MBR membrane permeability intelligent detection method based on self-organizing robust neural network

Abstract

The invention provides an intelligent MBR membrane permeability detection method based on a self-organizing robust neural network, and aims to solve the problem that the MBR membrane permeability is difficult to detect under the interference conditions of water quality and water quantity fluctuation and the like. For abnormal input data caused by external interference, a data immune mechanism is introduced to realize self-organization of a neural network structure, so that the model adapts to the change of the input data, the influence of noise and abnormal values on the model is reduced, model parameters are updated by utilizing a self-adaptive gradient descent algorithm, and the anti-interference capability of the model is effectively improved. Experimental results prove that the method can realize accurate prediction of the membrane permeability under the condition that abnormal input data exists, and relatively good stability and robustness are kept.

Description

Technical Field

[0001] Based on the high uncertainty of the membrane fouling process and the dynamic characteristics of the fuzzy neural network, the present invention designs an intelligent prediction method for membrane permeability using a fuzzy neural network based on self-organization, and realizes the accurate prediction of membrane permeability under external disturbances. Membrane permeability refers to the water flow rate passing through a unit membrane area per unit time, and is an important indicator for evaluating the degree of membrane fouling. The intelligent detection method for membrane permeability based on the self-organizing robust neural network can reduce the sensitivity of the network to uncertain inputs, maintain stable performance under external disturbances, accurately predict membrane permeability, and obtain more accurate prediction results, which belongs to the field of water science. Background Art

[0002] At present, water shortage and water pollution have become global environmental problems, seriously threatening human survival and development. With the acceleration of industrialization and urbanization, the demand for water resources continues to increase. However, many water bodies have become unusable due to pollution. Therefore, the development of efficient sewage treatment technology has become the key to solving this problem. As an advanced sewage treatment method, membrane bioreactor technology has gradually attracted widespread attention.

[0003] However, membrane bioreactor technology also faces some challenges in practical applications, the most prominent of which is membrane fouling. Membrane fouling refers to the accumulation of particulate matter, microorganisms and their metabolites in sewage on the membrane surface or in the pores, resulting in a decrease in membrane flux and a weakening of the separation effect. Membrane fouling will not only shorten the service life of the membrane and increase operating costs, but also affect the effluent quality and reduce the overall efficiency of sewage treatment. Therefore, it is particularly important to design corresponding methods to identify membrane fouling, which is of great practical significance.

[0004] The present invention designs a membrane fouling identification method based on a self-organizing fuzzy neural network, utilizes the dynamic characteristics of the fuzzy neural network, adjusts the network structure through a self-organizing strategy, makes the network adapt to noise and abnormal input, improves the robustness of the fuzzy neural network, reduces the influence of external interference on the prediction performance, and completes the safe and stable operation of the sewage treatment process. Summary of the invention

[0005] The present invention obtains a membrane pollution identification method based on a self-organizing robust neural network, which mainly uses a self-organizing strategy based on a data immune mechanism to adjust the network structure, so that the model can adapt to abnormal input data with a more compact structure, thereby accurately identifying membrane pollution and ensuring the safe and stable operation of the sewage treatment process;

[0006] The present invention adopts the following technical solutions and implementation steps:

[0007] 1. An intelligent detection method for MBR membrane permeability based on a self-organizing robust neural network, characterized in that an intelligent detection model based on a robust neural network is established, a data immune mechanism is designed to dynamically adjust the model structure, and an adaptive gradient descent algorithm is used to update the model parameters, thereby solving the problem that the MBR membrane permeability is difficult to detect under interference conditions such as water quality and water quantity fluctuations, and comprising the following steps:

[0008] (1) Establish an intelligent detection model of membrane permeability based on a robust neural network, as follows:

