CN115422714A - Knowledge condition hybrid driving running state monitoring method for gas turbine - Google Patents
Knowledge condition hybrid driving running state monitoring method for gas turbine Download PDFInfo
- Publication number
- CN115422714A CN115422714A CN202210917053.4A CN202210917053A CN115422714A CN 115422714 A CN115422714 A CN 115422714A CN 202210917053 A CN202210917053 A CN 202210917053A CN 115422714 A CN115422714 A CN 115422714A
- Authority
- CN
- China
- Prior art keywords
- condition
- monitoring
- sample
- static
- knowledge
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
Abstract
The invention discloses a knowledge condition hybrid driving operation state monitoring method for a gas turbine. The method is characterized in that a knowledge matrix is constructed by combining expert knowledge and correlation analysis between variables of the gas turbine, a plurality of sub-modules are divided from an integral object, and the selection of the sub-module variables is guided by the knowledge matrix. On the basis, a condition driving idea is introduced, a large-range non-stable continuous process is decomposed into a plurality of condition modes, and a state monitoring model is respectively established on each condition mode, so that the running state monitoring under a variable working condition scene is realized. And finally, determining a weight coefficient of the submodule by combining the condition indicating variable, and fusing monitoring results of the submodule to realize comprehensive monitoring of the running state of the complex industrial equipment. The invention not only can enable expert knowledge to be better applied to non-stationary process monitoring, but also enhances the accuracy and reliability of actual online monitoring of the gas turbine, and provides a new efficient and reliable idea for monitoring the running state of complex industrial equipment.
Description
Technical Field
The invention belongs to the field of monitoring of the running state of complex industrial equipment which runs non-stably in a large range, and particularly relates to a method which combines knowledge driving and condition driving, constructs a knowledge matrix through expert knowledge and variable correlation analysis, divides a plurality of subcomponents on an integral object, establishes condition-driven monitoring models respectively, and finally fuses monitoring results according to condition indication variable values to realize intelligent monitoring of the running state of a gas turbine.
Background
The gas turbine is an internal combustion type power machine which drives turbine blades to rotate through fuel combustion so as to generate useful work, is a heat-power conversion type power generation device with the highest efficiency, is widely applied to the fields of aviation, ships, power generation and the like, and plays an important role as a core device. In practical application, taking the field of power generation as an example, a gas turbine generator set often participates in peak shaving of a power grid, so that the load of a gas turbine is variable, the start and stop are frequent, a plurality of steady-state working conditions and transition working conditions are included in a normal running state of the gas turbine generator set, the gas turbine generator set has obvious variable working condition characteristics, and how to distinguish a fault running state from a normal running state under the background of the variable working conditions is a difficult problem in a gas turbine state monitoring task. In addition, the gas turbine is taken as a complex industrial system, the structure of the gas turbine comprises a plurality of sub-components and a large number of measuring points, and the components are mutually influenced during operation, so that certain coupling relations exist among the components, the measuring points and the measuring points, and the difficulty in condition monitoring of the gas turbine is greatly improved. However, the research on the gas turbine monitoring technology in China is not sufficient at present, the gas turbine industry is very immature, and the research work on the gas turbine monitoring technology needs to be promoted urgently.
The existing gas turbine operation state monitoring methods are mainly divided into a digital signal analysis method and a data driving method. The early state monitoring method is a digital signal analysis method, which adopts methods such as autocorrelation coefficient, fourier transform and the like to analyze time domain and frequency domain signals of pressure, vibration, sound and the like of the gas turbine, but the digital signal analysis method has higher requirements on sensor data, and needs to acquire the operation data of the gas compressor in a fault state, thus having certain damage on equipment; however, the types of the models are very diverse, such as a bayesian model, a Support Vector Machine (SVM), an Extreme Learning Machine (ELM), a stacked noise reduction self-encoder (SDAE), a fuzzy clustering algorithm model and the like, but the working conditions of multiple default objects in the conventional data-driven monitoring method are stable, namely, the monitoring is only carried out on a specific working condition, but because the actual operating state of the gas turbine is always a variable working condition and comprises multiple transition sections, the method has the typical large-range non-stationary characteristic, and the method cannot be suitable for monitoring the operating state of the gas turbine in a large-range non-stationary scene, and therefore, the method cannot be well applied to the actual industrial process.
In recent years, to solve the problem of monitoring the operating state in a wide range of non-stationary scenes, a condition-driven operating state monitoring method has received more and more attention. A large-range non-steady transient continuous process is divided into a plurality of condition modes, and refined process analysis is carried out on each condition mode to comprehensively describe the potential distribution of data, so that the model can adapt to variable input with large-range change, and a better monitoring effect is achieved under various different working conditions. State monitoring of large-range non-stationary processes has been studied in the industrial field, zhao et al reconstructs non-stationary sequences in the time dimension into stationary sequences in the condition dimension, and after analyzing static and dynamic information of process operation, defines a new Bayesian inference statistic for online monitoring. Although there are some successful cases in industrial application, for large complex industrial equipment such as gas turbine with high complexity and strong coupling, there is obvious mutual influence relationship between coupling between components and measured point variables, and establishing a monitoring model on an overall object by directly using all variables cannot well reflect the operation state of subcomponents, so that the monitoring result lacks accuracy and interpretability. In this case, the pure condition-driven monitoring method may not achieve a satisfactory operation state monitoring effect.
In summary, for an actual complex industrial process object, which often has a complex internal structure, the whole object is composed of a plurality of sub-components, and complicated coupling relationships exist among variables of each component; meanwhile, the operation state thereof is generally transient, and in a certain state, it stays for only a short time, and the state may continuously change, thereby exhibiting typical dynamic characteristics. Therefore, an operation state monitoring method combining knowledge driving and condition driving is designed, a molecule division module is divided by means of expert knowledge, and a knowledge matrix is constructed to guide the establishment of a condition monitoring model, so that the influence of coupling among components and large-range non-stationary characteristics on a monitoring result is reduced, and the accuracy and reliability of the monitoring result are improved.
Disclosure of Invention
The invention aims to provide a knowledge matrix and condition drive mixed running state monitoring method for a large-range non-stationary process, aiming at the defects of the existing running state monitoring technology for the large-range non-stationary process. The method comprises the steps of constructing a knowledge matrix among variables by combining expert knowledge and correlation relations among measured point variables, dividing sub-modules from an integral object, and selecting main variables and related variables of each sub-module by combining the knowledge matrix. On the basis, the method introduces a data reconstruction strategy, reconstructs data in a large-range non-stationary process into a stationary sequence on a condition dimension, extracts static and dynamic information from the stationary sequence, and further establishes a condition-driven monitoring model on each divided sub-module. The method determines the weight of the monitoring result of each submodule by comparing the correlation coefficient between the main variable of each submodule and the main variable of the whole object so as to obtain the final monitoring result. The invention provides a novel analysis angle for monitoring the running state of a large-range non-stationary process, not only introduces the mechanism knowledge of a gas turbine object in the real world, but also enhances the reliability of a monitoring result by combining knowledge driving and condition driving, is beneficial to an industrial engineer to accurately judge the running state of the large-range non-stationary transient continuous process of the gas turbine, and finds faults in time, thereby ensuring the safety of actual production.
The purpose of the invention is realized by the following technical scheme:
a knowledge condition hybrid driving operation state monitoring method for a gas turbine specifically comprises the following steps:
constructing a monitoring model, comprising the following steps:
(1) Acquiring N samples of a normal continuous operation process of a gas turbine system, wherein each sample comprises J measured variables;
(2) Calculating vector X consisting of N samples of any two measured variables m ，X n Coefficient of correlation between corr (X) m ，X n ) Constructing a knowledge matrix M according to the calculated correlation coefficient and by combining with the mechanism knowledge of the object, wherein the mth row and the nth column of elements M of the knowledge matrix mn Represents the variable X m And X n Whether the two are related; 1 represents association, 0 represents no association;
(3) Dividing the whole object into S submodules according to mechanism knowledge of the gas turbine and the physical structure of the object; and according to the established knowledge matrix M and the object mechanism knowledge, selecting a condition indicating variable X capable of indicating the process running state to the maximum extent on each submodule s s,c And other related variables X associated with the variable s,o Jointly forming a two-dimensional sample matrix X of submodules s s ＝[X s,c ，X s,o ](ii) a Simultaneously, according to the established knowledge matrix M and the mechanism knowledge of the object, selecting a condition indicating variable X capable of indicating the operation state of the whole process of the object to the maximum extent c ；
(4) Two-dimensional data matrix X for each submodule s Calculating the first difference of each sample from the previous sample
Combining each sample with its corresponding first order difference as a new sample
All new samples obtain a dynamic two-dimensional sample matrix
(5) For each sub-module sample matrix X s 、
Rearranging and dividing the samples into K according to the sequence of the condition indicating variable values from small to large to form K static condition pieces
And K dynamic condition pieces
Wherein N is k,s The number of samples in the kth conditional slice of the submodule s, K =1, \ 8230;, K; then, the mean value removing processing is carried out on each condition piece to obtain a data matrix of the standardized condition pieces
And
merging the standardized static condition pieces to obtain c s Each static condition section consists of a plurality of continuous standardized static condition pieces, and the control limit calculated by each static condition section based on the PCA model is smaller than the product of the control limit calculated by each static condition piece in the static condition section based on the PCA model and a constant proportionality factor gamma; establishing a monitoring model and a control limit Ctrl of each static condition section c,s ；
C is likewise marked off according to the breakpoint of the static condition segment s Each dynamic condition section, and establishing a monitoring model and a control limit of each dynamic condition section
(6) Dividing the integral condition indicating variable into Z intervals, and calculating the condition indicating variable X of each submodule in each interval Z s,c With an overall condition indicating variable X c The proportion of the correlation coefficient is used as the weight coefficient of each submodule when the overall condition indicating variable is in the interval z
The on-line monitoring comprises the following steps:
obtaining a sample x to be measured in real time new (1 XJ) and processed to obtain a first order difference
And its sample x on each sub-module new,s (ii) a Judging the section z of the sample according to the overall condition indicating variable value of the sample new And judging the static condition section and the dynamic condition section according to the sub-module condition indication variable value of the sample, and monitoring the model, control limit and the section z according to the static condition section and the dynamic condition section new Calculating the monitoring index and the control limit of the current sample to be tested by the weight coefficient, and considering that the gas turbine operates in a normal state when and only when the static and dynamic monitoring indexes are smaller than the control limit; otherwise, it is regarded that the abnormality occurs in the gas turbine
Further, the measured variables comprise a plurality of preset module natural gas volume flow, preset module natural gas mass flow, compressor inlet temperature, compressor outlet pressure, compressor outlet temperature, compressor bearing temperature, compressor thrust shoe bearing generator end temperature, compressor bearing vibration, compressor side large shaft vibration, compressor air inlet channel pressure difference, compressor anti-freezing device air inlet electric regulating valve position, gas turbine exhaust average temperature, gas turbine side large shaft vibration, combustion chamber pressure difference, gas turbine cooling air regulating valve position 1, gas turbine cooling air regulating valve position 2, gas turbine humming, gas turbine rotating speed, gas turbine power, gas turbine 2-stage stationary blade retaining ring cavity cooling air pressure and gas turbine 3-stage stationary blade retaining ring cavity cooling air pressure.
Further, the dividing the whole object into S sub-modules specifically includes:
the whole object is divided into three submodules of a gas compressor, a combustion chamber and a gas turbine.
Further, the establishing each static barMonitoring model and control limit Ctrl of segment c,s The method comprises the following specific steps:
establishing PCA model of each static condition segment
Wherein the content of the first and second substances,
is the data matrix for the c-th static condition segment of the s-th module,
to represent
The corresponding main component(s) is (are),
representing a PCA conversion matrix corresponding to the c static condition section of the s module;
for each static condition segment, establishing a Gaussian mixture GMM model in a feature space, and determining the number v of Gaussian elements by using an F-J algorithm c And then using EM algorithm to estimate GMM parameters
GMM parameter for the ith gaussian element of static condition segment c;
namely a monitoring model of the static condition section c;
constructing a probability density function g based on a PCA model of a static condition section and GMM parameters:
wherein the content of the first and second substances,
is the parameter of the probability density function, is the number of Gaussian elements in the feature space of the static condition section c,
is the prior probability of the ith gaussian element of conditional segment c,
GMM parameter for the ith gaussian element of conditional segment c;
calculating a Bayes inferred distance index of the static condition segment c:
wherein, N c,s For the number of samples of condition segment c in sub-module s,
is the posterior probability that sample j belongs to the ith gaussian component,
is T j,c,s Mahalanobis distance to the ith gaussian component; t is a unit of j,c,s To represent
Row j of (1); according to the confidence coefficient 1-alpha, calculating BID by a nuclear density estimation method c,s Control limit Ctrl of c,s 。
Further, the static condition section and the dynamic condition section which the sub-module condition indicating variable value of the sample belongs to are judged according to the sub-module condition indicating variable value of the sample, and the control limit and the section z which the sub-module condition indicating variable value belongs to are determined according to the monitoring model and the control limit of the static condition section and the dynamic condition section which the sub-module condition indicating variable value of the sample belongs to new When calculating the weight coefficient ofThe monitoring indexes and the control limits of the sample to be detected are as follows:
calculating the monitoring statistics of the sample to be detected on the submodule:
wherein, c s Is the condition segment serial number of the sample to be tested in the submodule s,
denotes c s The monitoring model corresponding to the condition section is used,
the characteristic value of the sample to be detected is represented as follows:
according to the interval z new Calculating the static monitoring index BID new Dynamic monitoring index
And corresponding control limit Ctrl new 、
Compared with the prior art, the invention has the beneficial effects that: the invention provides a new research idea for monitoring the state of a large-range non-steady continuous process. The method comprises the steps of designing a condition indication variable selection strategy based on knowledge driving, introducing a knowledge driving concept into a condition driving monitoring method for the first time, constructing a knowledge matrix by combining object expert knowledge and correlation relations among variables, dividing sub-modules, and determining condition variables and correlation variables of the sub-modules under the guidance of the knowledge matrix, thereby providing a basis for further analysis; by introducing a data reconstruction strategy, rearranging data along a condition direction, dividing a plurality of condition sections, and further establishing a monitoring model on each condition section to enable the algorithm to be suitable for various different working conditions; considering the difference of the influence of the submodule on the whole operation state under different working conditions, an intelligent fusion method of the submodule monitoring result is designed, the weight coefficient of the submodule is determined by utilizing the correlation among the condition variables under different working conditions, so that the whole monitoring result is obtained through fusion, and the intelligent monitoring of the operation state with the fusion of the knowledge matrix and the condition drive is realized. The method is successfully applied by detailed experimental research in the actual industrial process. The method enhances the application of priori knowledge to specific objects by establishing a knowledge matrix for a large-range non-stationary sequence, and combines knowledge driving and condition driving, thereby avoiding the problem of coupling among components in the traditional monitoring method, improving the accuracy and reliability of process monitoring, finally being applied to a gas turbine combined cycle power generation site and ensuring the safety and reliability of the operation process of the gas turbine.
Drawings
FIG. 1 is a flow chart of the present invention knowledge matrix and condition driven hybrid operational status monitoring;
FIG. 2 is a diagram of variable lists and knowledge matrices in accordance with an embodiment of the present invention, wherein (a) is the name and corresponding number of the measured variable and (b) is the knowledge matrix;
FIG. 3 is a schematic diagram of a condition indicating variable table for each component;
fig. 4 shows the monitoring results of the method of the present invention and the comparison results with the conventional monitoring method in the embodiment of the present invention, wherein (a) is the monitoring results of the static indicators of the fault cases, (b) is the monitoring results of the dynamic indicators of the fault cases, and (c) is the comparison results of the method of the present invention and the conventional monitoring method.
Detailed Description
The invention is further described with reference to the following drawings and specific examples.
Gas turbine combined cycle power plants are a highly complex industrial process with time varying, dynamic and non-stationary characteristics. The heavy-duty gas turbine for power generation consists of three core components, namely a gas compressor, a combustion chamber and a gas turbine, wherein the gas compressor and the gas turbine are designed in a multistage axial flow mode. The working principle is as follows: air enters from an air inlet of the gas turbine, and is accelerated and pressurized by the high-speed rotation of the multistage rotor blades of the gas compressor, and is sent into a combustion chamber after being compressed; then fuel (gas or liquid fuel) is injected into the combustion chamber to be mixed with high-temperature compressed air, and combustion is carried out under constant pressure; the generated high-temperature and high-pressure flue gas expands after being combusted and heated, enters a turbine area and pushes a power blade to rotate at a high speed until the flue gas is discharged from a gas outlet to become waste gas; the exhaust gas is reused, for example, in a gas turbine power plant, fed to a waste heat boiler and formed into a combined cycle. For a gas turbine, the change of the load of a unit causes the power of the gas turbine to change, so that the running state of the unit is changed continuously, and the typical wide-range non-smooth transient characteristic is shown. Aiming at the characteristics, the invention provides a method for monitoring the running state of the knowledge matrix and condition driving hybrid oriented to the large-range non-stationary process, and fig. 1 shows a flow chart of the running state monitoring of the knowledge matrix and condition driving hybrid, wherein the method comprises the following steps:
(1) Obtaining N samples of the normal continuous operation process of the gas turbine system, wherein each sample comprises J measured variables, and a two-dimensional data matrix Y (N multiplied by J) can be obtained;
in this example, about 3000 samples were collected from a power plant in Zhejiang for modeling during normal operation, with the following 22 variables measured: the method comprises the steps of pre-module natural gas volume flow, pre-module natural gas mass flow, compressor inlet temperature (environment temperature), compressor outlet pressure, compressor outlet temperature, compressor bearing temperature, compressor thrust tile bearing generator end temperature, compressor thrust tile bearing engine end temperature, compressor bearing vibration, compressor side large shaft vibration, compressor inlet passage pressure difference, compressor anti-freezing device air inlet electric regulating valve position, gas turbine exhaust average temperature, gas turbine side large shaft vibration, combustor pressure difference, gas turbine cooling air regulating valve position 1, gas turbine cooling air regulating valve position 2, gas turbine humming, gas turbine rotation speed, gas turbine power, gas turbine 2-stage stationary blade retaining ring chamber cooling air pressure and gas turbine 3-stage stationary blade retaining ring chamber cooling air pressure.
(2) The condition indicating variable selection based on knowledge driving specifically comprises the following sub-steps:
(2.1) dividing the sub-modules: dividing the whole object into S submodules according to the mechanism knowledge and the physical structure of the object;
in this example, the following 3 sub-modules are divided from the gas turbine: the gas compressor, the combustion chamber and the gas turbine.
(2.2) establishing a knowledge matrix: for the two-dimensional data matrix Y in the step 1, calculating any two measured variables X m ，X n (i.e., the m, n th column data of the two-dimensional data matrix Y):
wherein, E (-) and D (-) respectively represent the mean and variance of the measured variable in the N samples; corr (X) m ，X n )∈[-1,1]As a measured variable X m ，X n Positive values represent the correlation coefficient betweenThe two variables are positively correlated, a negative value represents that the two variables are negatively correlated, and the more the absolute value of the correlation coefficient is close to 1, the stronger the linear correlation between the two variables is.
According to the calculated correlation coefficient and the mechanism knowledge of the object, the incidence relation between the measured variables can be obtained, and then a knowledge matrix M = [ M ] is obtained mn ] J×J And has:
wherein M is mn Represents the variable X m And X n Whether the two are related; delta is a correlation coefficient threshold value, and when the value is exceeded, an obvious linear correlation relationship exists between the two variables; x m ∝X n Represents the variable X m And X n There is a mechanistic association;
(2.3) selecting submodule variables: for the S submodules divided in the step 2.1, according to the knowledge matrix M constructed in the step 2.2 and the object mechanism knowledge, selecting a condition indicating variable X capable of indicating the process running state of each submodule to the maximum extent on each submodule S s,c And other related variables X associated with the variable s,o Jointly forming a two-dimensional sample matrix of submodules s
Wherein J s Dimension of the variable on the sub-module s;
(2.4) selecting the overall condition indicating variable: according to the knowledge matrix M constructed in the step 2.2 and the mechanism knowledge of the object, selecting a condition indicating variable X which is related to most variables on the knowledge matrix and can indicate the running state of the whole process of the object on the mechanism c ；
(3) The data preprocessing specifically comprises the following substeps:
(3.1) calculating a sample difference: two-dimensional data matrix X for each sub-module described in step 2.3 s At time t, i.e. the t-th sample
The difference is calculated as follows:
wherein the content of the first and second substances,
is a sample x s,t The first-order difference of (a) is,
is J s Real number of dimension, step of sampling x s,t Is extended to
Of samples at all times
Form a matrix
Wherein, the sample of the first time is directly used as the original sample value.
Wherein the variable X is indicated to the condition s,c All time sample expansion obtains matrix
(3.2) reconstructing the data matrix: for each sub-module sample matrix X obtained in step 2.3 s 、
And its condition indicating variable X s,c 、
And rearranging the samples according to the sequence of the variable values of the condition indication variable from small to large. Determining a condition interval sigma, dividing the condition indicating variable into K condition regionsAnd forming a data matrix by using the samples with the condition indicating variable values belonging to the condition intervals respectively to form K static condition pieces
And K dynamic condition pieces
Wherein N is k,s The number of samples in the kth conditional slice of the submodule s, K =1, \ 8230;, K. Then, the mean value removing processing is carried out on each condition piece to obtain a standardized condition piece data matrix
And
(4) The condition-driven PCA modeling specifically comprises the following sub-steps:
(4.1) establishing a condition slice PCA model: normalized conditional piece data matrix for each submodule
Establishing a PCA model, wherein the formula is as follows:
wherein the content of the first and second substances,
represent
The corresponding main component(s) is (are),
the PCA conversion matrix is represented.
(4.2) determination of T 2 The formula for the statistical quantity control limit is as follows:
wherein Ctrl is k,s Represents the control limit of the sub-module s on the condition slice k, p is the number of the principal elements, F α (p,N k,s -p) is a value corresponding to a confidence of 1 α, a first degree of freedom p and a second degree of freedom N k,s -F distribution of p;
(4.3) merging condition pieces: from the first condition piece, combining the next condition piece with the previous condition piece to obtain the condition segment data matrix
Wherein l represents that the condition section is formed by combining l condition sheets,
then, referring to equations 6 and 7, PCA model of the condition segment is calculated
And control limit Ctrl l,s ；
(4.4) determining conditional segment breakpoints: taking a constant ratio example factor gamma, and comparing the control limit Ctrl of the condition section l,s And the control limit Ctrl of each condition piece in the condition section k,s If Ctrl is present l,s ＞γ·Ctrl k,s If so, it indicates that the control limit of the condition segment has changed greatly and is no longer close to the control limit of the condition piece, so that the condition segment needs to be divided; otherwise, continuing splicing the condition pieces until the condition sections need to be divided.
Assuming that there is l =forconditional segment partitioningl * Then l is obtained * For conditional segment breakpoints, add the first l * Dividing each condition piece into the same condition section, removing the divided condition pieces, and continuously dividing the rest condition pieces;
(4.5) conditional segment division: repeating the steps 4.3-4.4 until all data are divided, and finally obtaining C s A condition section;
(4.6) establishing a conditional segment PCA model: for the well-divided C s Calculating the PCA model of the condition section according to the formulas 6 and 7
(4.7) calculating a Bayesian inferred distance monitoring index: for each conditional segment, a Gaussian Mixture Model (GMM) is established in the feature space, GMM parameters are estimated using the EM algorithm, the number of Gaussian elements is determined using the F-J algorithm, and a probability density function is given by:
wherein the content of the first and second substances,
is a parameter of a probability density function, v c The number of gaussian elements in the feature space of the condition segment c,
is the prior probability of the ith gaussian element of conditional segment c,
for the GMM parameter of the ith Gaussian of conditional segment c, g (-) is a probability density function.
Further, a Bayesian inferred distance index of the condition segment c is calculated:
wherein, N x,s For the number of samples of condition segment c in sub-module s,
the posterior probability for a sample j belonging to the ith gaussian component is calculated from the probability density function,
is T j,c,s Mahalanobis distance to the ith gaussian component; t is j,c,s Is shown in
The jth row of (a); according to the confidence coefficient 1-alpha, the BID can be calculated by a nuclear density estimation method c,s Control limit Ctrl of c,s ；
(4.8) dividing C according to the break points of the static condition segments s The dynamic condition segments are combined, and the step 4.7 is referred to, so that the Bayesian inferred distance index of each submodule s on each dynamic condition segment c can be obtained
And control limits
(5) Determining sub-module weight coefficients: for the global condition indicating variable X obtained in step 2.4 c Dividing the sample into Z intervals at equal intervals, putting the sample with the whole condition indicating variable positioned in a certain interval into a corresponding sample set, and dividing Z groups of samples; combining the submodule condition indicating variable X obtained in the step 2.3 s,c And calculating the correlation coefficient under each interval:
wherein the content of the first and second substances,
is an overall condition indicating variable for the samples in the interval z,
variables are indicated for the sub-module conditions for the samples in interval z.
Further, a weight coefficient of the submodule s when the overall condition indicating variable is in the interval z can be calculated
Weight coefficient
I.e. representing the degree of contribution of the monitoring result on the submodule s to the final result;
(6) The online monitoring specifically comprises the following substeps:
(6.1) obtaining test sample x new (1 XJ) and processed in steps 2.3 and 3.1 to obtain differential samples
And its sample x on each sub-module new,s (ii) a Judging the section z = z according to the overall condition indication variable value of the sample new Judging the condition section of the sub-module in each model according to the condition indication variable value of the sub-module of the sample;
(6.2) assume that it belongs to the c-th sub-module s s And a condition segment, which can calculate the characteristic value:
(6.3) calculating the monitoring statistics of the sample on the submodule:
(6.4) according to the interval z new Calculating the fusion statistics and the control limit:
(6.5) judging the running state: if and only if the static and dynamic monitoring indexes are less than the control limits, the gas turbine is considered to be operated in a normal state; otherwise, the gas turbine is considered to have static deviation or dynamic abnormity, and further inspection is needed.
Firstly, by using the knowledge matrix construction method of the present invention, the knowledge matrix and the condition indicating variable which are established under the value of the correlation coefficient threshold value δ =0.6 are shown in fig. 2 (a) and (b), wherein fig. 2 (a) is the name and the corresponding number of the measured variable, fig. 2 (b) is the knowledge matrix, and fig. 3 is the condition indicating variable table. Taking the compressor part as an example, from the mechanical point of view, the pressure ratio of the compressor can intuitively embody the pressureThe key variable of the operation state of the gas compressor, and the pressure ratio of the gas compressor and the outlet pressure of the gas compressor have obvious positive correlation; from the knowledge matrix, the variable of the compressor outlet pressure and other variables have correlation, so that the compressor outlet pressure (X) can be selected 4 ) As a condition indicating variable for the compressor. Similarly, combining expert knowledge with the constructed knowledge matrix, the combustion engine power (X) is finally selected 20 ) As a whole condition indicating variable, and selecting a combustion chamber pressure differential (X) 15 ) Average temperature (X) of exhaust gas of turbine of combustion engine 13 ) Respectively as condition indicating variables of combustion chamber and turbine of combustion engine.
Next, the monitoring method of the present invention is used to monitor the operation state of an abnormal process, the cross validation method is used to determine the value of the constant ratio scaling factor γ, and the monitoring results under the value of γ =1.05 are shown in fig. 4 (a) and (b), where fig. 4 (a) is the static index monitoring result and fig. 4 (b) is the dynamic index monitoring result. In the present failure case, the actual occurrence time of the failure is 3267 th sample, and as can be seen from fig. 4 (a), the dynamic and static monitoring statistics of the first 3314 normal samples are all within the control limit, and from 3315 th sample, the static monitoring statistics start to continuously exceed the limit, and the time delay is 48 samples. The static monitoring statistic overrun represents that the current running state and the stable running state in the training data have static deviation, and the actual inspection shows that the fault is the abnormal air pressure of the cooling air of the air compressor caused by the dirt of the air compressor and the like. Fig. 4 (c) is a comparison result between the method of the present invention and the conventional monitoring method, and it can be seen from the comparison result that the detection delay of the method of the present invention is significantly lower than that of the conventional monitoring method. Generally, the monitoring strategy based on the knowledge matrix and condition driving mixing provided by the invention can introduce expert knowledge into the field of state monitoring, distinguish multiple operation conditions of a large-range non-stable process through the condition driving monitoring strategy, and finally fuse the monitoring results of all parts according to the condition indication variable values of all subcomponents to form an overall monitoring index, so that the change of the overall operation state of an object can be more accurately judged, the timeliness, the accuracy and the reliability of actual online monitoring are effectively improved, an industrial engineer is facilitated to accurately judge the process operation state of equipment, and the safe and reliable operation of an actual production process is ensured.
The invention provides a knowledge matrix and condition-driven mixed operation state monitoring method for a gas turbine, which is characterized in that on the basis of considering the structural complexity of a large-scale industrial equipment object and the large-scale non-stationary characteristic of the operation state of the large-scale industrial equipment object, a knowledge matrix is constructed through expert knowledge and relevant analysis, sub-modules are divided to reduce the coupling of models, then a condition-driven monitoring method is introduced to divide condition sections and establish a plurality of monitoring models, and therefore the problem of monitoring the large-scale non-stationary process is solved. By applying the method to the actual industrial process, the condition-driven model can be established and fused under the guidance of expert knowledge successfully proved, and the timeliness, the accuracy and the reliability of actual online monitoring are improved.
It should be understood that the present invention is not limited to the above-described specific implementation of the operation of the gas turbine, and that equivalent modifications or substitutions may be made by those skilled in the art without departing from the spirit of the present invention, and are intended to be included within the scope of the appended claims.
Claims (5)
1. A knowledge condition hybrid driving operation state monitoring method for a gas turbine is characterized by specifically comprising the following steps of:
the monitoring model is constructed, and the method comprises the following steps:
(1) Acquiring N samples of a normal continuous operation process of a gas turbine system, wherein each sample comprises J measured variables;
(2) Calculating vector X consisting of N samples of any two measured variables m ,X n Coefficient of correlation between corr (X) m ,X n ) Constructing a knowledge matrix M according to the calculated correlation coefficient and by combining with the mechanism knowledge of the object, wherein the mth row and the nth column of elements M of the knowledge matrix mn Represents the variable X m And X n Whether the two are related; 1 represents association, 0 represents no association;
(3) According to the mechanism of gas turbineKnowledge and the physical structure of an object, and dividing the whole object into S sub-modules; and selecting a condition indicating variable X capable of indicating the process running state to the maximum extent on each submodule s according to the constructed knowledge matrix M and the object mechanism knowledge s,c And other related variables X associated with the variable s,o Jointly combining two-dimensional sample matrices X of submodules s s ＝[X s,c ,X s,o ](ii) a Simultaneously, according to the established knowledge matrix M and the mechanism knowledge of the object, a condition indicating variable X capable of indicating the operation state of the whole process of the object to the maximum extent is selected c ；
(4) Two-dimensional data matrix X for each submodule s Calculating the first difference of each sample from the previous sample
Combining each sample with its corresponding first order difference as a new sample
All new samples obtain a dynamic two-dimensional sample matrix
(5) Sample matrix X for each sub-module s 、
Rearranging and dividing the samples into K according to the sequence of the variable values of the condition indication from small to large to form K static condition pieces
And K dynamic condition pieces
Wherein N is k,s K =1, \ 8230;, K, the number of samples within the kth conditional slice of submodule s; then, the averaging processing is performed for each condition piece,obtaining a data matrix of normalized condition pieces
And
combining the standardized static condition sheets to obtain C s Each static condition section consists of a plurality of continuous standardized static condition pieces, and the control limit calculated by each static condition piece based on the PCA model in each static condition section is smaller than the product of the control limit calculated by each static condition piece based on the PCA model in the static condition section and a constant proportionality factor gamma; establishing a monitoring model and a control limit Ctrl of each static condition section c,s ；
C is likewise marked off according to the breakpoint of the static condition segment s Each dynamic condition section, and establishing a monitoring model and a control limit of each dynamic condition section
(6) Dividing the integral condition indicating variable into Z intervals, and calculating the condition indicating variable X of each submodule in each interval Z s,c And the overall condition indicating variable X c The proportion of the correlation coefficient is used as the weight coefficient of each submodule when the overall condition indicating variable is in the interval z
The on-line monitoring comprises the following steps:
obtaining a sample x to be measured in real time new (1 XJ) and processed to obtain a first order difference
And its sample x on each sub-module new,s (ii) a Judging the section z of the sample according to the overall condition indicating variable value of the sample new And judging the static condition segment and the sub-module condition indicating variable value of the sample according to the sub-module condition indicating variable value of the sampleDynamic condition section, according to the monitoring model, control limit and belonged interval z of the belonged static condition section and dynamic condition section new Calculating the monitoring index and the control limit of the current sample to be tested by the weight coefficient, and if and only if the static and dynamic monitoring indexes are smaller than the control limit, considering that the gas turbine operates in a normal state; otherwise, the gas turbine is considered to be abnormal.
2. The method of claim 1, wherein the measured variables comprise a plurality of pre-module natural gas volumetric flow, pre-module natural gas mass flow, compressor inlet temperature, compressor outlet pressure, compressor outlet temperature, compressor bearing temperature, compressor thrust shoe bearing generator end temperature, compressor thrust shoe bearing combustor end temperature, compressor bearing vibration, compressor side large shaft vibration, compressor inlet duct differential pressure, compressor anti-freeze device inlet electrical control valve position, combustion engine turbine exhaust mean temperature, combustion engine turbine side large shaft vibration, combustor differential pressure, combustion engine cooling air control valve position 1, combustion engine cooling air control valve position 2, combustion engine hum, combustion engine speed, combustion engine power, combustion engine 2-stage retaining ring chamber cooling air pressure, combustion engine 3-stage stationary blade retaining ring chamber cooling air pressure.
3. The method according to claim 1, wherein the dividing of the whole object into S sub-modules is specifically: the whole object is divided into three submodules of a gas compressor, a combustion chamber and a gas turbine.
4. The method of claim 1, wherein the establishing of the monitoring model and the control limit Ctrl for each static condition segment is performed in a manner that is consistent with the present invention c,s The method comprises the following specific steps:
establishing PCA model of each static condition segment
Wherein the content of the first and second substances,
is the data matrix for the c-th static condition segment of the s-th module,
to represent
The corresponding main component(s) is (are),
representing a PCA conversion matrix corresponding to the c static condition section of the s module;
for each static condition segment, establishing a Gaussian mixture GMM model in a feature space, and determining the number v of Gaussian elements by using an F-J algorithm c And then using EM algorithm to estimate GMM parameters
GMM parameter for the ith gaussian element of static condition segment c;
namely a monitoring model of the static condition section c;
constructing a probability density function g based on a PCA model of the static condition section and GMM parameters:
wherein, the first and the second end of the pipe are connected with each other,
is the parameter of the probability density function, is the number of Gaussian elements in the feature space of the static condition section c,
the prior probability of the ith gaussian element of conditional segment c,
GMM parameter of the ith Gaussian element of the conditional segment c;
calculating a Bayes inferred distance index of the static condition segment c:
wherein N is c,s For the number of samples of condition segment c in sub-module s,
is the posterior probability that sample j belongs to the ith gaussian component,
is T j,c,s Mahalanobis distance to the ith gaussian component; t is a unit of j,c,s Represent
The jth row of (a); according to the confidence coefficient 1 alpha, calculating BID by a nuclear density estimation method c,s Control limit Ctrl of c,s 。
5. The method of claim 4, wherein the sub-module condition indicating variable values of the samples are used to determine the static condition segments and the dynamic condition segments thereof, and the control limits and the corresponding intervals z are determined according to the monitoring models, the control limits and the corresponding intervals of the static condition segments and the dynamic condition segments new The weight coefficient of (3) calculates the monitoring index and the control limit of the current sample to be measured, which are as follows:
calculating the monitoring statistics of the sample to be detected on the submodule:
wherein, c s Is the condition segment serial number of the sample to be tested in the submodule s,
denotes c s The monitoring model corresponding to the condition section is used,
the characteristic value of the sample to be detected is represented as follows:
according to the interval z new Calculating static monitoring index BID new Dynamic monitoring index
And corresponding control limit Ctrl new 、
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210917053.4A CN115422714A (en) | 2022-08-01 | 2022-08-01 | Knowledge condition hybrid driving running state monitoring method for gas turbine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210917053.4A CN115422714A (en) | 2022-08-01 | 2022-08-01 | Knowledge condition hybrid driving running state monitoring method for gas turbine |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115422714A true CN115422714A (en) | 2022-12-02 |
Family
ID=84196398
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210917053.4A Pending CN115422714A (en) | 2022-08-01 | 2022-08-01 | Knowledge condition hybrid driving running state monitoring method for gas turbine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115422714A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117575258A (en) * | 2023-11-27 | 2024-02-20 | 中国电力工程顾问集团有限公司 | Coal-fired power plant deep peak shaving method and device considering wastewater treatment |
| CN117575258B (en) * | 2023-11-27 | 2024-05-10 | 中国电力工程顾问集团有限公司 | Coal-fired power plant deep peak shaving method and device considering wastewater treatment |
-
2022
- 2022-08-01 CN CN202210917053.4A patent/CN115422714A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117575258A (en) * | 2023-11-27 | 2024-02-20 | 中国电力工程顾问集团有限公司 | Coal-fired power plant deep peak shaving method and device considering wastewater treatment |
| CN117575258B (en) * | 2023-11-27 | 2024-05-10 | 中国电力工程顾问集团有限公司 | Coal-fired power plant deep peak shaving method and device considering wastewater treatment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Hanachi et al. | A physics-based modeling approach for performance monitoring in gas turbine engines | |
| Kurz et al. | Degradation in gas turbine systems | |
| Barad et al. | Neural network approach for a combined performance and mechanical health monitoring of a gas turbine engine | |
| RU2658869C2 (en) | Estimation of health parameters in industrial gas turbines | |
| US6587737B2 (en) | Method for the monitoring of a plant | |
| Li et al. | Gas turbine gas path diagnosis under transient operating conditions: A steady state performance model based local optimization approach | |
| CN111159844B (en) | Abnormity detection method for exhaust temperature of gas turbine of power station | |
| CN113656906B (en) | Gas turbine-oriented non-stationary multivariable causal relationship analysis method | |
| CN112580267A (en) | Aero-engine surge prediction method based on multi-branch feature fusion network | |
| Li et al. | Study on gas turbine gas-path fault diagnosis method based on quadratic entropy feature extraction | |
| Aretakis et al. | Performance model “zooming” for in-depth component fault diagnosis | |
| Yildirim et al. | Engine health monitoring in an aircraft by using Levenberg-Marquardt feedforward neural network and radial basis function network | |
| Chen et al. | A time-series turbofan engine successive fault diagnosis under both steady-state and dynamic conditions | |
| Jombo et al. | Automated gas turbine sensor fault diagnostics | |
| CN112418306B (en) | Gas turbine compressor fault early warning method based on LSTM-SVM | |
| Salamat | Gas path diagnostics for compressors | |
| Zhang et al. | Physics-based model and neural network model for monitoring starter degradation of APU | |
| CN115422714A (en) | Knowledge condition hybrid driving running state monitoring method for gas turbine | |
| Tsalavoutas et al. | Identifying faults in the variable geometry system of a gas turbine compressor | |
| Wang et al. | Research on anomaly detection and positioning of marine nuclear power steam turbine unit based on isolated forest | |
| Barbosa et al. | Industrial gas turbine diagnostics using fuzzy logic | |
| García-Matos et al. | Hybrid model-based fault detection and diagnosis for the axial flow compressor of a combined-cycle power plant | |
| Borguet et al. | A Study on Sensor Selection for Efficient Jet Engine Health Monitoring | |
| Golberg et al. | THE USE OF SOFTWARE “VIRTUAL ENGINE” FOR TIMELY GAS-TURBINE ENGINE FAULT LOCALIZATIONS | |
| Goyal et al. | Prediction Enhancement of Machine Learning Using Time Series Modeling in Gas Turbines |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |