Membrane pollution dynamic prediction and early warning method based on real-time monitoring

Abstract

In order to solve the problems of frequent pollution events, relatively high energy consumption and the like in the technical field of membranes in the water treatment process, the invention discloses a real-time monitoring-based membrane pollution dynamic prediction and early warning method, which comprises the following steps of: acquiring various parameter data influencing membrane pollution in real time; inputting the obtained parameter data into a pre-trained LSTM neural network model; the membrane pollution degree changing along with time is output, and real-time membrane pollution monitoring is completed; the membrane pollution degree prediction result and the early warning system are combined, automatic early warning of membrane pollution is achieved, the hysteresis quality in the membrane pollution treatment process is overcome, and the treatment capacity of the membrane pollution problem is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of water treatment, and particularly relates to a method for dynamically predicting and warning membrane fouling based on real-time monitoring. Background Art

[0002] Due to its advantages such as low energy consumption and environmental friendliness, membrane technology is playing an increasingly important role in the fields of water treatment, membrane heat pumps, fuel cells, passive cooling, etc. However, the performance of the separation membrane directly affects its energy conversion efficiency and environmental friendliness. Especially the problem of membrane fouling seriously restricts the improvement of its efficiency.

[0003] Membrane fouling refers to the continuous accumulation of various pollutants on the membrane surface or in the membrane pores during the membrane separation process, resulting in a decrease in membrane flux and separation performance. This phenomenon not only increases the operation and maintenance costs of the system, but also may affect the quality of the effluent. Therefore, membrane fouling is a bottleneck problem restricting the further application and development of membrane-based water treatment technology. At present, the main methods to solve membrane fouling on the market are to add drugs, clean or replace the membrane module. These methods have obvious hysteresis, and repeated cleaning and replacement will increase energy consumption.

[0004] Therefore, in order to reduce the occurrence of membrane fouling, a general technical solution is needed, which can predict the degree of membrane fouling and give timely warnings, improve the efficiency of manual cleaning or replacement of the membrane module, so as to achieve the purpose of reducing the occurrence of membrane fouling. Summary of the Invention

[0005] According to one aspect of the present disclosure, the present invention aims to provide a method for dynamically predicting and warning membrane fouling based on real-time monitoring. This method takes the real-time operating parameters as input, outputs the prediction result of the membrane fouling degree, and gives timely warnings according to the prediction result, so as to take measures in advance to avoid the shutdown and performance loss of the membrane system.

[0006] Based on the above purpose, the present invention adopts the following technical solutions:

[0007] A method for dynamically predicting and warning membrane fouling based on real-time monitoring, comprising the following steps:

[0008] Data collection: Collect the real-time operating parameters of the membrane module in the device, including collecting the influent water quality parameters such as chemical oxygen demand (COD), biochemical oxygen demand (BOD), etc.; the physical property parameters of the membrane module, including membrane pore size, hydrophilicity / hydrophobicity, porosity, etc.; and the operating environment condition parameters such as operating pressure, flow rate, etc.;

[0009] Data preprocessing: Preprocess the collected data, including missing value processing and outlier processing, to ensure the integrity and accuracy of the data;

[0010] In the method of the present invention, the missing value processing method includes: deleting missing values, filling in missing values, interpolation methods, etc.;

[0011] In the method of the present invention, the outlier detection and processing method includes: box plot analysis method, scatter plot analysis method, Z-score method, binning processing method, etc.;

[0012] The present invention realizes the prediction of the membrane fouling degree by inputting the preprocessed parameters into the currently optimal long short-term memory (LSTM) neural network model. The LSTM neural network model is based on a training data set composed of historical data including influent water quality parameters (such as chemical oxygen demand (COD), biochemical oxygen demand (BOD), etc.), membrane property parameters such as membrane pore size, and environmental condition parameters such as operating pressure and flow rate, and is trained and optimized in combination with the LSTM neural network algorithm, thereby establishing the prediction model.

[0013] In the method of the present invention, the membrane fouling degree includes the fouling layer thickness, fouling layer density, fouling layer coverage rate, etc. that change with time;

[0014] In the method of the present invention, the method for establishing the currently optimal LSTM neural network model specifically includes:

[0015] Obtain data training samples, where the training samples include historical data of membrane property parameters, operating environment condition parameters, and membrane fouling degree, so that the model has the performance of multiple inputs and multiple outputs. Specifically, the inputs include membrane property parameters and operating environment condition parameters; the outputs are the membrane fouling degree, specifically including parameters such as the fouling layer thickness, fouling layer density, and fouling layer coverage rate that change with time;

[0016] Divide the data training samples into a training set and a test set with a ratio of 8:2;

[0017] Construct an LSTM neural network model, and the LSTM neural network model is a multiple-input multiple-output model;

[0018] In the method of the present invention, the LSTM neural network model has a three-layer network structure, including an input layer, a hidden layer, and an output layer. First, input the training sample data into the LSTM neural network model, and verify the prediction error of the LSTM neural network model by comparing the predicted result of the membrane fouling degree obtained based on the training sample data with the actual result in the training sample data. Subsequently, input the test set data into the LSTM neural network model for backpropagation training, so as to adjust and optimize the parameters of the LSTM neural network model and further improve the prediction accuracy.

[0019] In the method of the present invention, the activation function of the LSTM neural network model selects tanh as the activation function, the model adopts the Adam optimization algorithm as the optimization strategy, and a Dropout layer is added to prevent overfitting.

[0020] The formula of the tanh activation function is as follows:

[0021] Among them, the calculation formula of the Adam optimization algorithm is:

[0022] Calculate the gradient:

[0023] Among them, g t is the gradient of the loss function L with respect to the parameter θ;

[0024] Update the first-order momentum: m t = β 1 m t-1 +(1 - β 1 )g t

[0025] Update the second-order momentum:

[0026] Among them, β 1 is the first-order momentum decay rate, and β 2 is the second-order momentum decay rate;

[0027] Bias correction:

[0028] Among them, because m t and v t are both initialized to 0 in the initial stage, so bias correction is required to avoid bias;

[0029] Update the parameters through the corrected momentum:

[0030] Among them, α is the learning rate of the model, and ∈ is a very small constant to prevent division by zero error.

[0031] The number of neurons in the input layer of the established LSTM neural network structure is the membrane property parameter and the operating environment condition parameter; the number of neurons in the hidden layer of this LSTM neural network structure is determined according to the number of training iterations; the number of neurons in the output layer of this LSTM neural network structure is the degree of membrane fouling.

[0032] In the method of the present invention, the hidden layer of the LSTM neural network structure includes a forget gate, an input gate, an output gate, and a state update unit. The forget gate is used to control how much of the input information at the current time needs to be deleted from the memory unit; the input gate is used to control how much of the new information input at the current time is added to the cell state; and the output gate determines the output value of the hidden state at the current time.

[0033] Among them, the calculation formulas for the forget gate, input gate, and output gate of the LSTM neural network model are as follows:

[0034] The calculation formula for the forget gate is: f t = σ(W f · [h t-1 , x t + b f )

[0035] Among them, f t is the output of the forget gate, σ is the Sigmoid function, W f is the weight matrix, b f is the bias term, h t-1 is the hidden state of the previous time step, and x t is the input of the current time step;

[0036] The input gate mainly includes two parts: the input gate and the candidate memory. The calculation formula for the input gate is: i t = σ(W i · [h t-1 , x t + b i ),

[0037] Among them, i t is the output of the input gate, is the candidate memory of the current input;

[0038] Update the memory unit:

[0039] Among them, C t is the state of the new memory unit, that is, the weighted sum of the state C t-1 of the previous time step and the candidate memory of the current time step;

[0040] The calculation formula for the output gate is: o t = σ(W o · [h t-1 , x t + b o ), h t = o t · tanh(C t )

[0041] Among them, o t is the output of the output gate, and the hidden state h t is the output at the current time step.

[0042] During the training process of the model, the mean squared error (MSE) is used as the target loss function, and the Adam algorithm is combined as the optimization strategy to perform gradient descent. Through continuous iteration, backpropagation is performed to minimize the loss function, obtain qualified weights and biases, and finally obtain a qualified model.

[0043] Among them, the mean squared error MSE adopts the following calculation formula:

[0044]

[0045] Among them, n is the number of samples, and y i is the actual value of the i-th sample, is the predicted value of the i-th sample.

[0046] In the evaluation of the model, the test set data is used to evaluate the performance of the model, and the coefficient of determination R 2 and the root mean square error (RMSE) are selected as indicators to judge the model performance.

[0047] According to one aspect disclosed by the present invention, a membrane fouling early warning method based on real-time monitoring is provided. The method mainly includes the following steps:

[0048] Step 1) Obtain key parameter data. By analyzing historical membrane operation data, identify the critical points of membrane fouling and key parameters. The membrane includes data such as influent water quality (COD, BOD, etc.), membrane pore size, hydrophilicity / hydrophobicity, porosity, etc., and set the threshold range in the software according to the historical data of membrane fouling;

[0049] Step 2) Data analysis. Input the real-time parameter data into the constructed prediction model; based on the current optimal LSTM neural network model, predict the degree of membrane fouling, and analyze and judge the prediction results;

[0050] Step 3) Automatic early warning. If the prediction result of the degree of membrane fouling matches the set threshold, an early warning message will be displayed on the operation interface.

[0051] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are:

[0052] This solution reduces the occurrence of membrane fouling by constructing an LSTM neural network model: on the one hand, this LSTM neural network model can monitor the situation of the membrane module in real time and realize the dynamic prediction of the information on the degree of membrane fouling; on the other hand, based on the prediction results of this LSTM neural network model, timely early warning is realized, and warning information is prompted, which can effectively overcome the lag of methods such as dosing, cleaning or replacing the membrane module. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the internal structure diagram of LSTM for a method of dynamic prediction and early warning of membrane fouling based on real-time monitoring provided by the present invention;

[0054] Figure 2 It is the neural network structure diagram of a method of dynamic prediction and early warning of membrane fouling based on real-time monitoring provided by the present invention;

[0055] Figure 3 It is the working flow chart of a method of dynamic prediction and early warning of membrane fouling based on real-time monitoring provided by the present invention;

[0056] Figure 4 It is the neural network flow chart of a method of dynamic prediction of membrane fouling based on real-time monitoring provided by the embodiment of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] Embodiment

[0058] The purpose of the present invention is to provide a method for dynamic prediction and early warning of membrane fouling based on real-time monitoring, which can monitor the situation of the membrane surface in real time, predict the degree of fouling, and give accurate early warning, so as to reduce the losses caused by the occurrence of membrane fouling.

[0059] In the embodiment of the present invention, optical coherence tomography (OCT) technology is used to realize two-dimensional and three-dimensional real-time in-situ fouling monitoring and obtain the image of the pollution layer on the membrane surface. During the implementation process, since the two-dimensional image can only determine the thickness of the fouling layer when the fouling is serious, therefore, in this embodiment, three-dimensional images are selected to comprehensively monitor the formation of the fouling layer and realize more accurate real-time monitoring. At the same time, the operating parameters are monitored and recorded in real time through a sensor array.

[0060] In the embodiment of the present invention, data preprocessing adopts filtering technology. A low-intensity filter is used to set a low tolerance to screen out noise and the fouling layer, and only the membrane layer is retained; a high-intensity filter is used to set a relatively high tolerance to screen out noise, and only the fouling voxels and the membrane layer are retained. Fouling voxel screening: The voxels at positions where the intensity value is lower than the top of the membrane layer are set to zero to screen out the fouling voxels. Eliminate the irrelevant details of the image data, reduce the image dimension, and obtain the sequence data of the degree of pollution.

[0061] In an embodiment of the present invention, the LSTM (Long Short-Term Memory network) aims to solve the problems of vanishing gradients and exploding gradients faced by traditional Recurrent Neural Networks (RNNs) when processing long sequence data. Although RNNs are widely used in sequence data processing, their performance is poor when dealing with long sequences, mainly due to the phenomena of vanishing gradients and exploding gradients. By introducing a gating mechanism, LSTM effectively overcomes these deficiencies and can maintain long-term dependencies, making it possible to capture long-term dependencies in time series.

[0062] A specific implementation method of a dynamic prediction and early warning method for membrane fouling based on real-time detection

[0063] The present invention obtains a dynamic prediction method for membrane fouling based on real-time monitoring: The method includes: data collection, real-time monitoring and recording of the parameters of the membrane module operation through a sensor array, including membrane property parameters such as influent water quality (COD, BOD, etc.), membrane pore size, hydrophilicity / hydrophobicity, porosity, etc., and operation environment condition parameters such as operating pressure, flow rate, etc.; using Optical Coherence Tomography (OCT) technology for in-situ fouling monitoring, obtaining a three-dimensional image of the fouling layer on the membrane surface, and processing the three-dimensional image to obtain the degree of fouling, including data such as the thickness of the fouling layer, the density of the fouling layer, and the coverage rate of the fouling layer that change over time; at the same time, the degree of fouling can be evaluated in real time.

[0064] In an embodiment of the present invention, the images collected by using Optical Coherence Tomography (OCT) technology are enhanced through digital filtering technology, including contrast adjustment, noise filtering, and image sharpening, to eliminate irrelevant details, highlight the detailed features of the fouling layer, and reduce the image dimension. Each image has a corresponding time series parameter set.

[0065] In an embodiment of the present invention, a normalization method is used to clean the data, and the data is scaled to a certain range ([0, 1]) to improve the stability and convergence speed of training; at the same time, for data with missing items, the method of deleting missing values is adopted for processing.

[0066] In an embodiment of the present invention, the LSTM neural network model can achieve multiple inputs and multiple outputs. Taking the degree of membrane fouling such as the thickness of the fouling layer, the density of the fouling layer, and the coverage rate of the fouling layer as the output targets, and taking the membrane property parameters such as the influent water quality (COD, BOD, etc.), membrane pore size, etc., and the environmental parameters such as operating pressure, flow rate, etc. collected as input variables, specifically including:

[0067] As an example, the processed data set is randomly divided into a training set and a test set according to a ratio of 8:2.

[0068] As an example, based on the Python language, an LSTM neural network model is built by importing libraries such as NumPy, Pandas, scikit-learn, and Keras, and the tanh function is selected as the activation function.

[0069] The formula for the tanh activation function is as follows:

[0070] Based on the above model, the training set data is input into the LSTM neural network model. The gradient descent algorithm of the model is the Adam algorithm, and a Dropout layer is added to prevent overfitting.

[0071] Among them, the calculation formula of the Adam optimization algorithm is:

[0072] Calculate the gradient:

[0073] Among them, g t is the gradient of the loss function L with respect to the parameter θ;

[0074] Update the first-order momentum: m t = β 1 m t-1 +(1 - β 1 )g t

[0075] Update the second-order momentum:

[0076] Among them, β 1 is the first-order momentum decay rate, and β 2 is the second-order momentum decay rate;

[0077] Bias correction:

[0078] Among them, because m t and v t are both initialized to 0 in the initial stage, so bias correction is needed to avoid bias;

[0079] Update the parameters using the corrected momentum:

[0080] Among them, α is the learning rate of the model, and ∈ is a very small constant used to prevent division by zero errors.

[0081] As an example, the setting range of learning_rate is 0.01 to 0.2, that is, the setting range of the learning rate α of the model is 0.01 to 0.2. After continuous iteration and backpropagation, a model that meets the conditions is finally obtained and saved.

[0082] In this embodiment, the data acquisition process lasts for 24 hours. Specifically, the total time of the experiment is 24 hours, during which images are acquired once per hour in the first 4 hours; from the 4th hour to the 24th hour, images are acquired once every 4 hours. Other types of data also have time attributes, and their acquisition intervals are the same as those of image acquisition.

[0083] Based on the above model, according to the weights and biases, new data sets are constructed with membrane property parameters such as influent water quality (such as COD, BOD, etc.), membrane pore size, hydrophilicity / hydrophobicity, porosity, and operation environment condition parameters such as operation pressure and flow rate. Subsequently, the data sets containing these variables are input into the LSTM neural network model for multi-objective output prediction.

[0084] As an example, based on the LSTM neural network model for real-time monitoring, multiple relevant variables are input to obtain multi-objective outputs of the membrane fouling degree, and the coefficient of determination R 2 and the root mean square error (RMSE) are used as indicators to judge the performance of the model, improve the expression ability of the model, and complete the prediction of the membrane fouling degree.

[0085] Based on the above model, the present invention constructs a membrane fouling prediction and early warning system, which consists of an LSTM neural network, sensors, and monitoring and processing devices. The system stores the collected historical membrane operation data in a database and sets a membrane fouling threshold range. By analyzing the membrane fouling layer information in real time and generating prediction results, the system compares the predicted values of the LSTM model with the preset thresholds. When the membrane fouling degree exceeds the threshold, the system triggers an early warning mechanism and displays a warning message on the operation interface to prompt the operator to take necessary measures in a timely manner.

[0086] Through this early warning method, the membrane fouling condition can be effectively monitored, measures can be taken in a timely manner, and the damage and operation costs of the membrane can be reduced.

Claims

1. A method for dynamic prediction and early warning of membrane fouling based on real-time monitoring, characterized in that: The following steps are involved: Data collection, real-time operating parameters of membrane components in the collection equipment; Data preprocessing: preprocess the collected data to ensure the integrity and accuracy of the data; The parameters affecting the pretreatment membrane fouling are input into the current optimal LSTM neural network model to predict the degree of membrane fouling. The automatic early warning system prompts early warning information based on the model's prediction results and user-set thresholds.

2. A membrane pollution dynamic prediction and early warning method based on real-time monitoring as claimed in claim 1, characterized in that: The construction of the membrane fouling degree prediction model includes: Constructing an LSTM neural network model, wherein the LSTM neural network model is a multi-input multi-output model; Divide the dataset samples into training set and test set; The training set data information is input into the LSTM neural network model to verify the error of the LSTM neural network model, and the test set data information is input into the LSTM neural network model to adjust and optimize the parameters of the LSTM neural network model.

3. A membrane pollution dynamic prediction and early warning method based on real-time monitoring as claimed in claim 2, characterized in that: The multiple inputs refer to the input membrane property parameters and operating condition parameters, including: inlet water quality (COD, BOD, etc.), membrane pore size, hydrophilicity, porosity and other membrane property parameters, operating pressure, flow rate and other operating environment condition parameters; the multiple outputs refer to the degree of membrane contamination, including the thickness of the contamination layer, the density of the contamination layer, the coverage of the contamination layer, etc. that change with time.

4. A membrane pollution dynamic prediction and early warning method based on real-time monitoring as claimed in claim 2, characterized in that: The data set samples are divided into a training set and a test set in a ratio of 8:

2.

5. A membrane pollution dynamic prediction and early warning method based on real-time monitoring as claimed in claim 2, characterized in that: The LSTM neural network model has a three-layer network structure including an input layer, a hidden layer, and an output layer; The training sample data information is input into the LSTM neural network model, and the error of the LSTM neural network model is verified by comparing the predicted results of membrane contamination degree based on the training sample data with the actual results in the training sample data; then the test set data information is input into the LSTM neural network model for reverse training, so as to adjust and optimize the parameters of the LSTM neural network model.

6. A membrane fouling dynamic prediction and early warning method based on real-time monitoring as claimed in claim 5, characterized in that: The number of neurons in the input layer is the membrane property parameters and the operating environment condition parameters; the optimal number of neurons in the hidden layer is determined according to the number of training iterations; and the number of neurons in the output layer is the degree of membrane contamination.

7. A membrane pollution dynamic prediction and early warning method based on real-time monitoring according to claim 6, characterized in that: The hidden layer has a forget gate, an input gate, an output gate and a state update unit; the forget gate determines how much of the input information at the current moment needs to be deleted from the memory unit; the input gate determines how much of the new input information at the current moment is added to the cell state; the output gate determines the output value of the hidden state at the current moment.

8. A membrane pollution dynamic prediction and early warning method based on real-time monitoring according to claim 7, characterized in that: The neural network model adopts mean square error (MSE) as the target loss function and combines the Adam algorithm as an optimization strategy to perform gradient descent. Through continuous iteration, qualified weights and biases are obtained, and finally a qualified model is obtained.

9. A membrane fouling dynamic prediction and early warning method based on real-time monitoring as claimed in claims 1 to 8, characterized in that: The automatic early warning comprises the following steps: Obtain key parameter data, identify critical points and key parameters of membrane fouling through analysis of historical membrane operation, and set threshold ranges in the software based on historical data of membrane fouling; Data analysis: input the real-time parameter data into the constructed prediction model; analyze and judge the prediction results based on the prediction results of the membrane fouling layer information of the current optimal LSTM neural network model; Automatic warning: if the predicted result of the pollution layer information exceeds the threshold set by the user, a warning message will be prompted on the operation interface.

Images