CN115130564A - High-pressure heater online monitoring method based on parallel GMM-LCKSVD - Google Patents

Abstract

The invention discloses a high-pressure heater online monitoring method based on parallel GMM-LCKSVD, which comprises the following steps: collecting high-pressure heater operation parameter data, clustering the data set by adopting a parallel GMM algorithm so as to divide historical data into a plurality of groups of data sets which accord with Gaussian distribution, then establishing an offline model by using a Label consistency-based Dictionary Learning algorithm (LCKSVD), and then carrying out online fault diagnosis on real-time data by adopting a reconstruction-based fault isolation method. The system and the method provided by the invention can effectively carry out real-time online monitoring on the running state of the high-pressure heater of the power plant, and can quickly position the fault parameters after the fault occurs, thereby providing reasonable guidance for subsequent fault elimination and maintenance work.

Description

High-pressure heater online monitoring method based on parallel GMM-LCKSVD

Technical Field

The invention relates to the technical field of fault diagnosis of high-pressure heaters, in particular to an online monitoring method of a high-pressure heater based on parallel GMM-LCKSVD.

Background

The steam-water system is one of core systems of the coal-fired power plant, whether the steam-water system can safely and stably operate directly determines whether the whole unit can normally generate electricity, and the high-pressure heater is important equipment of the steam-water system, so that the heat efficiency of the power plant can be improved, and the coal consumption can be reduced.

If the high-pressure heater can not normally operate, the output of the generator is generally reduced by about 10%, and the heat consumption and the power supply coal consumption of the unit are increased. And meanwhile, the high-pressure fault can cause water impact of the steam turbine, so that the expansion difference of the steam turbine is increased, the safety and stability of boiler operation are reduced, and even the personal safety can be threatened in severe cases. Therefore, the state monitoring and fault early warning of the high-pressure heater have important significance for the safety and economic operation of the unit.

In the last two decades, along with the development of computer storage and communication technologies, thermal power generating units in China have built highly automated Systems such as Distributed Control Systems (DCS), Management Information Systems (MIS), Supervisory Information monitoring Systems (SIS), and the like.

The systems improve the automation level of the whole power plant on one hand, and also store the operation historical data of a large number of unit equipment on the other hand, and the data provides a solid foundation for the online monitoring and fault early warning of the unit equipment.

In recent years, researchers have studied data-driven modeling techniques such as PCA and deep learning to extract relationships between operating parameters from a large amount of historical data of high-voltage operation, and have established a fault diagnosis model by processing and analyzing the relationships.

Along with the continuous improvement of unit parameters, the times of simultaneously participating in deep peak regulation and starting and stopping equipment are more and more frequent, and real-time state monitoring and fault early warning of equipment running in the unit become important components in the construction process of an intelligent power plant.

However, these methods do not acquire sparse representation of data, and it is difficult to mine essential features of the data.

Meanwhile, as the thermal power generating unit frequently participates in deep peak shaving, the high-pressure heater can be operated in a wide load section for a long time, so that historical data of the operation of the high-pressure heater can show a plurality of different distribution modes, and an ideal effect can not be achieved even if a sample is not processed and the process monitoring model is established by directly using data of all histories.

Therefore, the distribution characteristics of the operation data of the high-pressure heater are deeply researched, and the essential characteristics of the data are fully extracted, so that the method has important practical significance for improving the monitoring effect of the model.

Disclosure of Invention

The technical problems to be solved by the invention are as follows: the invention provides a novel high-pressure heater online monitoring method and a novel fault diagnosis method based on a parallel GMM and label consistent dictionary learning algorithm, which aim to solve the problem that the essential characteristics of the operating data of a high-pressure heater are difficult to dig out in the prior art and meet the requirements of real-time state monitoring and fault early warning on the high-pressure heater.

The technical scheme of the invention is as follows: the invention discloses a high-pressure heater online monitoring method based on parallel GMM-LCKSVD, which comprises the following steps:

s1, fast clustering original data by adopting a parallel GMM algorithm, and dividing the data into a plurality of groups of data sets;

s2, establishing a model by using an LCKSVD dictionary learning algorithm, and carrying out online monitoring on the high-pressure heater;

and S3, after the occurrence of the fault is monitored, positioning the parameter of the fault occurrence by using a fault isolation method based on reconstruction.

Further, the step S1 specifically includes the following steps:

SA1. clustering a data set using a parallel GMM algorithm:

suppose data sample set Y ═ Y ₁ ,y ₂ ,...y _m }, in which: m represents the number of samples, sample y _i ＝{y _i1 ,y _i2 ,...y _in Subscript i represents a sample number, and n represents the number of parameters;

dividing a data sample set Y into N independent parts at random, and designating clustering number K;

for the

And

use of

And

updating the Gaussian mixture model parameters of each part;

in the formula: m is a unit of _t Represents the number of samples of the t-th part;

l _g,i the calculation method is as follows:

in the formula: omega _g A weight for each gaussian portion; sigma _g Is a covariance; mu.s _g As desired;

SA2. parallel computing N parts of mixed model parameters by using OPENMP, and after computing, using

To update the parameters of the entire GMM model;

SA3, repeating the calculation process until the iteration times are greater than the preset maximum iteration times or the error of the likelihood function is smaller than the given error and ending;

and SA4, after the calculation is finished, calculating the posterior probability of each sample through a Bayes formula, and defining the maximum posterior probability as the clustering label of the sample.

Further, the specific steps of S2 are as follows:

SB1, establishing an LCKSVD dictionary learning model:

according to the clustering result of the parallel GMM algorithm, a sample label matrix H of a data sample set Y is constructed, then Y is standardized and transposed, and an LCKSVD dictionary learning objective function about the data sample set Y is constructed, so that a dictionary D is obtained;

wherein the LCKSVD dictionary learning objective function for the data sample set Y is:

in the formula: d is a dictionary matrix; x is a sparse matrix; t is the number of elements in X which are not 0; a and W are matrixes to be solved;

q is a k × n dimensional matrix if y _i And d _k Is the same, then the corresponding element of Q is 1; otherwise, the element of Q is 0; k is the number of dictionaries;

SB2. statistical index of calculation model and threshold value delta thereof ² ：

Having obtained the dictionary D, y for each sample _i According to

Calculating the square prediction error SPE of the training data set, and obtaining the square prediction error set SPE of the whole training data set _list ；

Using nuclear density estimation method for the SPE _list Carrying out nonparametric estimation, and taking the corresponding value of 99% of the area of the probability density function as the threshold value delta of the model ² ；

And SB3, carrying out online monitoring on the real-time data:

for real-time data samples y _new Solving using FISTA algorithm

And obtaining a sparse matrix X based on the optimization problem

Calculating the square prediction error SPE;

if SPE>δ ² The occurrence of a failure is considered, and the flow proceeds to step S3;

if SPE is less than or equal to delta ² No failure is deemed to have occurred and the process loops to step S2.

Further, the specific steps of S3 are as follows:

sc 1: initializing a parameter set S {1,2, … n }, SPE set SPEs { }, and a fault parameter set fs { };

and (3) Sc 2: for the

Initializing beta, wherein beta _i ＝0，

In the formula: beta represents a direction vector in which the kth element is 0 and all other elements are 1;

solving for

And adding the obtained SPE into the SPES set;

sc 3: calculating the minimum value of SPES SPE _min Adding i into a fault parameter set fs with a parameter subscript i corresponding to the parameter subscript i in Sc 2;

fruit SPE _min >δ ² Or the number of fs elements is less than n, updating the parameter set by using S-fs, emptying the SPES error set, and returning to Sc2 for next calculation;

if SPE _min ＜δ ² Or if the number of the fs elements is equal to n, entering Sc 4;

sc 4: and outputting a fault parameter set fs.

Has the advantages that: compared with the prior art, the invention has the following advantages:

1. the system and the method provided by the invention can rapidly perform sample clustering on the historical data of the high-pressure heater, and fully extract the essential characteristics of the data by using a dictionary learning algorithm, thereby realizing the real-time monitoring and fault early warning of the running state of the high-pressure heater, and after the fault occurs, if the correct fault occurrence direction and fault amplitude are found, the SPE value of the whole system is inevitably smaller than the threshold value delta of the model ² (ii) a For sample y _i Can in turn assume that a parameter has failed, and if the assumption is correct, the parameter is changed from y _i After the removal, the calculated SPE value is necessarily smaller than the threshold δ for the remaining m-1 parameters ² Based on the above analysis, the fault isolation problem can be converted into equation

And using a FISTA algorithm to carry out rapid solution; improveEconomy and safety of plant operation.

2. The system and the method provided by the invention solve the problem of fault isolation of the dictionary learning model after monitoring a fault, and can provide reasonable guidance and decision for subsequent maintenance work.

Drawings

FIG. 1 is a flow chart of the parallel GMM-LCKSVD-based high-pressure heater online monitoring method of the invention;

FIG. 2 is a high tip differential and outlet temperature parameter distribution scatterplot in accordance with an embodiment of the present invention;

FIG. 3 is a diagram of FGMM algorithm monitoring results of high-voltage leakage faults in the embodiment of the present invention;

FIG. 4 is a diagram of a monitoring result of the GMM-LCKSVD algorithm with high plus leakage fault in the embodiment of the present invention;

FIG. 5 is a diagram of FGMM algorithm monitoring results of short circuit faults of the high water inlet chamber and the high water outlet chamber in the embodiment of the invention;

FIG. 6 is a diagram of the monitoring result of the GMM-LCKSVD algorithm for short circuit fault of the high water inlet and outlet chambers in the embodiment of the invention.

Detailed Description

For the understanding of the present invention, the following detailed description of the present invention is given with reference to the accompanying drawings, which are provided for illustration purposes only and are not intended to limit the scope of the present invention.

Taking a No. 1 high-pressure heater in a thermodynamic system of a 600MW power plant as an object to perform example analysis, selecting 14 measuring points shown in Table 1 as modeling parameters, wherein the sampling interval is 10s, respectively acquiring 2000 samples to perform model training and testing, and equally dividing the test samples into two groups, namely A and B.

Wherein, the first 750 samples of the group A are normal operation data, and the last 250 samples are high leakage fault data; the first 750 of group B are normal run samples and the last 250 are fault data for high charge inlet and outlet chamber shorts.

Figure 2 shows the results of clustering high plus end difference, outlet temperatures using the parallel GMM algorithm. It can be seen that along the direction of load change, the parallel GMM algorithm divides the data into three classes, which are completely consistent with the actual distribution characteristics of the data.

TABLE 1 survey Point List and statistical characteristics

The GMM-LCKSVD method of the invention is used for establishing a process monitoring model, and the monitoring result is compared with the FGMM model, the monitoring results of the two algorithms are shown in figures 3-6, the solid line in the figures represents the limit value obtained by calculating each sample, and the dotted line represents the threshold value delta of the model ² 。

Table 2 gives the fault diagnosis results.

TABLE 2 high-voltage fault isolation results

The results show that:

1. it can be seen from fig. 3 and 5 that after the fault occurs, the sample limit BIP of the FGMM model is lower than the model threshold δ ² Indicating that the algorithm cannot monitor the occurrence of the fault; meanwhile, it can be seen that, before the 250 th failure sample, the FGMM model has a situation where a large number of sample limits are higher than the model threshold, which indicates that the FGMM algorithm regards the normal sample as the failure data at this time, and the monitored error rate is very high.

2. As can be seen from FIGS. 4 and 6, the GMM-LCKSVD model further extracts data features due to the use of a dictionary learning algorithm.

Therefore, after the fault occurs, the sample limit SPE is higher than the model threshold delta ² The method shows that the algorithm well monitors the occurrence of the fault, can monitor the high voltage fault in real time, and has low monitoring error rate.

On the basis, the fault parameter can be accurately positioned by adopting a fault isolation method based on reconstruction, and a brand-new reliable method is provided for high-voltage online monitoring and fault diagnosis.

The present invention will be further described with reference to the accompanying drawings.

The parallel GMM-LCKSVD-based high-pressure heater online monitoring method comprises the following steps of:

(1) clustering the data set Y using a parallel GMM algorithm:

assume that the data sample set Y ═ Y ₁ ,y ₂ ,...y _m }, wherein: m represents the number of samples, sample y _i ＝{y _i1 ,y _i2 ,...y _in Subscript i represents a sample number, and n represents the variable number of the sample; the sample Y is randomly divided into N independent parts and the number of clusters K is specified. For the

Use of

And

the gaussian mixture model parameters of each section are updated. In the formula: m is _t Represents the number of samples of the t-th part; l _g,i The calculation method is as follows:

in the formula: omega _g A weight for each gaussian component; sigma _g Is a covariance; mu.s _g As desired. Parallel computing is carried out on N parts by using OPENMP, and after the N parts are all computed, the parallel computing is carried out

To update the parameters of the entire GMM model. The above calculation process is repeated until the iteration number is larger than the maximum iteration number or the error of the likelihood function is smaller than the given error and is terminated. After the calculation is finished, the posterior probability of each sample is calculated through a Bayes formula, and the maximum posterior probability is defined as the clustering label of the sample.

(2) Establishing an LCKSVD dictionary learning model:

firstly, a sample label matrix H of training data Y is constructed according to a clustering result of a parallel GMM algorithm, then Y is standardized and transposed, and an LCKSVD dictionary learning target function related to the training data Y is constructed:

in the formula: d is a dictionary matrix; x is a sparse matrix; t is the number of elements which are not 0 in X; q is a k × n dimensional matrix if y _i And d _k If the tags of the Q are the same, the corresponding element of Q is 1, otherwise, the element of Q is 0; and k is the number of dictionaries.

In order to solve the parameters of LCKSVD dictionary learning by using KSVD algorithm, the method can be used for solving the parameters of LCKSVD dictionary learning

The objective function can be written as:

it can be seen that the target function has only 2 unknown parameters after transformation and the form is completely consistent with the target function of the KSVD algorithm. However, the KSVD algorithm must be solved by firstly giving the initial value of the dictionary D. To this end, it is possible to construct

And obtaining an initialization value of a matrix as a ═ QX ^T (XX ^T +λI) ^-1 ；

Similarly, a ridge regression model for W is constructed, and the initial value of the W matrix is obtained as W ═ HX ^T (XX ^T +λI) ^-1 ；

After the initial values of A and W are obtained, the initial value of the dictionary D can be obtained, and the KSVD algorithm is used for solving the objective function so as to obtain the dictionary D.

(3) MeterStatistical index of calculation model and threshold value delta thereof ² ：

Having obtained the dictionary D, y for each sample _i According to | y _i -Dx _i || _F ² Calculating the square prediction error SPE of the training data set, and obtaining the square prediction error set SPE of the whole training data set _list For SPE, kernel density estimation _list Carrying out nonparametric estimation, and taking the value corresponding to 99% of the area of the probability density function as the threshold value delta of the model ² ；

(4) And (3) fault monitoring is carried out on the real-time data:

for real-time data samples y _new Solving using FISTA algorithm

And obtaining a sparse matrix X according to the problem of optimization and obtaining a sparse matrix X according to the value of y _i -Dx _i || _F ² Calculate its SPE, if SPE > δ ² Judging that a fault occurs, and judging that the sample is normal by the anti-regularities;

(5) the method for fault diagnosis of fault data by using a reconstruction-based method comprises the following specific steps:

step 1: initializing a parameter set S {1,2, … n }, SPE set SPEs { }, and a fault parameter set fs { };

step 2: for the

Initializing a direction vector β, wherein β _i ＝0，

Solving using FISTA algorithm

And adding the obtained SPE into the SPES set;

and step 3: calculating minimum SPES _min And adding i into a fault parameter set fs according to the corresponding parameter index i in the step 2, if SPE is used _min ＜δ ² Or elements of fsIf the number is equal to n, entering a step 4, otherwise, updating the parameter set by using S-fs, clearing the SPES of the error set, and returning to the step 2 to calculate the next time;

and 4, step 4: and outputting a fault parameter set fs.

The specific implementation steps of the online fault diagnosis method provided by the invention are described below by taking a number 1 high reheater in a 600MW power plant thermodynamic system as an example.

With reference to the flow shown in FIG. 1, the steps are as follows:

1. collecting historical data of operation of a large number of devices through sensors installed in the on-site high-pressure heater;

2. clustering original data by adopting a parallel GMM algorithm, and dividing the data into a plurality of groups of variable data sets;

3. constructing a model by using an LCKSVD dictionary learning algorithm, and monitoring real-time data of the high-pressure heater by using a trained LCKSVD model;

4. and after the occurrence of the fault is monitored, positioning the parameter of the fault occurrence by adopting a fault isolation method based on reconstruction.

It is further understood that modifications and equivalents of the disclosed embodiments may occur to persons skilled in the art without departing from the spirit and scope of the disclosed embodiments, and are to be included within the scope of the appended claims.

