Fault identifying method for sludge bulking based on a recurrent RBF neural network

Abstract

The wastewater treatment process by using activated sludge process often appear the sludge bulking fault phenomenon. Due to production conditions of wastewater treatment process, the correlation and restriction between variables, the characteristics of nonlinear and time-varying, which lead to hard identification of sludge bulking; Sludge bulking is not easy to detect and the reasons resulting in the sludge bulking are difficult to identify, are current RBF neural network is designed for detecting and identifying the causes of sludge volume index (SVI) in this patent. The method builds soft-computing model of SVI based on recurrent RBF neural network, it has been completed to the real-time prediction of SVI concentration and better accuracy were obtained. Once the fault of sludge bulking is detected, the identifying cause variables (CVI) algorithm can find the cause variables of sludge bulking. The method can effectively identify the fault of sludge bulking and ensure the safety operation of the wastewater treatment process.

Description

CROSS REFERENCE TO RELATED PATENT APPLICATIONS

This application claims priority to Chinese Patent Application No. 201710186738.5, filed on Mar. 27, 2017, entitled “an fault identifying method for sludge bulking based on a recurrent RBF neural network,” which is hereby incorporated by reference in its entirety.

TECHNICAL FIELD

In this present disclosure, an identifying method is designed for detecting the sludge bulking and find the fault reasons of sludge bulking in the urban wastewater treatment process (WWTP) by a recurrent RBF neural network. The sludge volume index-SVI is an important parameter of characterization of sludge bulking. The basic link of proposed identifying method is to predict SVI on the basis of the relationships between variables. The technology of this present disclosure is part of advanced manufacturing technology, belongs to both the field of control engineer and environment engineer. Therefore, the identifying method for detecting the sludge bulking and find the fault reasons of sludge bulking in WWTP is of great significance.

BACKGROUND

The urban WWTP also happen the sludge bulking in the activated sludge process. However, the various influencing factors for SVI are various and complex. Therefore, it is difficult to identify the cause variables for sludge bulking, which seriously affected the stable operation of the urban WWTP. The identifying method for sludge bulking, based on recurrent RBF neural network, is helpful to detect the fault phenomenon of sludge bulking and identify the cause variables that resulted in the sludge bulking, which strengthen delicacy management, ensure water quality effluent standards of urban WWTP. It has better economic benefit as well as significant environmental and social benefits. Thus, the research achievements have wide application prospect in this present disclosure.

At present, the activated sludge process has widely used in the urban wastewater treatment process of papermaking, printing and dyeing, chemical industry, and many other industrial wastewater. However, sludge bulking fault problem has always been a thorny problems existing in the activated sludge system. Due to the sludge and water can't normal separate, cause the failure of wastewater treatment process. Sludge bulking occurs frequently and it has basically different degrees of sludge bulking in the wastewater treatment process. High sludge bulking coverage, in many countries such as Germany, Britain, South Africa's wastewater treatment plants according to the survey, more than half of the wastewater treatment plants exist the situation of the excessive growth of filamentous bacteria. Thus, sludge bulking is a common problem in wastewater treatment plants at home and abroad. Many scholars of most countries have studied the prevention and control method of the sludge bulking. Although some progress has been made, but so far, there is no effective control measures of sludge bulking; Moreover, in the event of the sludge bulking, the reason is not easy to explore, and need longer time to treat the failure of wastewater treatment. To sum up, once in the event of sludge bulking, consequences cannot be ignored. For this failure phenomenon of sludge bulking, therefore, early diagnosis and prevention is the most effective methods to solve the problem of sludge bulking, so it has high practical significance.

In this present disclosure, an identifying method for sludge bulking, is presented by building a soft-computing model based on recurrent RBF neural network. The neural network uses fast gradient descent algorithm to ensure the accuracy of recurrent RBF neural network. Once the sludge bulking is detected, an identifying cause variables (CVI) algorithm will be exploited to implement the identification of fault variables. This method can effectively prevent the happen of sludge bulking and reduce the loss of the wastewater treatment plant.

SUMMARY

A fault identification method is designed for the sludge bulking based on a recurrent RBF neural network. Its characteristic and steps include following steps:

(1) Determine the Input and Output Variables of SVI:

For sewage treatment process of activated sludge system, by analyzing the detailed mechanism of sludge bulking, five process variables are analyzed and select the input variables of SVI soft-computing model: dissolved oxygen concentration-DO, mixed liquor suspended solids concentration-MLSS, temperature-T, chemical oxygen demand-COD and total nitrogen-TN. The output value of soft-computing model is detected SVI concentration.

(2) Initial Recurrent RBF Neural Network:

The structure of recurrent RBF neural network comprise three layers: input layer, hidden layer and output layer. The network is 5-J−1, named the number of input layer is 5 and hidden neurons is J. Connection weights between input layer and hidden layer are assigned 1, the connection weights between hidden layer and output layer randomly assign values, the assignment interal is [1, 1]. The number of the training sample is N and the input of recurrent RBF neural network is x(t)=[x₁(t), x₂(t), x₃(t), x₄(t), x₅(t)] at time t. The expectations output of neural network output is expressed as y_d(t) and the actual output is expressed as y(t). Soft-computing method of SVI can be described:

{circle around (1)} The input Layer: There are 5 neurons which represent the input variables in this layer. The output values of each neuron are as follows:u_i(t)=x_i(t)  (1)wherein u_i(t) is the ith output value at time t, i=1, 2, . . . , 5, and the input vector is x(t)=[x₁(t), x₂(t), . . . , x₅(t)].

