Intelligent evaluation system for reverse osmosis (RO) membrane health monitoring

Abstract

The application relates to the technical field of reverse osmosis (RO) membrane health monitoring, and discloses an intelligent evaluation system for RO membrane health monitoring. The system comprises a running data acquisition module for acquiring historical running data of multiple groups of RO membranes; a feature correlation modeling module for constructing a health feature map and initializing correlation weights according to the health influence parameters correlated with the historical running data; a state trend classification module for classifying the historical running data to generate health state trends and calling corresponding maps; an aging feature extraction module and an abnormal feature extraction module for respectively extracting core features through a health parameter analysis model and optimizing map weights; and a health state evaluation module for performing evaluation according to map calling parameters. The system integrates sliding average algorithm, Euclidean distance algorithm, self-encoder, GRU, LSTM and other technologies, realizes deep analysis of RO membrane running data and accurate evaluation of the health state, can early warn membrane performance attenuation and abnormalities, and reduces maintenance costs.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of reverse osmosis (RO) membrane health monitoring, and particularly to an intelligent evaluation system for RO membrane health monitoring. BACKGROUND

[0002] As a core component in water treatment processes, the performance of reverse osmosis (RO) membranes directly affects water quality and system operation efficiency. In actual applications, RO membranes may exhibit aging, pollution, and other problems due to long-term operation, resulting in decreased water production and reduced desalination rates. If not promptly monitored and addressed, this will seriously affect the normal operation of water treatment systems.

[0003] There are many deficiencies in the technical means for RO membrane health monitoring. Traditional monitoring methods rely heavily on manual analysis of operation data, which is not only time-consuming and labor-intensive, but also difficult to capture subtle changes in membrane performance in real time, making it difficult to meet the demand for rapid evaluation of RO membrane health status in actual production. Existing evaluation systems often use fixed evaluation models and weight settings, which cannot be dynamically adjusted according to different operating stages and maintenance frequencies of RO membranes, resulting in insufficient recognition accuracy of membrane aging and abnormal states.

[0004] Existing technologies lack deep mining of data features when processing RO membrane operation data, making it difficult to effectively extract core features reflecting the health status of the membrane, resulting in inaccurate health status trend analysis and inability to predict membrane performance degradation and potential failures in advance. Meanwhile, most evaluation systems do not consider the impact of maintenance time and frequency on membrane performance, and the data classification method is simple, resulting in an unscientific construction of health feature maps, which further affects the reliability of health status evaluation. These problems make the existing RO membrane health monitoring technology have the disadvantages of evaluation lag, high maintenance cost, and short service life of the membrane in actual applications, and there is an urgent need for a more intelligent and accurate health monitoring system to solve the above problems. SUMMARY

[0005] The present application aims to provide an intelligent evaluation system for reverse osmosis (RO) membrane health monitoring to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides an intelligent evaluation system for reverse osmosis (RO) membrane health monitoring, which comprises:

[0007] An operation data acquisition module acquires historical operation data of multiple groups of reverse osmosis (RO) membranes;

[0008] A feature correlation modeling module correlates a first health influence parameter and a second health influence parameter based on the historical operation data, constructs a health feature map, and initializes the correlation weight of the health feature map;

[0009] a state trend classification module, which classifies a plurality of sets of the historical operation data of the reverse osmosis RO membranes within an S period, generates health state trends, compares and labels the health state trends, and calls corresponding health feature maps;

[0010] an aging feature extraction module, which calls the first health impact parameter according to the health feature maps of the same category, extracts first and second core features of the first health impact parameter through a health parameter analysis model, and optimizes a first weight of the health feature map;

[0011] an abnormal feature extraction module, which calls the second health impact parameter according to the health feature maps of the same category, extracts third and fourth core features of the second health impact parameter through the health parameter analysis model, and optimizes a second weight of the health feature map;

[0012] a health state evaluation module, which evaluates the first and second health impact parameters according to the health feature maps during evaluation.

[0013] Preferably, classifying the plurality of sets of the historical operation data of the reverse osmosis RO membranes within an S period comprises: dividing the historical operation data according to the maintenance time S as a dividing point, and classifying the data according to the total number of the maintenance time S as a maintenance frequency.

[0014] Preferably, comparing the health state trends and calling the corresponding health feature maps comprises:

[0015] extracting the health state trends of the divided historical operation data through a sliding average algorithm, calculating the mean of the health state trends of the same maintenance frequency as a standard health trend;

[0016] calculating the deviation of a plurality of the health state trends of the same maintenance frequency from the standard health trend, labeling a plurality of trend categories according to the deviation, and calling the health feature maps of the corresponding categories;

[0017] comparing the health state trends of the same reverse osmosis RO membrane at the mth maintenance frequency and the (m+1)th maintenance frequency according to the Euclidean distance algorithm, and if the comparison result is less than a critical value of the health state trend, calling the health feature map of the reverse osmosis RO membrane at the (m+1)th maintenance frequency.

[0018] Preferably, calling the first health impact parameter according to the health feature maps of the same category, extracting the first and second core features of the first health impact parameter through the health parameter analysis model, and optimizing the first weight of the health feature map comprises:

[0019] The first core feature and the second core feature are obtained according to a performance attenuation analysis model of the health parameter analysis model; the performance attenuation analysis model comprises an impact parameter preprocessing unit, an impact parameter feature analysis unit, a first weight optimization unit and a first weight storage unit;

[0020] The impact parameter preprocessing unit performs data segmentation and data normalization on the first health impact parameter to generate input data; the impact parameter feature analysis unit predicts next segmented time data through a 1-layer autoencoder and a 5-layer GRU, and identifies the first core feature of the input data and the second core feature of the next segmented time data through a 3-layer deep convolutional layer, a 1-layer empty convolutional layer and a 4-layer GRU;

[0021] The first weight optimization unit optimizes the correlation weight of the health feature map according to the first core feature and the second core feature; and the first weight storage unit stores the correlation weight of the health feature map.

[0022] Preferably, the first weight optimization unit optimizing the correlation weight of the health feature map according to the first core feature and the second core feature comprises: linearly adjusting the initial correlation weight of the health feature map according to the feature importance degree output by the performance attenuation analysis model, combining the stability of the first core feature and the predictability of the second core feature.

[0023] Preferably, the calculation of the feature importance degree comprises: grouping the input core features through a hierarchical clustering algorithm, calculating the similarity between the features by combining a dynamic time warping algorithm, and determining the importance level label of each group of features according to the feature similarity distribution and the within-group dispersion.

[0024] Preferably, calling the second health impact parameter according to the health feature map of the same category, extracting the third core feature and the fourth core feature of the second health impact parameter through the health parameter analysis model and optimizing the second weight of the health feature map comprises:

[0025] The third core feature and the fourth core feature are obtained according to an abnormal diagnosis analysis model of the health parameter analysis model; the abnormal diagnosis analysis model comprises an abnormal parameter preprocessing unit, an abnormal parameter feature extraction unit, a second weight optimization unit and a second weight storage unit;

[0026] The abnormal parameter preprocessing unit performs data segmentation and data normalization on the second health influence parameter to generate to-be-recognized data; the abnormal parameter feature extraction unit extracts the third core feature of the to-be-recognized data according to a one-layer attention mechanism, a one-layer time convolution layer and four two-way LSTM units, and extracts the fourth core feature of the to-be-recognized data through two residual convolution layers, one hollow convolution layer and four LSTM units;

[0027] The second weight optimization unit optimizes the correlation weight of the second health influence parameter based on the third core feature and the fourth core feature; and the second weight storage unit stores the optimized correlation weight.

[0028] Preferably, the second weight optimization unit optimizing the correlation weight of the second health influence parameter based on the third core feature and the fourth core feature comprises: performing exponential adjustment on the initial correlation weight of the second health influence parameter according to the feature abnormality degree output by the abnormal diagnosis analysis model, combining the saliency of the third core feature and the persistence of the fourth core feature.

[0029] Preferably, when the total number of maintenance times according to the maintenance time S is used as the maintenance frequency for data classification, if the maintenance frequency is less than the set frequency threshold, it is classified into the initial maintenance category, and if the maintenance frequency is greater than or equal to the set frequency threshold, it is classified into the stable maintenance category.

[0030] Preferably, the way in which the health state evaluation module evaluates according to the health feature map to call the corresponding first health influence parameter and second health influence parameter comprises: extracting the first health influence parameter and the second health influence parameter in the current running data, respectively matching the optimized first weight and the second weight in the health feature map, and calculating a comprehensive health index as an evaluation basis through weighted summation.

[0031] Compared with the prior art, the present application has the following advantages:

[0032] The system acquires multiple groups of historical running data through the running data acquisition module, providing a rich data basis for subsequent feature analysis and state evaluation, ensuring the reliability of the evaluation results. The feature correlation modeling module constructs a health feature map and initializes the correlation weight, establishing a scientific framework for the evaluation of the health state of the RO membrane, making the evaluation process have rules to follow. The state trend classification module classifies the historical running data according to the maintenance time and the maintenance frequency, generates a health state trend, and through comparison and analysis by the sliding average algorithm and the Euclidean distance algorithm, can accurately identify the membrane state trend at different maintenance stages, improving the pertinence and accuracy of the evaluation.

[0033] The aging feature extraction module and the abnormal feature extraction module respectively utilize a performance degradation analysis model and an abnormal diagnosis analysis model, combine various machine learning algorithms such as self-encoder, GRU, attention mechanism, bidirectional LSTM, and the like, deeply mine core features of the first health influence parameter and the second health influence parameter, and dynamically optimize the correlation weight of the health feature spectrum according to the feature importance and the abnormality degree, so that the system can adaptively adjust the evaluation model, better adapt to the performance change of the RO membrane at different stages, and improve the recognition ability of the membrane aging and abnormal state.

[0034] The health state evaluation module calculates the comprehensive health index by extracting the influence parameters in the current running data, matching and weighting the optimized weight for summation, and realizes the precise evaluation of the RO membrane health state. The dynamic weight adjustment mechanism and the multi-model collaborative working mode not only improve the processing ability of the system to complex data and the evaluation precision of the membrane health state, but also can find the performance degradation trend and abnormal signs of the membrane in advance, provide a scientific basis for maintenance decision, reduce unnecessary maintenance cost, prolong the service life of the RO membrane, and improve the operation efficiency and stability of the water treatment system. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A working principle diagram of the intelligent evaluation system for reverse osmosis (RO) membrane health monitoring is provided.

[0036] Figure 2 A flowchart for health state trend comparison and atlas calling is provided.

[0037] Figure 3 A flowchart for first health influence parameter feature extraction is provided.

[0038] Figure 4 A flowchart for second health influence parameter feature extraction is provided. DETAILED DESCRIPTION

[0039] 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.

[0040] Please refer to Figures 1-4 The present application provides an intelligent evaluation system for reverse osmosis (RO) membrane health monitoring, which comprises.

[0041] The running data acquisition module acquires a plurality of groups of historical running data of reverse osmosis (RO) membranes.

[0042] The feature association modeling module constructs a health feature map by associating the first health impact parameter and the second health impact parameter with the historical operating data, and initializes the association weights of the health feature map.

[0043] The status trend classification module classifies the historical operating data of the multiple sets of reverse osmosis (RO) membranes within the S-cycle, generates health status trends, compares and labels the health status trends, and calls the corresponding health feature map.

[0044] The aging feature extraction module calls the first health impact parameter based on the health feature map of the same category, extracts the first core feature and the second core feature of the first health impact parameter through the health parameter analysis model, and optimizes the first weight of the health feature map.

