An ai multi-feature fusion diagnosis method and system for health status of a membrane module

Abstract

The application provides an AI multi-feature fusion diagnosis method and system for the health state of a membrane assembly, comprising the following steps: obtaining component features of wastewater to be treated, constructing a pulse coding sequence and a corresponding monitoring strategy; sequentially applying bidirectional pulse excitation to the membrane assembly and obtaining bidirectional response data sets; constructing a differential feature pair, combining wastewater working condition features and membrane assembly operation features to form a heterogeneous feature set; inputting the heterogeneous feature set into a preset AI fusion inference model to obtain a damage state vector of the membrane assembly; performing multi-task decoding based on the damage state vector to output a full-dimensional diagnosis result. According to the scheme, the weak damage signal of the membrane assembly is accurately captured by means of bidirectional pulse contrast collection, and the multi-type heterogeneous features and causal time sequence inference are combined to distinguish the external disturbance such as water quality fluctuation from the real damage of the membrane body, so that the sensitivity and diagnosis comprehensiveness of the membrane assembly health monitoring are significantly improved, the health attenuation change of the membrane assembly can be tracked throughout the whole process, and the health monitoring of the membrane assembly throughout the whole cycle is realized.

Description

Technical Field

[0001] This application relates to the field of wastewater treatment, and more specifically, to an AI multi-feature fusion diagnostic method and system for the health status of membrane modules. Background Technology

[0002] Membrane filtration systems are widely used in industrial wastewater treatment. During long-term operation, membrane modules are prone to damage problems such as membrane pore blockage, membrane surface adsorption and scaling, and filter cake accumulation in the flow channel. Damage to membrane modules will directly reduce the permeate flux and deteriorate the permeate water quality. In severe cases, it can cause the failure of a single membrane fiber or even the shutdown of the entire membrane module for maintenance.

[0003] Current conventional membrane module monitoring solutions rely solely on fixed unidirectional fluid sampling of single operating parameters such as differential pressure and flux to perform simple threshold judgments. This results in low overall monitoring sensitivity and a limited diagnostic scope. Faults can only be identified after severe membrane module damage and significant deviations in operating parameters, failing to predict the health degradation state of the membrane module in advance. Maintenance personnel can only rely on periodic disassembly and inspection to assess the internal health of the membrane module. This disassembly and inspection process is cumbersome and incurrs high downtime costs, making it difficult to achieve full-cycle, non-disassembly, and precise health monitoring of the membrane module. Summary of the Invention

[0004] Based on the problems existing in the prior art, this application provides an AI multi-feature fusion diagnostic method and system for the health status of membrane modules.

[0005] An AI-based multi-feature fusion diagnostic method for the health status of membrane modules, applied to a bidirectional flow membrane filtration device capable of switching between forward and reverse operating states, includes the following steps:

[0006] The component characteristics of the wastewater to be treated are obtained, and a pulse coding sequence and corresponding monitoring strategy are constructed for the membrane module to be monitored based on the component characteristics. Each coding unit of the pulse coding sequence contains a parameter combination of a positive detection pulse and a reverse control pulse. Bidirectional pulse excitation is sequentially applied to the membrane module according to the pulse coding sequence, and the multi-sensor array is controlled to perform forward and reverse differential synchronous acquisition according to the monitoring strategy to obtain the bidirectional response dataset corresponding to each coding unit; Feature extraction and differential operations are performed on the bidirectional response dataset to construct differential feature pairs. Combined with wastewater condition characteristics and membrane module operation characteristics, a heterogeneous feature set is formed. The heterogeneous feature set is input into a preset AI fusion inference model to perform cross-modal attention fusion and causal temporal inference to obtain the damage state vector of the membrane module; Multi-task decoding is performed based on the damage state vector, and a full-dimensional diagnostic result is output. An AI multi-feature fusion diagnostic system for the health status of membrane modules is applied to a bidirectional flow membrane filtration device that can switch between forward and reverse operating states. The system includes: The coding strategy unit is used to acquire the component characteristics of the wastewater to be treated, and to construct a pulse coding sequence and a corresponding monitoring strategy for the membrane module to be monitored based on the component characteristics. Each coding unit of the pulse coding sequence contains a parameter combination of a positive detection pulse and a reverse control pulse. A bidirectional excitation unit is used to sequentially apply bidirectional pulse excitation to the membrane module according to the pulse coding sequence, and control the multi-sensor array to perform forward and reverse differential synchronous acquisition according to the monitoring strategy to obtain a bidirectional response dataset corresponding to each coding unit. The differential parsing unit is used to perform feature extraction and differential operations on the bidirectional response dataset to construct differential feature pairs, and combine the wastewater condition characteristics with the membrane module operation characteristics to form a heterogeneous feature set; The model inference unit is used to input the heterogeneous feature set into a preset AI fusion inference model to perform cross-modal attention fusion and causal temporal inference to obtain the damage state vector of the membrane module. The diagnostic output unit is used to perform multi-task decoding based on the damage state vector and output full-dimensional diagnostic results.

[0007] Beneficial effects: This application proposes an AI multi-feature fusion diagnostic method and system for membrane module health status. It relies on bidirectional pulse comparison acquisition to accurately capture weak damage signals of membrane modules. It combines multiple heterogeneous features and causal time-series reasoning to distinguish between external disturbances such as water quality fluctuations and actual damage to the membrane itself. This significantly improves the sensitivity and comprehensiveness of membrane module health monitoring and diagnosis. It can continuously track changes in membrane module health degradation throughout the entire process, provide early warning of potential faults, eliminate the need for frequent disassembly and testing, and achieve full-cycle, refined health monitoring of membrane modules.

[0008] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the AI ​​multi-feature fusion diagnostic method of this application; Figure 2This is a schematic diagram illustrating the principle of the AI ​​multi-feature fusion diagnostic method of this application; Figure 3 This is a schematic diagram of the monitoring strategy and pulse-coded sequence acquisition process of this application; Figure 4 This is a schematic diagram of the damage state vector acquisition process in this application; Figure 5 This is a schematic diagram of the AI ​​multi-feature fusion diagnostic system module of this application.

[0011] Figure labeling: 1-Encoding strategy unit; 2-Bidirectional excitation unit; 3-Differential analysis unit; 4-Model inference unit; 5-Diagnostic output unit. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0013] This application proposes an AI-based multi-feature fusion diagnostic method for membrane module health status. First, a pulse coding sequence and adapted monitoring strategy are customized based on wastewater component characteristics. Bidirectional pulse excitation is implemented using coding units containing forward detection pulses and reverse control pulses. A multi-sensor array is used to complete forward and reverse differential synchronous acquisition to obtain a bidirectional response dataset. Differential feature pairs that eliminate fixed interference from pipeline valve groups are constructed through feature extraction and differential operations. Then, wastewater operating condition characteristics and membrane module operating characteristics are fused to form a standardized heterogeneous feature set. An AI fusion inference model performs cross-modal attention fusion and causal temporal inference to obtain a damage state vector. Finally, multi-task decoding outputs a full-dimensional diagnostic result. A schematic diagram of the AI ​​multi-feature fusion diagnostic method is attached. Figure 1 As shown in the attached diagram, the principle is as follows: Figure 2 As shown, the specific solution is as follows: An AI-based multi-feature fusion diagnostic method for the health status of membrane modules, applied to a bidirectional flow membrane filtration device capable of switching between forward and reverse operating states, includes the following steps: 101. Obtain the component characteristics of the wastewater to be treated, and construct a pulse coding sequence and corresponding monitoring strategy for the membrane module to be monitored based on the component characteristics. Each coding unit of the pulse coding sequence contains a parameter combination of a positive detection pulse and a reverse control pulse. 102. Apply bidirectional pulse excitation to the membrane module sequentially according to the pulse coding sequence, and control the multi-sensor array to perform forward and reverse differential synchronous acquisition according to the monitoring strategy to obtain the bidirectional response dataset corresponding to each coding unit; 103. Perform feature extraction and difference operations on the bidirectional response dataset to construct differential feature pairs, and combine the wastewater condition characteristics with the membrane module operation characteristics to form a heterogeneous feature set; 104. Input the heterogeneous feature set into the preset AI fusion inference model to perform cross-modal attention fusion and causal temporal inference to obtain the damage state vector of the membrane module; 105. Perform multi-task decoding based on the damage state vector and output full-dimensional diagnostic results.

[0014] The solution proposed in this application is applicable to specialized filtration devices, including membrane modules, fluid delivery units, pipeline valve groups, and electrical reversing control units. It can flexibly switch the fluid flow direction according to operational requirements, resulting in two stable working states with different functions. At the same time, the hardware structure and flow path design can be deeply matched with the bidirectional pulse excitation and synchronous data acquisition requirements of this method.

[0015] The forward operating mode is the mainstream operating condition for the device in daily operation, and it is also the working mode in which the membrane module completes the core function of solid-liquid separation. In this mode, the raw water to be treated flows in from the inlet pipe at the front end of the device and flows uniformly through the inside of the membrane module along the preset forward flow path. Impurities such as suspended particles, colloidal substances, heavy metal ions, and high molecular weight organic matter contained in the water are intercepted by the membrane layer and adhere to the membrane surface or remain inside the membrane pores. The clean water after membrane screening flows out from the product water pipe on the other side of the membrane module, completing the water treatment process.

[0016] The reverse operation mode primarily relies on the device's reversing structure to reverse the flow direction. The internal electrical reversing control unit drives the pipeline valve assembly to change its opening and closing status. The fluid delivery unit then adjusts the water flow direction, allowing water to flow from the original product water end into the membrane module and then out from the original feed water end. The overall flow path is completely opposite to the forward operation mode. In the device's routine operation and maintenance process, reverse water flow is mainly used for online flushing operations. The impact force of the reverse water flow removes loose impurities adhering to the membrane surface, slowing down the membrane fouling rate.

[0017] Step 101 is the initial stage of the diagnostic process, playing a core role in environmental perception and strategy formulation. First, the component characteristics of the wastewater to be treated are acquired; this step is fundamental to enabling targeted monitoring of the entire diagnostic method. Different types of industrial wastewater contain different kinds of impurities, and these impurities cause fundamentally different types and mechanisms of damage to membrane modules. For example, wastewater containing heavy metal ions easily causes rigid blockage of membrane pores, wastewater containing surfactants easily causes flexible adsorption and scaling on the membrane surface, and wastewater containing solid particles easily causes filter cake accumulation in the membrane channels. If a uniform monitoring strategy is used, it will be impossible to effectively excite characteristic signals corresponding to the damage type, leading to a significant decrease in monitoring sensitivity. After acquiring the wastewater component characteristics, a unique pulse coding sequence and corresponding monitoring strategy are constructed for the membrane module to be monitored based on these characteristics. The pulse coding sequence is a series of pre-set fluid excitation signals, and each coding unit contains two sets of parameter combinations: a forward probe pulse and a reverse control pulse. The forward probe pulse is used to excite potential damage inside the membrane module to generate a response signal, while the reverse control pulse is used to provide a reference signal from the same source. The corresponding monitoring strategy pre-plans the operating mode of the multi-sensor array based on the parameters of the pulse code sequence, ensuring accurate acquisition of the corresponding response data simultaneously with the application of pulse excitation. The core function of this step is to achieve dynamic adaptation between the diagnostic strategy and the wastewater conditions, making subsequent excitation and acquisition processes more targeted and laying the foundation for obtaining high-quality diagnostic data.

[0018] The control terminal connects to an external water quality detection hardware module to collect water quality parameters of the wastewater to be treated in real time. The hardware-detected data is analyzed to obtain the wastewater component characteristics. The terminal's built-in program matches the corresponding pulse parameter library based on these characteristics, generating a unique pulse coding sequence. This pulse coding sequence does not correspond to virtual electrical signals, but rather to fluid control parameters that can be recognized by the hardware, including fluid flow amplitude, pulse duration, and pulse interval parameters. The control terminal sends the coded parameters to the fluid pulse generation unit. This unit, relying on the coordinated operation of a variable frequency booster pump and a precision flow control valve, dynamically adjusts the pump's output power and valve opening to accurately generate forward detection fluid pulses and reverse control fluid pulses that meet the parameter requirements, realizing the hardware conversion from software coding strategy to physical fluid excitation. Simultaneously, the terminal generates a corresponding monitoring strategy, pre-setting the sensor's acquisition sequence, acquisition cycle, and operating status, providing a scheduling basis for subsequent hardware acquisition.

[0019] Step 102 is the active excitation and data acquisition stage, a crucial step in obtaining raw information about the membrane module's health status. The main control module strictly follows the pulse coding sequence constructed in step 101, sequentially applying bidirectional pulse excitation to the membrane module. When a forward detection pulse is applied, the fluid flows in the filtration direction during normal membrane module operation. Damage inside the membrane module will generate specific response signals under the influence of the forward fluid. When a reverse control pulse is applied, the fluid flows in the direction of reverse flushing of the membrane module. The response signals generated by non-membrane module components such as system pipelines and valve groups are essentially the same as during the forward pulse, while the response signals generated by membrane module damage will show significant differences. The application interval between the forward detection pulse and the reverse control pulse is strictly controlled within a very small range to ensure that the external operating conditions such as wastewater temperature, flow rate, and pressure remain consistent during both pulse applications. Simultaneously with the application of bidirectional pulse excitation, the main control module controls a multi-sensor array to perform forward and reverse differential synchronous acquisition according to a pre-defined monitoring strategy. Synchronous acquisition is achieved through global clock synchronization technology, ensuring that all sensors start acquisition at the same time, guaranteeing that the acquired data is aligned in the time dimension. The multi-sensor array simultaneously acquires response data from multiple physical dimensions, which together constitute a bidirectional response dataset corresponding to each coding unit. The bidirectional response dataset contains all response data under the action of a forward probe pulse and all baseline data under the action of a reverse control pulse, serving as the raw data source for subsequent feature processing and inference analysis.

[0020] The forward and reverse operation of the bidirectional flow membrane filtration device is achieved through hardware switching of the electrically controlled reversing valve assembly. The valve assembly receives commands from the control terminal and precisely switches the fluid pipeline path, ensuring a smooth switching process without pipeline blockage or flow fluctuations, guaranteeing consistency between the two pulse excitation conditions. During the forward detection pulse execution phase, the reversing valve assembly switches to the filtration path, and the fluid flows along the normal permeate filtration direction of the membrane module. The fluid pulse generation unit outputs forward fluid excitation with set parameters. During the reverse control pulse execution phase, the reversing valve assembly quickly switches to the backwash path, and the fluid flows along the reverse flushing direction of the membrane module, outputting reverse fluid excitation with consistent parameters but opposite flow direction. The entire hardware system is equipped with an industrial-grade hardware clock synchronization module, and all sensors are bound to the same clock reference. At the hardware level, this ensures that the sampling start time and sampling duration of all sensing units are aligned, resolving timing misalignment issues. A multi-sensor array is uniformly arranged on the inlet side of the membrane module, the permeate side, and the main pipeline path, comprehensively collecting physical response data of the fluid under pulse excitation. Finally, all hardware-collected data under both forward and reverse pulse conditions are summarized to form a complete bidirectional response dataset.

[0021] Step 103 is the feature processing and feature set construction stage, responsible for transforming the raw response data into effective features that reflect the health status of the membrane module. First, feature extraction is performed on the bidirectional response dataset obtained in Step 102. This process involves filtering and calculating key indicators that characterize different states of the membrane module from massive amounts of raw time-series data. The raw response data contains a large amount of noise and redundant information; directly using it for inference would significantly reduce diagnostic efficiency and accuracy. Feature extraction transforms high-dimensional raw data into low-dimensional effective features while retaining core information related to membrane module damage. After feature extraction, differential operations are performed to construct differential feature pairs. The basic principle of differential operations is to subtract the features corresponding to the forward detection pulse from the corresponding features of the same type in the reverse control pulse, aligning them one by one. Since the response signals generated by the inherent structures of the system pipelines and valve groups are essentially the same under forward and reverse pulse action, differential operations can effectively eliminate these fixed system interference components. Furthermore, damage inside the membrane module will produce different resistance effects on the forward and reverse fluids; therefore, differential operations will retain the differentiated features caused solely by membrane module damage. After constructing the differential feature pairs, wastewater condition features and membrane module operating features are further combined to form a heterogeneous feature set. Wastewater condition features characterize the impact of the current external environment on the membrane module's operating status, while membrane module operating features characterize the overall long-term operating status of the membrane module. Integrating these three types of features from different sources and with different physical meanings allows for a comprehensive reflection of the membrane module's true health status from multiple dimensions, providing rich and comprehensive input information for subsequent AI inference.

