Artificial intelligence-based ro membrane system failure prediction and maintenance method

Abstract

This invention relates to the field of artificial intelligence technology, specifically disclosing an AI-based method for fault prediction and maintenance of RO membrane systems. The method includes: collecting operational data from the main node of the RO membrane system; establishing mass conservation, solute conservation, and energy conservation equations to obtain a three-equation constraint set; assuming all membrane elements are in a healthy baseline state, calculating the theoretical output values ​​of each conservation equation to form a three-dimensional constraint residual vector; using the time-series evolution sequence of this vector as input, employing an AI model to jointly analyze the proportional relationship and time-series evolution pattern of the three conservation residuals, identifying failure mode types and failure location intervals, obtaining membrane element-level fault prediction results, generating a fault location report, and outputting maintenance instructions. This invention can accurately locate faults to specific membrane element location intervals under the objective existence of multi-level dilution effects of failure signals, improving the accuracy of RO membrane system fault prediction and maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for predicting and maintaining RO membrane system failures based on artificial intelligence. Background Technology

[0002] Reverse osmosis (RO) membrane separation technology has been widely used in municipal water supply, industrial pure water production, seawater desalination, and wastewater reuse, becoming a core process in modern water treatment engineering. With the increasing demands for water supply stability in industrial production, AI-based RO membrane system fault prediction and maintenance methods are gradually becoming a research hotspot and engineering practice direction in this field.

[0003] Existing AI-based methods for RO membrane system fault prediction and maintenance typically use data collected from sensors deployed at the main pipe node of the RO membrane system as model input to model the overall status of the RO membrane system and identify anomalies. However, industrial-grade RO membrane systems consist of dozens of pressure vessels connected in parallel, with multiple membrane elements connected in series within each vessel, forming a closed, multi-level topology. There are no independent sensors between the individual membrane elements. When a single membrane element experiences a local failure (such as O-ring breakage, membrane perforation, or endplate seal failure), its abnormal signal must undergo two stages of physical attenuation: inter-stage hydraulic equalization of the series-connected membrane elements within the vessel and lateral flow dilution by the parallel vessel array. By the time it reaches the system-level sensor, it has been submerged in normal operating noise fluctuations. This step-by-step physical attenuation mechanism inherent in the closed, series-parallel topology of the system constitutes a multi-stage dilution effect of the failure signal.

[0004] Existing AI-based methods for predicting and maintaining RO membrane system failures lack a feature reconstruction mechanism for the multi-level dilution effect of failure signals. The established AI models cannot effectively extract the discriminative features of local failures of individual membrane elements from system-level aggregated data. As a result, existing methods can only trigger system-level alarms when the failure range expands to affect the overall system performance in the early stages of local failure of membrane elements. They cannot further locate the fault to the specific pressure vessel and its internal membrane elements. After receiving the alarm, maintenance personnel still need to disassemble and inspect all pressure vessels and their internal membrane elements one by one, resulting in low maintenance efficiency, long downtime, and high maintenance costs. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an artificial intelligence-based method for RO membrane system fault prediction and maintenance, thereby addressing the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for fault prediction and maintenance of RO membrane systems based on artificial intelligence, comprising the following steps: The operating data of the main pipe node of the RO membrane system is collected synchronously at a preset sampling frequency. Based on the operating data, the mass conservation equation, solute conservation equation and energy conservation equation of the RO membrane system at the current operating moment are established respectively, and a three-equation constraint set is obtained. Assuming all membrane elements in the RO membrane system are in a healthy baseline state, the theoretical output values ​​of each conservation equation in the three-equation constraint group are calculated under the healthy baseline state. The difference between the measured output value and the corresponding theoretical output value of each conservation equation is obtained to obtain the mass conservation residual, solute conservation residual and energy conservation residual, which constitute a three-dimensional constraint residual vector. Using the time-series evolution sequence of the three-dimensional constrained residual vector as input, the trained artificial intelligence model is used to jointly analyze the proportional relationship between the three conserved residuals and their time-series evolution patterns, identify the failure mode type and failure location range corresponding to the proportional relationship and time-series evolution patterns, and obtain the membrane element-level failure prediction results. Based on the membrane element-level fault prediction results, a fault location report is generated, which includes the failure mode type, failure location range and corresponding failure probability. The maintenance priority of each failure location range is sorted according to the failure probability, and maintenance instructions are output.

[0007] Preferably, the steps for obtaining the three-equation constraint set are as follows: The operating data of the RO membrane system's main pipe node are collected synchronously at a preset sampling frequency. The main pipe node includes the main inlet pipe, the main permeate pipe, and the main concentrate pipe. The operating data includes the main inlet pipe pressure, main inlet pipe flow rate, main inlet pipe conductivity and temperature, main permeate pipe pressure, main permeate pipe flow rate, main permeate pipe conductivity, and main concentrate pipe pressure and main concentrate pipe flow rate. Based on the aforementioned operational data, the mass conservation equation, solute conservation equation, and energy conservation equation for the RO membrane system at the current operating moment are established. By combining the mass conservation equation, the solute conservation equation, and the energy conservation equation, we obtain a three-equation constraint set.

[0008] Preferably, the steps for obtaining the three-dimensional constraint residual vector are as follows: S21. Obtain the health baseline parameters for all membrane elements in the RO membrane system to be in a healthy baseline state. The health baseline state is based on the water permeability coefficient and salt permeability coefficient of the membrane element under the factory inspection conditions. Based on the temperature data in the operating data, the health baseline parameters are corrected according to the temperature correction relationship provided by the membrane element manufacturer to obtain the corrected health baseline parameters under the current temperature conditions. S22. Based on the corrected health baseline parameters, and combined with the total feed water pressure, total feed water flow rate, and total feed water conductivity from the operating data, calculate the theoretical total permeate flow rate of the RO membrane system under the health baseline state according to the membrane mass transfer relationship.Q p0 Theoretical total concentrate flow rate Q c0 Theoretical total water production pipe conductivity C p0 Theoretical total concentrate pipe conductivity C c0 and theoretical high-pressure pump input power W 0 We obtain the theoretical output values ​​of the three conservation equations under a healthy baseline state; S23. Before performing the calculation of each conserved residual, a steady-state judgment is made on the change amplitude of the running data at adjacent sampling times; S24. For the sampling time determined by steady-state judgment, obtain the measured output value of each conservation equation, and subtract the measured output value of each conservation equation from the corresponding theoretical output value to obtain the mass conservation residual. Solute conservation residual and energy conservation residual The mass conservation residual, solute conservation residual, and energy conservation residual are combined to form a three-dimensional constrained residual vector. ; S25. Repeat the calculation process from S21 to S24 at consecutive effective sampling times, and arrange the three-dimensional constraint residual vector R at each effective time in time sequence to obtain the time sequence evolution sequence of the three-dimensional constraint residual vector.

[0009] Preferably, the steady-state determination of the change amplitude of the running data at adjacent sampling times is as follows: Calculate the absolute change in total inlet pipe pressure and the relative percentage change in total inlet pipe flow rate between the current sampling time and the previous sampling time. Only when the absolute change in the total inlet water pressure is lower than the pressure steady-state threshold and the relative percentage change in the total inlet water flow is lower than the flow steady-state threshold, is the operating data at the current sampling time determined to be in steady state, and subsequent conserved residual calculation is performed. Otherwise, sampling time data that fail the steady-state judgment are marked and skipped, and are not included in the construction of the three-dimensional constrained residual vector.

[0010] Preferably, the steps for obtaining the membrane element-level fault prediction results are as follows: S31. Extract features from two independent dimensions of the temporal evolution sequence of the three-dimensional constrained residual vector to form a two-dimensional fingerprint feature vector; S32. The artificial intelligence model adopts a two-stage cascaded structure: the first stage is a failure mode classifier, which takes the two-dimensional fingerprint feature vector F as input, classifies four states: sealing failure, membrane perforation, scaling and normal operation, and outputs the current failure mode type and the corresponding confidence level. The second stage is the failure location interval classifier, which uses the two-dimensional fingerprint feature vector F and the failure mode type output from the first stage as joint inputs to locate the failure location interval under the constraint of the known failure mode type, and outputs the failure location interval and the corresponding failure probability. S33. During the formal operation phase of the RO membrane system, the two-dimensional fingerprint feature vector F of the three-dimensional constraint residual vector time-series evolution sequence is continuously extracted using a sliding time window. The vector is input into the trained artificial intelligence model and then processed through a two-stage process of failure mode classifier and failure location interval classifier. The failure mode type, failure location interval and corresponding failure probability of the current time window are output. The failure mode type, failure location range and corresponding failure probability after continuous window consistency confirmation are used as the membrane element level failure prediction result and output to step S4.

[0011] Preferably, the two-dimensional fingerprint feature vector is obtained in the following way: The first dimension is the three residual ratio feature: within the current sliding time window, calculate the ratio between the mass conservation residual, solute conservation residual, and energy conservation residual, and obtain the ratio feature vector P after normalization; The second dimension is the temporal evolution characteristic: for the three-dimensional constraint residual vectors at consecutive moments within the sliding time window, calculate the magnitude of the three-dimensional constraint residual vector. The growth rate over time and the current cumulative magnitude constitute the temporal evolution feature vector D; The proportional feature vector P is combined with the temporal evolution feature vector D to form a two-dimensional fingerprint feature vector F=[P,D], which is used as the input to the artificial intelligence model.

[0012] Preferably, the steps for outputting maintenance instructions are as follows: Based on the membrane element-level fault prediction results, a fault location report is generated; Based on the failure probability corresponding to each failure location interval in the fault location report, the maintenance priority of each failure location interval is sorted. Based on the maintenance priority ranking result, corresponding maintenance instructions are generated for each failure location interval.

[0013] As described above, the artificial intelligence-based RO membrane system fault prediction and maintenance method provided by the present invention has at least the following beneficial effects: This invention provides an AI-based method for RO membrane system fault prediction and maintenance. By constructing a three-dimensional constrained residual vector from the residuals of the mass, solute, and energy conservation equations, it fundamentally overcomes the inherent multi-level dilution effect of failure signals in the series-parallel closed topology of RO membrane systems, which overwhelms fault characteristics. This allows local failures of membrane elements at any location to be captured as non-diluted conservation deviations. The method explicitly separates the features of the three-dimensional constrained residual vector into residual ratio features and temporal evolution features, and uses a two-stage cascaded AI model to sequentially identify failure mode types and locate failure location intervals. This overcomes the limitation of existing methods that can only trigger system-level alarms and cannot locate specific membrane elements, achieving a leap from system-level alarms to precise membrane element-level location. By combining failure probability and the severity of failure mode type as maintenance priority ranking criteria, it generates targeted maintenance instructions covering three levels: immediate shutdown for repair, planned repair, and intensive monitoring. This precisely limits the maintenance inspection scope to specific failure location intervals and provides corresponding inspection methods, significantly shortening downtime, reducing maintenance costs, and improving the operational reliability and maintenance efficiency of the RO membrane system. Attached Figure Description

