CN113722977B - Gas turbine rotor fault early warning method based on hybrid prediction - Google Patents
Gas turbine rotor fault early warning method based on hybrid prediction Download PDFInfo
- Publication number
- CN113722977B CN113722977B CN202110576134.8A CN202110576134A CN113722977B CN 113722977 B CN113722977 B CN 113722977B CN 202110576134 A CN202110576134 A CN 202110576134A CN 113722977 B CN113722977 B CN 113722977B
- Authority
- CN
- China
- Prior art keywords
- prediction
- rotor
- gas turbine
- early warning
- threshold
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 42
- 238000012544 monitoring process Methods 0.000 claims abstract description 21
- 238000002485 combustion reaction Methods 0.000 claims abstract description 9
- 239000011159 matrix material Substances 0.000 claims description 27
- 238000004422 calculation algorithm Methods 0.000 claims description 23
- 238000000354 decomposition reaction Methods 0.000 claims description 21
- 230000008569 process Effects 0.000 claims description 9
- 238000006073 displacement reaction Methods 0.000 claims description 7
- 239000000203 mixture Substances 0.000 claims description 4
- 230000009467 reduction Effects 0.000 claims description 3
- 238000010438 heat treatment Methods 0.000 abstract 1
- 239000007789 gas Substances 0.000 description 24
- 230000006870 function Effects 0.000 description 21
- 238000012423 maintenance Methods 0.000 description 10
- 238000003066 decision tree Methods 0.000 description 7
- 230000008859 change Effects 0.000 description 6
- VNWKTOKETHGBQD-UHFFFAOYSA-N methane Chemical compound C VNWKTOKETHGBQD-UHFFFAOYSA-N 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 5
- 238000010276 construction Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 4
- 238000001514 detection method Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 239000013598 vector Substances 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 3
- 230000001965 increasing effect Effects 0.000 description 3
- 239000003345 natural gas Substances 0.000 description 3
- 238000010248 power generation Methods 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 241001391944 Commicarpus scandens Species 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000003449 preventive effect Effects 0.000 description 1
- 238000000513 principal component analysis Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
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
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- 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
- G06F17/11—Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
- G06F17/13—Differential equations
-
- 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
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/10—Numerical modelling
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Optimization (AREA)
- Computational Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Operations Research (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Geometry (AREA)
- Probability & Statistics with Applications (AREA)
- Computer Hardware Design (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention relates to a gas turbine rotor fault early warning method based on mixed prediction, which comprises the following steps: 1) Constructing a mixed prediction model combined with the LMD-XGBoost, and predicting the state parameters of the rotor of the combustion engine in real time to obtain a prediction signal; 2) Constructing a threshold model based on LMD-PCA, and acquiring a monitoring index threshold according to historical normal operation data; 3) Obtaining monitoring index of predicted signal including principal component statistics value T 2 And the square prediction error SPE is compared with a monitoring index threshold serving as an upper limit, so that early warning of the failure of the rotor of the combustion engine is realized. Compared with the prior art, the method has the characteristics of high hybrid prediction precision and combination of vibration and heating power, avoids the limitation of manually setting the monitoring threshold value, and can realize advanced early warning of the rotor fault of the gas turbine.
Description
Technical Field
The invention relates to the technical field of gas turbine equipment fault early warning, in particular to a gas turbine rotor fault early warning method based on mixed prediction.
Background
Gas turbine technology has been widely studied and utilized in various countries around the world as the current technology of improving the utilization rate of energy resources and thoroughly solving the environmental problems. According to the statistics of the national energy bureau, the installed total amount of the natural gas power generation industry in the 2020 China exceeds 9700 kilowatts, and the proportion of gas and electricity in the power generation installation breaks through 4.70%. It is expected that natural gas power generation will increase rapidly, 40-50 g W generator sets will be newly added by 2025, and the consumption of natural gas in the power industry will be doubled to 750-800 g cubic meters. As a large-scale rotary power equipment, the internal structure of a gas turbine is complex, and the gas turbine is easy to break down when running for a long time under the environment of high temperature, high pressure and high rotating speed, wherein the rotor is used as the core of the whole machine design of the gas turbine, and the fault rate of the rotor is up to more than 45 percent. Therefore, the method and the device perform effective early warning before the occurrence of the rotor fault of the gas turbine, timely send out alarm signals to remind operation and maintenance personnel to process, change post-maintenance into pre-prevention, and have important significance for improving the economic benefit of a power plant and prolonging the service life of the gas turbine.