Claims (4)

1. A high-pressure heater on-line monitoring method based on parallel GMM-LCKSVD is characterized in that: the method comprises the following steps:

s1, adopting a parallel GMM algorithm to perform rapid clustering on original data, and dividing the data into a plurality of groups of data sets;

s2, establishing a model by using an LCKSVD dictionary learning algorithm, and carrying out online monitoring on the high-pressure heater;

and S3, after the occurrence of the fault is monitored, positioning the parameter of the fault occurrence by using a fault isolation method based on reconstruction.

2. The parallel GMM-LCKSVD-based high-pressure heater online monitoring method according to claim 1, characterized in that: the step of S1 specifically includes the following steps:

SA1. clustering a data set using a parallel GMM algorithm:

assume that the data sample set Y ═ Y ₁ ,y ₂ ,...y _m }, wherein: m represents the number of samples, sample y _i ＝{y _i1 ,y _i2 ,...y _in Subscript i represents a sample number, and n represents the number of parameters;

dividing a data sample set Y into N independent parts at random, and designating clustering number K;

for the

And

use of

And

updating the parameters of the Gaussian mixture model of each part;

in the formula: m is _t Represents the number of samples of the t-th part;

l _g,i the calculation method of (A) is as follows:

in the formula: omega _g A weight for each gaussian component; sigma _g Is a covariance; mu.s _g As desired;

SA2. parallel computing N parts of mixed model parameters by using OPENMP, and after computing, using

To update the parameters of the entire GMM model;

SA3, repeating the calculation process until the iteration times are greater than the preset maximum iteration times or the error of the likelihood function is smaller than the given error and ending;

and SA4, after the calculation is finished, calculating the posterior probability of each sample through a Bayes formula, and defining the maximum posterior probability as the clustering label of the sample.

3. The parallel GMM-LCKSVD-based high-pressure heater online monitoring method according to claim 1, characterized in that: the specific steps of S2 are as follows:

SB1, establishing an LCKSVD dictionary learning model:

according to the clustering result of the parallel GMM algorithm, a sample label matrix H of a data sample set Y is constructed, then Y is standardized and transposed, and an LCKSVD dictionary learning target function about the data sample set Y is constructed, so that a dictionary D is obtained;

wherein the LCKSVD dictionary learning objective function for the data sample set Y is:

s.t.||x|| ₀ ≤T

in the formula: d is a dictionary matrix; x is a sparse matrix; t is the number of elements in X which are not 0; a and W are matrixes to be solved;

q is a k × n dimensional matrix if y _i And d _k Is the same, the corresponding element of Q is 1; otherwise, the element of Q is 0; k is the number of dictionaries;

SB2. statistical index of calculation model and its threshold value delta ² ：

Having obtained the dictionary D, y for each sample _i According to

Calculating the square prediction error SPE of the training data set, and obtaining the square prediction error set SPE of the whole training data set _list ；

Using nuclear density estimation method for the SPE _list To carry out non-Estimating parameters, and taking the corresponding value at 99% of the area of the probability density function as the threshold value delta of the model ² ；

SB3, carrying out on-line monitoring on real-time data:

for real-time data samples y _new Solving using FISTA algorithm

And obtaining a sparse matrix X based on the optimization problem

Calculating the square prediction error SPE;

if SPE>δ ² The occurrence of a failure is considered, and the flow proceeds to step S3;

if SPE is less than or equal to delta ² No failure is deemed to have occurred and the process loops to step S2.

4. The parallel GMM-LCKSVD-based high-pressure heater online monitoring method according to claim 1, characterized in that: the specific steps of S3 are as follows:

sc 1: initializing a parameter set S {1,2, … n }, SPE set SPEs { }, and a fault parameter set fs { };

and (3) Sc 2: for the

Initializing beta, wherein beta _i ＝0，

In the formula: β represents a direction vector in which the kth element is 0 and all other elements are 1;

solving for

And adding the obtained SPE into the SPES set;

sc 3: calculating the minimum value of SPES SPE _min And their correspondence in Sc2Adding i into a fault parameter set fs;

fruit SPE _min >δ ² Or the number of fs elements is less than n, updating the parameter set by using S-fs, emptying the SPES error set, and returning to Sc2 for next calculation;

if SPE _min ＜δ ² Or if the number of the fs elements is equal to n, entering Sc 4;

sc 4: and outputting a fault parameter set fs.

Images