CN111204867A

CN111204867A - Membrane bioreactor-MBR membrane pollution intelligent decision-making method

Info

Publication number: CN111204867A
Application number: CN202010154288.3A
Authority: CN
Inventors: 韩红桂; 张会娟; 王盈旭; 郭民
Original assignee: Beijing University of Technology
Current assignee: Beijing University of Technology
Priority date: 2019-06-24
Filing date: 2020-03-07
Publication date: 2020-05-29
Anticipated expiration: 2040-03-07
Also published as: CN111204867B

Abstract

The membrane bioreactor-MBR membrane fouling intelligent decision-making method belongs to the field of online early warning of water quality parameters of sewage treatment. First, the membrane fouling mechanism was deeply analyzed, and a membrane fouling prediction model based on a deep belief network was constructed to achieve accurate prediction of membrane water permeability. Intelligent early warning of membrane fouling; secondly, based on the main factors of membrane fouling, an intelligent decision-making model based on data and knowledge is established to provide decision support for operators, reduce the harm caused by membrane fouling, reduce the incidence of membrane fouling, and ensure MBR The sewage treatment process is safe, and the efficient and stable operation of the MBR sewage treatment plant is promoted.

Description

Membrane bioreactor-MBR membrane pollution intelligent decision-making method

Technical Field

The invention belongs to the field of online early warning of water quality parameters in sewage treatment, and provides an intelligent early warning system for membrane bioreactor-MBR membrane pollution for the first time. On the basis of real operation data of an MBR membrane sewage treatment process, extracting characteristic variables of MBR membrane water permeability by a characteristic analysis method, and establishing a soft measurement model by using a deep belief network to predict the membrane water permeability which is difficult to directly measure in the MBR membrane sewage treatment process; establishing a membrane pollution comprehensive evaluation model by utilizing the water permeability prediction value and combining other collectable process variables of a water plant, determining main factors of membrane pollution and realizing intelligent early warning of the membrane pollution; based on data and knowledge drive, an intelligent decision model is established, accurate identification of membrane pollution is realized, the incidence rate of membrane pollution is reduced, and the sewage treatment performance of the membrane is improved.

Background

MBR is a novel sewage treatment process combining membrane separation technology and biological treatment technology, has the obvious advantages of good solid-liquid separation effect, low sludge load, small occupied area and the like, and has great development potential. However, membrane fouling is a bottleneck problem in MBR wastewater treatment processes, which can lead to reduced effluent quality and increased operating costs, and even to a breakdown of the wastewater treatment process. Currently, the main approach to the membrane fouling problem is to periodically clean and replace the membranes. In the actual sewage treatment process, the cleaning and the replacement of the membrane do not have a strict, objective and quantitative standard and are mainly carried out according to human experience. However, frequent cleaning can lead to rupture and corrosion of membrane filaments, which can reduce membrane life, increase production energy consumption and operation cost, and greatly restrict popularization and application of MBR. Therefore, the method can accurately identify the membrane pollution, reduce the incidence rate of the membrane pollution, and become a key point for ensuring the stable operation of the MBR and popularizing the MBR technology. However, MBR sewage treatment process is complex and difficult to directly model, and the monitoring of pollution condition is a difficult problem in the current control field. Currently, an effective early warning decision system is not available in a membrane sewage treatment plant which is built and put into operation so as to realize intelligent early warning decision in the membrane sewage treatment process. Therefore, the research of new decision-making technology for solving the membrane pollution problem in the sewage treatment process becomes an important subject of research in the field of sewage control and has important practical significance.

The invention relates to an intelligent decision-making method for membrane bioreactor-MBR membrane pollution, which utilizes a characteristic analysis method to extract characteristic variables and establishes a soft measurement model of membrane water permeability based on a deep belief network to realize accurate prediction of the water permeability in the process of membrane sewage treatment, utilizes a water permeability prediction value to establish a membrane pollution comprehensive evaluation model in combination with other collectable process variables of a water plant, but an intelligent decision-making system for membrane pollution at home and abroad does not form a complete theoretical system, builds an intelligent decision-making method for MBR membrane pollution including soft and hard platforms based on an intelligent method, and has high development and application values in the aspects of filling technical blanks at home and abroad, integrating a sewage treatment industrial chain and the like.

Disclosure of Invention

The membrane bioreactor-MBR membrane pollution intelligent decision method comprises the steps of operation process data acquisition, operation process data preprocessing, membrane pollution intelligent prediction and membrane pollution intelligent decision, and specifically comprises the following steps:

(1) acquiring data of an operation process: the MBR membrane processing system is used as a research object, and the data of the operation process is collected through a collection instrument arranged on a process field, and the MBR membrane processing system comprises the following steps: the method comprises the following steps of (1) realizing data acquisition by using inlet water chemical oxygen demand, inlet water pH value, inlet water biological oxygen demand, outlet water chemical oxygen demand, outlet water pH value, outlet water biological oxygen demand, anaerobic zone oxidation-reduction potential, anoxic zone oxidation-reduction potential, aerobic zone nitrate, aerobic zone dissolved oxygen, water production pressure, water production turbidity, water production flow, sludge concentration and aeration quantity; the data collected by the instrument is transmitted to the programmable logic controller through a communication protocol, the programmable logic controller transmits the operation process data to the upper computer through the communication protocol, and the data in the upper computer is transmitted to the data processing server through the local area network;

(2) preprocessing operation process data: the operating data of the membrane pool is taken as a research object, a characteristic analysis model is established by utilizing a partial least square method, and 5 principal component variables are obtained, wherein the principal component variables are respectively as follows: the water production pressure, the water production turbidity, the water production flow, the sludge concentration and the aeration rate, and 5 main component variables are used as input variables of the membrane pollution intelligent prediction model block; the water permeability is used as an output variable of the membrane pollution intelligent prediction module;

(3) intelligent prediction of membrane pollution: the water permeability prediction is realized by intelligent prediction of membrane pollution, wherein: the water permeability is obtained by predicting a deep belief network-DBN, the DBN consists of 1 input layer, 2 hidden layers and 1 output layer, the number of neurons in the input layers is 5, the number of neurons in each hidden layer is M, M is a positive integer which is more than 2 and less than 30, and the number of neurons in the output layers is 1, namely the connection mode is 5-M-M-1; n 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)]，x₁(t) value x representing water pressure at time t₂(t) value x representing turbidity of water produced at time t₃(t) value x representing the water production flow at time t₄(t) value x representing sludge concentration at time t₅(t) represents the value of aeration at the time t, and the calculation modes of the soft measurement model for predicting the water permeability based on the DBN sequentially comprise the following steps:

① input layer h₀(t)＝x(t)

(1)

② first hidden layer:

③ second hidden layer:

④ output layer:

wherein h is₀(t) represents the output vector of the input layer at time t, h₁(t) represents the output vector of the first hidden layer at time t, h₂(t) represents the output vector of the second hidden layer at time t, w_0,1(t) represents the weight vector between the input layer and the first hidden layer at time t, w_1,2(t) represents the weight vector between the first hidden layer and the second hidden layer at time t, w_2,3(t) represents the weight vector between the second hidden layer and the output layer at time t, b₁(t) denotes the bias vector of the first hidden layer at time t, b₂(t) represents the bias vector of the second hidden layer at time t, and y (t) represents the actual output of the DBN at time t;

the DBN training is divided into two processes: unsupervised pre-training and supervised weight fine tuning; setting the iteration number of each layer of pre-training as 100, the iteration number of a back propagation algorithm as 1000, the expected error as 0.01, and the initial weight and bias as 0.01; the specific training steps are as follows:

① unsupervised pre-training, the update rule of the parameters obtained by the contrast divergence algorithm is as follows:

wherein, w_0,1(t +1) is the weight vector between the input layer and the first hidden layer at time t +1, w_1,2(t +1) is the weight vector between the first hidden layer and the second hidden layer at time t +1, b₁(t +1) is the bias vector for the first hidden layer at time t +1, b₂(t +1) is the bias vector for the second hidden layer at time t +1, E_dataIs a expectation of training data, E_modelIs an expectation of model definition, μ_w1E (0,0.02) is the learning rate of the weights of the input layer and the first hidden layer, mu_w2E (0,0.02) is the learning rate of the weights of the first hidden layer and the second hidden layer, mu_b1E (0,0.01) is the learning rate of the first hidden layer, μ_b2E (0,0.01) is the learning rate of the second hidden layer;

②, adjusting the weight value by a back propagation algorithm, namely training layer by layer to obtain the initial value of the DBN parameter, then fine-adjusting the weight value by the back propagation algorithm to obtain better model effect, and adjusting the weight value by adopting an error back propagation method as follows:

wherein, w_2,3(t +1) is the weight vector between the second hidden layer and the output layer at time t +1, y_d(t) is the desired output of DBN at time t, η_outE (0,3) is the weight learning rate between the second hidden layer and the output layer, η₂E (0,3) is the weight learning rate between the first hidden layer and the second hidden layer, η₁E (0,3) is a weight learning rate between the input layer and the first hidden layer;

(4) intelligent decision making of membrane pollution: the membrane pollution intelligent decision is to establish a membrane pollution comprehensive evaluation model by utilizing a water permeability prediction value and combining other membrane pollution related variables, and provide a decision aiming at membrane pollution, and the process is as follows:

1) determining an evaluation index of membrane fouling, u ═ u { [ u ]₁,u₂,u₃,u₄,u₅,u₆}，u₁Is the value of the water production flow, u₂Is the value of the water production pressure u₃Is the value of turbidity of the produced water, u₄Is the value of the sludge concentration u₅Is the value of aeration amount and u₆Is a predicted value of the water permeability;

2) establishing monitoring statistics, wherein the process variable is u epsilon R^6×nWherein 6 is a variable dimension, and n is the number of samples; and (3) carrying out independent component and principal component decomposition on the membrane pollution evaluation matrix u:

u＝As+Pt^T+f^T(7)

wherein A ∈ R^6×rFor the mixing matrix, s ∈ R^r×nFor independent matrices, t ∈ R^n×kAs principal component scoring matrix, load matrix P is formed by R^6×k，f∈R^n×6The final residual matrix is obtained, r represents the number of independent elements, and k is the number of selected principal elements; to estimate A and s, a unmixing matrix W is required to obtain a reconstructed signal

Is an estimate of the source signal, whitening uProcessing to obtain a score vector:

Z＝Λ^-1/2H^Tu＝Bs (8)

wherein Λ is a diagonal matrix containing all eigenvalues, H is a corresponding eigenvector matrix and the unit is orthogonal, and B is a unit orthogonal matrix; the unmixing matrix is:

W＝B^TΛ^-1/2H^T(9)

decomposing the original data by independent components and principal components to obtain independent component matrix and principal component scoring matrix, and establishing I²And T²Statistics, namely establishing SPE statistics for the residual information; independent component vector is s_l＝[s_1l, s_2l,…,s_rl]^T∈R^r×1And a pivot score vector t_l＝[t_l1,t_l2,…,t_lk]^T∈R^k×1Residual vector f_l＝[f_l1, f_l2,…,f_l6]^T∈

R

^6×11,2, …, n (n is the number of samples); establishment of I²、T²And the SPE statistics are as follows:

wherein

Is the control limit for the SPE statistic,

λ_j(j ═ 1, 2., k) is the eigenvalue corresponding to each principal component covariance matrix, k is the number of principal components, F_k.(n-k),αIs the upper limit value of F distribution with confidence coefficient of α and degree of freedom of k and (n-k), c_αCorresponding to the lower limit value of the normal distribution under the confidence coefficient α, determining I by adopting a kernel density estimation method²Control limit of statistic, determining T based on F distribution²A control limit for the statistic;

3) intelligent decision making: obtaining the membrane pollution evaluation index u (t) at the time t on line, and monitoring tStatistical amount of etching I²、T²In order to determine the main factors of membrane pollution, taking measures aiming at the determined factors and establishing a multi-classifier based on a kernel function to distinguish the membrane pollution factors, wherein the steps mainly comprise that ① produced water flow reaches the peak value of 460m³H, ② produced water pressure is lower than 20kp, ③ aeration rate is lower than 2400m³④ the sludge concentration is more than 13000mg/l, ⑤ the water production pressure is less than 20kp and the water permeability is less than 30LMH/bar, ⑥ the water production pressure is less than 20kp and the water permeability is less than 60LMH/bar, the turbidity of the produced water is more than 5NUT, 6 binary classifiers are combined into a multi-classifier, for the first binary classifier, the data not belonging to the first class is labeled as-1, the data belonging to the first class is labeled as +1, for the second binary classifier, the data not belonging to the second class is labeled as-1, the data belonging to the second class is labeled as +1, for the third binary classifier, the data not belonging to the third class is labeled as-1, the data belonging to the third class is labeled as +1, for the fourth binary classifier, the data not belonging to the fourth class is labeled as-1, the data belonging to the fourth class is labeled as +1, for the fifth binary classifier, the data not belonging to the fifth class is labeled as-1, the fifth binary classifier, the data belonging to the sixth class is labeled as +1, the sixth classifier, the q label is labeled as +1, the sixth classifier, the target function of the q 2, the sixth classifier, the target function of the target classifier, the target function of the data of the following steps:

wherein C ∈ (0,5) is a penalty coefficient, w^qIs the connection weight of the q classifier, T is the transpose symbol, ξ_p ^qRelaxation variables for the p-th data sample of the q-th classifier, b^qFor the bias of the q-th classifier,

is a Gaussian kernel function, σ ∈ (0,1) is the width of the kernel function, u_p ^qP-th data sample, u, of the q-th classifier, respectively_v ^qIs the v-th data sample of the q-th classifier (v ═ 1,2, …, n), y_pIs the category of the p sample; introducing a Lagrange multiplier, and converting the optimization problem into:

updating parameters:

wherein, y_vClass of the v-th sample, α_pThe Grenarian multiplier for the p-th data sample, α_vFor the evaluation index u, the lagrangian multiplier for the v-th data sample is as follows:

f(u)＝sgn(α_py_pK(u,u_p ^q)+b^q) (15)

wherein sgn represents a sign function, the input is positive, the output is 1, the input is negative, the output is-1, the output value can be obtained by inputting the sample to be tested into the decision function, the category corresponding to the function for obtaining the maximum value is the category to which the sample to be tested belongs, the 6-category corresponding operation suggestion comprises ①, the operation in the state is not suitable for exceeding 4h, and the water yield is reduced to 260m³Below/h, ② reducing the water yield of the membrane pool to 260m³Controlling transmembrane pressure difference to be less than 40kPa, ③ increasing aeration to 3000m³④ controlling the sludge concentration of the membrane pool to be 8000- & lt12000 mg/L, ⑤ adjusting the operation parameters, reducing the water yield to 260m³H, or increasing aeration amount to 5000m³And ⑥, performing online physical cleaning and online chemical cleaning within 24 hours, applying the trained decision model to the sewage treatment process, performing characteristic matching on the model for fault data, outputting fault types, performing operation suggestions corresponding to the fault types, and providing decision support for the production process.

Drawings

FIG. 1 is a diagram of a deep belief network architecture;

FIG. 2 is a diagram of 10-step prediction results of a water outlet water permeability soft measurement model, wherein black line circles are actual water permeability values, and black line points are predicted values of a deep belief network soft measurement model;

FIG. 3 is a diagram showing the MBR membrane contamination early warning result, wherein (a) is a non-Gaussian statistic I²Changing result, wherein the black line points are statistics change curve, the black line is control limit, and the graph (b) is Gaussian statistic T²Changing results, wherein black line band points are a statistic change curve, black lines are control limits, and a graph (c) is a non-Gaussian statistic SPE changing result, wherein the black line band points are the statistic change curve, and the black lines are the control limits;

FIG. 4 is a plot of MBR membrane fouling decision results, wherein black circles are the decision model outputs and black stars are the actual decision suggestion categories;

FIG. 5 is a data flow direction indicating diagram of the MBR membrane pollution intelligent early warning system;

Detailed Description

(1) Specific implementation of membrane pollution intelligent decision system design and software and hardware function integration

A hardware platform environment set up in an actual sewage treatment plant is shown in figure 2, operation process data are collected through a collection instrument installed in a process field, the data collected by the instrument are transmitted to a PLC through a Modbus communication protocol, the PLC transmits the operation process data to an upper computer through an RS232 communication protocol, the data in the upper computer are transmitted to a data processing Server through a local area network, the operation process data are issued to a water plant work manager through a Web Server in a Browser/Server mode, the prediction of water permeability and the early warning result of membrane pollution are displayed in a Client/Server mode, the main functions of the developed MBR membrane pollution intelligent early warning system can be achieved by ① inquiring operation parameters of a membrane pool, ② online prediction of water permeability, ③ early warning of membrane pollution, and ④ intelligent decision is made on membrane pollution.

The invention adopts component technology in software industry to package the membrane pollution data preprocessing module, the membrane pollution intelligent prediction module, the membrane pollution intelligent early warning module and the membrane pollution intelligent decision-making module into functional modules, thereby enhancing the reusability of the model and making up the blank of popularization of MBR membrane pollution intelligent decision-making technology to human-computer interaction interface in actual system operation at home and abroad; the NET platform is adopted for software development, so that an ActiveX control is conveniently created, and the usable environment range of software is expanded; a field bus technology is adopted to establish a full-flow system communication network to realize information transmission among modules; meanwhile, the MBR membrane pollution intelligent decision-making system provided by the invention realizes the connection of a central control room and each on-site data acquisition point, forms a centralized management early warning system, is easy to expand, has independent functions of each part, can add software and hardware modules according to actual prediction requirements and is fused with other systems, can realize the stability and reliability of the system and ensures the early warning precision of membrane pollution.

(2) Specific implementation of film pollution intelligent decision method research

The invention obtains an MBR membrane pollution intelligent decision-making method; the method 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 a deep belief network, intelligent detection of MBR membrane water permeability is realized, a membrane pollution comprehensive evaluation model is established by combining a predicted value of the membrane water permeability with other collectable process variables of a water plant, identification of membrane pollution and identification of main factors are realized, the intelligent level of the sewage treatment plant is improved, and normal operation of a sewage treatment process is guaranteed.

① input variables were collected by on-line instrumentation installed at the process site, including 5 variables to be collected, and the parameter information and collection location are shown in Table 1.

TABLE 1 Process variable types collected

② A soft measurement model is built by using the deep belief network, 510-step prediction of membrane water permeability is realized based on a multi-step prediction strategy, the deep belief network is trained and tested by adopting data collected in real time, 80 groups of data are selected for testing, and the collected data are shown in a table 2.

③ comprehensive evaluation of membrane pollution is carried out by using the predicted value of water permeability and the current values of other related collection variables (water production flow, water production pressure, aeration quantity and sludge concentration) so as to judge whether membrane pollution occurs.

④ making operation decision of membrane pollution according to relevant variable of membrane pollution.

TABLE 2 Soft measurement model test data

Claims

1. Membrane bioreactor-MBR membrane pollution intelligent decision-making method, is characterized in that, comprises the following steps:

(1) Operation process data collection: Take the MBR membrane treatment system as the research object, collect the operation process data through the collection instrument installed on the process site, including: influent chemical oxygen demand, influent pH, influent biological oxygen demand, Effluent chemical oxygen demand, effluent pH, effluent biological oxygen demand, anaerobic zone redox potential, anoxic zone redox potential, aerobic zone nitrate, aerobic zone dissolved oxygen, produced water pressure, produced water turbidity, Produced water flow, sludge concentration, aeration volume, to achieve data collection; data collected by the instrument is transmitted to the programmable logic controller through the communication protocol, and the programmable logic controller transmits the operation process data to the upper computer through the communication protocol. The data in the computer is transmitted to the data processing server through the local area network;

(2) Operation process data preprocessing: Taking the membrane tank operation data as the research object, the partial least squares method is used to establish a characteristic analysis model, and five principal component variables are obtained, namely: permeate pressure, permeate turbidity, permeate flow , sludge concentration, aeration volume, 5 principal component variables are used as the input variables of the membrane fouling intelligent prediction model block; water permeability is used as the output variable of the membrane fouling intelligent prediction module;

(3) Intelligent prediction of membrane fouling: The intelligent prediction of membrane fouling realizes the prediction of water permeability, in which: the water permeability is obtained by the prediction of the deep belief network-DBN. The DBN consists of 1 input layer, 2 hidden layers, and 1 output layer. There are 5 neurons in the layer, M neurons in each hidden layer, M is a positive integer greater than 2 and less than 30, and there is 1 neuron in the output layer, that is, the connection method is 5-MM-1; n groups of data As the training sample of the soft sensor model; the DBN input at time t is x(t)=[x ₁ (t),...,x ₅ (t)], x ₁ (t) represents the value of the water production pressure at time t, x ₂ (t) represents the value of produced water turbidity at time t, x ₃ (t) represents the value of produced water flow rate at time t, x ₄ (t) represents the value of sludge concentration at time t, and x ₅ (t) represents time t The value of the aeration amount, the calculation method of the soft sensing model based on DBN to predict the water permeability is as follows:

①Input layer: h ₀ (t)=x(t) (1)

②The first hidden layer:

③ The second hidden layer:

④Output layer:

Among them, h ₀ (t) represents the output vector of the input layer at time t, h ₁ (t) represents the output vector of the first hidden layer at time t, h ₂ (t) represents the output vector of the second hidden layer at time t, w _0,1 (t) represents the weight vector between the input layer and the first hidden layer at time t, and w _1,2 (t) represents the weight between the first hidden layer and the second hidden layer at time t value vector, w _2,3 (t) represents the weight vector between the second hidden layer and the output layer at time t, b ₁ (t) represents the bias vector of the first hidden layer at time t, b ₂ (t ) represents the bias vector of the second hidden layer at time t, and y(t) represents the actual output of the DBN at time t;

DBN training is divided into two processes: unsupervised pre-training and supervised weight fine-tuning; set the number of iterations of each layer of pre-training to 100, the number of iterations of the backpropagation algorithm to be 1000, the expected error to be 0.01, the initial weights and The bias is set to 0.01; the specific training steps are as follows:

①Unsupervised pre-training: The update rules for parameters obtained by the contrast divergence algorithm are:

Among them, w _0,1 (t+1) is the weight vector between the input layer and the first hidden layer at time t+1, and w _1,2 (t+1) is the first hidden layer at time t+1 The weight vector between the second hidden layer and the second hidden layer, b ₁ (t+1) is the bias vector of the first hidden layer at the time of t+1, and b ₂ (t+1) is the second hidden layer at the time of t+1. The bias vector of the containing layer, E _data is the expectation of the training data, E _model is the expectation defined by the model, μ _w1 ∈ (0, 0.02) is the learning rate of the connection weight between the input layer and the first hidden layer, μ _w2 ∈ (0, 0.02) is the learning rate of the connection weight between the first hidden layer and the second hidden layer, μ _b1 ∈ (0, 0.01) is the learning rate of the first hidden layer, μ _b2 ∈ (0, 0.01) is the learning rate of the second hidden layer;

②Adjust the weights by the back-propagation algorithm: The initial value of the parameters of the DBN is obtained by layer-by-layer training, and then the back-propagation algorithm is used to fine-tune the weights to obtain a better model effect. The error back-propagation method is used to adjust the weights to:

Among them, w _2,3 (t+1) is the weight vector between the second hidden layer and the output layer at time t+1, y _d (t) is the expected output of DBN at time t, η _out ∈(0, 3) is the weight learning rate between the second hidden layer and the output layer, η ₂ ∈(0,3) is the weight learning rate between the first hidden layer and the second hidden layer, η ₁ ∈ (0,3) is the weight learning rate between the input layer and the first hidden layer;

(4) Membrane fouling intelligent decision-making: Membrane fouling intelligent decision-making is to use the predicted value of water permeability, combined with other membrane fouling-related variables, to establish a comprehensive evaluation model of membrane fouling, and provide decision-making for membrane fouling. The process is as follows:

1) Determine the evaluation index of membrane fouling, u={u ₁ , u ₂ , u ₃ , u ₄ , u ₅ , u ₆ }, u ₁ is the value of permeate flow, u ₂ is the value of permeate pressure, u ₃ is the value of produced water turbidity, u ₄ is the value of sludge concentration, u ₅ is the value of aeration rate and u ₆ is the predicted value of water permeability;

2) Establish monitoring statistics, the process variable is u∈R ^6×n , where 6 is the dimension of the variable, and n is the number of samples; perform independent component and principal component decomposition on the membrane fouling evaluation matrix u:

u=As+Pt ^T +f ^T (7)

Among them, A∈R ^6×r is the mixed matrix, s∈R ^r×n is the independent matrix, t∈R ^n×k is the main element score matrix, the load matrix P∈R ⁶ ^×k , f∈R ^n×6 is The final residual matrix, r represents the number of independent elements, and k is the number of selected pivot elements; in order to estimate A and s, it is necessary to obtain the unmixing matrix W to obtain the reconstructed signal

is the estimated value of the source signal, whitening u to obtain the component vector:

Z = Λ ^-1/2 H ^T u = Bs (8)

where Λ is a diagonal matrix containing all eigenvalues, H is the corresponding eigenvector matrix and is unit-orthogonal, and B is a unit-orthogonal matrix; the unmixing matrix is:

W=B ^T Λ ^-1/2 H ^T (9)

After the original data is decomposed by independent components and principal components, the independent component matrix and the principal component score matrix are obtained, the I ² and T ² statistics are established, and the SPE statistics are established for the residual information; the independent component vector is s _l =[s _1l , s _2l ,…,s _rl ] ^T ∈ R ^r×1 and pivot score vector t _l =[t _l1 ,t _l2 ,…,t _lk ] ^T ∈ R ^k×1 , residual vector f _l =[f _l1 ,f _l2 ,…,f _l6 ] ^T ∈R ^6×1 , l=1,2,…,n (n is the number of samples); establish I ² , T ² and SPE statistics as follows:

in

is the control limit of the SPE statistic,

λ _j is the eigenvalue corresponding to each principal component covariance matrix, where j=1, 2, .., k; k is the number of principal elements, F _{k.(nk), α} is the confidence degree α, the degree of freedom is the upper limit of the F distribution of k and (nk); c _α corresponds to the lower limit of the normal distribution under the confidence α, the kernel density estimation method is used to determine the control limit of the I ² statistic, and T is determined based on the F distribution ² the control limit of the statistic;

3) Intelligent decision-making: obtain the evaluation index u(t) of membrane fouling at time t online, when monitoring statistics I ² , T ² and SPE at time t exceed the control limit, determine the occurrence of membrane fouling; in order to determine the main factors of membrane fouling , take measures to determine the factors, establish a multi-classifier based on kernel function to distinguish membrane fouling factors, mainly including: ① the permeate flow reaches a peak value of 460m ³ /h, ② the permeate pressure is lower than 20kp, ③ the aeration volume is lower than 2400m ³ /h, ④ sludge concentration greater than 13000mg/l, ⑤ product water pressure less than 20kp and water permeability less than 30LMH/bar, ⑥ product water pressure less than 20kp and water permeability less than 60LMH/bar, product water turbidity Greater than 5NUT; 6 binary classifiers are combined into a multi-classifier, for the first binary classifier, the data label that does not belong to the first class is -1, and the data label that belongs to the first class is +1, For the second binary classifier, label the data that does not belong to the second class as -1 and the data that belongs to the second class as +1, and for the third binary classifier, label the data that does not belong to the third class The label is -1, the data label belonging to the third class is +1, for the fourth binary classifier, the data label that does not belong to the fourth class is -1, the data label belonging to the fourth class is +1, for For the fifth binary classifier, the data label that does not belong to the fifth class is -1, and the data label that belongs to the fifth class is +1. For the sixth binary classifier, the data label that does not belong to the sixth class is is -1, the data label belonging to the sixth category is +1, and the optimization objective function of the qth classifier (q=1,2,...,6) is:

Among them, C∈(0,5) is the penalty coefficient, wq is the connection weight of the ^qth classifier, T is the transpose symbol, and ξpq is the slack variable of the _pth data sample of the ^qth classifier , b ^q is the bias of the qth classifier,

is the Gaussian kernel function, σ∈(0,1) is the width of the kernel function, u _p ^q are the p-th data sample of the q-th classifier, and u _v ^q are the v-th data sample of the q-th classifier (v=1,2,...,n), y _p is the category of the p-th sample; Lagrange multipliers are introduced, and the optimization problem is transformed into:

Parameter update:

Among them, y _v is the category of the v-th sample, α _p is the Grange multiplier of the p-th data sample, α _v is the Lagrangian multiplier of the v-th data sample, and for the evaluation index u, the decision function for:

f(u)=sgn(α _p y _p K(u,u _p ^q )+b ^q ) (15)

Among them, sgn represents the sign function, when the input is positive, the output is 1, and when the input is negative, the output is -1; the output value can be obtained by inputting the sample to be tested into this decision function, and the category corresponding to the function that obtains the maximum value is It is the category of the sample to be tested; the corresponding operation suggestions for the 6 categories include: ① It should not run for more than 4 hours in this state, reduce the water production to below 260m ³ /h, and ② reduce the membrane tank water production to 260 m ³ /h and control the transmembrane pressure difference Less than 40kPa, ③ increase the aeration to 3000m ³ /h, ④ control the sludge concentration of the membrane tank to 8000-12000mg/L, ⑤ adjust the operating parameters, reduce the water production to 260m ³ /h, or increase the aeration volume to 5000m ³ /h or more, ⑥ Carry out online physical cleaning and online chemical cleaning within 24 hours; use the trained decision-making model for the sewage treatment process, and for fault data, the model performs feature matching, outputs the fault category, and provides operation suggestions corresponding to the fault category, which are The production process provides decision support.