[0014] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the artificial intelligence-based RO membrane system fault prediction and maintenance method of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 As shown, this invention provides an artificial intelligence-based method for RO membrane system fault prediction and maintenance, comprising the following steps: S1. Synchronously collect the operating data of the main pipe node of the RO membrane system at a preset sampling frequency. Based on the operating data, establish the mass conservation equation, solute conservation equation and energy conservation equation of the RO membrane system at the current operating moment to obtain the three-equation constraint set. In this embodiment, the steps for obtaining the three-equation constraint set are as follows: Operating data of the RO membrane system's main pipe node is synchronously collected at a preset sampling frequency using pressure sensors, flow meters, conductivity meters, and temperature sensors deployed at the main pipe node. The main pipe node includes the main inlet pipe, the main permeate pipe, and the main concentrate pipe. The operating data includes the main inlet pipe pressure, main inlet pipe flow rate, main inlet pipe conductivity and temperature, the main permeate pipe pressure, main permeate pipe flow rate, main permeate pipe conductivity, and the main concentrate pipe pressure and main concentrate pipe flow rate. In one specific embodiment, the sampling frequency is set to once per minute, which can cover the time scale requirements of the high-pressure pump start-up and shutdown transient process and membrane surface fouling accumulation.

[0018] It should be specifically explained that the synchronization of the data acquisition time of the main node sensors is achieved in the following way: the data acquisition controller sends a unified acquisition trigger command to each sensor, each sensor responds to the same trigger command to complete data acquisition, and the data acquisition controller adds a unified timestamp to the data returned by each sensor; when the response delay of each sensor is not negligible, the data acquired by each node is linearly interpolated and aligned according to its actual acquisition timestamp to ensure the consistency of the data of each node in the time dimension at the same running time, and to avoid the introduction of timing errors when establishing the conservation equation due to asynchronous acquisition time.

[0019] Based on the aforementioned operational data, the mass conservation equation, solute conservation equation, and energy conservation equation for the RO membrane system at the current operating moment are established. The specific form of the mass conservation equation is as follows: ,in, Q in This represents the measured total inlet water flow rate. Q p This represents the measured total flow rate of the production water pipe. Q c The three values ​​represent the actual measured flow rates of the total concentrate pipe, all obtained directly from the flow meters at the corresponding main pipe nodes.

[0020] The mass conservation equation reflects the water balance state of the RO membrane system at the current operating moment. When the membrane element is in a healthy state, the mass conservation relationship holds within the allowable error range. When the membrane element fails to seal, resulting in bypass leakage between the inlet and product water sides, the sum of the measured total product water flow rate and the total concentrate flow rate will deviate from the total inlet water flow rate, and the mass conservation equation will show a quantifiable deviation. The specific form of the solute conservation equation is as follows: ,in, C in This is the measured value of the total inlet pipe conductivity. C p This is the measured value of the total permeable water pipe conductivity. C cThis is an estimated value for the conductivity of the total concentrate pipe, since conductivity meters are typically not installed on the total concentrate pipe. C c The formula is derived from the measured parameters based on the solute conservation principle. ; The solute conservation equation reflects the solute flux balance of the RO membrane system at the current operating moment. When the membrane element is in a healthy state, the solute conservation relationship holds within the allowable error range. When the membrane element experiences membrane perforation or an abnormal increase in salt permeability, the measured permeate conductivity will deviate from the constraint expectation of the solute conservation equation, and the solute conservation equation will show a quantifiable deviation.

[0021] The specific form of the energy conservation equation is as follows: Where W is the measured value of the high-pressure pump input power. P in This is the measured value of the main inlet water pipe pressure. P p This is the measured value of the total production water pipe pressure. P c This is the measured value of the total concentrate pipe pressure. The effective permeation power of an RO membrane system characterizes the effective work done by high-pressure feed water to drive permeate through the membrane surface. For the hydraulic dissipation power on the concentrate side, This is to account for additional hydraulic losses.

[0022] The additional hydraulic losses refer to energy losses in the RO membrane system other than the effective permeation power of the membrane surface and the hydraulic dissipation power on the concentrate side. These mainly include: frictional resistance losses along the pipeline, local resistance losses at bends and valves, and turbulent dissipation generated when water flows through pipeline accessories. These losses objectively exist during system operation and vary with the total influent flow rate.

[0023] The additional hydraulic loss is obtained through calibration during the system commissioning period. The calibration process is performed after the membrane element is brand new and in good health, as confirmed by inspection. Under multiple typical flow conditions covering the normal operating range of the system, the total inlet pipe flow rate and the corresponding measured high-pressure pump input power, total inlet pipe pressure, total product water pipe pressure, total concentrate pipe pressure, total product water pipe flow rate, and total concentrate pipe flow rate are recorded for each condition. The additional hydraulic loss value under each condition is then calculated using the energy conservation equation, thereby establishing a calibration curve for the additional hydraulic loss as a function of the total inlet pipe flow rate. After the system is officially operational, the corresponding additional hydraulic loss value is interpolated from the calibration curve based on the currently measured total inlet pipe flow rate and substituted into the energy conservation equation for subsequent calculations. Those skilled in the art can implement the calibration and value acquisition process based on the above description.

[0024] The energy conservation equation reflects the energy balance state of the RO membrane system at the current operating moment. When the membrane elements are in a healthy state, the energy conservation relationship holds within the allowable error range. When fouling occurs in the membrane elements, resulting in an abnormal increase in the hydraulic resistance on the membrane surface, the effective permeation power of the system decreases, and the input power of the high-pressure pump is abnormally high under the same feed water conditions. A quantifiable deviation appears in the energy conservation equation. The input power W of the high-pressure pump is obtained as follows: First, directly read the power output display value of the frequency converter supporting the high-pressure pump. When the frequency converter does not have a power output interface, measure the three-phase current and three-phase voltage of the high-pressure pump motor using a current transformer and a voltage sensor respectively, and calculate W according to the three-phase active power calculation formula. The specific calculation formula is: , where U is the line voltage, I is the line current, and cosφ is the power factor.

[0025] 联立质量守恒方程、溶质守恒方程及能量守恒方程,得到三方程约束组;所述三方程约束组中各守恒方程相互独立,分别从水量、溶质通量及能量三个物理维度对RO膜系统当前运行状态进行约束描述,为后续步骤中健康基准理论值的计算及守恒残差的构造提供完整的物理方程基础。 By联立质量守恒方程、溶质守恒方程及能量守恒方程,得到三方程约束组;所述三方程约束组中各守恒方程相互独立,分别从水量、溶质通量及能量三个物理维度对RO膜系统当前运行状态进行约束描述,为后续步骤中健康基准理论值的计算及守恒残差的构造提供完整的物理方程基础。 The mass conservation equation, solute conservation equation, and energy conservation equation are联立 to obtain a three-equation constraint group. Each conservation equation in the three-equation constraint group is independent, and respectively describes the current operating state of the RO membrane system from three physical dimensions of water volume, solute flux, and energy, providing a complete physical equation basis for the calculation of the healthy benchmark theoretical value and the construction of the conservation residual in the subsequent steps.

[0026] It should be specifically noted that, compared with the traditional method, in step S1, the conservation equations are used as the organizational framework for operating data, rather than directly using the original sensor data as the input of fault characteristics. The constraint relationship of the conservation equations is guaranteed by the basic laws of thermodynamics. Its deviation reflects the imbalance of the overall physical state of the system and is not affected by the multi-stage dilution effect of failure signals. That is, no matter where the failed membrane element is located in the series-parallel topology, the mass, solute, or energy deviation caused by it is reflected in the corresponding conservation equation in the form of the total amount and will not be erased due to multi-stage dilution. In contrast, the traditional method directly uses the original numerical changes of the main pipe conductivity or pressure difference as fault characteristics. The local failure signals of single membrane elements have extremely small amplitudes after series averaging and parallel dilution, and are submerged in the normal operating condition fluctuation noise, resulting in the inability to detect early faults.

[0027] S2. On the premise that all membrane elements in the RO membrane system are in the healthy benchmark state, calculate the theoretical output values of each conservation equation in the three-equation constraint group under the healthy benchmark state, and subtract the measured output value of each conservation equation from the corresponding theoretical output value to obtain the mass conservation residual, solute conservation residual, and energy conservation residual, forming a three-dimensional constraint residual vector. In this embodiment, the steps for obtaining the three-dimensional constraint residual vector are as follows: s S21. Obtain health baseline parameters for all membrane elements in the RO membrane system to be in a healthy baseline state. The health baseline state is based on the water permeability coefficient and salt permeability coefficient of the membrane elements under factory inspection conditions. The water permeability coefficient and salt permeability coefficient are determined by the product specifications provided by the membrane element manufacturer and are verified and corrected by measured operating data of the membrane elements in a brand-new state during system commissioning. Since the water permeability coefficient and salt permeability coefficient change with the operating temperature, the health baseline parameters are corrected according to the temperature data in the operating data and the temperature correction relationship provided by the membrane element manufacturer to obtain the corrected health baseline parameters under the current temperature conditions. It should be noted that, in order to ensure the continuous accuracy of the health baseline parameters in long-term operation, a regular update mechanism is established: three consecutive time periods in which the conservation residuals are all below the preset threshold are selected periodically and judged as a fault-free normal operation state. Based on the measured operating data within the time period, the current actual water permeability coefficient and salt permeability coefficient are calculated and used as the updated health baseline parameters.

[0028] In one specific embodiment, the preset threshold is set so that the absolute value of each residual is less than 1% of the full scale. This update mechanism enables the health baseline parameters to track the slow performance degradation of the membrane element due to normal aging, avoiding the introduction of systematic bias.

[0029] S22. Based on the corrected health baseline parameters, and combined with the total feed water pressure, total feed water flow rate, and total feed water conductivity from the operating data, calculate the theoretical total permeate flow rate of the RO membrane system under the health baseline state according to the membrane mass transfer relationship. Q p0 Theoretical total concentrate flow rate Q c0 Theoretical total water production pipe conductivity C p0 Theoretical total concentrate pipe conductivity C c0 and theoretical high-pressure pump input power W 0 We obtain the theoretical output values ​​of the three conservation equations under a healthy baseline state; It should be noted that calculating the theoretical permeate flow rate and theoretical permeate conductivity based on the membrane mass transfer relationship is common knowledge in the field of RO membrane systems. Those skilled in the art can calculate the theoretical output value based on the dissolution-diffusion model and the water permeability and salt permeability under healthy baseline conditions, combined with the current total inlet pressure, total inlet flow rate and total inlet conductivity. The specific calculation formula will not be elaborated here.

[0030] The theoretical output values ​​include: the theoretical output values ​​of the mass conservation equation. Theoretical output value of the solute conservation equation The theoretical output value W of the energy conservation equation0 .

[0031] S23. Before performing the calculation of each conserved residual, a steady-state judgment is made on the change amplitude of the running data at adjacent sampling times; In this embodiment, the steady-state determination of the change amplitude of running data at adjacent sampling times is as follows: Calculate the absolute change in total inlet pipe pressure and the relative percentage change in total inlet pipe flow rate between the current sampling time and the previous sampling time. Only when the absolute change in the total inlet water pressure is lower than the pressure steady-state threshold and the relative percentage change in the total inlet water flow is lower than the flow steady-state threshold, is the operating data at the current sampling time determined to be in steady state, and subsequent conserved residual calculation is performed. Otherwise, sampling time data that fail the steady-state judgment are marked and skipped, and are not included in the construction of the three-dimensional constrained residual vector.

[0032] It should be noted that the pressure steady-state threshold and the flow steady-state threshold are preset empirical parameters used to distinguish between steady-state system operation and operating condition fluctuations.

[0033] In one specific embodiment, the steady-state pressure threshold is set to 0.2 bar, and the steady-state flow rate threshold is set to 5%. The thresholds are selected based on the following: under healthy baseline conditions, pressure changes caused by normal operating condition fluctuations are typically within 0.1 bar, and flow rate changes are within 3%. Appropriately relaxing the thresholds to 0.2 bar and 5% can avoid misjudgments caused by measurement noise, while ensuring that non-steady-state operating conditions (such as high-pressure pump start-up and shutdown, valve regulation) are effectively eliminated.

[0034] It should be noted that the steady-state judgment step ensures that the residual calculation is performed only when the system is in a stable operating state, avoiding distortion of the theoretical output value calculation results when the membrane mass transfer process has not reached a steady state during periods of drastic fluctuations in the influent conditions, thereby ensuring the physical validity of the three-dimensional constraint residual vector at all effective moments.

[0035] S24. For the sampling time determined by steady-state judgment, obtain the measured output values ​​of each conservation equation. The measured output values ​​include: the measured output values ​​of the mass conservation equation. ; Measured output values ​​of the solute conservation equation The measured output value W of the energy conservation equation; The mass conservation residuals are obtained by subtracting the measured output values ​​from the corresponding theoretical output values ​​of each conservation equation. Solute conservation residual and energy conservation residual ; Specific calculation formula: ; ; ; in, It reflects the degree of deviation between the actual water balance of the system and the healthy baseline state. It reflects the degree of deviation between the actual solute flux balance of the system and the healthy baseline state. It reflects the degree of deviation between the system's actual energy balance and its healthy baseline state.

[0036] The mass conservation residual, solute conservation residual, and energy conservation residual are combined to form a three-dimensional constrained residual vector. ; The direction of the three-dimensional constrained residual vector in the three-dimensional residual space is determined by the proportional relationship of the three conserved residual components, and its magnitude is determined by the combined magnitude of the three conserved residual components.

[0037] The three-dimensional constraint residual vectors corresponding to different failure modes and failure location intervals have distinguishable characteristic combinations in terms of direction and modulus: when the membrane element experiences sealing failure leading to bypass leakage on the feed water side and product water side, the additional leakage carries both water and solute. and They increase in the same direction, while Because the hydraulic resistance of the leakage path is lower than that of the normal membrane mass transfer path, the input power of the high-pressure pump is relatively reduced. When membrane perforation occurs in the membrane element, leading to an abnormally high salt permeability, the solute flux increases while the water balance is less affected. Significantly increased and The changes are relatively limited; When fouling occurs on the membrane element, increasing the hydraulic resistance at the membrane surface, the driving force pressure required to maintain the same water production rate increases, and the input power of the high-pressure pump increases. Significantly increased and and The changes are relatively small. The differences in the proportional relationship of the conserved residuals constitute the physical basis for subsequent artificial intelligence models to identify failure modes and locate failure location intervals.

[0038] S25. Repeat the calculation process from S21 to S24 at consecutive effective sampling times, and arrange the three-dimensional constraint residual vector R at each effective time in time sequence to obtain the time sequence evolution sequence of the three-dimensional constraint residual vector.

[0039] It should be specifically noted that, compared with traditional methods, step S2 uses a three-dimensional constrained residual vector instead of a single-parameter residual as a structured representation of fault characteristics. Traditional methods typically only monitor the deviation of a single parameter, such as transmembrane pressure difference or product water conductivity. The response of a single-parameter residual to different failure modes is aliased, making it impossible to distinguish the failure type. The three-dimensional constrained residual vector constructed by step S2 simultaneously characterizes the deviation of the system state from three independent physical dimensions: water quantity, solute flux, and energy. The proportional relationship of the three conserved residual components has a unique distinguishability for different failure modes, fundamentally solving the problem of insufficient ability of single-parameter residuals to distinguish failure modes. This provides a complete feature input for subsequent artificial intelligence models to achieve simultaneous identification of failure mode types and failure location intervals.

[0040] S3. Using the time-series evolution sequence of the three-dimensional constrained residual vector as input, the trained artificial intelligence model is used to jointly analyze the proportional relationship between the three conserved residuals and their time-series evolution patterns, identify the failure mode type and failure location interval corresponding to the proportional relationship and time-series evolution patterns, and obtain the membrane element-level fault prediction results. In this embodiment, the steps for obtaining the membrane element-level fault prediction results are as follows: S31. Extract features from two independent dimensions of the temporal evolution sequence of the three-dimensional constrained residual vector to form a two-dimensional fingerprint feature vector; In this embodiment, the two-dimensional fingerprint feature vector is obtained as follows: The first dimension is the three residual ratio feature: within the current sliding time window, the ratio between the mass conservation residual, solute conservation residual and energy conservation residual is calculated, and after normalization, the ratio feature vector P is obtained. The ratio feature vector represents the direction of the three-dimensional constraint residual vector in the residual space. Specifically, the proportional feature vector has the ability to distinguish failure mode types: the three failure modes of sealing failure, membrane perforation and scaling correspond to different directions in the residual space, and the proportional relationship of the three residuals constitutes the main basis for failure mode type identification.

[0041] The second dimension is the temporal evolution characteristic: for the three-dimensional constraint residual vectors at consecutive moments within the sliding time window, calculate the magnitude of the three-dimensional constraint residual vector. The growth rate over time and the current cumulative magnitude constitute the temporal evolution feature vector D; Specifically, the temporal evolution feature vector has the ability to distinguish failure location intervals. The physical basis for this is that in a parallel topology RO membrane system, there is a pressure decreasing gradient along the series direction from the feed water end to the concentrate end. Membrane elements closer to the feed water end have a high net driving force, and when a membrane element fails at this location, the bypass leakage or abnormal salt permeation is relatively large. The magnitude growth rate of the three-dimensional constraint residual vector is fast, and the cumulative amplitude is large. Conversely, the net driving force at the concentrate end has decreased due to the consumption of the preceding membrane elements. When a membrane element fails at this location, the bypass leakage or abnormal salt permeation is relatively small, and the magnitude growth rate of the three-dimensional constraint residual vector is slow, and the cumulative amplitude is small. The difference between the growth rate and the cumulative amplitude forms a distinguishable location fingerprint in the temporal evolution path, constituting the physical basis for locating the failure location interval.

[0042] It should be noted that the growth rate can be calculated using the first difference of the modulus sequence within the current window or the slope of the linear regression, while the cumulative magnitude can be directly taken as the modulus value at the current moment. Those skilled in the art can achieve this by following the above guidelines, and the specific calculation formula will not be elaborated here.

[0043] The length of the sliding time window is selected within the range of the following: the lower bound is the shortest time required for the three-dimensional constraint residual vector to deviate from the healthy baseline state to the detectable amplitude, and the upper bound is the shortest time interval at which the system operating conditions change significantly.

[0044] In one specific embodiment, the time window length is set to 24 hours and the window sliding step size is set to 1 hour, which can meet the above constraints.

[0045] The proportional feature vector P is combined with the temporal evolution feature vector D to form a two-dimensional fingerprint feature vector F=[P,D], which is used as the input of the artificial intelligence model. S32. The artificial intelligence model adopts a two-stage cascaded structure: the first stage is a failure mode classifier, which takes the two-dimensional fingerprint feature vector F as input, classifies four states: sealing failure, membrane perforation, scaling and normal operation, and outputs the current failure mode type and the corresponding confidence level. The second stage is the failure location interval classifier, which uses the two-dimensional fingerprint feature vector F and the failure mode type output from the first stage as joint inputs to locate the failure location interval under the constraint of the known failure mode type, and outputs the failure location interval and the corresponding failure probability. The failure location intervals are defined based on the actual number of membrane elements connected in series in the RO system and the planned pressure vessel array. In one specific embodiment, the membrane elements inside the pressure vessel are divided into three segments according to the pressure decrease gradient along the series direction: the inlet segment, the middle segment, and the concentrate segment. The temporal evolution characteristics of each segment have distinguishable differences in growth rate and cumulative amplitude, meeting the resolution requirements for failure location interval positioning. In practical applications, the granularity of the failure location interval division can be adjusted according to the system scale and positioning accuracy requirements.

[0046] It is important to note that there is a physical coupling between failure mode type and failure location interval. If these are treated as a single-label classification problem, the AI ​​model would need to learn both overlapping relationships simultaneously, resulting in high training difficulty and low accuracy. The two-stage cascaded structure uses the failure mode type output from the first stage as a conditional constraint for the second stage. Given the failure mode type, failure location interval localization only requires distinguishing the temporal evolution characteristics of different locations under the same failure mode, effectively reducing the complexity of the localization task and improving localization accuracy.

[0047] In one specific embodiment, the failure mode classifier is constructed using a random forest algorithm, containing 100 decision trees. It takes five dimensions (3 dimensions of proportional features + 2 dimensions of temporal features) of the two-dimensional fingerprint feature vector F as input and outputs four types of probabilities. The failure location interval classifier is constructed using a multilayer perceptron, containing two hidden layers (with 32 and 16 neurons respectively), using ReLU as the activation function, and the output layer uses Softmax to output the probability distribution of each location interval.

[0048] It should be specifically explained that the training process of the artificial intelligence model is as follows: the two-dimensional fingerprint feature vector F corresponding to the events with confirmed failure mode types and failure location intervals in the historical operation records is used as the label training sample, and the corresponding failure mode type and failure location interval are used as labels to conduct supervised training on the failure mode classifier and the failure location interval classifier respectively; during training, the loss function of the failure mode classifier adopts cross-entropy loss, and the loss function of the failure location interval classifier adopts weighted cross-entropy loss, and different weights are assigned to samples of different location intervals to handle the class imbalance problem; the model evaluation indicators include classification accuracy, location accuracy (the overlap ratio between the predicted interval and the actual interval), and F1-score.

[0049] S33. During the formal operation phase of the RO membrane system, the two-dimensional fingerprint feature vector F of the three-dimensional constraint residual vector time-series evolution sequence is continuously extracted using a sliding time window. The vector is input into the trained artificial intelligence model and then processed through a two-stage process of failure mode classifier and failure location interval classifier. The failure mode type, failure location interval and corresponding failure probability of the current time window are output. Specifically, continuous window consistency verification is performed on the output results: when the output results of a consecutive preset number of time windows are consistent in the failure mode type and failure location range, the prediction results are confirmed to be valid; otherwise, they are filtered.

[0050] In one specific embodiment, the number of consecutive window confirmations is set to 3, that is, the prediction result is confirmed to be valid when the output results of 3 consecutive time windows are consistent.

[0051] The failure mode type, failure location range and corresponding failure probability after continuous window consistency confirmation are used as the membrane element level failure prediction result and output to step S4.

[0052] It should be specifically noted that, compared with traditional methods, step S3 replaces the original residual time series signal with a two-dimensional fingerprint feature vector and directly inputs it into the artificial intelligence model. This explicitly separates the failure mode information carried by the residual proportional relationship from the failure location information carried by the time series evolution features. This allows the artificial intelligence model to establish mapping relationships for the two types of features with different physical meanings, avoiding the problem of the model being unable to distinguish between failure modes and failure locations due to the mixed input of all features.

[0053] S4. Based on the membrane element-level fault prediction results, generate a fault location report that includes the failure mode type, failure location range and corresponding failure probability, and sort the maintenance priority of each failure location range based on the failure probability, and output maintenance instructions.

[0054] In this embodiment, the step of outputting the maintenance command is as follows: Based on the membrane element-level fault prediction results, a fault location report is generated. The fault location report includes the failure mode type, the corresponding failure location interval, and the failure probability of each failure location interval, which are confirmed by continuous window consistency. When there are multiple failure location interval prediction results at the same operating time, they are listed in the fault location report by failure location interval, and each record includes the corresponding failure mode type and failure probability. Based on the failure probability corresponding to each failure location interval in the fault location report, the maintenance priority of each failure location interval is sorted. Specifically, failure locations with higher failure probabilities are ranked higher; when failure probabilities are the same, they are further ranked according to the degree of impact of the failure mode type on the system's operational safety, with sealing failures and membrane perforation being ranked higher than scaling failures. The physical basis for this is that, under the same failure probability, different failure modes have varying degrees of harm to the system's permeate water quality and operational safety. Water quality deterioration caused by seal failure and membrane perforation has an immediate impact, which will continue to escalate if not addressed promptly; while the increase in hydraulic resistance caused by scaling failures has a gradual impact, which does not affect the permeate water quality within a certain range. Incorporating the severity of failure mode types into the ranking criteria ensures that the maintenance priority ranking matches the actual operational risks of the system, avoiding the shortcomings of simply ranking based on failure probability while ignoring the differences in the harm caused by different failure modes.

[0055] Based on the maintenance priority ranking result, corresponding maintenance instructions are generated for each failure location interval; Specifically, for the highest-ranked failure location range, an immediate shutdown and maintenance instruction is generated, which includes the range of pressure vessels that need to be isolated, the locations of membrane elements that need to be inspected, and the inspection methods determined based on the failure mode type.

[0056] For the second-highest ranked failure location range, a planned maintenance instruction is generated, which includes a suggestion to perform the inspection within the next planned shutdown window, and the failure probability is continuously monitored before the planned shutdown.

[0057] For failure location ranges that are ranked low and have a failure probability below the planned maintenance threshold, an encrypted monitoring instruction is generated, which includes increasing the monitoring frequency to a preset encrypted monitoring frequency and continuously tracking the evolution trend of failure probability.

[0058] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0059] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for fault prediction and maintenance of RO membrane systems based on artificial intelligence, characterized in that, Includes the following steps: The operating data of the main pipe node of the RO membrane system is collected synchronously at a preset sampling frequency. Based on the operating data, the mass conservation equation, solute conservation equation and energy conservation equation of the RO membrane system at the current operating moment are established respectively, and a three-equation constraint set is obtained. Assuming all membrane elements in the RO membrane system are in a healthy baseline state, the theoretical output values ​​of each conservation equation in the three-equation constraint group are calculated under the healthy baseline state. The difference between the measured output value and the corresponding theoretical output value of each conservation equation is obtained to obtain the mass conservation residual, solute conservation residual and energy conservation residual, which constitute a three-dimensional constraint residual vector. Using the time-series evolution sequence of the three-dimensional constrained residual vector as input, the trained artificial intelligence model is used to jointly analyze the proportional relationship between the three conserved residuals and their time-series evolution patterns, identify the failure mode type and failure location range corresponding to the proportional relationship and time-series evolution patterns, and obtain the membrane element-level failure prediction results. Based on the membrane element-level fault prediction results, a fault location report is generated, which includes the failure mode type, failure location range and corresponding failure probability. The maintenance priority of each failure location range is sorted according to the failure probability, and maintenance instructions are output.

2. The method for fault prediction and maintenance of RO membrane systems based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the three-equation constraint set are as follows: The operating data of the RO membrane system's main pipe node are collected synchronously at a preset sampling frequency. The main pipe node includes the main inlet pipe, the main permeate pipe, and the main concentrate pipe. The operating data includes the main inlet pipe pressure, main inlet pipe flow rate, main inlet pipe conductivity and temperature, main permeate pipe pressure, main permeate pipe flow rate, main permeate pipe conductivity, and main concentrate pipe pressure and main concentrate pipe flow rate. Based on the aforementioned operational data, the mass conservation equation, solute conservation equation, and energy conservation equation for the RO membrane system at the current operating moment are established. By combining the mass conservation equation, the solute conservation equation, and the energy conservation equation, we obtain a three-equation constraint set.

3. The method for fault prediction and maintenance of RO membrane systems based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the three-dimensional constraint residual vector are as follows: S21. Obtain the health baseline parameters for all membrane elements in the RO membrane system to be in a healthy baseline state. The health baseline state is based on the water permeability coefficient and salt permeability coefficient of the membrane element under the factory inspection conditions. Based on the temperature data in the operating data, the health baseline parameters are corrected according to the temperature correction relationship provided by the membrane element manufacturer to obtain the corrected health baseline parameters under the current temperature conditions. S22. Based on the corrected health baseline parameters, and combined with the total feed water pressure, total feed water flow rate, and total feed water conductivity from the operating data, calculate the theoretical total permeate flow rate of the RO membrane system under the health baseline state according to the membrane mass transfer relationship. Q p0 Theoretical total concentrate flow rate Q c0 Theoretical total water production pipe conductivity C p0 Theoretical total concentrate pipe conductivity C c0 and theoretical high-pressure pump input power W 0 We obtain the theoretical output values ​​of the three conservation equations under a healthy baseline state; S23. Before performing the calculation of each conserved residual, a steady-state judgment is made on the change amplitude of the running data at adjacent sampling times; S24. For the sampling time determined by steady-state judgment, obtain the measured output value of each conservation equation, and subtract the measured output value of each conservation equation from the corresponding theoretical output value to obtain the mass conservation residual. Solute conservation residual and energy conservation residual The mass conservation residual, solute conservation residual, and energy conservation residual are combined to form a three-dimensional constrained residual vector. ; S25. Repeat the calculation process from S21 to S24 at consecutive effective sampling times, and arrange the three-dimensional constraint residual vector R at each effective time in time sequence to obtain the time sequence evolution sequence of the three-dimensional constraint residual vector.

4. The method for fault prediction and maintenance of RO membrane systems based on artificial intelligence according to claim 3, characterized in that: The steady-state determination of the change amplitude of running data at adjacent sampling times is as follows: Calculate the absolute change in total inlet pipe pressure and the relative percentage change in total inlet pipe flow rate between the current sampling time and the previous sampling time. Only when the absolute change in the total inlet water pressure is lower than the pressure steady-state threshold and the relative percentage change in the total inlet water flow is lower than the flow steady-state threshold, is the operating data at the current sampling time determined to be in steady state, and subsequent conserved residual calculation is performed. Otherwise, sampling time data that fail the steady-state judgment are marked and skipped, and are not included in the construction of the three-dimensional constrained residual vector.

5. The method for fault prediction and maintenance of RO membrane systems based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the membrane element-level fault prediction results are as follows: S31. Extract features from two independent dimensions of the temporal evolution sequence of the three-dimensional constrained residual vector to form a two-dimensional fingerprint feature vector; S32. The artificial intelligence model adopts a two-stage cascaded structure: the first stage is a failure mode classifier, which takes the two-dimensional fingerprint feature vector F as input, classifies four states: sealing failure, membrane perforation, scaling and normal operation, and outputs the current failure mode type and the corresponding confidence level. The second stage is the failure location interval classifier, which uses the two-dimensional fingerprint feature vector F and the failure mode type output from the first stage as joint inputs to locate the failure location interval under the constraint of the known failure mode type, and outputs the failure location interval and the corresponding failure probability. S33. During the formal operation phase of the RO membrane system, the two-dimensional fingerprint feature vector F of the three-dimensional constraint residual vector time-series evolution sequence is continuously extracted using a sliding time window. The vector is input into the trained artificial intelligence model and then processed through a two-stage process of failure mode classifier and failure location interval classifier. The failure mode type, failure location interval and corresponding failure probability of the current time window are output. The failure mode type, failure location range and corresponding failure probability after continuous window consistency confirmation are used as the membrane element level failure prediction result and output to step S4.

6. The method for fault prediction and maintenance of RO membrane systems based on artificial intelligence according to claim 5, characterized in that: The two-dimensional fingerprint feature vector is obtained as follows: The first dimension is the three residual ratio feature: within the current sliding time window, calculate the ratio between the mass conservation residual, solute conservation residual, and energy conservation residual, and obtain the ratio feature vector P after normalization; The second dimension is the temporal evolution characteristic: for the three-dimensional constraint residual vectors at consecutive moments within the sliding time window, calculate the magnitude of the three-dimensional constraint residual vector. The growth rate over time and the current cumulative magnitude constitute the temporal evolution feature vector D; The proportional feature vector P is combined with the temporal evolution feature vector D to form a two-dimensional fingerprint feature vector F=[P,D], which is used as the input to the artificial intelligence model.

7. The method for fault prediction and maintenance of RO membrane systems based on artificial intelligence according to claim 1, characterized in that: The steps for outputting maintenance instructions are as follows: Based on the membrane element-level fault prediction results, a fault location report is generated; Based on the failure probability corresponding to each failure location interval in the fault location report, the maintenance priority of each failure location interval is sorted. Based on the maintenance priority ranking result, corresponding maintenance instructions are generated for each failure location interval.

Images