CN109473182B - MBR (Membrane biological reactor) membrane water permeability intelligent detection method based on deep belief network - Google Patents
MBR (Membrane biological reactor) membrane water permeability intelligent detection method based on deep belief network Download PDFInfo
- Publication number
- CN109473182B CN109473182B CN201811338058.1A CN201811338058A CN109473182B CN 109473182 B CN109473182 B CN 109473182B CN 201811338058 A CN201811338058 A CN 201811338058A CN 109473182 B CN109473182 B CN 109473182B
- Authority
- CN
- China
- Prior art keywords
- layer
- hidden layer
- time
- neuron
- output
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
Abstract
The invention relates to an MBR membrane water permeability intelligent detection method based on a deep belief network, which comprises the following steps: (1) determining a target variable and a characteristic variable; (2) designing a membrane water permeability soft measurement model, and establishing a soft measurement model for predicting the membrane water permeability by using a DBN (direct bonded network): (3) correcting the established outlet water permeability soft measurement model to obtain a simulation error curve chart and a prediction result chart; (4) and (5) predicting the water permeability. The soft measurement model of the membrane water permeability is established based on the DBN, the calculation complexity of the water permeability is reduced, online accurate measurement and real-time correction of the water permeability are realized, an effective method is provided for predicting the pollution state of the membrane in sewage treatment, and the working efficiency and the economic benefit of the MBR membrane sewage treatment process are improved.
Description
Technical Field
The invention relates to the technical field of online detection of water quality in sewage treatment, in particular to an MBR (membrane bioreactor) membrane water permeability intelligent detection method based on a deep belief network.
Background
In recent years, the number of sewage treatment plants and sewage treatment capacity in China are gradually increased, the increase speed of sewage treatment capacity is higher than that of sewage treatment capacity, the operation load rate is also rapidly increased, and the demand for sewage treatment is rapidly increased. Meanwhile, the national development planning puts forward a sewage treatment technology with low energy consumption and high efficiency to be researched and popularized, the annual average growth rate of the total value of the membrane industry is more than or equal to 20 percent, and the annual average growth rate is predicted to reach 2000-2500 billion yuan in 2020. The MBR membrane plays an important role in sewage treatment, and has a very wide prospect in sewage treatment due to the performance advantages of the MBR membrane, which is one of the backgrounds of the research of the invention.
The advantages of the MBR membrane are not described in detail, the defects of the application of the traditional activated sludge process treatment technology are overcome, and the sewage regeneration treatment technology is improved to a new level. However, membrane fouling is inevitable during the MBR process for treating sewage. Membrane fouling is an inevitable problem in the process of treating sewage by applying an MBR process, and can cause many adverse effects, such as:
the sewage treatment capacity of the MBR is reduced. During the actual operation of the MBR sewage treatment plant, the filtering performance of the MBR membrane is greatly reduced with the accumulation of time. The United states environmental protection agency has made experiments to find that when the MBR membrane is operated for 2-3 years, the water inlet flow rate exceeds 1.5-2 times of the daily average flow rate in order to keep the water permeability of the MBR membrane consistent with that of the MBR membrane when the MBR membrane is just started to operate.
Secondly, the investment cost and the operation cost are increased. Membranes are key components in MBR processes, and are the components that require the most technical support and cost, and their capital investment is the greatest in the plant. According to analysis, the domestic MBR investment cost is 2000-2500 yuan/m3The construction cost is about 1.5 times of the construction cost of the traditional activated sludge process project. The membrane pollution causes the service life of the membrane component to be shortened, and the membrane component needs to be replaced after the membrane component does not reach a certain service life, so that the use cost of the MBR is greatly increased.
In the long-term process of membrane treatment sewage, the membrane pollution problem can cause the reduction of membrane flux and water permeability, the increase of transmembrane pressure difference and membrane surface resistance, the water quality of effluent of the membrane is reduced, the service life of the membrane is shortened, and the further large-scale application of the membrane water treatment technology is hindered, so that the membrane is necessary to be cleaned and maintained in time by correctly predicting the pollution condition of the membrane. The water permeability can directly indicate the degree of fouling of the membrane, and thus the degree of fouling of the membrane can be predicted by predicting the magnitude of the water permeability. The water permeability can not be directly measured, and a water plant generally adopts a calculation method to estimate the water permeability, so that the water permeability has serious hysteresis, and the accurate online prediction of the membrane pollution condition can not be realized. Therefore, the research on new prediction technologies to solve the problem of real-time acquisition of process variables has become an important subject of research in the field of sewage control and has important practical significance.
Disclosure of Invention
In view of the above-mentioned deficiencies of the prior art membrane treatments, the present invention is directed to: the MBR membrane water permeability intelligent detection method based on the Deep Belief Network (DBN for short) is provided, a soft measurement model of the membrane water permeability is established based on the DBN, the calculation complexity of the water permeability is reduced, online accurate measurement and real-time correction of the water permeability are realized, an effective method is provided for predicting the pollution state of the membrane in sewage treatment, and the working efficiency and the economic benefit of the MBR membrane sewage treatment process are improved.