[0045] The abnormal feature extraction module calls the second health impact parameter based on the health feature map of the same category, extracts the third and fourth core features of the second health impact parameter through the health parameter analysis model, and optimizes the second weight of the health feature map;

[0046] The health status assessment module uses the health characteristic map to call the corresponding first health impact parameter and second health impact parameter for assessment. Example

[0047] The historical operating data of multiple sets of reverse osmosis (RO) membranes within the S-cycle are classified, and the specific implementation method is as follows:

[0048] The system first acquires historical operating data from multiple sets of reverse osmosis (RO) membranes. This historical operating data includes the operating status information of the RO membranes at different points in time. Next, based on all the historical operating data of the RO membranes, the system divides the historical operating data using maintenance time S as the dividing point. Here, maintenance time S refers to the specific time point when maintenance operations are performed on the RO membranes; the system needs to accurately identify and determine the exact moment of each maintenance time S.

[0049] After determining the maintenance time S, the system classifies the data according to the total number of maintenance times S as the maintenance frequency. In practice, for each set of historical operating data for reverse osmosis (RO) membranes, the system counts the number of maintenance times S contained in that set of data; this number is the maintenance frequency for that RO membrane. For example, if a set of historical operating data for a RO membrane records 3 maintenance times S, then the maintenance frequency for that set of data is 3 times.

[0050] After completing the statistics on maintenance frequency, the system needs to categorize the data according to the frequency. A frequency threshold is set here: if the maintenance frequency is less than this threshold, the corresponding historical operating data is classified as initial maintenance; if the maintenance frequency is greater than or equal to the threshold, it is classified as stable maintenance. Determining the frequency threshold requires considering factors such as the characteristics of the reverse osmosis (RO) membrane and maintenance strategies in the actual application scenario. For example, in some application scenarios, the frequency threshold might be set to 5 times. When the maintenance frequency is less than 5 times, the RO membrane is considered to be in the initial maintenance stage, and its operating status and performance may still be gradually stabilizing; when the maintenance frequency is greater than or equal to 5 times, the RO membrane is considered to have entered the stable maintenance stage, and its operating status is relatively stable.

[0051] In the specific implementation process, the system needs to efficiently process and analyze historical operating data. First, it needs to accurately extract the maintenance time S from a large amount of historical operating data, which may involve data parsing and filtering. For example, historical operating data may be stored in the form of log files or database records, and the system needs to be able to identify the timestamps or related records corresponding to maintenance operations. Then, the extracted maintenance times S are statistically analyzed to calculate the maintenance frequency for each reverse osmosis (RO) membrane. This step needs to ensure the accuracy of the statistics to avoid errors in the calculation of maintenance frequency due to data omissions or errors.

[0052] When classifying maintenance categories, the system needs to compare the maintenance frequency with a set frequency threshold. This comparison process must strictly follow the set logic to ensure the accuracy of data classification. For example, if a reverse osmosis (RO) membrane is maintained 4 times, and the frequency threshold is set to 5 times, the data should be correctly classified as initial maintenance; if the maintenance frequency is 5 times, it should be classified as stable maintenance.

[0053] Furthermore, the system needs to be capable of handling different data formats and storage methods. Since historical operational data may come from different data sources, their formats and storage methods may differ. The system needs to be able to process and transform this data uniformly to ensure the smooth extraction of maintenance time S and the statistical analysis of maintenance frequency. For example, for structured data, maintenance time S can be extracted using database queries; for unstructured data, text parsing techniques may be needed to identify maintenance time information.

[0054] In practical applications, some special situations may arise, such as incomplete or inaccurate records of maintenance time S. To address these issues, the system needs to possess a certain degree of fault tolerance and processing capabilities. For example, a certain time tolerance range can be set; maintenance time records within this range are considered valid. For missing maintenance time records, data interpolation or other methods can be used to complete the records, ensuring the accuracy of maintenance frequency statistics.

[0055] Furthermore, the frequency threshold should not be fixed; the system needs to allow for adjustments based on actual operating conditions and maintenance effectiveness. For example, after the system has been running for a period of time, analysis of the health status assessment results and maintenance effectiveness of the reverse osmosis (RO) membrane may reveal that the current frequency threshold setting is unreasonable, leading to inaccurate data classification. In this case, the frequency threshold can be reset and adjusted to improve the system's adaptability and accuracy.

[0056] Through the above series of operations and processes, the system can accurately classify the historical operating data of multiple sets of reverse osmosis (RO) membranes within the S-cycle into initial maintenance and stable maintenance categories according to maintenance frequency. This classification method provides an important foundation for subsequent health status trend analysis and health feature map retrieval, enabling the system to perform more targeted analysis and evaluation of RO membranes at different maintenance stages, thereby improving the accuracy and effectiveness of the entire intelligent evaluation system. Example

[0057] The specific implementation method is as follows: Compare health status trends and retrieve corresponding health feature maps.

[0058] After classifying multiple sets of historical operating data for reverse osmosis (RO) membranes, the system needs to compare health status trends and retrieve corresponding health feature maps. The health status trend of the segmented historical operating data is extracted using a moving average algorithm. During the application of the moving average algorithm, the system sets an appropriate sliding window size, which needs to consider factors such as the time span and data density of the historical operating data. For example, if the historical operating data has a large time span and dense data points, a smaller sliding window may be selected to capture data changes more precisely; if the data points are sparse, a larger sliding window may be selected to smooth out data fluctuations.

[0059] When applying the moving average algorithm, the system moves the sliding window sequentially over the divided historical operating data, averaging the data within each window to obtain the corresponding health trend value. This method effectively reduces random noise in the data and extracts a health trend that better reflects the actual operating status of the reverse osmosis (RO) membrane.

[0060] The average health status trend for maintenance frequencies is calculated as the standard health trend. The system first groups historical operating data according to maintenance frequency. For each maintenance frequency group, the arithmetic mean of all health status trend data is calculated. For example, if there are 10 health status trend data in a certain maintenance frequency group, the system adds these 10 data points together and divides by 10 to obtain the standard health trend for that maintenance frequency. This standard health trend represents the typical health status trend of the reverse osmosis (RO) membrane under normal operation at that maintenance frequency.

[0061] The system calculates the deviation of multiple health status trends with the same maintenance frequency from the standard health trend, classifies them into multiple trend categories based on the deviation, and calls up the corresponding health feature maps. The deviation can be calculated using methods such as Euclidean distance. Specifically, for each health status trend data, the system calculates the Euclidean distance between it and the standard health trend at that maintenance frequency; this distance value is the deviation.

[0062] After obtaining the deviation, the system needs to set different deviation ranges to classify the trend categories. For example, a deviation less than a certain threshold A can be defined as a normal trend category, a deviation between threshold A and threshold B as a slight deviation trend category, and a deviation greater than threshold B as a severe deviation trend category. The setting of these thresholds needs to be determined in conjunction with the actual operation of the reverse osmosis (RO) membrane and engineering experience to ensure that the trend category classification accurately reflects the health status of the RO membrane.

[0063] For each health status trend data point, the system determines its trend category based on its deviation and labels it with the corresponding category identifier. After labeling, the system retrieves the corresponding health feature map based on different trend categories. Different trend categories may correspond to different health feature maps, allowing for more targeted analysis of the health status of the reverse osmosis (RO) membrane.

[0064] The system compares the health status trends of the same reverse osmosis (RO) membrane at the m-th and (m+1)-th maintenance frequencies using the Euclidean distance algorithm. If the comparison result is less than the critical value of the health status trend, the system retrieves the health feature map of the RO membrane corresponding to the (m+1)-th maintenance frequency. In this process, the system first obtains the health status trend data of the same RO membrane at the m-th and (m+1)-th maintenance frequencies.

[0065] The Euclidean distance algorithm is used to calculate the distance between the two health status trends. In calculating the Euclidean distance, the two health status trend data points are treated as vectors in a multi-dimensional space, and the straight-line distance between them is calculated. The critical value for the health status trend is a preset standard value used to determine whether the health status trends under two maintenance frequencies have sufficient similarity.

[0066] If the calculated Euclidean distance is less than the critical value, it indicates that the health status trends at the m-th and m+1-th maintenance frequencies are quite similar. In this case, the system will call the health feature map corresponding to the m+1-th maintenance frequency. This is because, under these circumstances, the health status trend after the m+1-th maintenance is less different from the trend after the previous maintenance, and calling the health feature map of the m+1-th maintenance can more accurately reflect the current health status.

[0067] Throughout the implementation process, the system needs to ensure that the parameters of the moving average algorithm are set reasonably to accurately extract the health status trend; the calculation of the standard health trend should be based on a sufficient amount of valid data to ensure its representativeness; the calculation of deviation and the classification of trend categories should be scientific and reasonable to truly reflect the differences in the health status of the reverse osmosis (RO) membrane; and the calculation of Euclidean distance and the setting of critical values ​​should be combined with the actual situation to ensure the accuracy of the health feature map call.

[0068] In addition, the system needs to be capable of processing large amounts of data, efficiently performing operations such as moving average calculation, mean calculation, deviation calculation, and Euclidean distance calculation. When faced with missing or abnormal data, the system needs corresponding handling mechanisms, such as interpolating missing data and identifying and filtering abnormal data, to ensure the accuracy of the calculation results.

[0069] At the same time, the system should allow adjustments to relevant parameters based on actual operating conditions, such as sliding window size, deviation threshold, and critical value, to adapt to different reverse osmosis (RO) membrane operating scenarios and maintenance needs, thereby improving the system's flexibility and adaptability. Example

[0070] Based on the health feature maps of the same category, the first health impact parameter is invoked. The first and second core features are extracted through the health parameter analysis model, and the first weight of the health feature map is optimized. The specific implementation method is as follows:

[0071] Taking a set of reverse osmosis (RO) membranes that have been operating for six months as an example, the system first classifies their historical operating data into a stable maintenance category through the status trend classification module. At this point, the feature association modeling module has already constructed a health feature map for this category, where the first health impact parameter includes operating data such as influent pressure, temperature, and pH value. The aging feature extraction module calls the first health impact parameter from this category map, for example, to obtain the daily influent pressure, temperature, and pH value data of the RO membrane over the past three months. This data is stored in time series format, including sampling points every minute.

[0072] The preprocessing unit for impact parameters in the performance degradation analysis model processes these primary health impact parameters. Taking influent pressure data as an example, the preprocessing unit divides the three months of data into 12 data segments by week, with each segment corresponding to one week's pressure data. For each data segment, normalization is performed, mapping the pressure values ​​from 0.8MPa-1.2MPa to the 0-1 range; for example, an actual pressure value of 1.0MPa is converted to 0.5. Similarly, temperature data (range 20℃-30℃) and pH data (range 6.5-8.5) are segmented and normalized to generate an input data matrix. Each row in the matrix represents the normalized parameter value for a specific day, and each column corresponds to a different parameter.

[0073] After receiving the input data, the influencing parameter feature parsing unit performs dimensionality reduction on the data using a single-layer autoencoder. The autoencoder's encoding layer contains 10 neurons, compressing 7 days of influent pressure data (1440 sampling points per day) into a 50-dimensional feature vector. The decoding layer then attempts to reconstruct the original data to extract its core feature representation. Simultaneously, a 5-layer GRU network processes the time series data. For example, the first GRU layer receives the daily normalized pressure data, the second layer combines it with temperature data, the third layer adds pH data, the fourth layer predicts the pressure trend for the following week, and the fifth layer outputs the predicted pressure data sequence. In this way, the GRU network captures the time series patterns of influent pressure changes with temperature and pH, predicting the influent pressure data for the next time segment (i.e., the following week).

[0074] In identifying core features, three deep convolutional layers process the normalized pressure, temperature, and pH data. The first convolutional layer uses a 3×3 kernel with a stride of 1 to extract local features of the data, such as pressure fluctuation patterns over a certain time period. The second convolutional layer uses a larger kernel (5×5) to capture features over a longer time span. The third convolutional layer combines the features from the first two layers to generate a more abstract feature representation. A dilated convolutional layer processes the data with a dilation rate of 2, filling data gaps and capturing sparser feature patterns, such as weekly periodic pressure changes. A four-layer GRU network further processes the features output from the convolutional layers. The first GRU layer identifies short-term features of daily parameter changes, the second GRU layer captures medium-term features of weekly changes, the third GRU layer analyzes long-term features of monthly changes, and the fourth GRU layer integrates all features to identify the first core feature of the input data (such as the sensitivity parameter of influent pressure to temperature changes) and the second core feature of the next time segment (such as the amplitude parameter of pressure fluctuations in the next week).

[0075] The first weight optimization unit linearly adjusts the initial association weights of the health feature map based on the feature importance output by the performance degradation analysis model, combined with the stability of the first core feature and the predictability of the second core feature. For example, in feature importance calculation, the hierarchical clustering algorithm divides the first core feature (such as the pressure sensitivity parameter) and the second core feature (such as the weekly pressure fluctuation amplitude) into two groups. The dynamic time warping algorithm calculates that the similarity between the pressure sensitivity parameter and historical fault data is 0.75, and the similarity between the fluctuation amplitude parameter and historical fault data is 0.6. Based on the feature similarity distribution (setting similarity above 0.7 as high importance) and the intra-group dispersion (the dispersion of the pressure sensitivity parameter is 0.15, and the dispersion of the fluctuation amplitude parameter is 0.2), the importance level label of the pressure sensitivity parameter is determined to be "high", and the importance level label of the fluctuation amplitude parameter is "medium".

[0076] Based on this, the initial association weights are as follows: influent pressure 0.3, temperature 0.25, and pH 0.25. Since pressure sensitivity parameters are of high importance, their stability index is 0.8 (out of 1), and their predictive index is 0.7. The system linearly adjusts the weight of influent pressure: 0.3 + 0.3 × (0.8 × 0.75 + 0.7 × 0.25) × 0.1 (adjustment coefficient), resulting in a weight of 0.33. The weights of temperature and pH are also adjusted accordingly based on the importance and index of their respective core features. The adjusted weights are stored in the health feature map through the first weight storage unit, updating the original association weights.

[0077] Throughout the implementation process, the system must ensure the rationality of data segmentation. For example, weekly segmentation should consider the maintenance cycle of the RO membrane. Normalization processing must maintain the physical meaning of parameters and avoid over-mapping that leads to feature distortion. The parameter settings of the autoencoder and GRU network need to be adjusted according to the amount of data; for example, the number of neurons should be increased when the data volume is large. The number of categories in hierarchical clustering needs to be determined based on the feature distribution to avoid clustering that is too fine or too coarse. The distance metric of the dynamic time warping algorithm needs to match the feature type; for example, Euclidean distance should be used for continuous parameters. The setting of adjustment coefficients should refer to engineering experience; for example, a coefficient of 0.1 can avoid excessive weight adjustment. When data anomalies occur during a certain period (such as a sudden drop in influent pressure), the system needs to filter the data using an outlier detection algorithm before feature extraction to prevent abnormal data from affecting the identification of core features. If changes in maintenance frequency lead to a reclassification of data categories, the system will re-call the health feature map of the corresponding category, re-extract the core features, and optimize the weights to ensure that the weights always match the current operating status. Example

[0078] Based on the health feature maps of the same category, the second health impact parameter is invoked. The third and fourth core features are extracted through the health parameter analysis model, and the second weight of the health feature map is optimized. The specific implementation method is as follows:

[0079] Taking a reverse osmosis (RO) membrane system as an example, this system has been running for 8 months and has been classified into the stable maintenance category by the status trend classification module. The second health impact parameter in its health feature spectrum includes data such as permeate flow rate, desalination rate, and feed water flow rate. The abnormal feature extraction module calls the second health impact parameter in this category spectrum, for example, to obtain the daily permeate flow rate data, desalination rate data, and feed water flow rate data of the RO membrane for the past 2 months. These data are recorded hourly, with 24 sampling points per day.

[0080] The abnormal parameter preprocessing unit in the abnormal diagnostic analysis model processes these secondary health impact parameters. Taking permeate production data as an example, the preprocessing unit divides two months of data into eight data segments, each segment corresponding to five days of permeate production data. For each data segment, normalization is performed, mapping the permeate production from 80 m³ / h to 120 m³ / h to the 0-1 range; for example, an actual permeate production of 100 m³ / h is converted to 0.5. Similarly, the desalination rate data (range 98%-99.5%) and influent flow rate data (range 150 m³ / h-180 m³ / h) are segmented and normalized to generate a data matrix to be identified. Each row in the matrix represents the normalized parameter value for a specific hour, and each column corresponds to a different parameter.

[0081] After receiving the data to be identified, the anomaly parameter feature extraction unit processes the data through a single attention mechanism. This mechanism assigns different weights to different parameters such as permeate flow rate, desalination rate, and influent flow rate. For example, when permeate flow rate fluctuates abnormally, the system automatically increases the attention weight of the permeate flow rate parameter, making it play a more significant role in the feature extraction process. Next, a temporal convolutional layer processes the time-series data. This layer employs causal convolution to ensure that processing current data does not depend on data from future times, thus more accurately capturing the temporal dependencies of the data. For instance, the temporal convolutional layer can identify a continuous downward trend in permeate flow rate over the past few hours.

[0082] A four-layer bidirectional LSTM unit further extracts features from the data. The first layer of the bidirectional LSTM processes the daily water production data in both forward and reverse directions to capture the changing characteristics of water production within the day. The second layer of the bidirectional LSTM combines desalination rate data to analyze the correlation between water production and desalination rate. The third layer of the bidirectional LSTM incorporates influent flow rate data to explore the comprehensive influence characteristics among the three factors. The fourth layer of the bidirectional LSTM integrates all features to extract the third core feature of the data to be identified, such as the abnormal rate of decline of water production within a specific time period.

[0083] Simultaneously, the data is processed through two residual convolutional layers. The first residual convolutional layer retains some features of the original data through skip connections, avoiding the loss of important information during convolution, such as preserving the basic trend of water production changes. The second residual convolutional layer further extracts more complex features, such as the pattern features of abnormal fluctuations in water production. A single dilated convolutional layer processes the data with an expansion rate of 3, expanding the receptive field and capturing longer-term feature dependencies, such as the periodic variation features of weekly water production. Four LSTM units process the features output from the residual and dilated convolutional layers. The first LSTM identifies short-term abnormal features of water production, the second LSTM captures medium-term desalination rate changes, the third LSTM analyzes long-term influent flow fluctuations, and the fourth LSTM integrates all features to extract the fourth core feature of the data to be identified, such as the persistence parameter of the abnormal correlation between desalination rate and influent flow.

[0084] The second weight optimization unit optimizes the association weights of the second health impact parameters based on the third and fourth core features. Specifically, based on the feature anomaly degree output by the anomaly diagnosis analysis model, and combining the significance of the third core feature and the persistence of the fourth core feature, the initial association weights of the second health impact parameters are exponentially adjusted. For example, in the feature anomaly degree calculation, by comparing the difference between the current feature and historical normal features, the anomaly degree of the abnormal rate of decline in permeable water production is determined to be 0.8, and the anomaly degree of the abnormal association between desalination rate and influent flow rate is 0.6. The significance index of the third core feature is 0.9 (out of 1), indicating that the abnormal rate of decline in permeable water production is prominent in anomaly diagnosis; the persistence index of the fourth core feature is 0.7, indicating that the abnormal association between desalination rate and influent flow rate has persisted for some time.

[0085] In the initial association weights, the weight of permeate flow rate was 0.35, the desalination rate was 0.3, and the influent flow rate was 0.25. Due to the high anomaly and significance of the abnormal rate of decrease in permeate flow rate, the system exponentially adjusted the weight of permeate flow rate: 0.35 × exp(0.8 × 0.9 × 0.2) (adjustment coefficient of 0.2), resulting in a weight of approximately 0.35 × 1.15 = 0.4025. The weights of the desalination rate and influent flow rate were also adjusted accordingly based on the anomaly and indicators of their respective core features. The adjusted weights were stored in the health feature map through the second weight storage unit, updating the original second weight association.

[0086] Throughout the implementation process, the system must ensure that the time interval of data segmentation matches the anomaly monitoring cycle of the RO membrane. For example, segmenting by 5 days can effectively capture weekly anomaly patterns. During normalization, the physical dimension correspondence of the parameters must be preserved to avoid distortion of anomaly features due to normalization. The weight allocation of the attention mechanism needs to be trained based on historical anomaly data; for example, the attention priority of different parameters can be determined by learning from past failure cases. The kernel size and stride of the temporal convolutional layer need to be adjusted according to the data sampling frequency; for example, a 3×1 kernel can be used for hourly data. The number of neurons in the hidden layer of the bidirectional LSTM needs to be set according to the complexity of the data features; when the features are complex, it can be increased to 128 neurons.

[0087] The skip connection design of residual convolutional layers must ensure the effective propagation of original features and avoid the gradient vanishing problem in deep networks. The dilation rate setting of dilated convolutional layers must consider the time span of the data; for example, a dilation rate of 3 can capture feature dependencies across 3 time steps. The forget gate, input gate, and output gate parameters of LSTM units need to be trained using backpropagation to optimize feature extraction performance. The calculation of feature anomalies requires establishing a reasonable baseline of normal features, such as a baseline model based on normal operating data from the past 3 months. The adjustment coefficient in exponential adjustment should refer to industry standards; for example, a coefficient of 0.2 can keep the weight adjustment range within a reasonable range.

[0088] When the system detects missing data for a certain period (e.g., 2 hours of missing permeate flow data due to sensor malfunction), it first uses linear interpolation to fill in the missing data before feature extraction to prevent data loss from affecting the accuracy of core features. If maintenance operations on the RO membrane cause a significant change in its operating status, the system will reclassify the data and call the corresponding health feature map according to the new category, re-extract core features and optimize weights to ensure that the weights can reflect the current operating status of the RO membrane in real time. Example

[0089] The health status assessment module uses the health characteristic map to call the corresponding first and second health impact parameters during the assessment. The specific implementation method is as follows:

[0090] Taking a set of reverse osmosis (RO) membranes in a water treatment plant as an example, this set of RO membranes has been operating for 10 months and has been classified as a stable maintenance category by the status trend classification module. Its health characteristic spectrum shows that the first health impact parameters include influent pressure, temperature, and pH value, with optimized first weights of 0.35, 0.25, and 0.2, respectively; the second health impact parameters include permeate flow rate, desalination rate, and influent flow rate, with optimized second weights of 0.3, 0.3, and 0.25, respectively. The system collects the current operating data of this set of RO membranes in real time. For example, at a certain moment, the first health impact parameters obtained are: influent pressure 1.1 MPa, temperature 25℃, pH value 7.2; the second health impact parameters are: permeate flow rate 110 m³ / h, desalination rate 98.8%, and influent flow rate 165 m³ / h.

[0091] After extracting the first and second health impact parameters from the current operating data, it is necessary to match them with the optimized first and second weights in the health feature map. For the first health impact parameter, the influent pressure of 1.1 MPa needs to be normalized first. The normalized range of the influent pressure recorded in the health feature map is 0.8 MPa-1.2 MPa. Therefore, the calculation method for mapping 1.1 MPa to the 0-1 interval is: (1.1-0.8) / (1.2-0.8)=0.75. The normalized range of the temperature of 25℃ is 20℃-30℃, and after normalization, it is (25-20) / (30-20)=0.5. The normalized range of the pH value of 7.2 is 6.5-8.5, and after normalization, it is (7.2-6.5) / (8.5-6.5)=0.35.

[0092] For the second health impact parameter, the normalized range for a permeable flow rate of 110 m³ / h is 80 m³ / h - 120 m³ / h, and after normalization, it is (110 - 80) / (120 - 80) = 0.75. The normalized range for a desalination rate of 98.8% is 98% - 99.5%, and after normalization, it is (98.8 - 98) / (99.5 - 98) ≈ 0.533. The normalized range for an influent flow rate of 165 m³ / h is 150 m³ / h - 180 m³ / h, and after normalization, it is (165 - 150) / (180 - 150) = 0.5.

[0093] After normalization, the system begins weighted summation calculation of the comprehensive health index. The weighted calculation of the first health impact parameter is as follows: normalized influent pressure 0.75 × weight 0.35 = 0.2625, normalized temperature 0.5 × weight 0.25 = 0.125, normalized pH value 0.35 × weight 0.2 = 0.07. The sum of these three values ​​gives the weighted sum of the first health impact parameter as 0.2625 + 0.125 + 0.07 = 0.4575.

[0094] The weighted calculation of the second health impact parameter is as follows: normalized value of permeate flow rate 0.75 × weight 0.3 = 0.225, normalized value of desalination rate 0.533 × weight 0.3 ≈ 0.1599, normalized value of influent flow rate 0.5 × weight 0.25 = 0.125. The weighted sum of the three is 0.225 + 0.1599 + 0.125 ≈ 0.5099.

[0095] Adding the weighted sum of the first health impact parameter to the weighted sum of the second health impact parameter yields a comprehensive health index of 0.4575 + 0.5099 ≈ 0.9674. The system's preset health assessment threshold is 0.8. When the comprehensive health index is greater than or equal to 0.8, the current health status of the RO membrane group is considered good; when the comprehensive health index is between 0.5 and 0.8, it is considered a sub-healthy state requiring attention; and when the comprehensive health index is less than 0.5, it is considered an abnormal state requiring maintenance. Since the comprehensive health index of this group of RO membranes is approximately 0.9674, which is greater than 0.8, the system assesses its current health status as good.

[0096] In practical applications, the weights in the health feature map are not fixed but are updated in real time as the aging feature extraction module and the anomaly feature extraction module optimize the weights. For example, if the RO membrane group experiences a continuous decline in permeate flow during subsequent operation, the anomaly feature extraction module will re-extract the core features of the second health impact parameter and optimize the second weight. Assume the weight of permeate flow is adjusted from 0.3 to 0.35 after optimization. When the permeate flow is collected again at 110 m³ / h (normalized to 0.75), its weighted value will become 0.75 × 0.35 = 0.2625, an increase from the previous 0.225. This makes the calculation results of the comprehensive health index more reflective of the impact of permeate flow on the health status of the RO membrane.

[0097] When collecting current operating data, the system verifies the validity of the data. For example, if the inlet pressure data shows a significant jump at a certain moment, exceeding the normal range (e.g., suddenly rising from 1.1 MPa to 1.8 MPa), the system will first process the abnormal data using a data filtering algorithm to determine whether the anomaly is caused by a sensor malfunction or a genuine change in operating status. If it is confirmed to be a sensor malfunction, the valid data from the previous moment will be used for interpolation to ensure the accuracy of the input first and second health impact parameters.

[0098] Furthermore, the health status assessment module retrieves corresponding health feature maps based on different maintenance frequency categories. For example, when the RO membrane undergoes its 6th maintenance, increasing the maintenance frequency from 5 to 6, the system automatically retrieves the corresponding health feature map from the stable maintenance category for the 6th maintenance frequency. The weights in this map may differ from those in the map during the 5th maintenance due to previous optimizations. Therefore, even if the collected current operating data is the same, the calculated comprehensive health index will differ due to the change in weights, thus more accurately reflecting the health status of the RO membrane after maintenance.

[0099] When calculating the comprehensive health index, the system also considers the impact of different operating periods. For example, during the high-temperature period in summer, temperature may have a greater impact on RO membrane performance, and the temperature weight in the health feature map may be increased based on optimization of historical data. Assuming the temperature weight is adjusted from 0.25 to 0.3, when a temperature of 25℃ (normalized to 0.5) is collected, its weighted value will become 0.5 × 0.3 = 0.15, an increase from the previous 0.125, making the comprehensive health index more reflective of the impact of temperature on RO membrane health during high-temperature periods.

[0100] Through the above specific implementation methods, the health status assessment module can accurately calculate the first and second health impact parameters in the current operating data based on the optimized weights in the health feature map, thereby obtaining a comprehensive health index and achieving a scientific assessment of the health status of the reverse osmosis (RO) membrane. During this process, the system updates the weights in real time, verifies the validity of the data, matches maps with different maintenance frequencies, and considers the impact of operating periods to ensure the accuracy and reliability of the assessment results, providing a basis for RO membrane maintenance decisions.

[0101] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent assessment system for monitoring the health of reverse osmosis (RO) membranes, characterized in that, include: Run the data acquisition module to obtain multiple sets of historical operating data for the reverse osmosis (RO) membrane; The feature association modeling module constructs a health feature map by associating the first health impact parameter and the second health impact parameter with the historical operating data, and initializes the association weights of the health feature map. The status trend classification module classifies the historical operating data of the multiple sets of reverse osmosis (RO) membranes within the S-cycle, generates health status trends, compares and labels the health status trends, and calls the corresponding health feature map. The aging feature extraction module calls the first health impact parameter based on the health feature map of the same category, extracts the first core feature and the second core feature of the first health impact parameter through the health parameter analysis model, and optimizes the first weight of the health feature map. The abnormal feature extraction module calls the second health impact parameter based on the health feature map of the same category, extracts the third and fourth core features of the second health impact parameter through the health parameter analysis model, and optimizes the second weight of the health feature map; The health status assessment module uses the health characteristic map to call the corresponding first health impact parameter and second health impact parameter for assessment. Based on the health feature map of the same category, the first health impact parameter is invoked. The first core feature and second core feature of the first health impact parameter are extracted through a health parameter analysis model, and the first weight of the health feature map is optimized, including: The first core feature and the second core feature are obtained based on the performance degradation analysis model of the health parameter analysis model; the performance degradation analysis model includes an influencing parameter preprocessing unit, an influencing parameter feature parsing unit, a first weight optimization unit, and a first weight storage unit. The influence parameter preprocessing unit performs data segmentation and data normalization on the first health influence parameter to generate input data; the influence parameter feature parsing unit predicts the next segment time data through a 1-layer autoencoder and a 5-layer GRU, and identifies the first core feature of the input data and the second core feature of the next segment time data through a 3-layer deep convolutional layer, a 1-layer dilated convolutional layer and a 4-layer GRU. The first weight optimization unit optimizes the association weights of the health feature map based on the first core feature and the second core feature; the first weight storage unit stores the association weights of the health feature map; The process involves calling the second health impact parameter based on the health feature map of the same category, extracting the third and fourth core features of the second health impact parameter through the health parameter analysis model, and optimizing the second weight of the health feature map, including: The third and fourth core features are obtained based on the abnormal diagnosis analysis model of the health parameter analysis model; the abnormal diagnosis analysis model includes an abnormal parameter preprocessing unit, an abnormal parameter feature extraction unit, a second weight optimization unit, and a second weight storage unit; The abnormal parameter preprocessing unit performs data segmentation and data normalization on the second health impact parameter to generate data to be identified; the abnormal parameter feature extraction unit extracts the third core feature of the data to be identified based on a 1-layer attention mechanism, a 1-layer temporal convolutional layer and a 4-layer bidirectional LSTM unit, and extracts the fourth core feature of the data to be identified through a 2-layer residual convolutional layer, a 1-layer dilated convolutional layer and a 4-layer LSTM. The second weight optimization unit optimizes the correlation weights of the second health impact parameter based on the third core feature and the fourth core feature; the second weight storage unit stores the optimized correlation weights. The health status assessment module calls the corresponding first health impact parameter and second health impact parameter based on the health feature map during the assessment. The method includes: extracting the first health impact parameter and second health impact parameter from the current running data, matching the optimized first weight and second weight in the health feature map respectively, and calculating the comprehensive health index by weighted summation as the assessment basis.

2. The intelligent assessment system for monitoring the health of reverse osmosis (RO) membranes according to claim 1, characterized in that, Classifying the historical operating data of multiple sets of reverse osmosis (RO) membranes within a period S includes: dividing the historical operating data of all the reverse osmosis (RO) membranes with maintenance time S as the dividing point, and classifying the data according to the total number of maintenance times S as the maintenance frequency.

3. The intelligent assessment system for monitoring the health of reverse osmosis (RO) membranes according to claim 1, characterized in that, Comparing the health status trend and calling the corresponding health feature map includes: The health status trend of the partitioned historical operating data is extracted by the moving average algorithm, and the mean of the health status trend of the same maintenance frequency is calculated as the standard health trend. Calculate the deviation of multiple health status trends with the same maintenance frequency from the standard health trend, classify multiple trend categories based on the deviation, label them, and call the corresponding category's health feature map; The health status trend of the same reverse osmosis (RO) membrane at the m-th maintenance frequency and the (m+1)-th maintenance frequency is compared according to the Euclidean distance algorithm. If the comparison result is less than the critical value of the health status trend, the health feature map of the RO membrane corresponding to the (m+1)-th maintenance frequency is called.

4. The intelligent assessment system for monitoring the health of reverse osmosis (RO) membranes according to claim 1, characterized in that, The first weight optimization unit optimizes the association weights of the health feature map based on the first core feature and the second core feature by: linearly adjusting the initial association weights of the health feature map based on the feature importance output by the performance decay analysis model, combined with the stability of the first core feature and the predictability of the second core feature.

5. The intelligent assessment system for monitoring the health of reverse osmosis (RO) membranes according to claim 4, characterized in that, The calculation of feature importance includes: grouping the input core features using a hierarchical clustering algorithm, calculating the similarity between features using a dynamic time warping algorithm, and determining the importance level label of each group of features based on the feature similarity distribution and the dispersion within the group.

6. The intelligent assessment system for monitoring the health of reverse osmosis (RO) membranes according to claim 1, characterized in that, The second weight optimization unit optimizes the association weight of the second health impact parameter based on the third core feature and the fourth core feature by: exponentially adjusting the initial association weight of the second health impact parameter according to the feature anomaly degree output by the anomaly diagnosis analysis model, combined with the significance of the third core feature and the persistence of the fourth core feature.

7. The intelligent assessment system for monitoring the health of reverse osmosis (RO) membranes according to claim 2, characterized in that, When classifying data based on the total number of maintenance times S as the maintenance frequency, if the maintenance frequency is less than the set frequency threshold, it is classified as the initial maintenance category; if the maintenance frequency is greater than or equal to the set frequency threshold, it is classified as the stable maintenance category.

Images