Industrial wastewater membrane process intelligent optimization method and system based on multi-dimensional data modeling

Abstract

The invention relates to the field of wastewater treatment, and provides an industrial wastewater membrane process intelligent optimization method and system based on multi-dimensional data modeling. The method comprises the following steps: acquiring wastewater quality parameters of a target scene and a production line equipment state in real time through a multi-dimensional sensor network, and constructing a digital twinborn model of a wastewater treatment system in the target scene; the membrane surface pollution type of each node in the wastewater treatment system is identified through a digital twin model, and the inner membrane pollution evolution condition in the future preset duration is predicted; a multi-strategy evolutionary reinforcement learning algorithm is adopted, the wastewater quality parameters, the production line equipment state and the membrane pollution prediction result are combined, multi-target dynamic optimization is conducted on a wastewater treatment system through a multi-agent collaborative decision-making algorithm, an optimization strategy is generated based on the multi-target dynamic optimization result, and membrane process operation parameters and production control parameters are dynamically adjusted. According to the method, real-time intelligent adjustment of membrane process operation can be realized, the stability of membrane process operation is guaranteed, and the treatment efficiency of membrane process operation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wastewater treatment, and more particularly, the present application relates to an industrial wastewater membrane process intelligent optimization method and system based on multi-dimensional data modeling. BACKGROUND

[0002] With the rapid development of the industrial field, the problem of industrial wastewater treatment, the industrial wastewater membrane process can be applied to the fields of high-concentration industrial wastewater (such as wastewater generated by the printing and dyeing industry) treatment, seawater desalination, etc. due to its high separation efficiency and strong resource recycling capability, etc. and has become one of the key technologies to realize wastewater purification, pollutant resource disposal and carbon emission reduction.

[0003] However, in the existing industrial wastewater membrane process, on the one hand, membrane pollution phenomenon is common, and organic matter, colloid and other substances in wastewater are easy to adhere to the membrane surface or block the membrane holes, which causes the membrane flux to decrease and the service life to shorten. The traditional membrane cleaning and operation and maintenance mode relying on manual experience is difficult to dynamically and efficiently cope with the evolution of membrane pollution under complex water quality, which seriously affects the stability of the process operation. On the other hand, the design of the microstructure of the membrane material, the optimization of the winding process parameters and the production line control, etc. have long relied on manual experience for adjustment, lack of systematic modeling analysis of wastewater characteristics, and cannot adapt to the changing requirements of complex and variable wastewater quality and working conditions, resulting in that the anti-pollution performance, flux and other performances of the membrane assembly cannot be fully utilized, which further causes the process energy consumption to be high and the treatment efficiency to be unstable, and restricts the further breakthrough of the membrane process in the energy saving and emission reduction target. In summary, it is urgent to design an industrial wastewater membrane process intelligent optimization scheme to solve at least one of the above technical problems. SUMMARY

[0004] In this context, the embodiments of the present application expect to provide an industrial wastewater membrane process intelligent optimization method and system based on multi-dimensional data modeling, which can realize real-time intelligent adjustment of the operation of the industrial wastewater membrane process, guarantee the stability of the operation of the membrane process, and improve the treatment efficiency of the operation of the membrane process.

[0005] In a first aspect of the embodiments of the present application, an industrial wastewater membrane process intelligent optimization method based on multi-dimensional data modeling is provided, which comprises: acquiring wastewater quality parameters and production line equipment states of a target scene in real time through a multi-dimensional sensor network; the production line equipment states comprise production line control parameters, membrane process operation parameters, and real-time image data of a microstructure of a membrane material at a preset production node; based on the wastewater quality parameters and the production line equipment states, a digital twin model of a wastewater treatment system under the target scene is constructed; through the digital twin model, a membrane surface pollution type of each node in the wastewater treatment system is identified and a membrane pollution evolution situation in a future preset time period is predicted, to obtain a membrane pollution prediction result; a scenario memory driven multi-strategy evolutionary reinforcement learning algorithm is adopted, combined with the wastewater quality parameters, the production line equipment states, and the membrane pollution prediction result, a multi-agent collaborative decision algorithm is used to perform multi-objective dynamic optimization on the membrane process operation parameters and the production control parameters of the wastewater treatment system; a multi-agent collaborative structure in the multi-agent collaborative decision algorithm is constructed based on the optimization objectives of maximizing membrane flux, minimizing energy consumption, and prolonging chemical cleaning period; and an optimization strategy of the wastewater treatment system is generated based on the multi-objective dynamic optimization result, to dynamically adjust the membrane process operation parameters and the production control parameters in the wastewater treatment system.

[0006] In a second aspect of the embodiments of the present application, an industrial wastewater membrane process intelligent optimization system based on multi-dimensional data modeling is provided, which comprises the following modules: a data acquisition module, a modeling module, an optimization decision module, and an automatic control module; wherein the data acquisition module is configured to acquire wastewater quality parameters and production line equipment states of a target scene in real time through a multi-dimensional sensor network; the production line equipment states comprise production line control parameters, membrane process operation parameters, and real-time image data of a microstructure of a membrane material at a preset production node; the modeling module is configured to construct a digital twin model of a wastewater treatment system under the target scene based on the wastewater quality parameters and the production line equipment states; the optimization decision module is configured to identify a membrane surface pollution type of each node in the wastewater treatment system through the digital twin model and predict a membrane pollution evolution situation in a future preset time period, to obtain a membrane pollution prediction result; a scenario memory driven multi-strategy evolutionary reinforcement learning algorithm is adopted, combined with the wastewater quality parameters, the production line equipment states, and the membrane pollution prediction result, a multi-agent collaborative decision algorithm is used to perform multi-objective dynamic optimization on the membrane process operation parameters and the production control parameters of the wastewater treatment system; a multi-agent collaborative structure in the multi-agent collaborative decision algorithm is constructed based on the optimization objectives of maximizing membrane flux, minimizing energy consumption, and prolonging chemical cleaning period; and the automatic control module is configured to generate an optimization strategy of the wastewater treatment system based on the multi-objective dynamic optimization result, to dynamically adjust the membrane process operation parameters and the production control parameters in the wastewater treatment system.

[0007] An industrial wastewater membrane process intelligent optimization method and system based on multi-dimensional data modeling according to an embodiment of the present application. The wastewater quality full-dimensional parameters and production line equipment states of the target scene are collected in real time through a multi-dimensional sensor network, the membrane surface pollution type is accurately identified in combination with a digital twin model, the future membrane pollution evolution trend is predicted, and then relying on a scenario memory driven multi-strategy evolutionary reinforcement learning algorithm, the cleaning strategy agent in the multi-agent collaborative decision-making architecture is linked, the adaptive parameters of the pulse cleaning mode are dynamically adjusted, the dynamic adaptation of membrane cleaning and operation to membrane pollution evolution under complex water quality is realized, and then the stable operation of the industrial wastewater membrane process is ensured, the membrane assembly pollution resistance and flux are improved, and the service life is prolonged. At the same time, the scenario memory driven multi-strategy evolutionary reinforcement learning algorithm is used in combination with the multi-agent collaborative decision-making architecture to perform multi-objective dynamic optimization on the membrane process operation parameters and production control parameters, and the parameters are corrected in real time by means of the virtual-real synchronous characteristics of the digital twin model, the accurate matching of the membrane process parameters and complex water quality and working conditions is realized, the process performance is improved, the energy consumption is reduced, and the treatment efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 A flowchart of an industrial wastewater membrane process intelligent optimization method based on multi-dimensional data modeling is shown in the present application. Figure 2 A structural diagram of an industrial wastewater membrane process intelligent optimization system based on multi-dimensional data modeling is shown in the present application. DETAILED DESCRIPTION

[0009] Reference will now be made to Figure 1 , Figure 1 A flowchart of an industrial wastewater membrane process intelligent optimization method based on multi-dimensional data modeling is provided in an embodiment of the present application.

[0010] To solve the above at least one technical problem, the embodiment of the present application provides an industrial wastewater membrane process intelligent optimization method and system based on multi-dimensional data modeling. Specifically, in the embodiment of the present application, the multi-dimensional sensor network is used to collect the wastewater quality full-dimensional parameters and the production line equipment state (including the membrane material microstructure image) in the target scene in real time, the digital twin model is combined to accurately identify the membrane surface pollution type and predict the membrane pollution evolution trend in the future preset time, and the scenario memory driven multi-strategy evolutionary reinforcement learning algorithm is combined to drive the cleaning strategy agent in the multi-agent collaborative decision architecture to dynamically adjust the adaptive parameters (such as high-pressure backflush strength, reagent concentration, cleaning cycle, etc.) of the pulse cleaning mode, so that the membrane cleaning and operation mode is adapted to the membrane pollution evolution process under complex water quality, thereby ensuring the stability of the industrial wastewater membrane process operation, improving the anti-pollution and flux of the membrane assembly, and prolonging the service life of the membrane assembly. Moreover, by using the scenario memory driven multi-strategy evolutionary reinforcement learning algorithm and combining the multi-agent collaborative decision architecture, the multi-objective dynamic optimization of the membrane process operation parameters (such as the water inlet flow rate) and the production control parameters (such as the operating pressure) is realized, and the virtual-real synchronous characteristics of the digital twin model are used to realize real-time feedback of the optimization effect and correction of the parameters, so that the membrane process parameters are accurately matched with the complex and variable wastewater quality and working condition requirements, thereby improving the performance of the industrial wastewater membrane process under the complex and variable wastewater quality and working condition change requirements, reducing the energy consumption of the industrial wastewater membrane process, and improving the treatment efficiency of the industrial wastewater membrane process.

[0011] Figure 1 An embodiment of the present application provides a flow of an industrial wastewater membrane process intelligent optimization method based on multi-dimensional data modeling, which includes: In step S101, the wastewater quality parameters and the production line equipment state of the target scene are obtained in real time by the multi-dimensional sensor network. The principle of step S101 is to solve the limitation of traditional single-point static data collection through multi-type full-process sensing coverage, to capture the dynamic changes of wastewater from entering the system to flowing through the membrane assembly, and the linkage feedback of the production line from the control instruction execution to the actual running state of the membrane assembly in real time, especially for the image collection of the microstructure of the membrane material, which can capture the initial morphological changes that are difficult to identify by the naked eye, and provide direct basis for subsequent pollution precursor identification.

[0012] Exemplarily, the implementation of step S101 needs to plan a deployment scheme of the multi-dimensional sensor network for the target scene (such as a printing and dyeing wastewater treatment or seawater desalination system) first, and various functional sensors are arranged at key links of the wastewater treatment system. For example, water quality sensors are installed at the wastewater inlet pipeline, the inlet and outlet of the membrane assembly, the production line control unit and the membrane assembly body are integrated with working condition sensors, and high-magnification industrial cameras and other image acquisition devices are configured at the preset membrane material monitoring nodes. At the same time, all sensors and the back-end data processing unit are established in real-time communication link through the Internet of Things technology, so as to ensure that the collected data can be transmitted to the data storage and preprocessing module synchronously. In this way, the data obtained can have real-time, comprehensiveness and correlation, avoiding misjudgment of the water quality and equipment state due to data loss or lag, providing real and continuous basic data support for subsequent construction of a digital twin model, and helping to capture the microscopic changes of the membrane material in advance, which is helpful for subsequent pollution prediction and pre-adjustment of operation and maintenance strategies, and reduces the risk of sudden failure.

[0013] In the embodiment of the application, the production line equipment state includes production line control parameters, membrane process running parameters, and real-time image data of the microscopic structure of the membrane material at the preset production node.

[0014] From the data content, the wastewater quality parameters cover indexes reflecting the pollution degree and physical and chemical characteristics of the wastewater. Further, the data accuracy of the wastewater quality parameters determines the adaptability of the membrane process treatment scheme. From the pollution load dimension, it includes chemical oxygen demand (COD) and biological oxygen demand (BOD), the former is used to measure the total amount of organic matter that can be chemically oxidized and decomposed in wastewater, and the latter reflects the content of biodegradable organic matter, and the values of the two are directly related to the risk of organic matter adhesion on the membrane surface. From the dimension of inorganic pollutants, it covers salinity (such as chloride ion and sodium ion concentration), heavy metal ion content (such as copper ion in printing and dyeing wastewater and nickel ion in electroplating wastewater), and high salinity is easy to cause the osmotic pressure of the membrane assembly to rise, increasing the energy consumption, and the heavy metal ions may react with the membrane material, affecting the retention performance of the membrane. From the dimensions of physical and acid-base characteristics, it includes temperature, pH value and suspended matter concentration, temperature change will change the viscosity of the wastewater, and then affect the membrane flux, pH value deviating from the tolerance range of the membrane material (such as most organic membranes suitable for pH 2-11) will accelerate the membrane aging, and the suspended matter is easy to block the membrane holes and cause initial pollution. These parameters together constitute the water quality basic image of the membrane process treatment, and provide a basis for subsequent judgment of the membrane pollution inducement and adjustment of the operation strategy.

[0015] The production line control parameters include operation instruction related parameters for adjusting the operation of the production line. The production line equipment state is a key information set reflecting the operation dynamics of the membrane process system and the health status of the membrane module, which is divided into three categories of core data. The first is the production line control parameter, i.e., the operation instruction type parameter set by humans, including the operation pressure (the pressure difference between the membrane module inlet and outlet as the control reference), the water inlet flow rate (the flow control value of the wastewater into the membrane module), and the cleaning agent injection amount (the dosing ratio of acid, alkali or oxidizing agent during chemical cleaning). This type of parameter is an instruction signal for guiding the operation of the membrane process, which needs to be adjusted in real time according to the changes in water quality to avoid membrane damage. The second is the membrane process operation parameter, which is the feedback type data of the actual operation of the equipment, including the transmembrane pressure difference (TMP, a core indicator reflecting the degree of membrane hole blockage, and an increase in the difference usually means that the membrane pollution is aggravated), the real-time flux (the amount of water produced through the membrane per unit time, which directly reflects the filtration efficiency of the membrane), and the concentrated water discharge amount (the discharge amount of wastewater after the membrane retains pollutants, which is linked with the salinity and pollutant concentration). This type of data is a barometer for judging the current working state of the membrane and can timely warn the performance degradation of the membrane. The third is the real-time image data of the microstructure of the membrane material, which is collected by a high-power industrial camera or an electron microscope, showing the pore size distribution, defects (such as pinholes and scratches) and initial pollution marks (such as the attachment of small colloidal particles and the formation of organic film) on the membrane surface. This type of micro data can find subtle changes in the membrane earlier than the macro operation parameters, providing micro support for pollution precursor identification and membrane life assessment. The three types of data are interrelated and together constitute a complete state picture of the production line equipment, which is the core data input for subsequent construction of digital twin models and realization of intelligent optimization.

[0016] The membrane process operation parameter is a real-time indicator directly reflecting the working state of the membrane module. Among them, the real-time image data of the microstructure of the membrane material records the micro-morphological characteristics such as the pore size distribution, defect state and initial pollution marks on the membrane surface.

[0017] In step S102, a digital twin model of the wastewater treatment system under the target scenario is constructed based on the wastewater quality parameters and the production line equipment state.

[0018] In step S103, the membrane surface pollution type of each node in the wastewater treatment system is identified and the membrane pollution evolution in a future preset time period is predicted through the digital twin model, to obtain a membrane pollution prediction result.

[0019] In the embodiments of the present application, the digital twin model is a dynamic system for full-dimensional virtual mapping of the wastewater treatment system in the target scene, and is not a simple three-dimensional geometric replication, but a comprehensive body integrating physical laws, real-time data and simulation algorithms. The core is to realize virtual-real synchronization, real-time interaction and simulation prediction. The digital twin model not only restores the structure and spatial layout of physical devices such as membrane components, pipelines and sensors through geometric models, but also links real-time with sensors and controllers of the real system through data interfaces, inputs dynamic data such as wastewater quality parameters and production line equipment states into the virtual model, and embeds physical field simulation modules conforming to fluid mechanics and membrane separation principles, so that the virtual model can reproduce the physical processes of wastewater flow, pollutant migration and membrane pollution formation in the real system, and even simulate the response changes of the system under different parameter adjustments, becoming a digital mirror of the real wastewater treatment system.

[0020] In step S102, the construction of the digital twin model needs to start from the accurate mapping and dynamic correlation of the physical system. First, the physical information of the membrane components in the wastewater treatment system, such as the size, pipeline direction and sensor installation position, is collected through three-dimensional scanning technology to generate a geometric model that matches the real equipment 1:1, ensuring that the structure layout, connection relationship of the equipment in the virtual space are completely consistent with the real system. Then, a data interaction channel is established to continuously input the wastewater quality parameters (such as COD, pH, salinity) and production line equipment states (such as operating pressure, transmembrane pressure difference, membrane micro image) collected in step S101 into the virtual model through the Internet of Things protocol, realizing real-time binding of physical data and virtual model, and dynamically updating the state of the virtual model with the running of the real system. Finally, a physical field simulation module is embedded, such as a simulation unit based on computational fluid dynamics principles, to simulate the flow trajectory of wastewater in the membrane component, the shear force action and deposition process of pollutants on the membrane surface, and to construct the mathematical relationship between pollutants and membrane material interaction (such as organic matter adsorption and colloid clogging dynamics model) based on the physical and chemical laws of membrane separation, so that the virtual model can not only replicate the current state of the real system, but also deduce the trend of system changes based on physical laws, laying a foundation for subsequent pollution identification and prediction. Thus, through this step, a virtual platform that can reflect the state of the real system in real time and can be simulated without risk is formed, providing an intuitive and accurate research carrier for subsequent pollution analysis, and avoiding membrane damage or process interruption that may be caused by direct debugging in the real system.

[0021] After the digital twin model is completed, step S103 realizes the identification and future evolution prediction of membrane surface pollution through the digital twin model. First, the real-time updated wastewater quality parameters and production line equipment state data (including membrane microstructure images) are continuously input into the digital twin model, and the digital twin model automatically aligns these multi-source data in space and time. For example, the COD data of the membrane module inlet at a certain time, the transmembrane pressure difference data of the module at the same time, and the membrane microstructure image data of the same region are associated with the corresponding membrane module node in the virtual model, ensuring that the data and the spatial position and time node of the virtual equipment are accurately matched; then the model calls the built-in feature analysis algorithm, combines the pollutant motion trajectory and shear force distribution data output by the physical field simulation module, and comprehensively analyzes the pollutant form in the membrane microstructure image, the transmembrane pressure difference trend, and the pollutant type (such as organic matter and colloid) in the water quality parameter, such as judging the membrane surface pollution type of the node by identifying the form characteristics (such as flocculent for organic matter and granular for colloid) of the membrane surface attachments and combining the simulated pollutant deposition rate. On this basis, the digital twin model simulates the expansion process of the pollution type on the membrane surface within a preset time (such as 72 hours) based on the historical operation data (such as the evolution period of membrane pollution under the same water quality conditions in the past) and the physical field simulation law (such as the correlation between pollutant concentration and deposition rate). For example, the proportion of membrane pore blockage, the rising amplitude of transmembrane pressure difference, and the attenuation rate of flux in different time periods are predicted, and the pollution type, flux attenuation trend, and pollution area expansion range of each membrane module node in the whole system are output based on the pollution state of different nodes in the virtual model, forming a complete membrane pollution prediction result. The principle of this step is to use the virtual-real synchronization feature of the digital twin model to integrate scattered multi-source data into system-level information with spatial correlation and time continuity, and then combine physical simulation with historical data to realize the retrospective analysis and forward-looking prediction of the pollution process. Thus, this step solves the limitations of traditional pollution identification, which relies on manual observation and can only find the pollution that has occurred, and can accurately identify the pollution type of each node and predict the evolution direction of future pollution, providing an advance for subsequent development of targeted cleaning strategies and adjustment of operating parameters, avoiding membrane module damage or process shutdown due to pollution outbreaks, and ensuring the stability of the wastewater treatment system.

[0022] As an optional embodiment, in step S102, a geometric model of the membrane module and the pipeline in the target scene is established by three-dimensional scanning, and dynamic binding of the geometric model is performed in combination with real-time collected wastewater quality parameters and production line equipment states to generate a virtual mapping system with virtual-real synchronization characteristics with the target scene, forming a digital twin model of the wastewater treatment system. Further, a computational fluid dynamics module is embedded in the digital twin model to simulate the shear force distribution state and deposition motion path of the pollutants on the membrane surface, and the physical field of the membrane pollution process is restored in the digital twin model.

[0023] Specifically, first, for the membrane assembly and pipeline system in the target scene, an appropriate three-dimensional scanning device is selected, the scanning range is determined according to the size specifications of the membrane assembly, the layout of the pipeline and the connection mode, and it is ensured to cover the whole process core structure from the water inlet end to the water outlet end, including the shape of the membrane assembly shell, the arrangement density and spacing of the internal membrane filaments, the diameter and bending angle of the pipeline and other details, through multi-angle scanning and data splicing, the scanning blind area caused by equipment shielding is eliminated, and an initial geometric model that can accurately reproduce the structural characteristics of the physical equipment is generated. Then, a real-time data interaction interface is built, and the wastewater quality parameters collected in step S101 (such as chemical oxygen demand reflecting the degree of pollution and pH value reflecting the acid-base characteristics) and the production line equipment state (such as the real-time operating pressure of the membrane assembly and the water inlet flow rate) are bound to the corresponding positions of the virtual geometric model according to the spatial logic of the physical system. For example, the real-time water quality data of the membrane assembly water inlet is associated with the virtual node of the water inlet in the model, and the real-time transmembrane pressure difference of the membrane assembly is bound to the corresponding membrane assembly body in the model, and through the real-time data transmission protocol, the data changes of the physical system are synchronized to the virtual model within 1 second, thereby generating a virtual mapping system with virtual-real synchronization characteristics, i.e. the preliminary wastewater treatment system digital twin model. On this basis, a computational fluid dynamics module is further embedded in the digital twin model, first, according to the physical properties of the real wastewater (such as density, viscosity) and the microstructure parameters of the membrane assembly (such as membrane filament spacing, pore size), the matching simulation initial conditions are set in the module, then through the fluid motion algorithm built in the module, the flow trajectory of wastewater in the membrane assembly is simulated, the shear force size and distribution difference of pollutant particles in different regions of the membrane surface are calculated, and the complete motion path of the pollutant from the water body to the membrane surface and gradually deposited to form a pollution layer is tracked, and finally these simulation results are presented in the virtual model in a visual form. For example, gradient color is used to mark the strong and weak areas of shear force, and dynamic lines are used to show the deposition trajectory of pollutants, so that the physical field of the membrane pollution process is completely restored in the digital twin model. This process is essentially to reproduce the physical structure through three-dimensional scanning, to realize the real-time linkage of data and structure through dynamic binding, and to simulate the physical process through the computational fluid dynamics module, and finally to build a digital twin model with structural authenticity, data real-time and physical logic. The effect is that the virtual model can not only intuitively present the appearance and real-time data of the real system, but also clearly simulate the motion of pollutants and the formation process of membrane pollution that cannot be seen by the naked eye in the real system, providing analysis basis conforming to the physical law for subsequent accurate identification of pollution types and prediction of pollution evolution, avoiding judgment deviation caused by insufficient understanding of internal physical processes.

[0024] For example, in the process of establishing the microstructure mapping, first select a representative membrane material sample from the membrane assembly of the current production line, ensure that the microstructure of the sample is consistent with the actual running membrane assembly, fix the sample on the stage of the electron microscope, adjust the magnification and focusing accuracy of the microscope until the fine structure of the membrane surface can be clearly captured, then record the pore size, the proportion of different pore sizes and the uniformity of the pore size distribution at different positions on the membrane surface through regional scanning, and mark the defects on the membrane surface such as small pinholes, local scratches or structure damage to form complete membrane microstructure data. Then, the correlation mechanism between microstructure data and real-time operating parameters is established, and the pore size distribution and defect topology information of the membrane surface obtained by scanning are correlated with the real-time transmembrane pressure difference and water inlet flow rate collected in step S101. For example, the pore size data of a certain area on the membrane surface is bound to the real-time transmembrane pressure difference of the corresponding area. When the transmembrane pressure difference of the area in the physical system abnormally increases, the membrane microstructure data of the corresponding area in the virtual model will be associated with this change synchronously, which can help to determine whether the increase of the transmembrane pressure difference is related to the pore blockage of the area. The principle of this operation is to capture the microstructure of the membrane material by using the high-resolution imaging capability of the electron microscope, and to realize the dynamic linkage between the microstructure and the macro operating parameters through data correlation. The effect is that the digital twin model can reflect the real state of the membrane at the micro level, such as when the pore size of a certain area on the membrane surface is reduced due to blockage by pollutants, combined with the change of the real-time transmembrane pressure difference, the pollution precursor of the area can be identified in advance, which provides micro basis for subsequent accurate judgment of pollution type (such as pore blockage type pollution) and formulation of targeted cleaning strategy (such as local intensive cleaning of the blocked area), avoiding the problem of pollution cause misjudgment or incomplete cleaning caused by relying only on macro parameter judgment.

[0025] Further optionally, the above-mentioned virtual mapping system, as the core carrier of the digital twin model of the wastewater treatment system, has multi-dimensional functions to support accurate replication, real-time monitoring and deep analysis of the real system. First, it has the functions of accurate replication and real-time visualization of physical structure, which can present the spatial layout and structural details of the membrane assembly, pipeline and sensor based on three-dimensional scanning data, from the arrangement of the membrane filaments inside the membrane assembly, the bending angle of the pipeline, to the connection logic of the equipment in the entire wastewater treatment process, all are presented in a 1:1 virtual form, and the real-time collected wastewater quality parameters (such as COD, pH) and production line equipment status (such as operating pressure, transmembrane pressure difference) are superimposed in the form of dynamic annotation on the corresponding virtual device position, so that the operator can intuitively master the real-time operating data of each node of the system without on-site inspection, and clearly understand the overall working condition.

[0026] Secondly, the virtual mapping system has the functions of real-time data synchronization and dynamic association. Through the Internet of Things interface and the sensors and controllers of the real system, it can synchronize the dynamic changes of water quality and equipment state in the real system to the virtual model within seconds. All data are not presented in isolation, but are deeply bound to the virtual physical structure. For example, the salinity data at the inlet of the membrane module are directly associated with the corresponding node in the virtual model. When the salinity increases, the data label of the corresponding node in the virtual model is updated in real time, and the real-time flux changes of the downstream membrane module of the node can be viewed simultaneously, helping the operator to quickly establish the association cognition of "water quality change-equipment response" and avoid analysis faults caused by data fragmentation.

[0027] Furthermore, the virtual mapping system has the functions of physical field simulation and visualization. Based on the embedded computational fluid dynamics module, it can simulate the physical processes that cannot be seen by the naked eye in the real system. For example, it can simulate the flow trajectory of wastewater in the membrane module and show the distribution difference of water flow velocity through dynamic lines. It can simulate the force state of pollutants on the membrane surface and label the strong and weak areas of shear force with gradient colors to clearly present the high-risk areas where shear force is weak and pollutants are easy to deposit. At the same time, it can track the complete path of pollutants from water to membrane surface and gradually deposit to form a pollution layer, and restore the formation process of membrane pollution in a dynamic evolution form, so that the operator can intuitively understand the physical mechanism of pollution occurrence instead of relying on macro data inference.

[0028] In addition, the virtual mapping system also has the functions of state monitoring and abnormal warning. Based on the preset normal parameter range, it can continuously monitor the water quality, equipment state and microstructure data of each node in the virtual model. When the transmembrane pressure difference of a membrane module exceeds the normal threshold, or the pore size in the membrane microstructure mapping abnormally shrinks or the defect expands, the virtual model will automatically mark an abnormal warning at the corresponding node and display the recent water quality change trend and the simulation results of pollutant deposition associated with the node to help the operator quickly locate the abnormal cause (such as whether the pressure difference is caused by colloid blockage) and intervene in advance to avoid the expansion of the fault. At the same time, the system also supports the integration and tracing of historical data, which can retrieve the state data of the virtual model in any time period and compare the water quality, equipment operation and pollution in different periods to provide data support for analyzing pollution rules and optimizing operation strategies.

[0029] In the foregoing optional embodiment, one optional embodiment of step S103 is to extract real-time physical field simulation data from the digital twin model, the real-time physical field simulation data including a shear force distribution matrix of the pollutants on the membrane surface and deposition path trajectory data. Then, the real-time physical field simulation data, the real-time collected wastewater quality parameters, the operating pressure and the cleaning cycle in the production line control parameters, the transmembrane pressure difference and the real-time flux in the membrane process operation parameters, and the membrane surface defect image are aligned and fused in the time-space dimension to construct a multi-dimensional wastewater feature containing physical field characteristics. Next, a hybrid model of a convolutional neural network (CNN) and a long short-term memory network (LSTM) is used to identify the membrane surface pollution type based on the multi-dimensional wastewater feature. The membrane surface pollution type includes organic matter adhesion, colloidal blockage, and / or composite pollution. In combination with the identified membrane surface pollution type, the real-time physical field simulation data and the historical membrane process operation data, the hybrid model is used to predict the membrane flux decay rate and the pollution area expansion range in a future preset time period. The pollution area expansion range is represented as a membrane surface pollutant area ratio and a three-dimensional space distribution coordinate corresponding to a predicted coverage range of the pollutant. Finally, the membrane surface pollution type, the membrane flux decay rate and the pollution area expansion range are taken as the membrane pollution prediction result.

[0030] Specifically, in the foregoing optional embodiment, first, real-time physical field simulation data output by the embedded computational fluid dynamics module in the digital twin model constructed in step S102 is called. The data is generated by the module based on the physical properties (such as viscosity and density) of the real wastewater and the membrane assembly structure parameters, and specifically includes the distribution of the shear force on the membrane surface in different areas (the shear force at different positions is recorded in the form of a matrix), and the motion trajectory data of the pollutants in the process of migrating from the water body to the membrane surface and gradually depositing (the spatial position change of the pollutants at different time points is recorded). Then, the real-time physical field simulation data and other source data are aligned and fused in the time-space dimension. A uniform time stamp and a space coordinate label are added to all data (including the real-time collected wastewater quality parameters, the operating pressure and the cleaning cycle in the production line control parameters, the transmembrane pressure difference and the real-time flux in the membrane process operation parameters, and the membrane surface defect image), that is, the time stamp corresponds to the accurate time of data collection or generation, and the space coordinate corresponds to the specific node of the membrane assembly (such as the inlet end membrane wire and the middle region membrane wire), then different types of data are matched and integrated according to the time stamp and the space coordinate, for example, the shear force data of a certain time and a certain membrane assembly node are associated with the COD value, the operating pressure, and the membrane defect image at the same time and the same node, and finally a multi-dimensional wastewater feature containing numerical parameters, image information and physical field characteristics is constructed, ensuring that subsequent model analysis can be based on complete and time-space consistent information.

[0031] Subsequently, a hybrid model composed of a convolutional neural network and a long short-term memory network is used to process multi-dimensional wastewater characteristics to identify membrane surface pollution types. The membrane surface defect image in the multi-dimensional characteristics and the shear force distribution matrix (as two-dimensional data) are input into the convolutional neural network, which is good at extracting spatial features, to capture the morphological features of the pollutants (such as flocculent and granular) and the abnormal areas of the shear force distribution (such as the characteristics of the weak shear force deposition area). At the same time, the time-series data in the multi-dimensional characteristics (such as the transmembrane pressure difference and the real-time flux change curve over time) are input into the long short-term memory network, which relies on its processing capability for time-series information to mine the trend features of parameter changes (such as whether the pressure difference is slowly rising or rapidly rising). Then, the spatial features and time features extracted by the two networks are fused and input into the subsequent classification module, combined with the preset pollution type judgment rules (such as flocculent morphology + slow pressure difference rise corresponds to organic matter attachment, granular morphology + rapid pressure difference rise corresponds to colloid blockage, and both morphologies and trends coexist correspond to composite pollution), and finally the pollution type of the membrane surface is identified.

[0032] After the pollution type is identified, the hybrid model is further used to predict the membrane flux decay rate and the pollution area expansion range in the future preset time. The identified pollution type, real-time physical field simulation data (such as the pollution expansion speed area determined by the shear force distribution), and historical membrane process operation data (such as the flux decay law and pollution expansion speed under the same water quality and the same pollution type in the past) are input into the model. The model will first learn the evolution law of different pollution types under different physical field conditions based on historical data (such as the expansion speed of colloid blockage in the low shear force area is several times that in the high shear force area), and then adjust the prediction logic based on the current real-time data (such as the current shear force distribution and the initial deposition amount of pollutants), and finally output the decay trend of the membrane flux (i.e. flux decay rate) and the specific changes of the pollution area in the future preset time. Not only the overall degree of pollution is reflected by the area ratio of membrane surface pollutants, but also the specific location and diffusion range of pollution in different nodes of the membrane module are marked by three-dimensional spatial distribution coordinates, such as the pollution area of the left area of the inlet of a certain membrane module will expand to 20% in the next 3 days.

[0033] Thus, the fragmentation problem of multi-source data is eliminated through spatio-temporal alignment fusion, providing a comprehensive and consistent analysis basis for the hybrid model. The hybrid architecture of convolutional neural network and long short-term memory network is used to analyze both spatial morphological features and temporal evolution trends, achieving accurate identification of pollution types. Combined with the internal laws reflected by physical field simulation data and the experience laws accumulated from historical data, the prediction results not only conform to the physical logic but also fit the actual operation situation. Thus, the problem that the traditional pollution identification relies on manual observation and can only judge the surface phenomenon is solved. Not only can the pollution types be accurately distinguished, but also the development trend and specific impact range of pollution can be predicted in advance, providing accurate basis for subsequent targeted cleaning strategies (such as selecting oxidizing agent cleaning for organic matter attachment and strengthening high-pressure backflush for colloidal blockage) and adjusting operation parameters (such as increasing local water inflow velocity to enhance shear force in areas where pollution expands quickly), avoiding membrane damage aggravation and energy consumption increase caused by untimely response or improper measures, and ensuring the stability and economy of membrane process operation.

[0034] Further optionally, the hybrid network at least includes an identification layer and a prediction layer. Based on this structure, in step S103, a hybrid model of convolutional neural network CNN and long short-term memory network LSTM is used to identify the membrane surface pollution type based on multi-dimensional wastewater characteristics. That is, in the identification layer, the convolutional neural network CNN embedded with physical field prior knowledge is used to jointly extract features from the membrane surface defect image and the shear force distribution matrix in the multi-dimensional wastewater characteristics; the long short-term memory network LSTM is used to dynamically associate and analyze the time-series operation parameters and deposition path data in the joint feature extraction results, and the membrane surface pollution type is identified.

[0035] Then, based on the foregoing structure, in step S103, the membrane flux decay rate and the pollution area expansion range in the future preset time period are predicted by the hybrid model based on the identified membrane surface pollution type, the real-time physical field simulation data, and the historical membrane process operation data. That is, a spatio-temporal attention Transformer network that integrates the evolution constraint of physical field is introduced in the prediction layer; based on the virtual-real synchronization characteristics of the digital twin model, the shear force change gradient and the deposition path expansion rate that conform to the physical field evolution law are extracted from the real-time physical field simulation data based on the identified membrane surface pollution type, the extracted shear force change gradient and deposition path expansion rate are weighted and associated with the historical pollution evolution characteristics in the historical membrane process operation data, and a pollution diffusion constraint term derived from the CFD physical model is embedded in the network loss function; the membrane flux decay rate and the pollution area expansion range in the future preset time period are predicted by the spatio-temporal attention Transformer network.

[0036] In the implementation process of the mixed network identification layer, first, the convolutional neural network (CNN) embedding physical field prior knowledge is designed. Combined with the objective physical law of pollutant deposition on the membrane surface, such law is integrated into the feature extraction logic of CNN, such as in the network parameter setting, the area characteristics of the shear force distribution matrix where the pollutant deposition is more likely to occur are given higher response weight, to ensure that the spatial features extracted by the network always conform to the physical reality, and avoid feature deviation from the law. Subsequently, the membrane surface defect image in the multi-dimensional wastewater feature and the shear force distribution matrix are taken as the dual input of CNN. The membrane surface defect image provides intuitive spatial information such as the morphology and position of the pollutant, and the shear force distribution matrix provides the physical environment information of the pollutant deposition. Through multi-channel convolution operation, the spatial features of the two types of data are extracted synchronously, and then through pooling, activation and other processing steps, the joint spatial features with both pollutant morphology and deposition environment information are integrated. After completing the feature extraction of CNN, the time-series operating parameters and deposition path data in the multi-dimensional wastewater feature are input into the long short-term memory network (LSTM). With the dynamic adjustment of the gating mechanism, LSTM can not only retain the key associated information in long-term time-series data, but also capture the mutation characteristics of short-term data, and then mine the dynamic association rules between operating parameter changes and deposition path evolution, such as specific corresponding modes of operating parameters and deposition paths under different pollution types. Finally, the identification layer integrates the joint spatial features output by CNN and the dynamic features output by LSTM into the same feature space through a feature fusion unit, forms comprehensive features covering spatial, temporal and physical multi-dimensional information, and then inputs the comprehensive features into the classification module to complete the identification of the membrane surface pollution type combined with the preset pollution type feature rules.

[0037] The principle of the above-mentioned identification layer is that the CNN embedding physical field prior knowledge ensures that the spatial feature extraction conforms to the objective law, and the LSTM makes up for the limitations of CNN in processing time-series data. The combination of the two realizes the comprehensive fusion of multi-dimensional information, so that different pollution types can be accurately distinguished even under complex working conditions, providing an accurate basis for the subsequent prediction layer.

[0038] In the embodiments of the present application, the film surface defect image is a microscopic morphological image obtained by deploying a high-power industrial camera or an electron microscope at a preset production node to take real-time pictures of the surface of the film assembly. The core function is to intuitively present the physical state and early abnormal signs of the film material surface. This type of image can clearly capture various defect characteristics on the film surface, such as tiny pinholes, local scratches, structural damage areas, or microscopic cracks caused by material aging. It can also record the traces of initial contamination. For example, flocculent organic matter and particulate colloids attached to the film surface. The morphology, size, and distribution of these contaminants can be directly observed through the image. It is not a simple appearance record, but a key carrier of microscopic evidence for subsequent contamination identification. For example, through the flocculent morphology of the contaminants in the image, it can be preliminarily judged that organic matter may be attached. Combined with the contamination aggregation situation of the defect location, it can also analyze whether the defect has aggravated the local pollution deposition, providing intuitive microscopic basis for subsequent accurate judgment of contamination type and development of targeted cleaning strategies.

[0039] The shear force distribution matrix is a computational fluid dynamics module in the digital twin model, which is simulated based on the actual structure of the membrane assembly (such as membrane filament arrangement, water flow channel size) and real-time operating parameters (such as water inlet flow rate, wastewater viscosity), used to describe the spatial distribution of shear force generated by wastewater flowing on the membrane surface. Shear force is essentially the friction of water flow on the membrane surface and contaminants, and its size directly affects the deposition trend of contaminants. Generally, the higher the shear force, the stronger the scouring effect of water flow on contaminants, and the more difficult it is for contaminants to attach. The lower the shear force, the easier it is for contaminants to deposit stably. The matrix presents the shear force size of different spatial nodes on the membrane surface in a numerical form, which can clearly reflect the high shear force area and low shear force area on the membrane surface, such as the area with high shear force in the center of the membrane assembly where the water flow is fast, and the area with low shear force at the edge or water flow dead angle. Thus, it provides the physical environmental background for contamination deposition, such as combining the membrane surface defect image, if a large number of particulate contaminants are found to be aggregated in the low shear force area, it can be quickly associated that low shear force and defects are the double contamination causes, providing physical law support for subsequent contamination prediction (such as faster contamination expansion in low shear force area) and operating parameter optimization (such as increasing local flow rate to enhance shear force), avoiding the problem of incomplete understanding of contamination causes caused by relying only on macroscopic data.

[0040] In the prediction layer, firstly, a spatio-temporal attention Transformer network with physical field evolution constraints is introduced. Based on the virtual-real synchronization characteristics of the digital twin model, according to the membrane surface pollution type output by the recognition layer, key information that conforms to the physical field evolution law is filtered and extracted from real-time physical field simulation data, i.e. shear force change gradient and deposition path expansion rate. These information directly reflects the dynamic trend of pollutant diffusion and deposition. Then, the extracted shear force change gradient and deposition path expansion rate are associated with the historical pollutant evolution characteristics in the historical membrane process operation data. During the association process, the weight is adjusted according to the correlation between the characteristics and the current pollution prediction, to ensure that the historical experience can effectively adapt to the current working condition. At the same time, a pollutant diffusion constraint term derived from the CFD physical model is embedded in the loss function of the spatio-temporal attention Transformer network. Through the action of the constraint term, it is ensured that the network prediction process does not violate the physical law of pollutant diffusion, such as avoiding the prediction direction that the pollutant rapidly expands in the high shear force area which is inconsistent with the actual situation. Finally, the spatio-temporal attention Transformer network focuses on the feature information that has a significant impact on pollution prediction, such as the pollution expansion trend under a specific shear force change gradient and the historical evolution characteristics that are highly matched with the current pollution type, through the network's time series and spatial modeling capabilities, the membrane flux decay rate and the pollution area expansion range within a preset time are predicted.

[0041] The principle of the above prediction layer is that the spatio-temporal attention Transformer network can efficiently capture the key association of multi-dimensional features, the physical field evolution constraint ensures that the prediction conforms to the objective law, the historical data association provides experience support for the prediction, and the combination of the three realizes accurate prediction, so as to not only accurately predict the decay trend of the membrane flux, but also clearly determine the expansion range of the pollution area, providing accurate time and space basis for subsequent development of targeted operation strategies, and ensuring the stable operation of the membrane process.

[0042] It can be understood that the implementation of the hybrid network prediction layer first starts from the accurate integration and screening of multi-source information, relying on the virtual-real synchronization characteristics of the digital twin model, according to the membrane surface pollution type output by the recognition layer (such as organic matter adhesion, colloid blockage, etc.), the key features highly related to the current pollution type are extracted from the real-time physical field simulation data. That is, the shear force change gradient (reflecting the dynamic change trend of shear force over time) and the deposition path expansion rate (reflecting the speed of pollutant spatial diffusion), while the historical pollutant evolution characteristics (such as the flux decay rhythm under the same type of pollution in the past, the pollution area expansion mode) that match the current pollution type and similar water quality working conditions in the historical membrane process operation data are called, and the current real-time membrane process operation parameters (such as transmembrane pressure difference, inlet flow rate) are included to form a complete input information set for the prediction layer, ensuring that the input data not only fits the current system state, but also has the dual support of physical law and historical experience.

[0043] Subsequently, the feature processing and association stage is entered, and different types of input features are adapted and converted: classification features such as pollution types are converted into quantifiable feature vectors through embedding vectors, so that they can be calculated with numerical features. For physical field features such as shear force change gradient and deposition path expansion rate, the dimension is adjusted through feature mapping to keep the dimension consistent with the historical pollutant evolution feature and eliminate the difference in data format. On this basis, according to the similarity of the current working condition and the historical working condition (such as the matching degree of water quality parameters and operating pressure), the historical pollutant evolution feature is weighted. That is, the higher the similarity, the greater the weight of the historical data, ensuring that the historical experience can accurately adapt to the current prediction scenario and avoiding irrelevant historical data interference. At the same time, in this stage, pollutant diffusion constraint rules derived based on CFD physical models are embedded simultaneously, which are converted into constraint conditions at the feature level, such as the feature association logic of presetting a lower pollutant diffusion rate in areas with a larger shear force change gradient, which defines the physical boundary for subsequent network prediction.

[0044] Next, the core operation link of the spatio-temporal attention Transformer network. The Transformer network first focuses on key information of the processed multi-dimensional features through attention mechanism. For the time dimension, it focuses on features that are strongly related to membrane flux changes, such as shear force change gradient mutation points, continuous rising trend of transmembrane pressure difference, and historical flux decay feature fragments matching the current pollution type. For the spatial dimension, it focuses on capturing areas with abnormal deposition path expansion rate (such as an area with significantly higher expansion rate than other areas) and potential high-pollution risk areas with sustained low shear force, ensuring that the network concentrates computing resources on features that have the greatest impact on prediction results. Based on feature focusing, the network learns the time correlation rules between features through time modeling modules, such as learning the time sequence chain of shear force change gradient slowing down, deposition path expansion rate accelerating, and membrane flux decaying, and then deducing the continuous decay trend of membrane flux in the future preset time. At the same time, through the spatial modeling module, combined with the three-dimensional geometric structure of the digital twin model membrane component, the diffusion differences of pollutants in different shear force regions are simulated, such as in areas with small shear force change gradient and fast expansion rate, the area expansion ratio of pollutants over time is calculated, and the specific spatial location of pollution coverage is labeled based on the three-dimensional coordinate system of the membrane component.

[0045] In the network training and optimization process, the pollutant diffusion constraint term derived from the CFD physical model will be explicitly embedded in the loss function. When the network prediction result deviates from the physical law (such as the pollutant diffusion rate in high shear force area being higher than that in low shear force area), the constraint term will increase the loss value, guiding the network to adjust the parameters to ensure that the prediction logic always conforms to the objective physical mechanism of pollutant deposition and diffusion, avoiding the deviation from reality caused by pure data-driven. Finally, after multiple rounds of iteration optimization, the network outputs two types of core prediction results: the first is the membrane flux decay rate in the future preset time, which presents the trend of flux change over time in the form of continuous time series curve; the second is the expansion range of the pollution area, which marks the specific location and area ratio of pollution coverage in the three-dimensional space coordinates of the membrane module, forming a complete membrane pollution prediction output, providing accurate basis for subsequent operation and maintenance strategy.

[0046] Further optionally, after predicting the membrane pollution prediction result in step S103, the corresponding pulse cleaning mode can be selected according to the membrane surface pollution type and flux decay rate prediction value in the membrane pollution prediction result. The membrane surface pollution type includes organic matter attachment and / or colloidal blockage. The adaptive adjustment parameters in the pulse cleaning mode include at least high-pressure backflush intensity, pulse frequency, drug type, drug concentration, and cleaning cycle. Then, according to the real-time flux decay rate and the preset parameter threshold in the pulse cleaning mode, the adaptive adjustment parameters in the pulse cleaning mode are dynamically adjusted, and the membrane pollution recovery rate is monitored in real time to dynamically adjust the cleaning cycle of the pulse cleaning mode.

[0047] Specifically, first, the core direction of cleaning is determined according to the membrane surface pollution type in the result, and then the severity of the pollution is judged by combining the flux decay rate prediction value, and the corresponding pulse cleaning mode is matched. When the pollution type is organic matter attachment, the pulse mode of mainly assisted by drug and supplemented by high-pressure backflush is selected, because organic matter needs to be removed by chemical action of specific drugs. If the pollution type is colloidal blockage, the pulse mode of mainly high-pressure backflush and assisted by low-concentration drugs is selected, because colloidal particles are more easily dispersed from the membrane surface by the physical scouring force of high-pressure water flow. If it is mixed pollution of both, the intensity of drug action and high-pressure backflush will be balanced to form a composite pulse cleaning mode that takes into account chemical decomposition and physical scouring.

[0048] After determining the basic pulse cleaning mode, the actual flux decay rate in the membrane process operation is collected in real time, and it is continuously compared with the preset parameter threshold of the pulse cleaning mode. If the actual flux decay rate is higher than the preset threshold, it means that the current cleaning intensity is not enough to inhibit the pollution accumulation, and the high-pressure backflush intensity needs to be appropriately increased to enhance the physical flushing effect, or the pulse frequency needs to be increased to shorten the flushing interval, and the drug concentration also needs to be fine-tuned for the organic matter attachment scene to strengthen the chemical decomposition effect. If the actual flux decay rate is lower than the preset threshold, in order to avoid damage to the membrane structure caused by excessive cleaning (such as membrane filament rupture caused by high-frequency high-pressure backflush and excessive drug may corrode the membrane material), the high-pressure backflush intensity or the pulse frequency needs to be appropriately reduced, and the drug dosage needs to be controlled to ensure that the cleaning intensity is adapted to the pollution level.

[0049] During the entire pulse cleaning process, the membrane pollution recovery rate, i.e. the proportion of the membrane flux recovered to the normal operation level after cleaning, is continuously monitored. If the recovery rate reaches the expected standard, it means that the current cleaning period is matched with the pollution accumulation rate, and the existing period can be maintained or appropriately extended to reduce unnecessary cleaning times and reduce energy consumption and drug consumption in the cleaning process. If the recovery rate does not reach the expected standard, it may be that the pollution is excessively accumulated due to the long cleaning period, which exceeds the processing capacity of the current cleaning parameters. At this time, the cleaning period needs to be appropriately shortened, and the high-pressure backflush, drug and other parameters are further adjusted in combination with the real-time flux decay rate to form a closed-loop regulation of parameter adjustment, recovery rate monitoring and period optimization, ensuring the balance between cleaning effect and membrane protection.

[0050] Through the above steps, the targeted cleaning strategy is formulated based on the differences in the formation mechanism of different types of pollution, following the rules of adapting organic matter by chemical action and adapting colloids by physical flushing, avoiding the limitations of a single cleaning mode. The dynamic correction of cleaning parameters is realized through the comparison of real-time flux decay rate and threshold, which essentially adjusts the intervention intensity according to the real-time state of pollution development, which is in line with the characteristics of dynamic operation of membrane process. The monitoring of pollution recovery rate provides a direct basis for cleaning period optimization, ensuring that the cleaning frequency neither increases the cost due to being too short, nor causes pollution to intensify due to being too long. Finally, the pulse cleaning mode is accurately adapted to the membrane pollution state, effectively removing pollutants and recovering membrane flux, while avoiding excessive cleaning damage to the membrane module, prolonging the service life of the membrane, reducing cleaning energy consumption and drug consumption, ensuring the stability and economy of industrial wastewater membrane process operation, and avoiding process interruption or processing efficiency reduction caused by improper cleaning.

[0051] In step S104, a multi-strategy evolution reinforcement learning algorithm driven by episodic memory is used to combine the wastewater quality parameters, the production line equipment state, and the membrane pollution prediction results, and a multi-agent collaborative decision algorithm is used to perform multi-objective dynamic optimization on the membrane process operation parameters and the production control parameters of the wastewater treatment system. In the multi-agent collaborative decision algorithm, the multi-agent collaborative structure is constructed with the optimization objectives of maximizing the membrane flux, minimizing the energy consumption, and prolonging the chemical cleaning period.

[0052] Further optionally, a multi-agent collaborative structure adopting a double-layer master-slave structure is constructed. In constructing the multi-agent collaborative structure adopting the double-layer master-slave structure, first, the master agent is taken as a global optimizer, and according to the multi-objective optimization requirements of maximizing membrane flux, minimizing energy consumption, and prolonging chemical cleaning period, the internal correlation between each target and the key link of the wastewater treatment process is analyzed first, that is, the promotion of membrane flux depends on the microstructure adaptability of the membrane material, the control of energy consumption depends on the accurate adjustment of real-time operation parameters, and the prolongation of the chemical cleaning period requires the development of an adaptive cleaning strategy based on the pollution state. Based on this correlation, the multi-objective optimization problem is decomposed into three types of sub-tasks, namely material design, real-time control, and cleaning strategy. Then, algorithms are matched to the characteristics of different sub-tasks. The material design sub-task needs to find a multi-objective balance between membrane flux promotion and anti-pollution enhancement, NSGA-III algorithm is good at handling multi-objective optimization scenarios and can generate optimal solution sets that take into account different targets, so it is used for this task. The real-time control sub-task needs to dynamically fine-tune continuous variables such as operating pressure and inlet flow rate to quickly respond to water quality fluctuations and pollution changes. TD3 algorithm has precise and stable parameter output capability in continuous control scenarios, making it suitable for this task. The cleaning strategy sub-task needs to quickly match the cleaning rules based on pollution type, flux attenuation rate, and other working conditions. Rule engine can achieve efficient response through pre-set logical rules, so it is used to handle this task. Then, while decomposing the tasks, the master agent will monitor the load status of the edge computing nodes in real time by collecting real-time CPU and GPU occupancy rate data to determine the remaining computing resources of each node. According to the computing requirements of different sub-tasks, resources are dynamically allocated. The material design sub-task requires a large amount of computation when generating the Pareto solution set, so resources should be appropriately tilted on the premise of ensuring the normal operation of other tasks. The real-time control sub-task requires high response delay, so sufficient resources should be reserved to ensure that parameter adjustment instructions can be generated quickly. The cleaning strategy sub-task has relatively small computation, so the remaining resources can be allocated flexibly. During the entire resource allocation process, the master agent will continuously regulate to ensure that the overall CPU and GPU occupancy rate does not exceed the set upper limit, avoiding task delays caused by overloaded computing nodes, and ensuring the processing efficiency of each sub-task and the real-time performance of process adjustment. The agent acts as a special executor and performs tasks assigned by the master agent. The material optimization agent focuses on the material design sub-task and receives wastewater quality parameters (such as pollutant particle size and water viscosity) and membrane performance requirements from the master agent, and then performs multi-objective optimization on the membrane pore size distribution based on the NSGA-III algorithm. By adjusting parameters such as pore size and distribution density, the membrane flux is improved while ensuring that the membrane pores can effectively trap pollutant particles to enhance anti-pollution, and finally a Pareto solution set that meets the collaborative optimization of flux promotion and anti-pollution enhancement is output, providing the master agent with multiple optimal options for developing membrane material adjustment schemes.The control agent is responsible for executing real-time control subtasks, combining wastewater quality parameters (such as COD, salinity), production line equipment status (such as the current transmembrane pressure difference), and membrane pollution prediction results (such as future flux attenuation trends), and based on the TD3 algorithm, a parameter optimization model is constructed to generate collaborative control instructions for operating pressure and water flow rate. When generating collaborative control instructions for operating pressure and water flow rate, the discretization step of the action space is strictly controlled to ensure the smoothness of parameter adjustment and avoid the impact of pressure or flow rate on membrane components (such as membrane filament rupture and membrane hole damage). At the same time, the instructions can accurately adapt to the current working conditions. For example, for high pollution risk scenarios, the water flow rate is appropriately increased to enhance the shear force on the membrane surface and inhibit the deposition of pollutants, ultimately generating collaborative control instructions that meet the needs of energy minimization and flux maximization. In this way, the double-layer master-slave structure solves the task decomposition and resource coordination problem of multi-objective optimization through the global planning of the master agent, avoiding the low efficiency and target conflict when a single agent handles multiple tasks. The specific execution of the agent focuses on specific tasks and improves optimization accuracy and response speed through adaptive algorithms. Dynamic resource allocation ensures the stability of system operation and avoids the impact of computational bottlenecks on the real-time performance of process adjustment. Thus, the optimization of membrane process operating parameters and production control parameters can meet the multi-objective requirements simultaneously. The Pareto solution set of the material optimization agent supports the improvement of membrane material performance, directly contributing to flux growth and anti-pollution improvement. The collaborative instructions of the control agent accurately adjust the operating parameters, reduce invalid energy consumption, and inhibit pollution deposition. The rule-based processing of the cleaning strategy subtask prolongs the cleaning cycle and reduces operation and maintenance costs, ultimately realizing the efficient, low-cost, and stable operation of the wastewater treatment system, and avoiding the decline in treatment efficiency or damage to membrane components caused by unbalanced parameter optimization.

[0053] As an optional embodiment, in step S104, the historical optimal triplets containing states, actions and rewards in the historical running process are extracted from the historical membrane process running data by the episodic memory tree EMT to provide guidance for the exploration process of the reinforcement learning agent in the sparse reward scenario; by the reinforcement learning agent, the TD3 algorithm is adopted to combine the real-time collected wastewater quality parameters, the production line equipment state and the predicted membrane pollution prediction results to optimize the water inlet flow rate in the membrane process running parameter and the operating pressure in the production control parameter in real time, the input of the Critic network of the reinforcement learning agent being the membrane pollution prediction results predicted by the digital twin model; by the evolutionary algorithm population, the NSGA-III algorithm is adopted to combine the membrane pore size distribution in the membrane microstructure parameter to optimize and solve the hydrophilic and hydrophobic parameters and the production line process combination, and the optimization strategy parameters of the elite individuals are obtained; the Genetic Soft Update mechanism is adopted to migrate the optimization strategy parameters of the elite individuals in the evolutionary algorithm population to the reinforcement learning agent, and the final output multi-objective dynamic optimization result is obtained through multiple rounds of cooperation between the reinforcement learning agent and the evolutionary algorithm population, and the multi-objective dynamic optimization of the membrane process running parameter and the production control parameter is completed.

[0054] Specifically, first, the historical membrane process running data is systematically combed and mined by the episodic memory tree EMT, and the historical optimal triplets containing states, actions and rewards are selected from the massive historical data. The states cover the wastewater quality parameters and the production line equipment running state in the past different time periods, the actions refer to the adjustment strategies for the membrane process running parameters (such as the water inlet flow rate) and the production control parameters (such as the operating pressure) at that time, and the rewards are the comprehensive feedback results of the multi-objective such as the membrane flux, energy consumption and chemical cleaning period after the adjustment. The episodic memory tree stores and indexes these optimal triplets in a hierarchical structure, and when the reinforcement learning agent is in a sparse reward scenario (i.e. the current adjustment cannot quickly obtain obvious reward feedback, such as the slight parameter adjustment cannot intuitively reflect the flux change in the short term), the optimal action and reward information under the similar historical state can be retrieved from the episodic memory tree to provide guidance for the exploration direction of the agent, avoid the agent from falling into blind exploration, and reduce the number of invalid adjustments.

[0055] Subsequently, the reinforcement learning agent employs the TD3 algorithm for real-time parameter optimization, focusing on the feed flow rate in the membrane process operating parameters and the operating pressure in the production control parameters. During the optimization process, the agent continuously receives real-time collected wastewater quality parameters (such as COD, salinity), production line equipment status (such as current transmembrane pressure difference, real-time flux), and the membrane pollution prediction results (such as future preset time length of flux decay rate, pollution area expansion trend) output by the digital twin model as the core input of its Critic network. When evaluating the value of actions, the Critic network not only considers the current system state, but also predicts the impact of parameter adjustment on future pollution development in advance. For example, if a pressure adjustment is predicted to accelerate subsequent colloidal clogging, the Critic network will lower the value score of that action, guiding the agent to choose a better solution. Through this evaluation mechanism combining real-time data and forward-looking prediction, the agent can dynamically adjust the feed flow rate and operating pressure to ensure current flux while suppressing future pollution intensification, balancing short-term operating efficiency and long-term system stability.

[0056] At the same time, the evolutionary algorithm population uses the NSGA-III algorithm for global multi-objective optimization, covering membrane microstructure parameters such as membrane pore size distribution, membrane material hydrophilic / hydrophobic parameters, and production line process combinations (such as different combinations of feed pretreatment processes and membrane filtration processes). The algorithm will construct an initial population containing multiple parameter combinations, and through simulation of selection, crossover, and mutation operations in biological evolution, iteratively optimize individuals in the population. In each iteration, the performance of different parameter combinations is evaluated based on multiple objectives such as flux improvement, energy consumption reduction, and anti-pollution enhancement, and the elite individuals with the best overall performance are selected. The parameter combinations corresponding to these elite individuals are the optimization strategy parameters at the current iteration, for example, a certain elite individual may correspond to a specific pore size distribution, moderate hydrophobicity, and a combination of certain pretreatment processes, which can simultaneously satisfy the coordinated improvement of flux and anti-pollution.

[0057] To realize the collaborative optimization of the reinforcement learning agent and the evolutionary algorithm population, the Genetic Soft Update mechanism is adopted for policy parameter migration. This mechanism does not directly replace the parameter of the reinforcement learning agent with the optimization strategy parameter of the elite individual in the evolutionary algorithm population, but gradually migrates the features beneficial to multi-objective optimization (such as anti-pollution experience corresponding to specific hydrophobicity parameters) in the elite individual parameter to the policy network of the agent in a slow fusion manner, avoiding the instability of the agent's strategy caused by parameter mutation. On this basis, the reinforcement learning agent and the evolutionary algorithm population will collaborate for multiple rounds. The real-time parameter adjustment results of the reinforcement learning agent will be fed back to the evolutionary algorithm population, providing effective data support for the next round of iteration of the population. The new elite individual parameter of the evolutionary algorithm population will update the reinforcement learning agent through the Genetic Soft Update mechanism, helping the agent break through the local optimum and explore the parameter space that better meets the global multi-objective demand. After multiple rounds of such collaborative interaction, the multi-objective dynamic optimization result that balances membrane flux maximization, energy minimization, and chemical cleaning cycle extension is finally output, completing the comprehensive optimization of membrane process operating parameters and production control parameters.

[0058] In this way, the exploration efficiency problem under sparse rewards is solved by the scenario memory tree, the real-time dynamic adjustment of parameters is realized by relying on the TD3 algorithm, the global multi-objective macro-optimization is completed by means of the NSGA-III algorithm, and the collaborative fusion of local real-time optimization and global macro-optimization is realized through the Genetic Soft Update mechanism, forming a complete optimization logic of historical experience guidance, real-time dynamic adjustment, global optimization supplement, and collaborative iteration improvement. Thus, not only can the parameter adjustment quickly adapt to the changes in current water quality and equipment state, but also the membrane microstructure and process combination can be optimized from a global perspective, avoiding the limitations of single algorithm optimization. For example, the real-time adjustment of reinforcement learning can cope with short-term water quality fluctuations, and the global optimization of evolutionary algorithm can solve the problem of long-term membrane performance degradation, ultimately realizing the multi-objective balance of membrane process operating parameters and production control parameters, improving membrane flux and processing efficiency while reducing energy consumption, prolonging the chemical cleaning cycle, and ensuring the stable and efficient operation of the wastewater treatment system under complex and variable working conditions.

[0059] Further optionally, the multi-agent collaborative structure adopts a double-layer master-slave architecture, including a master agent and a slave agent. The master agent serves as a global optimizer, responsible for the decomposition of the multi-objective optimization task and resource scheduling. The slave agent serves as a special executor, including a material optimization agent, a control agent, and a cleaning strategy agent. Based on this structure, in the optional embodiment of step S104, the TD3 algorithm is used to combine the real-time collected wastewater quality parameters, production line equipment states, and predicted membrane pollution prediction results to optimize the water inlet flow rate in the membrane process operation parameters and the operating pressure in the production control parameters in real time, including: through the double-layer master-slave architecture of multi-agent collaboration, the control agent constructs an action decision model based on the TD3 algorithm, and takes the real-time wastewater quality parameters, the current operating pressure in the production line equipment state, the water inlet flow rate, the transmembrane pressure difference, and the flux decay rate and pollution area expansion range in the future preset time length in the membrane pollution prediction result as state input. Further, the action space is set as the water inlet flow rate adjustment range matched with the membrane module type and the operating pressure adjustment range conforming to the pressure bearing standard of the membrane module, and the action discretization step is set according to the control accuracy requirement; in the reward function, the reward value is obtained by weighted calculation of the real-time flux improvement rate, the energy consumption growth rate, and the cleaning cycle extension rate, the weight of the real-time flux improvement rate is higher than that of the energy consumption growth rate and the cleaning cycle extension rate, and the weights of the energy consumption growth rate and the cleaning cycle extension rate are the same, and the real-time flux improvement rate is calculated based on the difference between the predicted value and the current value of the digital twin model. Finally, the Actor network of the TD3 algorithm generates a cooperative control instruction of the operating pressure and the water inlet flow rate, the Critic network combines the risk coefficient in the membrane pollution prediction result to evaluate the value, and the pollution area expansion rate exceeds the preset threshold to trigger a penalty term, and iterative optimization is performed until convergence, realizing real-time dynamic adjustment of the water inlet flow rate and the operating pressure.

[0060] Exemplarily, in the double-layer master-slave architecture, the TD3 algorithm is used to combine the real-time collected wastewater quality parameters, production line equipment states, and predicted membrane pollution prediction results to optimize the water inlet flow rate in the membrane process operation parameters and the operating pressure in the production control parameters in real time, which can be realized as: Firstly, the task allocation and data synchronization of the agent. The master agent divides the real-time optimization process of the feed flow rate in the membrane process operating parameters and the operating pressure in the production control parameters into a special subtask, assigns it to the control agent, and synchronizes the real-time wastewater quality parameters, production line equipment status (current operating pressure, feed flow rate, transmembrane pressure difference), and membrane fouling prediction results (future flux decay rate, pollution area expansion range) output by the digital twin model. At the same time, it is clear that this subtask needs to match the multi-objective constraints of maximizing global membrane flux, minimizing energy consumption, and prolonging chemical cleaning cycle. Further, the model construction and parameter setting of the control agent. Based on the task and synchronized data assigned by the master agent, the control agent uses the TD3 algorithm to construct an action decision model, using the above data provided by the master agent as the state input of the model. The action space is pre-defined by the master agent according to the process adaptability (feed flow rate adjustment range) and safety pressure standard (operating pressure adjustment range) corresponding to the type of membrane module, and the control agent determines the action discretization step according to the control accuracy requirement set by the master agent. Optionally, the weight rule of the reward function is set by the master agent according to the global multi-objective priority. The real-time flux improvement rate weight is higher than the energy consumption growth rate and the cleaning cycle extension rate, and the energy consumption growth rate and the cleaning cycle extension rate weight is equal. Based on this rule, the control agent calculates the real-time flux improvement rate by combining the difference between the predicted value and the current value of the flux output by the digital twin model, and then determines the reward value. Then, the optimization iteration and result feedback of the control agent. The control agent generates a coordinated control instruction for the operating pressure and the feed flow rate through the Actor network of the TD3 algorithm, and then evaluates the value of the control instruction through the Critic network combined with the pollution risk threshold (triggering a penalty term when the pollution area expansion rate exceeds this threshold) pre-set by the master agent. The control agent continuously iterates and optimizes the model until it meets the convergence conditions set by the master agent, and then feeds back the optimization results to the master agent. Specifically, during the optimization iteration process of the control agent, when the Critic network evaluates the value, it will not only consider the penalty term associated with the pollution area expansion rate, but also consider the future membrane pollution evolution trend predicted by the digital twin model. If the evaluation finds that a control instruction can improve the flux in the short term but will lead to rapid expansion of the pollution area in the future, even if the current pollution expansion rate does not exceed the threshold, the value score of that instruction will be appropriately reduced, guiding the Actor network to generate control instructions with more long-term stability. During the iteration and optimization process, the control agent will also regularly feedback the optimization progress to the master agent. If the iteration falls into a local optimum (such as no significant change in flux improvement rate after continuous optimization for multiple rounds), the master agent can trigger the parameter reset mechanism to help the control agent escape from the local optimum and explore a better parameter space. Finally, the global verification and instruction issuance of the master agent.The master agent globally checks the optimization result fed back by the control agent, and after confirming that the checking result meets the multi-objective balance requirements of membrane flux, energy consumption, and cleaning cycle, issues a final adjustment instruction to the wastewater treatment system to realize real-time dynamic adjustment of the influent flow rate and operating pressure. Specifically, when performing global checking, the master agent not only checks whether the optimization result meets the multi-objective constraints, but also cooperatively checks it with the output results of the material optimization agent and the cleaning strategy agent. For example, if the membrane pore size distribution optimization scheme output by the material optimization agent tends to improve the anti-pollution property, the master agent will determine whether the operating pressure adjustment of the control agent is suitable for the scheme to avoid parameter conflicts. When issuing an instruction after passing the check, a gradual adjustment method is adopted instead of adjusting the parameters to the target value at once. For example, the influent flow rate is adjusted to the optimized value in multiple small steps to prevent the water flow state in the membrane module from changing dramatically due to parameter mutation, which may cause rapid deposition of local pollutants.

[0061] Thus, the double-layer master-slave architecture solves the problems of data consistency, resource coordination, and cross-agent parameter adaptation among multiple agents through global planning by the master agent, avoiding one-sidedness and limitations in single-agent decision-making. The double Critic network of the TD3 algorithm in the control agent effectively alleviates the value overestimation problem in traditional reinforcement learning, and the reward function combined with safety constraint logic and dynamic weights ensures that the optimization process is carried out around multi-objective requirements without breaking the safety operating boundary of the membrane module. The long-term pollution trend evaluation and local optimal breakthrough mechanism in the optimization iteration enable the optimization result to break away from short-term benefit limitations and have long-term stability. The cooperative checking of the master agent and the gradual issuance of instructions further reduce the impact of parameter adjustment on the system and ensure the synergy of the overall process. Thus, precise, safe, and dynamic adjustment of the influent flow rate and operating pressure is achieved, which prioritizes membrane flux improvement to meet the wastewater treatment efficiency requirements while effectively controlling energy consumption growth, extending the chemical cleaning cycle, and avoiding the sacrifice of other indicators due to the pursuit of a single target. The reward function and safety constraints that adapt to dynamic working conditions enable the optimization strategy to flexibly respond to complex scenarios such as wastewater quality fluctuations and treatment load changes, quickly outputting adaptive control instructions. The cooperative checking of multiple agents and gradual adjustment significantly reduce the risk of parameter mutation damaging the membrane module, ensuring long-term stable operation of the membrane process, while improving the anti-interference ability and multi-objective balanced optimization level of the entire system, providing reliable parameter regulation support for efficient, low-consumption, and stable operation of industrial wastewater membrane processes.

[0062] Further optionally, in the optional embodiment of step S104, the optimization strategy parameters of the elite individuals in the evolutionary algorithm population are migrated to the reinforcement learning agent by using a Genetic Soft Update mechanism, including: selecting elite individuals with fitness values meeting a preset ranking from the evolutionary algorithm population, and extracting the strategy parameters corresponding to the elite individuals; the fitness value is calculated by a weighted score of membrane flux, membrane process energy consumption and anti-pollution performance. Then, the optimization strategy parameters corresponding to the elite individuals are extracted, including the membrane pore size distribution interval, the hydrophilic-hydrophobic adjustment coefficient and the production line process combination scheme, and the optimization strategy parameters are weighted and fused with the current strategy parameters of the reinforcement learning agent. Finally, the fused strategy parameters are updated to the Actor network of the reinforcement learning agent through a sliding average filtering process.

[0063] Specifically, when migrating parameters using the Genetic Soft Update mechanism, elite individuals are first selected from the evolutionary algorithm population. This selection is not simply based on fitness value ranking, but rather a dynamic adjustment of the pre-set ranking criteria in combination with the current operating conditions of the wastewater treatment system. For example, in high pollution risk scenarios, the proportion of pollution resistance in the fitness weighted score is increased, making individuals with better pollution resistance performance more likely to enter the elite list. The weighted score calculation of the fitness value is not fixed weight, but the master agent adjusts the weight of each index in real time according to the global multi-objective priority (such as the current urgency of flux, energy consumption, and pollution resistance), ensuring that the selected elite individuals can adapt to the direction that the system needs to optimize the most. When extracting the optimization strategy parameters of elite individuals, the parameters are not directly used, but are first standardized. For example, the pore size distribution interval of different membrane modules is converted into a relative proportion value, and the hydrophilic and hydrophobic adjustment coefficient is uniformly mapped to the same numerical range, so that various parameters have consistent fusion consistency. For production line process combination schemes, which are non-numeric parameters, they are converted into a calculable vector form through feature encoding, ensuring that they can be fused with the policy parameters of the reinforcement learning agent in the same feature space. In the parameter weighted fusion stage, the weight distribution is not a fixed proportion, but is dynamically determined according to the fitness ranking of elite individuals and the stability of the existing policy of the reinforcement learning agent. The earlier the elite individual in the fitness ranking, the higher the weight of its parameters in the fusion, and if the current policy of the reinforcement learning agent is in a stable convergence state (such as small fluctuations in consecutive optimization results), the weight of the elite parameters will be appropriately reduced to avoid disrupting the existing stable policy, and vice versa. The weight will be increased to promote policy updates; during the fusion process, the parameters will also be checked for reasonableness, and if the elite parameters differ significantly from the current parameters of the agent (such as exceeding the physical feasibility range), the fusion weight of that part of the parameter will be automatically reduced to prevent the introduction of abnormal parameters. Finally, when performing sliding average filtering, instead of simply taking the average value of a fixed window, a dynamic window size is set according to the parameter type. For parameters such as membrane pore size distribution interval that need to be stable in the long term, a larger window is used to smooth fluctuations, and for parameters such as hydrophilic and hydrophobic adjustment coefficient that can be adjusted in the short term, a smaller window is used to maintain sensitivity. During the filtering process, the trend characteristics of the parameters are also preserved, for example, if the elite parameters show that the pore size needs to be gradually increased to improve the flux, the filtering will not eliminate this trend, but will make the change more gradual, ensuring that the parameters updated to the reinforcement learning agent Actor network have absorbed the optimization information of the elite individuals and maintained the continuity of the policy.

[0064] Thus, the Genetic Soft Update mechanism ensures that the migrated parameters can accurately match the optimization needs of the current system by dynamically screening elite individuals. Standardization and feature encoding solve the fusion barrier of different types of parameters, allowing the structure parameters of global optimization (such as membrane pore size) to be effectively combined with real-time control parameters (such as operating pressure); dynamic weighting and reasonableness verification balance the stability of global optimization information and local real-time strategy, avoiding system oscillation caused by parameter mutation. The moving average filter further smooths the parameter update process, ensuring the gradualness and reliability of policy adjustment. Thus, the reinforcement learning agent can not only retain the real-time dynamic adjustment capability obtained through the TD3 algorithm, but also absorb the global optimal strategy information explored by the evolutionary algorithm population through the NSGA-III algorithm, breaking through the limitations of a single algorithm. For example, when adjusting the inflow velocity, the agent will refer to the membrane pore size optimization scheme of the elite individual, making the flow rate adjustment and membrane structure characteristics more compatible, and avoiding flux loss or increased energy consumption due to parameter mismatch. The smoothness of parameter updating ensures that the system will not run erratically due to policy mutations, allowing the optimization of membrane process parameters to balance global multi-objective while accurately responding to real-time operating conditions, ultimately improving the optimization efficiency and operational stability of the entire wastewater treatment system in complex scenarios.

[0065] Optionally, the weight coefficient of the optimization strategy parameter is dynamically adjusted with the number of iterations. Further, the preset ranking, initial weight, and iteration cycle step value are dynamically set based on the membrane process operation requirements of the target scene, and the dynamically set adaptive rules are stored in the decision knowledge base of the main agent.

[0066] When the weight coefficient of the optimization strategy parameter is dynamically adjusted with the number of iterations, a fixed adjustment rule is not used, but the policy convergence state of the reinforcement learning agent and the diversity change of the evolutionary algorithm population are flexibly adjusted. In the early iterations, the policy of the reinforcement learning agent is still in the exploration stage, so the weight proportion of the evolutionary algorithm elite individual parameters is increased to allow the agent to quickly absorb global optimization information and avoid being trapped in local exploration. As the iteration progresses, when the agent's policy gradually converges (such as a reduction in the fluctuation amplitude of consecutive reward values), the weight of the elite parameters is gradually reduced, and the weight of the agent's own strategy parameters is increased, allowing the agent to perform fine-tuned optimization based on existing exploration and balance global information and local adaptability. If the diversity of the evolutionary algorithm population decreases (such as the convergence of most individual parameters) during the iteration process, the sensitivity of the weight adjustment is temporarily increased to introduce more differentiated information from the elite parameters, restore the exploration vitality of the population, and ensure that the optimization strategy can cover a wider parameter space.

[0067] The dynamic setting of the preset ranking, initial weight, and iteration cycle step value relies on the adaptive rules stored in the main agent decision knowledge base, which are pre-constructed based on the membrane process operation requirements of different target scenarios. For example, in the high-concentration organic wastewater treatment scenario, the risk of membrane fouling is high, and the decision knowledge base will tilt the screening criteria of the preset ranking towards individuals with better anti-fouling properties, while increasing the proportion of anti-fouling related parameters (such as hydrophilic-hydrophobic adjustment coefficient) in the initial weight, and setting the iteration cycle step value to a small value to speed up the parameter adjustment frequency, ensuring a quick response to pollution changes. In the low load operation scenario, energy consumption control becomes the core requirement, and the rules will adjust the preset ranking to focus more on individuals with excellent energy consumption performance, tilt the initial weight towards energy consumption related parameters, and appropriately increase the iteration cycle step value to reduce unnecessary frequent adjustments, balancing optimization accuracy and computational cost. During the operation of the main agent, it will monitor the current membrane process operation state (such as influent pollutant concentration, treatment load) in real time, match the corresponding adaptive rules from the decision knowledge base, and automatically update the preset ranking, initial weight, and iteration cycle step value, without the need for manual intervention to adapt to scenario changes. Thus, dynamic adjustment of weight coefficients addresses the contradiction between global exploration and local optimization by adapting to different requirements in the iteration phase, avoiding strategy bias caused by excessive reliance on local information in the early stage, or optimization stagnation caused by neglecting global information in the later stage. Dynamic parameter setting based on scenario requirements allows the optimization process to accurately interface with the core objectives under different working conditions, solving the limitations of fixed parameters in adapting to complex scenarios and improving the adaptive ability of the optimization strategy. During the iteration process, parameter migration can efficiently absorb global optimal information while maintaining the stability and refinement of local strategies. In different scenarios, the optimization direction can quickly focus on the most urgent requirements (such as anti-fouling and energy reduction), making the adjustment of membrane process operation parameters more targeted, and ultimately achieving a multi-objective balance of membrane flux, energy consumption, and cleaning cycle in various complex working conditions, ensuring long-term efficient and stable operation of the system.

[0068] As an optional embodiment, in step S104, a scenario memory driven multi-strategy evolutionary reinforcement learning algorithm is adopted to combine the wastewater quality parameters, production line equipment states, and membrane fouling prediction results, and after multi-objective dynamic optimization of the membrane process operation parameters and production control parameters of the wastewater treatment system by a multi-agent collaborative decision algorithm, a time sequence modeling method can be used to build a membrane process operation model in the current production cycle based on the multi-objective dynamic optimization results, wherein the inputs of the membrane process operation model are the real-time collected wastewater quality parameters and production line equipment states, and the outputs are the dynamic change curves of the membrane flux and energy consumption. Combined with the dynamic change curves output by the membrane process operation model, the membrane surface defect position, membrane surface defect area, and corresponding membrane surface microstructure change in the current production cycle are predicted to obtain the membrane defect structure prediction results. Further, based on the Bayesian optimization graph neural network model, the membrane surface microstructure is taken as the graph structure node, the stress concentration area caused by local blockage is identified combined with the membrane defect structure prediction results, and the optimal adjustment direction of the membrane material hydrophilicity and hydrophobicity parameters is determined through inverse calculation taking the membrane flux decay rate and energy consumption as observation variables, and the remaining life of the membrane module is predicted. If the remaining life of the membrane module is less than the length of the next production cycle, the optimal membrane module configuration including the membrane pore size distribution, membrane material hydrophilicity and hydrophobicity parameters, and membrane rolling process parameters are selected from the wastewater membrane material spare parts warehouse based on the real-time collected wastewater quality parameters, production line equipment states, membrane fouling prediction results, and membrane defect structure prediction results, which are used to indicate the membrane material and membrane module replacement parts for the next production cycle to the user.

[0069] Specifically, after completing the multi-objective dynamic optimization of the membrane process operation parameters and production control parameters, first, the membrane process operation model of the current production cycle is built based on the optimization results using the time sequence modeling method. During modeling, the historical membrane process operation data of the same type of production cycle and the real-time collected wastewater quality parameters (such as pollutant concentration, pH value) and production line equipment states (such as operating pressure, water inlet flow rate) in the current cycle are fully integrated, and these dynamic changing data are feature learned through time sequence modeling to capture the time sequence correlation between water quality fluctuations, equipment state adjustments, and membrane flux, energy consumption. For example, the change chain of the increase of the concentration of a certain type of pollutant, the lagging decline of the membrane flux, and the increase of the energy consumption is learned, so that the model can output the dynamic change curves of the membrane flux and energy consumption that fit the actual running trend of the current cycle according to the real-time input water quality and equipment state, providing continuous performance change basis for subsequent membrane defect analysis.

[0070] In combination with the dynamic curve output by the membrane process operation model, the surface defect condition of the membrane in the current production cycle is further predicted. This prediction process correlates the pollution types identified in the membrane pollution prediction results with the pollution area expansion trend. For example, when the curve shows that the membrane flux of a certain area is continuously lower than that of other areas and the energy consumption abnormally increases, in combination with the previous colloidal blockage prediction results of the area, it can be inferred that the area may be caused by colloidal deposition leading to membrane pore blockage, thereby causing local defects. In combination with the historical defect data of the membrane module (such as the position and area change law of defects under the same working condition in the past), the specific position of the membrane surface defect in the current cycle, the defect area size, and the membrane surface microstructure change (such as the degree of pore size reduction and the increase of membrane surface roughness) corresponding to the defect area are comprehensively judged, and finally a complete membrane defect structure prediction result is formed, which fills the gap that the micro-defect state is difficult to judge only through the macro-performance curve.

[0071] Subsequently, based on the graph neural network model of Bayesian optimization, the membrane surface microstructure is first decomposed into multiple graph structure nodes, each node corresponding to a specific micro area of the membrane surface (such as membrane pore areas of different pore sizes, surface areas of different hydrophilic and hydrophobic properties), and the connection relationship between the nodes is constructed according to the correlation between the microstructures (such as the correlation of pollutant diffusion between adjacent membrane pores, and the mutual influence of pollution deposition between hydrophilic and hydrophobic areas), forming a complete membrane microstructure graph. After inputting the membrane defect structure prediction result into the model, the graph neural network will identify the stress concentration area caused by local blockage (such as membrane pore blockage in a certain node area) through node feature learning and correlation analysis. This kind of area is prone to local stress overload due to uneven water flow resistance caused by blockage, thereby accelerating the expansion of membrane defects. At the same time, taking the membrane flux decay rate and energy consumption as observation variables, through the inversion calculation ability of Bayesian optimization, the key factors affecting these two variables are reversely deduced to determine the optimal adjustment direction of the hydrophilic and hydrophobic properties of the membrane material, such as appropriately increasing the hydrophobicity to reduce organic matter adhesion, or adjusting to moderate hydrophilicity to balance the flux and anti-pollution property, and combining the defect development trend and performance decay rate, the remaining life of the membrane module is comprehensively predicted to determine the cycle length of the current membrane module that can still be stably operated.

[0072] If the predicted remaining life of the membrane module is less than the length of the next production cycle, a membrane module replacement selection process is started. At this time, the information of various membrane modules in the wastewater membrane material spare parts library is called, including different membrane pore size distribution, membrane material hydrophilicity and hydrophobicity parameters, module types corresponding to the winding membrane process parameters, combined with real-time collected wastewater quality parameters such as pollutant particle size and organic matter content, used to match the adapted membrane pore size and hydrophilicity and hydrophobicity, production line equipment state such as equipment compatible component size and operating pressure range, used to screen winding membrane process parameters that meet the requirements of equipment operation, membrane pollution prediction results such as future cycle pollution type, used to preferentially select membrane materials resistant to this type of pollution, and membrane defect structure prediction results (such as current defects mainly being pore blockage, used to avoid selecting small pore size membranes prone to blockage), and a multi-objective optimization algorithm is used to comprehensively score the membrane module configurations in the spare parts library. The scoring dimensions include adaptability to the next cycle working condition, flux improvement potential, energy consumption control ability, and anti-pollution ability, etc. The optimal membrane module configuration with the highest score is finally selected, and the membrane pore size distribution, membrane material hydrophilicity and hydrophobicity parameters, and winding membrane process parameters corresponding to the configuration are explicitly output, indicating to the user the membrane material and membrane module replacement parts needed for the next production cycle, ensuring that the replaced membrane module can fully adapt to the operating requirements of the next cycle, and avoiding process fluctuations or performance losses caused by improper component selection.

[0073] Thus, the association between macroscopic performance and microscopic defects is realized through time series modeling, microscopic structure analysis, stress identification and parameter inversion are completed with the help of graph neural networks and Bayesian optimization, and finally precise selection of replacement components is realized through multi-objective optimization, forming a full-process closed loop of performance modeling, defect prediction, life evaluation, and replacement selection. Thus, potential defects and life risks of membrane modules can be discovered in advance, and process downtime caused by sudden failure of membrane modules can be avoided.

[0074] In step S105, an optimization strategy for the wastewater treatment system is generated based on the multi-objective dynamic optimization result, and the membrane process operating parameters and production control parameters in the wastewater treatment system are dynamically adjusted.

[0075] By way of example, a reinforcement learning agent is deployed on the edge computing nodes. Then, in step S105, the multi-objective dynamic optimization result is obtained through collaborative calculation of multiple edge computing nodes, the adjustment instructions of the membrane process operating parameters and production control parameters are generated based on the multi-objective dynamic optimization result, and the instruction execution result is synchronized and updated to the digital twin model, which is used to monitor the execution result of the optimization strategy.

[0076] Further optionally, before step S105, embedding safety constraint conditions in the optimization strategy output by the reinforcement learning agent deployed on the edge computing node, the safety constraint conditions at least including the pressure limit of the membrane component, wherein the pressure limit of the membrane component is determined based on the equipment component attributes where the edge node is located. Then, when calculating the multi-objective dynamic optimization result in step S105, the Lagrangian multiplier method is used to inject the violation penalty term of the safety constraint condition into the Critic network of the reinforcement learning agent, and when the policy parameters predicted by the Critic network exceed the constraint range, the optimization direction is corrected through the penalty mechanism to generate an optimization strategy that satisfies the safety constraint. Specifically, when deploying the reinforcement learning agent on the edge computing node, each edge node will first identify the equipment component attributes connected to itself, such as the model, material, service life, etc. of the membrane component corresponding to the node, and determine the pressure limit of the membrane component based on these attributes. Different models of membrane components have different pressure limits due to differences in material strength and structural design, and the pressure limit of an old component may be reduced due to aging. The edge node will convert these specific limit values into quantifiable safety constraint conditions and embed them into the optimization strategy framework of the reinforcement learning agent to ensure that the agent considers these constraints as decision boundaries from the initial stage when generating strategies. When calculating the multi-objective dynamic optimization result in step S105, the Lagrangian multiplier method is used to integrate the violation penalty term of the safety constraint condition into the Critic network of the reinforcement learning agent. When evaluating the value of the policy parameters (such as operating pressure and water flow rate), the Critic network will compare the parameter values with the preset safety constraint range in real time. If the predicted operating pressure is close to or exceeds the pressure limit of the membrane component, the penalty term will take effect immediately, and the value score of the strategy will be reduced to weaken its possibility of being selected. The penalty strength will dynamically change with the degree of parameter exceeding the constraint, and the more it exceeds, the stronger the penalty will be, thereby strongly guiding the Actor network to adjust the parameter optimization direction and prompting it to explore better solutions within the safety range. Multiple edge computing nodes will perform collaborative calculation based on their respective local optimization results to verify the consistency and safety of the parameters. For example, the operating pressure output by a certain node may be within its own constraint range, but it may affect the safety of the membrane components of adjacent nodes. Collaborative calculation will perform global verification to ensure that the overall strategy satisfies the safety constraints of all nodes, and finally generate unified membrane process operating parameters and production control parameter adjustment instructions. After these instructions are issued to the wastewater treatment system for execution, the execution results will be synchronized in real time to the digital twin model, and the model will reproduce the system state after the execution of the instructions through the virtual-real synchronization feature. The operator can intuitively monitor whether the parameter adjustment is within the safety range and actually improve the optimization strategy for membrane flux, energy consumption, etc.

[0077] In this way, the embedding of safety constraints from the source delineates the physical boundaries of parameter optimization, avoiding the agent ignoring equipment safety in pursuit of multi-objective optimization. The Lagrangian multiplier method organically integrates safety goals with flux maximization, energy minimization, and other objectives through a penalty mechanism, allowing the optimization process to achieve a balance between multiple objectives while meeting safety constraints. The collaborative computing of edge nodes addresses the global safety conflict that may arise from local optimization, ensuring the safety and collaboration of the overall strategy. On the basis of ensuring that membrane components and other equipment are not damaged by excessive pressure, the effective optimization of membrane process operating parameters and production control parameters is achieved, avoiding equipment failure or process interruption caused by parameter over-limiting. At the same time, through real-time monitoring of the digital twin model, operators can timely grasp the execution effect of the strategy and intervene manually when necessary, further improving the safety and reliability of the wastewater treatment system operation.

[0078] It is worth noting that in addition to the Lagrangian multiplier method, a heuristic penalty method can also be used to inject a violation penalty term into the Critic network. This method directly adds a penalty term related to the degree of constraint violation to the Critic network's value evaluation function. When the predicted strategy parameters (such as operating pressure) exceed the safety constraint range (such as the pressure limit of the membrane component), a corresponding penalty value is set according to the amount of excess. The more the excess, the greater the penalty value, directly reducing the Critic network's value score for the strategy.

[0079] The embodiments of the present application can realize real-time intelligent adjustment of industrial wastewater membrane process operation, ensure the stability of membrane process operation, and improve the processing efficiency of membrane process operation.

[0080] After introducing the method of the exemplary embodiments of the present application, next, with reference to Figure 2An industrial wastewater membrane process intelligent optimization system based on multi-dimensional data modeling is described for an exemplary embodiment of the present application. A data acquisition module is configured to acquire wastewater quality parameters and production line equipment states of a target scene in real time through a multi-dimensional sensor network; the production line equipment states include production line control parameters, membrane process operation parameters, and real-time image data of a microstructure of a membrane material in a preset production node; a modeling module is configured to construct a digital twin model of a wastewater treatment system under the target scene based on the wastewater quality parameters and the production line equipment states; an optimization decision module is configured to identify a membrane surface pollution type of each node in the wastewater treatment system and predict a membrane pollution evolution situation in a future preset time period through the digital twin model, and obtain a membrane pollution prediction result; a multi-strategy evolutionary reinforcement learning algorithm driven by scenario memory is used to combine the wastewater quality parameters, the production line equipment states, and the membrane pollution prediction result, and a multi-agent collaborative decision algorithm is used to perform multi-objective dynamic optimization of membrane process operation parameters and production control parameters of the wastewater treatment system; a multi-agent collaborative structure in the multi-agent collaborative decision algorithm is constructed based on the optimization objectives of maximizing membrane flux, minimizing energy consumption, and prolonging chemical cleaning period; and an automatic control module is configured to generate an optimization strategy of the wastewater treatment system based on the multi-objective dynamic optimization result, and dynamically adjust the membrane process operation parameters and the production control parameters in the wastewater treatment system. The above system can implement each step described in the method embodiments, and the specific implementation of each step will not be repeated here. After introducing the method and system of the exemplary embodiments of the present application, next, a terminal device of an exemplary embodiment of the present application is described, which can implement each step described in the method embodiments, and the specific implementation of each step will not be repeated here.

[0081] It should be noted that the above-described embodiments are merely specific implementations of the present application, which are used to illustrate the technical solutions of the present application, rather than limit the same, and the protection scope of the present application is not limited thereto. Although the present application is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions described in the foregoing embodiments, or make equivalent replacements to some technical features thereof, within the technical scope disclosed by the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An industrial wastewater membrane process intelligent optimization method based on multi-dimensional data modeling, characterized in that, The method comprises: real-time acquisition of wastewater quality parameters and production line equipment states of a target scene through a multi-dimensional sensor network; the production line equipment states include production line control parameters, membrane process operation parameters, and real-time image data of a microstructure of a membrane material in a preset production node; construction of a digital twin model of a wastewater treatment system under the target scene based on the wastewater quality parameters and the production line equipment states; identification of a membrane surface pollution type of each node in the wastewater treatment system and prediction of a membrane pollution evolution in a future preset time period through the digital twin model, to obtain a membrane pollution prediction result; multi-objective dynamic optimization of the membrane process operation parameters and the production control parameters of the wastewater treatment system through a multi-agent collaborative decision algorithm, in combination with the wastewater quality parameters, the production line equipment states, and the membrane pollution prediction result, by using a scenario memory driven multi-strategy evolutionary reinforcement learning algorithm; a multi-agent collaborative structure in the multi-agent collaborative decision algorithm is constructed based on maximization of a membrane flux, minimization of energy consumption, and prolongation of a chemical cleaning period as optimization objectives; generation of an optimization strategy of the wastewater treatment system based on a multi-objective dynamic optimization result, for dynamic adjustment of the membrane process operation parameters and the production control parameters in the wastewater treatment system.

2. The method for intelligent optimization of industrial wastewater membrane process based on multi-dimensional data modeling according to claim 1, characterized in that, After the membrane pollution prediction result is obtained by identifying the membrane surface pollution type of each node in the wastewater treatment system and predicting the membrane pollution evolution in the future preset time period through the digital twin model, the method further comprises: selection of a corresponding pulse cleaning mode according to the membrane surface pollution type and the flux decay rate prediction value in the membrane pollution prediction result; the membrane surface pollution type includes organic matter adhesion and / or colloid blockage; adaptive adjustment parameters in the pulse cleaning mode include at least high-pressure backflush intensity, pulse frequency, chemical type, chemical concentration, and cleaning period; dynamic adjustment of the adaptive adjustment parameters in the pulse cleaning mode according to a real-time flux decay rate and a preset parameter threshold in the pulse cleaning mode, and real-time monitoring of a membrane pollution recovery rate to dynamically adjust a cleaning period of the pulse cleaning mode.

3. The method for intelligent optimization of industrial wastewater membrane process based on multi-dimensional data modeling according to claim 1, characterized in that, The construction of the digital twin model of the wastewater treatment system under the target scene based on the wastewater quality parameters and the production line equipment states comprises: establishment of a geometric model of a membrane assembly and a pipeline in the target scene through three-dimensional scanning, dynamic binding of the geometric model in combination with real-time acquisition of the wastewater quality parameters and the production line equipment states, generation of a virtual mapping system with virtual-real synchronization characteristics with the target scene, and formation of the digital twin model of the wastewater treatment system; embedding of a computational fluid dynamics module in the digital twin model, simulation of a shear force distribution state and a deposition motion path of a pollutant on a membrane surface, and restoration of a physical field of a membrane pollution process in the digital twin model.

4. The method for intelligent optimization of industrial wastewater membrane process based on multi-dimensional data modeling according to claim 3, characterized in that, The method of identifying the membrane surface pollution type of each node in the wastewater treatment system and predicting the membrane pollution evolution in the future preset time period through the digital twin model, to obtain the membrane pollution prediction result, comprises: extraction of real-time physical field simulation data from the digital twin model, wherein the real-time physical field simulation data includes a shear force distribution matrix of a pollutant on a membrane surface and deposition path trajectory data; aligning and fusing the real-time physical field simulation data, the real-time collected wastewater quality parameters, the operating pressure in the production line control parameters, the cleaning cycle, the transmembrane pressure difference in the membrane process operation parameters, the real-time flux, and the membrane surface defect image in the time-space dimension, to construct a multi-dimensional wastewater feature containing physical field characteristics; a hybrid model of convolutional neural network (CNN) and long short-term memory network (LSTM) is used to identify the membrane surface pollution type based on the multi-dimensional wastewater feature, wherein the membrane surface pollution type includes organic matter adhesion, colloid blockage, and / or composite pollution; combined with the identified membrane surface pollution type, the real-time physical field simulation data and the historical membrane process operation data, the hybrid model is used to predict the membrane flux decay rate and the pollution area expansion range in a future preset time period, and the pollution area expansion range is represented as the membrane surface pollutant area ratio and the three-dimensional space distribution coordinates corresponding to the predicted coverage range of the pollutant; the membrane surface pollution type, the membrane flux decay rate and the pollution area expansion range are taken as the membrane pollution prediction result.

5. The method for intelligent optimization of industrial wastewater membrane process based on multi-dimensional data modeling according to claim 4, characterized in that, The hybrid network at least includes an identification layer and a prediction layer; the hybrid model of convolutional neural network (CNN) and long short-term memory network (LSTM) is used to identify the membrane surface pollution type based on the multi-dimensional wastewater feature, including: in the identification layer, a convolutional neural network (CNN) embedded with physical field prior knowledge is used to jointly extract features from the membrane surface defect image and the shear force distribution matrix in the multi-dimensional wastewater feature; combined with the long short-term memory network (LSTM), the dynamically associated analysis is performed on the time-sequenced operation parameters and deposition path data in the joint feature extraction result to identify the membrane surface pollution type; combined with the identified membrane surface pollution type, the real-time physical field simulation data and the historical membrane process operation data, the hybrid model is used to predict the membrane flux decay rate and the pollution area expansion range in a future preset time period, including: a space-time attention Transformer network is introduced in the prediction layer to fuse the physical field evolution constraint; based on the virtual-real synchronous characteristics of the digital twin model, the shear force change gradient and the deposition path expansion rate conforming to the physical field evolution law are extracted from the real-time physical field simulation data based on the identified membrane surface pollution type, the extracted shear force change gradient and deposition path expansion rate are associated with the historical pollutant evolution characteristics in the historical membrane process operation data, and a pollutant diffusion constraint term derived from the CFD physical model is embedded in the network loss function; the space-time attention Transformer network is used to predict the membrane flux decay rate and the pollution area expansion range in a future preset time period.

6. The method for intelligent optimization of industrial wastewater membrane process based on multi-dimensional data modeling as claimed in claim 1, wherein, the multi-strategy evolutionary reinforcement learning algorithm driven by episodic memory is used to perform multi-objective dynamic optimization of the membrane process operation parameters and the production control parameters of the wastewater treatment system through a multi-agent collaborative decision-making algorithm based on the wastewater quality parameters, the production line equipment state and the membrane pollution prediction result, including: the episodic memory tree (EMT) is used to extract historical optimal triplets containing state, action and reward from the historical membrane process operation data, and provide guidance for the exploration process of the reinforcement learning agent in the sparse reward scenario; The TD3 algorithm is used to optimize the water inlet flow rate in the membrane process operation parameter and the operating pressure in the production control parameter in real time by combining the real-time collected wastewater quality parameters, the production line equipment state, and the predicted membrane pollution prediction results. The NSGA-III algorithm is used to optimize and solve the hydrophilic and hydrophobic parameters and the production line process combination by combining the membrane pore size distribution in the membrane microstructure parameter, and the optimized strategy parameters of the elite individuals are obtained. The Genetic Soft Update mechanism is used to migrate the optimized strategy parameters of the elite individuals in the evolutionary algorithm population to the reinforcement learning agent, and the final output multi-objective dynamic optimization result is obtained through multiple rounds of collaboration between the reinforcement learning agent and the evolutionary algorithm population, completing the multi-objective dynamic optimization of the membrane process operation parameters and the production control parameters.

7. The method of intelligent optimization of industrial wastewater membrane process based on multi-dimensional data modeling as claimed in claim 6 wherein, The multi-agent collaborative structure adopts a double-layer master-slave architecture, including a master agent and a slave agent; the master agent serves as a global optimizer, responsible for the decomposition and resource scheduling of the multi-objective optimization task, and the slave agent serves as a specialized executor, including a material optimization agent, a control agent, and a cleaning strategy agent. The TD3 algorithm is used to optimize the water inlet flow rate in the membrane process operation parameter and the operating pressure in the production control parameter in real time by combining the real-time collected wastewater quality parameters, the production line equipment state, and the predicted membrane pollution prediction results. The control agent constructs an action decision model based on the TD3 algorithm, and takes the real-time wastewater quality parameters, the current operating pressure, the water inlet flow rate, the transmembrane pressure difference, and the flux decay rate and pollution area expansion range in the future preset time length in the membrane pollution prediction results as state inputs. The action space is set to the water inlet flow rate adjustment range matching the membrane module type and the operating pressure adjustment range meeting the membrane module pressure standard, and the action discretization step is set according to the control accuracy requirement; in the reward function, the reward value is calculated by weighting the real-time flux improvement rate, the energy consumption growth rate, and the cleaning cycle extension rate, the weight of the real-time flux improvement rate is higher than that of the energy consumption growth rate and the cleaning cycle extension rate, and the weights of the energy consumption growth rate and the cleaning cycle extension rate are the same, the real-time flux improvement rate is calculated based on the difference between the predicted value and the current value of the digital twin model; The TD3 algorithm is used to optimize the water inlet flow rate in the membrane process operation parameter and the operating pressure in the production control parameter in real time by combining the real-time collected wastewater quality parameters, the production line equipment state, and the predicted membrane pollution prediction results.

8. The method of intelligent optimization of industrial wastewater membrane process based on multi-dimensional data modeling as claimed in claim 7, wherein, The Genetic Soft Update mechanism is used to migrate the optimized strategy parameters of the elite individuals in the evolutionary algorithm population to the reinforcement learning agent, and the final output multi-objective dynamic optimization result is obtained through multiple rounds of collaboration between the reinforcement learning agent and the evolutionary algorithm population, completing the multi-objective dynamic optimization of the membrane process operation parameters and the production control parameters. Screening of elite individuals with a fitness value meeting a preset ranking from an evolutionary algorithm population, extracting the strategy parameters corresponding to the elite individuals; the fitness value is calculated by the weighted score of membrane flux, membrane process energy consumption and anti-pollution; Extracting the optimization strategy parameters corresponding to the elite individuals, the optimization strategy parameters including membrane pore size distribution interval, hydrophilic-hydrophobic adjustment coefficient and production line process combination scheme, and weighting and fusing the optimization strategy parameters with the current strategy parameters of the reinforcement learning agent; dynamically adjusting the weight coefficient of the optimization strategy parameters with the iteration number; The fused strategy parameters are updated to the Actor network of the reinforcement learning agent through sliding average filtering processing.

9. The method for intelligent optimization of industrial wastewater membrane process based on multi-dimensional data modeling as claimed in claim 1, wherein, The multi-strategy evolutionary reinforcement learning algorithm driven by episodic memory, combined with wastewater quality parameters, production line equipment state, and membrane pollution prediction results, uses a multi-agent collaborative decision-making algorithm to perform multi-objective dynamic optimization of membrane process operation parameters and production control parameters of the wastewater treatment system, and further includes: Based on the multi-objective dynamic optimization results, a time series modeling method is used to construct a membrane process operation model within the current production cycle, wherein the inputs of the membrane process operation model are real-time collected wastewater quality parameters and production line equipment state, and the outputs are dynamic change curves of membrane flux and energy consumption; Combined with the dynamic change curves output by the membrane process operation model, the membrane surface defect position, membrane surface defect area and corresponding membrane surface microstructure change in the current production cycle are predicted to obtain the membrane defect structure prediction results; Based on the Bayesian optimization graph neural network model, the membrane surface microstructure is taken as the graph structure node, the stress concentration area caused by local blockage is identified combined with the membrane defect structure prediction results, and the optimal adjustment direction of the membrane material hydrophilic-hydrophobicity parameters is determined through inversion calculation taking the membrane flux decay rate and energy consumption as observation variables, and the remaining life of the membrane module is predicted; If the remaining life of the membrane module is less than the length of the next production cycle, the optimal membrane module configuration is selected from the wastewater membrane material spare parts library based on real-time collected wastewater quality parameters, production line equipment state, membrane pollution prediction results and membrane defect structure prediction results, the optimal membrane module configuration including membrane pore size distribution, membrane material hydrophilic-hydrophobicity parameters and membrane rolling process parameters, and is used to indicate the membrane material and membrane module replacement parts for the next production cycle to the user.

10. An industrial wastewater membrane process intelligent optimization system based on multi-dimensional data modeling, characterized in that, The system includes the following modules: data acquisition module, modeling module, optimization decision module and automatic control module; Wherein The data acquisition module is used to acquire the wastewater quality parameters and production line equipment state of the target scene in real time through a multi-dimensional sensor network; the production line equipment state includes production line control parameters, membrane process operation parameters, and real-time image data of the microstructure of the membrane material in the preset production node; The modeling module is used to construct a digital twin model of the wastewater treatment system under the target scene based on the wastewater quality parameters and production line equipment state; An optimization decision module is configured to identify the membrane surface pollution type of each node in the wastewater treatment system through the digital twin model and predict the membrane pollution evolution in a future preset time period to obtain a membrane pollution prediction result; a scenario memory driven multi-strategy evolutionary reinforcement learning algorithm is adopted to combine the wastewater quality parameters, the production line equipment state and the membrane pollution prediction result, and a multi-agent collaborative decision algorithm is used to perform multi-objective dynamic optimization on the membrane process operation parameters and the production control parameters of the wastewater treatment system; In the multi-agent collaborative decision algorithm, a multi-agent collaborative structure is constructed with the optimization objectives of maximizing the membrane flux, minimizing the energy consumption and prolonging the chemical cleaning cycle; An automatic control module is configured to generate an optimization strategy of the wastewater treatment system based on the multi-objective dynamic optimization result and dynamically adjust the membrane process operation parameters and the production control parameters in the wastewater treatment system.

Images