[0009] Membrane permeability can reflect the operating status of membrane bioreactor wastewater treatment. However, affected by external interference factors such as fluctuations in influent water quality and quantity, it is difficult for operators to effectively detect membrane permeability. Therefore, an intelligent detection model based on self-organizing robust neural network is established to realize dynamic detection of membrane permeability. The intelligent detection model based on self-organizing robust neural network is divided into four layers: input layer, radial basis function layer, normalization layer, and output layer. The model adopts the initial connection mode of 4-10-10-1, that is, the number of neurons in the input layer is 4, corresponding to the physical quantity They are respectively water production flow, water production pressure, single pool membrane scrubbing gas volume, and water production turbidity. The number of neurons in the radial basis function layer is 10, the number of neurons in the normalization layer is 10, the initial value of the connection weight between the radial basis function layer and the normalization layer is 1, the number of neurons in the output layer is 1, the corresponding physical quantity is membrane permeability, and the initial value of the connection weight between the normalization layer and the output layer is 1; the training data samples of the neural network intelligent detection model are set to N, and the test data samples are set to M, where N>M; each layer in the detection model based on the self-organizing robust neural network is expressed as:

[0010] Input layer: This layer consists of 4 neurons, and the output of each neuron is:

[0011] x i (t) = u i (t) (1)

[0012] Among them, x i (t) is the output of the i-th neuron in the input layer at time t, i = 1,...,4; u i (t) is the input of the i-th neuron in the input layer at time t, u1(t) represents the water production flow rate at time t, u2(t) represents the water production pressure at time t, u3(t) represents the single pool membrane scrubbing gas volume at time t, and u4(t) represents the water production turbidity at time t;

[0013] Radial basis function layer: This layer consists of 10 neurons, and the output of each neuron is:

[0014]

[0015] in, is the output value of the jth neuron in the radial basis function layer at time t, j = 1,...,10; m ij (t) is the center of the i-th membership function of the j-th neuron at time t, which takes a random value in the interval (0,1], σ ij (t) is the width of the i-th membership function of the j-th neuron at time t, and takes a random value in the interval (0,1);

[0016] Normalization layer: This layer consists of 10 neurons, and the output of each neuron is:

[0017]

[0018] Among them, v j (t) is the output value of the jth neuron in the normalized layer at time t, w vj (t) is the weighted connection between the radial basis function layer and the normalization layer at time t;

[0019] Output layer: This layer consists of 1 neuron and the output is:

[0020]

[0021] in, is the output of the neural network intelligent detection model at time t, w j (t) is the weighted connection between the jth normalized layer neuron and the output layer at time t;

[0022] (2) Design a data immune mechanism to dynamically adjust the model structure, as follows:

[0023] The loss function is defined as:

[0024]

[0025] in, Represents the error of the model at time t; its loss function boundary is:

[0026]

[0027] Among them, Z b (t) is the loss function boundary at time t, μ is used to define the loss function boundary, which requires that the loss function boundary at time t is greater than the loss function value at time t, and is randomly selected between (0, 0.15];

[0028] The dynamic adjustment of the detection model structure based on the self-organizing robust neural network is as follows:

[0029] ① Growth stage: If Z(t) is greater than Z b (t), and the j-th radial basis function layer neuron satisfies the following conditions:

[0030] w vj (t)≥max(w v (t-1)) (7)

[0031] Then add a new radial basis function neuron, where w v (t-1) = [w v1 (t-1),…,w vP (t-1)]; the parameters of this neuron can be expressed as:

[0032]

[0033] Among them, m new (t) and σ new (t) are the center and width of the neuron in the radial basis function layer, m j (t) is the center matrix of the jth radial basis function layer neuron at time t, x(t) is the input variable matrix at time t, σ j (t) is the width matrix of the neurons in the j-th radial basis function layer at time t; w vnew (t) and w new (t) are the weights of the neurons in the radial basis function layer, w v (t) is the weight matrix of all radial basis function layer neurons at time t, g(t) is the expected output at time t, is the predicted output at time t, x i (t) is the input of the i-th neuron at time t, m inew (t) is the center of the i-th membership function of the new neuron at time t, σ inew (t) is the width of the i-th membership function of the new neuron at time t;

[0034] ② Pruning stage: If Z b (t-1) is greater than Z(t), and the weights of the neurons in the h-th radial basis function layer satisfy the following conditions:

[0035]

[0036] Then the hth radial basis function layer neuron will be pruned; at the same time, the parameters of the hth and h'th radial basis function layer neurons are set to:

[0037]

[0038] Among them, the h'th radial basis function layer neuron is the neuron with the closest Euclidean distance to the hth radial basis function layer neuron, m' h (t) and m' h’ (t) are the centers of the hth and h'th radial basis function layer neurons after pruning the neurons, respectively.h’ (t) is the center of the h' radial basis function layer neurons before the pruned neuron; σ' h (t) and σ' h’ (t) are the widths of the hth and h'th radial basis function layer neurons after pruning neurons, σ h’ (t) is the width of the h'th radial basis function layer neuron before pruning the neuron; w' vh’ (t) and w' h’ (t) are the weights of the h'th radial basis function layer neurons and normalized layer neurons after pruning neurons, respectively, w vh’ (t) and w h’ (t) are the weights of the h'th radial basis function layer neurons and normalized layer neurons before pruning neurons, respectively, w vh (t) and w h (t) are the weights of the hth radial basis function layer neurons and normalized layer neurons before pruning neurons; w' vh (t) and w' h (t) are the weights of the neurons in the hth radial basis function layer before and after pruning the neurons; m ih (t) and m ih’ (t) are the centers of the hth and h'th radial basis function layer neurons before pruning the neurons, respectively, σ ih (t) and σ ih’ (t) are the widths of the hth and h'th radial basis function layer neurons before pruning the neurons;

[0039] (3) Update the model parameters using the adaptive gradient descent algorithm as follows:

[0040] The parameter update gradient of the detection model based on the self-organizing robust neural network is:

[0041]

[0042] (4) Detection of membrane permeability

[0043] The membrane permeability is detected using the trained robust neural network model, and the collected water production pressure, water production flow, water production turbidity and single-cell membrane scrubbing gas volume are used as the input variables of the model to obtain the membrane permeability prediction output value of the model.

[0044] By utilizing the self-organizing mechanism of the model structure and the parameter updating strategy, the sensitivity of the detection model to external interference is reduced, the robustness of the detection model is improved, the accurate detection of membrane permeability is achieved, the dynamic monitoring of the membrane bioreactor state is completed, and the safe and stable operation of the urban sewage treatment process is ensured. The creativity of the present invention is mainly reflected in:

[0045] (1) The present invention aims at the abnormal input data caused by external disturbances in the urban sewage treatment process, which will lead to the decline of prediction accuracy and effectiveness of fuzzy neural network. A parameter learning algorithm with a penalty evaluation mechanism is designed to update the network parameters and reduce the sensitivity of the model to abnormal input data.

[0046] (2) The present invention designs a network structure self-organizing strategy based on the data immunity mechanism. By dynamically adjusting the number of neurons and the network structure, the network can adapt to abnormal input data with a more compact structure, thereby enhancing the robustness of the model and ensuring that the model still has good prediction accuracy and performance under external disturbances. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is the membrane permeability prediction result diagram of the present invention

[0048] Figure 2 The membrane permeability prediction error diagram of the present invention is DETAILED DESCRIPTION

[0049] 1. An intelligent detection method for MBR membrane permeability based on a self-organizing robust neural network, characterized in that an intelligent detection model based on a robust neural network is established, a data immune mechanism is designed to dynamically adjust the model structure, and an adaptive gradient descent algorithm is used to update the model parameters, thereby solving the problem that the MBR membrane permeability is difficult to detect under interference conditions such as water quality and water quantity fluctuations, and comprising the following steps:

[0050] (1) Establish an intelligent detection model of membrane permeability based on a robust neural network, as follows:

[0051] Membrane permeability can reflect the operating status of membrane bioreactor wastewater treatment. However, affected by external interference factors such as fluctuations in influent water quality and quantity, it is difficult for operators to effectively detect membrane permeability. Therefore, an intelligent detection model based on self-organizing robust neural network is established to realize dynamic detection of membrane permeability. The intelligent detection model based on self-organizing robust neural network is divided into four layers: input layer, radial basis function layer, normalization layer, and output layer. The model adopts the initial connection mode of 4-10-10-1, that is, the number of neurons in the input layer is 4, corresponding to the physical quantity They are respectively water production flow, water production pressure, single pool membrane scrubbing gas volume, and water production turbidity. The number of neurons in the radial basis function layer is 10, the number of neurons in the normalization layer is 10, the initial value of the connection weight between the radial basis function layer and the normalization layer is 1, the number of neurons in the output layer is 1, the corresponding physical quantity is membrane permeability, and the initial value of the connection weight between the normalization layer and the output layer is 1; the training data samples of the neural network intelligent detection model are set to N, and the test data samples are set to M, where N>M; each layer in the detection model based on the self-organizing robust neural network is expressed as:

[0052] Input layer: This layer consists of 4 neurons, and the output of each neuron is:

[0053] x i (t) = u i (t) (1)

[0054] Among them, x i (t) is the output of the i-th neuron in the input layer at time t, i = 1,...,4; u i (t) is the output of the i-th neuron in the input layer at time t, u1(t) represents the water production flow rate at time t, u2(t) represents the water production pressure at time t, u3(t) represents the single pool membrane scrubbing gas volume at time t, and u4(t) represents the water production turbidity at time t;

[0055] Radial basis function layer: This layer consists of 10 neurons, and the output of each neuron is:

[0056]

[0057] in, is the output value of the jth neuron in the radial basis function layer at time t, j = 1,...,10; m ij (t) is the center of the i-th membership function of the j-th neuron at time t, which takes a random value in the interval (0,1], σ ij (t) is the width of the i-th membership function of the j-th neuron at time t, and takes a random value in the interval (0,1);

[0058] Normalization layer: This layer consists of 10 neurons, and the output of each neuron is:

[0059]

[0060] Among them, v j (t) is the output value of the jth neuron in the normalized layer at time t, w vj (t) is the weighted connection between the radial basis function layer and the normalization layer at time t;

[0061] Output layer: This layer consists of 1 neuron and the output is:

[0062]

[0063] in, is the output of the neural network intelligent detection model at time t, w j (t) is the weighted connection between the jth normalized layer neuron and the output layer at time t;

[0064] (2) Design a data immune mechanism to dynamically adjust the model structure, as follows:

[0065] The loss function is defined as:

[0066]

[0067] in, Represents the error of the model at time t; its loss function boundary is:

[0068]

[0069] Among them, Z b (t) is the loss function boundary at time t, μ is used to define the loss function boundary, which requires that the loss function boundary at time t is greater than the loss function value at time t, and is randomly selected between (0, 0.15];

[0070] The dynamic adjustment of the detection model structure based on the self-organizing robust neural network is as follows:

[0071] ① Growth stage: If Z(t) is greater than Z b (t), and the j-th radial basis function layer neuron satisfies the following conditions:

[0072] w vj (t)≥max(w v (t-1)) (7)

[0073] Then add a new radial basis function neuron, where w v (t-1) = [w v1 (t-1),…,w vP (t-1)]; the parameters of this neuron can be expressed as:

[0074]

[0075] Among them, m new (t) and σ new (t) are the center and width of the neuron in the radial basis function layer, m j (t) is the central parameter matrix of the jth radial basis function layer neuron at time t, x(t) is the input variable matrix at time t, σ j (t) is the width parameter matrix of the neurons in the j-th radial basis function layer at time t; w vnew (t) and w new (t) are the weights of the neurons in the radial basis function layer, w v (t) is the weight matrix of all radial basis function layer neurons at time t, g(t) is the expected output at time t, is the predicted output at time t, x i (t) is the input of the i-th neuron at time t, m inew (t) is the center of the i-th membership function of the new neuron at time t, σinew (t) is the width of the i-th membership function of the new neuron at time t;

[0076] ② Pruning stage: If Z b (t-1) is greater than Z(t), and the weights of the neurons in the h-th radial basis function layer satisfy the following conditions:

[0077]

[0078] Then the hth radial basis function layer neuron will be pruned; at the same time, the parameters of the hth and h'th radial basis function layer neurons are set to:

[0079]