The utilization of massive historical data and the research on a pre-early warning mechanism are lack aiming at the rotor faults of the gas turbine, and the following problems still exist in the existing research results of the fault early warning based on the data and need to be solved:
(1) At present, most of rotor fault researches adopt a vibration analysis method, and the researches on rotor thermodynamic parameters are lacked.
(2) And (5) selecting a preferred method for intelligently predicting the vibration trend.
(3) And the data threshold value is acquired, so that the limitation of manually setting the threshold value is avoided.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide a gas turbine rotor fault early warning method based on mixed prediction.
The aim of the invention can be achieved by the following technical scheme:
a gas turbine rotor fault early warning method based on mixed prediction comprises the following steps:
1) Constructing a mixed prediction model combined with the LMD-XGBoost, and predicting the state parameters of the rotor of the combustion engine in real time to obtain a prediction signal;
2) Constructing a threshold model based on LMD-PCA, and acquiring a monitoring index threshold according to historical normal operation data;
3) Obtaining monitoring index of predicted signal including principal component statistics value T 2 And the square prediction error SPE is compared with a monitoring index threshold serving as an upper limit, so that early warning of the failure of the rotor of the combustion engine is realized.
In the step 1), the state parameters of the rotor of the combustion engine include vibration parameters and thermal parameters, specifically, the displacement of the bearing rotor X, Y at the compressor side to vibration, the displacement of the bearing rotor X, Y at the turbine side to vibration, and the bearing bush temperature.
In the step 1), a hybrid prediction model is constructed by combining a local mean decomposition method and an extreme gradient lifting algorithm, the PF component obtained by the local mean decomposition is used as the input of the extreme gradient lifting algorithm, and the output of the extreme gradient lifting algorithm is used as a prediction signal.
In the prediction process of the hybrid prediction model, a plurality of PF components obtained by adopting local mean decomposition are selected according to frequency and then are overlapped to obtain two subsequences, namely a high-frequency subsequence and a low-frequency subsequence, and an extreme gradient lifting algorithm is respectively adopted for prediction and then is reconstructed to obtain a complete prediction signal.
In the hybrid prediction model, the loss function in the function space is as follows:
wherein ,Gk For the sum of the first-order gradients of each leaf node k, H k Representing the sum of the second order gradients of each leaf node k, T being the number of leaf nodes and λ being the regularization coefficient.
The step 2) comprises the processes of extracting the characteristic components of the LMD and detecting faults based on PCA, and specifically comprises the following steps:
the PF component composition matrix obtained by carrying out local mean decomposition on the historical normal operation data is projected into a PCA model to obtain a square prediction error SPE and a principal component statistical value T serving as monitoring indexes 2 Is set to a threshold value of (2).
In the step 2), the principal component statistics value T 2 The expression of (2) is:
T 2 ＝t T Λ -1 t
where Λ is the covariance matrix of the rotor signal matrix X (m×n), and Λ=diag (λ) 1 ，λ 2 …λ j …)，λ j The j-th eigenvalue of the covariance matrix is marked with a superscript T, T is an m×r principal element matrix, and r is the number of principal elements;
principal component statistics T 2 Threshold of (2)Subject to degrees of freedom m andthe F distribution F (m, n-m) of n-m, and the confidence is a, then there are:
where m is the number of samples and n is the number of PF components.
In the step 2), the square prediction error SPE has the following formula:
wherein ,the threshold SPE representing the confidence level α, i.e. the square prediction error lim I is a unit array, and x is a sample after dimension reduction.
The calculation formula of the control limit is as follows:
wherein ,h0 As intermediate variable, θ 1 θ is the sum of corresponding covariance matrix eigenvalues 2 θ is the sum of squares of eigenvalues of the corresponding covariance matrix 3 C is the sum of the corresponding covariance matrix eigenvalues to the power of three α As a standard normal distribution statistic of confidence alpha, lambda j Is the j-th eigenvalue of the covariance matrix of the rotor signal matrix X (mxn).
In the step 3), when any one of the rotor states of the gas turbine corresponds to the principal component statistic value T of the prediction signal 2 Or square prediction error SPE in prediction time periodAnd when the number proportion of the inner exceeding threshold exceeds a set early warning value, early warning is carried out on the gas turbine rotor.
Compared with the prior art, the invention has the following advantages:
1. aiming at the characteristic that rotor faults are limited to vibration signal analysis, the rotor vibration signal and the thermal signal are combined, so that fault early warning is realized, and the reliability of the rotor fault early warning is improved.
2. The time sequence prediction usually adopts a direct prediction mode, the prediction precision is not high, and the characteristic information of the prediction signal cannot be reflected.
3. PF component is reconstructed into a high-low frequency sequence, the finer characteristic information of the rotor signal is reflected, the variable number of the prediction model is unified, and the construction of the hybrid prediction model is facilitated.
4. In order to avoid the limitation of manually setting the threshold, a threshold model is built by combining a signal decomposition method and a fault detection algorithm, and the threshold of the fault monitoring index can be calculated according to the historical normal operation data of the rotor of the SIS system.
Drawings
FIG. 1 is an LMD decomposition flow diagram.
FIG. 2 shows the result of the LMD decomposition of the X-direction vibration displacement of the bearing No. 1.
Fig. 3 shows the high and low frequency sequences after reconstruction, wherein fig. 3a shows the high frequency subsequence and fig. 3b shows the low frequency subsequence.
Fig. 4 is a graph showing the prediction effect of 3 prediction methods.
FIG. 5 shows a predicted signal failure index T 2 And (3) a change.
Fig. 6 shows the predicted signal failure index SPE change.
Fig. 7 shows the fault monitoring index T2 early warning time.
Fig. 8 shows the fault monitoring indicator SPE warning time.
Fig. 9 is a flow chart of the method of the present invention.
Detailed Description
The invention will now be described in detail with reference to the drawings and specific examples.
As shown in fig. 9, the invention provides a gas turbine rotor fault early warning method based on mixed prediction, which aims to change the operation and maintenance mode of post maintenance of a gas turbine power plant and provides guidance comments for the gas turbine power plant on a rotating equipment fault early warning scheme, and specifically comprises the following steps:
performing data threshold analysis on historical data of a rotor system of the combustion engine, and establishing a threshold model of the health state of equipment during historical operation;
establishing a mixed prediction model based on signal decomposition aiming at rotor vibration data and thermal data;
calculating the ratio of fault monitoring indexes exceeding a threshold value and the range of residual errors exceeding a range, and judging whether the current equipment deviates from normal or not; when the equipment deviates from the normal range, an early warning signal is sent out in time to remind operation and maintenance personnel to process.
The invention carries out state monitoring and fault early warning work on the rotor of the combustion engine, carries out early warning on faults in advance, carries out fault tolerance control, changes fault stopping into planned stopping, reduces stopping or avoids accident expansion, and enables the maintenance management of enterprises to gradually transition from planned maintenance and accident maintenance to preventive maintenance.
The principle and steps of the invention are described as follows:
1. rotor signal mixing prediction model
On the premise of realizing fault early warning in advance, a prediction model of a rotor signal is established to obtain a rotor prediction signal with higher fitting degree, in the prior art, single time sequence prediction is generally carried out on the rotor vibration signal, the prediction precision is not high, and the characteristic information of the rotor signal cannot be embodied.
1.1 construction of LMD-XGBoost hybrid prediction model
The LMD-XGBoost prediction algorithm is a novel combination algorithm, the characteristic components obtained by LMD decomposition are utilized to construct the input of the XGBoost algorithm, and the output of the XGBoost is reconstructed to obtain a prediction signal with higher fitting degree.
The rotor state parameters included in the rotor signal mixture prediction model are: the invention adopts LMD-XGBoost mixing algorithm to predict rotor vibration signals, wherein the vibration displacement comprises a compressor side and a turbine side, and the rotor signals have the characteristics of strong coupling, nonlinearity, large noise and the like.
1.1.1LMD Signal decomposition
The local mean decomposition (Local Mean Decomposition, LMD) method adaptively decomposes a complex non-stationary multi-component signal into a sum of Product Functions (PFs) having physical significance for a plurality of instantaneous frequencies, continuously subtracts the local mean function from an original signal and divides the Product functions by an envelope estimation function (i.e. demodulates the Product functions), and repeats until the envelope estimation function is approximately equal to 1, thereby obtaining pure fm signals, and multiplying the envelope signals by the pure fm signals after obtaining the pure fm signals, so as to obtain PF components, wherein each PF component is directly obtained from an envelope signal and a pure fm signal, the envelope signal is the instantaneous amplitude of the PF component, and the instantaneous frequency of the PF component is directly obtained from the pure fm signal, and further, the instantaneous amplitudes and the instantaneous frequencies of all PF components are combined, so as to obtain the complete time-frequency distribution of the original signal.
The LMD decomposition flow chart is shown in fig. 1, specifically: construction of rotor state parameter signals X (t), i.e. u K As the input of the signal decomposition algorithm, the fixed sliding window is 3 in size, the local maximum and minimum of the rotor signal are calculated by the sliding window mode, the average value and the difference value of all adjacent local extremum points are calculated by the sliding average method, and the local average function m is obtained by smoothing the average value and the difference value after all adjacent average value points are connected by straight lines by the sliding average method in And envelope mean function a in Continuously subtracting the local mean function from the original signal, dividing the local mean function by the envelope estimate mean (i.e. demodulating the local mean function) to construct an envelope estimate function, and repeatingUntil the envelope estimation function is approximately equal to 1, a pure frequency modulation signal S is obtained in And multiplying the envelope signal with the pure frequency modulation signal to obtain the PF component after the pure frequency modulation signal is obtained.
1.1.2XGBoost time sequence prediction model construction
The gradient lifting tree belongs to one of the integrated algorithms, is an iterative decision tree algorithm, and consists of a plurality of decision trees, wherein each decision tree is used for improving the defects of a preamble model, and the result is obtained by accumulating all the decision trees, so that a strong learner is obtained, but the problem of easy overfitting exists, and the extreme gradient lifting algorithm (XGBoost) is further improved on the basis of the gradient lifting algorithm and mainly comprises the following steps:
(1) The complexity of the tree model is added into the regular term, so that overfitting is avoided, and the generalization performance is good;
(2) The loss function is expanded by a Taylor expansion method, a first derivative and a second derivative are used, and the optimization speed can be increased;
(3) When searching the optimal division point, adopting an approximate greedy algorithm to accelerate calculation;
(4) And automatically learning the splitting direction of the sample with missing characteristic values by adopting a sparse perception algorithm.
The XGBoost time sequence prediction method specifically comprises the following steps:
the sliding time window is used for sequentially converting the high-frequency and low-frequency subsequence X (T) of the rotor signal into a characteristic diagram, and the characteristic diagram is input into an XGBoost model for prediction, wherein the input characteristic diagram is as follows:
the XGBoost is characterized in that prediction errors are reduced through a plurality of regression trees, and meanwhile, the tree group formed by the regression trees is guaranteed to have the capacity of being functionalized as large as possible, which can be regarded as optimization of a functional, and a loss function in a function space is as follows:
the right side of the equation consists of an error function and a regularization term, parameters in the error functionIs a rotor prediction signal output by the model, y i The regularization term can be used for punishing model complexity and enhancing generalization capability, and the expression is as follows:
wherein T is the number of leaf nodes, and ω is the node value. Lambda is the regularization coefficient of the XGBoost model. Gamma is the loss reduction threshold due to decision tree splitting. When predicting the rotor signal, the number of predicted values is 300, the depth of a general decision tree is between [2,10], in this example, the depth of the decision tree is 6, so the number T of leaf nodes is 50, the regularization coefficient range is [0,1], the regularization coefficient lambda of the rotor signal prediction model is 0.5, and gamma is 7.
The loss function of XGBoost is then subjected to a second taylor expansion:
wherein the first derivativeSecond derivative->Defining the j-th set of leaf nodes as I j Order G j ＝j，/>Substituting the above and removing constant term to simplify, can obtain:
deriving the aboveSubstituting the result into the original expression to obtain a final objective function as follows:
in the formula ,Gj Representing the sum of the first-order gradients of each leaf node, H j Representing the sum of the second order gradients of each leaf node, the XGBoost model in the present invention has a learning rate of 0.05 and a sampling rate of 0.8.
2. LMD-PCA early warning threshold model construction
The PF component composition matrix obtained by LMD decomposition is projected into a PCA model, and statistics (monitoring index) SPE and T are calculated 2 The process of calculating the fault monitoring index threshold by the PCA model is as follows:
there is a rotor signal matrix X (mxn), where m is the number of samples and n is the number of PF components, which is obtained after principal component analysis:
in the formula ,tj Is m multiplied by 1 vector, represents the j-th principal element, p j N×1 vectors represent the jth load vector; e is the residual matrix. Let the principal element matrix be T (m×r) and the load matrix be P (n×r), then X can be expressed as:
X＝TP T +E (8)
it projects on the residual subspace as:
statistics T 2 Reflecting the degree to which each principal component deviates from the model in terms of trend and magnitude, is a measure of model internalization, which can be used to monitor multiple principal components simultaneously,representing the degree of deviation of fault information from a normal model in the state monitoring process; the statistics SPE characterizes the degree of deviation of the measured value of the input variable from the principal component model, and is a measure of the external change of the model.
Definition T 2 The statistics of (2) are principal component scores, a principal component score value T is calculated, and an estimated value is measuredAnd residual error e, the calculation process is as follows:
principal component statistics T 2 The method comprises the following steps:
T 2 ＝t T Λ -1 t (11)
wherein: Λ=diag (λ) 1 ，λ 2 …), m is the number of principal elements, lambda i Is the ith eigenvalue of the covariance matrix.
T 2 Threshold of (2)Obeying the F distribution F (m, n-m) with degrees of freedom m and n-m, the confidence is a, where a is 0.95.
The square prediction error SPE measures the projection change of the sample vector in the residual space, and the calculation formula is as follows:
wherein ,indicating a control limit for confidence α. />The calculation formula of (2) is as follows:
wherein ： c α a standard normal distribution statistic value of the confidence coefficient alpha; lambda (lambda) j Is the j-th eigenvalue of the covariance matrix of X.
Examples
In this example, simulation analysis is performed on rotor operation data and fault examples of a certain gas power plant above the sea, and the relevant vibration and thermal measurement points of the rotor of the gas unit of the power plant are shown in table 1.
TABLE 1 Power plant rotor station name
Taking turbine side No. 1 bearing rotor X-direction vibration displacement data as an example to verify the model effect, selecting the sample number as 2000 and the data sampling time as 1s, carrying out LMD decomposition on an original rotor signal to obtain 5 PF components with different scales as shown in FIG. 2:
since the PF1, PF2 and PF3 have large fluctuation and high frequency, PF1, PF2 and PF3 in the PF component of the vibration shift are divided into high frequency subsequences, and PF4 and PF5 are divided into low frequency subsequences, and the high and low frequency subsequences are superimposed, respectively, as shown in fig. 3.
2000 sets of data with LMD decomposition were used as training sets for XGBoost model, and the last 300 sets of data were used as test sets. And respectively predicting the high-frequency subsequence and the low-frequency subsequence by using an XGBoost algorithm, reconstructing to obtain a final prediction signal, and comparing the final prediction signal with the direct prediction effect of the XGBoost algorithm and the LSTM model on the undegraded signal to verify the effect of the mixed prediction model, wherein a comparison graph of the effects of the 3 prediction methods is shown in figure 4.
To evaluate the performance of the rotor vibration signal hybrid prediction model, the example selects the mean absolute error (mean absulute error, MAE), root mean square error (root absulute error, RMSE) and fitness (goodness of fit, R 2 ) As model evaluation indexes, the results of the 3 prediction method evaluation indexes are shown in table 2.
Table 2 3 evaluation indexes of prediction methods
| Prediction algorithm | MAE | RMSE | R 2 |
| LMD-XGB | 0.056 | 0.006 | 0.85 |
| XGB | 0.402 | 0.248 | 0.57 |
| LSTM | 0.527 | 0.364 | 0.43 |
Experimental results show that for rotor signal prediction, the LMD-XGBoost hybrid prediction method provided by the invention is obviously superior to XGBoost direct prediction method and LSTM method, and the signal prediction precision is further improved.
The gas turbine vibration abnormal condition occurs in the power plant #3 gas turbine unit 2019/8/9 at 12 am, the vibration has an increasing trend in the subsequent operation process, in order to ensure safe operation of the unit and reduce the fault risk, the gas turbine is set to be in an emergency standby state, the rotor mass unbalance caused by the #3 unit vibration abnormal condition is found out through unit inspection and vibration test, and the unit vibration amplitude is effectively reduced through the rotor dynamic balance weight.
And selecting data of 5-6 months under the normal working condition, and calculating an average threshold under the normal working condition through an LMD-PCA model to serve as a subsequent fault detection index threshold, wherein the subsequent fault detection index threshold is shown in a table 3.
TABLE 3 failure monitoring indicator threshold
In order to verify the effectiveness of the LMD-XGB-PCA early warning model, operation data of 6 days before the occurrence of the fault is selected from a power plant SIS system for modeling, the early warning model is trained by using the data of 5 days (3 days of 8 months to 7 days of 8 months) as a training set, the data of 2 days (8 days of 8 months to 9 days of 8 months) are predicted, and prediction signals SPE and T before the occurrence of the fault are predicted 2 The variation is shown in fig. 5 and 6.
SPE and T under normal conditions 2 The proportion of statistics exceeding the threshold before the fault occurs is obviously increased, and the proportion P of the predicted signal exceeding the threshold per hour is calculated T2 and PSPE As shown in the following figures 7 and 8, according to the operation and maintenance experience of the gas power plant, when the threshold ratio is beyond 0.32, the degradation trend of the equipment can be indicated, if the equipment is setPut P as T2 >0.48，P SPE >Early warning is carried out at 0.48, and the early warning curve chart shows that T 2 The index early warning time is about 11 pm at 2019/8/8 pm, the SPE index early warning time is about 10 pm at 2019/8/8 pm, 13 hours earlier than the fault time, and the SPE index is higher than T 2 The index is more sensitive, and the fault detection is more sensitive.
Claims (5)
1. A gas turbine rotor fault early warning method based on mixed prediction is characterized by comprising the following steps:
1) Constructing a mixed prediction model combined with the LMD-XGBoost, and predicting the state parameters of the rotor of the combustion engine in real time to obtain a prediction signal;
2) Constructing a threshold model based on LMD-PCA, and acquiring a monitoring index threshold according to historical normal operation data;
3) Obtaining monitoring index of predicted signal including principal component statistics value T 2 And comparing the square prediction error SPE with a monitoring index threshold serving as an upper limit to realize early warning of the rotor fault of the combustion engine;
in the step 1), a mixed prediction model is constructed by combining a local mean decomposition method and an extreme gradient lifting algorithm, the PF component obtained by the local mean decomposition is used as the input of the extreme gradient lifting algorithm, and the output of the extreme gradient lifting algorithm is used as a prediction signal;
in the prediction process of the hybrid prediction model, selecting a plurality of PF components obtained by adopting local mean decomposition according to frequency, then superposing the selected PF components to obtain two subsequences, namely a high-frequency subsequence and a low-frequency subsequence, respectively adopting an extreme gradient lifting algorithm to predict and then reconstructing the high-frequency subsequence and the low-frequency subsequence to obtain a complete prediction signal;
in the hybrid prediction model, the loss function in the function space is as follows:
wherein ,Gk A step for each leaf node kSum of degrees, H k Representing the sum of second-order gradients of each leaf node k, wherein T is the number of the leaf nodes, and lambda is a regularization coefficient;
the step 2) comprises the processes of extracting the characteristic components of the LMD and detecting faults based on PCA, and specifically comprises the following steps:
the PF component composition matrix obtained by carrying out local mean decomposition on the historical normal operation data is projected into a PCA model to obtain a square prediction error SPE and a principal component statistical value T serving as monitoring indexes 2 A threshold value of (2);
in the step 2), the principal component statistics value T 2 The expression of (2) is:
T 2 ＝t T Λ -1 t
where Λ is the covariance matrix of the rotor signal matrix X (m×n), and Λ=diag (λ) 1 ，λ 2 …λ j …)，λ j The j-th eigenvalue of the covariance matrix is marked with a superscript T, T is an m×r principal element matrix, and r is the number of principal elements;
principal component statistics T 2 Threshold of (2)Obeying the F distribution F (m, n-m) with degrees of freedom m and n-m, and confidence a, there are:
where m is the number of samples and n is the number of PF components.
2. The gas turbine rotor fault pre-warning method based on hybrid prediction according to claim 1, wherein in the step 1), the state parameters of the gas turbine rotor include vibration parameters and thermal parameters, in particular, displacement of the compressor-side bearing rotor X, Y to vibration, displacement of the turbine-side bearing rotor X, Y to vibration, and bushing temperature.
3. The gas turbine rotor fault pre-warning method based on hybrid prediction according to claim 1, wherein in the step 2), the square prediction error SPE is calculated as:
wherein ,the threshold SPE representing the confidence level α, i.e. the square prediction error lim I is a unit array, and x is a sample after dimension reduction.
4. The gas turbine rotor fault pre-warning method based on hybrid prediction according to claim 3, wherein the control limit is calculated by the formula:
wherein ,h0 As intermediate variable, θ 1 θ is the sum of corresponding covariance matrix eigenvalues 2 θ is the sum of squares of eigenvalues of the corresponding covariance matrix 3 C is the sum of the corresponding covariance matrix eigenvalues to the power of three α As a standard normal distribution statistic of confidence alpha, lambda j Is the j-th eigenvalue of the covariance matrix of the rotor signal matrix X (mxn).
5. The method according to claim 1A gas turbine rotor fault early warning method based on mixed prediction is characterized in that in the step 3), when any gas turbine rotor state corresponds to a principal component statistic value T of a prediction signal 2 Or when the number proportion of the square prediction error SPE exceeding the threshold value in the prediction time period exceeds a set early warning value, early warning is carried out on the gas turbine rotor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110576134.8A CN113722977B (en) | 2021-05-26 | 2021-05-26 | Gas turbine rotor fault early warning method based on hybrid prediction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110576134.8A CN113722977B (en) | 2021-05-26 | 2021-05-26 | Gas turbine rotor fault early warning method based on hybrid prediction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113722977A CN113722977A (en) | 2021-11-30 |
| CN113722977B true CN113722977B (en) | 2023-08-18 |
Family
ID=78672803
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110576134.8A Active CN113722977B (en) | 2021-05-26 | 2021-05-26 | Gas turbine rotor fault early warning method based on hybrid prediction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113722977B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109084186A (en) * | 2018-08-17 | 2018-12-25 | 常州大学 | Pipeline leakage signal recognition methods based on improved ELMD multi-scale entropy |
| CN111737918A (en) * | 2020-06-24 | 2020-10-02 | 大连理工大学 | Integrated learning method for fault diagnosis of high-pressure rotor of aircraft engine |
| CN112365022A (en) * | 2020-09-22 | 2021-02-12 | 南京航空航天大学 | Engine bearing fault prediction method based on multiple stages |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11188691B2 (en) * | 2018-12-21 | 2021-11-30 | Utopus Insights, Inc. | Scalable system and method for forecasting wind turbine failure using SCADA alarm and event logs |
-
2021
- 2021-05-26 CN CN202110576134.8A patent/CN113722977B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109084186A (en) * | 2018-08-17 | 2018-12-25 | 常州大学 | Pipeline leakage signal recognition methods based on improved ELMD multi-scale entropy |
| CN111737918A (en) * | 2020-06-24 | 2020-10-02 | 大连理工大学 | Integrated learning method for fault diagnosis of high-pressure rotor of aircraft engine |
| CN112365022A (en) * | 2020-09-22 | 2021-02-12 | 南京航空航天大学 | Engine bearing fault prediction method based on multiple stages |
Non-Patent Citations (1)
| Title |
|---|
| 基于LMD和AR模型的转子系统故障诊断方法;杨宇等;湖南大学学报(自然科学版);第37卷(第9期);24-28 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113722977A (en) | 2021-11-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110259646B (en) | Wind generating set component state early warning method based on historical data | |
| CN107291991B (en) | Early defect early warning method for wind turbine generator based on dynamic network sign | |
| CN103489035A (en) | Power grid electric energy quality comprehensive evaluation method based on gray weighting correlation analysis algorithm | |
| CN108252873A (en) | A kind of wind power generating set online data monitoring and its system of Performance Evaluation | |
| CN103886518A (en) | Early warning method for voltage sag based on electric energy quality data mining at monitoring point | |
| CN112883634B (en) | DC measurement system state prediction method and system based on multi-dimensional analysis | |
| CN105065212A (en) | Checking method and system of wind generation sets of wind power plant | |
| CN111669123A (en) | Method and device for fault diagnosis of photovoltaic string | |
| CN111878322A (en) | Wind power generator device | |
| CN109613372B (en) | Power grid fault diagnosis method based on multi-element power grid database | |
| CN117728587A (en) | Real-time monitoring system and method for operation data of new energy power generation equipment | |
| CN109635430A (en) | Grid power transmission route transient signal monitoring method and system | |
| CN111563685B (en) | Power generation equipment state early warning method based on auto-associative kernel regression algorithm | |
| CN113722977B (en) | Gas turbine rotor fault early warning method based on hybrid prediction | |
| CN105741184A (en) | Transformer state evaluation method and apparatus | |
| CN115577854A (en) | Quantile regression wind speed interval prediction method based on EEMD-RBF combination | |
| CN111855383B (en) | Fatigue life prediction method for wind turbine blade under icing load | |
| Deng et al. | A fractional hilbert transform order optimization algorithm based DE for bearing health monitoring | |
| Zhao et al. | Application of SOM neural network in fault diagnosis of wind turbine | |
| CN113406537A (en) | Quantitative evaluation method for fault degree of power equipment | |
| Liu et al. | Failure Probability Model of Distribution Network Equipment Based on Improved Age-Reduction Model and Factor Correction | |
| CN112653158B (en) | Frequency response characteristic identification method for doubly-fed wind power plant with additional frequency control | |
| CN117951633B (en) | Photovoltaic power generation equipment fault diagnosis method and system | |
| Liu et al. | Wind Turbine Fault Diagnosis Based on Feature Selection and Stacking Model Fusion with Small-scale Data. | |
| Zhao et al. | Stochastic Degradation Model and Preventive Maintenance Decision of Wind 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |