Ceramic membrane pollution prediction method and device, storage medium and electronic equipment

Abstract

The invention relates to the technical field of membrane separation, in particular to a ceramic membrane pollution prediction method, a storage medium and electronic equipment. According to the ceramic membrane pollution prediction method provided by the invention, firstly, relevant parameters of a ceramic membrane filtration process are collected in real time, and the relevant parameters comprise a macroscopic operation parameter, a microscopic interface parameter and a feed liquid property parameter; then establishing a quantitative relationship between the related parameters and the component resistance; calculating the total resistance of the membrane based on the quantitative relation, and representing the total resistance of the membrane as a function of component resistance; and finally, predicting the pollution degree of the ceramic membrane according to the total resistance and / or component resistance of the membrane. By comprehensively considering the macroscopic operation parameters, the microcosmic interface parameters and the material liquid property parameters, the interaction between the membrane and the pollutants in the material liquid is further analyzed, the future process of ceramic membrane pollution is analyzed, and the accuracy of predicting the pollution degree of the ceramic membrane is improved.

Description

Technical Field

[0001] This application relates to the field of membrane separation technology, specifically to a method for predicting ceramic membrane fouling, a storage medium, and an electronic device. Background Technology

[0002] Ceramic membranes, also known as inorganic ceramic membranes, are asymmetric membranes formed from inorganic ceramic materials through a special process. The core structure of a ceramic membrane is a porous tube wall densely covered with nano- or micron-sized micropores. During operation, external pressure drives the feed liquid to flow inside or outside the membrane tube. Small molecules can pass through the membrane pores, while large molecules, colloidal particles, and suspended solids are retained by the membrane, thus achieving functions such as material separation, concentration, purification, and wastewater treatment. It is a key component in the field of high-efficiency separation technology. Due to its advantages such as high temperature resistance, acid and alkali resistance, high mechanical strength, and strong chemical stability, ceramic membranes have been widely used in wastewater treatment and drinking water purification.

[0003] However, during the filtration process, various substances in the feed solution accumulate on the membrane surface or inside the pores through different mechanisms, forming a fouling layer that is difficult to eliminate naturally. This fouling layer significantly impacts the membrane's filtration performance. On one hand, it significantly increases the membrane's filtration resistance, causing a continuous rise in transmembrane pressure. To maintain the set filtration flux, the system needs to continuously increase its operating pressure, which not only increases energy consumption but may also accelerate mechanical damage to the membrane due to excessive pressure. On the other hand, the fouling layer reduces the effective filtration area of ​​the membrane and clogs the pores, leading to a significant drop in filtration flux. When the flux decreases to less than half of its initial value, the membrane system will be unable to meet production demands. Furthermore, long-term adhered pollutants can chemically erode the surface structure of the ceramic membrane, damaging its pore size distribution and separation performance, and shortening its lifespan.

[0004] Ceramic membrane fouling is an unavoidable problem during its operation, directly determining the filtration performance, energy costs, and equipment lifespan of the membrane system. More importantly, with prolonged operation, pollutants gradually solidify through deep adsorption of organic matter and lattice growth of inorganic matter, forming a stubborn fouling layer. Conventional cleaning methods are insufficient to completely remove this layer and may even further damage the membrane structure due to improper cleaning. Therefore, predicting the fouling state of ceramic membranes in advance and developing corresponding flushing strategies is a prerequisite for preventing fouling deterioration and ensuring the stable operation of the membrane system. Current technologies for predicting ceramic membrane fouling typically rely solely on macroscopic fouling data, i.e., predicting future fouling by monitoring changes in macroscopic indicators such as membrane flux and transmembrane pressure. However, this method has significant limitations: the formation of ceramic membrane pollutants is not only a gradual accumulation over time but also closely related to microscopic reactions such as molecular-level physical adsorption and interionic chemical bonding on the membrane surface; monitoring only macroscopic changes cannot capture these microscopic fouling triggers, ultimately leading to inaccurate fouling prediction results.

[0005] Therefore, improving the accuracy of predicting the degree of fouling in ceramic membranes is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] To overcome the problems existing in related technologies, this application provides a ceramic membrane fouling prediction method, storage medium, and electronic device to solve the problem of how to improve the accuracy of predicting the degree of ceramic membrane fouling.

[0007] The first aspect of this application provides a method for predicting ceramic membrane fouling, comprising: S1: Real-time acquisition of relevant parameters of the ceramic membrane filtration process, including macroscopic operating parameters, microscopic interface parameters and feed liquid property parameters. The microscopic interface parameters include at least the dielectric constant of the membrane surface and the Zeta potential of the membrane surface. S2: Establish a quantitative relationship between the relevant parameter and the component resistance, which includes at least one of filter cake layer resistance, organic matter adsorption resistance and inorganic scaling resistance; S3: Calculate the total membrane resistance based on this quantification relationship, and characterize the total membrane resistance as a function of this component resistance; S4: Predict the degree of ceramic membrane fouling based on the total resistance and / or component resistance of the membrane.

[0008] In a first possible implementation of the first aspect of this application, the method of collecting the dielectric constant of the film surface includes: S1.1a: Microwaves are emitted to the membrane surface by microwave transmission and the phase change is detected; S1.1b: Based on this phase change, the dielectric constant value of the film surface is calculated; The Zeta potential of the membrane surface was collected, including: S1.2a: A weak electric field is applied to the feed liquid on the membrane surface using an in-situ sensor; S1.2b: Measure the potential difference generated when the feed liquid flows on the membrane surface; S1.2c: Based on this potential difference, the Zeta potential value of the membrane surface is calculated.

[0009] In the second possible implementation of the first aspect of this application, calculating the filter cake layer resistance includes: S3.1a: Based on the current crossflow velocity, establish a model of the velocity field and shear stress distribution inside the ceramic membrane through finite element analysis; S3.1b: Based on this shear stress distribution model, determine the shear stress correction coefficients for different regions of the membrane surface; S3.1c: Using this correction factor, the initial value of filter cake layer resistance calculated based on macroscopic parameters is regionally corrected.

[0010] In the third possible implementation of the first aspect of this application, S2 includes: The coupling effect between the relevant parameters is analyzed by partial least squares path model, and the static correlation between the resistance of each component is quantified to continuously output the real-time estimated value sequence of the resistance of each component. S4 includes: S41: Input the real-time estimated value sequence into the pre-trained long short-term memory network model for prediction to obtain the predicted total membrane drag and the predicted component drag. S42: Determine the future fouling level of the ceramic membrane based on the predicted total membrane resistance and / or predicted component resistance.

[0011] In conjunction with the fourth possible implementation of the first aspect, the fifth possible implementation of the first aspect of this application, when a change in operating conditions is detected, further includes, between S2 and S4: S2.1a: Acquire several new sample data collected under the changed operating conditions; S2.1b: Freeze the weights of the first few layers of the network representing the general contamination pattern in this long short-term memory network model; S2.1c: Using the new sample data, update and fine-tune the weights of the remaining network layers that have not been frozen to form a long short-term memory network model adapted to the new working conditions.

[0012] In conjunction with the fifth possible implementation of the first aspect, S2.1c in the sixth possible implementation of the first aspect of this application includes: S2.1c-1: Input the new sample data and a portion of the sample data from the pre-training scenario into the basic prediction model, and extract their representations in the high-dimensional feature space respectively; S2.1c-2: Calculate the distribution difference between the new working condition sample data and the pre-trained working condition sample data in the high-dimensional feature space, as a measure of domain difference; S2.1c-3: Construct a domain-adaptive loss function, which is composed of a weighted average of the network's main task prediction loss and the domain-specific difference metric; S2.1c-4: With the goal of minimizing the adaptive loss function of the domain, the weights of the remaining network layers that have not been frozen are iteratively updated only through the backpropagation algorithm.

[0013] In the seventh possible implementation of the first aspect of this application, S4 includes: S51: Compare the total resistance of the membrane with the preset resistance threshold range to obtain the preliminary fouling level; S52: Identify the type of resistance with the largest value or the highest proportion in this component of resistance as the dominant pollution type; S53: Combining the preliminary pollution level with the dominant pollution type, a final pollution level is determined.

[0014] In conjunction with the seventh possible implementation of the first aspect, in the eighth possible implementation of the first aspect of this application, the resistance threshold range is dynamically calibrated using a Bayesian update algorithm, including: S5.1a: Record historical early warning records and subsequent cleaning effect data; S5.1b: If the pollution level rapidly escalates after multiple low-level warnings, the upper limit of the resistance threshold corresponding to the low level will be lowered through Bayesian updates. S5.1c: If the cleaning effect is not significant after multiple high-level warnings, the lower limit of the resistance threshold corresponding to the high level will be increased through Bayesian updates.

[0015] A second aspect of this application provides an electronic device comprising: a processor; and a memory having executable code stored thereon, wherein when the executable code is executed by the processor, the processor performs a ceramic membrane fouling prediction method as described in the first aspect and any possible implementation thereof.

[0016] A third aspect of this application provides a non-transitory machine-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform a ceramic membrane fouling prediction method as described in the first aspect and any possible implementation thereof.

[0017] The technical solution provided in this application may include the following beneficial effects: This application provides a method for predicting ceramic membrane fouling. First, relevant parameters of the ceramic membrane filtration process are collected in real time. These parameters include macroscopic operating parameters, microscopic interface parameters, and feed liquid property parameters. The microscopic interface parameters include at least the membrane surface dielectric constant and the membrane surface Zeta potential. Then, a quantitative relationship is established between these parameters and component resistances, which include at least one of the following: filter cake layer resistance, organic matter adsorption resistance, and inorganic fouling resistance. Next, the total membrane resistance is calculated based on the quantitative relationship and characterized as a function of the component resistances. Finally, the degree of ceramic membrane fouling is predicted based on the total membrane resistance and / or component resistances. This application, by comprehensively considering macroscopic operating parameters, microscopic interface parameters, and feed liquid property parameters, further analyzes the interaction between the membrane and pollutants in the feed liquid, analyzes the future process of ceramic membrane fouling, and improves the accuracy of predicting the degree of ceramic membrane fouling.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0020] Figure 1 This is a schematic flowchart illustrating a ceramic membrane fouling prediction method according to an embodiment of this application; Figure 2 This is another schematic flowchart illustrating a ceramic membrane fouling prediction method according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application. Detailed Implementation

[0021] Preferred embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0022] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0023] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0024] To address the aforementioned issues, this application provides a method for predicting ceramic membrane fouling.

[0025] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0026] See Figure 1 This embodiment provides a method for predicting ceramic membrane fouling, including: S1: Real-time acquisition of relevant parameters of the ceramic membrane filtration process, including macroscopic operating parameters, microscopic interface parameters and feed liquid property parameters. Microscopic interface parameters include at least the dielectric constant of the membrane surface and the Zeta potential of the membrane surface. It should be noted that this embodiment predicts the future degree of ceramic membrane fouling by real-time acquisition of relevant parameters during the ceramic membrane filtration process and calculating the total and / or component resistance of the membrane based on the current state using an algorithm model. This step involves real-time acquisition of relevant parameters during the ceramic membrane filtration process, including macroscopic operating parameters, microscopic interface parameters, and feed liquid property parameters. The microscopic interface parameters include at least the membrane surface dielectric constant and the membrane surface Zeta potential, which serve as the basis for subsequent calculations.

[0027] Among them, macroscopic operating parameters reflect the external operating conditions during ceramic membrane filtration, directly affecting filtration efficiency and fouling evolution rate. They are core parameters for quantifying cake layer resistance and judging fouling development trends. Feed liquid property parameters reflect the physicochemical properties of the feed liquid itself, determining the form and migration ability of pollutants in the feed liquid. They are important parameters for quantifying inorganic scaling resistance and assisting in judging the source of cake layer fouling.

[0028] Microscopic interface parameters refer to parameters that reflect the interfacial interaction characteristics between the ceramic membrane surface and the feed liquid. They directly determine the interaction strength between the membrane and pollutants in the feed liquid and are the core basis for quantifying the adsorption resistance of organic matter and the resistance to inorganic fouling. These parameters include at least the membrane surface dielectric constant and the membrane surface zeta potential. The membrane surface dielectric constant refers to the polarization ability of the ceramic membrane surface material under the action of an electric field, with units of F / m. It is a physical quantity used to characterize the electrostatic interaction between the membrane surface and charged particles in the feed liquid. In the process of pollution prediction, the larger the dielectric constant, the stronger the electrostatic adsorption capacity of the membrane surface for charged pollutants. Organic matter with opposite charges in the feed liquid is more easily attached to the membrane surface through electrostatic attraction, leading to an increase in organic matter adsorption resistance. Simultaneously, changes in the dielectric constant can reflect the formation state of the fouling layer on the membrane surface, providing a microscopic basis for judging the degree of organic matter adsorption pollution. The zeta potential of a membrane surface refers to the potential difference between the sliding surface of the ceramic membrane and the feed liquid, measured in mV. It is a key parameter characterizing the electrical properties and intensity of the membrane surface. The zeta potential's effect is that the larger the absolute value of the potential, the stronger the electrostatic interaction between the membrane surface and charged contaminants in the feed liquid. If the zeta potential is opposite in charge to the contaminant, electrostatic attraction will promote contaminant adsorption, increasing the adsorption resistance of organic matter. If the zeta potential is the same as the contaminant charge, electrostatic repulsion will inhibit adsorption and scaling, reducing the adsorption resistance of organic matter and the scaling resistance of inorganic matter. Furthermore, changes in the zeta potential can reflect the type of membrane surface fouling, providing a basis for distinguishing the dominant fouling type.

[0029] It is worth noting that there are two methods for judging and predicting ceramic membrane fouling in existing technologies. One method relies solely on the macroscopic operating parameters of the ceramic membrane to determine its fouling level. This method, if it does not involve predicting future trends, cannot identify the initial generation and accumulation of pollutants in advance. Even if the fouling level is known in advance, the accuracy of the prediction is insufficient because it only uses macroscopic operating parameters as a single benchmark and ignores further reactions between pollutants. The other method infers the future fouling of the ceramic membrane based on the composition of the feed solution. However, due to the complex and variable composition of wastewater feed solutions, it is difficult to accurately obtain comprehensive and real-time composition data in practical applications, easily leading to inaccurate prediction results. Furthermore, when predicting fouling based on feed solution composition, the coupling effect of various factors results in too many influencing variables, leading to unstable final prediction results. This embodiment, however, predicts the fouling situation after a preset time by jointly obtaining macroscopic operating parameters, microscopic interface parameters, and feed solution property parameters. The macroscopic operating parameters reflect the overall severity of fouling, while the microscopic interface parameters reflect possible internal reactions and changes. In addition, the micro-interface parameters here refer to the micro-interface parameters of the ceramic membrane, rather than the micro-characteristics of the feed liquid. Obtaining these parameters is simpler and more stable than obtaining the micro-characteristics of the feed liquid.

[0030] Preferably, in actual operation, if the obtained macroscopic operating parameters and microscopic interface parameters are not highly correlated with the fouling status of the ceramic membrane, the predicted results will be inaccurate. To solve this problem, this embodiment provides a preferred solution. In a preferred method for predicting ceramic membrane fouling, the macroscopic operating parameters include transmembrane pressure difference, membrane flux, and crossflow velocity; the feed liquid property parameters include ionic strength and turbidity change rate. It should be noted that: transmembrane pressure difference refers to the pressure difference between the feed side and the permeate side of the ceramic membrane module. Its essential concept is the driving force for the feed liquid to pass through the membrane pores. In fouling prediction, the larger the transmembrane pressure difference, the faster the migration rate of suspended particles, organic matter, etc. in the feed liquid to the membrane surface, the higher the risk of filter cake formation and accumulation, and the corresponding proportion of filter cake resistance will also increase significantly. At the same time, a sudden change in transmembrane pressure difference can directly serve as an early warning signal for a sudden fouling event, indicating that a large number of particles may have deposited or scaled on the membrane surface. Membrane flux refers to the volume of feed liquid passing through a unit effective membrane area per unit time. Its concept is a key indicator for measuring the filtration efficiency of ceramic membranes and reflects the actual treatment capacity of the membrane module. Cross-flow velocity refers to the speed at which the feed liquid flows parallel to the surface of the ceramic membrane. It is a parameter that characterizes the shear strength of the feed liquid on the membrane surface. The higher the cross-flow velocity, the stronger the shear stress of the feed liquid on the membrane surface, which can effectively flush away the particles and filter cake layer already deposited on the membrane surface, inhibit the continuous thickening of the filter cake layer, and thus reduce the growth rate of filter cake layer resistance. If the cross-flow velocity is too low, the shear effect is insufficient, and the filter cake layer is prone to rapid accumulation, resulting in a significant increase in the total membrane resistance in a short period of time. Therefore, cross-flow velocity is a key operating parameter for controlling filter cake layer fouling.

[0031] Preferably, in actual operation, if contact-based acquisition or non-professional detection methods are used, data deviations are easily caused by interference with the flow field, contamination of the membrane surface, or insufficient detection sensitivity, which in turn affects the accuracy of subsequent component resistance calculations and contamination degree determination. The acquisition method of the membrane surface dielectric constant and membrane surface Zeta potential directly affects the data accuracy. To solve this technical problem, this embodiment also provides a preferred solution, in which the acquisition of the membrane surface dielectric constant includes: S1.1a: Microwaves are emitted to the membrane surface by microwave transmission and the phase change is detected; It should be noted that the microwave transmission method is a non-contact detection method for measuring the dielectric constant of a membrane surface, utilizing the physical phenomenon that the phase and amplitude of microwaves change when propagating through media with different dielectric properties. Specifically, a fixed-frequency microwave signal is first emitted to the surface of the ceramic membrane. After the microwaves penetrate the contaminant layer and reach the membrane body, part of the signal is reflected or absorbed. The relative change of the remaining transmitted signal is quantitatively correlated with the dielectric constant of the membrane surface; the larger the dielectric constant, the slower the microwave propagation speed and the more pronounced the phase lag. Phase change refers to the shift in the vibration phase of the microwave signal relative to the incident signal after passing through the membrane surface and the feed liquid interface; it is a direct physical quantity reflecting the dielectric properties of the membrane surface.

[0032] It is worth noting that, unlike existing technologies, this embodiment uses microwave projection instead of the traditional contact measurement method. This is because the traditional contact measurement method is prone to contaminating the membrane surface and interfering with the flow field. This ensures that the dynamic changes of the dielectric constant of the membrane surface are accurately captured without affecting the ceramic membrane filtration process, providing a microscopic basis for quantifying the adsorption resistance of organic matter.

[0033] S1.1b: The dielectric constant value of the film surface is calculated based on the phase change; It should be noted that calculating the dielectric constant of a film surface based on phase change utilizes the quantitative correlation between microwave phase change and the dielectric constant of the film surface. When microwaves propagate in a medium, the dielectric constant determines the propagation speed; the larger the dielectric constant, the slower the propagation speed. This difference in propagation speed is directly reflected in phase lag. By establishing a linear model of phase standardization and dielectric constant through experimental calibration, the accurate calculation of the dielectric constant can be achieved.

[0034] The Zeta potential on the membrane surface was collected, including: S1.2a: A weak electric field is applied to the feed liquid on the membrane surface using an in-situ sensor; It should be noted that in-situ sensors refer to dedicated electrode sensors directly embedded inside the ceramic membrane module and installed flush with the membrane surface. They measure through non-invasive contact with the feed liquid, ensuring direct contact between the electrode and the feed liquid on the membrane surface to apply an electric field, while avoiding interference with the flow field or scratching of the membrane structure by the electrode touching the membrane surface. This solves the problems of traditional external electrodes disrupting the flow field and causing large measurement deviations. A weak electric field refers to a low-intensity electric field applied to the feed liquid on the membrane surface. Its core function is to drive charged particles in the feed liquid to move along the direction of the electric field, while avoiding feed liquid electrolysis or membrane surface damage caused by strong electric fields. A weak electric field is the basis for measuring the zeta potential of the membrane surface. By using the potential difference of the feed liquid flow under the action of the electric field, the charging characteristics of the membrane surface can be inferred, providing a microscopic basis for quantifying the adsorption resistance of organic matter and the resistance to inorganic fouling.

[0035] S1.2b: Measure the potential difference generated when the feed liquid flows on the membrane surface; It should be noted that the potential difference generated when the feed liquid flows on the membrane surface refers to the potential difference formed between the sliding surface of the ceramic membrane and the feed liquid mass after a microfluidic electric field is applied, as charged particles in the feed liquid move along the direction of the electric field. This potential difference is also known as the flow potential difference. The membrane surface becomes charged due to the adsorption of ions from the feed liquid. When a weak electric field drives the feed liquid to flow, the charge distribution on both sides of the sliding surface becomes unbalanced, thus generating a potential difference. This difference is directly related to the zeta potential of the membrane surface and is an intermediate variable for quantifying the charged characteristics of the membrane surface.

[0036] S1.2c: The Zeta potential value of the membrane surface is calculated based on the potential difference.

[0037] It should be noted that the Zeta potential of the membrane surface is calculated based on the potential difference. According to the classical theoretical formula for charged particle flow in an electric field, a quantitative correlation is established between the potential difference of the feed liquid flow and the Zeta potential of the membrane surface. A weak electric field applied drives the movement of charged particles in the feed liquid, and a potential difference is generated on both sides of the sliding surface of the membrane due to the difference in charge distribution. This potential difference is linearly positively correlated with the Zeta potential. By experimentally calibrating the parameters of the correction equation, the Zeta potential value can be derived from the measured potential difference.

[0038] S2: Establish a quantitative relationship between relevant parameters and component resistance, where component resistance includes at least one of filter cake layer resistance, organic matter adsorption resistance, and inorganic scaling resistance. It should be noted that this step calculates the resistance of each component by establishing a quantitative relationship between relevant parameters and component resistance. Among them, component resistance is a subdivision of the total resistance of ceramic membranes, reflecting the degree of obstruction to the membrane filtration process by different types of fouling. In this embodiment, it includes at least one of the following three categories: cake layer resistance refers to the resistance generated by the deposition of suspended particles in the feed liquid on the membrane surface to form a cake layer, which is directly related to the transmembrane pressure difference and cross-flow velocity. The higher the transmembrane pressure difference, the lower the cross-flow velocity, and the thicker the cake layer, the higher the resistance value. Organic matter adsorption resistance refers to the filtration resistance caused by organic matter in the feed liquid adhering to the membrane surface or membrane pores through electrostatic adsorption and hydrophobic interaction. It is mainly affected by the dielectric constant of the membrane surface, the Zeta potential, and the ionic strength of the feed liquid. The higher the dielectric constant, the smaller the absolute value of the Zeta potential, and the higher the ionic strength, the stronger the organic matter adsorption and the higher the resistance value. Inorganic scaling resistance refers to the filtration resistance generated by the crystallization and deposition of inorganic ions in the feed liquid on the membrane surface or membrane pores. It is mainly determined by the ionic strength of the feed liquid and the transmembrane pressure difference. The higher the ionic strength and the greater the transmembrane pressure difference, the higher the risk of scaling and the higher the resistance value. It is worth noting that the specific types of component resistances are determined by the type of wastewater being treated. For example, when the wastewater being treated does not contain organic matter, organic matter adsorption resistance is not included. Quantitative relationships refer to establishing mathematical expressions for relevant parameters, latent variables, and component resistances through algorithms or models. The core of this approach is to clarify the influence coefficient of each relevant parameter on each component resistance, ultimately forming a functional relationship that can be directly used for calculation. This provides a quantitative basis for subsequent calculations of total membrane resistance and determination of the degree of fouling.

[0039] Preferably, in actual operation, during ceramic membrane filtration, there are significant differences in shear stress in different regions of the membrane surface, resulting in uneven filter cake thickness. The greater the shear stress, the stronger the scouring ability of the feed liquid on the membrane surface particles, resulting in a thinner filter cake and lower resistance; conversely, the smaller the shear stress, the thicker the filter cake and the greater the resistance. Traditional filter cake resistance calculation ignores this difference and directly uses the average shear stress, leading to significant errors. To solve this technical problem, this embodiment provides a preferred solution. In the ceramic membrane fouling prediction method of this preferred solution, the calculation of the filter cake resistance includes: S3.1a: Based on the current crossflow velocity, establish a model of the velocity field and shear stress distribution inside the ceramic membrane through finite element analysis; It should be noted that finite element analysis (FEM) refers to a numerical simulation method that discretizes the three-dimensional flow channel structure of a ceramic membrane module into a finite number of element cells and solves the fluid dynamics control equations through numerical calculations to simulate the flow state of the feed liquid inside the membrane. Its core advantage is its ability to accurately capture microscopic flow characteristics such as local turbulence and velocity gradients within the flow channel, solving the problem that traditional empirical formulas cannot reflect the non-uniform velocity distribution inside the membrane. The velocity field model describes the spatial distribution of the magnitude and direction of the feed liquid velocity inside the ceramic membrane, reflecting the differences in flow rate at different locations. The shear stress distribution model describes the spatial distribution of the shear force generated on the membrane surface by the feed liquid flow, reflecting the membrane surface's ability to resist cake buildup in different regions.

[0040] S3.1b: Determine the shear stress correction coefficients for different regions of the membrane surface based on the shear stress distribution model; It should be noted that the shear stress correction coefficient is a dimensionless coefficient based on the shear stress distribution model, which quantifies the difference in the influence of actual shear stress on the filter cake layer accumulation in different regions of the membrane surface. The initial value of the filter cake layer resistance, which assumes uniform shear stress on the membrane surface, is adjusted to the true resistance value that reflects the local shear stress difference through the correction coefficient.

[0041] S3.1c: Using correction coefficients, the initial value of filter cake layer resistance calculated based on macroscopic parameters is regionally corrected.

[0042] It should be noted that the initial value of the filter cake layer resistance refers to the value calculated using traditional empirical formulas based solely on the collected macroscopic operating parameters. Its limitation lies in assuming uniform shear stress on the membrane surface and neglecting differences in the flow field across different regions, leading to reduced calculation error and failing to reflect the true state of local filter cake layer accumulation. Regional correction refers to the process of dividing the initial value into regions based on determined shear stress correction coefficients for different regions of the membrane surface, correcting each region separately, and then obtaining the final filter cake layer resistance through area weighting.

[0043] S3: Calculate the total membrane resistance based on the quantization relationship, and characterize the total membrane resistance as a function of the component resistance; It should be noted that total membrane resistance is the overall resistance that prevents the feed liquid from passing through the membrane module during ceramic membrane filtration. It is the sum of various component resistances, such as filter cake layer resistance, organic matter adsorption resistance, and inorganic fouling resistance. It directly determines the stability of membrane flux and filtration efficiency. The higher the total membrane resistance, the more difficult it is for the feed liquid to pass through, and the more significant the membrane flux decline. It is a core indicator for judging the overall fouling degree of ceramic membranes. In this embodiment, the total membrane resistance needs to be calculated based on the relevant parameters and component resistance quantification relationships established in S2 to ensure the accuracy and relevance of the data source.

[0044] The component resistance function refers to a mathematical expression that clearly defines the quantitative relationship between the total membrane resistance and the filter cake layer resistance, organic matter adsorption resistance, and inorganic fouling resistance. Its core is to reflect the "proportion of each component resistance's contribution to the total resistance." Unlike the crude calculations of simple addition in existing technologies, this application constructs a function containing coupling correction terms based on the physical mechanism of ceramic membrane fouling. This reflects the interaction between component resistances; for example, after the filter cake layer forms, it further promotes the adsorption of organic matter within the pores of the filter cake layer, causing the rate of increase in total resistance to exceed the sum of individual resistances.

[0045] S4: Predict the degree of ceramic membrane fouling based on total membrane resistance and / or component resistance.

[0046] It should be noted that the degree of ceramic membrane fouling refers to the extent to which the membrane's permeability decreases during the filtration process due to factors such as filter cake buildup, organic matter adsorption, and inorganic scaling. It is primarily reflected in the combined effect of total membrane resistance and the proportion of contribution from the dominant fouling type. Unlike traditional methods that solely determine fouling based on membrane flux decline, this application combines total membrane resistance and component resistance to achieve a dual assessment of both the overall degree of fouling and the dominant type, providing a precise basis for subsequent pollution control.

[0047] It is worth noting that, unlike existing technologies that use a single judgment logic, this judgment logic is based on the classification of total membrane resistance, supplemented by the proportion of component resistance. The core is to divide the pollution level by a preset resistance threshold range, while simultaneously identifying the component resistance type that contributes the most to the total resistance, ultimately forming a comprehensive judgment result of pollution level and dominant type. For example, moderate pollution and organic matter adsorption dominance not only clearly define the overall pollution severity but also point out the core pollution cause, solving the pain point of traditional judgments that only know the pollution but not the cause.

[0048] Preferably, in actual operation, if only the severity of contamination is known but the type of contamination is unknown, the subsequent matching of appropriate cleaning strategies lacks specificity. To solve this technical problem, see [link to relevant documentation]. Figure 2 This embodiment provides a preferred solution, in which S4 of the ceramic membrane fouling prediction method includes: S51: Compare the total membrane resistance with the preset resistance threshold range to obtain the preliminary fouling level; It should be noted that the preset resistance threshold range refers to a range of multiple resistance critical values ​​determined through experimental calibration and dynamic calibration based on the characteristics of the ceramic membrane material, filtration process requirements, and cleaning costs. This range is used to classify different levels of contamination and serves as an objective standard for preliminary contamination level determination. The preliminary contamination level refers to a preliminary classification of the overall severity of ceramic membrane contamination based on the comparison between the total membrane resistance and the threshold range. This is used to quickly locate the overall level of contamination and provide a basic framework for subsequent identification and comprehensive judgment of dominant contamination types.

[0049] S52: Identify the type of resistance with the largest value or the highest proportion in the component resistance as the dominant pollution type; It should be noted that this step uses both numerical comparison and proportion calculation to select the type of resistance that contributes the most to the total resistance as the dominant pollution type. When a certain component resistance meets both the criteria of having the largest numerical value and having a proportion exceeding a preset threshold, it can be identified as the core pollution cause, providing a direct basis for subsequent targeted prevention and control. Component resistance refers to the detailed components of the calculated total pollution resistance of the ceramic membrane, including at least one of the following: filter cake layer resistance, organic matter adsorption resistance, and inorganic scaling resistance. These correspond to three pollution mechanisms: particle accumulation, organic matter adsorption, and inorganic particle crystallization, respectively, and are the core basis for distinguishing pollution types. The dominant pollution type refers to the pollution type corresponding to the resistance with the largest numerical value or the highest proportion of the total resistance, used to further pinpoint the core pollution cause based on the initial pollution level.

[0050] S53: Combine the preliminary pollution level with the dominant pollution type to comprehensively determine the final pollution level.

[0051] It should be noted that the final pollution level refers to a dual-dimensional judgment structure that integrates the initial pollution level and the dominant pollution type, forming a severity and cause assessment structure. The initial pollution level serves as the framework, while the details of the dominant pollution type are used to supplement it, thereby increasing the dimensionality of the final pollution level and providing a complete decision-making basis for subsequent differentiated pollution prevention and control.

[0052] Preferably, in actual operation, if the resistance threshold range remains fixed and is not dynamically adjusted in conjunction with the actual early warning effect and pollution evolution law, it will lead to a mismatch between the threshold and the actual situation, resulting in inaccurate judgment of the deterioration rate. To solve this technical problem, this embodiment provides a preferred solution: The resistance threshold range is dynamically calibrated using a Bayesian update algorithm, including: S5.1a: Record historical early warning records and subsequent cleaning effect data; It should be noted that this preferred solution establishes a closed loop for evaluating the effectiveness of early warnings. If the cleaning effect is poor after a certain type of early warning, or if the pollution escalates rapidly after a low-level early warning, it indicates that the current resistance threshold is not suitable. In this case, the threshold is adjusted according to the Bayesian update algorithm to improve the suitability of the resistance threshold.

[0053] Historical early warning records refer to complete data records when pollution early warnings are issued based on comprehensive judgment. These records include pollution state parameters, warning level, and warning timestamp at the time the warning was triggered. They reflect when the warning was triggered and based on what state, serving as the basis for verifying the rationality of the warning. Subsequent cleaning effect data refers to the entire process data of pollution control implemented after the warning, including cleaning scheme parameters, changes in membrane performance before and after cleaning, and stable operation time after cleaning. These data primarily reflect the effectiveness of the control measures after the warning and are a key basis for calibrating the resistance threshold.

[0054] S5.1b: If the pollution level rapidly escalates after multiple low-level warnings, the upper limit of the resistance threshold corresponding to the low level will be lowered through Bayesian updates. It should be noted that: a low-level warning refers to a mild pollution warning output by comprehensive judgment, corresponding to the initial resistance threshold range, reflecting an initial judgment state where the pollution level is low and high-intensity cleaning is not yet required. Rapid escalation of pollution level refers to the situation where, after a low-level warning, the total membrane resistance rapidly exceeds the upper limit of the current level threshold within a preset time, jumping to the highest pollution level, indicating that the pollution development rate after the warning far exceeds the process expectations. Bayesian updating, based on Bayesian probability theory, refers to dynamically adjusting the upper limit of the resistance threshold for low-level warnings according to the calculation logic of prior probability, likelihood probability, and posterior probability, based on the historical data of low-level warnings and rapid escalation recorded in the previous step. This addresses the pain point that traditional fixed thresholds cannot match changes in pollution rate, making the threshold more closely aligned with the actual pollution evolution pattern.

[0055] S5.1c: If the cleaning effect is not significant after multiple high-level warnings, the lower limit of the resistance threshold corresponding to the high level will be increased through Bayesian updates.

[0056] It should be noted that a high-level warning refers to a comprehensive assessment resulting in a severe or moderate pollution warning. This corresponds to a high-level initial resistance threshold range, where the degree of pollution has significantly impacted filtration efficiency, necessitating intensive cleaning. This is a critical threshold distinguishing between moderate and severe pollution and directly determines the triggering timing of the high-level warning. A situation where cleaning is ineffective means that after a high-level warning, the recommended cleaning protocol was followed, but the membrane performance did not meet the specified effective recovery standards.

[0057] The beneficial effects of this embodiment: ① In this embodiment of the ceramic membrane fouling prediction method, relevant parameters of the ceramic membrane filtration process are first collected in real time. These parameters include macroscopic operating parameters, microscopic interface parameters, and feed liquid property parameters. The microscopic interface parameters include at least the membrane surface dielectric constant and the membrane surface Zeta potential. Then, a quantitative relationship is established between these parameters and component resistances. The component resistances include at least one of the following: filter cake layer resistance, organic matter adsorption resistance, and inorganic fouling resistance. Subsequently, the total membrane resistance is calculated based on the quantitative relationship and characterized as a function of the component resistances. Finally, the degree of ceramic membrane fouling is predicted based on the total membrane resistance and / or component resistances. This embodiment, by comprehensively considering macroscopic operating parameters, microscopic interface parameters, and feed liquid property parameters, further analyzes the interaction between the membrane and pollutants in the feed liquid, analyzes the future process of ceramic membrane fouling, and improves the accuracy of predicting the degree of ceramic membrane fouling.

[0058] ② This invention, by introducing microscopic interface parameters such as the dielectric constant and Zeta potential of the membrane surface for real-time monitoring, can sensitively capture changes in the interfacial physicochemical properties caused by the initial adsorption and deposition of pollutants on the membrane surface before significant changes occur in macroscopic operating parameters such as hydraulic flux and transmembrane pressure difference. Subsequently, by establishing precise quantitative relationships between these microscopic, macroscopic, and feed parameters and the resistance of each pollutant component, the system can quantify the increase in organic matter adsorption resistance or initial filter cake layer resistance based on weak early signals. Finally, based on the analysis of the above component resistance sequences, the pollution trend can be predicted, thereby issuing early warnings at a stage when the total amount of pollution is extremely low and the macroscopic manifestations are not obvious. Compared with traditional methods that rely on significant deterioration of macroscopic parameters before issuing alarms, this greatly advances the intervention opportunity, creating conditions for implementing low-cost, high-efficiency light cleaning.

[0059] ③ Traditional methods treat each parameter independently, ignoring the coupling effect between them, leading to inaccurate predictions when the feed composition or operating conditions change. This invention constructs a quantitative relationship model capable of analyzing the coupling effect between multiple parameters, and for the first time quantifies the interaction of multiple physical fields and multiple chemical conditions, such as "transmembrane pressure difference - crossflow velocity - ionic strength," in mathematical form. This relationship model can dynamically reflect the changes in the contribution weight of each parameter to fouling resistance under different conditions. Based on this, the total membrane resistance is scientifically characterized as a function of the resistance of each component, rather than a general statistical value, enabling the model to accurately distinguish and quantify the contribution proportion of different fouling mechanisms such as filter cake layer, organic matter adsorption, and inorganic fouling. This decoupled modeling approach based on fouling mechanisms allows the prediction model to maintain a deep and accurate understanding of the fouling process even when facing complex operating conditions with strong parameter interactions, thus significantly improving the accuracy and stability of the prediction results. Example

[0060] In practical operation, the degree of matching between the algorithm and model responsible for predicting ceramic membrane fouling and the actual application requirements directly affects the accuracy and computational efficiency of the final fouling prediction results. If the algorithm model does not match the requirements, on the one hand, the computational speed may slow down due to the model's inefficient adaptation to data characteristics, thus prolonging the fouling prediction response time; on the other hand, the model may fail to accurately capture the fouling mechanism, resulting in a significant deviation between the computational results and the actual fouling level, failing to provide a reliable basis for cleaning early warning. To solve this technical problem, this embodiment optimizes upon embodiment one. In a ceramic membrane fouling prediction method of this embodiment, S2 includes: The coupling effect between the relevant parameters is analyzed by partial least squares path model, and the static correlation between the resistance components is quantified to continuously output the real-time estimated value sequence of each resistance component. It should be noted that the partial least squares path model is a multivariate statistical modeling method that integrates multiple regression and path analysis. Its core essence is to extract latent variables related to relevant parameters and component resistances, establishing linear or nonlinear mapping relationships between variables, while simultaneously addressing issues such as multicollinearity among parameters and insufficient sample size. In this embodiment, the partial least squares path model is used to simultaneously handle three types of high-dimensional inputs: macroscopic operating parameters, microscopic interface parameters, and feed liquid property parameters. This accurately quantifies the contribution weight of each parameter to filter cake resistance, organic matter adsorption resistance, and inorganic scaling resistance, avoiding the quantification bias caused by traditional models due to single parameters or neglect of interface effects. The coupling effect of relevant parameters refers to the mutual influence between macroscopic operating parameters, microscopic interface parameters, and feed liquid property parameters. This coupling relationship indirectly changes the contribution weight of each parameter to component resistances and needs to be analyzed through model separation. Static correlation refers to the quantitative mapping relationship between relevant parameters and filter cake resistance, organic matter adsorption resistance, and inorganic scaling resistance under stable operating conditions. This relationship does not change dynamically over time and only requires calibration based on stable operating data. To meet real-time estimation needs, the partial least squares path model continuously receives relevant parameters acquired in real time, enabling it to continuously calculate and output estimated values ​​for each component resistance, thus forming a continuous real-time estimation value sequence. Here, 'static' means that the internal correlation of the model remains unchanged under stable operating conditions, not that its output is singular. The real-time estimation value sequence refers to the time-series data formed based on the continuous estimation results of each output component resistance, providing continuous input for subsequent model evaluation.

[0061] S4 includes: S41: Input the real-time estimated value sequence into the pre-trained long short-term memory network model for prediction to obtain the predicted total membrane drag and the predicted component drag. It should be noted that the pre-trained Long Short-Term Memory (LSTM) model is a time-series training model based on historical time-series data under multiple operating conditions. The component resistance time-series labels in its training data are calculated by analyzing the coupling relationship between historical macroscopic, microscopic, and feed liquid parameters using a partial least squares path model. The LSTM gating mechanism captures the long-term dependencies of component resistance, addressing the pain point of traditional time-series models in handling long-series dependencies and adapting to the characteristics of slow accumulation and dynamic evolution of ceramic membrane fouling. Predicted total membrane resistance and predicted component resistance refer to the fouling resistance data output by the LSTM model after a preset time. Predicted component resistance is the future value of each individual resistance type, while predicted total membrane resistance is the result of the sum of the three, together constituting a quantitative basis for the future degree of ceramic membrane fouling, providing decision support for subsequent cleaning and early warning.

[0062] S42: Determine the future fouling level of the ceramic membrane based on the predicted total membrane resistance and / or predicted component resistance.

[0063] It should be noted that: the predicted total membrane resistance refers to the total resistance of the ceramic membrane after a preset time, output by the LSTM model. It is the superposition of the filter cake layer resistance, organic matter adsorption resistance, inorganic fouling resistance, and membrane body resistance, reflecting the overall degree of future pollution. The predicted component resistance refers to the future values ​​of the three types of pollution resistance output by the LSTM model alone, reflecting the core causes of future pollution. The future degree of ceramic membrane pollution refers to the comprehensive judgment result of the ceramic membrane pollution status after a preset time, based on the predicted resistance data, from the dual dimensions of overall severity level and main pollution type. This fundamentally addresses the pain point of traditional methods that rely solely on total resistance, which cannot predict future risks or identify the causes of pollution, and provides a basis for formulating cleaning strategies in advance. The judgment logic uses the predicted total membrane resistance to classify severity levels as a framework, supplemented by the predicted component resistance to identify the dominant type, forming a two-layer judgment system of overall and local. First, the severity of pollution is determined by the total resistance, and then the core causes are located by the proportion of component resistance, ensuring that the judgment results have both quantitative basis and practical guidance significance.

[0064] Optionally, a hypothetical path diagram of pollution formation can be constructed based on prior knowledge of pollution mechanisms using structural equation modeling. Then, through iterative fitting, the standardized coefficients of all the above paths are output, thereby accurately quantifying the direct and indirect effects of each parameter on the drag component, thus establishing a quantitative relationship. Subsequently, the real-time estimated values ​​of each component drag continuously calculated and output by the structural equation model are input into a temporal convolutional network model to predict the total membrane drag and component drag at one or more future time steps.

[0065] Preferably, in actual operation, when the scenario to which the pollution prediction model is applicable changes significantly, there may be situations where the original training data lacks relevant parameters. Therefore, it is necessary to retrain the model. However, existing methods typically involve retraining the entire model, which is costly and inefficient. To address this technical problem, this embodiment provides a preferred solution. In a pollution prediction method for ceramic membranes in this preferred solution, when a change in operating conditions is detected, the method further includes the following step after S2: S2.1a: Acquire several new sample data collected under the changed operating conditions; It should be noted that: "Changed operating conditions" refers to a state in which the operating conditions or the object being treated of the ceramic membrane filtration system have significantly changed, including abrupt changes in operating parameters, variations in feed properties, and switching of filtration stages. "New sample data" refers to characteristic data collected after the change in operating conditions, used to characterize the current filtration process. This data must contain parameter dimensions consistent with the training data of the Long Short-Term Memory (LSTM) network model and possess timeliness and representativeness. "Acquisition logic" refers to a closed-loop process of real-time detection, operating condition determination, data collection, and quality inspection to automatically capture changes in operating conditions and extract qualified samples. This ensures that the new sample data accurately reflects the contamination patterns under the changed operating conditions, providing reliable input for model adaptation.

[0066] S2.1b: Freeze the weights of the first few layers of the network representing the general contamination pattern in the Long Short-Term Memory network model; It should be noted that the Long Short-Term Memory (LSTM) network model is divided into bottom and top layers according to the granularity of feature extraction. The first few layers are the bottom layer, responsible for extracting general temporal features, which are common under different operating conditions. The latter few layers are the top layer, responsible for learning features specific to particular operating conditions, which vary significantly with changes in operating conditions. General contamination patterns refer to universal patterns in ceramic membrane filtration that are unaffected by specific operating conditions, such as temporal and correlational patterns. Weight freezing refers to locking the weight parameters of several layers in the LTM network model, preventing them from being updated during fine-tuning, and allowing only the top layer network modules to learn features for new operating conditions. Its purpose is to preserve general patterns, adapt to specific features, solve the problem of forgetting general knowledge caused by traditional full-scale fine-tuning, and meet the needs of efficient transfer learning.

[0067] S2.1c: Using new sample data, update and fine-tune the weights of the remaining network layers that have not been frozen to form a long short-term memory network model adapted to the new working conditions.

[0068] It should be noted that the remaining unfrozen network layers refer to the layers responsible for extracting specific features. Update fine-tuning refers to optimizing the weight parameters of the unfrozen layers using a backpropagation algorithm based on the acquired new sample data, minimizing the deviation between the model output and the new sample labels. Its core is to preserve general patterns and adapt to specific features, which, compared to full retraining and predefined models in existing technologies, is a crucial step in balancing model stability and adaptability. The Long Short-Term Memory (LSTM) network model adapted to the new operating conditions refers to the model formed after fine-tuning, capable of accurately predicting the pollution trend of the changed operating conditions. It must meet the prediction error of component drag under the new operating conditions while retaining the predictive ability for the original operating conditions.

[0069] Preferably, in actual operation, when the applicable scenario of the pollution prediction model changes significantly, there may be situations where relevant parameters are missing from the original training data. Therefore, it is necessary to retrain the model. However, existing methods typically involve retraining the entire model, which is costly and inefficient. To solve this technical problem, this embodiment provides a preferred solution. In the preferred solution of the pollution prediction method for ceramic membranes, S2.1c includes: S2.1c-1: Input the new sample data and some sample data from the pre-training scenario into the long short-term memory network model, and extract their representations in the high-dimensional feature space respectively; It should be noted that: new sample data refers to the acquired data under the changed operating conditions, containing features specific to the new operating conditions; the sample data for the pre-training operating conditions refers to samples selected from the basic training dataset that are related to the new operating conditions, and must cover the general contamination rules learned by the Long Short-Term Memory (LSTM) network model. The LTM network model is used to map element parameters to a high-dimensional feature space, outputting a feature representation with semantic information. The high-dimensional feature space representation refers to the high-dimensional vector formed after feature extraction from the sample data by the LTM network model. This vector encodes the temporal features, correlation features, and potential contamination rules of the parameters, and is the core output for subsequent cross-operating condition feature alignment and weight updates. Essentially, it transforms the original parameters from a physical quantity space into a feature semantic space, facilitating the quantification of the differences between the new and pre-training operating conditions, and providing interpretable feature basis for fine-tuning.

[0070] S2.1c-2: Calculate the distribution difference between the new working condition sample data and the pre-trained working condition sample data in the high-dimensional feature space, as a measure of domain difference; It should be noted that: the high-dimensional feature space refers to the space formed by the extracted feature vectors output from the frozen layer of the Long Short-Term Memory network model. This space encodes the general contamination patterns and specific features of the sample data. Distribution difference refers to the degree of deviation between the data distribution of the new operating condition samples and the pre-training operating condition samples in the high-dimensional feature space, primarily reflecting the difference in the contamination mechanisms of the two types of operating conditions. The domain difference measure quantifies the distribution difference into a value in the interval [0, 1], serving as an indicator to judge the domain similarity between the new operating condition and the pre-training operating condition; the smaller the value, the smaller the difference; the larger the value, the larger the difference. It provides a quantitative basis for the fine-tuning intensity of subsequent weight updates.

[0071] S2.1c-3: Construct a domain-adaptive loss function, which is composed of a weighted average of the network's main task prediction loss and a domain difference metric; It should be noted that the domain-adaptive loss function is a conformance loss function designed to balance the prediction accuracy of new operating conditions with the consistency of features across operating conditions. By integrating the main task prediction loss and the domain difference metric, it guides the update of the unfrozen layer weights to optimize in a direction that both adapts to the specific features of the new operating condition and retains the general rules of pre-training. This resolves the contradiction caused by the traditional single loss function, which may either overfit to the new operating condition and lose general knowledge or retain general knowledge but fail to adapt adequately. The main task prediction loss refers to the deviation between the model's predicted output for the new operating condition samples and the measured labels. It primarily reflects the model's direct adaptability to the new operating condition and is quantified using mean squared error. The domain difference metric, i.e., the calculated interval, quantifies the difference in distribution between the new operating condition and the pre-training operating condition in the high-dimensional feature space. It serves as an indicator to constrain the feature shift across operating conditions. The larger the interval, the more severe the feature deviation, and the stronger the constraint needs to be to retain the general rules. The weighted composition logic refers to the fusion of the main task prediction loss and the domain difference metric through dynamic weights. The weights are adaptively adjusted according to the level of domain difference. When the difference between the new working condition and the pre-training working condition is large, the weight of the domain difference metric is increased to strengthen the constraints of general rules; when the difference is small, the weight of the main task prediction loss is increased to prioritize the adaptation to the new working condition and achieve loss balance with difference perception.

[0072] S2.1c-4: With the goal of minimizing the domain adaptive loss function, the backpropagation algorithm is used to iteratively update only the weights of the remaining network layers that have not been frozen.

[0073] It's important to note that minimizing the domain adaptive loss function refers to adjusting the weights of the unfrozen layers to achieve the minimum value of the constructed domain adaptive loss function. The core objective is to balance the prediction accuracy under new conditions with the preservation of general patterns, ensuring that the model output closely resembles the actual labels of the new samples while avoiding a decrease in generalization ability due to feature distribution deviations from the pre-training conditions. The backpropagation algorithm, based on the principle of gradient descent, calculates the partial derivative of the adaptive loss function with respect to the weights of the unfrozen layers (the gradient) and iteratively adjusts the weight parameters along the negative gradient direction until the loss function gradually converges. This achieves a closed loop of error backtracking and parameter correction, ensuring that the weight update direction aligns with the loss minimization objective. Updating only the weights of the unfrozen layers strictly limits the iterative update scope to the unfrozen network layers, keeping the frozen layer weights fixed to avoid disrupting general contamination patterns.

[0074] The beneficial effects of this embodiment: This embodiment uses a partial least squares path model to analyze the coupling effect between relevant parameters and quantifies the static correlation between the resistance components. It continuously outputs real-time estimates of each resistance component for training, and inputs the sequence of real-time estimates into a pre-trained long short-term memory (LSTM) network model for prediction, obtaining the predicted total membrane resistance and predicted component resistances. Finally, based on the predicted total membrane resistance and / or predicted component resistances, the future fouling level of the ceramic membrane is determined. In this embodiment, the partial least squares path model effectively extracts the coupling relationship between different parameters, while the LSM network model effectively learns and updates the model, thereby improving the speed and accuracy of computation.

[0075] Specific application examples This specific application example uses "SiC ceramic membrane for industrial chemical wastewater treatment" as an application scenario to illustrate in detail the specific implementation process of the technical solution of this application. The construction of key models is based on relevant physical and chemical principles and engineering empirical formulas.

[0076] 1. Preliminary calibration and parameter setting Membrane inherent resistance determination: Deionized water was used as the test medium to measure membrane flux (J) under different transmembrane pressure differences (TMP). According to Darcy's law, the inherent resistance R of the ceramic membrane... mem Through formula R mem = TMP / (μ * J) is calculated, where μ is the dynamic viscosity of water. The R_mem of this SiC ceramic membrane was measured to be 5.2 × 10¹¹ m⁻¹. This value is a fixed parameter and remains constant throughout the membrane's lifespan.

[0077] Operating parameter range settings: The normal operating range set for this system is: TMP: 0.2 - 0.6 MPa, crossflow velocity (CFV): 1.0 - 2.5 m / s.

[0078] 2. Construction and Calculation of the Component Resistance Mathematical Model One of the cores of this method is to establish a mechanism-based component resistance calculation model.

[0079] a. Cake Layer Resistance (R cake ) Model Construct the formula R cake = k * (TMP n ) / (CFV m ): Formula Explanation and Principle Analysis: k: is the proportionality coefficient of the model, and its value is related to the concentration, size distribution, and specific resistance of suspended particles in the feed liquid, and is determined by fitting the filtration experimental data of a specific material system. TMP n : Transmembrane Pressure Difference (TMP) is the driving force for cake layer compression and formation. The exponent n (usually n≥1) is used to describe the non-linear relationship between pressure and the specific resistance of the cake layer. When n > 1, it indicates that as the pressure increases, the cake layer is compressed, the porosity decreases, resulting in a significant increase in the specific resistance, which is consistent with the theory of cake compressibility. CFV m : Cross-flow velocity (CFV) determines the shear stress on the membrane surface and has a scouring effect on the cake layer. The exponent m (usually 0<m<1) reflects that the inhibitory effect of shear stress on the cake layer thickness is non-linear. This power function form is a commonly used empirical model in fluid mechanics to describe shear-related processes. The coefficients k and exponents n, m are determined by experimental fitting to ensure the applicability of the model to a specific wastewater system.

[0080] FEA Correction: Based on the current CFV, calculate the shear stress distribution on the membrane surface through finite element analysis (FEA). For regions where the shear stress is lower than the average value (such as the feed end of the flow channel), introduce a correction coefficient greater than 1 (such as 1.1 to 1.3) to up-regulate the local R_cake to more realistically reflect the non-uniformity of the cake distribution.

[0081] b. Organic Matter Adsorption Resistance (R ads ) Model R ads = A * (I / |ζ|) * f(Δε) Formula Explanation and Principle Analysis: A: Proportional coefficient, related to the adsorption equilibrium constant of organic matter and membrane surface area, calibrated through adsorption experiments. I (Ionic Strength): According to DLVO theory, the higher the ionic strength of the solution, the stronger the compression effect on the colloidal double layer, thereby weakening the electrostatic repulsion between the membrane and charged organic pollutants and promoting adsorption. |ζ| (Absolute Value of Zeta Potential): Its value directly characterizes the electrostatic repulsion barrier between the membrane and the pollutant. The smaller |ζ| is, the weaker the repulsion. Proportional Relationship I / |ζ|: This combined term quantitatively reflects the regulation of electrostatic interactions by the chemical environment of the solution. Increasing I or decreasing |ζ| will lead to an enhancement of the adsorption driving force. f(Δε): A function of the change in dielectric constant of the membrane surface Δε (Δε = ε - ε0, where ε0 is the dielectric constant of the clean membrane). The adsorption of organic matter on the membrane surface will change the polarization properties and thickness of the interfacial layer, causing a change in the dielectric constant. f(Δε) serves as a dynamic calibration factor, feeding back the changes in physical properties caused by the adsorption process into the resistance calculation, realizing in-situ and dynamic sensing of the adsorption process.

[0082] c. Inorganic scaling resistance (R) scale ) Model Formula for construction: R scale = B * [Ca²⁺] * exp(C * T) Formula Explanation and Principle Analysis: B: Proportional coefficient, related to the crystal morphology of the scaling salt, film surface properties, etc. [Ca²⁺]: Calcium ion concentration is a necessary reactant for the formation of scale such as calcium carbonate. Under supersaturation, the scaling rate is usually positively correlated with the power of the reactant concentration. To simplify the model, the first power is taken as the basic relationship. exp(C * T): Temperature (T) has a significant impact on the crystallization process of inorganic salts. The crystallization rate constant usually follows the Arrhenius equation and has an exponential relationship with temperature. The exponential term exp(C * T) can effectively describe the accelerating effect of temperature increase on crystal nucleation and growth rate, and is a manifestation of chemical reaction kinetics in the scaling process. The coefficient C reflects the sensitivity of the scaling process to temperature.

[0083] 3. System Operation and Prediction Process ① Real-time data acquisition: The system continuously collects parameters such as TMP, CFV, J, I, ζ, ε, T, and [Ca²⁺].

[0084] ②Real-time resistance calculation: Substitute the real-time parameters into the above model to calculate R. cake , R ads , R scale And through FEA to R cake Make corrections. Then, calculate the total resistance R. total = R mem+ R cake + R ads + R scale .

[0085] ③ Time series prediction: R cake , R ads , R scale The real-time estimated value sequence (such as data from the past 4 hours) is input into a pre-trained Long Short-Term Memory (LSTM) network model. This LSTM model has learned the temporal patterns of resistance evolution and can output predicted resistance values ​​for a future time period (such as 2 hours).

[0086] Pollution level assessment and early warning: This involves predicting the R... total The system compares the predicted values ​​with a preset dynamic threshold range to determine the contamination level. Simultaneously, it identifies the most prevalent component, resistance, to determine the dominant contamination type. The system then outputs a comprehensive warning, such as "Contamination will develop into Level 3 within the next 2 hours, with inorganic scaling as the dominant type," and recommends corresponding cleaning strategies.

[0087] 4. Implementation of cross-condition migration When a significant change in influent water quality or operating conditions is detected (e.g., ionic strength I continuously deviates from the original operating range by more than 20%), the system triggers a transfer learning process: collect a small amount of sample data under the new operating conditions (e.g., 500 sets) → freeze the first few layers of the pre-trained LSTM model (responsible for learning general pollution time-series dynamics), and fine-tune the last few layers of the model only with the new data (responsible for learning specific mappings related to the new operating conditions) → in this way, the model can quickly adapt to the new operating conditions using a small amount of new data, significantly reducing the amount of data and time required for retraining. Example

[0088] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0089] Figure 3 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application.

[0090] See Figure 3 The electronic device 1000 includes a memory 1010 and a processor 1020.

[0091] The processor 1020 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0092] Memory 1010 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by the processor 1020 or other modules of the computer. Permanent storage devices may be read-write storage devices. Permanent storage devices may be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use mass storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices may be removable storage devices (e.g., floppy disks, optical drives). System memory may be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory may store some or all of the instructions and data required by the processor during operation. Furthermore, memory 1010 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some implementations, memory 1010 may include a removable storage device that is readable and / or writable, such as a laser disc (CD), a read-only digital multifunction optical disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-high density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.

[0093] The memory 1010 stores executable code, which, when processed by the processor 1020, can cause the processor 1020 to execute part or all of the methods described above.

[0094] The solution of this application has been described in detail above with reference to the accompanying drawings. In the above embodiments, the descriptions of each embodiment have different emphases; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. Those skilled in the art should also understand that the actions and modules involved in the specification are not necessarily essential to this application. Furthermore, it is understood that the steps in the method of this application's embodiments can be adjusted, combined, and deleted according to actual needs, and the modules in the device of this application's embodiments can be combined, divided, and deleted according to actual needs.

[0095] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.

[0096] Alternatively, this application may be implemented as a non-transitory machine-readable storage medium (or computer-readable storage medium, or machine-readable storage medium) storing executable code (or computer program, or computer instruction code) thereon, which, when executed by a processor of an electronic device (or electronic device, server, etc.), causes the processor to perform part or all of the steps of the methods described above according to this application.

[0097] Those skilled in the art will also understand that the various exemplary logic blocks, modules, circuits, and algorithm steps described in connection with the present application can be implemented as electronic hardware, computer software, or a combination of both.

[0098] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems and methods according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0099] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for predicting ceramic membrane fouling, characterized in that, include: S1: Real-time acquisition of relevant parameters of the ceramic membrane filtration process, including macroscopic operating parameters, microscopic interface parameters and feed liquid property parameters, wherein the microscopic interface parameters include at least the dielectric constant of the membrane surface and the Zeta potential of the membrane surface; S2: Establish a quantitative relationship between the relevant parameters and component resistance, wherein the component resistance includes at least one of filter cake layer resistance, organic matter adsorption resistance and inorganic scaling resistance; S3: Calculate the total membrane resistance based on the quantification relationship, and characterize the total membrane resistance as a function of the component resistance; S4: Predict the degree of ceramic membrane fouling based on the total membrane resistance and / or the component resistance.

2. The ceramic membrane fouling prediction method according to claim 1, characterized in that, The dielectric constant of the film surface is collected including: S1.1a: Microwaves are emitted to the membrane surface by microwave transmission and the phase change is detected; S1.1b: Based on the phase change, the dielectric constant value of the film surface is calculated; The Zeta potential on the membrane surface was collected, including: S1.2a: A weak electric field is applied to the feed liquid on the membrane surface using an in-situ sensor; S1.2b: Measure the potential difference generated when the feed liquid flows on the membrane surface; S1.2c: Based on the potential difference, the Zeta potential value of the membrane surface is calculated.

3. The method for predicting ceramic membrane fouling according to claim 1, characterized in that, The calculation of the filter cake layer resistance includes: S3.1a: Based on the current crossflow velocity, establish a model of the velocity field and shear stress distribution inside the ceramic membrane through finite element analysis; S3.1b: Based on the shear stress distribution model, determine the shear stress correction coefficients for different regions of the membrane surface; S3.1c: Using the correction coefficient, the initial value of the filter cake layer resistance calculated based on macroscopic parameters is regionally corrected.

4. The ceramic membrane fouling prediction method according to claim 1, characterized in that, S2 include: The coupling effect between the relevant parameters is analyzed by partial least squares path model, and the static correlation between the resistance components is quantified to continuously output the real-time estimated value sequence of each resistance component. S4 includes: S41: Input the real-time estimated value sequence into the pre-trained long short-term memory network model for prediction to obtain the predicted total membrane drag and the predicted component drag. S42: Determine the future fouling level of the ceramic membrane based on the predicted total membrane resistance and / or predicted component resistance.

5. The ceramic membrane fouling prediction method according to claim 4, characterized in that, When a change in operating condition is detected, the process after S2 and before S4 also includes: S2.1a: Acquire several new sample data collected under the changed operating conditions; S2.1b: Freeze the weights of the first few layers of the network representing the general contamination pattern in the Long Short-Term Memory network model; S2.1c: Using the new sample data, update and fine-tune the weights of the remaining network layers that have not been frozen to form a long short-term memory network model adapted to the new working conditions.

6. The ceramic membrane fouling prediction method according to claim 5, characterized in that, S2.1c includes: S2.1c-1: Input the new sample data and some sample data from the pre-training scenario into the basic prediction model, and extract their representations in the high-dimensional feature space respectively; S2.1c-2: Calculate the distribution difference between the new working condition sample data and the pre-trained working condition sample data in the high-dimensional feature space, and use it as a domain difference measure; S2.1c-3: Construct a domain-adaptive loss function, which is composed of a weighted average of the network's main task prediction loss and the domain difference metric; S2.1c-4: With the goal of minimizing the domain adaptive loss function, the weights of the remaining network layers that have not been frozen are iteratively updated only through the backpropagation algorithm.

7. The ceramic membrane fouling prediction method according to claim 1, characterized in that, S4 include: S51: Compare the total membrane resistance with a preset resistance threshold range to obtain a preliminary pollution level; S52: Identify the resistance type with the largest value or the highest proportion among the component resistances as the dominant contamination type; S53: Combining the preliminary pollution level with the dominant pollution type, determine the final pollution level.

8. The ceramic membrane fouling prediction method according to claim 7, characterized in that: The resistance threshold range is dynamically calibrated using a Bayesian update algorithm, including: S5.1a: Record historical early warning records and subsequent cleaning effect data; S5.1b: If the pollution level rapidly escalates after multiple low-level warnings, the upper limit of the resistance threshold corresponding to the low level will be lowered through Bayesian updates. S5.1c: If the cleaning effect is not significant after multiple high-level warnings, the lower limit of the resistance threshold corresponding to the high level will be increased through Bayesian updates.

9. An electronic device, characterized in that, include: processor; as well as A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform a ceramic membrane fouling prediction method as described in any one of claims 1-8.

10. A non-transitory machine-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform a ceramic membrane fouling prediction method as described in any one of claims 1-8.

Images