realized. The application improves the wastewater oil liquid separation efficiency and prolongs the service life of the filtering device.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wastewater treatment, and particularly relates to a wastewater oil liquid separation multi-stage filtration treatment method and system based on coalescence materials. BACKGROUND

[0002] With the acceleration of industrialization, the treatment of oil-water mixed wastewater has become a key problem in environmental protection. Oil-water separation technology has been widely used in petrochemical industry, food processing, shipping and aviation industries. Traditional oil-water separation methods mainly include gravity separation, centrifugal separation, air flotation separation and membrane separation, etc., while coalescence separation as a high-efficiency oil-water separation technology has attracted widespread attention due to its high treatment efficiency and wide application range. Coalescence filtration is a process of using the lipophilicity of the surface of the filter material to make small oil droplets adhere and coalesce into larger oil droplets on the surface of the filter material, and finally separate out the water phase by gravity. Multi-stage coalescence filtration technology can effectively treat complex wastewater with high oil content and small oil droplet size by setting filter layers with different properties and gradually improving the oil-water separation efficiency.

[0003] However, the existing multi-stage coalescence filtration technology has obvious deficiencies in practical application. The design of traditional coalescence filter layer lacks accurate analysis of the micro oil droplet motion behavior, and the setting of filter structure parameters mainly relies on experience and simplified models, which makes it difficult to accurately predict the coalescence behavior of oil droplets in complex fiber networks, resulting in unstable filtration efficiency. The existing technology cannot monitor and evaluate the use state of each layer of the multi-stage filtration device in real time, and lacks a quantitative evaluation method for the degree of filter layer blockage and remaining service life, often leading to premature replacement or excessive use of filter materials, increasing operating costs. The parameter optimization process of the existing coalescence filtration device lacks systematicness and adaptability, and it is difficult to dynamically adjust the filter layer structure parameters according to different wastewater characteristics and filtration states, which cannot realize the continuous optimization of filtration efficiency, especially when dealing with industrial wastewater with complex composition and uneven oil droplet size distribution, the adaptability is poor. SUMMARY

[0004] The embodiment of the present application provides a wastewater oil liquid separation multi-stage filtration treatment method and system based on coalescence materials, which can solve the problems in the prior art.

[0005] In a first aspect, the embodiment of the present application provides a wastewater oil liquid separation multi-stage filtration treatment method based on coalescence materials, comprising:

[0006] Obtaining the structure information and operating parameters of the multi-stage filtration device;

[0007] Using a deep reinforcement learning method to track and analyze oil droplet particles, obtaining oil droplet motion trajectories, constructing a filter layer space fiber network structure based on the oil droplet motion trajectories, and obtaining three-dimensional pore distribution parameters of the coalescence filter layer through pore division based on bubble growth.

[0008] inputting the three-dimensional pore distribution parameters into the graph neural network, constructing an oil droplet coalescence network topology, analyzing oil droplet migration strength to determine a coalescence channel, calculating spatial coordinates and coalescence rate of an oil droplet coalescence point, and determining an oil droplet coalescence position and coalescence expansion trend of the coalescence filter layer;

[0009] According to the oil droplet coalescence position and coalescence expansion trend, combined with the structural information and operating parameters, the residual service life and oil-water separation efficiency of each coalescence filter layer are calculated to generate filter state evaluation data;

[0010] Based on the filter state evaluation data, an evolution potential field is constructed to optimize the calculation of the fiber arrangement angle and the fiber spacing, and improved filter layer parameters are obtained.

[0011] The improved filter layer parameters are applied to the multi-stage coalescence filter device, and real-time oil-water separation data are collected as new operating parameters for optimization.

[0012] In an alternative embodiment, a deep reinforcement learning method is used to track and analyze oil droplet particles to obtain oil droplet motion trajectories, including:

[0013] Collecting motion state parameters and interface action parameters of oil droplet particles in a flow field to generate initial feature data; calculating the force state of the oil droplet particles in the flow field according to the initial feature data to generate force state data;

[0014] Input the initial feature data and the force state data into the deep reinforcement learning network to predict the motion trajectory and generate trajectory prediction data; collect the actual motion trajectory and compare it with the trajectory prediction data to calculate the trajectory evaluation value, which is used as the feedback signal of the deep reinforcement learning network; according to the feedback signal, optimize the deep reinforcement learning network and adjust the training parameters until the prediction deviation of the trajectory prediction data is less than a preset deviation threshold, and determine the optimized deep reinforcement learning network;

[0015] Using the optimized deep reinforcement learning network to continuously predict the motion trajectory and generate continuous trajectory prediction data; real-time collection of actual motion trajectory data, comparison of the actual motion trajectory data with the continuous trajectory prediction data, correction of the continuous trajectory prediction data, and finally obtaining the oil droplet motion trajectory.

[0016] In an alternative embodiment, a filter layer space fiber network structure is constructed based on the oil droplet motion trajectory, and three-dimensional pore distribution parameters of the coalescence filter layer are obtained through pore division based on bubble growth, including:

[0017] Extracting feature parameters of the oil droplet motion trajectory, obtaining the spatial distribution of the trajectory turning points and the frequency of the motion direction changes, constructing a trajectory density distribution map, and determining the oil droplet motion channel distribution;

[0018] establishing a fiber distribution probability density field of the filter layer according to the trajectory density distribution map, taking the spatial distribution of the trajectory turning points as a prior distribution, combining the frequency of the change of the motion direction as a conditional probability, and obtaining a posterior probability distribution of the fiber intersection nodes through Bayesian iteration calculation, and determining the fiber orientation angle according to the posterior probability distribution;

[0019] discretizing the fiber distribution probability density field into grid cells, randomly generating fiber position samples in each grid cell according to the posterior probability distribution of the fiber intersection nodes, performing Markov chain iteration update on the fiber position samples, calculating an acceptance probability of each iteration, obtaining fiber spatial distribution data when the acceptance probability converges, optimizing the connection relationship of the fiber intersection nodes based on the fiber spatial distribution data, and constructing a filter layer spatial fiber network structure;

[0020] performing pore division on the filter layer spatial fiber network structure based on bubble growth to obtain pore cells, calculating pore volume distribution and pore throat ratio data, and establishing a pore connectivity relationship;

[0021] calculating three-dimensional pore distribution parameters of the coalesced filter layer according to the pore connectivity relationship.

[0022] In an optional embodiment, performing pore division on the filter layer spatial fiber network structure based on bubble growth to obtain pore cells, calculating pore volume distribution and pore throat ratio data, and establishing a pore connectivity relationship includes:

[0023] calculating distance values from each grid point in the filter layer spatial fiber network structure to surrounding fibers to obtain distance field data, finding a local maximum value position in the distance field data, determining the local maximum value position as a bubble core point, and determining the distance value as an initial radius of a corresponding bubble;

[0024] expanding the bubble with the bubble core point as the center, the expansion rate of the bubble being inversely proportional to the local density of the fibers in the filter layer spatial fiber network structure, recording contact position points of the bubble and the fibers, and stopping the expansion when adjacent bubbles contact;

[0025] constructing a common boundary at the contact position of the adjacent bubbles to obtain an initial pore cell, optimizing the curvature of the common boundary according to the spatial distribution characteristics of the fibers, and merging the initial pore cell with a volume less than a preset volume threshold and an adjacent pore cell to obtain a final pore cell;

[0026] calculating the volume and surface area of the final pore cell, identifying a minimum cross-section position between adjacent pore cells as a pore throat, calculating the ratio of the size of the pore throat to the volume of the corresponding pore cell to obtain pore throat ratio data;

[0027] According to the space connection relationship between the final pore units and the pore throat ratio data, a pore connectivity relationship is established, and the connectivity number and connectivity strength of each pore unit are determined.

[0028] In an alternative embodiment, three-dimensional pore distribution parameters are input into a graph neural network to construct an oil droplet coalescence network topology, analyze oil droplet migration strength to determine coalescence channels, calculate the spatial coordinates and coalescence rate of oil droplet coalescence points, and determine the oil droplet coalescence position and coalescence expansion trend of the coalescence filter layer, including:

[0029] Three-dimensional pore distribution parameters are input into a graph neural network, and an oil droplet coalescence network topology is constructed according to the pore connectivity relationship. The pore units are set as network nodes, the volume and surface area of the pore units are taken as node features, and the pore throat ratio data and connectivity strength are taken as edge features.

[0030] The oil droplet coalescence network topology is calculated by a graph neural network, node feature information is transmitted based on a directional convolution operator, and coalescence feature representation of the nodes is obtained by feature fusion according to the pore throat ratio data.

[0031] The oil droplet migration strength between adjacent pore units is calculated according to the coalescence feature representation of the nodes, pore unit sequences are formed by pore units with oil droplet migration strength greater than a preset strength threshold, and the pore unit sequences are determined as oil droplet coalescence channels.

[0032] The collision frequency of oil droplets is calculated in the oil droplet coalescence channels according to the connectivity strength, the spatial coordinates of the oil droplet coalescence point are determined as the position with the highest collision frequency, and the oil droplet coalescence rate is determined according to the collision frequency.

[0033] The attenuation value of the coalescence rate is calculated along the direction of the pore connectivity relationship with the oil droplet coalescence point as the starting position, the influence range of oil droplet coalescence is determined according to the attenuation value, and the oil droplet coalescence position distribution and coalescence expansion trend of the coalescence filter layer are obtained.

[0034] In an alternative embodiment, the oil droplet coalescence network topology is calculated by a graph neural network, node feature information is transmitted based on a directional convolution operator, and coalescence feature representation of the nodes is obtained by feature fusion according to the pore throat ratio data, including:

[0035] A directional convolution operator is constructed according to the spatial distribution characteristics of the pore units, the volume and surface area of the pore units are projected onto the principal axis direction, the anisotropy degree is calculated, and the anisotropy degree is taken as the shape parameter of the convolution kernel.

[0036] The node features are directionally convoluted by the directional convolution operator, local structure features are extracted along the main axis direction, and anisotropic weights are set for feature information in different directions, and the anisotropic weights are positively correlated with the anisotropy degree.

[0037] The convolution results in different directions are adaptively fused, the direction importance score is established based on the pore throat ratio data, the fusion weight is dynamically adjusted according to the direction importance score, and the agglomeration feature representation of the node is obtained.

[0038] In an optional embodiment, an evolution potential field is constructed based on the filtration state evaluation data, and the fiber arrangement angle and the fiber spacing are optimized to obtain improved filter layer parameters, including:

[0039] According to the filtration state evaluation data, an evolution potential field is constructed, regions with oil-water separation efficiency greater than a preset efficiency threshold are set as positive potential energy regions, and regions with oil-water separation efficiency less than the preset efficiency threshold are set as negative potential energy regions, and the evolution path direction is determined according to the remaining use time of the agglomeration filter layer;

[0040] The potential energy values of the positive potential energy regions and the negative potential energy regions are determined, the search space is determined according to the potential energy distribution of the evolution potential field, the fiber arrangement angle and the fiber spacing are used as optimization variables, and the search step of the optimization variables is set based on the potential energy values, wherein the corresponding search step of the positive potential energy region is greater than the corresponding search step of the negative potential energy region;

[0041] The potential energy gradient in the evolution potential field is calculated, the potential energy gradient is used as the adjustment direction of the fiber structure, the fiber spacing of the positive potential energy region is adjusted to a first preset spacing according to the potential energy gradient, and the fiber spacing of the negative potential energy region is adjusted to a second preset spacing, wherein the first preset spacing value is greater than the second preset spacing value;

[0042] The fiber arrangement angle is calculated based on the evolution path direction, so that the included angle between the fiber arrangement angle and the evolution path direction is kept within a preset angle range;

[0043] According to the optimization results of the fiber arrangement angle and the fiber spacing, the pore distribution structure of the agglomeration filter layer is adjusted to obtain improved filter layer structure parameters.

[0044] In a second aspect of the embodiment of the application, a wastewater oil liquid separation multi-stage filtration processing system based on agglomeration material is provided, including:

[0045] The first unit is used for acquiring structure information and operation parameters of the multi-stage filtration device.

[0046] The second unit is configured to track and analyze the oil droplet particles by using a deep reinforcement learning method, to obtain an oil droplet motion trajectory, to construct a filter layer space fiber network structure based on the oil droplet motion trajectory, and to obtain three-dimensional pore distribution parameters of the coalescence filter layer by pore division based on bubble growth.

[0047] The third unit is configured to input the three-dimensional pore distribution parameters into a graph neural network, to construct an oil droplet coalescence network topology, to analyze oil droplet migration strength to determine a coalescence channel, to calculate spatial coordinates and coalescence rates of oil droplet coalescence points, and to determine oil droplet coalescence positions and coalescence expansion trends of the coalescence filter layer.

[0048] The fourth unit is configured to calculate residual use time and oil-water separation efficiency of each coalescence filter layer according to the oil droplet coalescence positions and coalescence expansion trends, in combination with structure information and operating parameters, to generate filter state evaluation data.

[0049] The fifth unit is configured to construct an evolution potential field based on the filter state evaluation data, to perform optimization calculation on fiber arrangement angles and fiber spacings, and to obtain improved filter layer parameters.

[0050] The sixth unit is configured to apply the improved filter layer parameters to a multi-stage coalescence filter device, to collect real-time oil-water separation data as new operating parameters, and to perform optimization in a cyclic iteration manner.

[0051] In a third aspect, an electronic device is provided, including:

[0052] a processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0055] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0056] In the embodiment of the present application, the oil droplet coalescence network topology is constructed by deep reinforcement learning and graph neural network, which can accurately capture the motion characteristics and coalescence behavior of oil droplets in the coalescence filter layer, improve the accuracy of oil-water separation process modeling, reduce system error, and thus realize more efficient wastewater treatment in practical application; based on the filter state evaluation data, the evolution potential field is constructed to optimize the calculation of the fiber arrangement angle and the fiber spacing, realizing the adaptive adjustment of the coalescence filter layer structure, significantly improving the oil-water separation efficiency, prolonging the service life of the filter device, reducing the maintenance cost and operation complexity; the method combining multi-stage filtration with cyclic iteration optimization is adopted to realize the continuous self-improvement of the wastewater treatment system, which can dynamically adjust the filtration parameters according to the real-time operation data, adapt to the oil-water separation requirements under different working conditions, has strong universality and adaptability, and provides an efficient and reliable technical solution for industrial wastewater treatment. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 FIG. 1 is a flowchart of a wastewater oil liquid separation multi-stage filtration method based on coalescence materials according to an embodiment of the present application.

[0058] Figure 2 FIG. 4 is a directional convolution and adaptive fusion logic flowchart. DETAILED DESCRIPTION

[0059] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0060] The technical scheme of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0061] Figure 1 FIG. 1 is a flowchart of a wastewater oil liquid separation multi-stage filtration method based on coalescence materials according to an embodiment of the present application, as shown in the figure, the method comprises: Figure 1

[0062] Obtaining structure information and operation parameters of a multi-stage filtration device;

[0063] Using a deep reinforcement learning method to track and analyze oil droplet particles, obtaining oil droplet motion trajectories, constructing a filter layer space fiber network structure based on the oil droplet motion trajectories, and obtaining three-dimensional pore distribution parameters of the coalescence filter layer through pore division based on bubble growth;​

[0064] The three-dimensional pore distribution parameters are input into a graph neural network to construct an oil droplet coalescence network topology, analyze oil droplet migration strength to determine coalescence channels, calculate the spatial coordinates and coalescence rate of oil droplet coalescence points, and determine the oil droplet coalescence position and coalescence expansion trend of the coalescence filter layer;

[0065] According to the oil droplet coalescence position and coalescence expansion trend, combined with the structural information and operating parameters, the remaining service life and oil-water separation efficiency of each coalescence filter layer are calculated to generate filter state evaluation data;

[0066] Based on the filter state evaluation data, an evolution potential field is constructed to optimize the calculation of the fiber arrangement angle and fiber spacing, and improved filter layer parameters are obtained;

[0067] The improved filter layer parameters are applied to the multi-stage coalescence filter device, and real-time oil-water separation data are collected as new operating parameters for optimization.

[0068] In a specific embodiment, the implementation process of the wastewater oil separation multi-stage filtration method based on coalescence materials first needs to obtain the structural information and operating parameters of the multi-stage filtration device, including the size specifications of the device, the number of filter layers, material properties, and current operating parameters such as flow rate, pressure, and temperature. These information will be used as the basis data for subsequent analysis, providing the necessary boundary conditions and initial state for the entire processing process.

[0069] Deep reinforcement learning method is used to track and analyze oil droplet particles. High-speed camera system is used to capture the motion images of oil droplets in the flow field, extract the position, velocity, acceleration and other motion state parameters of oil droplets, as well as the interfacial interaction parameters between oil droplets, water molecules and fiber materials, and generate initial feature data. Then, according to these data, the force state of oil droplets in the flow field, such as gravity, buoyancy, drag force and surface tension, is calculated. These feature data and force state data are input into the pre-designed deep reinforcement learning network, and the motion trajectory of the oil droplet is predicted through the multi-layer neural network structure. At the same time, the actual motion trajectory of the oil droplet is collected and compared with the predicted trajectory to calculate the prediction accuracy, and this evaluation value is used as a feedback signal to optimize the deep reinforcement learning network, and the network parameters are continuously adjusted until the prediction deviation reaches an acceptable range. For example, when processing oil-containing wastewater from a refinery, 500 frames of images per second are captured by a high-speed camera to track oil droplet particles with a diameter of 50-200 μm. After 50 rounds of iterative training, the average deviation between the predicted trajectory and the actual trajectory is reduced to less than 5%.

[0070] Based on the obtained oil droplet motion trajectory, the spatial fiber network structure of the filter layer is constructed, the characteristic parameters of the oil droplet motion trajectory such as the spatial distribution of the trajectory turning points and the frequency of the motion direction change are extracted, the distribution map reflecting the motion density of the oil droplets is constructed, and thus the main motion channel of the oil droplets is determined. According to the distribution information, the probability density field of the fiber distribution of the filter layer is established, the trajectory turning points are taken as the prior knowledge, the frequency of the direction change is taken as the conditional probability, the posterior probability distribution of the fiber cross nodes is obtained through the Bayesian iterative calculation, and thus the orientation angle of the fiber is determined. The probability density field is discretized into grid cells, the fiber position samples are generated in each cell according to the posterior probability, and the spatial distribution of the fiber is updated and optimized through the Markov chain iteration.

[0071] The obtained three-dimensional pore distribution parameters are input into a graph neural network, and the oil droplet coalescence network topology structure is constructed. In this structure, the pore cells are set as network nodes, the volume and surface area of the pores are taken as node features, and the pore throat ratio data and connectivity strength are taken as edge features. The graph neural network calculation is performed on the topology structure, the node feature information is transmitted through the directional convolution operator, the coalescence feature representation of each node is obtained through the feature fusion according to the pore throat ratio data, the oil droplet migration strength between adjacent pore cells is calculated based on this, the pore cells with high migration strength are combined into a sequence, and the oil droplet coalescence channels are determined. In these channels, the collision frequency of the oil droplets is calculated according to the connectivity strength, the spatial coordinates of the oil droplet coalescence points and the coalescence rate are determined, and thus the oil droplet coalescence position distribution and the expansion trend of the coalescence filter layer are obtained.

[0072] According to the coalescence position and the expansion trend, the remaining service life and the oil-water separation efficiency of each coalescence filter layer are calculated based on the device structure information and the operating parameters, and the filter state evaluation data is generated. Based on these evaluation data, the evolution potential field is constructed, the high-efficiency area is set as the positive potential area, the low-efficiency area is set as the negative potential area, and the evolution path direction is determined by the remaining service life. According to the potential energy distribution, the search space is determined, the fiber arrangement angle and the spacing are taken as the optimization variables, the potential energy gradient is calculated as the adjustment direction of the fiber structure, the fiber spacing and the arrangement angle in different areas are optimized, and finally the improved filter layer structure parameters are obtained.

[0073] The improved filter layer parameters are applied to the multi-stage coalescence filter device, the real-time oil-water separation data is collected as new operating parameters, and the optimization is iterated again to realize the continuous improvement of the oil-water separation and filtration efficiency of the wastewater. For example, in the wastewater treatment of a petrochemical enterprise, after three rounds of optimization iteration, the oil-water separation efficiency of the device is improved from 85% to 97%, the service life of the filter layer is prolonged by 40%, and the operating cost is greatly reduced.

[0074] In an alternative embodiment, a deep reinforcement learning method is used to track and analyze oil droplet particles, and the oil droplet motion trajectory is obtained, including:

[0075] Collecting the motion state parameters and interface action parameters of the oil droplet particles in the flow field to generate initial feature data; calculating the force state of the oil droplet particles in the flow field according to the initial feature data to generate force state data;

[0076] Inputting the initial feature data and the force state data into a deep reinforcement learning network to predict the motion trajectory and generate trajectory prediction data; collecting the actual motion trajectory and comparing it with the trajectory prediction data to calculate a trajectory evaluation value, which is used as a feedback signal of the deep reinforcement learning network; optimizing the deep reinforcement learning network according to the feedback signal, adjusting the training parameters, until the prediction deviation of the trajectory prediction data is less than a preset deviation threshold, and determining the optimized deep reinforcement learning network;

[0077] Using the optimized deep reinforcement learning network to continuously predict the motion trajectory and generate continuous trajectory prediction data; collecting real-time actual motion trajectory data, comparing the actual motion trajectory data with the continuous trajectory prediction data, and correcting the continuous trajectory prediction data to obtain the final oil droplet motion trajectory.

[0078] According to the technical scheme of tracking the motion trajectory of the oil droplet particles by deep reinforcement learning, the embodiment describes the specific implementation process in detail. In this embodiment, the deep reinforcement learning network adopts a double neural network structure, including a policy network and a value network, for predicting the motion trajectory of the oil droplet particles.

[0079] In actual application, first, a high-speed camera system is used to collect videos of the oil droplet particles in the flow field in the microchannel, the frame rate is set to 1000 frames / s, and the resolution is 1920×1080 pixels. During the collection process, the system records the position coordinates (x, y, z), velocity (vx, vy, vz), acceleration (ax, ay, az), and angular velocity (ωx, ωy, ωz) of the oil droplet particles as the motion state parameters; at the same time, the surface tension coefficient (0.072 N / m), contact angle (120°), and interface potential (-30 mV) of the oil droplet particles are recorded as the interface action parameters to generate initial feature data.

[0080] For the initial feature data, the system calculates the force state of the oil droplet particles in the flow field, including gravity (9.8×10 -15 N), buoyancy (8.5×10 -15 N), viscous drag (3.2×10 -14 N), surface tension (4.7×10 -14 N), and electrostatic force (1.8×10-14 N) and Brownian force (random distribution, mean 5.0 x 10 -16 N). The resultant force determines the motion state of the oil droplet particle, and the system calculates the force state data through Newton's second law.

[0081] The initial feature data and force state data are combined into a feature vector and input into the deep reinforcement learning network. The network consists of an input layer, four hidden layers, and an output layer. The number of input layer nodes is 15, corresponding to the dimension of the feature vector; the number of nodes in the four hidden layers is 64, 128, 256, and 128 respectively; the number of output layer nodes is 6, corresponding to the predicted position coordinates and velocity components of the oil droplet particle within the next 30 frames. The network uses ReLU activation function and Adam optimizer for parameter update, with a learning rate of 0.001.

[0082] During the training process, the predicted position coordinates are used to generate trajectory prediction data, and the actual motion trajectory of the oil droplet particle is collected through a high-speed camera system. The Euclidean distance between the predicted trajectory and the actual trajectory is calculated as the trajectory evaluation value. When the trajectory evaluation value is less than 0.5 microns, it is considered that the prediction accuracy meets the requirements; otherwise, the trajectory evaluation value is used as a feedback signal to update the network parameters through the reward mechanism of deep reinforcement learning.

[0083] The trajectory evaluation value is converted into a reward signal R, and the calculation method is R = 10 - 20 x (trajectory evaluation value), that is, the more accurate the prediction, the higher the reward. The system uses the time difference learning method to calculate the Q value, and updates the network parameters through the gradient descent method. In each training batch, 100 samples are randomly selected, each containing a 30-frame oil droplet motion sequence. The system performs trajectory prediction on each sample, calculates the reward value, and updates the network parameters according to the reward value.

[0084] After 10,000 training batches, the prediction deviation of the network is reduced to 0.3 microns, which is less than the preset threshold of 0.5 microns, at which point the optimized deep reinforcement learning network is determined. After completing network optimization, the optimized deep reinforcement learning network is used to perform continuous trajectory prediction on the oil droplet particle. Each prediction predicts the motion trajectory of the next 30 frames, generating continuous trajectory prediction data. At the same time, the actual motion trajectory data of the oil droplet particle is collected in real time, with a collection interval of 10 frames.

[0085] The actual motion trajectory data and the continuous trajectory prediction data are compared, and the deviation between them is calculated. If the deviation exceeds 0.5 microns, the actual trajectory data is used to correct the predicted trajectory. The correction method uses weighted averaging, that is, the corrected trajectory point coordinates are equal to the weighted average of the predicted coordinates and the actual coordinates, with a weight ratio of 3:7. In this way, the predicted trajectory can be adjusted in real time to ensure the accuracy of the trajectory prediction.

[0086] After multiple corrections, the final oil droplet trajectory is obtained. The deep reinforcement learning method used to track the oil droplet particle trajectory can not only achieve high-precision trajectory prediction, but also improve the robustness of tracking through real-time correction mechanism, effectively solving the problem of easy loss of target and insufficient precision of traditional tracking methods in complex flow field, providing reliable data support for the research of microfluidic field.

[0087] In an optional embodiment, a filter layer space fiber network structure is constructed based on the oil droplet motion trajectory, and the three-dimensional pore distribution parameters of the coalescence filter layer are obtained by pore division based on bubble growth, including:

[0088] The feature parameters of the oil droplet motion trajectory are extracted, the trajectory turning point spatial distribution and the motion direction change frequency are obtained, the trajectory density distribution map is constructed, and the oil droplet motion channel distribution is determined;

[0089] According to the trajectory density distribution map, the fiber distribution probability density field of the filter layer is established, the trajectory turning point spatial distribution is taken as the prior distribution, the motion direction change frequency is taken as the conditional probability, the posterior probability distribution of the fiber cross node is obtained through Bayesian iteration calculation, and the fiber orientation angle is determined according to the posterior probability distribution;

[0090] The fiber distribution probability density field is discretized into grid cells, and fiber position samples are randomly generated in each grid cell according to the posterior probability distribution of the fiber cross node. The fiber position samples are updated by Markov chain iteration, the acceptance probability of each iteration is calculated, and the fiber spatial distribution data is obtained when the acceptance probability converges. The connection relationship of the fiber cross node is optimized based on the fiber spatial distribution data, and the filter layer space fiber network structure is constructed;

[0091] The pore division based on bubble growth is performed on the filter layer space fiber network structure to obtain pore cells, the pore volume distribution and pore throat ratio data are calculated, and the pore connectivity relationship is established;

[0092] The three-dimensional pore distribution parameters of the coalescence filter layer are calculated according to the pore connectivity relationship.

[0093] In one embodiment, a high-speed camera system is used to capture the motion process of oil droplets in the filter layer, and the spatial position coordinate sequence (x_i, y_i, z_i) of each oil droplet is recorded, where i represents the time sequence point. Based on these position coordinates, the characteristic parameters of the oil droplet motion trajectory are extracted. The characteristic parameters include trajectory turning points and motion direction change frequency. The trajectory turning point is defined as the position point where the oil droplet motion direction changes more than 30 degrees. In one specific case, for a certain oil droplet trajectory, 200 time sequence points are recorded, and a total of 15 turning points are identified, with their spatial distribution ranging from X axis 1.2-5.8mm, Y axis 0.8-4.2mm, and Z axis 0.5-3.5mm. The motion direction change frequency is calculated as the number of direction changes per unit distance, and for this oil droplet trajectory, the average change frequency is 3.2 times / mm.

[0094] By analyzing multiple oil droplet trajectories, the three-dimensional space is divided into 0.5mm x 0.5mm x 0.5mm voxel units, and the frequency of oil droplets passing through each voxel is counted to generate a trajectory density distribution map. The density value is represented by a color gradient, for example, the density value at voxel (2.5, 3.0, 1.5) is 52 times / cubic millimeter, indicating that this area is the main motion channel of the oil droplets.

[0095] According to the trajectory density distribution map, a fiber distribution probability density field of the filter layer is established. The density field is represented as P(f|x, y, z), representing the probability of the presence of fibers at the spatial position (x, y, z), where f represents the event of the presence of fibers, and x, y, z are spatial coordinates. The spatial distribution of trajectory turning points is taken as the prior distribution P(f), representing the probability of the presence of fibers at each space without other information; in this embodiment, this distribution shows that the probability value is higher in the dense area of turning points, for example, in the region (2.0-3.0, 2.5-3.5, 1.0-2.0) mm, the average value of the prior probability is 0.75. Combined with the motion direction change frequency as the conditional probability P(d|f), where d represents the event of the change of the direction of the oil droplets, representing the probability of the change of the direction of the oil droplets under the condition of the presence of fibers. In this example, the conditional probability is set to 0.82 in the region where the direction change frequency is greater than 3 times / mm.

[0096] The posterior probability distribution P(f|d) of the fiber intersection nodes is obtained through Bayesian iterative calculation, representing the probability of the presence of fibers at the location under the condition of observing the change in the direction of the oil droplets. During the calculation process, the posterior probability is updated every iteration until the probability changes by less than 0.01 in two consecutive iterations. After 8 iterations, a stable posterior probability distribution is obtained. The fiber orientation angle is determined according to the posterior probability distribution, that is, at each high-probability point (probability value greater than 0.7), the main orientation angle of the fiber at that location is determined according to the statistical characteristics of the change in the direction of the oil droplets. In the embodiment, a total of 247 high-probability fiber nodes are determined, and the orientation angle of each node is represented as a three-dimensional vector (θx, θy, θz), indicating the angle of the fiber relative to the x, y, and z coordinate axes in space, such as the orientation angle of (45°, 30°, 60°) at the node (2.5, 3.2, 1.8).

[0097] The fiber distribution probability density field is discretized into 0.1 mm x 0.1 mm x 0.1 mm grid cells, a total of 10 x 10 x 10 = 1000 cells. In each grid cell, a fiber position sample is randomly generated according to the posterior probability distribution of the fiber intersection nodes. The initial sample size is set to 5000 points, each containing spatial coordinates and orientation information.

[0098] The fiber position sample is iteratively updated using Markov chain. In each iteration, a sample point is randomly selected, a new candidate position is generated within its neighborhood (radius 0.2 mm), and the acceptance probability a is calculated. The acceptance probability a is determined according to the posterior probability ratio of the new and old positions, a = min(1, P(f_new|d) / P(f_old|d)), where P(f_new|d) represents the posterior probability of the new position and P(f_old|d) represents the posterior probability of the old position. When a is greater than a randomly generated threshold (between 0 and 1), the new position is accepted; otherwise, the old position is retained. After 1000 iterations, the average value of the acceptance probability decreases from the initial 0.78 to 0.32 and stabilizes, indicating that the iteration converges, and the fiber spatial distribution data is obtained at this time.

[0099] Based on the fiber spatial distribution data, the connection relationship of the fiber intersection nodes is optimized. For any two fiber nodes with a distance of less than 0.5 mm, if their orientation angle difference is less than 15 degrees, they are considered to belong to the same fiber and a connection is established. For node pairs with an orientation angle difference between 15 and 45 degrees, a connection is established with a probability of 0.5, indicating the bending or intersection of the fiber. In this way, a filter layer spatial fiber network structure containing 372 fibers and 689 intersection points is constructed.

[0100] The constructed filter layer spatial fiber network structure is subjected to pore division based on bubble growth. 583 initial bubble cores are uniformly distributed in the network space, each bubble expands from the core position to the surrounding, and stops when encountering a fiber or other bubble boundary. The expansion process is controlled through discrete time steps, with an expansion of 0.02 mm per step, and a total of 50 iteration steps. After the expansion is completed, a set of pore cells is obtained, and the volume of each pore cell and the connected area (pore throat) with adjacent pores are calculated.

[0101] Statistical analysis shows that the pore volume distribution ranges from 0.01 to 0.25 cubic millimeters, with an average of 0.08 cubic millimeters and a standard deviation of 0.04 cubic millimeters. The pore throat ratio data (the ratio of pore throat area to pore volume) ranges from 2.1 to 8.7, with an average of 4.3. A pore connectivity relationship diagram is established, with an average of 4.2 adjacent pores connected to each pore. According to the pore connectivity relationship, the three-dimensional pore distribution parameters of the coalesced filter layer are calculated, including porosity (0.72), specific surface area (12.5 square millimeters / cubic millimeter), average pore size (0.32 millimeters), and pore size distribution curve (peak at 0.28-0.35 millimeters). These parameters can be used to evaluate the performance of the filter layer and optimize the design of the filter layer.

[0102] In an alternative embodiment, the filter layer spatial fiber network structure is subjected to pore division based on bubble growth to obtain pore cells, calculate pore volume distribution and pore throat ratio data, and establish pore connectivity relationship, including:

[0103] The distance values of each grid point in the filter layer spatial fiber network structure to the surrounding fibers are calculated to obtain distance field data, and the local maximum position is found in the distance field data, which is determined as the bubble core point, and the distance value is determined as the initial radius of the corresponding bubble;

[0104] Bubble inflation is performed with the bubble core point as the center, the inflation rate is inversely proportional to the local density of fibers in the filter layer spatial fiber network structure, and the contact position points of the bubble and the fibers are recorded, and the inflation stops when adjacent bubbles contact;

[0105] A common boundary is constructed at the contact position of adjacent bubbles to obtain initial pore cells, the curvature of the common boundary is optimized according to the spatial distribution characteristics of the fibers, and the initial pore cells with a volume less than a preset volume threshold are merged with adjacent pore cells to obtain final pore cells;

[0106] The volume and surface area of the final pore cells are calculated, the minimum cross-section position between adjacent pore cells is identified as the pore throat, the size of the pore throat and the volume of the corresponding pore cell are calculated to obtain the pore throat ratio data;

[0107] According to the space connection relationship between the final pore units and the pore throat ratio data, a pore connectivity relationship is established, and the connectivity number and connectivity strength of each pore unit are determined.

[0108] In a specific embodiment, for an established filter layer spatial fiber network structure, the distance value of each grid point in the network to the surrounding fibers is calculated. Using the Euclidean distance calculation method, the distance from each point in the regular grid division in space to the nearest fiber surface is calculated, forming a three-dimensional distance field data. For example, in a 100x100x100 micron filter layer space, with a grid spacing of 1 micron, 1 million grid point distance field data is obtained. By scanning the distance field data, identify the local maximum position, that is, the position whose distance value is greater than that of all its adjacent grid points, and determine these local maximum positions as bubble core points, and the corresponding distance value as the initial radius of the bubble. In practical applications, for example, for a fiber network structure with a porosity of 75%, about 1000 bubble core points may be identified, with initial radii distributed between 0.5-10 microns.

[0109] Bubble inflation process is carried out with bubble core points as the center. The inflation rate of the bubble is inversely proportional to the local fiber density, that is, in areas with lower fiber density, the bubble inflation speed is faster; in areas with higher fiber density, the bubble inflation speed is slower. The bubble surface is discretized into a uniformly distributed point set, and each point moves in the radial direction. The moving step is dynamically adjusted according to the fiber density near the point, which can be determined by calculating the volume proportion of fibers in a unit volume. For example, when the local fiber density is 15%, the inflation rate of the bubble in this area is 0.8 times the standard rate; when the local fiber density is 5%, the bubble inflation rate is 1.2 times the standard rate. During the bubble inflation process, the contact position points between the bubble and the fiber are recorded, which will affect the shape of the final pore unit. When adjacent bubbles come into contact, the contact position is recorded and the further inflation of the two bubbles in the contact area is stopped.

[0110] A common boundary is constructed at the contact position of adjacent bubbles to form an initial pore unit. At the contact area of two adjacent bubbles, a plane or curved surface is constructed as their common boundary. The curvature of the common boundary is optimized according to the spatial distribution characteristics of the fibers, that is, the shape of the boundary is adjusted according to the density and distribution of the fibers near the boundary. In the curvature optimization process, a local weighted average method can be used to make the boundary smoother and more consistent with the actual fiber distribution. For example, when the fiber distribution near the boundary is uneven, the boundary will protrude to the side with high fiber density. For initial pore units with a volume less than a preset volume threshold, they are merged with adjacent pore units. The preset volume threshold can be determined according to the overall pore volume distribution, for example, it can be set to 10% of the median of the volume of all pore units. Through this merging operation, small pore units are eliminated, making the distribution of the final pore units more reasonable and avoiding excessive subdivision.

[0111] The volume and surface area of the final pore unit are calculated. The volume calculation uses a spatial integration method to accumulate the volume of the grid points inside the pore unit; the surface area calculation is realized by summing the areas of the discrete triangular patches on the surface of the pore unit. In actual cases, a typical filter layer may contain about 800 effective pore units, with a pore volume distribution ranging from 100 to 5000 cubic microns, and an average volume of about 1200 cubic microns. The minimum cross-sectional position between adjacent pore units is identified as the pore throat, which usually occurs on the common boundary of the two pore units. The calculation method is to find the position with the smallest cross-sectional area on the common boundary, which is the pore throat. The size of the pore throat is calculated by dividing the volume of the corresponding pore unit, and the pore throat ratio data is obtained. For example, a pore unit with a volume of 2000 cubic microns has a connected pore throat cross-sectional area of 50 square microns, so the pore throat ratio is 0.025.

[0112] According to the spatial connection relationship between the final pore units and the pore throat ratio data, the pore connectivity relationship is established. For each pore unit, record the number of all pore units directly connected to it to determine the connectivity number; according to the size of the pore throat at the connection, the connectivity strength is evaluated, the larger the pore throat size, the higher the connectivity strength. In practical applications, the average connectivity number of a pore unit is about 4-6, and the connectivity strength is normalized according to the pore throat ratio data, ranging from 0 to 1. For example, for a pore unit with a connectivity number of 5, its connectivity strength is 0.82, 0.65, 0.53, 0.47, and 0.31, respectively. These data can be used for subsequent analysis of fluid permeation behavior and filtration efficiency in the filter layer.

[0113] Through the above method, the pore division of the spatial fiber network structure of the filter layer is completed, the detailed pore volume distribution and pore throat ratio data are obtained, and the complete pore connectivity relationship is established, providing an important basis for performance evaluation and optimization design of filter materials.

[0114] In an alternative embodiment, three-dimensional pore distribution parameters are input into a graph neural network to construct an oil droplet coalescence network topology, analyze oil droplet migration strength to determine coalescence channels, calculate the spatial coordinates and coalescence rate of oil droplet coalescence points, and determine the oil droplet coalescence position and coalescence expansion trend of the coalescence filter layer, comprising:

[0115] Three-dimensional pore distribution parameters are input into a graph neural network to construct an oil droplet coalescence network topology based on pore connectivity relationships, with pore units set as network nodes, the volume and surface area of the pore units as node features, and pore throat ratio data and connectivity strength as edge features.

[0116] The oil droplet coalescence network topology is calculated using a graph neural network, with node feature information transmitted based on a directional convolution operator and coalescence feature representation of the nodes obtained through feature fusion based on pore throat ratio data.

[0117] The oil droplet migration strength between adjacent pore units is calculated based on the coalescence feature representation of the nodes, and pore units with oil droplet migration strength greater than a preset strength threshold are combined into a pore unit sequence, with the pore unit sequence determined as an oil droplet coalescence channel.

[0118] The collision frequency of oil droplets is calculated in the oil droplet coalescence channel based on the connectivity strength, with the position with the highest collision frequency determined as the spatial coordinates of the oil droplet coalescence point, and the oil droplet coalescence rate determined based on the collision frequency.

[0119] The decay value of the coalescence rate is calculated along the direction of the pore connectivity relationship from the oil droplet coalescence point as the starting position, the influence range of oil droplet coalescence is determined based on the decay value, and the oil droplet coalescence position distribution and coalescence expansion trend of the coalescence filter layer are obtained.

[0120] In a specific embodiment, three-dimensional pore distribution parameters are input into a graph neural network for processing. These parameters include microstructure features such as the volume, surface area, spatial position of the pore units, and the connectivity relationship between pores. Based on the pore connectivity relationship, an oil droplet coalescence network topology is constructed, with each pore unit set as a network node and edge connections established between interconnected pore units. Node features include the volume and surface area of the pore units, such as the volume of 0.082 cubic millimeters and the surface area of 0.46 square millimeters of pore unit P143. Edge features include pore throat ratio data and connectivity strength. Pore throat ratio data is defined as the ratio of the minimum cross-sectional area at the connection between two adjacent pore units to the volume of the smaller pore, and connectivity strength is defined as the fluid passing capacity between two pore units. For example, the pore throat ratio between pore units P143 and P144 is 0.32, and the connectivity strength is 0.78.

[0121] The constructed oil droplet coalescence network topology is calculated by a graph neural network. A directional convolution operator is used to transfer node feature information. This operator can effectively capture the directionality of oil droplet migration between pores based on the flow characteristics of fluid in porous media. The directional convolution operator considers the spatial relationship between pore cells and the direction of fluid flow, making the feature transfer more consistent with the actual oil droplet migration law. During feature transfer, feature fusion is performed based on the pore-throat ratio data. The larger the pore-throat ratio, the higher the efficiency of feature information transfer. For node i, the feature update formula considers the feature information from adjacent node j, and the fusion weight is proportional to the pore-throat ratio. The graph convolution network is set to a 3-layer structure, with a hidden feature dimension of 64 at each layer. ReLU activation function is used in the middle to ensure the non-linear ability of feature expression. Through multi-layer graph convolution operation, each node finally obtains a coalescence feature representation containing the surrounding pore environment information, with a dimension of 32.

[0122] Based on the node coalescence feature representation calculated by the graph neural network, the oil droplet migration intensity between adjacent pore cells is calculated. The oil droplet migration intensity reflects the possibility of oil droplet migration from one pore cell to an adjacent pore cell, and is related to the similarity of the coalescence feature representation of the two pore cells. The cosine similarity of the coalescence feature vectors of two adjacent pore cells is calculated, and adjusted combined with the pore-throat ratio to obtain the final oil droplet migration intensity. In practical applications, the feature cosine similarity between pore cells P143 and P144 is 0.83, and after adjusting combined with the pore-throat ratio 0.32, the final migration intensity is 0.72. Set the migration intensity threshold to 0.65, and the adjacent pore cells with migration intensity greater than the threshold form a pore cell sequence, which is the oil droplet coalescence channel. In a certain test sample, a total of 57 oil droplet coalescence channels are identified, with an average of 8.3 pore cells per channel.

[0123] In the determined oil droplet coalescence channel, the collision frequency of oil droplets is calculated according to the connectivity intensity. The oil droplet collision frequency is related to the connectivity intensity of each node in the channel, the fluid flow rate and the oil droplet concentration. For each pore cell node in the channel, the possible collision times of all oil droplets passing through the node are calculated, and the collision frequency is proportional to the connectivity intensity. Take a coalescence channel as an example, which contains pore cell sequence P143-P144-P157-P162-P165-P170. By analyzing the connectivity intensity distribution, it is calculated that the collision frequency at P157 position is the highest, reaching 78 times per minute. The position with the highest collision frequency is determined as the spatial coordinates of the oil droplet coalescence point, which is the area where oil droplets are most likely to coalesce to form larger oil droplets. At the same time, the collision frequency at this position is used to determine the oil droplet coalescence rate, which represents the number of larger oil droplets formed per unit time at this position. For the above coalescence point, the coalescence rate is 42 per minute, and the spatial coordinates are X=12.5 mm, Y=8.3 mm, and Z=4.2 mm.

[0124] The attenuation value of the coalescence rate is calculated along the direction of the pore connectivity relationship starting from the oil droplet coalescence point. The attenuation value reflects the degree of weakening of the coalescence effect as the distance from the coalescence point increases. The attenuation calculation takes into account the distance factor and the medium characteristics, and generally shows an exponential decay trend. For a position 1 millimeter away from the coalescence point, the coalescence rate is attenuated to 85% of the initial value; for a distance of 3 millimeters, the attenuation is 62% of the initial value; for a distance of 5 millimeters, the attenuation is 35% of the initial value; for a distance of 7.5 millimeters, the attenuation is 10% of the initial value. The influence range of oil droplet coalescence is determined according to the calculated attenuation value, and the influence range is defined as the area where the coalescence rate is attenuated to 10% of the initial value. By this method, the oil droplet coalescence position distribution and coalescence expansion trend of the coalescence filter layer are obtained. In the test sample, the influence range of the main coalescence point is a spherical area with a radius of about 7.5 millimeters, and within this influence range, the average coalescence rate is 18 per minute.

[0125] Through the analysis of multiple coalescence points, a thermal map of oil droplet coalescence for the entire coalescence filter layer can be drawn. The thermal map shows that the coalescence activity is mainly concentrated in the central region of the filter layer and exhibits an uneven distribution characteristic. These coalescence regions play a key role in the actual wastewater oil-water separation process and are the main areas where oil droplets coalesce into larger droplets and accelerate the separation process. For a typical coalescence filter layer, test data show that about 65% of the oil droplet coalescence effect is generated within the first 30% of the flow distance, which is an important guiding significance for optimizing the structure of the filter layer.

[0126] Experimental verification shows that the oil droplet coalescence position distribution predicted by this method is highly consistent with the results observed in the actual wastewater oil-water separation process, providing an effective theoretical basis for optimizing the design of the coalescence filter layer and guiding the optimization of the structure of the coalescence material and the adjustment of the parameters of the filtration system, thereby improving the efficiency of wastewater oil-water separation. By adjusting the pore structure of the coalescence material, the high-efficiency coalescence area can be targeted to be strengthened, further improving the oil-water separation performance.

[0127] In an alternative embodiment, a graph neural network is used to calculate the coalescence network topology of the oil droplets, and a directional convolution operator is used to transfer node feature information based on the pore-throat ratio data to obtain a coalescence feature representation of the node, including:

[0128] A directional convolution operator is constructed according to the spatial distribution characteristics of the pore units, the volume and surface area of the pore units are projected onto the principal axis direction, and the anisotropy degree is calculated as the shape parameter of the convolution kernel;

[0129] The node features are directionally convoluted by the directional convolution operator, local structure features are extracted along the principal axis direction, and anisotropic weights are set for feature information in different directions, and the anisotropic weights are positively correlated with the anisotropy degree;

[0130] The convolution results in different directions are adaptively fused, the direction importance score is established based on the pore-throat ratio data, the fusion weight is dynamically adjusted according to the direction importance score, and the coalescence feature representation of the node is obtained.

[0131] In a specific embodiment, the pore structure data of the porous medium is obtained, including the spatial position, volume, surface area, connectivity relationship and pore-throat ratio data of the pore unit. Based on these data, an oil droplet coalescence network is constructed, in which the pore unit is taken as the network node and the connectivity relationship between the pores is taken as the network edge. For example, for a porous medium sample containing 200 pore units, a network topology structure with 200 nodes can be established, each node has an initial feature vector containing the volume (such as 0.005mm 3 ), surface area (such as 0.02mm 2 ), connectivity (such as connected to 6 other pores) and other information of the pore unit.

[0132] When constructing the directional convolution operator according to the spatial distribution characteristics of the pore unit, the principal axis direction of the pore unit needs to be determined. Specifically, the pore unit is regarded as a three-dimensional geometric body, and its inertia matrix is calculated. Through eigenvalue decomposition of the inertia matrix, three principal axis directions are obtained, which are denoted as x, y and z directions respectively. The volume and surface area of the pore unit are projected onto the three principal axis directions to obtain the projection values in each direction. For example, the volume projection values of a pore unit in x, y and z directions are 0.003mm 3 , 0.004mm 3 and 0.002mm 3 respectively, and the surface area projection values are 0.012mm 2 , 0.015mm 2 and 0.008mm 2 .

[0133] The anisotropy degree is calculated based on the projection values, specifically the difference degree between the projection values in different directions. The anisotropy degree can be represented by the ratio of the maximum projection value to the minimum projection value. In the above example, the anisotropy degree of volume projection is 0.004 / 0.002=2.0, and the anisotropy degree of surface area projection is 0.015 / 0.008=1.875. These anisotropy degrees are taken as the shape parameters of the convolution kernel to construct the directional convolution operator. The higher the anisotropy degree, the flatter the convolution kernel shape in that direction, and the more prominent the convolution effect.

[0134] When performing directional convolution calculation on node features by directional convolution operator, local structure features are extracted along the principal axis direction. For each node, directional convolution operators are applied in x, y, z directions respectively to extract feature information in this direction. During the convolution process, the neighbor nodes of the node are considered, and different weights are given according to the relative position of the neighbor nodes in the corresponding direction. For example, when performing convolution in the x direction, the nodes adjacent to the current node in the x direction will obtain higher weights (such as 0.8), while the nodes adjacent in the y and z directions will obtain lower weights (such as 0.3 and 0.2).

[0135] Anisotropic weights are set for feature information in different directions, and the anisotropic weights are positively correlated with the anisotropy. Specifically, if the anisotropy of the x direction is 2.0, the anisotropy of the y direction is 1.875, and the anisotropy of the z direction is 1.5, the corresponding anisotropic weights can be set to 0.5, 0.3, and 0.2, ensuring that the sum of the weights is 1. In this way, in the direction with higher anisotropy, the convolution result will obtain higher weight, thereby playing a more important role in feature fusion.

[0136] When adaptively fusing convolution results in different directions, direction importance scores are established based on pore throat ratio data. Pore throat ratio represents the ratio of throat diameter connecting two pores to pore diameter, which is an important factor affecting oil droplet coalescence. For each direction, the average value of the pore throat ratio in the direction is calculated and used as the importance index of the direction. For example, if the average pore throat ratios in x, y, z directions are 0.6, 0.4, and 0.3 respectively, the x direction has the highest importance score.

[0137] When dynamically adjusting the fusion weight according to the direction importance score, a soft attention mechanism can be used. First, the direction importance score is converted into an initial weight through normalization processing, and then combined with the anisotropic weight to calculate the final fusion weight. For example, the weight obtained by normalizing the direction importance score is 0.46, 0.31, and 0.23, and combined with the anisotropic weight 0.5, 0.3, and 0.2, the final fusion weight 0.48, 0.305, and 0.215 is obtained by weighted average.

[0138] The convolution results in different directions are weighted and summed using final fusion weights to obtain the aggregation feature representation of the node. For example, if the convolution results in x, y, and z directions are feature vectors [0.8, 0.6, 0.4], [0.7, 0.5, 0.3], and [0.6, 0.4, 0.2] respectively, the weighted and fused feature representation is [0.48*0.8+0.305*0.7+0.215*0.6, 0.48*0.6+0.305*0.5+0.215*0.4, 0.48*0.4+0.305*0.3+0.215*0.2] = [0.74, 0.53, 0.33].

[0139] Through the above process, each node obtains an aggregation feature representation that fuses multi-directional information, which fully considers the spatial distribution characteristics of the pore structure and the pore throat ratio data, and can more accurately reflect the coalescence behavior of oil droplets in the porous medium.

[0140] As shown in FIG. 1, a directional convolution and adaptive fusion logic flowchart is shown. Figure 2

[0141] In an optional implementation, an evolution potential field is constructed based on the filtration state evaluation data, and the fiber arrangement angle and the fiber spacing are calculated to obtain improved filter layer parameters, including:

[0142] According to the filtration state evaluation data, an evolution potential field is constructed, a region with an oil-water separation efficiency greater than a preset efficiency threshold is set as a positive potential energy region, a region with an oil-water separation efficiency less than the preset efficiency threshold is set as a negative potential energy region, and an evolution path direction is determined according to a remaining use time of the coalescence filter layer;

[0143] The potential energy values of the positive potential energy region and the negative potential energy region are determined, a search space is determined according to the potential energy distribution of the evolution potential field, the fiber arrangement angle and the fiber spacing are used as optimization variables, and the search step of the optimization variables is set based on the potential energy values, wherein the corresponding search step of the positive potential energy region is greater than the corresponding search step of the negative potential energy region;

[0144] The potential energy gradient in the evolution potential field is calculated, the potential energy gradient is used as the adjustment direction of the fiber structure, the fiber spacing of the positive potential energy region is adjusted to a first preset spacing, and the fiber spacing of the negative potential energy region is adjusted to a second preset spacing, wherein the first preset spacing value is greater than the second preset spacing value;

[0145] The fiber arrangement angle is calculated based on the evolution path direction, so that the included angle between the fiber arrangement angle and the evolution path direction is kept within a preset angle range;

[0146] ​According to the optimization results of the fiber arrangement angle and the fiber spacing, the pore distribution structure of the coalescence filter layer is adjusted to obtain improved filter layer structure parameters.

[0147] In practical applications, the present application realizes the optimization of the fiber arrangement angle and the fiber spacing by constructing an evolutionary potential field and performing optimization calculation, so as to improve the oil-water separation efficiency of the coalescence filter layer. The specific implementation process is as follows:

[0148] In one specific embodiment, the evolutionary potential field is constructed based on the filter state evaluation data, and the oil-water separation efficiency is compared with a preset efficiency threshold of 95%. When the oil-water separation efficiency of a certain region is higher than 95%, it is marked as a positive potential region; when it is lower than 95%, it is marked as a negative potential region. For example, when the separation efficiency of region A is detected to be 97%, it is set as a positive potential region; when the separation efficiency of region B is 92%, it is set as a negative potential region. At the same time, the evolutionary path direction is determined by using the remaining service time data of the coalescence filter layer, such as when the remaining service time is 500 hours, the evolutionary path direction is determined to be the direction of extending the service time.

[0149] In the potential value determination stage, the potential value is calculated according to the difference between the oil-water separation efficiency and the preset threshold. Specifically, when the separation efficiency is 97%, the positive potential value is (97%-95%) / 95%=2.1%; when the separation efficiency is 92%, the negative potential value is (92%-95%) / 95%=-3.2%. Based on these potential values, the search space boundary is determined, the search range of the fiber arrangement angle is set to 0° to 90°, and the search range of the fiber spacing is set to 5μm to 50μm. The positive potential region uses a larger search step, such as an angle step of 5° and a spacing step of 5μm; the negative potential region uses a smaller search step, such as an angle step of 2° and a spacing step of 2μm, so as to more finely explore the solution space.

[0150] Potential gradient calculation is a key step in the optimization process, which analyzes the potential change rate between adjacent regions and determines the direction with the fastest potential change as the adjustment direction. In practical applications, the potential gradient calculated from the transition zone from the negative potential region (-3.2%) to the positive potential region (2.1%) is 0.53% / μm. According to this gradient value, the fiber structure is automatically adjusted. For the positive potential region, the fiber spacing is adjusted to a first preset spacing of 30μm; for the negative potential region, the fiber spacing is adjusted to a second preset spacing of 15μm. This differential setting can optimize the fluid passing performance while ensuring the filtering efficiency.

[0151] The optimization of the fiber arrangement angle is closely related to the evolution path direction, and the optimal fiber arrangement angle is calculated according to the evolution path direction (with the horizontal direction as the reference, and the direction angle is 60°). In order to ensure that the angle between the fiber and the evolution path is within the preset range (such as 30° to 60°), the fiber arrangement angle in the positive potential energy region is adjusted to 25°, and the fiber arrangement angle in the negative potential energy region is adjusted to 85°. Experimental verification shows that when the angle between the fiber and the evolution path is 45°, the oil-water separation efficiency can be increased to 98.5%, and the remaining service life is extended to 750 hours.

[0152] According to the optimized fiber arrangement angle and fiber spacing parameters, the pore distribution structure of the coalescence filter layer is adjusted. In engineering implementation, by precisely controlling the fiber weaving process, the positive potential energy region adopts a fiber spacing of 30 μm and an arrangement angle of 25°, and the negative potential energy region adopts a fiber spacing of 15 μm and an arrangement angle of 85°. This structure adjustment makes the overall filter layer form a gradient-changing microstructure at the oil-water interface, and the oil droplet coalescence efficiency is increased to 98%, the pressure loss is reduced by 25%, and the service life of the filter layer is extended by 40%.

[0153] In order to verify the optimization effect, the test is carried out under the condition that the oil content in the oil-water mixture is 5%. The oil-water separation efficiency of the filter layer before optimization decreases from 96% to 91% after working for 100 hours; and the oil-water separation efficiency of the filter layer after optimization remains above 95% after working for 150 hours. The pressure difference test shows that the pressure loss of the filter layer after optimization is reduced from 0.5 MPa to 0.38 MPa under the same flow condition, and the fluid passing performance is significantly improved.

[0154] The optimization method is not only suitable for conventional oil-water separation scenes, but also can be extended to be applied to emulsion treatment, high-viscosity oil-water mixture separation and other complex working conditions. By adjusting the preset efficiency threshold, potential energy calculation parameters and search step length and other parameters, the best filter layer structure can be customized for different application scenes, and the overall performance of the filter system is improved.

[0155] The wastewater oil liquid separation multi-stage filter processing system based on the coalescence material embodiment of the application comprises:

[0156] The first unit is used for acquiring structure information and operation parameters of the multi-stage filter device;

[0157] The second unit is used for tracking and analyzing oil droplet particles by using a deep reinforcement learning method, acquiring an oil droplet motion trajectory, constructing a filter layer space fiber network structure based on the oil droplet motion trajectory, and obtaining three-dimensional pore distribution parameters of the coalescence filter layer through pore division based on bubble growth;

[0158] The third unit is configured to input the three-dimensional pore distribution parameters into the graph neural network, construct an oil droplet coalescence network topology, analyze oil droplet migration strength to determine a coalescence channel, calculate spatial coordinates and coalescence rate of an oil droplet coalescence point, and determine an oil droplet coalescence position and coalescence expansion trend of the coalescence filter layer.

[0159] The fourth unit is configured to calculate a remaining use time and an oil-water separation efficiency of each coalescence filter layer according to the oil droplet coalescence position and coalescence expansion trend, in combination with structure information and operation parameters, and generate filter state evaluation data.

[0160] The fifth unit is configured to construct an evolution potential field based on the filter state evaluation data, perform optimization calculation on fiber arrangement angles and fiber spacings, and obtain improved filter layer parameters.

[0161] The sixth unit is configured to apply the improved filter layer parameters to the multi-stage coalescence filter device, collect real-time oil-water separation data as new operation parameters, and perform optimization in a cycle iteration manner.

[0162] In a third aspect, an electronic device is provided, including:

[0163] a processor;

[0164] a memory for storing processor-executable instructions;

[0165] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0166] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0167] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0168] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-stage filtration method for separating oily wastewater based on coalescing materials, characterized in that, include: Obtain structural information and operating parameters of the multi-stage filtration device; Deep reinforcement learning is used to track and analyze oil droplet particles to obtain their trajectories. Based on these trajectories, a spatial fiber network structure for the filter layer is constructed. By dividing the pores based on bubble growth, the three-dimensional pore distribution parameters of the coalescing filter layer are obtained. The three-dimensional pore distribution parameters are input into a graph neural network to construct the oil droplet coalescence network topology. The oil droplet migration intensity is analyzed to determine the coalescence channel. The spatial coordinates and coalescence rate of the oil droplet coalescence point are calculated to determine the oil droplet coalescence location and coalescence expansion trend of the coalescence filter layer. Based on the location and expansion trend of oil droplet coalescence, combined with structural information and operating parameters, the remaining service time and oil-water separation efficiency of each coalescing filter layer are calculated to generate filtration status assessment data. An evolution potential field is constructed based on the filtration status assessment data, and the fiber arrangement angle and fiber spacing are optimized and calculated to obtain improved filter layer parameters. The improved filter layer parameters were applied to the multi-stage coalescing filter device, and real-time oil-water separation data were collected as new operating parameters for iterative optimization.

2. The method according to claim 1, characterized in that, Deep reinforcement learning was used to track and analyze oil droplet particles, and the trajectory of the oil droplets was obtained, including: Collect the motion state parameters and interface interaction parameters of oil droplets in the flow field to generate initial feature data; calculate the force state of oil droplets in the flow field based on the initial feature data to generate force state data; The initial feature data and the force state data are input into a deep reinforcement learning network to predict the motion trajectory and generate trajectory prediction data. The actual motion trajectory is collected and compared with the trajectory prediction data to calculate the trajectory evaluation value. The trajectory evaluation value is used as a feedback signal for the deep reinforcement learning network. The deep reinforcement learning network is optimized based on the feedback signal, and the training parameters are adjusted until the prediction deviation of the trajectory prediction data is less than a preset deviation threshold. The optimized deep reinforcement learning network is then determined. An optimized deep reinforcement learning network is used to continuously predict motion trajectories and generate continuous trajectory prediction data. Actual motion trajectory data is collected in real time, and the actual motion trajectory data is compared with the continuous trajectory prediction data to correct the continuous trajectory prediction data, thus obtaining the final oil droplet motion trajectory.

3. The method according to claim 1, characterized in that, A spatial fiber network structure for the filter layer was constructed based on the oil droplet motion trajectory. The three-dimensional pore distribution parameters of the coalescing filter layer were obtained through pore partitioning based on bubble growth, including: Extract the feature parameters of the oil droplet motion trajectory, obtain the spatial distribution of trajectory turning points and the frequency of motion direction changes, construct a trajectory density distribution map, and determine the distribution of oil droplet motion channels; The fiber distribution probability density field of the filter layer is established based on the trajectory density distribution map. The spatial distribution of trajectory turning points is taken as the prior distribution, and the frequency of motion direction change is taken as the conditional probability. The posterior probability distribution of fiber intersection nodes is obtained through Bayesian iterative calculation. The fiber orientation angle is determined based on the posterior probability distribution. The fiber distribution probability density field is discretized into grid cells. In each grid cell, fiber position samples are randomly generated according to the posterior probability distribution of fiber cross nodes. Markov chain iteration is performed on the fiber position samples to update them. The acceptance probability of each iteration is calculated. When the acceptance probability converges, the fiber spatial distribution data is obtained. The connection relationship of fiber cross nodes is optimized based on the fiber spatial distribution data to construct the filter layer spatial fiber network structure. The pore structure of the filter layer is divided into pore units based on bubble growth. The pore volume distribution and pore throat ratio data are calculated, and the pore connectivity relationship is established. The three-dimensional pore distribution parameters of the coalescing filter layer are calculated based on the pore connectivity relationship.

4. The method according to claim 3, characterized in that, The filter layer spatial fiber network structure is divided into pore units based on bubble growth. Pore volume distribution and pore-throat ratio data are calculated, and pore connectivity relationships are established, including: Calculate the distance from each grid point in the filter layer's spatial fiber network structure to the surrounding fibers to obtain distance field data. Find the location of the local maximum value in the distance field data, determine the location of the local maximum value as the bubble core point, and determine the distance value as the initial radius of the corresponding bubble. Bubble expansion is performed with the bubble core as the center. The rate of bubble expansion is inversely proportional to the local density of fibers in the fiber network structure of the filter layer. The contact points between the bubble and the fibers are recorded. Expansion stops when adjacent bubbles come into contact. A common boundary is constructed at the contact position of adjacent bubbles to obtain an initial pore unit. The curvature of the common boundary is optimized according to the spatial distribution characteristics of the fibers. The initial pore units with a volume smaller than a preset volume threshold are merged with the adjacent pore units to obtain the final pore unit. Calculate the volume and surface area of ​​the final pore unit, identify the minimum cross-sectional position between adjacent pore units as the pore throat, calculate the ratio of the size of the pore throat to the volume of the corresponding pore unit, and obtain the pore throat ratio data. Based on the spatial connection relationship between the final pore units and the pore-throat ratio data, a pore connectivity relationship is established to determine the number of connections and the connectivity strength of each pore unit.

5. The method according to claim 1, characterized in that, The three-dimensional pore distribution parameters are input into a graph neural network to construct the oil droplet coalescence network topology. The oil droplet migration intensity is analyzed to determine the coalescence channels. The spatial coordinates and coalescence rate of the oil droplet coalescence points are calculated. The determination of the oil droplet coalescence location and coalescence expansion trend in the coalescence filter layer includes: The three-dimensional pore distribution parameters are input into a graph neural network. Based on the pore connectivity relationship, an oil droplet coalescence network topology is constructed. Pore units are set as network nodes. The volume and surface area of ​​the pore units are used as node features, and the pore-throat ratio data and connectivity strength are used as edge features. The graph neural network topology of the oil droplet coalescence network is calculated, and node feature information is transmitted based on the directional convolution operator. Feature fusion is performed based on the pore-throat ratio data to obtain the coalescence feature representation of the nodes. The oil droplet migration intensity between adjacent pore units is calculated based on the coalescence characteristics of the nodes. Pore units with oil droplet migration intensity greater than a preset intensity threshold are grouped into a pore unit sequence, and the pore unit sequence is determined as an oil droplet coalescence channel. In the oil droplet coalescence channel, the collision frequency of oil droplets is calculated based on the connectivity strength, and the position with the highest collision frequency is determined as the spatial coordinate of the oil droplet coalescence point. The oil droplet coalescence rate is determined based on the collision frequency. Starting from the point where the oil droplets coalesce, the attenuation value of the coalescence rate is calculated along the direction of the pore connectivity relationship. Based on the attenuation value, the influence range of oil droplet coalescence is determined, and the distribution of oil droplet coalescence locations and the coalescence expansion trend of the coalescing filter layer are obtained.

6. The method according to claim 5, characterized in that, The graph neural network topology of the oil droplet coalescing network is calculated, and node feature information is transmitted based on the directional convolution operator. Feature fusion is performed based on the pore-throat ratio data to obtain the coalescing feature representation of the nodes, including: An oriented convolution operator is constructed based on the spatial distribution characteristics of the pore units. The volume and surface area of ​​the pore units are projected onto the principal axis direction, the anisotropy is calculated, and the anisotropy is used as the shape parameter of the convolution kernel. The node features are subjected to directional convolution calculations by the directional convolution operator, and local structural features are extracted along the main axis. Anisotropic weights are set for feature information in different directions, and the anisotropic weights are positively correlated with the anisotropic degree. The convolution results in different directions are adaptively fused, and a directional importance score is established based on the aperture-throat ratio data. The fusion weights are dynamically adjusted according to the directional importance score to obtain the clustering feature representation of the nodes.

7. The method according to claim 1, characterized in that, An evolution potential field was constructed based on the filtration status assessment data. The fiber alignment angle and fiber spacing were optimized and calculated to obtain improved filter layer parameters, including: An evolution potential field is constructed based on the filtration status assessment data. Regions with oil-water separation efficiency greater than a preset efficiency threshold are set as positive potential energy regions, and regions with efficiency less than the preset efficiency threshold are set as negative potential energy regions. The evolution path direction is determined by the remaining usage time of the coalescing filter layer. Determine the potential energy values ​​of the positive and negative potential energy regions, determine the search space based on the potential energy distribution of the evolved potential field, use the fiber arrangement angle and fiber spacing as optimization variables, and set the search step size of the optimization variables based on the potential energy values, wherein the corresponding search step size of the positive potential energy region is greater than the corresponding search step size of the negative potential energy region. Calculate the potential energy gradient in the evolution potential field, use the potential energy gradient as the adjustment direction of the fiber structure, adjust the fiber spacing in the positive potential energy region to a first preset spacing according to the potential energy gradient, and adjust the fiber spacing in the negative potential energy region to a second preset spacing, wherein the first preset spacing value is greater than the second preset spacing value. The fiber arrangement angle is calculated based on the evolution path direction, so that the angle between the fiber arrangement angle and the evolution path direction is kept within a preset angle range; Based on the optimization results of the fiber arrangement angle and the fiber spacing, the pore distribution structure of the coalescing filter layer is adjusted to obtain the improved filter layer structure parameters.

8. A multi-stage filtration system for separating oil and wastewater based on coalescing materials, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire structural information and operating parameters of the multi-stage filtration device; The second unit is used to track and analyze oil droplet particles using deep reinforcement learning methods, obtain the oil droplet trajectory, construct a spatial fiber network structure for the filter layer based on the oil droplet trajectory, and obtain the three-dimensional pore distribution parameters of the coalescing filter layer by pore partitioning based on bubble growth. The third unit is used to input the three-dimensional pore distribution parameters into the graph neural network, construct the oil droplet coalescence network topology, analyze the oil droplet migration intensity to determine the coalescence channel, calculate the spatial coordinates and coalescence rate of the oil droplet coalescence point, and determine the oil droplet coalescence location and coalescence expansion trend of the coalescence filter layer. The fourth unit is used to calculate the remaining service time and oil-water separation efficiency of each coalescing filter layer based on the location and expansion trend of oil droplet coalescence, combined with structural information and operating parameters, and to generate filtration status assessment data. The fifth unit is used to construct an evolution potential field based on the filter state assessment data, optimize the fiber arrangement angle and fiber spacing, and obtain improved filter layer parameters. The sixth unit is used to apply the improved filter layer parameters to the multi-stage coalescing filter device, collect real-time oil-water separation data as new operating parameters, and perform iterative optimization.

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

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

Images