An artificial intelligence-based membrane component data management system and method
By using an AI-based membrane component data management system, accurate real-time prediction of membrane component status and lifespan is achieved, solving the problem of inaccurate prediction in existing technologies and ensuring the stable operation of the demineralized water process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-27
- Publication Date
- 2026-03-27
AI Technical Summary
Existing technologies cannot accurately predict the condition and lifespan of membrane components in real time, leading to substandard water quality or equipment failure in demineralized water processes. Furthermore, existing models cannot adapt to dynamic changes in equipment condition.
An AI-based membrane component data management system is adopted. By acquiring historical operating data, preprocessing and spatiotemporal alignment are performed to construct multidimensional feature vectors and health models. Combined with lifespan analysis data, a lifespan prediction model is constructed and updated in real time using Kalman filtering.
It improves the accuracy of membrane component condition assessment and life prediction, reduces prediction errors, dynamically updates parameters, and ensures the normal operation of the demineralized water process.
Smart Images
Figure CN120408289B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of film component data management, and particularly relates to a film component data management system and method based on artificial intelligence. BACKGROUND
[0002] In the thermal power industry, the desalination process is an important link to ensure the normal operation of equipment and prolong the service life. The desalination process mainly removes salt and other impurities in water through membrane separation technologies such as reverse osmosis (RO) membranes and nanofiltration (NF) membranes. As the core component of the desalination process, the performance of the membrane component directly affects the efficiency and quality of water treatment. The performance of the membrane component will decrease with the increase of the use time, which may lead to substandard water quality or equipment failure.
[0003] Therefore, timely diagnosis of the state of the membrane component and maintenance and replacement are the key to ensure the normal operation of the desalination process. The degradation of the membrane component is affected by multiple factors, and the degradation process is usually nonlinear, which cannot be accurately described by the existing model. The existing model cannot update the parameters in real time, and it is difficult to adapt to the dynamic changes of the equipment state, resulting in low prediction accuracy.
[0004] Therefore, the application discloses a film component data management system and method based on artificial intelligence to solve the above problems. SUMMARY
[0005] The application aims to provide a film component data management system and method based on artificial intelligence to solve the problems in the prior art.
[0006] To achieve the above-mentioned purpose, the application provides the following technical scheme: a film component data management method based on artificial intelligence, which comprises the following steps:
[0007] S1: obtaining historical operation data of the membrane component, preprocessing the historical operation data of the membrane component, realizing space-time alignment of operation parameters of different collection periods, and performing spatial interpolation;
[0008] S2: analyzing data features based on the preprocessed time series of each operation parameter of the membrane component, assembling a multi-dimensional feature vector, analyzing reconstruction errors of the multi-dimensional feature vector, and constructing a health model based on the reconstruction errors;
[0009] S3: constructing a life analysis data pair based on the historical operation data of the membrane component from the starting use state to the failure state, and constructing a corresponding life prediction model based on the life analysis data pair;
[0010] S4: obtaining life prediction model parameters of the related historical membrane component of the to-be-predicted membrane component, constructing a real-time life prediction model of the to-be-predicted membrane component, and performing early warning analysis according to the real-time residual life and health index of the to-be-predicted membrane component.
[0011] According to the above scheme, in S1, the following is included:
[0012] S101: Obtain the historical operation data of the membrane component, generate the time series of each operation parameter, including online monitoring data and offline detection data; the online monitoring data includes transmembrane pressure difference, product water quality, flow rate and inlet water temperature; the product water quality is equal to the product water conductivity divided by the product water turbidity; the offline detection data includes membrane surface topography image and pollutant composition concentration; the acquisition period of each operation parameter data is different and is preset by the system; the offline detection data is laboratory detection data;
[0013] S102: Perform image binarization on the membrane surface topography image, perform full coverage of the foreground area based on the window set preset by the system, and count the number of each size window that completes the full coverage of the foreground area; form a data pair based on the size of each window and the corresponding window number, and perform linear fitting based on the data pair; the slope of the fitted straight line is recorded as the image feature value;
[0014] S103: Extract the time series of any two operation parameters, denoted as X={x i |i∈[1,I]} and Y={y j |j∈[1,J]}, wherein x i represents the i-th operation parameter value in the time series X, y j represents the j-th operation parameter value in the time series Y, I represents the number of values in the time series X, and J represents the number of values in the time series Y; the operation parameter acquisition frequency of the time series Y is lower than that of the time series X; an accumulated distance matrix D of size I×J is constructed, and the elements in the accumulated distance matrix D are initialized to 0; for each matrix element D(i, j), the corresponding distance d(i, j) is calculated and the accumulated distance matrix is filled, and if the accumulated distance is greater than a preset threshold, the calculation is terminated; wherein D(i, j) represents the minimum distance from the beginning of the sequence to x i and y j , and d(i, j) is the Euclidean distance between x i and y j ; starting from D(N, M), the shortest path is recorded by backtracking to (1, 1) according to the recursive path; according to the obtained shortest path, the operation parameter values in the time series Y are interpolated in the time scale of the time series X to align them, and the Kriging method is used for spatial interpolation; traverse all kinds of operation parameters to form an extended time series of each operation parameter.
[0015] The application realizes time alignment by preprocessing the time sequence of the operation parameters, solves the problem of inconsistent time scales of the sensor data and the offline detection data, does not need to unify the collection cycle of each sensor, reduces the overall detection resources, and improves the overall analysis efficiency when filling the accumulated distance matrix, and terminates the calculation when the accumulated distance is greater than the preset threshold.
[0016] According to the above scheme, in S2, the following content is included:
[0017] S201: Obtain an extended time sequence of the transmembrane pressure difference, analyze the coefficient of variation CV of the transmembrane pressure difference, the coefficient of variation being equal to the standard deviation of the extended time sequence of the transmembrane pressure difference divided by the average value of the extended time sequence of the transmembrane pressure difference; analyze the theoretical value of the membrane flux according to the extended time sequence of the transmembrane pressure difference and Darcy's law, generate residual characteristics ε by analyzing the theoretical value and the actual value of the membrane flux, and analyze the frequency spectrum energy E and the main frequency component f of the extended time sequence of the membrane flux by using Fourier transform dominant .
[0018] S202: Obtain extended time sequences of water production quality, flow rate and inlet water temperature, analyze the mean and standard deviation of each extended time sequence; based on the coefficient of variation CV of the transmembrane pressure difference, the mean and standard deviation of each extended time sequence, the residual characteristics ε, the frequency spectrum energy E and the main frequency component f dominant , construct a multi-dimensional feature vector A, A = [CV, μ T , σ T , μ Q , σ Q , μ C , σ C , ε, E, f dominant ]; wherein μ T represents the mean of the extended time sequence of the inlet water temperature, σ T represents the standard deviation of the extended time sequence of the inlet water temperature; μ Q represents the mean of the extended time sequence of the flow rate, σ Q represents the standard deviation of the extended time sequence of the flow rate; μ C represents the mean of the extended time sequence of the water production quality, σ C represents the standard deviation of the extended time sequence of the water production quality;
[0019] The performance of the membrane component is the result of the joint action of multiple operation parameters; by only performing Fourier transform on the comprehensive performance index, the periodic characteristics of the joint action of multiple operation parameters can be captured, the analysis effect is guaranteed while the calculation complexity is reduced, redundancy and errors caused by separate analysis of the operation parameters are avoided, and the frequency domain features (such as frequency spectrum energy and main frequency component) extracted by one Fourier transform have clear physical meaning, which can effectively support the health state evaluation and fault diagnosis of the membrane component.
[0020] S203: data compression is performed on the multi-dimensional feature vector A to form a latent space representation, and the latent space representation is reconstructed into original data; the encoder expression corresponding to the data compression of the multi-dimensional feature vector A to form the latent space representation is h = Ψ(W e A + b e ); wherein h represents the latent space representation, W e represents the encoder weight matrix, b e represents the encoder bias vector, and Ψ() represents the activation function ReLU; the decoder expression corresponding to the reconstruction of the latent space representation into the original data is A' = Ψ(W d h + b d ); wherein A' represents the reconstructed feature vector, W d represents the decoder weight matrix, and b d represents the decoder bias vector; the weight matrices W e and W d are initialized by randomly sampling from a normal distribution with a mean of 0 and a variance of 2 / N, N being the number of neurons in the previous layer; the bias vectors b e and b d are initialized as zero vectors.
[0021] The parameters of the encoder and the decoder are optimized by back propagation to minimize the reconstruction error; wherein the loss function is mean square error; based on the reconstruction error, a health model HI = 1 - L AE / L AE max ; wherein L AE max represents the maximum reconstruction error of the training set, and L AE represents the reconstruction error of the current sample.
[0022] The application compresses high-dimensional input data into low-dimensional latent representation through the encoder, extracts key features in the data, learns the essential features of the data, and removes noise in the sensor data; the denoised data can be used for analysis, enhancing the accuracy of the remaining useful life prediction;
[0023] According to the above scheme, in S3, the following contents are included:
[0024] S301: obtain historical operation data of the film component from the start of use state to the failure state, record the remaining useful life corresponding to the failure state as 0, and construct life analysis data pair (t, HI t , RUL t ) based on the historical operation data, wherein t represents time, HI t represents the health index corresponding to time t, and RUL trepresents the remaining useful life corresponding to time t; constructing a life prediction model according to the life analysis data pair: d RUL t / dt=-k×exp(β×HI t ); wherein k represents a degradation rate coefficient, β represents a health index influence coefficient, and exp() represents an exponential function with a natural number as a base;
[0025] S302: discretizing the life prediction model into a difference equation: RUL t+1 =RUL t -△t×k×exp(β×HI t ), wherein △t represents a time step, and the initialized values of k and β are preset by the system; the life analysis data pair corresponding to the historical operation data is substituted into the discretized difference equation, the prediction error corresponding to each time node is calculated, the prediction error is equal to the predicted remaining useful life minus the actual service life, the total error is analyzed based on the prediction error, the total error is equal to the sum of squares of the prediction errors of all time points; the degradation rate coefficient and the health index influence coefficient are updated using the gradient descent method until the total error reaches a minimum value or reaches a maximum number of iterations; and the latest degradation rate coefficient and health index influence coefficient are bound to the membrane component.
[0026] The application groups life analysis data pairs based on historical operation data and constructs a life prediction model, combines the advantages of physical models and data-driven models, and reduces prediction errors; in the case of insufficient data, the physical model provides reliable support and reduces the dependence on data.
[0027] According to the above scheme, in S4, the following contents are included:
[0028] S401: extracting the degradation rate coefficient and the health index influence coefficient corresponding to a preset number of historical membrane components, calculating the average value of the degradation rate coefficient and the average value of the health index influence coefficient of the historical membrane components as the initialized values of the degradation rate coefficient and the health index influence coefficient of the membrane component to be predicted; the production coordinates of the extracted preset number of historical membrane components are the same as those of the membrane component to be predicted;
[0029] S402: constructing a real-time life prediction model d RUL / dt=-k×exp(β×HI) based on the initialized values of the degradation rate coefficient and the health index influence coefficient of the membrane component to be predicted; updating the degradation rate coefficient and the health index influence coefficient of the membrane component to be predicted in real time by using Kalman filtering; and predicting the real-time remaining useful life based on the real-time operation data of the membrane component to be predicted;
[0030] S403: If the real-time residual life of the membrane component to be predicted is less than the corresponding threshold value and the health index of the membrane component to be predicted is less than the corresponding threshold value, a first-level early warning signal is issued; if the real-time residual life of the membrane component to be predicted is less than the corresponding threshold value and the health index of the membrane component to be predicted is greater than or equal to the corresponding threshold value, a second-level early warning signal is issued; if the real-time residual life of the membrane component to be predicted is greater than or equal to the corresponding threshold value and the health index of the membrane component to be predicted is less than the corresponding threshold value, a third-level early warning signal is issued; if the real-time residual life of the membrane component to be predicted is greater than or equal to the corresponding threshold value and the health index of the membrane component to be predicted is greater than or equal to the corresponding threshold value, no early warning signal is issued.
[0031] When the system identifies a first-level early warning signal, an instruction is sent to notify the administrator to replace the membrane component; when the system identifies a second-level early warning signal, an instruction is sent to notify the administrator to perform routine maintenance on the membrane component; when the system identifies a third-level early warning signal, an instruction is sent to notify the administrator to perform routine maintenance or repair on the membrane component.
[0032] The application can improve the accuracy of prediction by replacing the model coefficient of the membrane component to be predicted with the model coefficient of the historical membrane component, dynamically update the parameters, and further improve the accuracy of real-time residual life prediction.
[0033] Another aspect of the application provides a membrane component data management system based on artificial intelligence, which is applied to the above-mentioned membrane component data management method based on artificial intelligence, and comprises a data preprocessing module, a health analysis module, a life prediction model construction module and a real-time early warning module.
[0034] The data preprocessing module is used to obtain the historical operation data of the membrane component, pre-process the historical operation data of the membrane component, realize the space-time alignment of the operation parameters of different collection periods, and perform spatial interpolation.
[0035] The health analysis module analyzes the data characteristics of the pre-processed time series of each operation parameter of the membrane component, constructs a multi-dimensional feature vector, analyzes the reconstruction error according to the multi-dimensional feature vector, and constructs a health model based on the reconstruction error.
[0036] The life prediction model construction module constructs a life analysis data pair according to the historical operation data of the membrane component from the starting state to the failure state, and constructs a corresponding life prediction model based on the life analysis data pair.
[0037] The real-time early warning module obtains the life prediction model parameters of the relevant historical membrane component of the membrane component to be predicted, constructs a real-time life prediction model of the membrane component to be predicted, and performs early warning analysis according to the real-time residual life and the health index of the membrane component to be predicted.
[0038] According to the scheme, the data preprocessing module comprises a data acquisition unit and a space-time alignment expansion unit.
[0039] The data acquisition unit is configured to acquire historical operation data of the membrane component, generate a time sequence of each operation parameter, perform image binarization on the membrane surface morphology image, perform full coverage of the foreground area based on a system preset window set, and count the number of full coverage of the foreground area for each size of window; form a data pair based on the size of each window and the corresponding number of windows, perform linear fitting based on the data pair, and record the slope of the fitted straight line as an image feature value;
[0040] The space-time alignment expansion unit is configured to extract time sequences of any two operation parameters, analyze the shortest path of each operation parameter value of the two time sequences, perform interpolation based on the shortest path, and perform spatial interpolation using a Kriging method; and traverse all types of operation parameters to form an expanded time sequence of each operation parameter.
[0041] According to the scheme, the health analysis module comprises a feature analysis unit and a health model construction unit.
[0042] The feature analysis unit is configured to acquire an expanded time sequence of the transmembrane pressure difference, analyze the coefficient of variation of the transmembrane pressure difference, analyze the theoretical value of the membrane flux based on the expanded time sequence of the transmembrane pressure difference and Darcy's law, generate residual features by analyzing the theoretical value and actual value of the membrane flux, perform Fourier transform on the expanded time sequence of the membrane flux, analyze the frequency spectrum energy and main frequency component, and acquire the expanded time sequences of the water quality, flow rate, and inlet water temperature, and analyze the mean and standard deviation of each expanded time sequence.
[0043] The health model construction unit constructs a multi-dimensional feature vector based on the coefficient of variation of the transmembrane pressure difference, the mean and standard deviation of each expanded time sequence, the residual features, the frequency spectrum energy, and the main frequency component; performs data compression on the multi-dimensional feature vector to form a latent space representation, and reconstructs the latent space representation into original data; optimizes the parameters of the encoder and the decoder through back propagation to minimize the reconstruction error; and constructs a health model based on the reconstruction error.
[0044] According to the scheme, the life prediction model construction module comprises a life prediction model construction unit and a parameter optimization unit.
[0045] The life prediction model construction unit acquires historical operation data of the entire process of the membrane component from the start of use to the failure state, constructs a life analysis data pair based on the historical operation data, and constructs a life prediction model according to the life analysis data pair.
[0046] The parameter optimization unit is configured to discretize the life prediction model into a difference equation, update the degradation rate coefficient and the health index influence coefficient using a gradient descent method until a total error reaches a minimum value or a maximum iteration number is reached, and bind the latest degradation rate coefficient and the health index influence coefficient to the membrane component.
[0047] According to the above scheme, the real-time warning module comprises a real-time model construction unit and a warning unit.
[0048] The real-time model construction unit is configured to extract degradation rate coefficients and health index influence coefficients corresponding to a preset number of historical membrane components, calculate average values of the degradation rate coefficients and the health index influence coefficients of the historical membrane components as initial values of the degradation rate coefficients and the health index influence coefficients of a membrane component to be predicted, construct a real-time life prediction model based on the initial values of the degradation rate coefficients and the health index influence coefficients of the membrane component to be predicted, and perform real-time residual life prediction.
[0049] The warning unit grades warning signals according to the real-time residual life and the health index, and prompts the system according to the warning signal grade.
[0050] Compared with the prior art, the application has the beneficial effects that: the application realizes time alignment by preprocessing the time sequence of the operating parameters; solves the problem of inconsistent time scales of the sensor data and the offline detection data, does not need to unify the collection cycles of each sensor, reduces the overall detection resources; when filling the cumulative distance matrix, if the cumulative distance is greater than a preset threshold, the calculation is terminated; improves the overall analysis efficiency; by performing Fourier transform on only the comprehensive performance index, the periodic characteristics of the combined action of multiple operating parameters can be captured, the analysis effect is guaranteed while the computational complexity is reduced; redundancy and errors caused by separate analysis of the operating parameters are avoided; the frequency domain features extracted by one Fourier transform have clear physical meaning, which can effectively support the health state evaluation of the membrane component; the application compresses high-dimensional input data into low-dimensional potential representation through an encoder, extracts key features in the data; learns the essential features of the data, removes noise in the sensor data; the de-noised data can be used for analysis, enhancing the accuracy of the residual life prediction; the application establishes a life analysis data pair based on historical operation data and constructs a life prediction model, combining the advantages of the physical model and the data-driven model, reducing the prediction error; in the case of insufficient data, the physical model provides reliable support, reducing the dependence on data; the application can improve the prediction accuracy by using the model coefficients of the historical membrane components to replace the model coefficients of the membrane component to be predicted, dynamically updating the parameters, and further improving the accuracy of real-time residual life prediction. BRIEF DESCRIPTION OF DRAWINGS
[0051] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application, and together with the description serve to explain the application, and do not limit the application. In the drawings:
[0052] Figure 1 A flowchart of a film component data management method based on artificial intelligence according to the present application;
[0053] Figure 2 A structural diagram of a film component data management system based on artificial intelligence according to the present application. DETAILED DESCRIPTION
[0054] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0055] Please refer to Figure 1 The present application provides a technical solution: a film component data management method based on artificial intelligence, which comprises the following steps:
[0056] S1: obtaining historical operation data of the film component, preprocessing the historical operation data of the film component, realizing space-time alignment of operation parameters of different collection periods, and performing spatial interpolation;
[0057] In S1, the following contents are included:
[0058] S101: obtaining historical operation data of the film component, generating time series of each operation parameter, and the operation parameters including online monitoring data and offline detection data; the online monitoring data including transmembrane pressure difference, water production quality, flow rate and inlet water temperature; the water production quality being equal to the water production conductivity divided by the water production turbidity; the offline detection data including membrane surface topography images and pollutant composition concentration; the acquisition periods of each operation parameter data being different and being preset by the system;
[0059] S102: performing image binarization on the membrane surface topography images, performing full coverage of the foreground area based on a window set preset by the system, and counting the number of each size window completing the full coverage of the foreground area; forming a data pair based on the size of each window and the corresponding window number, and performing linear fitting based on the data pair; and recording the slope of the fitted straight line as an image feature value;
[0060] S103: extracting time series of any two operation parameters, denoted as X={x i |i∈[1,I]} and Y={y jwhere x i represents the i-th operating parameter value in time series X, y j represents the j-th operating parameter value in time series Y, I represents the number of values in time series X, and J represents the number of values in time series Y; the operating parameter acquisition frequency of time series Y is lower than that of time series X; an accumulated distance matrix D of size I x J is constructed, and the elements in the accumulated distance matrix D are initialized to 0; for each matrix element D(i, j), the corresponding distance d(i, j) is calculated and the accumulated distance matrix is filled, and if the accumulated distance is greater than a preset threshold, the calculation is terminated; where D(i, j) represents the minimum distance from the start of the sequence to x i and y j , and d(i, j) is the Euclidean distance between x i and y j ; starting from D(N, M), the shortest path is recorded by backtracking according to the recursive path to (1, 1); according to the obtained shortest path, the operating parameter values in time series Y are interpolated on the time scale of time series X so that they are aligned, and the Kriging method is used for spatial interpolation; all kinds of operating parameters are traversed to form an extended time series of each operating parameter.
[0061] S2: Analyzing the data characteristics of the preprocessed time series of each operating parameter of the membrane component, constructing a multi-dimensional feature vector, analyzing the reconstruction error of the multi-dimensional feature vector, and constructing a health model based on the reconstruction error;
[0062] In S2, the following contents are included:
[0063] S201: Obtaining the extended time series of the transmembrane pressure difference, analyzing the coefficient of variation CV of the transmembrane pressure difference, which is equal to the standard deviation of the extended time series of the transmembrane pressure difference divided by the average value of the extended time series of the transmembrane pressure difference; analyzing the theoretical value of the membrane flux according to the extended time series of the transmembrane pressure difference and Darcy's law, generating residual features ε by analyzing the theoretical value and actual value of the membrane flux; obtaining the extended time series of the membrane flux and analyzing the frequency spectrum energy E and the main frequency component f dominant ;
[0064] S202: Obtaining the extended time series of the water production quality, flow rate, and inlet water temperature, analyzing the mean and standard deviation of each extended time series; constructing a multi-dimensional feature vector A based on the coefficient of variation CV of the transmembrane pressure difference, the mean and standard deviation of each extended time series, the residual features ε, the frequency spectrum energy E, and the main frequency component f dominant ; T ; T ; Q ; Q ; C ;C , ε, E, f dominant ] ; wherein μ T represents the mean of the inflow temperature augmented time series, σ T represents the standard deviation of the inflow temperature augmented time series; μ Q represents the mean of the flow augmented time series, σ Q represents the standard deviation of the flow augmented time series; μ C represents the mean of the produced water quality augmented time series, σ C represents the standard deviation of the produced water quality augmented time series;
[0065] S203: data compression is performed on the multi-dimensional feature vector A to form a latent space representation, and the latent space representation is reconstructed into original data; the encoder expression corresponding to the data compression of the multi-dimensional feature vector A to form a latent space representation is h = Ψ(W e A + b e ); wherein h represents the latent space representation, W e represents the encoder weight matrix, b e represents the encoder bias vector, and Ψ() represents the activation function ReLU; the decoder expression corresponding to the reconstruction of the latent space representation into original data is A' = Ψ(W d h + b d ); wherein A' represents the reconstructed feature vector; W d represents the decoder weight matrix; b d represents the decoder bias vector; the weight matrices W e and W d are initialized by randomly sampling from a normal distribution with a mean of 0 and a variance of 2 / N, N being the number of neurons in the previous layer; the bias vectors b e and b d are initialized as zero vectors;
[0066] The parameters of the encoder and the decoder are optimized through back propagation to minimize the reconstruction error; wherein the loss function is mean square error; based on the reconstruction error, a health model HI = 1 - L AE / L AE max ; wherein L AE max represents the maximum reconstruction error of the training set, and L AE represents the reconstruction error of the current sample.
[0067] S3: constructing a life analysis data pair based on the historical operation data of the membrane component from the start-of-use state to the failure state, and constructing a corresponding life prediction model based on the life analysis data pair;
[0068] In S3, the following contents are included:
[0069] S301: Obtain the historical operation data of the membrane component from the beginning of use to the failure state, record the remaining life corresponding to the failure state as 0, and construct a life analysis data pair (t, H t , RUL t ) based on the historical operation data, wherein t represents time, H t represents the health index corresponding to time t, and RUL t represents the remaining life corresponding to time t; construct a life prediction model based on the life analysis data pair: d RUL t / dt = -k x exp(β x H t ); wherein k represents a degradation rate coefficient, β represents a health index influence coefficient, and exp() represents an exponential function with a natural number as the base;
[0070] S302: Discretize the life prediction model into a difference equation: RUL t+1 = RUL t -△t x k x exp(β x H t ), wherein △t represents a time step, and the initialized values of k and β are system presets; substitute the life analysis data pair corresponding to the historical operation data into the discretized difference equation, calculate the prediction error corresponding to each time node, and the prediction error is equal to the predicted remaining life minus the actual service life; analyze the total error based on the prediction error, and the total error is equal to the sum of the squares of the prediction errors at all time points; update the degradation rate coefficient and the health index influence coefficient using the gradient descent method until the total error reaches a minimum value or reaches a maximum iteration number; and bind the latest degradation rate coefficient and health index influence coefficient to the corresponding membrane component.
[0071] Embodiment 1: In this embodiment, the life analysis data pair exists (0, 1.0, 1000), (100, 0.9, 900), (200, 0.8, 800), (300, 0.7, 700) and (400, 0.6, 600); wherein the units of time t and remaining life RUL t are hours;
[0072] In this embodiment, the time step △t = 100h, the initialized parameters k = 0.01, and β = 0.1;
[0073] Therefore, the discretization formula is: RUL t+1 = RUL t -100 x 0.01 x exp(0.1 x H t );
[0074] For each time point t, calculate the prediction error e t = RUL true，T+1 - [RUL texp(0.1 x HI t )];
[0075] Total error Using gradient descent method to optimize iteratively, finally k = 0.1, β = 0.5;
[0076] S4: Obtain the life prediction model parameters of the related historical membrane components of the to-be-predicted membrane component, construct a real-time life prediction model of the to-be-predicted membrane component, and perform early warning analysis according to the real-time residual life and health index of the to-be-predicted membrane component.
[0077] In S4, the following contents are included:
[0078] S401: Extract the degradation rate coefficient and health index influence coefficient corresponding to a preset number of historical membrane components, calculate the average value of the degradation rate coefficient and the average value of the health index influence coefficient of the historical membrane components as the initialization value of the degradation rate coefficient and the health index influence coefficient of the to-be-predicted membrane component; extract the production coordinates corresponding to a preset number of historical membrane components which are the same as the to-be-predicted membrane component;
[0079] S402: Construct a real-time life prediction model dRUL / dt = -k x exp(β x HI) based on the initialization value of the degradation rate coefficient and the health index influence coefficient of the to-be-predicted membrane component; use Kalman filter to update the degradation rate coefficient and the health index influence coefficient of the to-be-predicted membrane component in real time; predict the real-time residual life based on the real-time running data of the to-be-predicted membrane component;
[0080] S403: If the real-time residual life of the to-be-predicted membrane component is less than the corresponding threshold value and the health index of the to-be-predicted membrane component is less than the corresponding threshold value, a first-level early warning signal is issued; if the real-time residual life of the to-be-predicted membrane component is less than the corresponding threshold value and the health index of the to-be-predicted membrane component is greater than or equal to the corresponding threshold value, a second-level early warning signal is issued; if the real-time residual life of the to-be-predicted membrane component is greater than or equal to the corresponding threshold value and the health index of the to-be-predicted membrane component is less than the corresponding threshold value, a third-level early warning signal is issued; if the real-time residual life of the to-be-predicted membrane component is greater than or equal to the corresponding threshold value and the health index of the to-be-predicted membrane component is greater than or equal to the corresponding threshold value, no early warning signal is issued;
[0081] When the system identifies a first-level early warning signal, an instruction is sent to notify the administrator to replace the membrane component; when the system identifies a second-level early warning signal, an instruction is sent to notify the administrator to perform daily maintenance on the membrane component; when the system identifies a third-level early warning signal, an instruction is sent to notify the administrator to perform daily maintenance or repair on the membrane component.
[0082] Please refer to Figure 2The application provides a technical scheme: a membrane component data management system based on artificial intelligence, which comprises a data preprocessing module, a health analysis module, a life prediction model construction module and a real-time early warning module.
[0083] The data preprocessing module is used for obtaining historical operation data of the membrane component, preprocessing the historical operation data of the membrane component, realizing space-time alignment of operation parameters of different collection periods, and performing space interpolation.
[0084] The health analysis module is based on the time series analysis data characteristics of the pretreated operation parameters of the membrane component, assembles a multi-dimensional feature vector, analyzes the reconstruction error according to the multi-dimensional feature vector, and constructs a health model based on the reconstruction error.
[0085] The life prediction model construction module constructs a life analysis data pair according to the historical operation data of the whole process from the starting use state to the failure state of the membrane component, and constructs a corresponding life prediction model based on the life analysis data pair.
[0086] The real-time early warning module obtains the life prediction model parameters of the related historical membrane component of the to-be-predicted membrane component, constructs a real-time life prediction model of the to-be-predicted membrane component, and performs early warning analysis according to the real-time residual life and health index of the to-be-predicted membrane component.
[0087] The data preprocessing module comprises a data collection unit and a space-time alignment expansion unit.
[0088] The data collection unit is used for obtaining the historical operation data of the membrane component, generating time series of various operation parameters, performing image binarization on the membrane surface morphology image, performing full coverage of the foreground area based on a system preset window set, counting the number of full coverage of the foreground area of each size window, forming a data pair based on the size of each window and the corresponding window number, performing linear fitting based on the data pair, and recording the slope of the fitted straight line as an image feature value.
[0089] The space-time alignment expansion unit is used for extracting time series of any two operation parameters, analyzing the shortest path of the values of the two time series of various operation parameters, performing interpolation based on the shortest path, and performing space interpolation by using a Kriging method; and all kinds of operation parameters are traversed to form an expanded time series of each operation parameter.
[0090] The health analysis module comprises a feature analysis unit and a health model construction unit.
[0091] The characteristic analysis unit is configured to obtain an extended time series of the transmembrane pressure difference, analyze a coefficient of variation of the transmembrane pressure difference, analyze a theoretical value of the membrane flux according to the extended time series of the transmembrane pressure difference and Darcy's law, generate residual characteristics by analyzing the theoretical value and the actual value of the membrane flux, obtain an extended time series of the membrane flux by using Fourier transform, analyze frequency spectrum energy and a main frequency component, obtain extended time series of the product water quality, flow and feed water temperature, and analyze the mean and standard deviation of each extended time series.
[0092] The health model construction unit is configured to construct a multi-dimensional feature vector based on the coefficient of variation of the transmembrane pressure difference, the mean and standard deviation of each extended time series, the residual characteristics, the frequency spectrum energy and the main frequency component, compress the multi-dimensional feature vector to form a latent space representation, reconstruct the latent space representation into original data, optimize parameters of an encoder and a decoder through back propagation to minimize reconstruction error, and construct a health model based on the reconstruction error.
[0093] The life prediction model construction module includes a life prediction model construction unit and a parameter optimization unit.
[0094] The life prediction model construction unit is configured to obtain historical operation data of a membrane component from a starting use state to a failure state, construct life analysis data pairs based on the historical operation data, and construct a life prediction model according to the life analysis data pairs.
[0095] The parameter optimization unit is configured to discretize the life prediction model into a difference equation, update a degradation rate coefficient and a health index influence coefficient using a gradient descent method until a total error reaches a minimum value or a maximum iteration number is reached, and bind the latest degradation rate coefficient and the health index influence coefficient to the membrane component.
[0096] The real-time early warning module includes a real-time model construction unit and a warning unit.
[0097] The real-time model construction unit is configured to extract a degradation rate coefficient and a health index influence coefficient corresponding to a preset number of historical membrane components, calculate a degradation rate coefficient average value and a health index influence coefficient average value of the historical membrane components as initialization values of the degradation rate coefficient and the health index influence coefficient of a to-be-predicted membrane component, extract production coordinates corresponding to the preset number of historical membrane components that are the same as those of the to-be-predicted membrane component, construct a real-time life prediction model based on the initialization values of the degradation rate coefficient and the health index influence coefficient of the to-be-predicted membrane component, and perform real-time residual life prediction.
[0098] The warning unit is configured to grade a warning signal according to the real-time residual life and the health index, and provide a system prompt according to the warning signal grade.
[0099] It is to be understood that the phrases such as first and second, and the like, refer only to different instances of an entity or to different operations, and do not necessarily imply that the entities or operations are in any way causally related or that one of the entities or operations precedes another of the entities or operations. Also, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.
[0100] It will be apparent to those skilled in the art that the application is not limited to the details of the foregoing exemplary embodiments and that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. No feature of the application is considered critical unless expressly stated in the claims.
Claims
1. An artificial intelligence-based film component data management method, characterized by, The method includes the following steps: S1: Acquire historical operating data of the membrane component, preprocess the historical operating data of the membrane component, achieve spatiotemporal alignment of operating parameters for different acquisition cycles, and perform spatial interpolation; S2: Based on the preprocessed time series analysis data characteristics of each operating parameter of the membrane component, construct a multidimensional feature vector, analyze the reconstruction error of the multidimensional feature vector, and build a health model based on the reconstruction error; S3: Construct a lifetime analysis data pair based on the historical operating data of the membrane component from the start of use to the failure state, and construct a corresponding lifetime prediction model based on the lifetime analysis data pair; S4: Obtain the life prediction model parameters of the membrane component to be predicted based on the historical membrane component, construct the real-time life prediction model of the membrane component to be predicted, and perform early warning analysis based on the real-time remaining life and health indicators of the membrane component to be predicted. S1 contains the following: S101: Acquire historical operating data of the membrane components and generate time series of various operating parameters. The operating parameters include online monitoring data and offline detection data. The online monitoring data includes transmembrane pressure difference, permeate quality, flow rate, and feed water temperature. The permeate quality is equal to the permeate conductivity divided by the permeate turbidity. The offline detection data includes membrane surface morphology images and pollutant component concentrations. The acquisition period for each operating parameter data is different and preset by the system. S102: Binarize the membrane surface morphology image, perform full coverage of the foreground region based on the system's preset window set, and count the number of windows of each size that complete full coverage of the foreground region. Data pairs are formed based on the size of each window and the corresponding number of windows. Linear fitting is performed based on the data pairs, and the slope of the fitted line is recorded as the image feature value. S103: Extract the time series of any two operating parameters, denoted as X={x i |i∈[1,I]} and Y={y j |j∈[1,J]}, where x i y represents the value of the i-th running parameter in the time series X. j Let I represent the j-th running parameter value in time series Y, I represent the number of values in time series X, and J represent the number of values in time series Y. The sampling frequency of the running parameters in time series Y is lower than that of the running parameters in time series X. Construct a cumulative distance matrix D of size I×J, and initialize all elements in the cumulative distance matrix D to 0. For each matrix element D(i,j), calculate the corresponding distance d(i,j) and fill the cumulative distance matrix. If the cumulative distance is greater than a preset threshold, terminate the calculation. where D(i,j) represents the minimum distance from the start of the sequence to x i and y j , d(i,j) is the Euclidean distance between x i and y j ; starting from D(N,M), the shortest path is recorded by backtracking from (1,1) according to the recursive path; according to the obtained shortest path, the values of the operating parameters in the time series Y are interpolated on the time scale of the time series X so that they are aligned, and the Kriging method is used for spatial interpolation; all kinds of operating parameters are traversed to form an extended time series of each operating parameter.
2. The method of claim 1, wherein: S2 contains the following: S201: Obtain the extended time series of the trans-membrane pressure difference, analyze the coefficient of variation CV of the trans-membrane pressure difference, which is equal to the standard deviation of the extended time series of the trans-membrane pressure difference divided by the average value of the extended time series of the trans-membrane pressure difference; analyze the theoretical value of the membrane flux according to the extended time series of the trans-membrane pressure difference and Darcy's law, analyze the residual characteristics ε generated by the theoretical value and the actual value of the membrane flux; obtain the extended time series of the membrane flux using Fourier transform, analyze the frequency spectrum energy E and the main frequency component f dominant ; S202: Obtain the extended time series of water production quality, flow rate and influent temperature, analyze the mean and standard deviation of each extended time series; based on the coefficient of variation CV of transmembrane pressure difference, the mean and standard deviation of each extended time series, residual characteristic ε, spectrum energy E and main frequency component f dominant Construct a multi-dimensional feature vector A, A=[CV, µ T , σ T , µ Q , σ Q , µ C , σ C , ε, E, f dominant ]; wherein µ T represents the mean of the extended time series of influent temperature, σ T represents the standard deviation of the extended time series of influent temperature; µ Q represents the mean of the extended time series of flow rate, σ Q represents the standard deviation of the extended time series of flow rate; µ C represents the mean of the extended time series of water production quality, σ C represents the standard deviation of the extended time series of water production quality; S203: data compression is performed on the multi-dimensional feature vector A to form a latent space representation, and the latent space representation is reconstructed into original data; the encoder expression corresponding to the data compression of the multi-dimensional feature vector A to form the latent space representation is h = Ψ (W e A + b e ); wherein h represents the latent space representation, W e represents the encoder weight matrix, b e represents the encoder bias vector, and Ψ () represents the activation function ReLU; the decoder expression corresponding to the reconstruction of the latent space representation into the original data is A' = Ψ (W d h + b d ); wherein A' represents the reconstructed feature vector, W d represents the decoder weight matrix, b d represents the decoder bias vector; the weight matrices W e and W d are randomly sampled from a normal distribution with a mean of 0 and a variance of 2 / N, N being the number of neurons of the previous layer; the bias vectors b e and b d are zero vectors; The parameters of the encoder and decoder are optimized through backpropagation to minimize the reconstruction error; the loss function is the mean squared error; a healthy model is constructed based on the reconstruction error.
3. The method of claim 2, wherein: S3 includes the following: S301: Obtain historical operation data of the film component from the beginning of use to the failure state, record the remaining life corresponding to the failure state as 0, and construct a life analysis data pair (t, H t , RUL t ) based on the historical operation data, wherein t represents time, H t represents a health index corresponding to time t, and RUL t represents the remaining life corresponding to time t; construct a life prediction model according to the life analysis data pair; S302: Discretize the life prediction model into a difference equation, substitute the life analysis data corresponding to the historical operating data into the discretized difference equation, calculate the prediction error corresponding to each time node, the prediction error is equal to the predicted remaining life minus the actual life, and analyze the total error based on the prediction error, the total error is equal to the sum of the squares of the prediction errors at all time points. The degradation rate coefficient and health indicator influence coefficient are updated using the gradient descent method until the total error reaches the minimum or the maximum number of iterations is reached. The latest degradation rate coefficient and health indicator impact coefficient are bound to the corresponding membrane components.
4. The method of claim 3, wherein: S4 includes the following: S401: Extract the degradation rate coefficient and health index influence coefficient corresponding to a preset number of historical membrane components, calculate the average value of the degradation rate coefficient and the health index influence coefficient of the historical membrane components as the initialization value of the degradation rate coefficient and the health index influence coefficient of the to-be-predicted membrane component; the production coordinates corresponding to the to-be-predicted membrane component are the same as those of the to-be-predicted membrane component; S402: Construct a real-time life prediction model based on the initialization value of the degradation rate coefficient and the health index influence coefficient of the to-be-predicted membrane component; update the degradation rate coefficient and the health index influence coefficient of the to-be-predicted membrane component in real time by using Kalman filtering; predict the real-time residual life based on the real-time operation data of the to-be-predicted membrane component; S403: If the real-time residual life of the to-be-predicted membrane component is less than the corresponding threshold value and the health index of the to-be-predicted membrane component is less than the corresponding threshold value, a first-level warning signal is issued; if the real-time residual life of the to-be-predicted membrane component is less than the corresponding threshold value and the health index of the to-be-predicted membrane component is greater than or equal to the corresponding threshold value, a second-level warning signal is issued; if the real-time residual life of the to-be-predicted membrane component is greater than or equal to the corresponding threshold value and the health index of the to-be-predicted membrane component is less than the corresponding threshold value, a third-level warning signal is issued; if the real-time residual life of the to-be-predicted membrane component is greater than or equal to the corresponding threshold value and the health index of the to-be-predicted membrane component is greater than or equal to the corresponding threshold value, no warning signal is issued; When the system identifies a first-level warning signal, an instruction is sent to notify the administrator to replace the membrane component; when the system identifies a second-level warning signal, an instruction is sent to notify the administrator to perform daily maintenance on the membrane component; when the system identifies a third-level warning signal, an instruction is sent to notify the administrator to perform daily maintenance or repair on the membrane component.
5. An artificial intelligence-based membrane component data management system, which is applied to the implementation of an artificial intelligence-based membrane component data management method according to any one of claims 1-4, characterized in that, The system comprises a data preprocessing module, a health analysis module, a life prediction model construction module, and a real-time warning module; The data preprocessing module is used to obtain historical operation data of the membrane component, preprocess the historical operation data of the membrane component, realize time-space alignment of operation parameters of different collection periods, and perform spatial interpolation; The health analysis module analyzes data characteristics based on time series of preprocessed operation parameters of the membrane component, assembles a multi-dimensional feature vector, analyzes reconstruction error according to the multi-dimensional feature vector, and constructs a health model based on the reconstruction error; The life prediction model construction module constructs a life analysis data pair based on historical operation data of the membrane component from a starting use state to a failure state, and constructs a corresponding life prediction model based on the life analysis data pair; The real-time warning module obtains life prediction model parameters of a to-be-predicted membrane component related to historical membrane components, constructs a real-time life prediction model of the to-be-predicted membrane component, and performs warning analysis according to the real-time residual life and the health index of the to-be-predicted membrane component.
6. The artificial intelligence-based membrane component data management system of claim 5, wherein: The data preprocessing module comprises a data collection unit and a time-space alignment and expansion unit; The data acquisition unit is configured to acquire historical operation data of the membrane component, generate time series of each operation parameter, perform image binarization on a membrane surface topography image, perform full coverage of a foreground region based on a system preset window set, and count the number of full coverage of the foreground region for each window size; form a data pair based on the size of each window and the corresponding number of windows, perform linear fitting based on the data pair, and record the slope of the fitted straight line as an image feature value; The space-time alignment and expansion unit is configured to extract time series of any two operation parameters, analyze the shortest path of each operation parameter value of the two time series, perform interpolation based on the shortest path, and perform spatial interpolation using a Kriging method; and traverse all types of operation parameters to form an expanded time series of each operation parameter.
7. The artificial intelligence-based membrane component data management system of claim 5, wherein: The health analysis module includes a feature analysis unit and a health model construction unit. The feature analysis unit is configured to acquire the expanded time series of the transmembrane pressure difference, analyze the coefficient of variation of the transmembrane pressure difference, analyze the theoretical value of the membrane flux based on the expanded time series of the transmembrane pressure difference and Darcy's law, generate residual features by analyzing the theoretical value and the actual value of the membrane flux, perform Fourier transform on the expanded time series of the membrane flux, analyze the frequency spectrum energy and the main frequency component, and acquire the expanded time series of the water quality, flow rate, and inlet water temperature, and analyze the mean and standard deviation of each expanded time series. The health model construction unit constructs a multi-dimensional feature vector based on the coefficient of variation of the transmembrane pressure difference, the mean and standard deviation of each expanded time series, the residual features, the frequency spectrum energy, and the main frequency component. The multi-dimensional feature vector is compressed to form a latent space representation, and the latent space representation is reconstructed into original data; the multi-dimensional feature vector A is compressed to form a latent space representation, and the corresponding encoder expression of the latent space representation is h=Ψ(W e A+b e ); wherein h represents the latent space representation, W e represents the encoder weight matrix, b e represents the encoder bias vector, and Ψ() represents the activation function ReLU; the corresponding decoder expression of the latent space representation is A'=Ψ(W d h+b d ); wherein A' represents the reconstructed feature vector, W d represents the decoder weight matrix, b d represents the decoder bias vector; the weight matrices W e and W d are initialized by randomly sampling from a normal distribution with a mean of 0 and a variance of 2 / N, N being the number of neurons in the previous layer; the bias vectors b e and b d are initialized as zero vectors; The parameters of the encoder and the decoder are optimized by back propagation to minimize the reconstruction error. A health model is constructed based on the reconstruction error.
8. The artificial intelligence-based membrane component data management system of claim 5, wherein: The life prediction model construction module includes a life prediction model construction unit and a parameter optimization unit. The life prediction model construction unit acquires historical operation data of the entire process of the membrane component from the start of use to failure, constructs life analysis data pairs based on the historical operation data, and constructs a life prediction model according to the life analysis data pairs. The parameter optimization unit is configured to discretize the life prediction model into a difference equation; The gradient descent method is used to update the degradation rate coefficient and the health index influence coefficient until the total error reaches a minimum value or a maximum number of iterations is reached; The latest degradation rate coefficient and health index influence coefficient are bound to the corresponding membrane component.
9. The artificial intelligence-based membrane component data management system of claim 5, wherein: The real-time warning module includes a real-time model construction unit and a warning unit. The real-time model construction unit is configured to extract the degradation rate coefficient and the health index influence coefficient of a preset number of historical membrane components corresponding to the same production coordinates as the membrane component to be predicted, calculate the average value of the degradation rate coefficient and the average value of the health index influence coefficient of the historical membrane components as the initialization values of the degradation rate coefficient and the health index influence coefficient of the membrane component to be predicted, and construct a real-time life prediction model based on the initialization values of the degradation rate coefficient and the health index influence coefficient of the membrane component to be predicted. Real-time residual life prediction is performed. The pre-warning unit grades the pre-warning signals according to the real-time residual life and health index, and prompts the system according to the pre-warning signal grade. The pre-warning unit grades the pre-warning signals according to the real-time residual life and health index, and prompts the system according to the pre-warning signal grade.
Citation Information
Patent Citations
Aero-engine residual life prediction method based on transfer learning
CN113297787A
METHOD AND SYSTEM FOR ACCELERATED LIFE TESTING (ALT) OF PROTON EXCHANGE MEMBRANE FUEL CELLS (PEMFCs)
US20250087733A1