AI-based intelligent optimization method and system for chemical cleaning of ultrafiltration membranes

Abstract

This invention relates to the field of water treatment technology, and more particularly to an AI-based intelligent optimization method and system for chemical cleaning of ultrafiltration membranes. It acquires operational data and adaptively decomposes it into multi-scale time-series components. Based on an interaction influence matrix, it quantifies the interaction contribution to generate a pollution characteristic vector. Through source separation and decoupling, it identifies independent pollution source components and determines their types and weights. Based on the pollution source types, it establishes a collaborative relationship network for cleaning agents, generating multi-cleaning agent combination formulations and phased cleaning time-series schemes. During cleaning, it collects membrane flux response in real time and dynamically adjusts the dosing rate and time-series parameters based on deviations from preset targets until the membrane flux reaches the target. This invention achieves intelligent optimization of chemical cleaning, improves cleaning efficiency and membrane flux recovery, and reduces reagent usage and operating costs.

Description

Technical Field

[0001] This invention relates to the field of water treatment technology, and in particular to an AI-based intelligent optimization method and system for chemical cleaning of ultrafiltration membranes. Background Technology

[0002] In the actual operation and maintenance of ultrafiltration membrane systems, chemical cleaning is a key means of restoring membrane flux. Current conventional practices typically rely on operator experience or preset fixed parameters to perform cleaning operations. For example, cleaning is triggered when the inter-membrane pressure difference or permeate flow rate drops to a set threshold; the type and concentration of cleaning agents often use standard formulations; and the cleaning process follows a pre-defined sequence, such as a fixed order of alkaline washing followed by acid washing. While these practices have accumulated a wealth of empirical data through long-term industrial practice, they are essentially open-loop control strategies, lacking the ability to perceive and respond to real-time changes in membrane fouling status.

[0003] Because membrane fouling is usually caused by a combination of multiple contaminants, which may have synergistic or antagonistic effects, fixed formulations cannot adapt to the dynamic evolution of fouling components in actual operating conditions. For example, when organic and inorganic fouling are mixed on the membrane surface, simple alkaline or acid washing is insufficient to effectively remove the complex fouling layer, resulting in low cleaning efficiency and even accelerated membrane aging due to residual contaminants. Existing cleaning solutions cannot be dynamically adjusted based on membrane flux recovery feedback during the cleaning process. When the actual cleaning effect deviates from the expected target, operators can only rely on post-cleaning sampling or visual judgment to correct parameters, which often leads to excessively long cleaning times, wasted reagents, or frequent shortening of cleaning cycles due to incomplete cleaning.

[0004] Furthermore, conventional methods for identifying pollution sources primarily rely on offline detection techniques, such as microscopic observation of membrane surface morphology or chemical analysis of pollutants. These methods are time-consuming and difficult to apply in real-time during the cleaning process. Therefore, existing technologies generally lack the ability to quantitatively analyze multi-component pollutants, making it impossible to support the generation and online optimization of targeted cleaning strategies. Summary of the Invention

[0005] This invention provides an AI-based intelligent optimization method and system for chemical cleaning of ultrafiltration membranes, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an AI-based intelligent optimization method for chemical cleaning of ultrafiltration membranes, comprising:

[0007] The system acquires the operating data of the ultrafiltration membrane system, and automatically determines the decomposition parameters based on the signal characteristics of the operating data using an adaptive decomposition algorithm, and decomposes the operating data into multi-scale time-series components.

[0008] A correlation analysis was performed on the multi-scale time series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time series component was quantified. The interaction contribution was then fused with the time series components to generate a pollution feature characterization vector.

[0009] Source separation processing is performed on the pollution feature characterization vector to decouple the composite pollution signal into multiple independent pollution source components, and the pollution source type and pollution contribution weight corresponding to each independent pollution source component are identified.

[0010] Establish a synergistic relationship network between cleaning agents based on the type of pollution source, and generate multi-cleaning agent combination formulas and phased cleaning sequence schemes based on the synergistic relationship network and pollution contribution weights;

[0011] The ultrafiltration membrane cleaning system is controlled to start chemical cleaning according to the phased cleaning sequence plan. Membrane flux response data is collected in real time during the cleaning process. The deviation between the membrane flux response data and the preset recovery target trajectory is calculated. The dosing rate of the multi-cleaning agent combination formula and the cleaning sequence parameters are dynamically adjusted according to the deviation. Chemical cleaning is continuously performed until the membrane flux reaches the preset recovery target.

[0012] An adaptive decomposition algorithm automatically determines the decomposition parameters based on the signal characteristics of the running data and decomposes the running data into multi-scale time-series components, including:

[0013] The frequency domain transformation of the running data is performed to obtain the spectral distribution characteristics, and the energy peak position and energy distribution width are extracted from the spectral distribution characteristics;

[0014] The set of center frequencies of the time series components is determined based on the energy peak location. The bandwidth parameters corresponding to each center frequency are calculated based on the energy distribution width. The number of decomposition levels is determined as the decomposition parameters based on the set of center frequencies and the bandwidth parameters.

[0015] Multiple bandpass filters are constructed based on the decomposed parameters. The passband range of each bandpass filter is determined by the center frequency and bandwidth parameters.

[0016] The running data is filtered sequentially through multiple sets of bandpass filters. Each set of bandpass filters outputs a time component, and all time components constitute a multi-scale time component.

[0017] Orthogonality verification is performed on the multi-scale time series components. When the correlation coefficient between time series components exceeds the preset threshold, the bandwidth parameter of the corresponding bandpass filter is adjusted and the filtering process is re-executed until the correlation coefficient between all time series components is lower than the preset threshold, and the orthogonalized multi-scale time series components are output.

[0018] A correlation analysis is performed on the multi-scale time-series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time-series component is quantified. The interaction contribution is then fused with the time-series components to generate a pollution feature representation vector, including:

[0019] Time delay correlation analysis is performed on multi-scale time series components. The cross-correlation coefficients of any two time series components under different time delays are calculated, and the peak cross-correlation coefficient when the cross-correlation coefficient reaches its peak value is identified as the correlation strength of the time series component pair.

[0020] Construct an interaction influence matrix, where the rows and columns of the matrix correspond to multi-scale time series components. Fill the corresponding positions of the interaction influence matrix with the correlation strength. Calculate the sum of the elements in each row of the interaction influence matrix as the output influence strength of the corresponding time series component, and calculate the sum of the elements in each column of the interaction influence matrix as the input influence strength of the corresponding time series component.

[0021] The interaction contribution of the corresponding time series component is obtained by adding the output influence strength to the input influence strength;

[0022] The numerical sequence of each time series component is multiplied element-wise with the corresponding interaction contribution, and the multiplied time series components are concatenated to form a pollution feature representation vector.

[0023] Source separation processing is performed on the pollution feature representation vector to decouple the composite pollution signal into multiple independent pollution source components. The pollution source type and pollution contribution weight corresponding to each independent pollution source component are identified, including:

[0024] Standard evolution patterns are extracted from a pre-defined pollution source feature library as source separation constraints. The pollution feature representation vector is then subjected to source decomposition using these constraints to obtain a set of candidate pollution source components.

[0025] The reconstructed signal is obtained by linearly superimposing the candidate pollution source component sets. The difference between the reconstructed signal and the pollution feature representation vector is calculated to obtain the separation residual signal. When the energy of the separation residual signal exceeds the preset residual threshold, the residual pattern is extracted from the separation residual signal and matched with the preset pollution source feature library to obtain the supplementary pollution source component.

[0026] The supplementary pollution source components are incorporated into the candidate pollution source component set and the constraint source decomposition is re-executed. The iteration is performed until the energy of the separated residual signal converges, and the converged candidate pollution source component set is output as the independent pollution source component.

[0027] Temporal morphological features are extracted from each independent pollution source component. The temporal morphological features are then matched with the standard evolution patterns in the preset pollution source feature library to determine the pollution source type corresponding to each independent pollution source component. The pollution contribution weight of the corresponding pollution source type is obtained by numerically integrating each independent pollution source component in the time dimension.

[0028] A synergistic effect network among cleaning agents is established based on the type of pollution source. Based on this network and pollution contribution weights, multi-cleaning agent combination formulations and phased cleaning sequence plans are generated, including:

[0029] The removal efficiency of cleaning agents was determined for each type of pollution source when acting alone, in combination, and in stages. The synergy coefficient was calculated based on the removal efficiency of combined and individual actions, and the influence coefficient was calculated based on the impact of the preceding cleaning agent on the removal efficiency of the subsequent cleaning agent.

[0030] The synergy coefficient is used as the strength of the synergy between cleaning agent nodes, and the influence coefficient is used as the strength of the temporal interaction between cleaning agent nodes to construct a synergy relationship network of cleaning agents.

[0031] The pollution contribution weights are mapped to the target removal amount of each pollution source type. Based on the target removal amount and removal efficiency, cleaning agent nodes are selected from the cleaning agent synergy relationship network, and the synergy strength and time-series effect strength between the selected cleaning agent nodes are extracted.

[0032] When the synergistic effect strength is greater than the preset synergistic threshold, the corresponding cleaning agents are combined into the same cleaning stage and the concentration ratio is calculated based on the target removal amount and removal efficiency. When the synergistic effect strength is less than the preset inhibition threshold, the corresponding cleaning agents are separated into different cleaning stages. The order of cleaning stages is determined according to the temporal effect strength, and a multi-cleaning agent combination formula and a phased cleaning sequence plan are generated.

[0033] Calculate the deviation between the membrane flux response data and the preset recovery target trajectory. Based on the deviation, dynamically adjust the dosing rate and cleaning sequence parameters of the multi-cleaning agent combination formulation, and continuously perform chemical cleaning until the membrane flux reaches the preset recovery target, including:

[0034] Trend fitting is performed on the membrane flux response data to extract the instantaneous evolution slope sequence, and trend fitting is performed on the preset recovery target trajectory to extract the target evolution slope sequence;

[0035] The numerical difference between the membrane flux response data and the preset recovery target trajectory is calculated as the numerical deviation, and the difference between the instantaneous evolution slope sequence and the target evolution slope sequence is calculated as the trend deviation. The numerical deviation and the trend deviation are weighted and fused to obtain the deviation amount.

[0036] The cleaning response time constant is extracted from each independent pollution source component. The ratio of trend deviation to numerical deviation is calculated as the deviation dominance coefficient. Based on the deviation dominance coefficient, the weight of the deviation in the acceleration rate adjustment dimension and the cleaning time sequence parameter adjustment dimension is determined. The acceleration rate adjustment and cleaning time sequence parameter adjustment for each cleaning agent in the multi-cleaning agent combination formula are calculated by combining the cleaning response time constant and the weight. The acceleration rate and cleaning time sequence parameters of the multi-cleaning agent combination formula are adjusted according to the acceleration rate adjustment and the cleaning time sequence parameter adjustment.

[0037] The instantaneous evolution slope sequence is extrapolated, extended, and integrated to reconstruct the predicted membrane flux trajectory. When the fluctuation amplitude of the predicted membrane flux trajectory meets the stability condition, the membrane flux is determined to have reached the preset recovery target and chemical cleaning is stopped.

[0038] Extrapolating and integrating the instantaneous evolution slope sequence to reconstruct the predicted membrane flux trajectory includes:

[0039] The re-pollution rate after cleaning is extracted from each independent pollution source component, and the instantaneous evolution slope sequence is decoupled according to the pollution source type to obtain the slope component set.

[0040] The recontamination rate after cleaning is converted into an extension correction coefficient. Based on the extension correction coefficient, each slope component in the slope component set is extrapolated and extended. The extrapolated slope components are then superimposed to obtain the extension slope sequence.

[0041] Extract the switching time of each cleaning stage in the phased cleaning sequence scheme, introduce the synergistic coefficient of the cleaning agent interaction at the switching time, and enhance and correct the slope value of the extension slope sequence at the switching time according to the synergistic coefficient to obtain the synergistic slope sequence.

[0042] The flux increment sequence is obtained by integrating the efficiency slope sequence. The flux value at the starting point of the membrane flux response data is added to the flux increment sequence to obtain the predicted membrane flux trajectory.

[0043] A second aspect of this invention provides an AI-based intelligent optimization system for chemical cleaning of ultrafiltration membranes, comprising:

[0044] Adaptive sub-units are used to acquire the operating data of the ultrafiltration membrane system. The adaptive decomposition algorithm automatically determines the decomposition parameters based on the signal characteristics of the operating data and decomposes the operating data into multi-scale time-series components.

[0045] The pollution feature unit is used to perform correlation analysis on multi-scale time series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time series component is quantified, and the interaction contribution is fused with the time series components to generate a pollution feature characterization vector.

[0046] The source separation unit is used to perform source separation processing on the pollution feature characterization vector, decouple the composite pollution signal into multiple independent pollution source components, and identify the pollution source type and pollution contribution weight corresponding to each independent pollution source component.

[0047] The cleaning scheme unit is used to establish a synergistic relationship network between cleaning agents based on the type of pollution source, and to generate multi-cleaning agent combination formulas and phased cleaning sequence schemes based on the synergistic relationship network and pollution contribution weights.

[0048] The dynamic control unit is used to control the ultrafiltration membrane cleaning system to start chemical cleaning according to the phased cleaning sequence plan, collect membrane flux response data in real time during the cleaning process, calculate the deviation between the membrane flux response data and the preset recovery target trajectory, and dynamically adjust the dosing rate and cleaning sequence parameters of the multi-cleaning agent combination formula according to the deviation, and continuously perform chemical cleaning until the membrane flux reaches the preset recovery target.

[0049] A third aspect of the present invention provides an electronic device, comprising:

[0050] processor;

[0051] Memory used to store processor-executable instructions;

[0052] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0053] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0054] This invention employs an adaptive decomposition algorithm to automatically determine decomposition parameters based on signal characteristics, decomposing operational data into multi-scale time-series components. This avoids signal distortion and detail loss caused by traditional fixed decomposition methods, fully preserving the dynamic characteristics of contamination signals at different time scales. The interaction matrix constructed through correlation analysis quantifies the interactions between time-series components, and the generated contamination feature representation vector comprehensively reflects the multi-dimensional coupling characteristics of membrane fouling, providing a reliable foundation for accurate contamination type identification. Source separation processing decouples the composite contamination signal into multiple independent contamination source components, accurately identifying each contamination source type and its contribution weight. Based on a synergistic relationship network reflecting the synergistic or antagonistic relationships between different cleaning agents, the resulting multi-cleaning agent combination formulation and phased cleaning sequence scheme are highly targeted and synergistic, achieving the most efficient contamination removal with the minimum cleaning agent dosage, avoiding the problems of insufficient effect of a single cleaning agent or mutual interference between multiple cleaning agents. During the chemical cleaning process, membrane flux response data is collected in real time, compared with the preset recovery target trajectory to calculate the deviation, and the cleaning agent dosing acceleration rate and timing parameters are dynamically adjusted based on the deviation, ensuring that the cleaning process always approaches the target along the optimal path. This closed-loop feedback control effectively responds to changes in pollution and unpredictable factors, precisely controls membrane flux to achieve the recovery target, and eliminates over-cleaning or under-cleaning caused by fixed cleaning schemes. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating an AI-based intelligent optimization method for chemical cleaning of ultrafiltration membranes.

[0056] Figure 2 A flowchart is generated for the synergistic effect of cleaning agents and the timing scheme. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0058] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0059] Figure 1 This is a schematic diagram of the AI-based intelligent optimization method for chemical cleaning of ultrafiltration membranes according to an embodiment of the present invention.

[0060] AI-based intelligent optimization methods for chemical cleaning of ultrafiltration membranes include:

[0061] The system acquires the operating data of the ultrafiltration membrane system, and automatically determines the decomposition parameters based on the signal characteristics of the operating data using an adaptive decomposition algorithm, and decomposes the operating data into multi-scale time-series components.

[0062] A correlation analysis was performed on the multi-scale time series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time series component was quantified. The interaction contribution was then fused with the time series components to generate a pollution feature characterization vector.

[0063] Source separation processing is performed on the pollution feature characterization vector to decouple the composite pollution signal into multiple independent pollution source components, and the pollution source type and pollution contribution weight corresponding to each independent pollution source component are identified.

[0064] Establish a synergistic relationship network between cleaning agents based on the type of pollution source, and generate multi-cleaning agent combination formulas and phased cleaning sequence schemes based on the synergistic relationship network and pollution contribution weights;

[0065] The ultrafiltration membrane cleaning system is controlled to start chemical cleaning according to the phased cleaning sequence plan. Membrane flux response data is collected in real time during the cleaning process. The deviation between the membrane flux response data and the preset recovery target trajectory is calculated. The dosing rate of the multi-cleaning agent combination formula and the cleaning sequence parameters are dynamically adjusted according to the deviation. Chemical cleaning is continuously performed until the membrane flux reaches the preset recovery target.

[0066] An adaptive decomposition algorithm automatically determines the decomposition parameters based on the signal characteristics of the running data and decomposes the running data into multi-scale time-series components, including:

[0067] The frequency domain transformation of the running data is performed to obtain the spectral distribution characteristics, and the energy peak position and energy distribution width are extracted from the spectral distribution characteristics;

[0068] The set of center frequencies of the time series components is determined based on the energy peak location. The bandwidth parameters corresponding to each center frequency are calculated based on the energy distribution width. The number of decomposition levels is determined as the decomposition parameters based on the set of center frequencies and the bandwidth parameters.

[0069] Multiple bandpass filters are constructed based on the decomposed parameters. The passband range of each bandpass filter is determined by the center frequency and bandwidth parameters.

[0070] The running data is filtered sequentially through multiple sets of bandpass filters. Each set of bandpass filters outputs a time component, and all time components constitute a multi-scale time component.

[0071] Orthogonality verification is performed on the multi-scale time series components. When the correlation coefficient between time series components exceeds the preset threshold, the bandwidth parameter of the corresponding bandpass filter is adjusted and the filtering process is re-executed until the correlation coefficient between all time series components is lower than the preset threshold, and the orthogonalized multi-scale time series components are output.

[0072] After acquiring the operating data of the ultrafiltration membrane system, it is first subjected to frequency domain transformation. Specifically, a Fast Fourier Transform (FFT) is applied to the collected operating data (including time-series signals such as transmembrane pressure difference, membrane flux, influent and effluent turbidity, and temperature) to convert the time-domain signals into a frequency-domain representation, obtaining the spectral distribution characteristics of each operating parameter. The spectral distribution characteristics fully characterize the energy distribution of the operating data in different frequency ranges, providing a basis for subsequent adaptive parameter determination. During the spectral analysis, two key features are extracted: first, the energy peak position, i.e., the frequency coordinates of local energy maxima in the spectrum, reflecting the main periodic or quasi-periodic components present in the operating data; second, the energy distribution width, i.e., the frequency range around each energy peak where energy is significantly concentrated, reflecting the degree of frequency diffusion of the corresponding periodic components.

[0073] Based on the extracted energy peak positions, determine the set of center frequencies of the time-series components. Assume that a total of [number] frequencies are detected in the spectrum. The energy peaks, and their corresponding center frequencies, are denoted as follows: Then the set of center frequencies is Each center frequency ( Each frequency band corresponds to an independent time component, representing a physical process within a specific frequency range of the operational data. For example, the low-frequency band corresponds to a long-term trend of membrane fouling accumulation, the mid-frequency band corresponds to periodic fluctuations caused by cleaning operations, and the high-frequency band corresponds to short-term noise or transient disturbances. Simultaneously, based on the energy distribution width at each energy peak... Calculate the corresponding bandwidth parameters Bandwidth parameters With energy distribution width Between ,in This is the bandwidth adjustment factor, typically set between 0.8 and 1.2, used to balance filter bandwidth and frequency resolution. When the frequency interval between two adjacent energy peaks is small, it should be appropriately reduced. To prevent filter passband overlap; when the energy distribution is relatively dispersed, appropriately increase the [value / size]. To ensure the complete capture of the target components, the number of decomposition levels is determined based on the center frequency set and bandwidth parameters. As the core decomposition parameter of the adaptive decomposition algorithm, it completes the automatic determination process of decomposition parameters without the need for manual pre-setting of a fixed number of decomposition layers.

[0074] Based on the defined decomposition parameters, construct Groups of bandpass filters. The passband range of each group of bandpass filters is defined by its corresponding center frequency. and bandwidth parameters The cutoff frequency in the passband was determined to be... The cutoff frequency in the passband is The filter design employs a finite impulse response structure, and the filter coefficients are determined using window function methods (such as the Hanning or Kaiser window) to ensure the filter has linear phase characteristics, avoid introducing nonlinear phase distortion into the time-series components, and ensure the time-series alignment accuracy of each component. In the ultrafiltration membrane fouling analysis scenario, time-series alignment accuracy is crucial for subsequent fouling source decoupling; any phase deviation may lead to incorrect attribution of fouling characteristics.

[0075] The running data will be passed sequentially. A group of bandpass filters performs filtering. Each group of bandpass filters performs a convolution operation on the input running data and outputs the time-series components of the corresponding frequency band. ,in Indicates a time index. (After) After processing by the group filter, the following is obtained One time-series component The entire set constitutes a multi-scale time series component. Each time series component is physically independent and corresponds to the operating characteristics at different frequency scales: the low-frequency component captures the long-term evolution trend of membrane fouling, the mid-frequency component reflects the impact of periodic changes in operating conditions on membrane performance, and the high-frequency component characterizes the rapid response of membrane flux caused by short-term shock events (such as sudden changes in influent water quality).

[0076] After the multi-scale time series component set is formed, it needs to be orthogonally checked to ensure that the information redundancy between each component meets the requirements. Orthogonality checking is performed by calculating the orthogonality of any two time series components. and ( Pearson correlation coefficient between ) To achieve this. The calculation formula is: ,in and Components and The time mean. When a correlation coefficient between a pair of time series components is detected. Exceeding the preset threshold When the value is set to an empirical value between 0.15 and 0.25 (usually an empirical value), it indicates that the two sets of bandpass filters overlap in frequency coverage, resulting in crosstalk between components.

[0077] For components that fail the orthogonality check, the bandwidth parameters of the corresponding bandpass filters are adjusted for those with excessive correlation coefficients. The adjustment strategy is as follows: if the component... and The correlation coefficient exceeded the standard, and Then The corresponding bandpass filter cutoff frequency shrinks towards lower frequencies, while simultaneously... The corresponding bandpass filter's cutoff frequency narrows towards higher frequencies, introducing a sufficient frequency gap between the two filters. The adjustment amplitude is reduced to the correlation coefficient. The following is the convergence objective, with the step size adjusted each time set to the current energy distribution width. The orthogonality target is gradually approximated by small iterative steps, ranging from 5% to 10%, to avoid over-adjustment that could lead to signal loss in the target frequency band. After adjustment, the running data is re-filtered, and the correlation coefficients between all component pairs are recalculated. This process is repeated until all components are properly filtered. All are below the preset threshold .

[0078] The final output is orthogonalized multi-scale time-series components, laying the foundation for subsequent construction of the interaction influence matrix and pollution feature representation vector. Orthogonalization ensures the statistical independence of each time-series component, providing a clear physical interpretation for the subsequent quantification of the interaction contribution of each component and avoiding confusion or weight bias in pollution source identification due to information overlap between components. In practical ultrafiltration membrane systems, operating data often contains multiple superimposed pollution mechanism signals. Through the aforementioned adaptive frequency domain decomposition and orthogonalization verification process, these composite signals can be effectively separated into independent components with clear frequency characteristics, providing high-quality input features for subsequent accurate identification of different types of pollution sources such as biological pollution, organic pollution, and inorganic scaling.

[0079] A correlation analysis is performed on the multi-scale time-series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time-series component is quantified. The interaction contribution is then fused with the time-series components to generate a pollution feature representation vector, including:

[0080] Time delay correlation analysis is performed on multi-scale time series components. The cross-correlation coefficients of any two time series components under different time delays are calculated, and the peak cross-correlation coefficient when the cross-correlation coefficient reaches its peak value is identified as the correlation strength of the time series component pair.

[0081] Construct an interaction influence matrix, where the rows and columns of the matrix correspond to multi-scale time series components. Fill the corresponding positions of the interaction influence matrix with the correlation strength. Calculate the sum of the elements in each row of the interaction influence matrix as the output influence strength of the corresponding time series component, and calculate the sum of the elements in each column of the interaction influence matrix as the input influence strength of the corresponding time series component.

[0082] The interaction contribution of the corresponding time series component is obtained by adding the output influence strength to the input influence strength;

[0083] The numerical sequence of each time series component is multiplied element-wise with the corresponding interaction contribution, and the multiplied time series components are concatenated to form a pollution feature representation vector.

[0084] After obtaining the multi-scale time-series components, it is necessary to further explore the dynamic correlations between these components to accurately describe the interactions between different fouling mechanisms. Because operating signals at different frequency scales often exhibit time lag propagation effects during ultrafiltration membrane fouling—for example, changes in influent turbidity will cause a membrane pressure difference response after a certain time delay—simply calculating the synchronization correlation coefficient cannot capture this causal delay characteristic. Therefore, time-delay correlation analysis is performed on each time-series component. By sliding the calculation of the cross-correlation coefficient between two time-series components within a certain time delay range, the cross-correlation function curve is obtained. The peak cross-correlation coefficient corresponding to the peak value of the cross-correlation coefficient is then identified, and this peak cross-correlation coefficient is defined as the correlation strength between the two time-series components.

[0085] Specifically, suppose that the multi-scale time series component set has a total of ... and In time delay Calculate the cross-correlation coefficient within the range ,in Within the preset maximum delay window Get the value from the inside. For all After value scanning, take The value corresponding to the maximum absolute value is taken as the component. With weight The strength of the correlation between ,Right now This processing method can simultaneously capture the extreme values ​​of both positive and negative correlations, ensuring that any positive or negative promoting relationship between pollutant components can be effectively quantified. The value of is usually determined based on the operating cycle of the ultrafiltration system and the data sampling frequency. It is generally set to 10% to 20% of the total data length to ensure the rationality of the delayed search.

[0086] After obtaining the association strength of all component pairs, an interaction influence matrix is ​​constructed. The rows and columns of the matrix correspond to Each time-series component, matrix element Fill in the components With weight The strength of the correlation between Diagonal elements Set to 0 to exclude the influence of the component itself on itself, and to avoid autocorrelation interfering with subsequent contribution calculations. Matrix for A non-negative symmetric matrix, where each non-zero element reflects the maximum correlation between two time-series components after time delay optimization.

[0087] Based on the interaction influence matrix The influence characteristics of each time-series component are extracted from both row and column dimensions. For the matrix... Summing all elements in the row yields the first... The total intensity of the influence exerted by one time-series component on other components is defined as the output influence intensity. ,Right now . The larger the value, the stronger the driving effect of that component on other components, and the more active its role in the pollution evolution process. For the matrix... Summing all elements of the column yields the other components relative to the first element. The total intensity of the influence exerted by each time-series component is defined as the input influence intensity. ,Right now . The larger the value, the stronger the combined driving force of the component from other components, and the more passive the response characteristics are during the pollution process.

[0088] The intensity of the output influence Intensity of input influence Add them together to get the first one. Interaction contribution of each time-series component ,Right now Interaction contribution This comprehensively reflects the overall participation of this component in the entire multi-scale time-series system. Regardless of whether it mainly manifests as an influencing source or a response end, as long as it has a strong correlation with other components, it will obtain a high interaction contribution. The rationale for this design is that in the complex dynamic system of ultrafiltration membrane fouling, highly involved components often carry richer information about the fouling mechanism and should be given greater weight in subsequent feature representations.

[0089] After obtaining the interaction contribution of each time-series component Then, these are used as weighting coefficients and fused with the numerical sequences of the corresponding time-series components. For the first... One time-series component Its weighted component sequence is This involves multiplying the value at each time point in the component sequence by the interaction contribution scalar, achieving element-wise multiplication. Through this operation, time-series components with higher contributions will obtain larger amplitude representations in the feature space, thus occupying a more important position in the subsequent pollution source separation and identification process; while components with lower contributions and weaker correlation with other components are relatively compressed, reducing their interference with the final feature representation.

[0090] After performing element-wise weighted summation, all Each weighted time series component The components are concatenated according to their index order to form a pollution feature representation vector. The concatenation method preserves the original sequence structure in the time dimension and arranges the weighted components sequentially in the feature dimension, so that... The dimension is The length is twice that of a single component. This pollution feature representation vector simultaneously contains multi-scale temporal dynamic information and inter-component interaction information, which can comprehensively describe the operating state of the ultrafiltration membrane under the combined effects of different pollution mechanisms.

[0091] In practical applications, for a certain ultrafiltration membrane water treatment system, the operating data, after adaptive decomposition, yields several time-series components, corresponding to pollution signals at different time scales, such as low-frequency trend changes, mid-frequency periodic fluctuations, and high-frequency noise disturbances. Time-delay correlation analysis reveals a lag correlation of approximately several hours between the low-frequency trend component and the mid-frequency periodic component, which aligns with the physical mechanism of organic pollutants gradually accumulating on the membrane surface and triggering periodic differential pressure responses. After constructing the interaction influence matrix, the output influence intensity of the low-frequency trend component is significantly higher than other components, indicating its dominant driving role in the entire pollution evolution process. Therefore, it receives a higher interaction contribution and is given greater weight in the pollution feature representation vector. The pollution feature representation vector formed by weighted concatenation not only retains the time-series characteristics of the original operating data but also incorporates structured information about the interaction correlations between components, providing high-quality feature input for subsequent pollution source separation and cleaning scheme generation.

[0092] Source separation processing is performed on the pollution feature representation vector to decouple the composite pollution signal into multiple independent pollution source components. The pollution source type and pollution contribution weight corresponding to each independent pollution source component are identified, including:

[0093] Standard evolution patterns are extracted from a pre-defined pollution source feature library as source separation constraints. The pollution feature representation vector is then subjected to source decomposition using these constraints to obtain a set of candidate pollution source components.

[0094] The reconstructed signal is obtained by linearly superimposing the candidate pollution source component sets. The difference between the reconstructed signal and the pollution feature representation vector is calculated to obtain the separation residual signal. When the energy of the separation residual signal exceeds the preset residual threshold, the residual pattern is extracted from the separation residual signal and matched with the preset pollution source feature library to obtain the supplementary pollution source component.

[0095] The supplementary pollution source components are incorporated into the candidate pollution source component set and the constraint source decomposition is re-executed. The iteration is performed until the energy of the separated residual signal converges, and the converged candidate pollution source component set is output as the independent pollution source component.

[0096] Temporal morphological features are extracted from each independent pollution source component. The temporal morphological features are then matched with the standard evolution patterns in the preset pollution source feature library to determine the pollution source type corresponding to each independent pollution source component. The pollution contribution weight of the corresponding pollution source type is obtained by numerically integrating each independent pollution source component in the time dimension.

[0097] In obtaining the pollution feature characterization vector Next, source separation processing is required to decouple the multiple superimposed pollutant components in the composite pollution signal into individual pollutant source components. During actual operation, ultrafiltration membranes often experience a combination of multiple pollution mechanisms simultaneously on the membrane surface and inside the pores, including organic adsorption, inorganic salt scaling, microbial contamination, and colloidal clogging. These mechanisms superimpose on the time-series signal, making direct differentiation difficult. The goal of source separation processing is to reconstruct the evolution trajectory of each independent pollutant source from the mixed pollution characteristic vector, providing precise information on the pollutant components for subsequent cleaning strategy development.

[0098] A pre-defined pollution source feature library forms the core prior knowledge foundation for source separation treatment. This library was established through statistical analysis of extensive historical cleaning experimental data and laboratory standard pollutant test data. For typical pollution types such as organic pollution, inorganic scaling, biological pollution, and colloidal pollution, standard evolution patterns are extracted from multi-dimensional operational signals including membrane flux decay curves, transmembrane pressure difference curves, and operating time. These standard evolution patterns are stored as time-series templates, describing the signal morphological characteristics of various pollution types at different development stages, including typical morphological parameters such as rise slope, periodic fluctuation characteristics, and decay rate. During source separation, a set of standard evolution patterns matching the current operating conditions is extracted from the feature library and used as constraints to guide the decomposition process, avoiding the generation of components with unclear physical meaning in unsupervised blind separation methods due to a lack of physical priors.

[0099] Constraint source decomposition method is used for During processing, the standard evolution model is used as the basis function constraint, and the linear combination of each candidate pollution source component is optimized to approximate the target as closely as possible. Simultaneously, it is required that each candidate component maintains morphological consistency with the corresponding standard model. Let the set of candidate pollution source components include... The component, denoted as the first component. The candidate components are The corresponding mixing coefficient is The constraint source decomposition problem is then transformed into solving the problem under the constraints of the standard evolutionary pattern. and The optimization problem involves introducing non-negativity constraints during the solution process to ensure the physical interpretability of each component, i.e., the contribution intensity of each pollution source is not negative. After the optimization solution is completed, the initial set of candidate pollution source components is obtained. .

[0100] The reconstructed signal is obtained by linearly superimposing all components in the candidate pollution source component set. ,calculate Compared with the original pollution feature representation vector The difference is used to obtain the separated residual signal. The energy of the separated residual signal is calculated and denoted as . and compared with the preset residual threshold Compare. If This indicates that the current set of candidate pollution source components has not yet fully covered the data. All contaminants contained therein, and valid contamination information still remains in the residual signal. At this point, for... Residual patterns are extracted and their temporal morphological characteristics are analyzed. A similarity search is then performed in a pre-defined pollution source feature database to find the standard evolutionary pattern that is closest in morphology to the residual pattern. The corresponding pollution type is then used as the newly discovered pollution component to generate supplementary pollution source components. .

[0101] Will Including candidate pollution source components expands the set size to For each component, the constrained source decomposition process is re-executed, and joint optimization is performed on the updated candidate component set. The morphological parameters and mixing coefficients of each component are recalculated, and the reconstructed signal is recalculated again. residual energy The above process constitutes an iterative loop, and each iteration determines... Does the convergence condition meet? The convergence criterion is that the change in residual energy between two consecutive iterations is lower than the preset convergence accuracy. ,Right now ,in Indicates the first The residual energy of the next iteration Indicates the first The residual energy of each iteration. When the convergence condition is met, the iteration stops, and the current set of candidate pollution source components is output as the final independent pollution source component. The introduction of the iterative mechanism can effectively address the problem of unknown pollution component quantity in complex pollution scenarios. By gradually supplementing the implicit pollution information in the residual, complete decoupling of complex complex pollution signals is achieved.

[0102] After obtaining the set of independent pollution source components, temporal morphological features are extracted from each component to determine its corresponding pollution source type. The extraction of temporal morphological features encompasses multiple dimensions: in terms of amplitude features, the peak amplitude, mean amplitude, and amplitude variation range of the component are extracted; in terms of trend features, the overall upward or downward slope of the component and the inflection point of the piecewise linear fit are calculated; in terms of periodicity features, spectral analysis is used to detect the presence of significant periodic components and their period lengths within the component; in terms of morphological similarity features, the dynamic time warp distance between the component and each standard evolutionary pattern is calculated to quantify the degree of morphological matching. The extracted temporal morphological feature vectors are then matched with the feature vectors of each standard evolutionary pattern in a pre-defined pollution source feature library, using weighted Euclidean distance as the similarity metric. The pollution type corresponding to the standard pattern with the smallest distance is the pollution source type of that independent pollution source component. Through this matching process, each independent pollution source component is assigned a clear physical label, such as organic adsorption pollution, calcium carbonate scaling, biofilm pollution, or colloidal blockage.

[0103] The pollution contribution weight is calculated by numerically integrating each independent pollution source component over time. Let the... Each independent pollution source component is Its observation time window Pollution contribution within Through points The calculated value reflects the cumulative contribution of the pollution source to membrane performance degradation throughout the entire observation period. The pollution contribution of all components is normalized to obtain the pollution contribution weight corresponding to each pollution source type. The calculation method is as follows ,in This represents the total number of independent pollution source components after convergence. The normalized pollution contribution weights. satisfy This directly reflects the proportion of each type of fouling in the current membrane fouling state, providing a quantitative basis for the design of subsequent cleaning agent synergistic formulations.

[0104] like Figure 2 As shown, Figure 2 This is a flowchart illustrating the synergistic effect of cleaning agents and the generation of a timing scheme in an embodiment of the present invention.

[0105] A synergistic effect network among cleaning agents is established based on the type of pollution source. Based on this network and pollution contribution weights, multi-cleaning agent combination formulations and phased cleaning sequence plans are generated, including:

[0106] The removal efficiency of cleaning agents was determined for each type of pollution source when acting alone, in combination, and in stages. The synergy coefficient was calculated based on the removal efficiency of combined and individual actions, and the influence coefficient was calculated based on the impact of the preceding cleaning agent on the removal efficiency of the subsequent cleaning agent.

[0107] The synergy coefficient is used as the strength of the synergy between cleaning agent nodes, and the influence coefficient is used as the strength of the temporal interaction between cleaning agent nodes to construct a synergy relationship network of cleaning agents.

[0108] The pollution contribution weights are mapped to the target removal amount of each pollution source type. Based on the target removal amount and removal efficiency, cleaning agent nodes are selected from the cleaning agent synergy relationship network, and the synergy strength and time-series effect strength between the selected cleaning agent nodes are extracted.

[0109] When the synergistic effect strength is greater than the preset synergistic threshold, the corresponding cleaning agents are combined into the same cleaning stage and the concentration ratio is calculated based on the target removal amount and removal efficiency. When the synergistic effect strength is less than the preset inhibition threshold, the corresponding cleaning agents are separated into different cleaning stages. The order of cleaning stages is determined according to the temporal effect strength, and a multi-cleaning agent combination formula and a phased cleaning sequence plan are generated.

[0110] For each type of pollution source, removal efficiency experiments were conducted to determine the removal efficiency of cleaning agents acting alone, in combination, and in stages. In the single-action experiments, each candidate cleaning agent was independently applied to the contaminated membrane sample under standard test conditions, and the pollutant removal rate was recorded as the baseline value for individual removal efficiency. In the combined-action experiments, two or more cleaning agents were added simultaneously, and the combined removal efficiency was measured within the same time window. In the staged-action experiments, cleaning agents were added sequentially according to different addition orders, and the overall removal efficiency under each combination of orders and the intermediate state removal efficiency after each stage of cleaning were recorded. Based on the above experimental data, the synergy coefficient was calculated. It is defined as a cleaning agent. With cleaning agent The ratio of the removal efficiency of the combined action to the sum of the removal efficiencies of the two actions alone, i.e. ,in To improve the removal efficiency of the combined action, and Cleaning agents and cleaning agent The individual removal efficiency. When This indicates a positive synergistic effect between the two. This indicates the presence of an antagonistic or inhibitory effect. Influence coefficient Description of pre-cleaning agent For subsequent cleaning agents The degree of efficiency gain or reduction is determined by comparing the effects of applying the first method. Apply again Compared to applying directly alone The difference in removal efficiency and The ratio is used to calculate, that is ,in In cleaning agent Pretreatment cleaning agent The actual removal efficiency. A positive influence coefficient indicates that the preceding cleaning agent has a promoting effect on the subsequent cleaning agent, while a negative value indicates an inhibitory effect.

[0111] Using each cleaning agent as a node, the synergy coefficient is... As the edge weight representing the strength of the collaborative effect between nodes, it will influence the coefficient. A synergistic effect network of cleaning agents is constructed using directed edges representing the temporal interaction strength between nodes. This network is a weighted directed graph structure, where the synergistic effect strength edges are undirected, reflecting the combined synergistic effect when two cleaning agents are applied simultaneously. The directed edges representing the temporal interaction strength point from preceding cleaning agent nodes to subsequent cleaning agent nodes, carrying directional information to describe the order dependency. Each node in the network is also associated with a classification removal efficiency attribute for different pollution source types, facilitating node selection and formulation calculation in specific pollution scenarios. For cleaning agent combinations with scarce experimental data, historical data from existing cleaning agents with similar chemical structures can be interpolated to complete the network, ensuring that the synergistic effect network covers all possible cleaning agent pairings.

[0112] Obtaining the normalized pollution contribution weights for each pollution source type Then, it is mapped to the target removal amount for each type of pollution source. The calculation of the target removal amount comprehensively considers the gap between the current membrane fouling baseline and the preset recovery target, as well as the proportion of each type of fouling source in the total fouling amount, i.e. ,in This represents the overall target removal rate, calculated from the difference between the current membrane flux and the target recovery value. Target removal rates are based on each pollution source type. The individual removal efficiency of cleaning agents for this type of pollution source The process involves selecting a set of cleaning agent nodes from the synergistic effect network of cleaning agents that can cover the main pollution source types. The selection criterion is that the individual removal efficiency of the candidate cleaning agent for the target pollution source type is not lower than a preset minimum efficiency threshold. Furthermore, the node has effective collaborative or temporally related edges with other selected nodes in the network. Extract the collaborative strength between all node pairs within the selected node set. and temporal action intensity This serves as the core basis for subsequent formula generation and timing scheme arrangement.

[0113] synergistic strength With preset collaboration threshold and preset suppression threshold Comparisons are made to determine the grouping and phase arrangement strategy for cleaning agents. When cleaning agents... With cleaning agent The strength of the synergistic effect between them satisfies When a significant positive synergistic effect is determined between the two, they are combined and added within the same cleaning stage. Within the same cleaning stage, the concentration ratio of each cleaning agent is calculated based on its removal efficiency and target removal amount for the target pollution source type. The concentration ratio is calculated with the target removal rate as a constraint and minimizing the total amount of cleaning agent added as the optimization objective. This is achieved by solving a simultaneous system of removal balance equations for each pollution source type to obtain the optimal concentration ratio of each cleaning agent at the current stage. When the cleaning agent... With cleaning agent The strength of the synergistic effect between them satisfies When significant antagonism or chemical compatibility issues are identified between the two substances, they are forcibly separated into different cleaning stages to avoid simultaneous addition leading to reduced removal efficiency or the generation of harmful byproducts. For synergistic effects with strengths between... and The choice between cleaning agent pairs should be flexibly made based on the actual pollution load and process constraints, deciding whether to combine the stages.

[0114] The order of the cleaning stages is based on the intensity of the temporal action. Confirmed. For all cleaning agent nodes assigned to different cleaning stages, establish the temporal priority relationship between stages: if the cleaning agent... Cleaning agent Temporal action intensity If the value is positive and the absolute value is large, then priority will be given to including cleaning agents. The cleaning stage is performed first to take advantage of the positive effect of the preceding cleaning on the efficiency of subsequent cleaning; if If the value is negative, avoid using the cleaning agent. Arranged in cleaning agent Previously, measures were taken to prevent the inhibition effect from reducing the subsequent cleaning effect. For multi-stage sorting problems, the temporal effect strength of each stage node is summarized to construct a stage priority matrix, and a topological sorting method is used to determine the final cleaning stage order, ensuring that the overall temporal scheme maximizes the synergistic gain between each stage.

[0115] The final multi-cleaning agent combination formulation includes the type and concentration of each cleaning agent in each cleaning stage. The solution includes the application method; the phased cleaning sequence plan includes the execution order of each cleaning phase and the preset duration of each phase. The rinsing intervals between stages are also considered. The formulation and timing scheme together constitute the initial control parameter set for the subsequent chemical cleaning process, which is called upon by the cleaning system when starting chemical cleaning and dynamically adjusted and optimized under real-time membrane flux response feedback. For typical scenarios where organic and inorganic fouling coexist, the alkaline cleaning agent stage is usually placed before the acidic cleaning agent stage because alkaline cleaning can effectively dissolve organic pollutants and loosen the scale structure, thereby improving the removal efficiency of inorganic scale by subsequent acid washing. This sequence arrangement is highly consistent with the positive value characteristics of the corresponding directional edges in the timing action intensity matrix, verifying the effectiveness of the method in practical engineering scenarios.

[0116] Calculate the deviation between the membrane flux response data and the preset recovery target trajectory. Based on the deviation, dynamically adjust the dosing rate and cleaning sequence parameters of the multi-cleaning agent combination formulation, and continuously perform chemical cleaning until the membrane flux reaches the preset recovery target, including:

[0117] Trend fitting is performed on the membrane flux response data to extract the instantaneous evolution slope sequence, and trend fitting is performed on the preset recovery target trajectory to extract the target evolution slope sequence;

[0118] The numerical difference between the membrane flux response data and the preset recovery target trajectory is calculated as the numerical deviation, and the difference between the instantaneous evolution slope sequence and the target evolution slope sequence is calculated as the trend deviation. The numerical deviation and the trend deviation are weighted and fused to obtain the deviation amount.

[0119] The cleaning response time constant is extracted from each independent pollution source component. The ratio of trend deviation to numerical deviation is calculated as the deviation dominance coefficient. Based on the deviation dominance coefficient, the weight of the deviation in the acceleration rate adjustment dimension and the cleaning time sequence parameter adjustment dimension is determined. The acceleration rate adjustment and cleaning time sequence parameter adjustment for each cleaning agent in the multi-cleaning agent combination formula are calculated by combining the cleaning response time constant and the weight. The acceleration rate and cleaning time sequence parameters of the multi-cleaning agent combination formula are adjusted according to the acceleration rate adjustment and the cleaning time sequence parameter adjustment.

[0120] The instantaneous evolution slope sequence is extrapolated, extended, and integrated to reconstruct the predicted membrane flux trajectory. When the fluctuation amplitude of the predicted membrane flux trajectory meets the stability condition, the membrane flux is determined to have reached the preset recovery target and chemical cleaning is stopped.

[0121] During the chemical cleaning phase, real-time monitoring and closed-loop control of the membrane flux recovery process are necessary. The real-time collected membrane flux response data constitutes a numerical sequence evolving over time. Comparing this sequence with a pre-defined recovery target trajectory is the core basis for dynamic adjustment. Simply comparing numerical values ​​cannot fully reflect the dynamic characteristics of the recovery process; therefore, trend information is also introduced. By performing polynomial local trend fitting on the membrane flux response data, the instantaneous evolution slope is extracted step-by-step, forming an instantaneous evolution slope sequence. The same trend fitting operation is performed on the preset recovery target trajectory to extract the target evolution slope sequence. The instantaneous evolution slope describes the rate of change of membrane flux at the current moment, while the target evolution slope describes the expected rate of change of the recovery trajectory at the corresponding moment; together, they form the basis for trend-level comparison.

[0122] The numerical deviation is obtained by calculating the difference between the measured value of the membrane flux response data at the current moment and the target value of the preset recovery target trajectory at that moment. This value reflects the absolute distance between the current membrane flux and the target value; the difference between the instantaneous evolution slope sequence and the target evolution slope sequence at the current moment is calculated to obtain the trend deviation. This value reflects the deviation between the membrane flux recovery rate and the desired rate. The numerical deviation and trend deviation are weighted and fused to obtain the comprehensive deviation. The fusion relationship can be represented as ,in The fusion weights are for numerical bias. The fusion weights for trend deviations satisfy the following conditions: The fusion weights are dynamically set based on the cleaning stage: in the initial stage of cleaning, changes in the membrane flux recovery rate better reflect the immediate effect of the cleaning agent, with trend deviation weights... Relatively high; in the later stages of cleaning, when the membrane flux approaches the target value, the numerical deviation weighting... The value is relatively high to ensure the final convergence accuracy.

[0123] In obtaining the comprehensive deviation Next, it is necessary to further determine how this deviation should be allocated across the two dimensions of acceleration rate adjustment and cleaning timing parameter adjustment. The cleaning response time constant should be extracted from each independent pollution source component. This parameter describes the characteristic response time of flux recovery for a specific pollutant under the action of the current cleaning agent, and the corresponding values ​​for different pollution sources. Since differences exist, the equivalent time constant after weighted averaging of all pollution sources is used as a global reference. The ratio of trend deviation to numerical deviation is calculated and defined as the deviation dominance coefficient. ,Right now (when (When not zero). When the absolute value is large, it indicates that the deviation of the recovery rate from the target is much greater than the numerical deviation. In this case, the dominant factor of the deviation lies in the kinetic characteristics of the cleaning process, and more adjustment weight should be allocated to the adjustment dimension of the cleaning sequence parameters, such as extending or shortening the duration of each cleaning stage; when When the absolute value is small, it indicates that the numerical deviation is the main problem. More adjustment weight should be allocated to the acceleration rate adjustment dimension. By increasing or decreasing the acceleration rate of the cleaning agent, the gap between the membrane flux and the target value can be quickly narrowed.

[0124] Let the weighting of the acceleration rate adjustment dimension be... The weighting of the cleaning time sequence parameter adjustment dimension is as follows: ,satisfy Both are about The monotonic function, with the specific mapping relationship calibrated based on historical cleaning data. For the first [aspect] in a multi-cleaning agent combination formulation... The dosage of this cleaning agent is adjusted according to its dosing rate. From the comprehensive deviation Assigning weights Cleaning response time constant The cleaning agent's removal efficiency against the current dominant pollution source was jointly determined; the adjustment amount of the cleaning sequence parameters was also considered. From the comprehensive deviation Assigning weights and The adjustments to the injection acceleration rate and cleaning timing parameters are jointly determined. Both adjustments are limited within preset safety ranges to prevent membrane material damage or cleaning agent waste due to excessive adjustments. After adjustment, the updated injection acceleration rate and cleaning timing parameters are sent to the cleaning control execution unit in real time, driving the cleaning system to operate according to the new parameters. Membrane flux response data is then collected again in the next sampling cycle, forming a complete closed-loop control circuit.

[0125] To determine whether chemical cleaning can be terminated, a prediction-based stopping criterion is used, rather than simply waiting for the instantaneous membrane flux value to exceed a target threshold. The current instantaneous evolution slope sequence is extrapolated and extended, using local linear extrapolation or low-order polynomial extrapolation to extend the slope sequence several future time steps, resulting in a predicted slope sequence. This predicted slope sequence is then integrated and reconstructed using the current measured membrane flux value as the initial condition, and the cumulative integral yields the predicted membrane flux trajectory. The predicted membrane flux trajectory describes the expected evolution path of membrane flux over a future period, assuming that the current cleaning parameters remain constant. The fluctuation amplitude of the predicted membrane flux trajectory within the prediction time window is calculated and defined as the difference between the maximum and minimum values ​​of the predicted trajectory within that window. .when Less than the preset stability threshold If the average value of the predicted membrane flux trajectory is not lower than the preset recovery target value, it is determined that the membrane flux has reached the preset recovery target and tends to stabilize, and an instruction to stop chemical cleaning is issued; if the predicted trajectory still has large fluctuations or the average value does not meet the target, cleaning is continued and re-evaluated in the next cycle.

[0126] This termination criterion based on predicted trajectory stability has significant advantages over the instantaneous threshold criterion: the instantaneous threshold criterion is prone to misjudgment due to short-term fluctuations in membrane flux, leading to premature termination of cleaning or unnecessary extension of cleaning time; while the predicted trajectory stability criterion, by examining whether the future evolution trend converges, can effectively identify the true stable state of membrane flux, avoiding erroneous termination of cleaning during transient fluctuations, thereby ensuring cleaning effectiveness while reducing the consumption of cleaning agents and the waste of cleaning time. The entire closed-loop control process organically integrates numerical deviation, trend deviation, contamination response time constant, deviation dominance coefficient, and predicted stability criterion, achieving refined control over the entire chemical cleaning process.

[0127] Extrapolating and integrating the instantaneous evolution slope sequence to reconstruct the predicted membrane flux trajectory includes:

[0128] The re-pollution rate after cleaning is extracted from each independent pollution source component, and the instantaneous evolution slope sequence is decoupled according to the pollution source type to obtain the slope component set.

[0129] The recontamination rate after cleaning is converted into an extension correction coefficient. Based on the extension correction coefficient, each slope component in the slope component set is extrapolated and extended. The extrapolated slope components are then superimposed to obtain the extension slope sequence.

[0130] Extract the switching time of each cleaning stage in the phased cleaning sequence scheme, introduce the synergistic coefficient of the cleaning agent interaction at the switching time, and enhance and correct the slope value of the extension slope sequence at the switching time according to the synergistic coefficient to obtain the synergistic slope sequence.

[0131] The flux increment sequence is obtained by integrating the efficiency slope sequence. The flux value at the starting point of the membrane flux response data is added to the flux increment sequence to obtain the predicted membrane flux trajectory.

[0132] Before obtaining the extended slope sequence, it is necessary to extract the recontamination rate information after cleaning from each independent contaminant source component. Each independent contaminant source component is not static during the chemical cleaning process; it is gradually removed with the continuous action of the cleaning agent. However, at the same time, some contaminants may redeposit on the membrane surface during cleaning intervals or when the cleaning agent concentration decreases, resulting in a recontamination effect. For each independent contaminant source component, a linear regression is performed on its time series in the tail end of the cleaning phase to extract the time-varying slope of that contaminant source component. The positive value portion of this slope is identified as the recontamination rate, denoted as . ,in This is an index for the pollution source components. The re-pollution rate of each pollution source component is then obtained. Subsequently, the instantaneous evolution slope sequence was decoupled according to the pollution source type, decomposing the overall instantaneous evolution slope sequence into slope components corresponding to each pollution source type, thus forming a set of slope components. The component decoupling was based on the contribution ratio of each pollution source type to the membrane flux evolution, utilizing known pollution contribution weights. The total slope sequence is proportionally distributed to the slope components corresponding to each pollution source type, ensuring that the sum of each slope component is consistent with the overall instantaneous evolution slope sequence.

[0133] The rate of recontamination In the process of converting to the extension correction factor, it is necessary to consider the inhibitory effect of the recontamination rate on the future membrane flux evolution trend. Define the... The extension correction factor corresponding to each pollution source component is: Its calculation method is to use the re-pollution rate After normalization, a weighted fusion is performed with the benchmark correction factor, specifically expressed as follows: ,in To reference the baseline value for recontamination rate, This is the recontamination inhibition intensity coefficient, used to control the degree of influence of recontamination on extension correction. When When the value is close to 1, it indicates that the repollution effect of this pollution source component is weak, and the slope extension is almost unaffected by the correction; when When the value is significantly less than 1, it indicates that the pollution source component has a clear re-pollution trend, and the extrapolated value of the corresponding slope component needs to be corrected downward to avoid overly optimistic estimates of the degree of membrane flux recovery.

[0134] Based on the continuation correction factor For each slope component in the slope component set, an extrapolation extension is performed. The starting point of the extrapolation extension is the latest moment of the current cleaning process, i.e., the extension start point, and the extension range covers all moments within the prediction time window. For the th... For each slope component, at each time point after the extension start point, the slope value of this component at the extension start point is multiplied by the corresponding extension correction factor. The exponential decay model is then applied, incorporating the time decay characteristics of this component, to ensure that the slope component of the extrapolated extension gradually stabilizes over time, reflecting the physical law that the membrane flux recovery rate gradually slows down after pollutant removal. After completing the independent extrapolation of each slope component, the extrapolated slope components corresponding to all pollution source types are summed at each time step to obtain the extrapolated slope sequence covering the prediction time window. During the summation process, it is necessary to ensure the alignment of each component on the time axis to avoid temporal misalignment caused by inconsistent component indices.

[0135] Extracting the switching moments of each cleaning stage in a phased cleaning sequence is a key step in introducing synergistic correction. During the phased execution of a multi-cleaner combination formulation, there are switching moments between adjacent cleaning stages. At these moments, a brief synergistic effect occurs between the residual cleaner from the preceding stage and the subsequent cleaner, resulting in a step-like accelerated recovery of membrane flux near the switching moment. Each switching moment is denoted as... ,in This is the index for the switching time. At each switching time... Based on the synergy coefficient between the preceding and subsequent cleaning agents... Calculate the efficiency enhancement factor corresponding to this switching moment. Efficiency coefficient The calculation basis is: if the synergy coefficient Exceeding the preset collaboration threshold If so, then a significant synergistic effect is considered to exist at that switching moment, and... Set the enhancement factor to be greater than 1; if the synergy coefficient is at the suppression threshold With collaboration threshold If the synergy coefficient is between 1 and 1, the synergy coefficient is set to 1, and no slope correction is applied; if the synergy coefficient is below the inhibition threshold... If the synergy coefficient is taken as a suppression factor less than 1, the slope value near the switching moment is corrected downward to reflect the negative impact of the antagonistic effect between cleaning agents on flux recovery.

[0136] Based on the efficiency enhancement coefficient corresponding to each switching moment The slope value of the extended slope sequence at the switching time is enhanced and corrected. The correction method is to perform the correction at the switching time. Within the neighborhood of , the slope value at the corresponding time in the extended slope sequence is multiplied by an enhancement factor. The neighborhood width is based on the cleaning response time constant. Confirmation, meaning the correction effect applies before and after the switching time. The flux linearly decays to an uncorrected state within the time range. After synergistic corrections at all switching points, an synergistic slope sequence is obtained. Compared with the extended slope sequence, the synergistic slope sequence more accurately reflects the actual evolution trend of membrane flux under the synergistic effect of multiple cleaning agents, especially the accelerated recovery behavior of flux near the cleaning agent switching node.

[0137] When integrating the enhancement slope sequence, a time-by-time cumulative numerical integration method is used. The enhancement slope sequence is integrated step-by-step along the time axis to obtain the flux increment sequence. The value at each time step in the flux increment sequence represents the cumulative increment of membrane flux relative to the extension start point from that point. After calculating the flux increment sequence, the measured flux values ​​of the membrane flux response data at the extension start point are used. As an integration benchmark, it is added to the cumulative increment of the flux increment sequence at each time step to obtain the predicted membrane flux trajectory. This predicted trajectory covers the entire timeframe from the start of the extension to the end of the prediction time window. It includes the recontamination suppression effect of each pollution source component and incorporates the synergistic correction at the switching times of multiple cleaning agents. It can accurately predict the future recovery trajectory of membrane flux under the current cleaning parameter configuration, providing a basis for subsequent assessment of membrane flux stability and calculation of predicted fluctuations. Provides the data foundation.

[0138] A second aspect of this invention provides an AI-based intelligent optimization system for chemical cleaning of ultrafiltration membranes, comprising:

[0139] Adaptive sub-units are used to acquire the operating data of the ultrafiltration membrane system. The adaptive decomposition algorithm automatically determines the decomposition parameters based on the signal characteristics of the operating data and decomposes the operating data into multi-scale time-series components.

[0140] The pollution feature unit is used to perform correlation analysis on multi-scale time series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time series component is quantified, and the interaction contribution is fused with the time series components to generate a pollution feature characterization vector.

[0141] The source separation unit is used to perform source separation processing on the pollution feature characterization vector, decouple the composite pollution signal into multiple independent pollution source components, and identify the pollution source type and pollution contribution weight corresponding to each independent pollution source component.

[0142] The cleaning scheme unit is used to establish a synergistic relationship network between cleaning agents based on the type of pollution source, and to generate multi-cleaning agent combination formulas and phased cleaning sequence schemes based on the synergistic relationship network and pollution contribution weights.

[0143] The dynamic control unit is used to control the ultrafiltration membrane cleaning system to start chemical cleaning according to the phased cleaning sequence plan, collect membrane flux response data in real time during the cleaning process, calculate the deviation between the membrane flux response data and the preset recovery target trajectory, and dynamically adjust the dosing rate and cleaning sequence parameters of the multi-cleaning agent combination formula according to the deviation, and continuously perform chemical cleaning until the membrane flux reaches the preset recovery target.

[0144] A third aspect of the present invention provides an electronic device, comprising:

[0145] processor;

[0146] Memory used to store processor-executable instructions;

[0147] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0148] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0149] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based intelligent optimization method for chemical cleaning of ultrafiltration membranes, characterized in that, include: The system acquires the operating data of the ultrafiltration membrane system, and automatically determines the decomposition parameters based on the signal characteristics of the operating data using an adaptive decomposition algorithm, and decomposes the operating data into multi-scale time-series components. A correlation analysis was performed on the multi-scale time series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time series component was quantified. The interaction contribution was then fused with the time series components to generate a pollution feature characterization vector. Source separation processing is performed on the pollution feature characterization vector to decouple the composite pollution signal into multiple independent pollution source components, and the pollution source type and pollution contribution weight corresponding to each independent pollution source component are identified. Establish a synergistic relationship network between cleaning agents based on the type of pollution source, and generate multi-cleaning agent combination formulas and phased cleaning sequence schemes based on the synergistic relationship network and pollution contribution weights; The ultrafiltration membrane cleaning system is controlled to start chemical cleaning according to the phased cleaning sequence plan. Membrane flux response data is collected in real time during the cleaning process. The deviation between the membrane flux response data and the preset recovery target trajectory is calculated. The dosing rate of the multi-cleaning agent combination formula and the cleaning sequence parameters are dynamically adjusted according to the deviation. Chemical cleaning is continuously performed until the membrane flux reaches the preset recovery target.

2. The method according to claim 1, characterized in that, An adaptive decomposition algorithm automatically determines the decomposition parameters based on the signal characteristics of the running data and decomposes the running data into multi-scale time-series components, including: The frequency domain transformation of the running data is performed to obtain the spectral distribution characteristics, and the energy peak position and energy distribution width are extracted from the spectral distribution characteristics; The set of center frequencies of the time series components is determined based on the energy peak location. The bandwidth parameters corresponding to each center frequency are calculated based on the energy distribution width. The number of decomposition levels is determined as the decomposition parameters based on the set of center frequencies and the bandwidth parameters. Multiple bandpass filters are constructed based on the decomposed parameters. The passband range of each bandpass filter is determined by the center frequency and bandwidth parameters. The running data is filtered sequentially through multiple sets of bandpass filters. Each set of bandpass filters outputs a time component, and all time components constitute a multi-scale time component. Orthogonality verification is performed on the multi-scale time series components. When the correlation coefficient between time series components exceeds the preset threshold, the bandwidth parameter of the corresponding bandpass filter is adjusted and the filtering process is re-executed until the correlation coefficient between all time series components is lower than the preset threshold, and the orthogonalized multi-scale time series components are output.

3. The method according to claim 1, characterized in that, A correlation analysis is performed on the multi-scale time-series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time-series component is quantified. The interaction contribution is then fused with the time-series components to generate a pollution feature representation vector, including: Time delay correlation analysis is performed on multi-scale time series components. The cross-correlation coefficients of any two time series components under different time delays are calculated, and the peak cross-correlation coefficient when the cross-correlation coefficient reaches its peak value is identified as the correlation strength of the time series component pair. Construct an interaction influence matrix, where the rows and columns of the matrix correspond to multi-scale time series components. Fill the corresponding positions of the interaction influence matrix with the correlation strength. Calculate the sum of the elements in each row of the interaction influence matrix as the output influence strength of the corresponding time series component, and calculate the sum of the elements in each column of the interaction influence matrix as the input influence strength of the corresponding time series component. The interaction contribution of the corresponding time series component is obtained by adding the output influence strength to the input influence strength; The numerical sequence of each time series component is multiplied element-wise with the corresponding interaction contribution, and the multiplied time series components are concatenated to form a pollution feature representation vector.

4. The method according to claim 1, characterized in that, Source separation processing is performed on the pollution feature representation vector to decouple the composite pollution signal into multiple independent pollution source components. The pollution source type and pollution contribution weight corresponding to each independent pollution source component are identified, including: Standard evolution patterns are extracted from a pre-defined pollution source feature library as source separation constraints. The pollution feature representation vector is then subjected to source decomposition using these constraints to obtain a set of candidate pollution source components. The reconstructed signal is obtained by linearly superimposing the candidate pollution source component sets. The difference between the reconstructed signal and the pollution feature representation vector is calculated to obtain the separation residual signal. When the energy of the separation residual signal exceeds the preset residual threshold, the residual pattern is extracted from the separation residual signal and matched with the preset pollution source feature library to obtain the supplementary pollution source component. The supplementary pollution source components are incorporated into the candidate pollution source component set and the constraint source decomposition is re-executed. The iteration is performed until the energy of the separated residual signal converges, and the converged candidate pollution source component set is output as the independent pollution source component. Temporal morphological features are extracted from each independent pollution source component. The temporal morphological features are then matched with the standard evolution patterns in the preset pollution source feature library to determine the pollution source type corresponding to each independent pollution source component. The pollution contribution weight of the corresponding pollution source type is obtained by numerically integrating each independent pollution source component in the time dimension.

5. The method according to claim 1, characterized in that, A synergistic effect network among cleaning agents is established based on the type of pollution source. Based on this network and pollution contribution weights, multi-cleaning agent combination formulations and phased cleaning sequence plans are generated, including: The removal efficiency of cleaning agents was determined for each type of pollution source when acting alone, in combination, and in stages. The synergy coefficient was calculated based on the removal efficiency of combined and individual actions, and the influence coefficient was calculated based on the impact of the preceding cleaning agent on the removal efficiency of the subsequent cleaning agent. The synergy coefficient is used as the strength of the synergy between cleaning agent nodes, and the influence coefficient is used as the strength of the temporal interaction between cleaning agent nodes to construct a synergy relationship network of cleaning agents. The pollution contribution weights are mapped to the target removal amount of each pollution source type. Based on the target removal amount and removal efficiency, cleaning agent nodes are selected from the cleaning agent synergy relationship network, and the synergy strength and time-series effect strength between the selected cleaning agent nodes are extracted. When the synergistic effect strength is greater than the preset synergistic threshold, the corresponding cleaning agents are combined into the same cleaning stage and the concentration ratio is calculated based on the target removal amount and removal efficiency. When the synergistic effect strength is less than the preset inhibition threshold, the corresponding cleaning agents are separated into different cleaning stages. The order of cleaning stages is determined according to the temporal effect strength, and a multi-cleaning agent combination formula and a phased cleaning sequence plan are generated.

6. The method according to claim 1, characterized in that, Calculate the deviation between the membrane flux response data and the preset recovery target trajectory. Based on the deviation, dynamically adjust the dosing rate and cleaning sequence parameters of the multi-cleaning agent combination formulation, and continuously perform chemical cleaning until the membrane flux reaches the preset recovery target, including: Trend fitting is performed on the membrane flux response data to extract the instantaneous evolution slope sequence, and trend fitting is performed on the preset recovery target trajectory to extract the target evolution slope sequence; The numerical difference between the membrane flux response data and the preset recovery target trajectory is calculated as the numerical deviation, and the difference between the instantaneous evolution slope sequence and the target evolution slope sequence is calculated as the trend deviation. The numerical deviation and the trend deviation are weighted and fused to obtain the deviation amount. The cleaning response time constant is extracted from each independent pollution source component. The ratio of trend deviation to numerical deviation is calculated as the deviation dominance coefficient. Based on the deviation dominance coefficient, the weight of the deviation in the acceleration rate adjustment dimension and the cleaning time sequence parameter adjustment dimension is determined. The acceleration rate adjustment and cleaning time sequence parameter adjustment for each cleaning agent in the multi-cleaning agent combination formula are calculated by combining the cleaning response time constant and the weight. The acceleration rate and cleaning time sequence parameters of the multi-cleaning agent combination formula are adjusted according to the acceleration rate adjustment and the cleaning time sequence parameter adjustment. The instantaneous evolution slope sequence is extrapolated, extended, and integrated to reconstruct the predicted membrane flux trajectory. When the fluctuation amplitude of the predicted membrane flux trajectory meets the stability condition, the membrane flux is determined to have reached the preset recovery target and chemical cleaning is stopped.

7. The method according to claim 6, characterized in that, Extrapolating, extending, and integrating the instantaneous evolution slope sequence to reconstruct the predicted membrane flux trajectory includes: The re-pollution rate after cleaning is extracted from each independent pollution source component, and the instantaneous evolution slope sequence is decoupled according to the pollution source type to obtain the slope component set. The recontamination rate after cleaning is converted into an extension correction coefficient. Based on the extension correction coefficient, each slope component in the slope component set is extrapolated and extended. The extrapolated slope components are then superimposed to obtain the extension slope sequence. Extract the switching time of each cleaning stage in the phased cleaning sequence scheme, introduce the synergistic coefficient of the cleaning agent interaction at the switching time, and enhance and correct the slope value of the extension slope sequence at the switching time according to the synergistic coefficient to obtain the synergistic slope sequence. The flux increment sequence is obtained by integrating the efficiency slope sequence. The flux value at the starting point of the membrane flux response data is added to the flux increment sequence to obtain the predicted membrane flux trajectory.

8. An AI-based intelligent optimization system for chemical cleaning of ultrafiltration membranes, used to implement the method as described in any one of claims 1-7, characterized in that, include: Adaptive sub-units are used to acquire the operating data of the ultrafiltration membrane system. The adaptive decomposition algorithm automatically determines the decomposition parameters based on the signal characteristics of the operating data and decomposes the operating data into multi-scale time-series components. The pollution feature unit is used to perform correlation analysis on multi-scale time series components to construct an interaction influence matrix. Based on the interaction influence matrix, the interaction contribution of each time series component is quantified, and the interaction contribution is fused with the time series components to generate a pollution feature characterization vector. The source separation unit is used to perform source separation processing on the pollution feature characterization vector, decouple the composite pollution signal into multiple independent pollution source components, and identify the pollution source type and pollution contribution weight corresponding to each independent pollution source component. The cleaning scheme unit is used to establish a synergistic relationship network between cleaning agents based on the type of pollution source, and to generate multi-cleaning agent combination formulas and phased cleaning sequence schemes based on the synergistic relationship network and pollution contribution weights. The dynamic control unit is used to control the ultrafiltration membrane cleaning system to start chemical cleaning according to the phased cleaning sequence plan, collect membrane flux response data in real time during the cleaning process, calculate the deviation between the membrane flux response data and the preset recovery target trajectory, and dynamically adjust the dosing rate and cleaning sequence parameters of the multi-cleaning agent combination formula according to the deviation, and continuously perform chemical cleaning until the membrane flux reaches the preset recovery target.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images