In order to achieve the purpose, the invention adopts the following technical scheme:
an MBR membrane water permeability intelligent detection method based on a deep belief network is characterized in that: the detection method comprises the following steps:
(1) determining a characteristic variable and a target variable; taking an MBR membrane sewage treatment process as a research object, performing characteristic analysis on water quality data, extracting produced water flow, produced water pressure, single-pool membrane scrubbing gas flow, anaerobic zone oxidation-reduction potential ORP and aerobic zone nitrate as characteristic variables, wherein the 5 variables are respectively used as 5 inputs of a deep belief network, and effluent water permeability is used as a target variable;
(2) designing a membrane water permeability soft measurement model, and establishing a soft measurement model for predicting the membrane water permeability by using a DBN (database network), wherein the DBN comprises 1 input layer, 2 hidden layers and 1 output layer, the number of neurons of the input layer is 5, the number of neurons of each hidden layer is N, N is a positive integer greater than 2, and the number of neurons of the output layer is 1, namely the connection mode is 5-N-N-1; the input to the DBN is x (1), x (2), …, x (t), …, x (k), and the corresponding desired output is yd(1)，yd(2)，…，yd(t)，…，yd(k) The k groups of data are used as training samples of the soft measurement model; DBN input at time t is x (t) ═ x1(t),…,x5(t)]Wherein x is1(t) is the water production flow at time t, x2(t) is the water production pressure at time t, x3(t) is the scrubbing gas amount of the single cell membrane at the time of the t, x4(t) is the anaerobic zone reduction potential ORP, x at time t5(t) nitrate in the aerobic zone at time t, and the desired water permeability output of the DBN is expressed as yd(t), the actual water permeability output is represented as y (t); the soft measurement model calculation mode based on the DBN prediction membrane water permeability sequentially comprises the following steps:
inputting a layer: this layer consists of 5 neurons with the input vector:
x(t)＝[x1(t),x2(t),…,x5(t)]T (1)
v(t)＝x(t) (2)
where x (t) is the input vector at time t, x1(t) value x representing the water production flow at time t2(t) value x representing water pressure at time t3(t) represents the value of the scrubbing gas amount of the single cell membrane at the time t, x4(t) value of reduction potential ORP in anaerobic zone at time t, x5(t) represents the value of nitrate in the aerobic zone at time t, and v (t) is the output vector of the input layer at time t;
② first hidden layer: this layer consists of N neurons, the output of each neuron being:
wherein, w1ij(t) is a connection weight between the ith neuron of the input layer and the jth neuron of the first hidden layer at the time t, wherein i is 1, 2, …, 5; j ═ 1, 2, …, N; b1j(t) is the bias of the jth neuron of the first hidden layer at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t;
third, the second hidden layer: this layer consists of N neurons, the output of each neuron being:
wherein, w2jq(t) is the connection weight of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the time t, q is 1, 2, …, N; b2q(t) is the bias of the qth neuron of the second hidden layer at time t, h2q(t) is the output of the qth neuron of the second hidden layer at time t;
output layer: the network output is:
wherein, woutq(t) is a connection weight between the qth neuron of the second hidden layer and the output layer at the moment t;
(3) the water permeability soft measurement model correction process is as follows:
the DBN training includes two processes: unsupervised layer-by-layer pre-training and adjusting the network weight by using a back propagation algorithm; w (t) ═ wout(t)，w2(t)，w1(t)) is the weight vector of DBN at time t, where wout(t) is the weight vector between the second hidden layer and the output layer at time t, w2(t) is the weight vector between the first hidden layer and the second hidden layer at time t, w1(t) is a weight vector between the input layer and the first hidden layer at time t; (b) (t) ═ b2(t),b1(t)), wherein, b2(t) a bias vector for the second hidden layer at time t, b1(t) is the bias vector of the first hidden layer at time t; setting the iteration number of each layer of pre-training as 100, the iteration number of a back propagation algorithm as 10000, and setting the initial weight and bias as 0.01;
unsupervised pre-training: the energy function between the input layer and the first hidden layer is defined as:
wherein h is1(t) is the output vector of the first hidden layer at time t, w1ij(t) is the connection weight between the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment t, b1j(t) is the bias of the jth neuron in the first hidden layer at time t, and c (t) ═ c1(t)，c2(t)，…，ci(t)) is the offset vector of the input layer at time t, ci(t) is the bias of the ith input layer neuron at time t, vi(t) is the output of the ith input layer neuron at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t; based on the energy function, calculating a joint probability distribution between the input layer and the first hidden layer as:
the edge probability distribution of the input layer is:
let theta1(t)＝(w1(t)，c(t)，b1(t)), defining a likelihood function:
wherein K is the number of samples; parameter theta1(t) can be obtained by maximizing the log-likelihood function, which is usually obtained by a numerical method of gradient ascent, and maximizing L (theta) by random gradient ascent1(t)) model parameters were obtained:
wherein the content of the first and second substances,
representing a mathematical expectation of the distribution defined by the training sample set,
representing an expectation of a distribution defined by the model; the parameter iteration formula is as follows:
wherein, theta1(t +1) represents the value of the model parameter at time t + 1; the update rule of the parameters obtained by the contrast divergence algorithm is as follows:
wherein, Δ w1ij(t) is the connection weight adjustment quantity between the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment t, and deltaci(t) is the offset adjustment for the ith neuron in the input layer at time t, Δ b1j(t) is the bias modulation of the jth neuron of the first hidden layer, μw1∈(0，0.02]，μc∈(0，0.01]And mub1∈(0，0.01]Learning rates of weight, input layer neuron bias and first hidden layer neuron bias respectively;
the energy function between the first hidden layer and the second hidden layer is:
wherein h is2(t) is the output vector of the second hidden layer at time t, w2jq(t) is the connection weight of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the time of t, b2q(t) is the bias of the qth neuron of the second hidden layer at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t, h2q(t) is the output of the qth neuron of the second hidden layer at time t; based on the energy function (13), a joint probability distribution between the input layer and the first hidden layer is calculated as:
the edge probability distribution of the first hidden layer is:
the likelihood function is defined as:
wherein K is the number of samples; by calculating L (theta)2(t)) the partial derivative yields the model parameters:
wherein, theta2(t)＝(w2(t)，b1(t)，b2(t)) are parameters of the model; the parameter iteration formula is as follows:
wherein, theta2(t +1) represents the value of the model parameter at time t + 1; the parameter regulating quantity of the first hidden layer neuron and the second hidden layer neuron obtained by adopting a contrast divergence algorithm is as follows:
wherein, Δ w2jq(t) is the jth neuron and the second neuron of the first hidden layer at time tConnection weight adjustment, Δ b, between the qth neurons of the hidden layer1j(t) is the bias modulation of the jth neuron in the first hidden layer, Δ b2q(t) is the bias modulation of the qth neuron of the second hidden layer, μw2∈(0，0.02]，μb1∈(0，0.01]And mub2∈(0，0.01]Learning rates of the weight, the first hidden layer neuron bias and the second hidden layer neuron bias respectively;
adjusting the weight value by BP algorithm: obtaining the initial value of the DBN parameter through the layer-by-layer training of the formulas (11) - (12) and (18) - (19); then fine tuning the weight value through a BP algorithm to obtain a better model effect; the time t loss function is defined as:
wherein, y (t) and yd(t) actual output and expected output of the network are respectively obtained, and K is the total number of training samples; and (3) adjusting the weight and the bias by adopting an error back propagation method, then:
wherein, woutq(t +1) is the connection weight of the qth neuron of the second hidden layer and the output layer at the moment of t + 1; w is a2jq(t +1) is the connection weight of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the moment of t +1, w1ij(t +1) is the connection weight of the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment of t + 1; etaout∈(0,3]Is the weight learning rate, η, between the second hidden layer and the output layer2∈(0，3]Learning rate, η, of weights between the first hidden layer and the second hidden layer1∈(0，3]Learning the weight value between the input layer and the first hidden layer;
(4) and (3) water permeability prediction:
randomly selecting the water production flow, the water production pressure, the single-pool membrane scrubbing gas flow, the anaerobic zone oxidation-reduction potential ORP and the aerobic zone nitrate data which are counted in a period of 100-150 water plants to be detected as training data, randomly selecting the test sample data of the water production flow, the water production pressure, the single-pool membrane scrubbing gas flow, the anaerobic zone oxidation-reduction potential ORP and the aerobic zone nitrate which are counted in a period of 60-100 water plants to be detected from the 100-150 training data as the input of the DBN after training, wherein the output of the DBN is the soft measurement value of the water permeability.
Compared with the MBR membrane treatment process in the prior art, the MBR membrane water permeability intelligent detection method based on the deep belief network has the following advantages that:
(1) the invention applies the intelligent detection method to the MBR membrane sewage treatment process, realizes the online intelligent detection of the membrane water permeability, obtains the pollution condition of the membrane on line according to the water permeability, improves the effluent quality of the membrane and prolongs the service life of the membrane.
(2) Aiming at the problem that the membrane water permeability can not be measured on line in the MBR membrane sewage treatment process, the invention provides a membrane water permeability prediction method based on a deep belief network by extracting characteristic variables related to the membrane water permeability, realizes the prediction of the membrane water permeability, solves the problem that the membrane water permeability is difficult to measure in real time, and has the advantages of high prediction precision and simple operation.
(3) According to the invention, the current MBR membrane sewage treatment process is a complex and dynamic time-varying process, the relationship between the membrane water permeability and the related variable has the characteristics of nonlinearity, strong coupling and the like, and is difficult to describe by using an accurate mathematical model, so that the prediction of the membrane water permeability is realized by adopting a deep belief network based on the actual measured data of a sewage treatment plant, and the MBR membrane sewage treatment process has the characteristics of high prediction precision, good adaptability to environmental differences and the like.
Drawings
FIG. 1 is a deep belief network-based intelligent feature modeling topology of the present invention;
FIG. 2 is a graph of fitting results of the present invention, wherein black line points are actually calculated values of water permeability, and black line stars are fitted values of water permeability;
FIG. 3 is a graph of the error of the fit results of the present invention;
FIG. 4 is a diagram of the prediction result, wherein the black line with points is the actual calculated value of the water permeability, and the black line with stars is the predicted value of the intelligent detection model of the water permeability;
FIG. 5 is a graph of the error of the intelligent prediction results of the present invention.
Detailed Description
The details of the MBR membrane water permeability intelligent detection method based on the deep belief network are described in detail below with reference to the attached figures 1-5.
An MBR membrane water permeability intelligent detection method based on a deep belief network is characterized in that: the detection method comprises the following steps:
(1) determining a characteristic variable and a target variable; taking an MBR membrane sewage treatment process as a research object, performing characteristic analysis on water quality data, and extracting water production flow, water production pressure, single-pool membrane scrubbing gas flow, anaerobic zone reduction potential ORP and aerobic zone nitrate as characteristic variables, wherein the 5 variables respectively correspond to 5 input neurons of the DBN and serve as 5 inputs of the DBN; taking the water permeability of the effluent as a target variable;
(2) designing a membrane water permeability soft measurement model, and establishing a soft measurement model for predicting the membrane water permeability by using a DBN (database network), wherein the DBN comprises 1 input layer, 2 hidden layers and 1 output layer, the number of neurons of the input layer is 5, the number of neurons of each hidden layer is N, N is a positive integer greater than 2, and the number of neurons of the output layer is 1, namely the connection mode is 5-N-N-1; the input to the DBN is x (1), x (2), …, x (t), …, x (k), and the corresponding desired output is yd(1)，yd(2)，…，yd(t)，…，yd(k) The k groups of data are used as training samples of the soft measurement model; DBN input at time t is x (t) ═ x1(t),…,x5(t)]The desired output of the DBN is denoted as yd(t), the actual output is represented as y (t); the soft measurement model calculation mode based on the DBN prediction membrane water permeability sequentially comprises the following steps:
inputting a layer: this layer consists of 5 neurons with the input vector:
x(t)＝[x1(t),x2(t),…,x5(t)]T (1)
v(t)＝x(t) (2)
where x (t) is the input vector at time t, x1(t) value x representing the water production flow at time t2(t) value x representing water pressure at time t3(t) represents the value of the scrubbing gas amount of the single cell membrane at the time t, x4(t) value of reduction potential ORP in anaerobic zone at time t, x5(t) represents the value of nitrate in the aerobic zone at time t, and v (t) is the output vector of the input layer at time t;
② first hidden layer: this layer consists of N neurons, the output of each neuron being:
wherein, w1ij(t) is a connection weight between the ith neuron of the input layer and the jth neuron of the first hidden layer at the time t, wherein i is 1, 2, …, 5; j ═ 1, 2, …, N; b1j(t) is the bias of the jth neuron of the first hidden layer at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t;
third, the second hidden layer: this layer consists of N neurons, the output of each neuron being:
wherein, w2jq(t) is the connection weight of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the time t, q is 1, 2, …, N; b2q(t) is the bias of the qth neuron of the second hidden layer at time t, h2q(t) is the output of the qth neuron of the second hidden layer at time t;
output layer: the network output is:
wherein, woutq(t) is a connection weight between the qth neuron of the second hidden layer and the output layer at the moment t;
(3) the water permeability soft measurement model correction process is as follows:
the DBN training includes two processes: unsupervised layer-by-layer pre-training and adjusting the network weight by using a back propagation algorithm; w (t) ═ wout(t)，w2(t)，w1(t)) is the weight vector of DBN at time t, where wout(t) is the weight vector between the second hidden layer and the output layer at time t, w2(t) is the weight vector between the first hidden layer and the second hidden layer at time t, w1(t) is a weight vector between the input layer and the first hidden layer at time t; (b) (t) ═ b2(t),b1(t)), wherein, b2(t) a bias vector for the second hidden layer at time t, b1(t) is the bias vector of the first hidden layer at time t; setting the iteration number of each layer of pre-training as 100, the iteration number of a back propagation algorithm as 10000, and setting the initial weight and bias as 0.01;
unsupervised pre-training: the energy function between the input layer and the first hidden layer is defined as:
wherein h is1(t) is the output vector of the first hidden layer at time t, w1ij(t) is the connection weight between the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment t, b1j(t) is the bias of the jth neuron in the first hidden layer at time t, and c (t) ═ c1(t)，c2(t)，…，ci(t)) is the offset vector of the input layer at time t, ci(t) is the bias of the ith input layer neuron at time t, vi(t) is the output of the ith input layer neuron at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t; based on the energy function, calculating a joint probability distribution between the input layer and the first hidden layer as:
the edge probability distribution of the input layer is:
let theta1(t)＝(w1(t)，c(t)，b1(t)), defining a likelihood function:
wherein K is the number of samples; parameter theta1(t) can be obtained by maximizing the log-likelihood function, which is usually obtained by a numerical method of gradient ascent, and maximizing L (theta) by random gradient ascent1(t)) model parameters were obtained:
wherein the content of the first and second substances,
representing a mathematical expectation of the distribution defined by the training sample set,
representing an expectation of a distribution defined by the model; the parameter iteration formula is as follows:
wherein, theta1(t +1) represents the value of the model parameter at time t + 1; the update rule of the parameters obtained by the contrast divergence algorithm is as follows:
wherein, Δ w1ij(t) is the connection weight adjustment quantity between the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment t, and deltaci(t) is the offset adjustment for the ith neuron in the input layer at time t, Δ b1j(t) is the bias modulation of the jth neuron of the first hidden layer, μw1∈(0，0.02]，μc∈(0，0.01]And mub1∈(0，0.01]Learning rates of weight, input layer neuron bias and first hidden layer neuron bias respectively;
the energy function between the first hidden layer and the second hidden layer is:
wherein h is2(t) is the output vector of the second hidden layer at time t, w2jq(t) is the connection weight of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the time of t, b2q(t) is the bias of the qth neuron of the second hidden layer at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t, h2q(t) is the output of the qth neuron of the second hidden layer at time t; based on the energy function (13), a joint probability distribution between the input layer and the first hidden layer is calculated as:
the edge probability distribution of the first hidden layer is:
the likelihood function is defined as:
wherein K is the number of samples; by calculating L: (θ2(t)) the partial derivative yields the model parameters:
wherein, theta2(t)＝(w2(t)，b1(t)，b2(t)) are parameters of the model; the parameter iteration formula is as follows:
wherein, theta2(t +1) represents the value of the model parameter at time t + 1; the parameter regulating quantity of the first hidden layer neuron and the second hidden layer neuron obtained by adopting a contrast divergence algorithm is as follows:
wherein, Δ w2jq(t) is the connection weight adjustment between the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at time t, Δ b1j(t) is the bias modulation of the jth neuron in the first hidden layer, Δ b2q(t) is the bias modulation of the qth neuron of the second hidden layer, μw2∈(0，0.02]，μb1∈(0，0.01]And mub2∈(0，0.01]Learning rates of the weight, the first hidden layer neuron bias and the second hidden layer neuron bias respectively;
adjusting the weight value by BP algorithm: obtaining the initial value of the DBN parameter through the layer-by-layer training of the formulas (11) - (12) and (18) - (19); then fine tuning the weight value through a BP algorithm to obtain a better model effect; the time t loss function is defined as:
wherein, y (t) and yd(t) actual output and expected output of the network, and K is a training sampleThe total number of the samples; and (3) adjusting the weight and the bias by adopting an error back propagation method, then:
wherein, woutq(t +1) is the connection weight of the qth neuron of the second hidden layer and the output layer at the moment of t + 1; w is a2jq(t +1) is the connection weight of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the moment of t +1, w1ij(t +1) is the connection weight of the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment of t + 1; etaout∈(0,3]Is the weight learning rate, η, between the second hidden layer and the output layer2∈(0，3]Learning rate, η, of weights between the first hidden layer and the second hidden layer1∈(0，3]Learning the weight value between the input layer and the first hidden layer;
(4) and (3) water permeability prediction:
randomly selecting the water production flow, the water production pressure, the single-pool membrane scrubbing gas flow, the anaerobic zone oxidation-reduction potential ORP and the aerobic zone nitrate data which are counted in a period of 100-150 water plants to be detected as training data, randomly selecting the test sample data of the water production flow, the water production pressure, the single-pool membrane scrubbing gas flow, the anaerobic zone oxidation-reduction potential ORP and the aerobic zone nitrate which are counted in a period of 60-100 water plants to be detected from the 100-150 training data as the input of the DBN after training, wherein the output of the DBN is the soft measurement value of the water permeability.
Next, a certain regeneration water plant is used for testing, and the DBN is trained and tested according to actually collected data. 120 sets of data were selected for training and 80 sets of data were selected for testing. Table 1 is the training data for the present invention and table 2 is the test data. The method comprises the following specific steps:
an MBR membrane water permeability intelligent detection method based on a deep belief network is characterized in that characteristic variables of MBR membrane water permeability are obtained through characteristic analysis, a soft measurement model of MBR membrane water permeability is established by utilizing the deep belief network, and the MBR membrane water permeability intelligent detection is realized, and the method comprises the following steps:
(1) determining a target variable and a characteristic variable; taking a membrane bioreactor-MBR sewage treatment system as a research object, performing characteristic analysis on water quality data, extracting the produced water flow, the produced water pressure, the single-tank membrane scrubbing air flow, the reduction potential ORP of an anaerobic zone and the nitrate of an aerobic zone as characteristic variables, and taking the effluent permeability as a target variable; the parameter information and the collection position are shown in table 1.
TABLE 1 Process variable types collected
| Parameter name | Unit of | Acquisition position | Collection instrument |
| Flow of produced water | m3/h | MBR tank head end | ViSolid700IQ |
| Scrubbing gas flow of single pool membrane | m3/h | Air washing pump | SensoLyt700IQ |
| Nitrate in aerobic zone | mg/l | End of aerobic tank | NitraLyt700IQ |
| Pressure of produced water | kPa | MBR tank end | SensoLyt700IQ |
| Reduction potential OPR of anaerobic zone | mV | End of anaerobic tank | SensoLyt700IQ |
(2) Designing a membrane water permeability soft measurement model, and establishing a soft measurement model for predicting the membrane water permeability by using a DBN (direct bonded network):
and training and testing the DBN by adopting data actually collected by a water plant in 2017. 120 sets of data were selected for training. 80 sets of data were selected for testing. Table 1 is the training data for the present invention and table 2 is the test data.
(3) The established outlet water permeability soft measurement model is corrected, and the obtained simulation error curve diagram and the prediction result diagram are respectively shown in fig. 3 and fig. 5.
(4) Predicting the water permeability;
and taking test sample data as the input of the trained DBN, wherein the output of the DBN is the measured value of the membrane water permeability of the intelligent detection method.
As can be seen by combining the attached figures 3 and 5, the ranges of the training errors and the testing errors are respectively about +/-0.05 and +/-0.01, and the method has high prediction precision and good adaptability to environmental differences.
TABLE 1 training data
Table 2 test data:
the embodiments described above are intended to facilitate one of ordinary skill in the art in understanding and using the present invention. It will be readily apparent to those skilled in the art that various modifications to these embodiments may be made, and the generic principles described herein may be applied to other embodiments without the use of the inventive faculty. Therefore, the present invention is not limited to the embodiments described herein, and those skilled in the art should make improvements and modifications within the scope of the present invention based on the disclosure of the present invention.
Claims (1)
1. An MBR membrane water permeability intelligent detection method based on a deep belief network is characterized in that: the detection method comprises the following steps:
(1) determining a characteristic variable and a target variable; taking an MBR membrane sewage treatment process as a research object, performing characteristic analysis on water quality data, extracting water production flow, water production pressure, single-pool membrane scrubbing gas flow, anaerobic zone oxidation-reduction potential ORP and aerobic zone nitrate as characteristic variables, taking the 5 variables as 5 inputs of a deep belief network respectively, and taking effluent water permeability as a target variable;
(2) designing a soft measurement model of the water permeability of the membrane, and establishing a soft measurement model for predicting the water permeability of the membrane by using a DBN (database-based network), wherein the DBN is wrappedThe method comprises the following steps of 1 input layer, 2 hidden layers and 1 output layer, wherein 5 neurons of the input layer are provided, N neurons of each hidden layer are provided, N is a positive integer larger than 2, and 1 neuron of the output layer is provided, namely the connection mode is 5-N-N-1; the input to the DBN is x (1), x (2), …, x (t), …, x (k), and the corresponding desired output is yd(1)，yd(2)，…，yd(t)，…，yd(k) The k groups of data are used as training samples of the soft measurement model; DBN input at time t is x (t) ═ x1(t),…,x5(t)]Wherein x is1(t) is the water production flow at time t, x2(t) is the water production pressure at time t, x3(t) is the scrubbing gas amount of the single cell membrane at the time of the t, x4(t) is the anaerobic zone reduction potential ORP, x at time t5(t) nitrate in the aerobic zone at time t, and the desired water permeability output of the DBN is expressed as yd(t), the actual water permeability output is represented as y (t); the soft measurement model calculation mode based on the DBN prediction membrane water permeability sequentially comprises the following steps:
inputting a layer: this layer consists of 5 neurons with the input vector:
x(t)＝[x1(t),x2(t),…,x5(t)]T (1)
v(t)＝x(t) (2)
where x (t) is the input vector at time t, x1(t) value x representing the water production flow at time t2(t) value x representing water pressure at time t3(t) represents the value of the scrubbing gas amount of the single cell membrane at the time t, x4(t) value of reduction potential ORP in anaerobic zone at time t, x5(t) represents the value of nitrate in the aerobic zone at time t, and v (t) is the output vector of the input layer at time t;
② first hidden layer: this layer consists of N neurons, the output of each neuron being:
wherein, w1ij(t) the ith neuron and the first hidden neuron of the input layer at the time tThe connection weight between the jth neuron of the layer, i ═ 1, 2, …, 5; j ═ 1, 2, …, N; b1j(t) is the bias of the jth neuron of the first hidden layer at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t;
third, the second hidden layer: this layer consists of N neurons, the output of each neuron being:
wherein, w2jq(t) is the connection weight of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the time t, q is 1, 2, …, N; b2q(t) is the bias of the qth neuron of the second hidden layer at time t, h2q(t) is the output of the qth neuron of the second hidden layer at time t;
output layer: the network output is:
wherein, woutq(t) is a connection weight between the qth neuron of the second hidden layer and the output layer at the moment t;
(3) the water permeability soft measurement model correction process is as follows:
the DBN training includes two processes: unsupervised layer-by-layer pre-training and adjusting the network weight by using a back propagation algorithm; w (t) ═ wout(t)，w2(t)，w1(t)) is the weight vector of DBN at time t, where wout(t) is the weight vector between the second hidden layer and the output layer at time t, w2(t) is the weight vector between the first hidden layer and the second hidden layer at time t, w1(t) is a weight vector between the input layer and the first hidden layer at time t; (b) (t) ═ b2(t),b1(t)), wherein, b2(t) a bias vector for the second hidden layer at time t, b1(t) is the bias vector of the first hidden layer at time t; setting per-layer pretrainingThe iteration number of (2) is 100, the iteration number of the back propagation algorithm is 10000, and the initial weight and the bias are set to be 0.01;
unsupervised pre-training: the energy function between the input layer and the first hidden layer is defined as:
wherein h is1(t) is the output vector of the first hidden layer at time t, w1ij(t) is the connection weight between the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment t, b1j(t) is the bias of the jth neuron in the first hidden layer at time t, and c (t) ═ c1(t)，c2(t)，…，ci(t)) is the offset vector of the input layer at time t, ci(t) is the bias of the ith input layer neuron at time t, vi(t) is the output of the ith input layer neuron at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t; based on the energy function, calculating a joint probability distribution between the input layer and the first hidden layer as:
the edge probability distribution of the input layer is:
let theta1(t)＝(w1(t)，c(t)，b1(t)), defining a likelihood function:
wherein K is the number of samples; parameter theta1(t) can be obtained by maximizing the log-likelihood functionTo this end, a common numerical method for maximizing the likelihood function is gradient ascent, where L (θ) is maximized by random gradient ascent1(t)) model parameters were obtained:
wherein the content of the first and second substances,
representing a mathematical expectation of the distribution defined by the training sample set,
representing an expectation of a distribution defined by the model; the parameter iteration formula is as follows:
wherein, theta1(t +1) represents the value of the model parameter at time t + 1; the update rule of the parameters obtained by the contrast divergence algorithm is as follows:
wherein, Δ w1ij(t) is the connection weight adjustment quantity between the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment t, and deltaci(t) is the offset adjustment for the ith neuron in the input layer at time t, Δ b1j(t) is the bias modulation of the jth neuron of the first hidden layer, μw1∈(0，0.02]，μc∈(0，0.01]And mub1∈(0，0.01]Learning rates of weight, input layer neuron bias and first hidden layer neuron bias respectively;
the energy function between the first hidden layer and the second hidden layer is:
wherein h is2(t) is the output vector of the second hidden layer at time t, w2jq(t) is the connection weight of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the time of t, b2q(t) is the bias of the qth neuron of the second hidden layer at time t, h1j(t) is the output of the jth neuron of the first hidden layer at time t, h2q(t) is the output of the qth neuron of the second hidden layer at time t; based on the energy function (13), a joint probability distribution between the input layer and the first hidden layer is calculated as:
the edge probability distribution of the first hidden layer is:
the likelihood function is defined as:
wherein K is the number of samples; by calculating L (theta)2(t)) the partial derivative yields the model parameters:
wherein, theta2(t)＝(w2(t)，b1(t)，b2(t)) are parameters of the model; the parameter iteration formula is as follows:
wherein, theta2(t +1) represents the value of the model parameter at time t + 1; the parameter regulating quantity of the first hidden layer neuron and the second hidden layer neuron obtained by adopting a contrast divergence algorithm is as follows:
wherein, Δ w2jq(t) is the connection weight adjustment between the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at time t, Δ b1j(t) is the bias modulation of the jth neuron in the first hidden layer, Δ b2q(t) is the bias modulation of the qth neuron of the second hidden layer, μw2∈(0，0.02]，μb1∈(0，0.01]And mub2∈(0，0.01]Learning rates of the weight, the first hidden layer neuron bias and the second hidden layer neuron bias respectively;
adjusting the weight value by BP algorithm: obtaining the initial value of the DBN parameter through the layer-by-layer training of the formulas (11) - (12) and (18) - (19); then fine tuning the weight value through a BP algorithm to obtain a better model effect; the time t loss function is defined as:
wherein, y (t) and yd(t) actual output and expected output of the network are respectively obtained, and K is the total number of training samples; and (3) adjusting the weight and the bias by adopting an error back propagation method, then:
wherein, woutq(t +1) is the connection weight of the qth neuron of the second hidden layer and the output layer at the moment of t + 1; w is a2jq(t +1) of the jth neuron of the first hidden layer and the qth neuron of the second hidden layer at the moment of t +1Connection weight, w1ij(t +1) is the connection weight of the ith neuron of the input layer and the jth neuron of the first hidden layer at the moment of t + 1; etaout∈(0,3]Is the weight learning rate, η, between the second hidden layer and the output layer2∈(0，3]Learning rate, η, of weights between the first hidden layer and the second hidden layer1∈(0，3]Learning the weight value between the input layer and the first hidden layer;
(4) and (3) water permeability prediction:
randomly selecting the water production flow, the water production pressure, the single-pool membrane scrubbing gas flow, the anaerobic zone oxidation-reduction potential ORP and the aerobic zone nitrate data which are counted in a period of 100-150 water plants to be detected as training data, randomly selecting the test sample data of the water production flow, the water production pressure, the single-pool membrane scrubbing gas flow, the anaerobic zone oxidation-reduction potential ORP and the aerobic zone nitrate which are counted in a period of 60-100 water plants to be detected from the 100-150 training data as the input of the DBN after training, wherein the output of the DBN is the soft measurement value of the water permeability.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811338058.1A CN109473182B (en) | 2018-11-12 | 2018-11-12 | MBR (Membrane biological reactor) membrane water permeability intelligent detection method based on deep belief network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811338058.1A CN109473182B (en) | 2018-11-12 | 2018-11-12 | MBR (Membrane biological reactor) membrane water permeability intelligent detection method based on deep belief network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109473182A CN109473182A (en) | 2019-03-15 |
| CN109473182B true CN109473182B (en) | 2021-10-19 |
Family
ID=65671659
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811338058.1A Active CN109473182B (en) | 2018-11-12 | 2018-11-12 | MBR (Membrane biological reactor) membrane water permeability intelligent detection method based on deep belief network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109473182B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111204867B (en) * | 2019-06-24 | 2021-12-10 | 北京工业大学 | Membrane bioreactor-MBR membrane pollution intelligent decision-making method |
| CN110668562B (en) * | 2019-10-25 | 2022-05-13 | 中信环境技术（广州）有限公司 | Control method and system for eliminating membrane bioreactor pollution in real time and storage medium |
| CN112488286A (en) * | 2019-11-22 | 2021-03-12 | 大唐环境产业集团股份有限公司 | MBR membrane pollution online monitoring method and system |
| CN111204842A (en) * | 2019-12-23 | 2020-05-29 | 光大环境科技（中国）有限公司 | Method, device and system for realizing ultrafiltration membrane pollution evaluation through neural network |
| CN111573782B (en) * | 2020-05-25 | 2022-05-20 | 青岛涌利兴水科技有限公司 | Method and device for controlling membrane pollution of water treatment equipment |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101080384A (en) * | 2004-12-16 | 2007-11-28 | 诺瓦提斯公司 | Manufacture process of N-substituted salicylamide |
| CN202179933U (en) * | 2010-08-13 | 2012-04-04 | 旭化成化学株式会社 | Composite porous hollow fiber membrane, membrane component and membrane filtration apparatus |
| CN106096730A (en) * | 2016-06-09 | 2016-11-09 | 北京工业大学 | A kind of intelligent detecting method of MBR film permeability rate based on Recurrent RBF Neural Networks |
| CN106706491A (en) * | 2016-11-21 | 2017-05-24 | 北京工业大学 | Intelligent detection method for water permeation rate of membrane bioreactor MBR |
| CN106769748A (en) * | 2016-11-21 | 2017-05-31 | 北京工业大学 | The intelligent checking system of membrane bioreactor MBR water outlet permeability rates |
| CN107578007A (en) * | 2017-09-01 | 2018-01-12 | 杭州电子科技大学 | A kind of deep learning face identification method based on multi-feature fusion |
| CN108375534A (en) * | 2018-02-06 | 2018-08-07 | 北京工业大学 | MBR fouling membrane intelligent early-warning methods |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| BR112012011129A2 (en) * | 2009-11-10 | 2017-06-20 | Novozymes Biologicals Inc | method for improving the permeability of a membrane used in a process, or the flow through a membrane used in a process, method for increasing the critical flow of a membrane used in a process, and method for producing or preventing fouling of a membrane. membrane used in a process. |
| CN105510546B (en) * | 2015-12-27 | 2017-06-16 | 北京工业大学 | A kind of biochemical oxygen demand (BOD) BOD intelligent detecting methods based on self-organizing Recurrent RBF Neural Networks |
| CN106295800B (en) * | 2016-07-28 | 2019-03-05 | 北京工业大学 | A kind of water outlet total nitrogen TN intelligent detecting method based on recurrence Self organizing RBF Neural Network |
-
2018
- 2018-11-12 CN CN201811338058.1A patent/CN109473182B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101080384A (en) * | 2004-12-16 | 2007-11-28 | 诺瓦提斯公司 | Manufacture process of N-substituted salicylamide |
| CN202179933U (en) * | 2010-08-13 | 2012-04-04 | 旭化成化学株式会社 | Composite porous hollow fiber membrane, membrane component and membrane filtration apparatus |
| CN106096730A (en) * | 2016-06-09 | 2016-11-09 | 北京工业大学 | A kind of intelligent detecting method of MBR film permeability rate based on Recurrent RBF Neural Networks |
| CN106706491A (en) * | 2016-11-21 | 2017-05-24 | 北京工业大学 | Intelligent detection method for water permeation rate of membrane bioreactor MBR |
| CN106769748A (en) * | 2016-11-21 | 2017-05-31 | 北京工业大学 | The intelligent checking system of membrane bioreactor MBR water outlet permeability rates |
| CN107578007A (en) * | 2017-09-01 | 2018-01-12 | 杭州电子科技大学 | A kind of deep learning face identification method based on multi-feature fusion |
| CN108375534A (en) * | 2018-02-06 | 2018-08-07 | 北京工业大学 | MBR fouling membrane intelligent early-warning methods |
Non-Patent Citations (2)
| Title |
|---|
| An early warning system for MBR based on multi-step prediction and deep belief network classifier;Hong-Gui Han;《 2017 Chinese Automation Congress (CAC)》;20180101;全文 * |
| 基于递归RBF神经网络的MBR膜透水率软测量;韩红桂;《北京工业大学学报》;20170731;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109473182A (en) | 2019-03-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109473182B (en) | MBR (Membrane biological reactor) membrane water permeability intelligent detection method based on deep belief network | |
| CN103197544B (en) | Sewage disposal process multi-purpose control method based on nonlinear model prediction | |
| CN102411308B (en) | Adaptive control method of dissolved oxygen (DO) based on recurrent neural network (RNN) model | |
| CN108898215B (en) | Intelligent sludge bulking identification method based on two-type fuzzy neural network | |
| CN104360035B (en) | A kind of sewage total phosphorus TP flexible measurement method based on self-organization population-radial base neural net | |
| CN111291937A (en) | Method for predicting quality of treated sewage based on combination of support vector classification and GRU neural network | |
| CN109828089B (en) | DBN-BP-based water quality parameter nitrous acid nitrogen online prediction method | |
| CN101576734B (en) | Dissolved oxygen control method based on dynamic radial basis function neural network | |
| CN106802563B (en) | A kind of sewage procedure optimization control method based on drosophila optimization and LSSVM | |
| CN102854296A (en) | Sewage-disposal soft measurement method on basis of integrated neural network | |
| CN103728431A (en) | Industrial sewage COD (chemical oxygen demand) online soft measurement method based on ELM (extreme learning machine) | |
| CN109657790B (en) | PSO-based recursive RBF neural network effluent BOD prediction method | |
| CN108536106B (en) | Aeration system dissolved oxygen online regulation and control method based on Kalman filtering-extreme learning machine | |
| CN108375534A (en) | MBR fouling membrane intelligent early-warning methods | |
| CN102262147A (en) | Soft measurement method and system for effluent chemical oxygen demand (COD) of waste water treatment system | |
| CN112417765B (en) | Sewage treatment process fault detection method based on improved teacher-student network model | |
| CN106096730A (en) | A kind of intelligent detecting method of MBR film permeability rate based on Recurrent RBF Neural Networks | |
| CN103809436A (en) | Method for intelligent modeling sewage disposal process using activated sludge process | |
| CN111125907B (en) | Sewage treatment ammonia nitrogen soft measurement method based on hybrid intelligent model | |
| CN106706491B (en) | Intelligent detection method for membrane bioreactor-MBR water permeability | |
| CN112149905A (en) | Photovoltaic power station short-term power prediction method based on wavelet transformation and wavelet neural network | |
| CN111762958A (en) | Deep well aeration process optimization method and device for sewage treatment plant based on ASM2D model | |
| CN110642393B (en) | Aeration control system based on neural network model | |
| CN106769748B (en) | Intelligent detection system for water permeability of membrane bioreactor-MBR (Membrane bioreactor) | |
| CN111204867B (en) | Membrane bioreactor-MBR membrane pollution intelligent decision-making method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |