Method for intelligently identifying fault types of roll-type reverse osmosis and nanofiltration membrane elements

Abstract

The invention discloses a method for intelligently identifying the fault type of a roll-type reverse osmosis and nanofiltration membrane element, which is combined with an artificial intelligence technology of machine learning, establishes a fault type identification model and is used for fault diagnosis of reverse osmosis and nanofiltration membranes of a water treatment membrane system. According to the model, the fault type can be quickly identified under the condition that destructive disintegration analysis is not carried out only through simple membrane element test data and field use information, and useful information is provided for technicians. The problems that in traditional fault diagnosis, disintegration analysis items are numerous, and the period is long are solved, and rapid troubleshooting can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of failure analysis of reverse osmosis and nanofiltration spiral wound membrane elements, and particularly to a method for intelligently identifying the failure types of spiral wound reverse osmosis and nanofiltration membrane elements using machine learning algorithms. Background Art

[0002] Semi-permeable membranes represented by reverse osmosis and nanofiltration membranes are widely used in the fields of surface water desalination and wastewater treatment and reuse. However, due to various reasons, performance failures of membrane elements often occur, manifested as a decrease in rejection rate, a decrease in water production, an increase in differential pressure, etc., directly affecting the operation efficiency of the system. Through the inspection and analysis of failed membrane elements, the types of failures can be identified and corresponding improvement measures can be taken.

[0003] Common types of reverse osmosis and nanofiltration membranes include: ① physical damages such as separation layer damage and sealing defects caused by factors such as foreign matter inflow, channel flushing, and backpressure problems; ② chemical damages caused by oxidizing substances such as chlorine compounds; ③ inorganic scale deposits caused by crystallization and precipitation of supersaturated salts in water; ④ biological contamination caused by microbial proliferation; ⑤ organic chemical contaminations such as oil and surfactants; ⑥ other failures, etc.

[0004] Since the commonly available membrane elements on the market adopt a spiral wound structure and are protected by a non-removable outer shell, in the failure detection of the prior art, in addition to the performance test of the membrane element, it is often necessary to perform destructive disassembly on the membrane element to directly observe and analyze the internal membrane surface or its attachments to judge the failure type and cause. The disassembly analysis items include performance tests at various positions on the membrane surface, dyeing experiments, infrared analysis, XPS analysis, SEM analysis, sampling and composition analysis of the attachments on the membrane surface, etc., with an average time-consuming of more than 20 working days. On the one hand, this is time-consuming and laborious and difficult to match the urgency of troubleshooting; on the other hand, it causes irreversible damage to the membrane element and it cannot be reused by cleaning and repairing methods. Summary of the Invention

[0005] Therefore, the present invention proposes a method for intelligently identifying the failure types of spiral wound reverse osmosis and nanofiltration membrane elements. By combining artificial intelligence technology of machine learning, an identification model of failure types is established. This model can quickly identify the existing failure types based on simple membrane element test data and on-site usage information without performing destructive disassembly analysis, providing useful information for technicians.

[0006] The specific technical solution of the present invention includes the following contents:

[0007] A method for intelligently identifying the failure types of spiral wound reverse osmosis and nanofiltration membrane elements, comprising the following steps:

[0008] Step 1: Determine the input feature variables and output target labels for fault type identification;

[0009] Step 2: Model training;

[0010] Step 3: Use the above-trained model to identify the fault type;

[0011] In the above-mentioned Step 1, it includes:

[0012] Step 101: Obtain the evaluation data during the fault detection of different faulty membrane elements;

[0013] Step 102: Through feature correlation analysis, the feature variables are selected as the performance test data of the faulty membrane elements, including: salt passage rate, water production, weight, and the test data of the central tube probe; and the usage information, including: the filling position of the original system, usage time, raw water type, membrane type, and the performance of the filling system;

[0014] Step 103: The target labels include: component damage, separation layer breakage, chemical damage, inorganic scale formation, metal oxide pollution, biological pollution, non-biological organic pollution, no abnormality, or other unknown faults.

[0015] The salt passage rate, water production, and weight data will be processed by comparing with the factory value or the manual nominal value to represent the change trend. Different types of fault causes may lead to different change trends;

[0016] The central tube probe test data refers to the conductivity data of the produced water at different spatial positions inside the central water production pipe of the membrane element during the performance test of the membrane element, which can be non-destructively detected using a diversion tube. Dimensionality reduction processing is achieved by extracting variance, difference, skewness, etc. through statistical methods for feature optimization.

[0017] Preferably, in the above-mentioned Step 2, it includes:

[0018] Step 201: Establish a multi-label classification model;

[0019] Step 202: Use the feature variables as input parameters and the labels as output targets for model training and development of machine learning, and optimize the parameters;

[0020] Step 203: Normalize the numerical feature data in the feature variables to [0, 1]. The formula is: X n =(X - X min ) / (X max - X min ), where X n is the normalized value, X is the original sample data, X min and X maxThey are the minimum and maximum values of various types of test data for the feature variables respectively; perform one-hot encoding on the categorical feature data in the feature variables;

[0021] The numerical features refer to performance data of tests, usage time, etc.; the categorical features refer to information such as the type of raw water and the type of membrane element. For example, when the membrane type is divided into two categories, nanofiltration membrane and reverse osmosis membrane, the encoding is [1,0] and [0,1].

[0022] Step 204: Randomly divide the dataset into two categories, a training set and a test set. Among them, 70% of the data is used for training to find the best model, and the remaining 30% of the data does not participate in the training process and is used to test the effectiveness of the model;

[0023] Step 205: Use the method of classifier chain: Train a binary classifier for each label, and at the same time construct a classifier chain to capture the dependencies between the labels; each classifier is responsible for one label in the chain and takes the prediction results of the previous classifier as additional input features;

[0024] In this process, since there is a high probability that each faulty membrane element has several failure reasons, a multi-label classification model is established for model development. A binary classifier is established for each failure label, and binary classification data indicating the presence / absence of the label is output. Then, the dependencies between the labels are captured through the classifier chain to improve the generalization of the model.

[0025] Step 206: Use algorithms such as random forest algorithm, XGBoots algorithm or neural network algorithm to train the binary classifier for each label, and use grid search and random search to perform cross-validation on the best parameters of the algorithm; at the same time, use the method of random order to sort the classifier chain, and conduct multiple experiments to determine the best sorting;

[0026] Step 207: Use metrics such as Hamming Loss and Jaccard Similarity Coefficient to evaluate the performance of the model.

[0027] In this process, the Hamming Loss is the main evaluation metric. It represents the proportion of prediction errors among all labels, which is obtained by calculating the number of inconsistencies between the predicted value and the true value of each label and then dividing by the total number of labels. In Step 206, preferentially select the algorithm parameters and classifier order that minimize the Hamming Loss.

[0028] The beneficial effects of the present invention are as follows: Based on only simple membrane element test data and on-site usage information, it is possible to identify the possible types of faults without performing destructive disassembly analysis, providing useful information for technicians. It can also be used for preliminary diagnosis before disassembly analysis, providing a favorable reference for technicians to select targeted detection methods, greatly improving the efficiency of fault diagnosis for spiral wound reverse osmosis and nanofiltration membrane elements. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a schematic diagram of the fault type identification model of the present invention

[0030] Figure 2 is the macro-average confusion matrix diagram of the multi-label classification model in the embodiment of the present invention

[0031] Figure 3 is the evaluation diagram of the accuracy, precision, recall rate, and F1 score of each label of the multi-label classification model in the embodiment of the present invention

[0032] Figure 4 is the evaluation result of the multi-label classification model in the embodiment of the present invention DETAILED DESCRIPTION OF THE INVENTION

[0033] The following further describes the present invention in conjunction with embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The invention can also be implemented or applied through other different specific examples, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0034] Embodiment 1

[0035] Taking the model constructed based on the annual test data of a certain nanofiltration and reverse osmosis membrane detection center as an example for illustration.

[0036] (1) Determine the input feature variables and output target labels for fault type identification

[0037] Through feature correlation analysis, the input feature variables are selected as the performance test data of the faulty membrane element, including: salt rejection rate change, water production change, weight change, and central tube probe test data; and usage information, including: filling position of the original system, usage time, raw water type, and membrane type;

[0038] Data such as salt rejection rate, water production, and weight are processed by comparing with the factory value or the manual nominal value to represent the change trend. Different types of fault causes may lead to different change trends;

[0039] The central tube probe test data refers to the conductivity data of the produced water inside the central product water pipe of the membrane element at different spatial positions during the performance test of the membrane element. By using statistical methods, the central tube variance, difference, skewness, etc. are extracted to capture local anomalies of membrane internal pollution or damage, achieve dimensionality reduction processing, and optimize features.

[0040] The output classification labels of the model are common membrane element fault types, including: component damage, separation layer breakage, chemical damage, inorganic scale formation, metal oxide pollution, biological pollution, non-biological organic pollution, no anomaly, or other unknown faults. In the dataset, the correctness and rationality of all classification label results have been verified by professional technicians through disassembly analysis.

[0041] (2) Model training

[0042] Use the dataset to establish a multi-label classification model, perform model training and development of machine learning, and optimize parameters. The specific steps are as follows:

[0043] ① Normalize the numerical feature data in the input feature variables to [0,1]. The formula is: X n =(X - X min ) / (X max - X min ), where X n is the normalized value, X is the original sample data, and X min and X max are the minimum and maximum values of each type of test data in the dataset respectively; perform one-hot encoding on the categorical feature data in the input feature variables;

[0044] Numerical features refer to performance data obtained from tests, etc., including salt rejection rate, water production, weight, data of the central tube probe, etc.; categorical features refer to the types of raw water, types of membrane elements, etc. For example, membrane element types are divided into nanofiltration membranes, freshwater reverse osmosis membranes, and seawater reverse osmosis membranes.

[0045] ② Randomly divide the dataset into two categories, the training set and the test set, in a ratio of 7:3. Among them, 70% of the data is used for training to find the best model, and the remaining 30% of the data does not participate in the training process and is used to test the effectiveness of the model;

[0046] ③ Use the method of classifier chain: train a binary classifier for each label, and at the same time construct a classifier chain to capture the dependence relationship between labels. Each classifier is responsible for one label in the chain and uses the prediction results of the previous classifier as additional input features;

[0047] ④ Use the random forest algorithm to train the binary classifiers for each label. Use grid search and random search to perform cross-validation on the tree depth (max_depth) and the number of decision trees (n_estimators) of the random forest. At the same time, use the method of random order to sort the classifier chain and conduct 50 experiments to determine the best sorting.

[0048] ⑤ Use metrics such as Hamming loss and Jaccard similarity coefficient to evaluate the performance of the model. Prioritize selecting the model parameters and conditions with the smallest Hamming loss for model training.

[0049] (3) Use the developed model to identify the fault types. Obtain the data of faulty membrane elements outside the 4 sets of training sets and test sets, use this model for prediction, and output the probabilities of each fault label in each case. As shown in Table 1.

[0050] Table 1. Example of model output results: Probabilities (%) of each type of fault label being true

[0051] Case 1 2 3 4 Component damage 10 6 44 25 Membrane separation layer damage 15 1 7 18 Backpressure problem 3 0 2 2 Chemical damage 30 22 49 37 Inorganic salt scaling 28 33 30 22 Heavy metal pollution 31 2 23 12 Polymeric silica 4 15 3 5 Biological contamination 42 4 66 59 Organic chemical contamination 18 53 2 3 No abnormality or other unknown fault 10 31 14 14

[0052] According to the output results, in Case 1, the possibilities of chemical damage, inorganic salt scaling, heavy metal pollution, and biological pollution are the greatest, and these can be prioritized for investigation. In Case 2, the probability of organic chemical pollution is much higher than other types, and it can be inferred as the main fault type. In order to compare the consistency between the model prediction results and manual diagnosis, at the same time, a disassembly analysis was carried out on the above 4 cases, and the fault diagnosis was carried out by professional technicians. The results are shown in Table 2.

[0053] Table 2. Diagnostic results of professional technicians

[0054] Case 1 2 3 4 Component damage Yes Membrane separation layer damage Backpressure problem Chemical damage Yes Yes Inorganic salt scaling Yes Yes Yes Heavy metal pollution Yes Polymeric silica Biological contamination Yes Yes Yes Organic chemical contamination Yes No abnormality or other unknown fault

[0055] Comparing the results in Table 1 and Table 2, for the faults with high predicted probabilities by the model in each case, the vast majority have been actually confirmed.

[0056] An identification model for the fault types of spiral wound reverse osmosis and nanofiltration membrane elements was established by combining artificial intelligence technology of machine learning. This model can identify and predict the causes of faults based on simple membrane element test data and on-site use information without performing destructive disassembly analysis, providing useful information for technicians.

Claims

1. A method for intelligently identifying the failure types of spiral wound reverse osmosis and nanofiltration membrane elements, characterized in that, Including the following steps: Step 1: Determine the input feature variables and output target labels for fault type recognition; Step 2: Model training; Step 3: Use the above-trained model to identify the fault type; The Step 1 includes: Step 101: Obtain the evaluation data during the fault detection of different faulty membrane elements; Step 102: Through feature correlation analysis, the feature variables are selected as the performance test data of the faulty membrane elements, including: salt passage rate, water production, weight, and central tube probe test data; and the usage information, including: filling position of the original system, usage time, raw water type, membrane type, and performance performance of the operating system; Step 103: The target labels include: component damage, separation layer breakage, chemical damage, inorganic fouling, metal oxide pollution, biological pollution, non-biological organic pollution, no abnormality or other unknown faults.

2. The method for intelligently identifying the failure types of spiral wound reverse osmosis and nanofiltration membrane elements according to claim 1, wherein The Step 2 includes: Step 201: Establish a multi-label classification model; Step 202: Use the feature variables as input parameters and the labels as output targets to perform model training and development of machine learning and optimize the parameters; Step 203: Normalize the numerical feature data in the feature variables to [0, 1]. The formula is: X n =(X - X min ) / (X max - X min ), where X n is the normalized value, X is the original sample data, and X min and X max are the minimum and maximum values of each type of test data of the feature variable respectively; perform one-hot encoding on the categorical feature data in the feature variables; Step 204: Randomly divide the data set into two categories: training set and test set, where 70% of the data is used for training to find the best model, and the remaining 30% of the data does not participate in the training process and is used to test the effectiveness of the model; Step 205: Use the method of classifier chain: Train a binary classifier for each label, and at the same time construct a classifier chain to capture the dependencies between the labels; Each classifier is responsible for one label in the chain and uses the prediction results of the previous classifier as additional input features; Step 206: Use the random forest algorithm, XGBoots algorithm or neural network algorithm to train the binary classifier for each label, and at the same time use grid search and random search to determine the best training parameters; At the same time, use the method of random order to sort the classifier chain and conduct multiple experiments to determine the best sorting; Step 207: Use the Hamming loss and Jaccard similarity coefficient indicators to evaluate the performance of the model.

Images