[0022] The raw data output by the hardware acquisition device is a continuous analog physical signal. This signal is converted into computable digital time-series data by the terminal hardware analog-to-digital converter, providing the hardware data foundation for subsequent feature extraction and differential operations. The physical signals acquired by different sensor hardware are dimensionally independent, corresponding to different physical properties such as fluid resistance, interface state, and fluid steady-state parameters. Therefore, the extracted features possess inherent heterogeneity, ensuring the physical authenticity and effectiveness of the heterogeneous feature set. Simultaneously, system-fixed interferences such as pipeline resistance and valve damping are inherent physical characteristics of the hardware equipment. Their parameters remain constant under both forward and reverse flow conditions. Therefore, inherent hardware interferences can be accurately eliminated through differential operations, retaining only the physical response differences caused by damage to the membrane module itself.

[0023] Heterogeneous feature sets refer to sets of features that originate from different sources, have different physical dimensions, different numerical units, and independent representational meanings, and cannot be directly combined for calculation. Among them, differential feature pairs are specific to membrane damage and represent equipment state parameters; wastewater operating condition features are external environmental parameters; and membrane module operating features are long-term time-series degradation parameters. These features complement each other, and there is no feature redundancy. Traditional monitoring only uses a single homogeneous operating parameter, which cannot simultaneously account for the three influencing factors: environmental interference, long-term aging, and instantaneous damage, resulting in an inherent lack of monitoring dimensions.

[0024] Step 104 is the AI ​​fusion reasoning stage, the core intelligent part of the entire diagnostic method, responsible for deep analysis and reasoning of the heterogeneous feature set. First, the heterogeneous feature set formed in step 103 is input into the pre-trained AI fusion reasoning model. This model first performs a cross-modal attention fusion operation. Cross-modal attention fusion can automatically learn the correlation between different types of features and assign different weights to different features based on their contribution to the diagnosis of membrane module damage. Features that effectively reflect membrane module damage are given higher weights by the model. Interfering features unrelated to damage are given lower weights or even ignored. Through cross-modal attention fusion, the scattered information in the heterogeneous feature set can be integrated into a unified comprehensive feature, fully exploring the complementary information between different features. After completing cross-modal attention fusion, the model performs a causal temporal reasoning operation. Causal temporal reasoning not only analyzes the feature data at the current moment but also combines historical operating data of the membrane module over a period of time to analyze the development process and evolution of membrane module damage. Through causal temporal reasoning, it is possible to effectively distinguish between instantaneous operating condition fluctuations and actual membrane module damage, and at the same time reveal the causal relationship between different factors on membrane module damage. After the above two steps of reasoning, the model finally outputs the damage state vector of the membrane module. The damage state vector is a multi-dimensional numerical vector, where the value of each dimension corresponds to a certain aspect of the health status information of the membrane module, which is the basis for subsequent multi-task decoding.

[0025] Cross-modal attention fusion can dynamically adjust feature weights based on real-time operating conditions, enhancing key features and weakening interfering features. Causal temporal reasoning relies on the physical mechanism of damage occurrence to identify the root cause of data fluctuations, distinguishing between short-term parameter changes caused by instantaneous water quality fluctuations and long-term parameter shifts caused by irreversible damage to the membrane module. The damage state vector is a standardized multi-dimensional state representation vector generated after fusing all heterogeneous features. The values ​​of each dimension within the vector correspond to core implicit state information such as the degree of membrane module damage, damage response characteristics, and state decay rate. It does not directly display diagnostic results but is refined intermediate state data obtained from model reasoning, providing comprehensive and detailed raw state basis for multi-task parallel decoding.

[0026] Step 105 is the result output stage, responsible for transforming the intermediate results output by the AI ​​model into comprehensive diagnostic results that maintenance personnel can directly understand and use. The main control module performs multi-task decoding based on the damage state vector obtained in step 104. Multi-task decoding refers to processing multiple interrelated diagnostic tasks simultaneously and in parallel, rather than processing a single task sequentially. Each diagnostic task extracts corresponding information from the damage state vector to complete a specific dimension of diagnostic analysis. Through multi-task decoding, multiple dimensions of membrane module health status information can be obtained simultaneously. Finally, the outputs of all decoding tasks are integrated to output comprehensive diagnostic results. These comprehensive diagnostic results fully reflect the current health status of the membrane module, providing accurate basis for maintenance personnel to formulate scientific and reasonable maintenance strategies, avoiding the limitations of traditional monitoring methods that can only provide single fault alarms.

[0027] Steps 104 and 105 are implemented using the built-in computing hardware of the industrial control terminal. The terminal is equipped with a high-performance computing chip, which can perform algorithm calculations for cross-modal attention fusion, causal temporal reasoning and multi-task decoding in real time without the need for external offline computing devices. It can realize online real-time diagnosis of the health status of membrane components at the hardware level, avoid the data lag problem caused by offline analysis, and ensure the real-time and effective nature of the diagnostic results.

[0028] In some embodiments, the component characteristics of the wastewater to be treated include at least the coordination characteristics of organic complexed heavy metals. When the component characteristics indicate enrichment of organic complexed heavy metals, a bidirectional pulse with a sawtooth waveform is configured in the pulse coding sequence. Relying on the pulse rise and fall timing variation structure to match the excitation characteristics of membrane pore occlusion damage, a dielectric sensing channel and a differential pressure sensing channel are configured in the monitoring strategy to continuously collect dielectric response and fluid resistance response data inside the membrane pores. The monitoring strategy and the pulse coding sequence acquisition process are attached. Figure 3 As shown.

[0029] This embodiment is specifically adapted to industrial wastewater scenarios containing organically complexed heavy metals, and is used to accurately detect membrane pore clogging damage induced by such pollutants, effectively compensating for the shortcomings of conventional monitoring methods in identifying hidden damage within membrane pores. Organically complexed heavy metal pollutants are not free metal ions, but rather stable composite colloidal particles formed through coordination. The size of these particles is highly matched to the micropore diameter of the membrane module. During long-term wastewater filtration, these colloidal particles easily penetrate into the membrane pores and continuously accumulate. Ultimately, this causes membrane pore shrinkage, blockage, and decreased flow performance, resulting in irreversible membrane pore clogging damage. This type of damage occurs within the microporous structure of the membrane, and in the early stages, there are no obvious abrupt changes in flux or pressure differential, making it difficult to effectively identify using conventional single-parameter monitoring methods.

[0030] When the main control module detects that the wastewater composition exhibits an enrichment state of organic complexed heavy metals, it will specifically configure a bidirectional pulse with a sawtooth waveform as the excitation signal. Unlike conventional constant waveform pulses, the sawtooth waveform possesses a continuously rising and falling temporal dynamic change characteristic, allowing fluid pressure and flow velocity to rise and fall smoothly according to a preset time sequence. This dynamic change characteristic can fully adapt to the excitation pattern of membrane pore clogging damage. The obstructive effect of heavy metal complex particles accumulated inside the membrane pores on the fluid will show regular differences with the dynamic changes in fluid parameters. This can effectively amplify the weak response signal caused by microscopic membrane pore blockage, overcoming the limitation that static pulses cannot excite early clogging damage characteristics. Simultaneously, this pulse continues the structural design of the basic coding unit, adopting a bidirectional combination of forward detection pulses and reverse control pulses to ensure parameter uniformity and opposite flow directions. This provides a stable common-source reference for subsequent differential data processing, accurately eliminating fixed system interference from equipment pipelines and valve groups.

[0031] Simultaneously with the pulse excitation configuration, a dedicated monitoring strategy is updated, enabling the dielectric sensing channel and differential pressure sensing channel to collaboratively conduct full-time data acquisition. The differential pressure sensing channel continuously acquires fluid resistance response data. Membrane pore blockage directly reduces the fluid flow cross-section, altering the overall flow path resistance of the membrane module. During the dynamic fluid changes of the sawtooth waveform pulse, the differential pressure sensing channel can fully capture the dynamic fluctuation characteristics of the flow path resistance, providing feedback on the severity of membrane pore blockage from a macroscopic fluid perspective. The dielectric sensing channel continuously acquires dielectric response data within the membrane pores. The membrane itself possesses fixed dielectric parameters, and changes in the internal medium directly alter the dielectric response value. When heavy metal complex particles embed into the membrane pores, the internal medium structure changes, resulting in a significant shift in the dielectric parameters, allowing for a direct microscopic characterization of the particle accumulation state within the membrane pores. The two types of sensing channels simultaneously and continuously collect data, enabling complementary acquisition of macroscopic resistance characteristics and microscopic dielectric characteristics to form complete operating condition response data. This provides accurate and targeted raw data support for subsequent feature extraction, differential calculation, and AI intelligent diagnosis, significantly improving the early identification accuracy of heavy metal-induced membrane pore blockage damage.

[0032] In some embodiments, the component characteristics of the wastewater to be treated include at least the polymerization characteristics of polymeric surfactants. When the component characteristics indicate enrichment of polymeric surfactants, a bidirectional pulse of gradient square wave is configured in the pulse-coded sequence. Relying on the pulse-step flow velocity change structure to match the excitation characteristics of adsorption damage on the membrane surface, an interfacial tension sensing channel and a fluid temporal sensing channel are configured in the monitoring strategy to continuously collect data on membrane interfacial viscous damping and fluid steady-state offset response. The monitoring strategy and the pulse-coded sequence acquisition process are attached. Figure 3 As shown.

[0033] In industrial wastewater systems, polymeric surfactants possess unique polymerization and adhesion properties, allowing them to polymerize in water to form large molecular micelle structures. Unlike membrane pore-clogging damage caused by heavy metals, polymeric surfactants do not penetrate deep into the pores but instead continuously adhere and accumulate on the surface of the membrane module. Over long-term operation, this forms a flexible adsorption layer, directly altering the fluid adhesion characteristics and flow stability of the membrane interface, resulting in surface adsorption damage. This type of damage is a flexible surface damage, lacking obvious characteristics such as pore blockage or sudden pressure changes. Conventional monitoring methods using constant fluid parameters cannot elicit effective damage signals, making it highly susceptible to missed detections and misdiagnosis.

[0034] When analyzing the characteristics of wastewater components and determining that the water body is enriched with high-molecular-weight surfactants, the excitation waveform of the pulse-coded sequence is actively replaced, and a bidirectional pulse with a gradient square wave structure is configured to complete the fluid excitation. The core characteristic of the gradient square wave is the step-like graded velocity change. The fluid velocity and pressure rise and fall and stabilize step by step, unlike the continuous and smooth changes of the sawtooth wave. This step-like dynamic structure is highly adapted to the excitation mechanism of adsorption damage on the membrane surface. The high-molecular-weight adsorption layer attached to the membrane surface has viscous damping characteristics, and its response to abrupt or uniform fluid changes is minimal. However, it will produce significant resistance hysteresis and flow deviation phenomena in the case of step-like gradual changes in fluid conditions. The gradient square wave can disturb the flexible adsorption layer at the membrane interface layer by layer, fully amplifying the differences in fluid response corresponding to different adsorption thicknesses and different adsorption densities, and accurately exciting the unique characteristic signals of adsorption damage on the membrane surface. Meanwhile, the pulse strictly follows the architecture of the basic coding unit, retaining the bidirectional combination of the forward detection pulse and the reverse control pulse, ensuring that the excitation parameters are unified and the flow direction is opposite. Relying on the bidirectional reference control structure, interference from fixed systems such as pipelines and valve groups is eliminated, ensuring the accuracy of subsequent differential data processing.

[0035] Based on a matched dedicated excitation pulse, the monitoring strategy is updated synchronously, and the interfacial tension sensing channel and the fluid timing sensing channel are activated in a targeted manner to achieve the directional acquisition of adsorption damage characteristics on the membrane surface. The interfacial tension sensing channel is specifically used to capture viscous damping response data at the membrane interface. When there is no adsorption layer on the membrane surface, the viscous parameters at the membrane interface are stable and fixed. When a polymeric surfactant forms an adsorption layer, the viscous resistance and interfacial tension at the membrane interface will continuously shift. This sensing channel can detect microscopic interfacial viscosity changes in real time, directly characterizing the severity of adsorption on the membrane surface. The fluid timing sensing channel is specifically used to collect steady-state displacement response data of the fluid. The adsorption layer on the membrane surface disrupts the steady-state characteristics of fluid flow, causing temporal anomalies such as flow velocity lag and flow displacement under constant excitation. This channel can record the temporal changes in fluid flow throughout the entire process, providing feedback on the flow disturbances caused by adsorption at the membrane interface from a macroscopic flow perspective.

[0036] Two types of sensing channels continuously and synchronously acquire data throughout the entire process, covering all temporal stages of gradient square wave changes, and fully recording the microscopic interface parameters and macroscopic fluid parameters corresponding to the operating conditions, forming a dedicated bidirectional response dataset. This embodiment achieves directional excitation and precise acquisition of flexible adsorption damage on the membrane surface through precise matching of pollutant characteristics, excitation waveforms, and sensing acquisition strategies. This effectively compensates for the shortcomings of conventional monitoring methods in identifying surface flexible damage, significantly improving the precision and diagnostic accuracy of membrane module health monitoring under complex organic wastewater conditions.

[0037] In some embodiments, within each encoding unit, the forward probe pulse is the excitation applied along the flow path corresponding to the forward operating state of the membrane module, and the reverse control pulse is the excitation applied along the flow path corresponding to the reverse operating state of the membrane module. The reverse control pulse provides a common reference for the response data corresponding to the forward probe pulse. During the execution of forward and reverse differential synchronous acquisition, the forward probe pulse and reverse control pulse applied in the current encoding unit are used as reference signals to extract the same frequency response component that matches the excitation frequency of such pulses.

[0038] This embodiment clearly defines the dual-pulse excitation logic within the pulse coding unit. In each independent coding unit, two types of pulse excitation signals with different flow directions are configured. The core excitation parameters such as amplitude, frequency, waveform, and duration of the two types of pulses remain consistent; only the fluid application paths are distinguished. The forward detection pulse is applied to the forward working flow path of the membrane module's conventional permeate, primarily used to excite the membrane module's response signal under normal filtration conditions, carrying information about the membrane module's true health status and damage characteristics. The reverse control pulse is applied to the reverse working flow path of the membrane module's backwashing, not for direct damage detection, but specifically as a reference signal.

[0039] The two types of pulses are applied sequentially within the same coding unit, ensuring uniform operating conditions and equipment status, thus achieving the technical effect of common-source reference. Since the membrane filtration system experiences relatively constant performance in both the forward and reverse flow paths due to fixed interferences such as pipeline resistance, valve losses, and inherent equipment response deviations, the response data acquired by the reverse control pulse can accurately characterize the baseline level of inherent system interference. Using this as a reference, fixed equipment interference can be precisely isolated, and the differentiated responses caused by membrane module damage can be screened out separately, fundamentally solving the technical deficiency of traditional unidirectional monitoring in distinguishing between equipment interference and membrane damage.

[0040] Building upon the bidirectional pulse excitation deployment, this embodiment further defines the signal purification strategy during the forward and reverse differential synchronous acquisition process. The data acquisition process no longer directly accepts the raw response data, but instead uses the forward probe pulse and reverse control pulse applied in real-time by the current encoding unit as precise reference signals. Based on the fixed frequency characteristics of the excitation signal, the synchronous response component matching the pulse excitation frequency in the acquired signal is selectively extracted. Industrial water treatment sites commonly experience broadband random noise caused by mechanical vibration, circuit clutter, and water quality fluctuations. This type of noise has a chaotic and irregular spectrum, which can severely mask weak early membrane damage response signals.

[0041] By employing frequency-locked loop purification, the effective membrane response signal induced by pulse excitation can be accurately preserved, while irrelevant broadband interference and noise signals are filtered out, significantly improving the signal-to-noise ratio of the acquired data. This embodiment utilizes a combined design of bidirectional co-source reference excitation and co-frequency signal purification to eliminate interference from operating condition variables at the excitation end and purify effective damage signals at the acquisition end. This greatly enhances the purity and effectiveness of the bidirectional response dataset, providing reliable data support for subsequent high-precision differential calculations, heterogeneous feature fusion, and early-stage weak damage diagnosis of membrane modules.

[0042] In some embodiments, the process of obtaining differential feature pairs includes: parsing the forward response dataset corresponding to the forward probe pulse and the reverse response dataset corresponding to the reverse control pulse from the bidirectional response dataset, and extracting three types of dimensional features—time domain features, frequency domain features, and energy domain features—from the forward response dataset and the reverse response dataset; in the process of constructing differential feature pairs, matching the same-dimensional features extracted from the forward response dataset and the reverse response dataset one-to-one, performing the same-dimensional difference operation, eliminating the fixed response deviation components caused by the pipeline and valve group in the bidirectional flow path, and retaining only the differentiated features caused by the asymmetric damage of the membrane module, thus completing the structured construction of differential feature pairs.

[0043] First, the original bidirectional response dataset was split and analyzed, and divided into two independent subsets according to the type of excitation pulse. One subset is the positive response dataset generated under the action of the positive probe pulse. This dataset integrates the inherent response of equipment such as pipeline valve groups as well as the comprehensive response information brought about by the health status and damage status of the membrane module itself. The other subset is the reverse response dataset generated under the action of the reverse control pulse. This dataset mainly reflects the benchmark response law of the same equipment structure under the reverse flow path. Both datasets are collected and generated by the same set of coding units, and the external operating conditions and equipment operating status are consistent, which provides the basic conditions for data comparison and analysis.

[0044] After splitting the dataset, multi-dimensional feature extraction is performed on both the forward and reverse response datasets, extracting three core feature categories: time-domain features, frequency-domain features, and energy-domain features. Time-domain features record the real-time changes in the signal over time, intuitively reflecting the dynamic fluctuations of fluid flow and medium state throughout the pulse excitation process. They can capture real-time resistance changes caused by localized blockage and surface adsorption in the membrane module. Frequency-domain features are the frequency distribution information obtained after transforming the original signal. Different types of membrane damage will cause differentiated frequency fluctuations in the response signal. Frequency-domain features can distinguish between regular vibrations caused by damage and random noise interference. Energy-domain features mainly statistically analyze the energy loss during pulse excitation transmission. When the membrane module is damaged, obstructed fluid flow causes additional energy consumption. Energy-domain features can quantify the degree of this loss, characterizing the severity of the damage from the perspective of energy change. These three types of features interpret the response data from different physical perspectives, complementing each other and avoiding information gaps caused by incomplete coverage of a single feature category.

[0045] After obtaining complete multi-dimensional features, differential operations are initiated to complete the structured construction of differential feature pairs. The operation process strictly follows the principle of one-to-one matching within the same dimension. That is, the time-domain features extracted from the forward dataset are only operated on with the time-domain features of the reverse dataset. Frequency domain features and energy domain features are also processed according to the same rules to ensure the consistency of the operation logic and prevent invalid data from being generated by cross-dimensional operations. From the working principle perspective, pipelines and valve assemblies are fixed structures of the entire fluid system. Regardless of whether the fluid flows along the forward or reverse flow path, the resistance and signal deviation generated by these components remain basically stable. By performing difference calculations on values ​​within the same dimension, the response deviations of this fixed system can be completely eliminated. However, membrane module damage mostly exhibits asymmetrical distribution characteristics. Problems such as membrane pore blockage and surface adsorption are often concentrated in local areas. When the fluid flows forward and backward, the resistance effect of the damaged parts on the fluid is significantly different. This difference is not eliminated by differential operations and will be completely retained. After the above operation processing, the final differential feature pairs no longer contain inherent interference components of the equipment and only retain effective features that can truly reflect the asymmetric damage of the membrane module.

[0046] In some embodiments, the extracted instantaneous time-domain features include the signal steady-state response amplitude, time-series attenuation characteristics, and step response offset characteristics; the frequency-domain features include the response fundamental frequency component, inherent frequency offset characteristics, and spectral energy distribution ratio; and the energy-domain features include the impulse response cumulative energy and energy dissipation distribution characteristics.

[0047] The steady-state response amplitude refers to the constant output value of the signal during the stable phase of pulse excitation. It objectively reflects the steady-state resistance and media response level of fluid flow in the membrane module, and can directly demonstrate the attenuation of steady-state flow performance caused by membrane pore blockage and surface adsorption. The time-series decay characteristic characterizes the rate at which the fluid response signal dissipates over time after the pulse excitation ends. In a healthy membrane module, the fluid damping is uniform and stable, and the signal decay curve is regular and smooth. However, damaged membrane modules will exhibit erratic decay rates and delayed decay due to abnormal local resistance. The step response offset characteristic specifically captures the instantaneous signal abrupt deviation during pulse rise and fall switching. It is highly sensitive to local asymmetric damage in the membrane module and can capture early, weak, instantaneous damage signals that cannot be identified by conventional steady-state parameters, compensating for the insufficient ability of steady-state monitoring to identify subtle damage.

[0048] The fundamental frequency component of the response is the effective dominant frequency signal that matches the bidirectional pulse excitation frequency. It can accurately filter out the effective response information induced by the pulse excitation and filter out random broadband noise caused by field equipment vibration and circuit clutter. The inherent frequency offset characteristic refers to the degree of offset of the inherent vibration frequency of the membrane fluid system. When the membrane module flow path structure is intact, the fluid vibration frequency remains constant. Damage such as membrane pore blockage and membrane surface adsorption will change the flow path structure and fluid constraint state, causing a regular shift in the inherent frequency. The spectral energy distribution ratio is used to statistically analyze the signal energy distribution ratio in different frequency ranges. After damage to the membrane module, signal energy will diffuse from the fundamental frequency to the cluttered frequencies, showing a decrease in the fundamental frequency energy ratio and an increase in the cluttered frequency band energy ratio, which can quantify the existence of damage.

[0049] Pulse response cumulative energy refers to the total effective response energy of the fluid acting on the membrane module within a single complete pulse excitation cycle. Damage to the membrane module increases fluid flow resistance, resulting in increased ineffective energy loss and reduced effective cumulative energy. Energy dissipation distribution characteristics are used to characterize the energy consumption distribution ratio of the entire flow path system. They can distinguish the changes in the proportion of pipeline dissipation, valve body dissipation, and membrane body dissipation, and accurately pinpoint the additional energy loss caused by the membrane module itself.

[0050] The time-domain, frequency-domain, and energy-domain features extracted from the forward response data are rigorously matched one-to-one with the corresponding features of the reverse response data to ensure uniformity in computational dimensions and feature types. Since hardware structures such as pipelines and valve assemblies are fixed components of the system, the resistance deviations, signal offsets, and energy losses they generate remain essentially consistent in both the forward and reverse flow paths. After alignment and differential computation, these fixed system deviations can be canceled out. However, membrane module damage such as pore blockage and surface adsorption exhibits spatial asymmetry, resulting in significant differences in the obstruction effect experienced by the fluid during forward and reverse flow. This differential signal caused by membrane damage is not canceled out and can be completely and accurately preserved.

[0051] In some embodiments, wastewater operating characteristics include water quality component concentration shift characteristics, fluid temperature adaptation characteristics, and acid / alkali environment steady-state characteristics; membrane module operating characteristics include permeate flux decay trend characteristics, transmembrane flow resistance fluctuation characteristics, and pipeline fluid steady-state shift characteristics.

[0052] This embodiment explicitly defines the auxiliary feature dimensions of the heterogeneous feature set. Based on the differential damage characteristics of the membrane body represented by the differential feature pairs, it supplements the structure with two major auxiliary feature systems: wastewater operating condition characteristics and membrane module operation characteristics. Differential feature pairs can only reflect the relative damage differences of membrane modules under forward and reverse pulse excitation, failing to reflect real-time water quality environmental disturbances and long-term equipment operating baseline shifts, resulting in a single diagnostic dimension. This embodiment improves the modal integrity of the heterogeneous feature set by introducing multi-dimensional operating condition and operation characteristics. This provides environmental and operational benchmarks for subsequent cross-modal attention fusion, causal temporal reasoning, and multi-task decoding, effectively avoiding misdiagnosis and missed diagnosis caused by water quality fluctuations and equipment baseline drift, and comprehensively improving the stability and accuracy of membrane module health diagnosis.

[0053] Wastewater operating condition characteristics characterize the external environmental features of the membrane module's real-time operating water environment, used to quantify external disturbance variables during the diagnostic process. These characteristics comprise three core sub-categories. The first is water component concentration shift characteristics. This feature quantifies the deviation of the real-time concentration of various target pollutants in the wastewater from the normal baseline operating conditions, accurately reflecting the enrichment degree and fluctuation state of damage-inducing factors such as heavy metals and surfactants. In subsequent model inference, this feature provides the AI ​​model with operating condition tracing evidence, helping the model distinguish between signal anomalies caused by instantaneous pollutant concentration fluctuations and feature shifts caused by permanent damage to the membrane module itself, fundamentally avoiding misjudgments of faults caused by water quality disturbances. The second is fluid temperature adaptation characteristics. This feature characterizes the steady-state level and dynamic fluctuation range of the wastewater's real-time temperature. Water temperature directly affects water viscosity, fluid flowability, and the physical response characteristics of the membrane material, and is a crucial environmental factor altering the baseline of the sensing signal. In subsequent feature standardization and cross-modal fusion stages, this feature can be used to unify the feature dimensions under different temperature conditions, correct baseline shift errors caused by temperature, and ensure consistency in damage identification standards under different operating conditions. The third category is the steady-state characteristics of acid-base environments. These characteristics characterize the stable state and degree of pH deviation in wastewater. Acid-base environments directly affect the polymerization and adsorption characteristics of pollutants and the charge characteristics of membrane surfaces, indirectly affecting the formation rate and manifestation of membrane damage. In subsequent causal time-series reasoning, these characteristics can serve as environmental prior constraints, assisting the model in analyzing the causes of damage such as membrane surface adsorption and membrane pore blockage, thereby improving the traceability of diagnostic results.

[0054] Membrane module operating characteristics are the intrinsic baseline features characterizing the long-term operating status of the equipment. They are used to quantify baseline deviations caused by equipment aging and long-term operational drift, and also include three core sub-features. The first type is the permeate flux decay trend feature. This feature records the continuous decay pattern of water permeate during the long-term operation of the membrane module, which is different from the instantaneous flux value and focuses on reflecting the cumulative state of overall membrane aging and long-term clogging. In subsequent time-series causal inference, this feature can effectively distinguish between temporary reversible clogging and permanent membrane damage, providing core time-series basis for the model to quantify the severity of damage and predict the damage decay trend. The second type is the transmembrane flow resistance fluctuation feature. This feature characterizes the dynamic fluctuation pattern of the transmembrane pressure difference upstream and downstream of the membrane module, which can intuitively reflect the change in the flow resistance of the overall flow path of the membrane. This feature can complement the local micro-damage features characterized by differential features, forming a high-low dimension complement, allowing the AI ​​model to simultaneously grasp the local differential damage of the membrane module and the overall flow path resistance change, greatly improving the accuracy of damage quantification. The third category is the pipeline fluid steady-state offset feature. This feature is used to characterize the fluid steady-state baseline deviation generated by the long-term operation of the entire fluid pipeline system, corresponding to the long-term aging error of equipment that cannot be eliminated by differential calculation. In the subsequent feature fusion process, this feature can achieve secondary correction of the system baseline, further removing interference caused by pipeline aging and equipment drift, and maximizing the purification of damage features that belong only to the membrane module itself.

[0055] In some embodiments, within a pre-defined AI fusion inference model, for heterogeneous feature sets, firstly, single-modal deep semantic features are extracted using a parallel modality-specific network; then, a directional interactive fusion of cross-modal features is achieved through an inter-modal causal attention mechanism; finally, the fused features are mapped to a prior causal graph constructed based on the composite damage physical mechanism, and temporal causal inference is performed to obtain the damage state vector. The damage state vector acquisition process is attached. Figure 4 As shown.

[0056] This embodiment specifically defines a three-layer progressive inference architecture for a pre-defined AI fusion inference model, which is the core and key step in realizing intelligent parsing of heterogeneous feature sets and diagnosis of physical mechanism constraints. This step inherits the multimodal heterogeneous feature set constructed at the front end, abandoning the traditional crude processing method of simple feature splicing and weighted fusion. Through a progressive logic of single-modal deep mining, cross-modal directional fusion, and physical mechanism temporal inference, it transforms discrete, multidimensional, and heterogeneous basic features into standardized, quantifiable, and decodeable membrane component damage state vectors. The entire process takes into account both data feature mining capabilities and actual damage physical mechanisms, effectively solving the shortcomings of traditional intelligent diagnostic models such as insufficient feature utilization, mixed modal interference, lack of physical basis for inference results, and inaccurate early damage identification.

[0057] The first level of the model execution is single-modal deep semantic feature extraction, employing multiple parallel modality-specific networks to process the heterogeneous feature set in different domains. The heterogeneous feature set contains three types of modal data with different physical properties: differential features characterizing membrane damage differences, wastewater operating condition features characterizing external environmental disturbances, and membrane module operating features characterizing the long-term state of the equipment. The dimensions, variation patterns, and physical meanings of these three types of features are different, making it impossible to extract effective information uniformly through a general network. Therefore, this embodiment configures a dedicated feature extraction network independently for each type of modality feature. These networks operate in parallel, without interference, and each performs its specific function. The dedicated network can perform in-depth mining of the data characteristics of the corresponding modality, discarding shallow redundant information and random noise in the original features, and extracting the deep semantic features hidden within the original data. Among them, the differential modal network focuses on extracting the specific semantics of subtle asymmetric damage to the membrane module, the operating condition modal network focuses on extracting the environmental disturbance semantics of water quality, temperature, and acid-base fluctuations, and the running modal network focuses on extracting the temporal semantics of membrane module aging and degradation and operating baseline shift, providing pure, effective, and domain-independent single-modal deep features for subsequent cross-modal fusion.

[0058] The second level of model execution involves intermodal causal attention and cross-modal directional interaction fusion. After obtaining three independent single-modal deep semantic features, the model no longer performs indiscriminate feature splicing and fusion, but instead introduces an intermodal causal attention mechanism to achieve precise interaction. This mechanism relies on the physical correlation logic of membrane module damage, automatically calculating the causal correlation weights between different modal features, and distinguishing between effective and ineffective correlations between features. For feature combinations with physical mechanism correlations, such as water concentration shifts and membrane adsorption damage, or transmembrane resistance fluctuations and membrane pore blockage damage, the model automatically assigns higher attention weights to strengthen information interaction and complementary fusion between features. For random noise features without physical correlations and stray signals from abnormal operating conditions, the model suppresses their weights and blocks ineffective information interaction. Through this directional fusion method, false feature interference caused by external operating condition fluctuations can be accurately removed, retaining the effective fusion information corresponding to the actual damage of the membrane module, and finally integrating to obtain a global comprehensive fusion feature that takes into account environmental conditions, equipment operating baseline, and membrane damage status.

[0059] The third level of model execution involves temporal causal reasoning constrained by physical mechanisms, used to output the final damage state vector. This level innovatively introduces a priori causal graph pre-constructed based on the physical mechanisms of complex damage. Unlike traditional data-driven black-box reasoning, this priori causal graph internally encapsulates the formation mechanisms, characteristic change patterns, temporal evolution logic of various membrane damages, and the coupling effects of multiple factors. The model maps the fused global comprehensive features onto the priori causal graph, using real physical laws as constraints, and combines continuously acquired feature data to perform temporal causal reasoning. It not only analyzes the static feature state at the current moment but also traces the temporal change trend of features, accurately decoupling the causal relationship between environmental disturbances, equipment baseline drift, and the actual damage to the membrane itself, effectively distinguishing between instantaneous operating condition fluctuations and permanent membrane damage. Through this level of mechanism-constrained reasoning, a damage state vector with regular dimensions, clear semantics, and explicit physical meaning is finally generated. This vector contains high-dimensional quantitative data on various damage types, damage degrees, and evolution trends of the membrane components, providing standardized and highly reliable core input for subsequent multi-task parallel decoding and comprehensive, accurate diagnostic results output.

[0060] The pre-defined AI fusion inference model is a multimodal time-series diagnostic model built by combining deep learning and causal reasoning theory. It is not a general algorithm framework, but rather a three-stage progressive structure designed for single-modal feature extraction, cross-modal fusion, and time-series causal reasoning. It is specifically adapted to application scenarios involving multiple types of membrane module damage characteristics, complex water quality conditions, and mixed equipment operating baselines. The model's construction and acquisition methods are briefly described below. During the development phase, historical operating data of the membrane filtration device throughout its entire lifecycle, as well as various damage samples such as membrane pore blockage and membrane surface adsorption induced by different pollutants, were collected in batches. Operating data under different temperatures, pH levels, and pollutant concentrations were also included. The raw data was cleaned, labeled, and divided into training, validation, and test sets. The network core was built according to a predetermined three-layer architecture. A prior causal graph was drawn based on the physical laws governing various complex damages to the membrane module and embedded into the model. Subsequently, supervised training and iterative parameter optimization were conducted based on labeled samples until the model's diagnostic accuracy reached the preset standard. After model training, the network structure and parameters were solidified, and the model could be deployed directly to the field diagnostic system for use.

[0061] In some embodiments, damage semantic information in the damage state vector is used to complete multi-task decoding, including damage type determination, damage degree quantification, damage location localization, and damage development trend prediction, and integrate them to obtain a full-dimensional membrane module diagnostic result that includes damage category, damage severity level, damage distribution location, and future evolution trend.

[0062] This embodiment defines the final output stage of the entire diagnostic method. It is the core step in transforming the abstract damage state vector obtained from AI model inference into a visualized diagnostic result that maintenance personnel can directly read and apply. The damage state vector output by the model mentioned earlier is high-dimensional structured digital semantic information, densely containing various implicit features and quantitative information of membrane module damage, which cannot directly correspond to the specific equipment fault state. This embodiment uses a multi-task parallel decoding mechanism to decompose, parse, and target the multi-dimensional semantic information of the damage state vector, completing four core diagnostic tasks in one go. This changes the traditional membrane monitoring model, which can only output fault alarms or single parameter anomalies, ultimately forming a complete diagnostic conclusion covering all dimensions of damage.

[0063] The multi-task decoding module is a pre-trained parallel parsing branch structure. Each branch shares the damage state vector input resource, operates independently without interference, and corresponds to four types of dedicated diagnostic functions. The first type is the damage type determination task. This branch relies on the damage semantic feature map learned in the previous training to match the feature distribution pattern in the state vector, accurately distinguishing mainstream damage types such as membrane pore clogging damage, membrane surface adsorption and fouling damage, and flow channel filter cake accumulation damage. This solves the problem that traditional technologies cannot automatically identify the cause of damage, providing a basis for fault tracing for subsequent targeted operation and maintenance.

[0064] The second category is the damage severity quantification task. This branch quantifies and grades the severity of membrane module damage based on the amplitude offset and feature proportion of damage features in the state vector, combined with preset damage level evaluation criteria. It abandons the traditional fuzzy fault determination method and precisely classifies severity levels such as normal, minor damage, moderate damage, and severe damage through feature offset amplitude. This clearly identifies the difference between early-stage minor damage and late-stage severe damage, achieving a refined quantitative assessment of damage severity.

[0065] The third category is damage location tasks. This branch utilizes the asymmetric damage difference information retained in the state vector to analyze the characteristic differences of different regions of the membrane module. Based on the spatial response difference characteristics formed by bidirectional pulse acquisition, it can accurately determine whether the damage is concentrated at the front end, rear end, or distributed throughout the membrane module, locate the locally damaged area of ​​the membrane, solve the shortcoming of traditional overall monitoring that cannot locate the damage location, and avoid maintenance personnel blindly carrying out overall maintenance.

[0066] The fourth category is damage development trend prediction. This branch combines the temporal feature change patterns embedded in the state vector, compares the historical temporal feature baseline with the current feature offset rate, and infers the accumulation speed and evolution direction of damage. It can predict whether damage will continue to worsen in the short term and whether there is a risk of rapid failure, realizing an upgrade from passive fault detection to proactive trend prediction, and supporting maintenance personnel to formulate preventive maintenance plans in advance.

[0067] After the four decoding tasks are independently analyzed, the output data from all branches are structured and integrated to form a standardized, full-dimensional membrane module diagnostic result. The final output fully includes four core types of information: specific damage category, damage severity level, damage distribution location, and future damage evolution trend. This transforms the membrane module health status from single-point monitoring and qualitative judgment to a full-dimensional, quantitative, and predictable intelligent diagnosis. This output result can comprehensively, accurately, and precisely reflect the real-time health status of the membrane module, effectively avoiding the shortcomings of traditional monitoring and diagnosis, such as being one-sided, lagging, and lacking predictive capabilities. It provides accurate and comprehensive data support for the refined operation and maintenance, preventive repair, and downtime planning of membrane filtration systems.

[0068] In some embodiments, the multi-sensor array includes a dielectric sensor and a differential pressure sensor adapted for detecting membrane pore clogging damage, and an interfacial tension sensor and a fluid timing sensor adapted for detecting membrane surface adsorption damage. The sensing channel referred to in this application is an integrated dedicated signal acquisition link serving membrane module health monitoring, not simply a physical sensing element. It uses a physical sensor as its core component, integrating signal transmission lines, signal conditioning circuits, timing control units, data buffers, and output interfaces to form a complete working path with independent control, signal acquisition, and front-end preprocessing capabilities. The sensor is the core sensing unit of the sensing channel; the sensing channel emphasizes the entire functional link from signal sensing, transmission, preprocessing to data output, and is a combination of hardware, circuitry, and control logic.

[0069] The multi-sensor array is a hardware array for in-situ synchronous acquisition of multiple physical quantities, customized for typical damage mechanisms of membrane modules. The overall layout is based on the differences in the flow path structure, damage occurrence area, and physical response characteristics of the membrane module. All sensors adopt a non-invasive in-situ installation method, which does not damage the membrane fiber structure or interfere with the normal filtration and backwashing operation of the equipment. The entire array includes dielectric and differential pressure sensors adapted for membrane pore clogging damage detection, as well as interfacial tension and fluid timing sensors adapted for membrane surface adsorption damage detection. The four types of sensors perform their respective functions and are time-synchronized, which can match the bidirectional excitation rhythm of the pulse coding sequence of this invention to complete forward and reverse differential synchronous acquisition, providing multi-dimensional homogeneous physical data for refined health monitoring of membrane modules. The entire sensor array follows the principle of zoned corresponding detection. Membrane pore occlusion damage occurs inside the micropores of the membrane fibers and belongs to the structural resistance damage of the membrane. Correspondingly, differential pressure sensors and dielectric sensors are deployed. Membrane surface adsorption damage occurs at the fluid contact interface on the outer surface of the membrane fibers and belongs to the interfacial flow characteristic damage. Correspondingly, interfacial tension sensors and fluid timing sensors are deployed. The hardware layout corresponds one-to-one with the location of the damage, making the detection highly targeted.

[0070] The differential pressure sensor is installed in a transmembrane pair configuration, fixedly mounted on the straight sections of the main feed water inlet and main product water outlet pipes of the membrane module. These two installation points are close to both ends of the membrane module, allowing direct acquisition of real-time fluid pressure parameters at the inlet and outlet. Accurate transmembrane differential pressure data is obtained based on the pressure difference between the two ends. When heavy metal clogging occurs in the membrane pores, the flow cross-sectional area of ​​the membrane fibers decreases, and the fluid penetration resistance increases significantly. This directly alters the steady-state value and dynamic fluctuation amplitude of the transmembrane differential pressure under pulse excitation. This sensor can completely capture the differential pressure response changes under both forward probe pulse filtration and reverse control pulse flushing conditions, accurately reflecting the resistance damage characteristics caused by membrane pore blockage. It is the core hardware for identifying rigid pore clogging damage.

[0071] The dielectric sensor adopts an external wall-mounted structure, which is integrally attached and fixed to the outer wall of the middle section of the membrane module's membrane shell, corresponding to the effective filtration area of ​​the membrane fibers inside the membrane module. This is a non-contact, non-destructive testing layout, which can complete parameter acquisition without contacting the internal fluid and membrane fibers. Under normal operating conditions, the dielectric parameters of the membrane fibers and the wastewater medium remain stable. When heavy metal particles accumulate inside the membrane pores, the composition of the medium inside the pores changes, and the overall dielectric response characteristics show a slight shift. These slight changes in physical properties cannot be identified by pressure parameters. The dielectric sensor can accurately capture the difference in dielectric response inside the membrane during bidirectional pulse excitation, and can identify early, slight membrane pore blockage damage that cannot be detected by traditional monitoring methods, thus compensating for the shortcomings of differential pressure detection in detecting early, slight structural damage.

[0072] An interfacial tension sensor is installed on the straight section of the pressure-stabilizing pipeline at the inlet of the membrane module, adjacent to the inlet face of the membrane module. It can detect the solid-liquid interfacial tension parameters of the fluid entering the membrane module in real time. When high molecular weight surfactants in wastewater accumulate and adsorb onto the surface of the membrane fibers, they alter the interfacial characteristics of the membrane fibers in contact with the fluid, causing a continuous shift in the interfacial tension parameters. These parameters exhibit differentiated response characteristics under forward filtration pulses and reverse flushing pulses. This sensor can collect the steady-state value and dynamic changes of the interfacial tension under bidirectional pulse excitation throughout the entire process, accurately characterizing the adhesion and accumulation state of flexible adsorption damage on the membrane surface, and specifically adapting to the monitoring needs of surfactant adsorption-related damage.

[0073] A fluid timing sensor is fixedly installed on a straight section of the permeate outlet pipe of the membrane module to collect fluid velocity, flow stability, and fluid response timing parameters in real time. After an adsorption layer forms on the membrane fiber surface, it alters the smoothness of the inner wall of the flow channel, causing problems such as fluid flow hysteresis, steady-state velocity shift, and pulse response timing delay. This sensor can accurately capture the differences in dynamic flow characteristics of the fluid during bidirectional pulse excitation, record fluid timing shift data under forward detection and reverse control conditions, quantify the degree of interference of membrane surface adsorption on the fluid flow pattern, and provide evidence of the severity of membrane surface adsorption damage from the perspective of fluid motion.

[0074] In practical applications, all four types of sensors are connected to a unified hardware clock synchronization system, possessing a consistent acquisition timing reference. They can strictly follow the working rhythm of the pulse coding sequence, synchronously starting and stopping data acquisition during the forward and reverse pulse excitation phases of a single coding unit, ensuring that all sensor data are time-aligned and operate under the same conditions. All raw physical signals acquired by the sensor array are converted into digital signals through an analog-to-digital converter, forming a complete bidirectional response dataset. Data from different physical dimensions acquired by different sensors complement each other.

[0075] An AI-based multi-feature fusion diagnostic system for the health status of membrane modules is applied to a bidirectional flow membrane filtration device that can switch between forward and reverse operating states. A schematic diagram of the system modules is attached. Figure 5 As shown, the system includes: Encoding strategy unit 1 is used to obtain the component characteristics of the wastewater to be treated, and to construct a pulse coding sequence and corresponding monitoring strategy for the membrane module to be monitored based on the component characteristics. Each coding unit of the pulse coding sequence contains a parameter combination of a positive detection pulse and a reverse control pulse. The bidirectional excitation unit 2 is used to sequentially apply bidirectional pulse excitation to the membrane module according to the pulse coding sequence, and control the multi-sensor array to perform forward and reverse differential synchronous acquisition according to the monitoring strategy to obtain the bidirectional response dataset corresponding to each coding unit. Differential parsing unit 3 is used to perform feature extraction and differential operations on the bidirectional response dataset to construct differential feature pairs, and combine wastewater condition characteristics with membrane module operation characteristics to form a heterogeneous feature set; Model inference unit 4 is used to input heterogeneous feature sets into a preset AI fusion inference model to perform cross-modal attention fusion and causal temporal inference to obtain the damage state vector of the membrane module. Diagnostic output unit 5 is used to perform multi-task decoding based on the damage state vector and output full-dimensional diagnostic results.

[0076] Those skilled in the art will understand that the components of this application described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage system for execution by the computing system. Alternatively, they can be fabricated as separate integrated circuit components, or multiple components or steps can be fabricated as a single integrated circuit component. Thus, this application is not limited to any particular combination of hardware and software.

[0077] Note that the above description is merely a preferred embodiment and the technical principles employed in this application. Those skilled in the art will understand that this application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments. Many other equivalent embodiments may be included without departing from the concept of this application, and the scope of this application is determined by the scope of the appended claims.

[0078] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. An AI-based multi-feature fusion diagnostic method for the health status of membrane modules, characterized in that, A bidirectional flow membrane filtration device applicable to switchable forward and reverse operating states, the method comprising: The component characteristics of the wastewater to be treated are obtained, and a pulse coding sequence and corresponding monitoring strategy are constructed for the membrane module to be monitored based on the component characteristics. Each coding unit of the pulse coding sequence contains a parameter combination of a positive detection pulse and a reverse control pulse. Bidirectional pulse excitation is sequentially applied to the membrane module according to the pulse coding sequence, and the multi-sensor array is controlled to perform forward and reverse differential synchronous acquisition according to the monitoring strategy to obtain the bidirectional response dataset corresponding to each coding unit; Feature extraction and differential operations are performed on the bidirectional response dataset to construct differential feature pairs. Combined with wastewater condition characteristics and membrane module operation characteristics, a heterogeneous feature set is formed. The heterogeneous feature set is input into a preset AI fusion inference model to perform cross-modal attention fusion and causal temporal inference to obtain the damage state vector of the membrane module; Multi-task decoding is performed based on the damage state vector to output a full-dimensional diagnostic result.

2. The AI ​​multi-feature fusion diagnostic method according to claim 1, characterized in that, The compositional characteristics of the wastewater to be treated include at least the coordination characteristics of organic complexed heavy metals; When the component characteristics indicate the enrichment of organic complex heavy metals, a bidirectional pulse with a sawtooth waveform is configured in the pulse coding sequence. The excitation characteristics of membrane pore occlusion damage are matched by the pulse rise and fall timing change structure. In the monitoring strategy, a dielectric sensing channel and a differential pressure sensing channel are configured to continuously collect dielectric response and fluid resistance response data inside the membrane pore.

3. The AI ​​multi-feature fusion diagnostic method according to claim 1, characterized in that, The compositional characteristics of the wastewater to be treated include at least the polymerization characteristics of polymeric surfactants; When the component characteristics indicate the enrichment of polymeric surfactants, a bidirectional pulse of gradient square wave is configured in the pulse coding sequence. Relying on the pulse step-type flow velocity change structure to match the excitation characteristics of adsorption damage on the membrane surface, an interfacial tension sensing channel and a fluid time sequence sensing channel are configured in the monitoring strategy to continuously collect membrane interfacial viscous damping and fluid steady-state offset response data.

4. The AI ​​multi-feature fusion diagnostic method according to claim 1, characterized in that, Within each coding unit, the forward probe pulse is the excitation applied along the flow path corresponding to the forward operating state of the membrane assembly, and the reverse reference pulse is the excitation applied along the flow path corresponding to the reverse operating state of the membrane assembly. The reverse reference pulse provides a common reference for the response data corresponding to the forward probe pulse. During the forward and reverse differential synchronous acquisition process, the forward probe pulse and reverse control pulse applied in the current coding unit are used as reference signals to extract the same frequency response component that matches the excitation frequency of the pulse.

5. The AI ​​multi-feature fusion diagnostic method according to claim 1, characterized in that, The process of obtaining the difference feature pairs includes: The forward response dataset corresponding to the forward detection pulse and the reverse response dataset corresponding to the reverse control pulse are parsed from the bidirectional response dataset. Three types of dimensional features—time domain features, frequency domain features, and energy domain features—are extracted from the forward and reverse response datasets. In the process of constructing differential feature pairs, the same-dimensional features extracted from the forward and reverse response datasets are matched one-to-one, and the same-dimensional difference operation is performed to remove the fixed response deviation components caused by the pipeline and valve group in the bidirectional flow path. Only the differential features caused by the asymmetric damage of the membrane module are retained, thus completing the structured construction of the differential feature pairs.

6. The AI ​​multi-feature fusion diagnostic method according to claim 1, characterized in that, The wastewater operating characteristics include water quality component concentration shift characteristics, fluid temperature adaptation characteristics, and acid-base environment steady-state characteristics. The membrane module operating characteristics include permeate flux decay trend characteristics, transmembrane flow resistance fluctuation characteristics, and pipeline fluid steady-state shift characteristics.

7. The AI ​​multi-feature fusion diagnostic method according to claim 1, characterized in that, In the preset AI fusion inference model, for the heterogeneous feature set, the single-modal deep semantic features are first extracted through a parallel modality-specific network, and then the cross-modal feature directional interaction fusion is achieved through an inter-modal causal attention mechanism. Finally, the fused features are mapped to a priori causal graph constructed based on the composite damage physical mechanism, and temporal causal inference is performed to obtain the damage state vector.

8. The AI ​​multi-feature fusion diagnostic method according to claim 1, characterized in that, By utilizing the semantic information of damage in the damage state vector, multi-task decoding is completed, including damage type determination, damage degree quantification, damage location localization, and damage development trend prediction. The results are integrated to obtain a full-dimensional membrane module diagnostic result that includes damage category, damage severity level, damage distribution location, and future evolution trend.

9. The AI ​​multi-feature fusion diagnostic method according to claim 1, characterized in that, The multi-sensor array includes a dielectric sensor and a differential pressure sensor adapted for detecting membrane pore clogging damage, as well as an interfacial tension sensor and a fluid timing sensor adapted for detecting membrane surface adsorption damage.

10. An AI multi-feature fusion diagnostic system for the health status of membrane modules, characterized in that, A bidirectional flow membrane filtration device applicable to switchable forward and reverse operating states, the system includes: The coding strategy unit is used to acquire the component characteristics of the wastewater to be treated, and to construct a pulse coding sequence and a corresponding monitoring strategy for the membrane module to be monitored based on the component characteristics. Each coding unit of the pulse coding sequence contains a parameter combination of a positive detection pulse and a reverse control pulse. A bidirectional excitation unit is used to sequentially apply bidirectional pulse excitation to the membrane module according to the pulse coding sequence, and control the multi-sensor array to perform forward and reverse differential synchronous acquisition according to the monitoring strategy to obtain a bidirectional response dataset corresponding to each coding unit. The differential parsing unit is used to perform feature extraction and differential operations on the bidirectional response dataset to construct differential feature pairs, and combine the wastewater condition characteristics with the membrane module operation characteristics to form a heterogeneous feature set; The model inference unit is used to input the heterogeneous feature set into a preset AI fusion inference model to perform cross-modal attention fusion and causal temporal inference to obtain the damage state vector of the membrane module. The diagnostic output unit is used to perform multi-task decoding based on the damage state vector and output full-dimensional diagnostic results.

Images