[0080] Among them, the h'th radial basis function layer neuron is the neuron with the closest Euclidean distance to the hth radial basis function layer neuron, m' h (t) and m' h’ (t) are the centers of the hth and h'th radial basis function layer neurons after pruning the neurons, respectively. h’ (t) is the center of the h' radial basis function layer neurons before the pruned neuron; σ' h (t) and σ' h’ (t) are the widths of the hth and h'th radial basis function layer neurons after pruning neurons, σ h’ (t) is the width of the h'th radial basis function layer neuron before pruning the neuron; w' vh’ (t) and w' h’ (t) are the weights of the h'th radial basis function layer neurons and normalized layer neurons after pruning neurons, respectively, w vh’ (t) and w h’ (t) are the weights of the h'th radial basis function layer neurons and normalized layer neurons before pruning neurons, respectively, w vh (t) and w h (t) are the weights of the hth radial basis function layer neurons and normalized layer neurons before pruning neurons; w' vh (t) and w' h (t) are the weights of the neurons in the hth radial basis function layer before and after pruning the neurons; m ih (t) and m ih’ (t) are the centers of the hth and h'th radial basis function layer neurons before pruning the neurons, respectively, σ ih (t) and σ ih’ (t) are the widths of the hth and h'th radial basis function layer neurons before pruning the neurons;

[0081] (3) Update the model parameters using the adaptive gradient descent algorithm as follows:

[0082] The parameter update gradient of the detection model based on the self-organizing robust neural network is:

[0083]

[0084] (4) Detection of membrane permeability

[0085] The membrane permeability is detected using the trained robust neural network model, and the collected water production pressure, water production flow, water production turbidity and single-cell membrane scrubbing gas volume are used as the input variables of the model to obtain the membrane permeability prediction output value of the model.

[0086] By utilizing the self-organizing mechanism and parameter updating strategy of the above-mentioned model structure, the sensitivity of the detection model to external interference is reduced, the robustness of the detection model is improved, the accurate detection of membrane permeability is achieved, the dynamic monitoring of the membrane bioreactor status is completed, and the safe and stable operation of the urban sewage treatment process is ensured. Figure 1 Displays the predicted results of membrane permeability. X-axis: test sample, unit is piece, Y-axis: membrane permeability, unit is L / (m 2 *h), the solid line is the actual value of membrane permeability, and the dotted line is the predicted value of membrane permeability; the predicted error of membrane permeability is shown in Figure 2 , X-axis: test sample, unit is piece, Y-axis: membrane permeability prediction error, unit is L / (m 2 *h).

Claims

1. An intelligent detection method for MBR membrane permeability based on a self-organizing robust neural network, characterized in that The following steps are involved: (1) Establish an intelligent detection model of membrane permeability based on a robust neural network, as follows: An intelligent detection model based on a self-organizing robust neural network is established. The model is divided into four layers: input layer, radial basis function layer, normalization layer, and output layer. The model adopts an initial connection mode of 4-10-10-1, that is, the number of neurons in the input layer is 4, and the corresponding physical quantities are water production flow, water production pressure, single pool membrane scrubbing gas volume, and water production turbidity. The number of neurons in the radial basis function layer is 10, the number of neurons in the normalization layer is 10, the initial value of the connection weight between the radial basis function layer and the normalization layer is 1, the number of neurons in the output layer is 1, the corresponding physical quantity is membrane permeability, and the initial value of the connection weight between the normalization layer and the output layer is 1. The training data samples of the neural network intelligent detection model are set to N, and the test data samples are set to M, where N>M. Each layer in the detection model based on the self-organizing robust neural network is expressed as: Input layer: This layer consists of 4 neurons, and the output of each neuron is: x i (t)=u i (t) (1) Among them, x i (t) is the output of the i-th neuron in the input layer at time t, i = 1,...,4; u i (t) is the input of the i-th neuron in the input layer at time t, u1(t) represents the water production flow rate at time t, u2(t) represents the water production pressure at time t, u3(t) represents the single pool membrane scrubbing gas volume at time t, and u4(t) represents the water production turbidity at time t; Radial basis function layer: This layer consists of 10 neurons, and the output of each neuron is: in, is the output value of the jth neuron in the radial basis function layer at time t, j = 1,...,10; m ij (t) is the center of the i-th membership function of the j-th neuron at time t, which takes a random value in the interval (0,1], σ ij (t) is the width of the i-th membership function of the j-th neuron at time t, and takes a random value in the interval (0,1); Normalization layer: This layer consists of 10 neurons, and the output of each neuron is: Among them, v j (t) is the output value of the jth neuron in the normalized layer at time t, w vj (t) is the weighted connection between the radial basis function layer and the normalization layer at time t; Output layer: This layer consists of 1 neuron and the output is: in, is the output of the neural network intelligent detection model at time t, w j (t) is the weighted connection between the jth normalized layer neuron and the output layer at time t; (2) Design a data immune mechanism to dynamically adjust the model structure, as follows: The loss function is defined as: in, Represents the error of the model at time t; its loss function boundary is: Among them, Z b (t) is the loss function boundary at time t, μ is used to define the loss function boundary, which requires that the loss function boundary at time t is greater than the loss function value at time t, and is randomly selected between (0, 0.15]; The dynamic adjustment of the detection model structure based on the self-organizing robust neural network is as follows: ① Growth stage: If Z(t) is greater than Z b (t), and the j-th radial basis function layer neuron satisfies the following conditions: w vj (t)≥max(w v (t-1)) (7) Then add a new radial basis function neuron, where w v (t-1) = [w v1 (t-1),…,w vP (t-1)]; the parameters of this neuron can be expressed as: Among them, m new (t) and σ new (t) are the center and width of the neuron in the radial basis function layer, m j (t) is the center matrix of the jth radial basis function layer neuron at time t, x(t) is the input variable matrix at time t, σ j (t) is the width matrix of the neurons in the j-th radial basis function layer at time t; w vnew (t) and w new (t) are the weights of the neurons in the radial basis function layer, w v (t) is the weight matrix of all radial basis function layer neurons at time t, g(t) is the expected output at time t, is the predicted output at time t, x i (t) is the input of the i-th neuron at time t, m inew (t) is the center of the i-th membership function of the new neuron at time t, σ inew (t) is the width of the i-th membership function of the new neuron at time t; ② Pruning stage: If Z b (t-1) is greater than Z(t), and the weights of the neurons in the h-th radial basis function layer satisfy the following conditions: Then the hth radial basis function layer neuron will be pruned; at the same time, the parameters of the hth and h'th radial basis function layer neurons are set to: Among them, the h'th radial basis function layer neuron is the neuron with the closest Euclidean distance to the hth radial basis function layer neuron, m' h (t) and m' h’ (t) are the centers of the hth and h'th radial basis function layer neurons after pruning neurons, respectively. h’ (t) is the center of the h' radial basis function layer neurons before the pruned neuron; σ' h (t) and σ' h’ (t) are the widths of the hth and h'th radial basis function layer neurons after pruning neurons, σ h’ (t) is the width of the h'th radial basis function layer neuron before pruning the neuron; w' vh’ (t) and w' h’ (t) are the weights of the h'th radial basis function layer neurons and normalized layer neurons after pruning neurons, respectively, w vh’ (t) and w h’ (t) are the weights of the h'th radial basis function layer neurons and normalized layer neurons before pruning neurons, respectively, w vh (t) and w h (t) are the weights of the hth radial basis function layer neurons and normalized layer neurons before pruning neurons; w' vh (t) and w' h (t) are the weights of the neurons in the hth radial basis function layer before and after pruning the neurons; m ih (t) and m ih’ (t) are the centers of the hth and h'th radial basis function layer neurons before pruning the neurons, respectively, σ ih (t) and σ ih’ (t) are the widths of the hth and h'th radial basis function layer neurons before pruning the neurons; (3) Update the model parameters using the adaptive gradient descent algorithm, as follows: The parameter update gradient of the detection model based on the self-organizing robust neural network is: (4) Detection of membrane permeability The membrane permeability is detected using the trained robust neural network model, and the collected water production pressure, water production flow, water production turbidity and single-cell membrane scrubbing gas volume are used as the input variables of the model to obtain the membrane permeability prediction output value of the model.

Images