{circle around (2)} The Hidden Layer: There are J neurons of hidden layer. The outputs of hidden neurons are:

$\begin{array}{cc}{\theta }_j\left(t\right)=e^{-\frac{{h_j\left(t\right)-c_j\left(t\right)}^2}{2{\sigma }_j^2\left(t\right)}},j=1,2,\dots ,J& \left(2\right)\end{array}$ c_j(t) denotes the center vector of the jth hidden neuron and c_j(t)=[c_j1(t), c_j2(t), . . . , c_jn+1(t)]^T at time t, ∥h_j(t)−c_j(t)∥ is the Euclidean distance between h_j(t) and c_j(t), and σ_j(t) is the radius or width of the jth hidden neuron at time t, h_j(t) is input vector of the jth hidden neuron at time t described ash_j(t)=[u₁(t),u₂(t), . . . u₅(t),v_j(t)×y(t−1)]^T  (3)y(t−1) is the output value of the output layer at time t−1, 1), v_j(t) denotes the connection weight from output layer to the jth hidden neuron at time t, and v(t)=[v₁(t), v₂(t), . . . , v_J(t)]^T, T represents transpose.

{circle around (3)} The Output Layer: There is only one node in this layer, the output is:

$\begin{array}{cc}y\left(t\right)=f\left(w\left(t\right),\theta \left(t\right)\right)=\sum _{j=1}^J{w}_j\left(t\right)\times {\theta }_j\left(t\right),j=1,L,J& \left(4\right)\end{array}$ wherein w(t)=[w₁(t), w₂(t), . . . , w_J(t)]^T is the connection weights between the hidden neurons and output neuron at time t, θ(t)=[θ₁(t), θ₂(t), . . . , θ_J(t)]^T is the output vector of the hidden layer, y(t) represents the output of recurrent RBF neural network at time t.

The error of recurrent RBF neural network is:

$\begin{array}{cc}E\left(t\right)=\frac{1}{N}\sum _{t=1}^N{\left(y_d\left(t\right)-y\left(t\right)\right)}^2& \left(5\right)\end{array}$ y_d(t) is the expectation output of neural network and the actual output is expressed as y(t).

(3) Train Recurrent RBF Neural Network:

{circle around (1)} Given the recurrent RBF neural network, the initial number of hidden layer neurons is J, J>2 is a positive integer. The input of recurrent RBF neural network is x(1), x(2), . . . , x(t), . . . , x(N), correspondingly, the expectation output is y_d(1), y_d(2), . . . , y_d(t), . . . , y_d(N), expected error value is set to E_d, E_d∈(0, 0.01). The every variable of initial centre value c₁(1)∈(−2, 2), width value σ_j(1)∈(0, 1), initial feedback weight v_j(1)∈(0, 1), j=1, 2, . . . , J. Initial weight w(1)∈(0, 1).

{circle around (2)} Set the learning step s=1;

{circle around (3)} t=s; According to Eqs. (1)-(4), calculate the output of recurrent RBF neural network, exploiting fast gradient descent algorithm:

$\begin{array}{cc}{c}_j\left(t+1\right)=c_j\left(t\right)-{\eta }_c\frac{1}{{\sigma }_j^2}\left(y_d\left(t\right)-y\left(t\right)\right)w_j\left(t\right)\times \theta \left(t\right)\left[h_j\left(t\right)-c_j\left(t\right)\right]& \left(6\right)\\ {\sigma }_j\left(t+1\right)={\sigma }_j\left(t\right)-{\eta }_{\sigma }\frac{1}{{\sigma }_j^3}\left(y_d\left(t\right)-y\left(t\right)\right)w_j\left(t\right)\times \theta \left(t\right){h_j\left(t\right)-c_j\left(t\right)}^2& \left(7\right)\\ v_j\left(t+1\right)=v_j\left(t\right)-{\eta }_v\left(y_d\left(t\right)-y\left(t\right)\right)w_j\left(t\right)\theta \left(t\right)y\left(t-1\right)& \left(8\right)\\ w_j\left(t+1\right)=w_j\left(t\right)-{\eta }_w\left(y_d\left(t\right)-y\left(t\right)\right){\theta }_j\left(t\right)& \left(9\right)\end{array}$ η_c, η_σ, η_v, η_w are the learning rate of centre, width, feedback connection weight from output layer to hidden layer and the connection weight between hidden layer and output layer, respectively. In addition, η_c∈(0, 0.01], η_σ∈(0, 0.01], η_v ∈(0, 0.02], η_w∈(0, 0.01]. c_j(t+1)=[c_1j(t+1), c_2j (t+1), . . . , c_5j(t+1)] denotes the center vector of the jth hidden neuron at time t+1. σ_j(t+1) is the radius or width of the jth hidden neuron at time t+1. v_j(t+1) denotes the connection weight from output layer to the jth hidden neuron at time t+1. w_j(t+1) is the connection weights between the hidden neurons and output neuron at time t+1.

{circle around (4)} increase 1 learning step s, if s<N, then turn to step {circle around (3)}. If s=N, turn to step {circle around (5)}.

{circle around (5)} according to Eq. (5), calculate the performance of recurrent RBF neural network. If E(t)≥E_d, then turn to step {circle around (3)}; If E(t)<E_d, stop the training process.

(4) SVI Concentration Prediction:

The testing samples used as the input of recurrent RBF neural network, the output of neural network is the soft-computing values of SVI.

(5) CVI Algorithm for Sludge Bulking:

{circle around (1)} Calculate the residual of the expectation output and the output of recurrent RBF neural network, if

$\begin{array}{cc}\{\begin{array}{c}y\left(t\right)-y_d\left(t\right)\ge 5\\ y\left(t\right)\ge 150\end{array}& \left(10\right)\end{array}$ then turn to step {circle around (2)}, otherwise, stop the process of fault identification for sludge bulking.

{circle around (2)} Define two formula:

$\begin{array}{cc}\{\begin{array}{c}{\mathrm{IC}}_1\left(t\right)={\delta }^T\left(t\right){\Lambda }_M^{-1}\left(t\right)\delta \left(t\right)\\ {\mathrm{IC}}_2\left(t\right)={{\theta }_M\left(t\right)}^2-{\delta }^T\left(t\right)\delta \left(t\right)\end{array}& \left(11\right)\end{array}$ wherein IC₁(t) is the Mahalanobis distance of input variables at time t and IC₂(t) is the squared prediction error at time t, M is the numbers of principal component of input data, θ_M(t) is the output vector of the hidden layer for M principal components training data at time t, K is the number of the front samples. δ(t)=[δ₁(t), . . . , δ_m(t), . . . , δ_M(y)]^T is the projection of the training data, and δ_m(t) is

$\begin{array}{cc}{\delta }_m\left(t\right)=\sum _{k=1}^K{a}_k\left(t\right)\left({\stackrel{\_}{\theta }}_m\left(t\right)·{\stackrel{\_}{\theta }}_{\mathrm{new}}\left(t\right)\right),k=1,L,K;m=1,\dots ,M& \left(12\right)\end{array}$ wherein θ_new(t) is the mean-centered output vector of the hidden layer for M principal components training data at time t, θ_m(t) is the output vector of the hidden layer for the mth principal component training data at time t, a_k(t) is a constant, and a_k(t)∈(0, 0.01]. And the diagonal matrix of eigenvalues associated with M principal components is defined as

$\begin{array}{cc}{\Lambda }_M\left(t\right)=\left[\begin{array}{cccc}{\lambda }_1& & & \\ & {\lambda }_2& & \\ & & O& \\ & & & {\lambda }_M\end{array}\right]\left({\lambda }_1\ge {\lambda }_2\ge L\ge {\lambda }_M\ge 0\right)& \left(13\right)\end{array}$ wherein Λ_M(t) is the diagonal matrix of eigenvalues at time t and it satisfies{tilde over (C)}(t)=Z(t)Λ_M(t)Z^T(t)+l′(t)(I(t)−Z(t)Z^T(t))  (14)wherein l′(t) is a constant value, I(t) is a unit matrix, {tilde over (C)}(t) is the regularized covariance matrix of C(t) at time t:

$\begin{array}{cc}\Omega \left(t\right)={\theta \left(t\right)}^T{\stackrel{\sim }{C}}^{-1}\left(t\right)\theta \left(t\right)& \left(15\right)\\ \Omega \left(t\right)={\mathrm{IC}}_1\left(t\right)+l^{\prime -1}\left(t\right){\mathrm{IC}}_2\left(t\right)& \left(16\right)\\ C\left(t\right)=\frac{1}{K}\sum _{k=1}^K{\theta }_k\left(t\right){{\theta }_k\left(t\right)}^T& \left(17\right)\end{array}$ wherein Ω(t) is the energy of each variable. θ_k(t) is the hidden output vector of the kth principal component at time t.

The constant vector a(t)=[a₁(t), . . . a_k(t), . . . a_K(t)]^T is given as

$\begin{array}{cc}\lambda \left(t\right)a\left(t\right)=\frac{1}{K}G\left(t\right)a\left(t\right)& \left(18\right)\end{array}$ G(t) is the Gaussian matrix and λ(t) denotes the eigenvalueG(t)={θ_i(t)·θ_j(t)}_K×K  (19)λ(t)P(t)=C(t)p(t)  (20)p(t) denotes eigenvector of the covariance matrix C(t) at time t.

{circle around (3)} For the ith input variable, the contribution degree index satisfy:

$\begin{array}{cc}{G}^i\left(t\right)=\frac{{\kappa }^i\left(t\right)}{\sum _{i=1}^5{\kappa }^i\left(t\right)}& \left(21\right)\end{array}$ wherein Gⁱ(t) is the contribution degree index of ith variable at time t, a hidden layer neuron corresponds to an input variable, κⁱ(t) is the contribution degree which is calculate by the mutual information between this testing samples and difference sets for testing samples xⁱ(t) of ith variable at time t, which can be expressed asκⁱ(t)=I(xⁱ(t),V_Δ(t))  (22)I(xⁱ(t), V_Δ(t)) is the mutual information of xⁱ(t) and V_Δ(t) at time t, V_Δ(t) is the difference matrix sets of data of training set and testing set at time t, which is expressed asV_Δ(t)=V_tr(t)−V_te(t)  (23)wherein V_tr(t) and V_te(t) is independent data sets of training set and testing set at time t, respectively.V_tr(t)=D⁻¹(t)G(t)  (24)V_te(t)=D_te⁻¹(t)G_te(t)  (25)D(t) is the covariance matrix of Φ(t) at time tD(t)=E{Φ(t)Φ^T(t)}  (26)Φ(t)=[θ(t−K+1), . . . ,θ(t−1),θ(t−1),θ(t)]^T  (27)Φ(t) is output matrix of hidden layer at time t, θ(t−K+1) is the hidden output vector at time t−K+1, K is the number of the front samples.

{circle around (4)} For the ith input variable, a hidden layer neuron corresponds to an input variable, if the Gⁱ(t) at time t satisfies:G¹(t)+ . . . Gⁱ(t)≥0.8  (28)

the variables 1, . . . , i is the cause variables resulted in sludge bulking.

The Novelties of this Present Disclosure Contain:

(1) To detect the sludge bulking and identify the cause variables that resulted in the sludge bulking, an identifying method for sludge bulking is developed in this present disclosure. The results demonstrate that the SVI trends in WWTP can be predicted with acceptable accuracy using the DO, MLSS, T, TN and COD as input variables. This method can not only solve the problem of measured online for SVI concentration with acceptable accuracy but also detect the happen of sludge bulking.

(2) This identifying fault variables method is based on the CVI algorithm. And it identify the fault variables of sludge bulking in the WWTP with high identifying precision. Thus, it can realize the effective regulation of the sludge bulking control in advance.

Attention: this present disclosure utilizes five input variables in this identifying method to predict the SVI. In fact, it is in the scope of this present disclosure that any of the variables: DO, T, MLSS, COD and TN, are used to predict the SVI concentration. Moreover, this identifying method is also able to predict the others variables in urban WWTP.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows the structure of identifying method based on the recurrent RBF neural network in this present disclosure.

FIG. 2 shows the testing result of the identifying method.

FIG. 3 shows the testing error of the identifying method.

FIG. 4 shows the IC₁(t) outputs of the identifying method.

FIG. 5 shows the IC₂(t) outputs of the identifying method.

FIG. 6 shows the marked fault points of IC₁(t) and IC₂(t) outputs of the identifying method.

FIG. 7 shows contribution degree index of input variables of identifying method.

DETAILED DESCRIPTION

This invention takes MLSS, DO, T, COD and TN as characteristic variables for SVI, the above unit is mg/L;

The experimental data comes from water quality analysis statement of a wastewater treatment plant in 2014; choosing data of MLSS, DO, T, COD and TN as experimental samples, after eliminating abnormal sample, 100 groups of data are available, and the group of 60 used as training samples, the remaining 40 groups as test samples.

This present disclosure adopts the following technical scheme and implementation steps:

A fault identification method is designed for the sludge bulking based on a recurrent RBF neural network. Its characteristic and steps include following steps:

(1) Determine the Input and Output Variables of SVI:

For sewage treatment process of activated sludge system, by analyzing the detailed mechanism of sludge bulking, five process variables are analyzed and select the input variables of SVI soft-computing model: dissolved oxygen concentration-DO, mixed liquor suspended solids concentration-MLSS, temperature-T, chemical oxygen demand-COD and total nitrogen-TN. The output value of soft-computing model is detected SVI concentration.

(2) Initial Recurrent RBF Neural Network:

The structure of recurrent RBF neural network comprise three layers: input layer, hidden layer and output layer in FIG. 1. The network is 5-5-1, named the number of input layer is 5 and hidden neurons is 5. Connection weights between input layer and hidden layer are assigned 1, the connection weights between hidden layer and output layer randomly assign values, the assignment interal is [1, 1]. The number of the training sample is N and the input of recurrent RBF neural network is x(t)=[x₁(t), x₂(t), x₃(t), x₄(t), x₅(t)] at time t. The expectations output of neural network output is expressed as y_d(t) and the actual output is expressed as y(t). Soft-computing method of SVI can be described:

{circle around (1)} The input Layer: There are 5 neurons which represent the input variables in this layer. The output values of each neuron are as follows:u_i(1)=x_i(t)  (29)wherein u_i(t) is the ith output value at time t, i=1, 2, . . . , 5, and the input vector is x(t)=[x₁(t), x₂(t), . . . , x₅(t)].

{circle around (2)} The Hidden Layer: There are J neurons of hidden layer. The outputs of hidden neurons are:

$\begin{array}{cc}{\theta }_j\left(t\right)=e^{-\frac{{h_j\left(t\right)-c_j\left(t\right)}^2}{2{\sigma }_j^2\left(t\right)}},j=1,2,\dots ,J& \left(30\right)\end{array}$ c_j (t) denotes the center vector of the jth hidden neuron and c_j(t)=[c_j1(t), c_j2(t), . . . , c_jn+1(t)]^T at time t, ∥h_j(t)−c_j(t)∥ is the Euclidean distance between h_j(t) and c_j(t), and σ_j(t) is the radius or width of the jth hidden neuron at time t, h_j(t) is input vector of the jth hidden neuron at time t described ash_j(t)=[u₁(t),u₂(t), . . . u₅(t),v_j(t)×y(t−1)]^T  (31)y(t−1) is the output value of the output layer at time t−1, v_j(t) denotes the connection weight from output layer to the jth hidden neuron at time t, and v(t)=[v₁(t), v₂(t), . . . , v_J(t)]^T, T represents transpose.

{circle around (3)} The Output Layer There is only one node in this layer, the output is:

$\begin{array}{cc}y\left(t\right)=f\left(w\left(t\right),\theta \left(t\right)\right)=\sum _{j=1}^J{w}_j\left(t\right)\times {\theta }_j\left(t\right),j=1,\dots ,J& \left(32\right)\end{array}$ wherein w(t)=[w₁(t), w₂(t), . . . , w_J(t)]^T is the connection weights between the hidden neurons and output neuron at time t, θ(t)=[θ₁(t), θ₂(t), . . . , θ_J(t)]^T is the output vector of the hidden layer, y(t) represents the output of recurrent RBF neural network at time t.

The error of recurrent RBF neural network is:

$\begin{array}{cc}E\left(t\right)=\frac{1}{N}\sum _{t=1}^N{\left(y_d\left(t\right)-y\left(t\right)\right)}^2& \left(33\right)\end{array}$

y_d(t) is the expectation output of neural network and the actual output is expressed as y(t).

(3) Train Recurrent RBF Neural Network:

{circle around (1)} Given the recurrent RBF neural network, the initial number of hidden layer neurons is J, J>2 is a positive integer. The input of recurrent RBF neural network is x(1), x(2), . . . , x(t), . . . , x(N), correspondingly, the expectation output is y_d(1), y_d(2), . . . , y_d(t), . . . , y_d(N), expected error value is set to E_d, E_d∈(0, 0.01). The every variable of initial centre value c_j(1)∈(−2, 2), width value σ_j(1)∈(0, 1), initial feedback weight v_j(1)∈(0, 1), j=1, 2, . . . , J. Initial weight w(1)∈(0, 1).

{circle around (2)} Set the learning step s=1;

{circle around (3)} t=s; According to Eqs. (1)-(4), calculate the output of recurrent RBF neural network, exploiting fast gradient descent algorithm:

$\begin{array}{cc}{c}_j\left(t+1\right)=c_j\left(t\right)-{\eta }_c\frac{1}{{\sigma }_j^2}\left(y_d\left(t\right)-y\left(t\right)\right)w_j\left(t\right)\times \theta \left(t\right)\left[h_j\left(t\right)-c_j\left(t\right)\right]& \left(34\right)\\ {\sigma }_j\left(t+1\right)={\sigma }_j\left(t\right)-{\eta }_{\sigma }\frac{1}{{\sigma }_j^3}\left(y_d\left(t\right)-y\left(t\right)\right)w_j\left(t\right)\times \theta \left(t\right){h_j\left(t\right)-c_j\left(t\right)}^2& \left(35\right)\\ v_j\left(t+1\right)=v_j\left(t\right)-{\eta }_v\left(y_d\left(t\right)-y\left(t\right)\right)w_j\left(t\right)\theta \left(t\right)y\left(t-1\right)& \left(36\right)\\ w_j\left(t+1\right)=w_j\left(t\right)-{\eta }_w\left(y_d\left(t\right)-y\left(t\right)\right){\theta }_j\left(t\right)& \left(37\right)\end{array}$ η_c, η_σ, η_v, η_w are the learning rate of centre, width, feedback connection weight from output layer to hidden layer and the connection weight between hidden layer and output layer, respectively. In addition, η_c∈(0, 0.01], η_σ∈(0, 0.01], η_v∈(0, 0.02], η_w∈(0, 0.01]. c_j(t+1)=[c_1j(t+1), c_2j(t+1), . . . , c_5j(t+1)] denotes the center vector of the jth hidden neuron at time t+1. σ_j(t+1) is the radius or width of the jth hidden neuron at time t+1. v_j(t+1) denotes the connection weight from output layer to the jth hidden neuron at time t+1. w_j(t+1) is the connection weights between the hidden neurons and output neuron at time t+1.

{circle around (4)} increase 1 learning step s, if s<N, then turn to step {circle around (3)}. If s=N, turn to step {circle around (5)}.

{circle around (5)} according to Eq. (5), calculate the performance of recurrent RBF neural network. If E(t)≥E_d, then turn to step {circle around (3)}; If E(t)<E_d, stop the training process.

(4) SVI Concentration Prediction:

The testing samples used as the input of recurrent RBF neural network, the output of neural network is the soft-computing values of SVI. The testing result is shown in FIG. 2. X axis indicates the number of samples. Y axis shows SVI. The unit of Y axis is mg/L. The red solid line presents the real values of SVI. The blue dot line shows the outputs of recurrent RBF neural network in the testing process. The errors between the real values and the outputs of recurrent RBF neural network in the testing process are shown in FIG. 3. X axis shows the number of samples. Y axis shows the testing error. The unit of Y axis is mg/L.

(5) CVI Algorithm for Sludge Bulking:

{circle around (1)} Calculate the residual of the expectation output and the output of recurrent RBF neural network, if

$\begin{array}{cc}\{\begin{array}{c}y\left(t\right)-y_d\left(t\right)\ge 5\\ y\left(t\right)\ge 150\end{array}& \left(38\right)\end{array}$ then turn to step {circle around (2)}, otherwise, stop the process of fault identification for sludge bulking.

{circle around (2)} Define two formula:

$\begin{array}{cc}\{\begin{array}{c}{\mathrm{IC}}_1\left(t\right)={\delta }^T\left(t\right){\Lambda }_M^{-1}\left(t\right)\delta \left(t\right)\\ {\mathrm{IC}}_2\left(t\right)={{\theta }_M\left(t\right)}^2-{\delta }^T\left(t\right)\delta \left(t\right)\end{array}& \left(39\right)\end{array}$ wherein IC₁(t) is the Mahalanobis distance of input variables at time t and IC₂(t) is the squared prediction error at time t, M is the numbers of principal component of input data. K is the number of the front samples, θ_M(t) is the output vector of the hidden layer for M principal components training data at time t. δ(t)=[δ₁(t), . . . , δ_m(t), . . . , δ_M(t)]^T is the projection of the training data, and δ_m(t) is

$\begin{array}{cc}{\delta }_m\left(t\right)=\sum _{k=1}^K{a}_k\left(t\right)\left({\stackrel{\_}{\theta }}_m\left(t\right)·{\stackrel{\_}{\theta }}_{\mathrm{new}}\left(t\right)\right),k=1,L,K;m=1,\dots ,M& \left(40\right)\end{array}$ wherein θ_new(t) is the mean-centered output vector of the hidden layer for M principal components training data at time t, θ_m(t) is the output vector of the hidden layer for the mth principal component training data at time t, K is the number of the front samples, a_k(t) is a constant, and a_k(t)∈(0, 0.01]. And the diagonal matrix of eigenvalues associated with M principal components is defined as

$\begin{array}{cc}{\Lambda }_M\left(t\right)=\left[\begin{array}{cccc}{\lambda }_1& & & \\ & {\lambda }_2& & \\ & & O& \\ & & & {\lambda }_M\end{array}\right]\left({\lambda }_1\ge {\lambda }_2\ge L\ge {\lambda }_M\ge 0\right)& \left(41\right)\end{array}$ wherein Λ_M(t) is the diagonal matrix of eigenvalues at time t and it satisfies{tilde over (C)}(t)=Z(t)Λ_M(t)Z^T(t)+l′(t)(I(t)−Z(t)Z^t(t))  (42)wherein l′(t) is a constant value, I(t) is a unit matrix, {tilde over (C)}(t) is the regularized covariance matrix of C(t) at time t:

$\begin{array}{cc}\Omega \left(t\right)={\theta \left(t\right)}^T{C}^{\%-1}\left(t\right)\theta \left(t\right)& \left(43\right)\\ \Omega \left(t\right)={\mathrm{IC}}_1\left(t\right)+l^{\prime -1}\left(t\right){\mathrm{IC}}_2\left(t\right)& \left(44\right)\\ C\left(t\right)=\frac{1}{K}\sum _{k=1}^K{\theta }_k\left(t\right){{\theta }_k\left(t\right)}^T& \left(45\right)\end{array}$ wherein Ω(t) is the energy of each variable. θ_k(t) is the hidden output vector of the kth principal component at time t.

The constant vector a(t)=[a₁(t), . . . a_j(t), . . . a_K(t)]^T is given as

$\begin{array}{cc}\lambda \left(t\right)a\left(t\right)=\frac{1}{K}G\left(t\right)a\left(t\right)& \left(46\right)\end{array}$ G(t) is the Gaussian matrix and λ(t) denotes the eigenvalueG(t)={θ_i(t)·θ_j(t)}_K×K  (47)λ(t)p(t)=C(t)p(t)  (48)p(t) denotes eigenvector of the covariance matrix C(t) at time t.

{circle around (3)} For the ith input variable, the contribution degree index satisfy:

$\begin{array}{cc}{G}^i\left(t\right)=\frac{{\kappa }^i\left(t\right)}{\sum _{i=1}^5{\kappa }^i\left(t\right)}& \left(49\right)\end{array}$ wherein Gⁱ(t) is the contribution degree index of ith variable at time t, a hidden layer neuron corresponds to an input variable, κⁱ(t) is the contribution degree which is calculate by the mutual information between this testing samples and difference sets for testing samples xⁱ(t) of ith variable at time t, which can be expressed asκⁱ(t)=I(xⁱ(t),V_Δ(t))  (50)I(xⁱ(t), V_Δ(t)) is the mutual information of xⁱ(t) and V_Δ(t) at time t, V_Δ(t) is the difference matrix sets of data of training set and testing set at time t, which is expressed asV_Δ(t)=V_tr(t)−V_te(t)  (51)wherein V_tr(t) and V_te(t) is independent data sets of training set and testing set at time t, respectively.V_tr(t)=D⁻¹(t)G(t)  (52)V_te(t)=D_te⁻¹(t)G_te(t)  (53)D(t) is the covariance matrix of Φ(t) at time tD(t)=E{Φ(t)Φ^T(t)}  (54)Φ(t)=[(t−K+1), . . . ,θ(t−1),θ(t)]^T  (55)Φ(t) is output matrix of hidden layer at time t, θ(t−K+1) is the hidden output vector at time t−K+1, K is the number of the front samples.

{circle around (4)} For the ith input variable, a hidden layer neuron corresponds to an input variable, if the Gⁱ(t) at time t satisfies:G¹(t)+ . . . Gⁱ(t)≥0.8  (56)

the variables 1, . . . , i is the cause variables resulted in sludge bulking.

The IC₁(t) outputs is shown in FIG. 4. X axis shows the number of samples. Y axis shows the IC₁(t) outputs. The IC₂(t) outputs is shown in FIG. 5. X axis shows the number of samples. Y axis shows the IC₂(t) outputs. The marked fault points of IC₁(t) and IC₂(t) is shown in FIG. 6. X axis shows the number of samples. Y axis shows the marked fault points of IC₁(t) and IC₂(t). The green solid line shows the marked fault points of IC₁(t) in the testing process. The red solid line shows the marked fault points of IC₂(t) in the testing process. The contribution degree index of process variables is shown in FIG. 7, X axis shows the process variables. Y axis shows the contribution degree index.

Tables 1-12 show the experimental data in this present disclosure. Tables 1-5 show the training samples of COD, DO, T, MLSS and TN. Tables 6-10 show the testing samples of COD, DO, T, MLSS and TN. Table 11 shows real output values of SVI. Table 12 shows the outputs of the recurrent RBF neural network in the predicting process. Moreover, the samples are imported as the sequence from the tables. The first data is in the first row and the first column. Then, the second data is in the first row and the second column. Until all of data is imported from the first row, the data in the second row and following rows are inputted as the same way.

Training samples are provided as follows:

TABLE 1
The input of chemical oxygen demand-COD (mg/L)
317.655	319.9375	322.25	324.5625	326.875	329.1875	331.5	333.8125	336.125	338.4375
340.75	343.0345	347.6875	350	350.1875	350.375	350.5625	350.75	350.9375	351.125
351.3125	351.5	351.6875	351.875	352.0625	352.25	352.4375	352.625	352.8125	353
348.0625	343.125	338.1875	333.25	328.3125	323.375	318.4375	313.5	308.5625	303.625
305.6875	306.75	306.8125	307.875	308.9375	309	310.875	311.75	320.625	329.5
328.375	327.25	326.125	325	323.875	325.75	326.625	328.5	330.375	331.25

TABLE 2
The input of dissolved oxygen concentration-DO (mg/L)
6.845	7.246	6.659	7.239	6.255	6.735	6.481	6.724	6.944	7.434
6.917	7.842	6.128	7.901	7.405	7.647	7.879	7.434	7.179	7.234
6.532	6.543	6.554	6.088	6.974	6.222	7.893	6.058	6.664	7.753
6.695	6.38	7.751	6.112	6.935	7.038	7.506	6.355	6.152	6.222
7.974	6.129	6.853	7.138	6.178	6.555	6.617	7.151	7.924	6.525
6.899	7.194	7.555	7.29	7.044	6.975	6.311	7.014	6.556	6.108

TABLE 3
The input of T
24.1907	24.1592	24.1163	23.9890	23.9090	23.8377	23.7450	23.6837	23.6623	23.6367
23.6253	23.6295	23.6295	23.6239	23.6253	23.6310	23.6253	23.6224	23.6082	23.5968
23.5769	23.5470	23.5114	23.4745	23.4318	23.3835	23.3310	23.2770	23.2146	23.1451
23.0870	23.0403	23.0007	22.9625	22.9851	23.0191	23.0559	23.0743	23.1409	23.1877
23.2032	23.1565	23.1834	23.1707	23.1551	23.1409	23.0956	23.1027	23.0630	23.0134
22.9738	22.9115	22.8593	22.8098	22.7420	22.7448	22.7504	22.8084	22.9144	23.0191

TABLE 4
The input of mixed liquor suspended solids concentration-MLSS (mg/L)
1286.885	1287.375	1286.5	1285.625	1284.75	1283.875	1283	1282.125	1281.25	1283.375
1289.5	1288.625	1287.75	1287.875	1286	1287.313	1284.625	1283.938	1285.25	1282.563
1281.875	1281.188	1282.5	1283.813	1285.125	1284.438	1287.75	1287.063	1286.375	1285.688
1285	1285.188	1285.375	1287.563	1288.75	1289.938	1290.125	1296.313	1296.5	1296.688
1294.875	1297.063	1297.25	1297.438	1297.625	1297.813	1298	1297.375	1296.75	1296.125
1295.5	1293.875	1284.25	1293.625	1293	1302.375	1311.75	1321.125	1319.5	1329.875

TABLE 5
The input of total nitrogen-TN (mg/L)
41.275	41.3125	41.35	41.3875	41.425	42.4625	41.5	43.5375	41.575	44.6125
42.65	41.6875	41.725	44.7625	45.8	42.3125	42.825	43.3375	43.85	44.3625
44.875	45.3875	45.9	46.4125	46.925	47.4375	48.95	48.4625	48.975	49.4875
47.241	49.5375	49.075	48.6125	48.15	47.6875	47.225	46.7625	46.3	45.8375
47.375	48.9125	47.45	43.9875	43.525	44.0625	42.6	42.88125	43.1625	43.44375
43.725	44.00625	44.2875	44.56875	45.85	44.13125	47.4125	46.69375	45.975	46.25625
Testing samples:

TABLE 6
The input of chemical oxygen demand-COD (mg/L)
332.125	334	334.8125	335.625	336.4375	337.25	338.0625	338.875	339.6875	340.5
341.3125	342.125	342.9375	341.75	342.5625	343.375	344.1875	343	343.875	344.75
345.625	344.5	346.375	347.25	348.125	349	349.875	349.75	349.625	349.5
349.375	350.25	351.125	359	356.375	353.75	351.125	348.5	345.875	346.25

TABLE 7
The input of dissolved oxygen concentration-DO (mg/L)
6.59875	6.693125	6.7875	7.281875	7.17625	7.070625	6.965	6.859375	6.75375	6.648125
7.5425	6.936875	6.93125	6.725625	7.12	7.178125	6.83625	6.994375	6.9525	6.410625
6.46875	6.526875	7.085	6.643125	6.80125	6.659375	6.8175	6.875625	6.93375	6.991875
7.25	7.3375	7.125	7.1125	7	6.9875	6.975	6.9625	6.95	6.9375

TABLE 8
The input of T
23.1381	23.2841	23.4446	23.5598	23.6338	23.6680	23.6751	23.7307	23.7393	23.7535
23.7378	23.7193	23.6766	23.6295	23.6096	23.5570	23.4958	23.4489	23.4048	23.4006
23.3949	23.4048	23.4190	23.4915	23.8191	23.8477	23.8662	23.8905	23.0148	27.2087
27.2933	27.3334	27.317	27.3022	27.2888	27.2681	27.2354	27.1983	27.1584	27.0976

TABLE 9
The input of mixed liquor suspended solids concentration-MLSS (mg/L)
1329.25	1338.625	1348	1346.25	1344.5	1342.75	1345	1350.25	1357.5	1355.75
1354.09	1352.25	1350.5	1348.75	1347	1345.25	1343.5	1341.75	1340	1338.25
1336.5	1334.75	1333	1331.25	1329.5	1327.75	1326	1324.25	1322.5	1320.75
1319	1317.25	1315.5	1313.75	1312	1308.688	1305.375	1302.063	1298.75	1295.438

TABLE 10
The input of total nitrogen-TN (mg/L)
46.5375	46.81875	47.1	47.3875	49.675	47.9625	48.25	49.5375	50.825	49.1125
48.4	47.6875	46.975	45.2625	46.55	47.8375	46.125	47.4125	45.7	44.6375
44.575	43.5125	41.45	41.3875	41.325	41.2625	43.2	42.1375	40.075	40.0125
40.95	39.8875	40.825	39.7625	38.7	36.475	36.25	36.025	36.8	37.575

TABLE 11
The real output of sludge volume index-SVI (mg/L)
102.1318	104.5506	104.9427	105.2887	106.4809	107.2111	113.4479	119.0132	121.5167	115.047
119.9333	124.776	128.374	125.9713	130.4835	129.933	133.8057	135.4856	135.7096	134.7779
132.6632	136.047	135.7843	133.8904	131.9964	138.4294	144.1786	150.3707	152.1785	149.7829
154.45	154.7836	152.1172	157.832	155.5825	157.7072	158.4782	159.0304	160.8222	157.2703

TABLE 12
The output of sludge volume index-SVI in the testing process (mg/L)
104.6798	103.4415	106.0121	102.9707	104.3378	105.9182	112.4079	118.0338	122.0289	118.0726
122.6676	127.6083	133.1356	128.6487	133.2567	132.7159	137.5757	138.2312	140.2654	139.2904
137.7872	137.7449	141.9649	139.9922	136.1045	143.3388	147.961	155.0361	159.5865	157.078
159.0946	160.3922	161.725	164.3612	166.4679	166.6599	165.2393	164.7875	163.0705	162.5019

Claims

1. A fault identification method for sludge bulking based on a recurrent radial basis function neural network (RRBFNN) comprising the following steps: (1) determine input and output variables of sludge volume index (SVI):for sewage treatment process of an activated sludge system, by analyzing detailed mechanism of the sludge bulking, analyzing five process variables and selecting the input variables of SVI soft-computing model: dissolved oxygen concentration-DO, mixed liquor suspended solids concentration-MLSS, temperature-T, chemical oxygen demand-COD and total nitrogen-TN, wherein an output value of the SVI soft-computing model is SVI value;(2) initial RRBFNN: