CN111880090A - Distribution layered online fault detection method for million-kilowatt ultra-supercritical unit - Google Patents

Abstract

The invention discloses a distributed layered on-line fault detection method for a million-kilowatt ultra-supercritical unit. Aiming at the problems of numerous process variables and complex working conditions of mega-kilowatt ultra-supercritical units, comprehensively considering the correlation between variables and the distribution of variables in the sample direction, the multi-layer information theory decomposition method is used to divide the variables into blocks. Based on the block results, combined with the Gaussian mixture model method and Bayesian theory, a multi-layer distributed monitoring algorithm for millions of ultra-supercritical units is established. This method fully explores the relevant information among the process variables, which is beneficial to the understanding of the complex process characteristics of the million-kilowatt ultra-supercritical unit. The multi-layer distributed monitoring method can not only effectively mine the local information of the process, but also analyze the different variables. The correlation between the blocks greatly improves the fault detection performance of the complex process of the million-kilowatt ultra-supercritical unit, thereby ensuring the safe and reliable operation of large coal-fired generating units.

Description

A distributed layered online fault detection for million kilowatt ultra-supercritical units method

technical field

The invention belongs to the field of thermal power process fault detection, in particular to a distributed layered on-line process monitoring method for a million-kilowatt ultra-supercritical unit with numerous variables and frequent fluctuations in operating conditions.

Background technique

Electric power industry is an important basic industry of the national economy and a key project in the national economic development strategy. With the rapid economic development, the demand for electricity has also increased rapidly. Coal resources are the main energy in my country, so it is difficult to fundamentally change the coal-dominated energy structure for a long period of time in the future. As China's main power source, the installed capacity of coal-fired power generation has always been above 70%. According to statistics, driven by the strong demand for electricity, the total electricity consumption of the whole society from January to August 2018 reached 4,529.6 billion kWh, a year-on-year increase of 9.0%. Among them, the cumulative power generation of thermal power was 3,310.3 billion kWh, accounting for about 73.1% of the country's total power generation, a year-on-year increase of 7.2%. In recent years, in order to realize the sustainable development of electric power, the thermal power industry has actively carried out structural adjustment, "superpowering the large and suppressing the small", replacing the high-energy consumption small thermal power units with large-capacity, low-energy-consumption ultra (super) critical units, basically forming hundreds of The power energy structure is mainly composed of large-scale coal-fired generating units such as 10,000-kilowatt ultra-supercritical units. Therefore, the analysis and research on the million kilowatt ultra-supercritical unit has great practical significance and application value.

Compared with traditional generator sets, the mega-kilowatt ultra-supercritical unit has a large scale, diverse equipment, numerous parameters and mutual influences, and the entire power generation process has a long industrial process, many unit devices, wide space distribution, and high safety requirements. Condition monitoring and fault diagnosis of 10,000-kilowatt ultra-supercritical units have brought difficulties. In addition, due to different environmental conditions, fuel characteristics and load sizes, the 1000MW ultra-supercritical unit may operate under different working conditions. Especially in recent years, due to grid load fluctuations, increased peak-to-valley difference, and changes in user-side demand caused by the integration of new energy sources such as wind power and photovoltaics, the new normals such as frequent and deep peak shaving have led to the occurrence of frequent and deep peak shaving. full operating mode. In addition, the large-scale coal-fired power generation process has a complex environment and a long industrial process. Even under the same working conditions, many variables still show different data distribution characteristics. All these pose great challenges to the fault detection and diagnosis of large coal-fired generating units.

Aiming at the problem of fault detection of thermal power generating units, predecessors have done corresponding researches and discussions from different angles, and put forward corresponding online process monitoring methods. However, most of the existing methods are mainly centralized monitoring methods for a single working condition. Faced with the characteristics of long process flow, numerous variables, complex correlations, and dynamic working conditions of mega-kilowatt ultra-supercritical units, the centralized monitoring method of single working condition cannot obtain a good monitoring effect. The content of the present invention deeply considers the complex correlation among many variables of the million-kilowatt ultra-supercritical unit and the multi-distribution of variables in the sample direction, and proposes a new multi-layer distributed system for the million-kilowatt ultra-supercritical unit. Online fault detection method.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a multi-layer distributed monitoring algorithm for 10,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 ultra-supercritical ultra-supercritical unit fault detection methods that the local information cannot be accurately described. The method comprehensively considers the correlation between variables and the distribution of variables in the sample direction, and uses the multi-layer information theory decomposition method to divide the variables into blocks, fully exploring the process information between the process variables, which is conducive to the analysis of the million kilowatt ultra-supercritical Knowledge of complex process characteristics of units. The multi-layer distributed monitoring method can not only effectively mine the local information of the process, but also analyze the correlation between different variable sub-blocks, which greatly improves the fault detection performance of the complex process of the million-kilowatt ultra-supercritical unit, thereby ensuring Safe and reliable operation of large coal-fired generating units.

The object of the present invention is achieved through the following technical solutions: a distributed layered online fault detection method for a million kilowatt ultra-supercritical unit, the method comprises the following steps:

(1) Obtaining normal data to be analyzed: Suppose a million kilowatt ultra-supercritical unit has J measurement variables and operating variables, each sampling can obtain a 1 × J vector, and the data obtained after N sampling is expressed as a two Dimensional matrix X=[X ₁ , X ₂ ,...,X _J ]∈R ^N×J , wherein the measured variables are the state parameters that can be measured during the normal operation of the unit, including flow, voltage, current, temperature , speed, etc.; the operating variables include air inlet volume, feeding volume, valve opening, etc.;

(2) Using the spectral clustering method based on mutual information, the process variables are divided into different sub-blocks, the variables in the same sub-block have strong correlation, and the correlation between different sub-blocks is weak. This step is implemented by the following sub-steps:

(2.1) Find the mutual information between variables:

I(X _i ,X _j )=H(X _i )+H(X _j )-H(X _i ,X _j ) (1)

Among them, X _i (i=1,2,...,J) represents the ith variable, and H(X _i ) is the information entropy of variable X _i :

H(X _i )=-∫ _x p(X _i )logp(X _i )dx (2)

H(X _i , X _j ) is the joint information entropy of variables X _i and X _j :

p(X _i ) and p(X _j ) represent the probability density functions of variables X _i and X _j , and p(X _i , X _j ) is the joint probability density function.

(2.2) Based on the mutual information obtained by formula (1), obtain the generalized correlation coefficient between the two variables:

where r _ij ∈ [0,1].

(2.3) Based on formula (4), obtain the correlation matrix of variables:

(2.4) Based on the correlation matrix R defined by formula (5), obtain the diagonal diagonal matrix D:

D=diag{ _Dii } (6)

Among them, D _ii is the sum of all elements in the i-th row in formula (5):

(2.5) Find the Laplace matrix of the diagonal diagonal matrix D

L=D ^-1/2 RD ^-1/2 (8)

(2.6) Spectral decomposition of the Laplace matrix

L=PΛP ^T (9)

Among them, P=[P ₁ , P ₂ , . . . , P _J ] are orthogonal eigenvectors.

(2.7) Select the eigenvectors corresponding to the k largest eigenvalues to form a matrix E=[P ₁ , P ₂ ,...,P _k ]∈R ^J×k , and normalize each row in the matrix E to obtain matrix Y

(2.8) Use kmeans clustering algorithm to cluster Y, if the i-th row belongs to the b-th class, the variable X _i is divided into the b-th sub-block X _b . In this way, many operating variables of the million kilowatt ultra-supercritical unit are divided into B variable blocks according to the degree of correlation.

X=[X ₁ X ₂ … X _B ] (11)

是第b(b＝1,2,...,B)个变量块，J_b表示X_b中包含的变量个数。in,

is the bth (b=1,2,...,B) variable block, and J _b represents the number of variables contained in X _b .

(3) Use the information theory decomposition method based on Gaussian mixture model to further decompose the variables in the variable block according to the distribution in the sample direction. This step is realized by the following sub-steps:

(3.1) The variable block is randomly divided into W _b sub-blocks:

(3.2) Use the Gaussian mixture model method to obtain the probability density of the wth (w=1, 2,...,W _b ) variable sub-block:

是子高斯成分的个数；

是第m个子高斯成分的先验概率，满足

以及

为包含子高斯成分的均值

和协方差矩阵

的参数集。

为多元高斯概率密度：in,

is the number of sub-Gaussian components;

is the prior probability of the mth sub-Gaussian component, satisfying

as well as

is the mean with sub-Gaussian components

and covariance matrix

parameter set.

is the multivariate Gaussian probability density:

中变量的个数。Among them, J _b,w is

The number of variables in .

(3.3) Obtain the probability density distribution function of each variable in the sub-block:

J_b,w是变量块

中变量的个数，

表示X_b,i属于子块

时X_b,i的条件概率密度。where the variable

J _{b, w} are variable blocks

the number of variables in ,

Indicates that X _b,i belongs to the sub-block

The conditional probability density of X _b,i when .

和

的KL散度(w,v∈[1,W_b])：(3.4) Obtain variable sub-block

and

The KL divergence of (w, _v∈ [1,Wb]):

和

分别是

和

的概率密度函数，可以利用公式(13)计算。in,

and

respectively

and

The probability density function of , can be calculated using formula (13).

(3.5) Use the ant colony algorithm to optimize the random block of step (3.1) so that the following objective function is maximized:

(3.6) By repeating steps (3.2)-(3.5), each variable block (b=1, 2, . . . , B) is further divided into several sub-blocks. The original dataset X is divided into different sub-blocks:

中所有变量具有很强的相关关系，变量子块

中的变量既有很强相关关系且具有类似的数据分布。where the variable block

All variables in have a strong correlation, the variable sub-block

The variables in are both strongly correlated and have similar data distributions.

中各个变量的相关关系(4) Based on the variable block results obtained in steps (2) and (3), first use Principal Component Analysis (PCA) to describe the variable sub-blocks

The relationship between the various variables in

Among them, P _b,w is the load matrix, and T _b,w is the pivot matrix.

(5) Use the Gaussian Mixture Model (GMM) method to establish the distribution of the principal matrix T _{b, w} :

是高斯分量的个数；

表示第m个分量的权重，

为包含子高斯成分的均值

和协方差矩阵

的参数集。in,

is the number of Gaussian components;

represents the weight of the mth component,

is the mean with sub-Gaussian components

and covariance matrix

parameter set.

建立BIP统计量(6) For each variable sub-block

Build BIP Statistics

表示

属于第m个分量

的概率，

为主元矩阵T_b,w第n(n＝1,2,..,N)行向量。

是基于局部马氏距离的概率，其定义为in,

express

belongs to the mth component

The probability,

is the nth (n=1,2,..,N) row vector of the principal matrix T _b,w .

is the probability based on the local Mahalanobis distance, which is defined as

为

到第m个高斯分量的马氏距离，t为T_b,w中任意一行。in,

for

Mahalanobis distance to the mth Gaussian component, t is any row in T _b,w .

(7) Using the Gaussian Mixture Model (GMM) method to monitor the relationship between each sub-block in each variable block, this step is realized by the following sub-steps.

(7.1) Combine the first columns of the pivot matrix of each sub-block in each variable block X _b :

Among them, t _b,w (w=1,2,...,W _b ) is the first column vector of the principal element matrix T _b,w .

的分布情况：(7.2) Using GMM to describe main metadata

The distribution of :

为第b个变量子块中高斯分量的个数；

表示第m个分量的权重，

为包含子高斯成分的均值

和协方差矩阵

的参数集。in,

is the number of Gaussian components in the b-th variable sub-block;

represents the weight of the mth component,

is the mean with sub-Gaussian components

and covariance matrix

parameter set.

建立BIP统计量：(7.3) Main metadata for each variable block

Build BIP statistics:

为主元矩阵

第n(n＝1,2,..,N)行向量。

是基于局部马氏距离的概率。

是基于局部马氏距离的概率，

为

到第m个高斯分量的马氏距离，

为

中任意一行。in,

pivot matrix

The nth (n=1,2,..,N) row vector.

is the probability based on the local Mahalanobis distance.

is the probability based on the local Mahalanobis distance,

for

Mahalanobis distance to the mth Gaussian component,

for

any line in the .

(8) During online fault detection, the process is monitored from three levels of variable sub-block, variable block and the entire unit. This step is achieved by the following sub-steps.

(8.1) Acquiring new data: According to step (1), collect the value of each measuring point variable, and record it as z(1×J).

(8.2) According to the variable block results obtained in steps (2) and (3), decompose the new data into sub-blocks:

z=[z ₁ z ₂ … z _b … z _B ] (26)

where z _b (b=1,2,...,B) is the bth variable sub-block.

的数据向对应子块的主元方向进行投影：(8.3) At the bottom layer, that is, the variable sub-block layer, put each sub-block in

The data is projected to the pivot direction of the corresponding sub-block:

是负载矩阵。in

is the load matrix.

的在线统计量指标：(8.4) Find each sub-block

Online statistic indicators for :

表示

属于第m个分量

的概率。

是基于局部马氏距离的概率，

为

到第m个高斯分量的马氏距离，t为T_b,w中任意一行。The meanings of the parameters in the above formula are similar to those in formula (22).

express

belongs to the mth component

The probability.

is the probability based on the local Mahalanobis distance,

for

Mahalanobis distance to the mth Gaussian component, t is any row in T _b,w .

(8.5) At the variable block layer, first combine the pivots of each variable sub-block in z _b :

(8.6) Obtain the online statistics index of each variable block z _b :

表示

属于第m个分量

的概率。

是基于局部马氏距离的概率，

为

到第m个高斯分量的马氏距离，

为

中任意一行。The meanings of the parameters in the above formula are similar to those in formula (25).

express

belongs to the mth component

The probability.

is the probability based on the local Mahalanobis distance,

for

Mahalanobis distance to the mth Gaussian component,

for

any line in the .

(8.7) In order to analyze the relationship between different variable sub-blocks, to monitor the operation status of the million kilowatt ultra-supercritical unit from the whole unit level, first convert the BIP index of each variable block to normal (marked as 'N') with the probability of failure (marked 'F'):

表示第b个变量块正常的条件概率；

表示第b个变量块发生故障的条件概率。Among them, BIP _{b, lmt} is the control limit of the statistic BIP indicator;

represents the normal conditional probability of the b-th variable block;

represents the conditional probability of failure of the bth variable block.

(8.8) Calculate the posterior probability of failure of the bth variable block by Bayesian rule

Among them, P _b (F) = α; P _b (N) = 1-α represent the prior probability of failure or normality of the process at the significance level of α, respectively.

(8.9) Comprehensively consider the failure probability of all variable blocks, and calculate the global monitoring statistics

(9) Judging the process running state: analyze the process state from three levels of variable sub-block, variable block and the whole unit. Compare statistics and control limits at three levels in real time:

中，如果BIP_b,w＞1-α，说明在子块

中的变量发生了故障，否则认为子块中的变量运行在正常范围内。(a) in each variable subblock

, if BIP _b,w > 1-α, it means that in the sub-block

The variables in the sub-block are considered faulty, otherwise the variables in the sub-block are considered to be operating within the normal range.

(b) At the variable block level, if BIP _b > 1-α, it means that the correlation of each variable sub-block in the variable block is abnormal; otherwise, it means that all variables in the bth sub-block are running normally.

(c) At the unit level, if PF _z >α, it means that an abnormality or failure has occurred during the operation of the million kilowatt ultra-supercritical unit; otherwise, it means that the unit is operating normally as a whole.

Compared with the prior art, the beneficial effects of the present invention are as follows: the purpose of the present invention is to aim at the large scale of the million kilowatt ultra-supercritical unit, various equipment, numerous parameters and mutual influence, and the entire power generation process has a long industrial process and many unit devices. , wide spatial distribution and frequent switching of operating conditions, a multi-layer distributed monitoring algorithm for millions of ultra-supercritical units is proposed. The method comprehensively considers the correlation between variables and the distribution of variables in the sample direction, and uses the multi-layer information theory decomposition method to divide the variables into blocks, fully exploring the process information between the process variables, which is conducive to the analysis of the million kilowatt ultra-supercritical Knowledge of complex process characteristics of units. The multi-layer distributed monitoring method can not only effectively mine the local information of the process, but also analyze the correlation between different variable sub-blocks, which greatly improves the fault detection performance of the complex process of the million-kilowatt ultra-supercritical unit, thereby ensuring Safe and reliable operation of large coal-fired generating units.

Description of drawings:

Fig. 1 is the explanatory diagram of the distributed layered on-line fault detection method of the present invention facing one million kilowatts of ultra-supercritical units;

2 is the monitoring result of the method of the present invention in the variable sub-block in the specific embodiment, (a) is the monitoring result of the two variable sub-blocks in the fifth variable block, (b) is the monitoring result in the seventh variable block The monitoring results of the three variable sub-blocks in the block, (c) is the monitoring results of the three variable sub-blocks in the ninth variable block.

Fig. 3 is the monitoring result in the variable layer of the method of the present invention in the specific embodiment, (a) is the monitoring result in the 5th variable block, (b) is the monitoring result in the 7th variable block, (c) For the monitoring results in the 9th variable block.

FIG. 4 is a monitoring result at the unit level of the method of the present invention in a specific embodiment.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific examples.

The present invention takes the No. 3 unit of Jiahua Power Plant under Zheneng Group as an example. This unit is a million kilowatt ultra-supercritical unit with a power of 600 MW and includes a total of 154 process variables. These variables involve pressure, temperature, flow, and flow rate. Wait.

As shown in Figure 1, the present invention is a kind of on-line monitoring method for the collaborative analysis of dynamic and static characteristics of a million-kilowatt ultra-supercritical unit, comprising the following steps:

(1) Obtaining normal data to be analyzed: Suppose a million kilowatt ultra-supercritical unit has J measurement variables and operating variables, each sampling can obtain a 1 × J vector, and the data obtained after N sampling is expressed as a two A dimensional matrix X=[X ₁ , X ₂ ,...,X _J ]∈R ^N×J . In this example, the sampling period is 1 minute, and a total of 2940 sample data are collected during the normal operation of the thermal power unit for variable block and offline modeling, and 154 process variables, that is, the modeling data is X (2940×154). Wherein, the measured variables are state parameters that can be measured during the normal operation of the unit, including flow, voltage, current, temperature, speed, etc.; the operating variables include air intake volume, feed volume, valve opening, etc.;

(2) Using the spectral clustering method based on mutual information, the process variables are divided into different sub-blocks, the variables in the same sub-block have strong correlation, and the correlation between different sub-blocks is weak. This step is implemented by the following sub-steps:

(2.1) Find the mutual information between variables:

I(X _i ,X _j )=H(X _i )+H(X _j )-H(X _i ,X _j ) (1)

Among them, X _i (i=1,2,...,J) represents the ith variable, and H(X _i ) is the information entropy of variable X _i :

H(X _i )=-∫ _x p(X _i )logp(X _i )dx (2)

H(X _i , X _j ) is the joint information entropy of variables X _i and X _j :

p(X _i ) and p(X _j ) represent the probability density functions of variables X _i and X _j , and p(X _i , X _j ) is the joint probability density function.

(2.2) Based on the mutual information obtained by formula (1), obtain the generalized correlation coefficient between the two variables:

where r _ij ∈ [0,1].

(2.3) Based on formula (4), obtain the correlation matrix of variables:

(2.4) Based on the correlation matrix R defined by formula (5), obtain the diagonal diagonal matrix D:

D=diag{ _Dii } (6)

Among them, D _ii is the sum of all elements in the i-th row in formula (5)

(2.5) Find the Laplace matrix of the diagonal diagonal matrix D

L=D ^-1/2 RD ^-1/2 (8)

(2.6) Spectral decomposition of the Laplace matrix

L=PΛP ^T (9)

Among them, P=[P ₁ , P ₂ , . . . , P _J ] are orthogonal eigenvectors.

(2.7) Select the eigenvectors corresponding to the k largest eigenvalues to form a matrix E=[P ₁ , P ₂ ,...,P _k ]∈R ^J×k , and normalize each row in the matrix E to obtain matrix Y

(2.8) Use kmeans clustering algorithm to cluster Y, if the i-th row belongs to the b-th class, the variable X _i is divided into the b-th sub-block X _b . In this way, many operating variables of the million kilowatt ultra-supercritical unit are divided into B variable blocks according to the degree of correlation.

X=[X ₁ X ₂ ... X _B ] (11)

是第b(b＝1,2,...,B)个变量块，J_b表示X_b中包含的变量个数。in,

is the bth (b=1,2,...,B) variable block, and J _b represents the number of variables contained in X _b .

In this example, according to the correlation, the 154 process variables are divided into 11 sub-blocks. As shown in Table 1, the variables in each sub-block have strong correlation, and the correlation between different sub-blocks is weak.

Table 1. Variable block situation in 1 million kilowatt ultra-supercritical units

(3) Use the information theory decomposition method based on Gaussian mixture model to further decompose the variables in the above 11 variable blocks according to the distribution in the sample direction. This step is realized by the following sub-steps

(3.1) The variable block is randomly divided into W _b sub-blocks:

(3.2) Use the Gaussian mixture model method to obtain the probability density of the wth (w=1, 2,...,W _b ) variable sub-block:

是子高斯成分的个数；

是第m个子高斯成分的先验概率，满足

以及

为包含子高斯成分的均值

和协方差矩阵

的参数集。

为多元高斯概率密度：in,

is the number of sub-Gaussian components;

is the prior probability of the mth sub-Gaussian component, satisfying

as well as

is the mean with sub-Gaussian components

and covariance matrix

parameter set.

is the multivariate Gaussian probability density:

中变量的个数。Among them, J _b,w is

The number of variables in .

(3.3) Obtain the probability density distribution function of each variable in the sub-block:

J_b,w是变量块

中变量的个数，

表示X_b,i属于子块

时X_bi的条件概率密度。where the variable

J _{b, w} are variable blocks

the number of variables in ,

Indicates that X _b,i belongs to the sub-block

The conditional probability density of X _bi when .

和

的KL散度(w,v∈[1,W_b])：(3.4) Obtain variable sub-block

and

The KL divergence of (w, _v∈ [1,Wb]):

和

分别是

和

的概率密度函数，可以利用公式(13)计算。in,

and

respectively

and

The probability density function of , can be calculated using formula (13).

(3.5) Use the ant colony algorithm to optimize the random block of step (3.1) so that the following objective function is maximized:

(3.6) By repeating steps (3.2)-(3.5), each variable block (b=1, 2, . . . , B) is further divided into several sub-blocks. The original dataset X is divided into different sub-blocks:

中所有变量具有很强的相关关系，变量子块

中的变量既有很强相关关系且具有类似的数据分布。where the variable block

All variables in have a strong correlation, the variable sub-block

The variables in are both strongly correlated and have similar data distributions.

In this example, according to the distribution of variables, the 11 variable blocks obtained in step (2) are further divided into 27 variable sub-blocks. The variables in each sub-block have strong correlations and the same distribution.

Table 2. The division of variable sub-blocks in the million kilowatt ultra-supercritical unit

中各个变量的相关关系(4) Based on the variable block results obtained in steps (2) and (3), first use Principal Component Analysis (PCA) to describe the variable sub-blocks

The relationship between the various variables in

Among them, P _b,w is the load matrix, and T _b,w is the pivot matrix.

(5) Use the Gaussian Mixture Model (GMM) method to establish the distribution of the principal matrix T _{b, w} :

是高斯分量的个数；

表示第m个分量的权重，

为包含子高斯成分的均值

和协方差矩阵

的参数集。in,

is the number of Gaussian components;

represents the weight of the mth component,

is the mean with sub-Gaussian components

and covariance matrix

parameter set.

建立BIP统计量(6) For each variable sub-block

Build BIP Statistics

表示

属于第m个分量

的概率，

为主元矩阵T_b,w第n(n＝1,2,..,N)行向量。

是基于局部马氏距离的概率，其定义为in,

express

belongs to the mth component

The probability,

is the nth (n=1,2,..,N) row vector of the principal matrix T _b,w .

is the probability based on the local Mahalanobis distance, which is defined as

为

到第m个高斯分量的马氏距离，t为T_b,w中任意一行。in,

for

Mahalanobis distance to the mth Gaussian component, t is any row in T _b,w .

(7) Using the Gaussian Mixture Model (GMM) method to monitor the relationship between each sub-block in each variable block, this step is realized by the following sub-steps.

(7.1) Combine the first columns of the pivot matrix of each sub-block in each variable block X _b :

Among them, t _b,w (w=1,2,...,W _b ) is the first column vector of the principal element matrix T _b,w .

的分布情况：(7.2) Using GMM to describe main metadata

The distribution of :

为第b个变量子块中高斯分量的个数；

表示第m个分量的权重，

为包含子高斯成分的均值

和协方差矩阵

的参数集。in,

is the number of Gaussian components in the b-th variable sub-block;

represents the weight of the mth component,

is the mean with sub-Gaussian components

and covariance matrix

parameter set.

建立BIP统计量：(7.3) Main metadata for each variable block

Build BIP statistics:

为主元矩阵

第n(n＝1,2,..,N)行向量。

是基于局部马氏距离的概率。

是基于局部马氏距离的概率，

为

到第m个高斯分量的马氏距离，

为

中任意一行。in,

pivot matrix

The nth (n=1,2,..,N) row vector.

is the probability based on the local Mahalanobis distance.

is the probability based on the local Mahalanobis distance,

for

Mahalanobis distance to the mth Gaussian component,

for

any line in the .

(8.1) Fault data preparation: The fault data collected here contains a total of 460 samples, and the data is recorded as Z (460×154). The fault is that the outlet pressure of the circulating water pump increases, and the fault occurs at the 121st sampling point.

(8.2) According to the variable block results obtained in steps (2) and (3), the new data is denoted as z(1×154) for sub-block decomposition. In this example, B=11:

z=[z ₁ z ₂ ... z _b ... z _B ] (26)

where z _b (b=1,2,...,B) is the bth variable sub-block.

的数据向对应子块的主元方向进行投影：(8.3) At the bottom layer, that is, the variable sub-block layer, put each sub-block in

The data is projected to the pivot direction of the corresponding sub-block:

的在线统计量指标：(8.4) Find each sub-block

Online statistic indicators for :

表示

属于第m个分量

的概率。

是基于局部马氏距离的概率，

为

到第m个高斯分量的马氏距离，t为T_b,w中任意一行。The meanings of the parameters in the above formula are similar to those in formula (22).

express

belongs to the mth component

The probability.

is the probability based on the local Mahalanobis distance,

for

Mahalanobis distance to the mth Gaussian component, t is any row in T _b,w .

(8.5) At the variable block layer, first combine the pivots of each variable sub-block in z _b :

(8.6) Obtain the online statistics index of each variable block z _b :

表示

属于第m个分量

的概率。

是基于局部马氏距离的概率，

为

到第m个高斯分量的马氏距离，

为

中任意一行。The meanings of the parameters in the above formula are similar to those in formula (25).

express

belongs to the mth component

The probability.

is the probability based on the local Mahalanobis distance,

for

Mahalanobis distance to the mth Gaussian component,

for

any line in the .

(8.7) In order to analyze the relationship between different variable sub-blocks, to monitor the operation status of the million kilowatt ultra-supercritical unit from the whole unit level, first convert the BIP index of each variable block to normal (marked as 'N') with the probability of failure (marked 'F'):

表示第b个变量块正常的条件概率；

表示第b个变量块发生故障的条件概率。Wherein, BIP _{b, lmt} is the control limit of the statistic BIP, which is taken as 0.5 in this specific embodiment;

represents the normal conditional probability of the b-th variable block;

represents the conditional probability of failure of the bth variable block.

(8.8) Calculate the posterior probability of failure of the bth variable block by Bayesian rule

Among them, P _b (F)=α; P _b (N)=1-α respectively represent the prior probability of failure or normality of the process at the significance level α, in this example, the significance level α=0.5.

(8.9) Comprehensively consider the failure probability of all variable blocks, and calculate the global monitoring statistics

(9) Judging the process running state: analyze the process state from three levels of variable sub-block, variable block and the whole unit. Compare statistics and control limits at three levels in real time:

中，如果BIP_b,w＞1-α，说明在子块

中的变量发生了故障，否则认为子块中的变量运行在正常范围内。(a) in each variable subblock

, if BIP _b,w > 1-α, it means that in the sub-block

The variables in the sub-block are considered faulty, otherwise the variables in the sub-block are considered to be operating within the normal range.

(b) At the variable block level, if BIP _b > 1-α, it means that the correlation of each variable sub-block in the variable block is abnormal; otherwise, it means that all variables in the bth sub-block are running normally.

(c) At the unit level, if PF _z >α, it means that an abnormality or failure has occurred during the operation of the million kilowatt ultra-supercritical unit; otherwise, it means that the unit is operating normally as a whole.

Using the monitoring method of the present invention to monitor the thermal power process on-line, the results are shown in Figures 2-4. Figure 2 shows the monitoring results of the method of the present invention in 8 variable sub-blocks. As can be seen from Figure 2(a), the statistics of the two variable sub-blocks of the fifth variable block are basically below the control limit, indicating that the current fault does not affect each variable in variable block #5. are operating normally. As can be seen from Figure 2(b), in the first 120 samples, the statistics of the first two sub-blocks of variable block #7 are basically below the control limit, indicating that the process is running normally. Starting from the 121st sample, the variable subblock statistics BIP _sub7,1 and BIP _sub7,2 rapidly exceed the control limits, and a fault is detected, indicating that the fault clearly affects the variables in these two subblocks. Similarly, analyzing the monitoring results in Figure 2(c), it can be found that the variables in the first two sub-blocks of variable block #9 are basically operating normally, but BIP _sub9,3 effectively detects the occurrence of faults.

Figure 3 shows part of the monitoring results of the method of the present invention in the variable layer. As can be seen in Figure 3(a), the fifth variable block is basically below the control limit, indicating that the correlation between the two variable sub-blocks in variable block #5 is not affected by the fault. The process variable is operating normally. The monitoring results of Fig. 3(b) and Fig. 3(c) respectively show that the correlation of variable sub-blocks in variable block #7 and variable block #9 is affected by the fault, and abnormality occurs. Figure 4 shows the monitoring results of the inventive method at the unit level. It can be seen that the statistics are basically operating below the control limit in the first 120 samples, indicating that the million-kilowatt ultra-supercritical unit is operating under normal conditions. Starting from the 121st sample, the statistic PF _z immediately exceeds the control limit, effectively detecting the occurrence of a fault. In general, the layered distributed fault detection method of the present invention has superior fault detection performance when monitoring a typical large-scale multi-working condition process of one million kilowatts of ultra-supercritical, and its block results can effectively analyze the relationship between many variables. The complex correlation can not only deepen the operator's understanding of the process, but also provide high-precision online process monitoring results for the technical management department of the actual thermal power plant industrial site, and provide reliable information for judging the process operating status in real time and identifying whether there is a fault. Based on this, the safety, reliability and effectiveness of the operation of the million-kilowatt ultra-supercritical unit are further improved.

Claims (1)

1. A distribution layered online fault detection method for a million kilowatt ultra-supercritical unit is characterized by comprising the following steps:

(1) acquiring normal data to be analyzed: a million-kilowatt ultra-supercritical unit is provided with J measurement variables and operation variables, a vector of 1 XJ can be obtained by sampling every time, and data acquired after sampling N times is expressed as a two-dimensional matrix X ═ X₁,X₂,...,X_J]∈R^N×JThe measured variables are state parameters which can be measured in the normal operation process of the unit, and comprise flow, voltage, current, temperature, speed and the like; the operation variables comprise air intake, feeding amount, valve opening and the like;

(2) the process variable is divided into different sub-blocks by using a spectral clustering method based on mutual information, the variable in the same sub-block has stronger correlation, and the correlation between different sub-blocks is weaker. This step is realized by the following substeps:

(2.1) obtaining mutual information among variables:

I(X_i,X_j)＝H(X_i)+H(X_j)-H(X_i,X_j) (1)

wherein, X_i(i ═ 1, 2.., J) denotes the ith variable, H (X)_i) Is a variable X_iThe information entropy of (2):

H(X_i)＝-∫_xp(X_i)logp(X_i)dx (2)

H(X_i,X_j) Is a variable X_iAnd X_jJoint information entropy of (a):

p(X_i) And p (X)_j) Represents variable X_iAnd X_jP (X) is a probability density function of_i,X_j) Is a joint probability density function.

(2.2) solving the generalized correlation coefficient between every two variables based on the mutual information solved by the formula (1):

wherein r is_ij∈[0,1]。

(2.3) based on equation (4), a correlation matrix of the variables is found:

(2.4) solving a diagonal matrix D based on the correlation matrix R defined by the formula (5):

D＝diag{D_ii} (6)

wherein D is_iiIs the sum of all elements in row i in equation (5):

(2.5) solving a Laplace matrix of the diagonal matrix D:

L＝D^-1/2RD^-1/2(8)

(2.6) performing spectral decomposition on the Laplace matrix:

L＝PΛP^T(9)

wherein, P ═ P₁,P₂,...,P_J]Are orthogonal eigenvectors.

(2.7) selecting the eigenvectors corresponding to the k maximum eigenvalues to form a matrix E ═ P₁,P₂,...,P_k]∈R^J×kNormalizing each row in the matrix E to obtain a matrix Y:

(2.8) clustering Y by using a kmeans clustering algorithm, and if the ith row belongs to the b-th class, carrying out variable X_iDivision into sub-block b X_b. Thus, a plurality of operation variables of the million-kilowatt ultra-supercritical unit are divided into B variable blocks according to the correlation degree:

X＝[X₁X₂… X_B](11)

wherein,

is the (B ═ 1, 2.., B) th variable block, J_bRepresents X_bThe number of variables contained in (1).

(3) The variables in the variable block are further decomposed according to the distribution situation in the sample direction by using an information theory decomposition method based on a Gaussian mixture model, and the step is realized by the following substeps.

(3.1) randomly dividing the variable block into W_bIndividual block:

(3.2) using gaussian mixture model method to obtain W (W1, 2.., W)_b) Probability density of individual variable subblocks:

wherein,

is the number of sub-gaussian components;

is the prior probability of the mth sub-Gaussian component, satisfies

And

is a mean value containing sub-Gaussian components

Sum covariance matrix

Of the parameter set (c).

Is a multivariate gaussian probability density:

wherein, J_b,wIs composed of

The number of medium variables.

(3.3) solving probability density distribution functions of all variables in the sub-blocks:

wherein the variable

J_b,wIs a variable block

The number of the medium variables is equal to or greater than the total number of the medium variables,

represents X_b,iBelonging to sub-blocks

When X_b,iThe conditional probability density of (2).

(3.4) solving variable subblocks

And

KL divergence of (W, v ∈ [1, W ]_b])：

Wherein,

and

are respectively

And

the probability density function of (2) can be calculated by using the formula (13).

(3.5) optimizing the random partitions of step (3.1) using an ant colony algorithm such that the following objective function is maximized:

(3.6) each variable block (B ═ 1, 2.., B) is further divided into sub-blocks by repeating steps (3.2) - (3.5). The original data set X is divided into different sub-blocks:

wherein, the variable block

(4) Based on the variable block result obtained in the steps (2) and (3), firstly, describing the variable sub-block by using a Principal Component Analysis (PCA)

The correlation relationship of each variable in (1):

wherein, P_b,wIs a load matrix, T_b,wIs a principal component matrix.

(5) Principal component matrix T established by Gaussian Mixture Model (GMM) method_b,wThe distribution of (c):

wherein,

is the number of gaussian components;

the weight of the m-th component is represented,

is a mean value containing sub-Gaussian components

Sum covariance matrix

Of the parameter set (c).

(6) For each variable sub-block

Establishing BIP statistics

Wherein,

to represent

Belonging to the m-th component

The probability of (a) of (b) being,

as a principal component matrix T_b,wAn nth (N ═ 1, 2.., N) row vector.

Is based on the probability of local Mahalanobis distance, which is defined as

Wherein,

is composed of

Mahalanobis distance to mth gaussian component, T being T_b,wAny one of the rows.

(7) The relationship between each sub-block in each variable block is monitored by a Gaussian Mixture Model (GMM) method, which is implemented by the following sub-steps.

(7.1) blocking each variable block X_bThe first columns of the pivot matrices of the respective sub-blocks are combined together:

wherein, t_b,w(w＝1,2,...,W_b) As a principal component matrix T_b,wThe 1 st column vector.

(7.2) describing principal metadata with GMM

The distribution of (c):

wherein,

the number of Gaussian components in the b variable sub-block is shown;

the weight of the m-th component is represented,

is composed of sub-Gaussian componentsMean value of

Sum covariance matrix

Of the parameter set (c).

(7.3) Main metadata for each variable Block

Establishing BIP statistic:

wherein,

as a principal component matrix

An nth (N ═ 1, 2.., N) row vector.

To represent

Belonging to the m-th component

The probability of (c).

Is based on the probability of the local mahalanobis distance,

is composed of

Mahalanobis distance to the mth gaussian component,

is composed of

Any one of the rows.

(8) During online fault detection, the process is monitored from three levels of variable subblocks, variable blocks and the whole unit. This step is realized by the following substeps.

(8.1) acquiring new data: and (4) collecting the variable values of the measuring points according to the step (1) and recording as z (1 multiplied by J).

(8.2) according to the variable blocking results obtained in the step (2) and the step (3), carrying out sub-block decomposition on the new data:

z＝[z₁z₂... z_b... z_B](26)

wherein z is_b(B ═ 1, 2.., B) is the B variable sub-block.

(8.3) at the bottom layer, i.e., variable sub-block layer, each sub-block

The data of (2) are projected to the principal element direction of the corresponding sub-block:

wherein

Is a load matrix.

(8.4) obtaining each sub-block

The online statistic index of (1):

wherein, the meaning of each parameter in the above formula is similar to that in the formula (22).

To represent

Belonging to the m-th component

The probability of (c).

Is based on the probability of the local mahalanobis distance,

is composed of

Mahalanobis distance to mth gaussian component, T being T_b,wAny one of the rows.

(8.5) at the variable block level, z is first put_bThe main elements of each variable sub-block are combined together:

(8.6) obtaining each variable block z_bThe online statistic index of (1):

wherein, the meaning of each parameter of the above formula is similar to that in the formula (25).

To represent

Belonging to the m-th component

The probability of (c).

Is based on the probability of the local mahalanobis distance,

is composed of

Mahalanobis distance to the mth gaussian component,

is composed of

Any one of the rows.

(8.7) converting the BIP index of each variable block into the probability of normal (labeled 'N') and fault (labeled 'F'):

wherein BIP_b,lmtA control limit for the statistical BIP indicator;

representing the normal conditional probability of the b variable block;

indicating the conditional probability of the failure of the b-th variable block.

(8.8) calculating the posterior probability of the b variable block failing by the Bayes rule

Wherein, P_b(F)＝α；P_b(N) ═ 1- α represents the prior probability of the process failing or being normal, respectively, at a level of significance α.

(8.9) comprehensively considering the fault probability of all variable blocks and calculating the global monitoring statistic

(9) Judging the running state of the process: and analyzing the process state from three levels of the variable subblocks, the variable blocks and the whole unit. Three levels of statistics are compared with the control limits in real time:

(a) at each variable sub-block

In, if BIP_b,w1-alpha, in sub-blocks

The variable in (b) fails, otherwise the variable in the sub-block is considered to be operating within the normal range.

(b) At the variable block level, if BIP_bIf the value is more than 1-alpha, the related relation of each variable sub-block in the variable block is abnormal, otherwise, all the variables in the b-th sub-block are normally operated.

(c) At the unit level, if PF_zIf the measured value is more than alpha, the abnormal condition or the fault occurs in the running process of the million kilowatt ultra-supercritical unit, otherwise, the whole unit runs normally.

Images