Methods and Systems for Predicting Membrane Fouling in Membrane Bioreactors

Abstract

This invention discloses a method and system for predicting membrane fouling in membrane bioreactors, relating to the field of fouling detection technology. The method includes: setting a frequency detection strategy; connecting the membrane bioreactor and acquiring membrane feedback signals based on the frequency detection strategy; performing time-series characteristic analysis of feedback signal fluctuations based on the membrane feedback signals to obtain feedback signal characteristics; importing the feedback signal characteristics into a mechanistic hybrid prediction model in conjunction with reactor operating parameters, wherein a mechanistic constraint layer applies mechanistic constraints based on the feedback signal characteristics, and a data-driven prediction layer performs data parsing on the time-series data characteristics based on the mechanistic constraints to obtain membrane state prediction results; and performing membrane fouling state analysis and progress prediction based on the membrane state prediction results to generate a membrane fouling prediction data chain. This method solves the technical problems of existing technologies where membrane fouling prediction relies on empirical thresholds, leading to recognition lag and insufficient accuracy, and achieves the technical effect of improving the accuracy and real-time performance of membrane fouling prediction.

Description

Technical Field

[0001] This invention relates to the field of pollution detection technology, specifically to a method and system for predicting membrane fouling in membrane bioreactors. Background Technology

[0002] Membrane bioreactors (MBRs) are widely used in wastewater treatment and reuse due to their high efficiency in solid-liquid separation and excellent effluent quality. However, membrane fouling remains a persistent challenge to the long-term stable operation of this technology. With increasing operating time, the membrane surface becomes susceptible to adhesion and blockage by particulate matter, organic matter, and microorganisms, leading to increased transmembrane pressure and decreased flux. Existing technologies typically rely on empirical thresholds or fixed detection indicators to determine membrane fouling status, but this approach has significant limitations. On the one hand, membrane fouling involves a complex dynamic evolution process, and a single threshold cannot comprehensively reflect the membrane state under different operating conditions, easily resulting in predictive lag. On the other hand, monitoring data fluctuates significantly, and traditional methods lack effective dynamic analysis and trend identification, leading to insufficient accuracy in prediction results and difficulty in providing timely guidance for operation control and maintenance cleaning. Summary of the Invention

[0003] This application provides a method and system for predicting membrane fouling in membrane bioreactors, which solves the technical problems of recognition lag and insufficient accuracy caused by the reliance on empirical thresholds in the prior art for membrane fouling prediction.

[0004] A first aspect of this application provides a method for predicting membrane fouling in membrane bioreactors, the method comprising:

[0005] A frequency detection strategy is set up, and the membrane bioreactor is connected to acquire membrane feedback signals based on the frequency detection strategy. The frequency detection strategy includes a set reaction current below a preset threshold and a detection frequency. The time-series characteristics of feedback signal fluctuations are analyzed based on the membrane feedback signals to obtain feedback signal characteristics. The feedback signal characteristics are combined with reactor operating parameters and imported into a mechanistic hybrid prediction model. The mechanistic hybrid prediction model includes a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer applies mechanistic constraints based on the feedback signal characteristics, and the data-driven prediction layer performs data parsing on the time-series data characteristics based on the mechanistic constraints to obtain membrane state prediction results. Based on the membrane state prediction results, membrane fouling state analysis and progress prediction are performed to generate a membrane fouling prediction data chain.

[0006] A second aspect of this application provides a membrane fouling prediction system for membrane bioreactors, the system comprising:

[0007] Signal Acquisition Module: Sets a frequency detection strategy, connects to the membrane bioreactor, and acquires membrane feedback signals based on the frequency detection strategy. The frequency detection strategy includes a set reaction current below a preset threshold and a detection frequency. Feature Analysis Module: Performs time-series feature analysis of feedback signal fluctuations based on the membrane feedback signals to obtain feedback signal features. State Prediction Module: Imports the feedback signal features into a mechanistic hybrid prediction model combined with reactor operating parameters. The mechanistic hybrid prediction model includes a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer performs mechanistic constraints based on the feedback signal features, and the data-driven prediction layer performs data parsing on the time-series data features based on the mechanistic constraints to obtain membrane state prediction results. Progress Prediction Module: Performs membrane fouling state analysis and progress prediction based on the membrane state prediction results, generating a membrane fouling prediction data chain.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] First, a frequency detection strategy is set up, and the membrane bioreactor is connected to acquire membrane feedback signals based on this strategy. The frequency detection strategy includes a preset reaction current below a predefined threshold and a detection frequency. Next, the temporal characteristics of the feedback signal fluctuations are analyzed to obtain the feedback signal features. Then, the feedback signal features are combined with reactor operating parameters and imported into a mechanistic hybrid prediction model. This model includes a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer applies mechanistic constraints based on the feedback signal features, while the data-driven prediction layer analyzes the temporal data features based on these constraints to obtain membrane state prediction results. Finally, membrane fouling state analysis and progress prediction are performed based on the membrane state prediction results, generating a membrane fouling prediction data chain. This solves the technical problems of existing technologies where membrane fouling prediction relies on empirical thresholds, leading to recognition lag and insufficient accuracy, and achieves the technical effect of improving the accuracy and real-time performance of membrane fouling prediction. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the membrane fouling prediction method for membrane bioreactors provided in the embodiments of this application;

[0012] Figure 2 This is a schematic diagram of the membrane fouling prediction system for membrane bioreactors provided in an embodiment of this application.

[0013] Explanation of reference numerals in the attached diagram: Signal acquisition module 11, Feature analysis module 12, State prediction module 13, Progress prediction module 14. Detailed Implementation

[0014] This application provides a method and system for predicting membrane fouling in membrane bioreactors, which solves the technical problems of recognition lag and insufficient accuracy caused by the reliance on empirical thresholds in the prior art for membrane fouling prediction.

[0015] 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 them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0016] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0017] Example 1, as Figure 1 As shown, this application provides a method for predicting membrane fouling in membrane bioreactors, wherein the method includes:

[0018] A frequency detection strategy is set, and the membrane bioreactor is connected to obtain membrane feedback signals based on the frequency detection strategy. The frequency detection strategy includes a set reaction current below a preset threshold and a detection frequency.

[0019] In this embodiment, a frequency detection strategy for operational monitoring is established. This strategy includes setting the reaction current and detection frequency to ensure continuous acquisition of valid data without affecting the normal operation of the membrane. The membrane bioreactor is connected to the frequency detection strategy, and membrane feedback signals are acquired according to a preset detection method. These feedback signals reflect the electrochemical response characteristics of the membrane under different operating conditions.

[0020] Furthermore, a frequency detection strategy is set, and the membrane bioreactor acquires membrane feedback signals based on the frequency detection strategy, including:

[0021] Based on historical reaction samples of the membrane structure, a current threshold and excitation frequency range are set. The current threshold and excitation frequency range are numerical requirements that can reflect the electrochemical state of the membrane without affecting the membrane structure and causing biological interference. Based on the current threshold and excitation frequency range, the reaction current and detection frequency are configured, and the frequency detection strategy is set to apply a micro-amplitude multi-band AC excitation current to the membrane structure component according to the reaction current and detection frequency. The electrochemical impedance spectral response signal of the membrane structure component is collected in real time to obtain the membrane feedback signal.

[0022] Based on historical response samples of the membrane structure under different operating conditions, typical response patterns in electrochemical impedance spectroscopy (EIS) were extracted, and current thresholds and excitation frequency ranges were set accordingly. The current threshold should be lower than the limit that the membrane material can withstand during long-term operation to ensure that the membrane structure is not damaged during testing, while still sensitively reflecting changes in the electrochemical properties of the membrane surface and pores. The excitation frequency range is limited to a range that can characterize both membrane pore resistance and double-layer capacitance, as well as capture key characteristic parameters such as diffusion impedance, thereby avoiding unnecessary biological interference to the membrane-biointerface.

[0023] Based on the current threshold and excitation frequency range, the reaction current and detection frequency for periodic detection are configured to form a frequency-based detection strategy. This strategy controls the application of micro-amplitude, multi-frequency AC excitation current to the membrane structure components by the detection module, ensuring effective acquisition of the membrane electrochemical response without altering the reactor's normal operating conditions.

[0024] Under the frequency detection strategy, the electrochemical impedance spectral response signals of the membrane structure component at different frequency points are acquired in real time and output as membrane feedback signals. The membrane feedback signals include raw detection data such as impedance amplitude and phase angle.

[0025] Based on the membrane feedback signal, the timing characteristics of the feedback signal fluctuation are analyzed to obtain the feedback signal characteristics.

[0026] Furthermore, based on the membrane feedback signal, the timing characteristics of the feedback signal fluctuation are analyzed to obtain the feedback signal characteristics, including:

[0027] The acquired raw electrochemical impedance spectroscopy response signal is simulated using an equivalent circuit model to extract electrochemical characteristic parameters, including membrane pore resistance, double-layer capacitance, and diffusion impedance. Based on the electrochemical characteristic parameters, instantaneous characteristics and dynamic correlation time-series characteristics are calculated to obtain the feedback signal characteristics, which include real-time measured values, first-order differences, and sliding window standard deviations.

[0028] Specifically, firstly, the acquired raw electrochemical impedance spectroscopy response signal is modeled and analyzed. An equivalent circuit model is used to fit the impedance spectrum curve to extract electrochemical characteristic parameters. These parameters include at least membrane pore resistance, double-layer capacitance, and diffusion impedance. Membrane pore resistance reflects the degree of membrane pore blockage, double-layer capacitance characterizes the accumulation of charge on the membrane surface, and diffusion impedance reflects the transport resistance of solutes on the membrane surface and in the pores. Then, instantaneous features are extracted from the real-time acquired electrochemical parameters to obtain direct measurements at each monitoring time. The first-order difference is further calculated to characterize the rate of change between adjacent time points, thus reflecting the dynamic evolution trend of the fouling state. Based on this, the standard deviation of the parameters is calculated using a sliding time window to measure the degree of fluctuation within a certain time scale, revealing the stability or abrupt changes in the membrane fouling state. Finally, the real-time measurements, first-order difference, and sliding window standard deviation are combined to form the feedback signal characteristics.

[0029] The feedback signal characteristics are combined with reactor operating parameters and imported into a mechanistic hybrid prediction model. The mechanistic hybrid prediction model includes a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer performs mechanistic constraints based on the feedback signal characteristics, and the data-driven prediction layer performs data analysis on the time series data characteristics based on the mechanistic constraints to obtain membrane state prediction results.

[0030] Reactor operating parameters include transmembrane pressure difference, membrane flux, temperature, and aeration intensity. Feedback signal characteristics and reactor operating parameters are input into a mechanistic hybrid prediction model. This model consists of a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer is designed based on the physical mechanism equations of membrane fouling, transforming feedback signal characteristics and operating parameters into physically meaningful intermediate variables, such as fouling resistance and concentration polarization factor, to constrain the model's prediction process and ensure it conforms to the basic laws of membrane fouling evolution. The data-driven prediction layer uses a deep learning framework to analyze temporal features, such as using temporal convolutional networks or attention-enhanced recurrent neural networks, to perform multi-scale feature extraction and correlation modeling of the input feedback signal characteristics and operating parameters. Through the data-driven prediction layer, the dynamic trends of membrane fouling evolution over time and the nonlinear relationships between multiple parameters can be captured.

[0031] The intermediate variables output by the mechanism constraint layer and the temporal feature analysis results of the data-driven prediction layer work together after fusion to generate the membrane state prediction results.

[0032] Furthermore, the feedback signal characteristics, combined with reactor operating parameters, are imported into a mechanistic hybrid prediction model, which includes the following steps:

[0033] A training dataset is obtained, which includes historical time-series data of membrane bioreactors, including electrochemical impedance spectroscopy (EIS) characteristic parameters, key operating parameters, and corresponding membrane fouling state labels. A physical information neural network architecture is constructed, including a mechanism constraint layer based on the physical mechanism equation of membrane fouling and a data-driven prediction layer based on time-series data. Using the training dataset, the time-series data of the EIS characteristic parameters and key operating parameters are used as input, the membrane fouling state is used as the prediction target, and the conservation of the physical mechanism equation is used as a constraint. The physical information neural network is jointly trained so that the network framework satisfies the constraint of the mechanism law while minimizing the prediction error, resulting in a mechanistic hybrid prediction model that has completed training convergence. This model is used to receive real-time data and output membrane fouling prediction results.

[0034] First, a training dataset is acquired, derived from historical time-series data collected during the long-term operation of the membrane bioreactor. This historical time-series data includes at least electrochemical impedance spectroscopy (EIS) characteristic parameters (such as membrane pore resistance, double-layer capacitance, and diffusion impedance), key operating parameters (such as transmembrane pressure difference, membrane flux, temperature, and aeration intensity), and corresponding membrane fouling status labels. These labels can be calibrated based on the recovery status after periodic physical or chemical cleaning. Next, a physical information neural network architecture is constructed, consisting of a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer is designed based on the physical mechanism equations of membrane fouling, mapping the input characteristic parameters to physically meaningful intermediate variables to introduce constraints on the evolution of membrane fouling during training. The data-driven prediction layer is constructed based on time-series data-driven methods, such as recurrent neural networks, temporal convolutional networks, or attention-based enhancement networks, to perform nonlinear analysis and predictive modeling of the input time-series data. Finally, using the training dataset, the time series of electrochemical impedance spectroscopy characteristic parameters and key operating parameters were used as input, membrane fouling state was used as the prediction target, and the conservation of the physical mechanism equations was used as a constraint to jointly train the physical information neural network. During training, the model reduced data error by minimizing the difference between the predicted values ​​and the true labels, and ensured physical consistency by minimizing the residuals of the mechanistic equations. After iterative optimization until convergence, the trained hybrid mechanistic prediction model was obtained, which can be used in practical applications to receive real-time data and output membrane fouling prediction results.

[0035] Furthermore, the physical information neural network is jointly trained, wherein the loss function consists of a data loss term, a physical loss term, and a hyperparameter balancing the two loss weights. The data loss term is used to calculate the error between the model's predicted value and the true label; the physical loss term is used to calculate whether the model's predicted output satisfies the preset physical mechanism equation.

[0036] In the process of jointly training the physical information neural network, the constructed loss function consists of a data loss term, a physical loss term, and hyperparameters used to balance the weights of the two losses.

[0037] The data loss term measures the difference between the model's predictions and the true labels. It can be calculated using common methods such as mean squared error, mean absolute error, or cross-entropy loss. For example, when using membrane fouling status levels or fouling resistance values ​​as labels, the mean squared error function can be used to calculate the deviation between the predicted and true values, ensuring the model has high prediction accuracy.

[0038] The physical loss term is used to constrain whether the model output conforms to the preset physical mechanism equation. Specifically, the output of the intermediate layer of the neural network is defined as an intermediate variable with physical meaning. This variable is substituted into the membrane fouling physical mechanism equation (such as the relationship between fouling resistance and transmembrane pressure difference and flux decay), and the residual value of the equation is calculated. By taking the norm of the residual value and normalizing it, the physical loss term is obtained, which reflects whether the model output satisfies the mechanism law.

[0039] By introducing hyperparameters as weight balancing factors, the data loss term and the physical loss term are weighted and summed proportionally to form a joint loss function: ,in, Indicates data loss items, Represents the physical loss term. This is a hyperparameter used to adjust the balance between data-driven and physical constraints.

[0040] Furthermore, the methods for calculating the physical loss term include:

[0041] The intermediate layer outputs of the neural network are defined as intermediate variables with physical meaning; the obtained intermediate layer output intermediate variables are substituted into the physical mechanism equation, the equation residuals are calculated, and the physical loss term is obtained according to the norm of the equation residuals.

[0042] In the design of a physical information neural network, firstly, the output results of some intermediate layers are explicitly defined as physically meaningful intermediate variables, such as fouling resistance, membrane flux attenuation coefficient, or diffusion impedance parameters. Then, these intermediate layer output variables are substituted into the physical mechanism equations related to membrane fouling. These physical mechanism equations may include equations relating transmembrane pressure difference and membrane resistance, flux calculation equations based on Darcy's law, or membrane fouling kinetic evolution equations. By substituting the intermediate variables into the mechanism equations, the difference between the predicted output and the mechanism equations is calculated, yielding the corresponding equation residuals. Finally, norm operations are performed on the equation residuals, for example, using the L2 norm to measure the overall residual magnitude, and the calculation result is used as the physical loss term.

[0043] Furthermore, obtaining membrane state prediction results includes:

[0044] A time-series alignment relationship is established between the feedback signal features and reactor operating parameters, including transmembrane pressure difference, membrane flux, temperature, and aeration intensity. Using the time-series aligned preprocessed feedback signal features and reactor operating parameters, data feature transformation is performed through a mechanistic constraint layer to obtain physically meaningful intermediate variables. These intermediate variables and the original time-series features are jointly constructed into a model input feature sequence according to time windows. Under physical mechanism constraints, the input feature sequence of the mechanistic hybrid prediction model is used for analytical prediction through a data-driven prediction layer using a time-series feature analysis algorithm, outputting membrane state prediction results.

[0045] Specifically, a time-series alignment relationship is established between feedback signal features and reactor operating parameters. Specifically, timestamp matching and linear interpolation methods are used to synchronize data from different sampling frequencies. When sampling is missing, sliding window averaging and Kalman filtering are used to estimate the missing points, ensuring continuous availability of input data at a unified time scale. The time-series aligned feedback signal features and reactor operating parameters are then input into the mechanistic constraint layer for data feature transformation. Specifically, the input data is analyzed and constrained based on physical mechanism equations to extract physically meaningful intermediate variables, such as fouling resistance, mass transfer coefficient, or pore attenuation factor, to ensure that the model prediction process conforms to the basic laws of membrane fouling evolution. Next, the intermediate variables and original time-series features are jointly constructed into a model input feature sequence according to time windows. Specifically, using a preset time window as a unit, intermediate variables and original features at consecutive moments are combined into a feature vector sequence, forming a time-series input dataset that reflects the dynamic changes in membrane state. Finally, the input feature sequence is imported into the mechanistic hybrid prediction model. Under the constraints of physical mechanisms, the data-driven prediction layer uses a time-series feature parsing algorithm for analysis and prediction, and outputs the membrane state prediction results.

[0046] Furthermore, the data-driven prediction layer utilizes a temporal feature parsing algorithm for analytical prediction, outputting membrane state prediction results, including:

[0047] An attention-enhanced spatiotemporal graph convolutional network is used as the data-driven prediction layer algorithm framework. A spatiotemporal graph structure is constructed using time-series data features as input, where electrochemical sensors at different locations are used as graph nodes. The node features are multi-time-step electrochemical characteristic parameters, operating parameters, and intermediate variable vectors generated by the mechanistic constraint layer. The edge weights are determined by a weighted combination of spatial topological distance and signal correlation. The spatiotemporal graph convolutional module captures the spatial propagation characteristics of membrane fouling, and the temporal attention module dynamically emphasizes the influence of key time steps, outputting the membrane state prediction results for multiple future time steps, including: the trend of fouling resistance change, the distribution of fouling layer thickness at key locations, and the membrane flux decay curve.

[0048] Specifically, an attention-enhanced spatiotemporal graph convolutional network is employed as the algorithmic framework for the data-driven prediction layer. This framework can simultaneously model the spatial diffusion correlation and temporal dynamic evolution of membrane fouling, thus providing a more comprehensive reflection of the fouling process. Then, a spatiotemporal graph structure is constructed using time-series data of feedback signal features and reactor operating parameters as input. In the spatiotemporal graph, electrochemical sensors at different locations are defined as graph nodes. The feature vector of each graph node is composed of electrochemical characteristic parameters (such as membrane pore resistance, double-layer capacitance, and diffusion impedance) at multiple time points, operating parameters (such as transmembrane pressure difference, membrane flux, temperature, and aeration intensity), and intermediate variables generated by the mechanistic constraint layer. The weights of each edge in the graph are determined by a weighted combination of the spatial topological distance and signal correlation between sensors to reflect the spatial proximity and dynamic coupling of the membrane fouling propagation path. Next, the spatiotemporal graph convolutional module extracts features from the graph structure to capture the spatial propagation characteristics of fouling between different sensing nodes and to uncover the diffusion patterns of fouling on the membrane surface and in the pore structure. Simultaneously, a time attention module is introduced to dynamically assign weights to the time-series input data, highlighting the contribution of key time steps to pollution prediction and weakening the interference of irrelevant or redundant time slices. Finally, based on the above multi-dimensional feature analysis, the membrane state prediction results for multiple future time steps are output. The membrane state prediction results include the trend of pollution resistance over time, the spatial distribution of pollution layer thickness at key locations, and the membrane flux decay curve.

[0049] Based on the membrane state prediction results, membrane fouling state analysis and progress prediction are performed to generate a membrane fouling prediction data chain.

[0050] The membrane state prediction results output by the mechanism hybrid prediction model are analyzed. The membrane state prediction results include at least the trend of fouling resistance over time, the distribution of fouling layer thickness at key locations, and the membrane flux decay curve. By quantitatively analyzing the membrane state prediction results, the current state stage of membrane fouling can be identified, such as the initial deposition stage, the stable development stage, or the severe clogging stage.

[0051] Based on the identification of the fouling status, the prediction results are used to predict the progress. Specifically, by combining the rate of increase of transmembrane pressure difference and the rate of decrease of membrane flux, the development trend of membrane fouling in the future time period is estimated; by setting an early warning threshold, it is determined whether the membrane fouling status will reach the critical point of needing cleaning or replacement within a certain period of time, and the corresponding progress prediction results are output.

[0052] The results of the pollution status analysis and the progress prediction are integrated in a time series format to form a membrane fouling prediction data chain. The membrane fouling prediction data chain uses a time axis as an index to sequentially record the pollution status level, pollution development rate, critical time point, and corresponding operational risk assessment results within the prediction period.

[0053] Furthermore, it also includes:

[0054] Based on the membrane fouling prediction data chain, fouling intervention targets are obtained. Target electrical signal formulations are generated by matching from a pre-set electrical signal intervention strategy library. These target electrical signal formulations are combinations of specific waveforms, frequencies, current intensities, and durations used to suppress the currently diagnosed membrane fouling state. According to the target electrical signal formulation, corresponding intervention electrical signals are applied to the membrane module to change the physicochemical properties at the membrane-liquid interface and inhibit the deposition and adhesion of pollutants on the membrane surface. The electrical signal intervention strategy library is a training sample set containing various membrane fouling types, different electrical signal formulations, and their corresponding intervention effect data. Based on the training sample set, with membrane fouling type and risk level as input and the optimization of the fouling suppression effect after intervention as the objective, a reinforcement learning algorithm is used to train and generate the electrical signal intervention strategy library, which is used to output the most cost-effective electrical signal formulation.

[0055] Based on the membrane fouling prediction data chain, the development trend of membrane fouling in the future time period is analyzed to identify the fouling intervention targets under the current operating conditions. For example, when the prediction results show that the transmembrane pressure difference will rise rapidly within a preset threshold time, or the membrane flux decay rate exceeds the normal fluctuation range, it is determined that the target state requires triggering electrical signal intervention.

[0056] The system retrieves an electrical signal formulation that matches the pollution intervention target from a pre-set library of electrical signal intervention strategies. The target electrical signal formulation consists of a combination of specific waveform (such as square wave, pulse wave, or superimposed wave), frequency, current intensity, and duration parameters. It aims to slow down the development of membrane fouling by creating an electric field disturbance on the surface of the membrane module, thereby altering the electrochemical environment at the membrane-liquid interface, weakening the deposition and adhesion of pollutants, and thus mitigating the effects of membrane fouling.

[0057] Based on the target electrical signal formulation, a corresponding intervention electrical signal is applied to the membrane module. Specifically, an alternating current or pulsed current is applied through an external electrochemical control module to dynamically disturb the membrane surface potential distribution, thereby inhibiting the continuous accumulation of particulate matter, organic matter, and microorganisms and reducing the membrane fouling rate.

[0058] The establishment of the electrical signal intervention strategy library includes: constructing a training sample set containing various membrane fouling types (such as particulate blockage, organic deposition, and biofouling) and different electrical signal formulations and their corresponding intervention effects; based on this, using membrane fouling type and risk level as input parameters, and the optimization of the fouling inhibition effect after intervention as the objective function, a reinforcement learning algorithm is used for training iteration to form the electrical signal intervention strategy library.

[0059] In summary, the embodiments of this application have at least the following technical effects:

[0060] First, a frequency detection strategy is set up, and the membrane bioreactor is connected to acquire membrane feedback signals based on this strategy. The frequency detection strategy includes a preset reaction current below a predefined threshold and a detection frequency. Next, the temporal characteristics of the feedback signal fluctuations are analyzed to obtain the feedback signal features. Then, the feedback signal features are combined with reactor operating parameters and imported into a mechanistic hybrid prediction model. This model includes a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer applies mechanistic constraints based on the feedback signal features, while the data-driven prediction layer analyzes the temporal data features based on these constraints to obtain membrane state prediction results. Finally, membrane fouling state analysis and progress prediction are performed based on the membrane state prediction results, generating a membrane fouling prediction data chain. This solves the technical problems of existing technologies where membrane fouling prediction relies on empirical thresholds, leading to recognition lag and insufficient accuracy, and achieves the technical effect of improving the accuracy and real-time performance of membrane fouling prediction.

[0061] Example 2, based on the same inventive concept as the membrane bioreactor membrane fouling prediction method in the foregoing examples, such as... Figure 2 As shown, this application provides a membrane fouling prediction system for membrane bioreactors, wherein the system includes:

[0062] Signal acquisition module 11: Sets a frequency detection strategy, connects to the membrane bioreactor, and acquires membrane feedback signals based on the frequency detection strategy. The frequency detection strategy includes a set reaction current below a preset threshold and a detection frequency. Feature analysis module 12: Performs time-series feature analysis of feedback signal fluctuations based on the membrane feedback signals to obtain feedback signal features. State prediction module 13: Imports the feedback signal features into a mechanistic hybrid prediction model in combination with reactor operating parameters. The mechanistic hybrid prediction model includes a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer performs mechanistic constraints based on the feedback signal features, and the data-driven prediction layer performs data parsing on the time-series data features based on the mechanistic constraints to obtain membrane state prediction results. Progress prediction module 14: Performs membrane fouling state analysis and progress prediction based on the membrane state prediction results to generate a membrane fouling prediction data chain.

[0063] Furthermore, the signal acquisition module 11 is used to perform the following method:

[0064] Based on historical reaction samples of the membrane structure, a current threshold and excitation frequency range are set. The current threshold and excitation frequency range are numerical requirements that can reflect the electrochemical state of the membrane without affecting the membrane structure and causing biological interference. Based on the current threshold and excitation frequency range, the reaction current and detection frequency are configured, and the frequency detection strategy is set to apply a micro-amplitude multi-band AC excitation current to the membrane structure component according to the reaction current and detection frequency. The electrochemical impedance spectral response signal of the membrane structure component is collected in real time to obtain the membrane feedback signal.

[0065] Furthermore, the feature analysis module 12 is used to perform the following methods:

[0066] The acquired raw electrochemical impedance spectroscopy response signal is simulated using an equivalent circuit model to extract electrochemical characteristic parameters, including membrane pore resistance, double-layer capacitance, and diffusion impedance. Based on the electrochemical characteristic parameters, instantaneous characteristics and dynamic correlation time-series characteristics are calculated to obtain the feedback signal characteristics, which include real-time measured values, first-order differences, and sliding window standard deviations.

[0067] Furthermore, the state prediction module 13 is used to perform the following method:

[0068] A training dataset is obtained, which includes historical time-series data of membrane bioreactors, including electrochemical impedance spectroscopy (EIS) characteristic parameters, key operating parameters, and corresponding membrane fouling state labels. A physical information neural network architecture is constructed, including a mechanism constraint layer based on the physical mechanism equation of membrane fouling and a data-driven prediction layer based on time-series data. Using the training dataset, the time-series data of the EIS characteristic parameters and key operating parameters are used as input, the membrane fouling state is used as the prediction target, and the conservation of the physical mechanism equation is used as a constraint. The physical information neural network is jointly trained so that the network framework satisfies the constraint of the mechanism law while minimizing the prediction error, resulting in a mechanistic hybrid prediction model that has completed training convergence. This model is used to receive real-time data and output membrane fouling prediction results.

[0069] Furthermore, the state prediction module 13 is used to perform the following method:

[0070] The physical information neural network is jointly trained, wherein the loss function consists of a data loss term, a physical loss term, and a hyperparameter balancing the two loss weights. The data loss term is used to calculate the error between the model's predicted value and the true label; the physical loss term is used to calculate whether the model's predicted output satisfies the preset physical mechanism equation.

[0071] Furthermore, the state prediction module 13 is used to perform the following method:

[0072] The intermediate layer outputs of the neural network are defined as intermediate variables with physical meaning; the obtained intermediate layer output intermediate variables are substituted into the physical mechanism equation, the equation residuals are calculated, and the physical loss term is obtained according to the norm of the equation residuals.

[0073] Furthermore, the state prediction module 13 is used to perform the following method:

[0074] A time-series alignment relationship is established between the feedback signal features and reactor operating parameters, including transmembrane pressure difference, membrane flux, temperature, and aeration intensity. Using the time-series aligned preprocessed feedback signal features and reactor operating parameters, data feature transformation is performed through a mechanistic constraint layer to obtain physically meaningful intermediate variables. These intermediate variables and the original time-series features are jointly constructed into a model input feature sequence according to time windows. Under physical mechanism constraints, the input feature sequence of the mechanistic hybrid prediction model is used for analytical prediction through a data-driven prediction layer using a time-series feature analysis algorithm, outputting membrane state prediction results.

[0075] Furthermore, the state prediction module 13 is used to perform the following method:

[0076] An attention-enhanced spatiotemporal graph convolutional network is used as the data-driven prediction layer algorithm framework. A spatiotemporal graph structure is constructed using time-series data features as input, where electrochemical sensors at different locations are used as graph nodes. The node features are multi-time-step electrochemical characteristic parameters, operating parameters, and intermediate variable vectors generated by the mechanistic constraint layer. The edge weights are determined by a weighted combination of spatial topological distance and signal correlation. The spatiotemporal graph convolutional module captures the spatial propagation characteristics of membrane fouling, and the temporal attention module dynamically emphasizes the influence of key time steps, outputting the membrane state prediction results for multiple future time steps, including: the trend of fouling resistance change, the distribution of fouling layer thickness at key locations, and the membrane flux decay curve.

[0077] Furthermore, the progress prediction module 14 is used to perform the following method:

[0078] Based on the membrane fouling prediction data chain, fouling intervention targets are obtained. Target electrical signal formulations are generated by matching from a pre-set electrical signal intervention strategy library. These target electrical signal formulations are combinations of specific waveforms, frequencies, current intensities, and durations used to suppress the currently diagnosed membrane fouling state. According to the target electrical signal formulation, corresponding intervention electrical signals are applied to the membrane module to change the physicochemical properties at the membrane-liquid interface and inhibit the deposition and adhesion of pollutants on the membrane surface. The electrical signal intervention strategy library is a training sample set containing various membrane fouling types, different electrical signal formulations, and their corresponding intervention effect data. Based on the training sample set, with membrane fouling type and risk level as input and the optimization of the fouling suppression effect after intervention as the objective, a reinforcement learning algorithm is used to train and generate the electrical signal intervention strategy library, which is used to output the most cost-effective electrical signal formulation.

[0079] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0080] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0081] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for predicting membrane fouling in membrane bioreactors, characterized in that, The method includes: A frequency detection strategy is set, and the membrane bioreactor is connected to obtain membrane feedback signals based on the frequency detection strategy. The frequency detection strategy includes a set reaction current below a preset threshold and a detection frequency. Based on the membrane feedback signal, the timing characteristics of the feedback signal fluctuation are analyzed to obtain the feedback signal characteristics. The feedback signal characteristics are combined with reactor operating parameters and imported into a mechanistic hybrid prediction model. The mechanistic hybrid prediction model includes a mechanistic constraint layer and a data-driven prediction layer. The mechanistic constraint layer performs mechanistic constraints based on the feedback signal characteristics, and the data-driven prediction layer performs data analysis on the time series data characteristics based on the mechanistic constraints to obtain membrane state prediction results. Based on the membrane state prediction results, membrane fouling state analysis and progress prediction are performed to generate a membrane fouling prediction data chain. The obtained membrane state prediction results include: Establish a time-series alignment relationship between the feedback signal characteristics and reactor operating parameters, wherein the reactor operating parameters include transmembrane pressure difference, membrane flux, temperature, and aeration intensity; By utilizing the feedback signal characteristics of time-series aligned preprocessing and reactor operating parameters, data feature transformation is performed through a mechanistic constraint layer to obtain physically meaningful intermediate variables; The intermediate variables and the original time-series features are jointly constructed into a model input feature sequence according to a time window. Under the constraints of physical mechanism, the input feature sequence of the mechanism hybrid prediction model is analyzed and predicted by the data-driven prediction layer using a time-series feature parsing algorithm, and the membrane state prediction result is output.

2. The method for predicting membrane fouling in membrane bioreactors according to claim 1, characterized in that, The frequency detection strategy, which connects the membrane bioreactor to obtain membrane feedback signals based on the frequency detection strategy, includes: The current threshold and excitation frequency range are set based on historical reaction samples of the membrane structure. The current threshold and excitation frequency range are numerical requirements that can reflect the electrochemical state of the membrane and do not affect the biological interference generated by the membrane structure. Based on the current threshold and excitation frequency range, the reaction current and detection frequency are configured, and the frequency detection strategy is set to apply a micro-amplitude multi-band AC excitation current to the membrane structure component according to the reaction current and detection frequency, and to collect the electrochemical impedance spectral response signal of the membrane structure component in real time to obtain the membrane feedback signal.

3. The method for predicting membrane fouling in membrane bioreactors according to claim 2, characterized in that, Based on the membrane feedback signal, the timing characteristics of the feedback signal fluctuation are analyzed to obtain the feedback signal characteristics, including: The original electrochemical impedance spectroscopy response signal was simulated using an equivalent circuit model to extract electrochemical characteristic parameters, including membrane pore resistance, double layer capacitance, and diffusion impedance. The instantaneous characteristics and dynamic correlation time-series characteristics are calculated based on the electrochemical characteristic parameters to obtain the feedback signal characteristics, which include real-time measured values, first-order differences, and sliding window standard deviations.

4. The method for predicting membrane fouling in membrane bioreactors according to claim 1, characterized in that, The feedback signal characteristics, combined with reactor operating parameters, are imported into a mechanistic hybrid prediction model, which includes the following: Obtain a training dataset, which includes historical time-series data of membrane bioreactors, including electrochemical impedance spectroscopy characteristic parameters, key operating parameters, and corresponding membrane fouling status labels. A physical information neural network architecture is constructed, including a mechanism constraint layer based on the physical mechanism equation of membrane fouling and a data-driven prediction layer based on time-series data. Using the training dataset, the time series data of electrochemical impedance spectroscopy characteristic parameters and key operating parameters are used as input, membrane fouling state is used as the prediction target, and the conservation of physical mechanism equations is used as a constraint. The physical information neural network is jointly trained so that the network framework satisfies the constraint of the mechanism law while minimizing the prediction error. The resulting hybrid mechanism prediction model is used to receive real-time data and output membrane fouling prediction results.

5. The method for predicting membrane fouling in membrane bioreactors according to claim 4, characterized in that, The physical information neural network is jointly trained, wherein the loss function consists of a data loss term, a physical loss term, and a hyperparameter balancing the two loss weights. The data loss term is used to calculate the error between the model's predicted value and the true label; the physical loss term is used to calculate whether the model's predicted output satisfies the preset physical mechanism equation.

6. The method for predicting membrane fouling in a membrane bioreactor according to claim 5, characterized in that, The calculation method for the physical loss term includes: Define some of the intermediate layer outputs of the neural network as intermediate variables with physical meaning; The intermediate variables obtained from the intermediate layer output are substituted into the physical mechanism equation to calculate the equation residuals. Based on the norm of the equation residuals, the physical loss term is obtained.

7. The method for predicting membrane fouling in membrane bioreactors according to claim 1, characterized in that, The data-driven prediction layer utilizes a time-series feature parsing algorithm for analytical prediction, outputting membrane state prediction results, including: An algorithmic framework is proposed that uses a spatiotemporal graph convolutional network enhanced with an attention mechanism as the data-driven prediction layer. A spatiotemporal graph structure is constructed using time-series data features as input, where electrochemical sensors at different locations are used as graph nodes. The node features are electrochemical characteristic parameters, operating parameters, and intermediate variable vectors generated by the mechanism constraint layer at multiple time points. The edge weights are determined by a weighted combination of spatial topological distance and signal correlation. The spatial propagation characteristics of membrane fouling are captured by the spatiotemporal graph convolution module, and the impact of key time steps is dynamically emphasized by the time attention module. The results of membrane state prediction for multiple future time steps are output, including: the trend of fouling resistance, the distribution of fouling layer thickness at key locations, and the membrane flux decay curve.

8. The method for predicting membrane fouling in a membrane bioreactor according to claim 1, characterized in that, Also includes: Based on the membrane fouling prediction data chain, the fouling intervention target is obtained, and the target electrical signal formula is generated by matching from the preset electrical signal intervention strategy library. The target electrical signal formula is used to suppress the combination of specific waveform, frequency, current intensity and duration parameters of the currently diagnosed membrane fouling state. According to the target electrical signal formulation, a corresponding intervention electrical signal is applied to the membrane module to change the physicochemical properties at the membrane-liquid interface and inhibit the deposition and adhesion of pollutants on the membrane surface. The electrical signal intervention strategy library is generated by establishing a training sample set containing data on various membrane fouling types, different electrical signal formulations, and their corresponding intervention effects. Based on the training sample set, the library is trained using a reinforcement learning algorithm with membrane fouling type and risk level as input and optimization of the fouling inhibition effect after intervention as the objective. This library is used to output the most cost-effective electrical signal formulation.

9. A membrane fouling prediction system for membrane bioreactors, characterized in that, For implementing the membrane bioreactor membrane fouling prediction method according to any one of claims 1-8, the system comprises: Signal acquisition module: Sets a frequency detection strategy, connects to the membrane bioreactor, and acquires membrane feedback signals based on the frequency detection strategy. The frequency detection strategy includes a set reaction current below a preset threshold and a detection frequency. Feature analysis module: Based on the membrane feedback signal, perform time-series feature analysis of feedback signal fluctuations to obtain feedback signal features; State prediction module: The feedback signal characteristics are combined with reactor operating parameters and imported into the mechanism hybrid prediction model. The mechanism hybrid prediction model includes a mechanism constraint layer and a data-driven prediction layer. The mechanism constraint layer performs mechanism constraints based on the feedback signal characteristics, and the data-driven prediction layer performs data analysis on the time series data characteristics based on the mechanism constraints to obtain the membrane state prediction results. Progress prediction module: Based on the membrane state prediction results, it performs membrane fouling state analysis and progress prediction, and generates a membrane fouling prediction data chain.

Images