Gas turbine typical equipment data missing-oriented ordered segment filling and real-time monitoring method

Abstract

The application discloses an ordered segment filling and real-time monitoring method for typical equipment data loss of a gas turbine, innovatively considers natural gas volume flow as an indicator variable to perform slice segment operation on the data containing missing values, and proposes a segment filling method for missing data of a high-rank or even full-rank matrix; a nonlinear matrix filling method is used to evaluate the newly defined data restoration error in each piece of data to realize clustering of similar working conditions; the clustered data can be filled with the overall nonlinear data filling method to realize data filling and monitoring; finally, according to the clustering result, online data is classified in real time, and segment missing value filling and monitoring are realized. The application finely evaluates the working condition characteristics of the gas turbine, solves the problem that it is difficult to accurately restore missing data and perform monitoring by using mean filling and other filling methods due to the complex and changeable working conditions of the combustion chamber, the compressor and the turbine equipment in the gas turbine, and the strong nonlinearity of the data.

Description

Technical Field

[0001] This invention belongs to the field of gas turbine unit process monitoring, and in particular relates to an orderly segmented filling and real-time monitoring method for missing data of typical gas turbine equipment. Background Technology

[0002] With the continuous growth of the global population and economy, global energy demand is booming, and oil and gas resources, as major fossil fuels, are widely used in various industrial and residential sectors. The energy sector is characterized by large investments, long cycles, numerous interconnected factors, and strong inertia. The next two to three decades will be a critical period for the adjustment and transformation of energy production and consumption patterns and energy structure. A new energy landscape will emerge with natural gas, oil, coal, nuclear energy, and renewable energy as the five pillars. Among these, due to my country's abundant natural gas resources, gas-fired power generation remains the primary source of electricity in my country, and it is also one of the main methods of thermal power generation. Furthermore, my country's production status quo, dominated by fossil fuels such as natural gas, will not change for a considerable period. In recent years, my country has continuously adjusted and optimized its energy supply system, and its total installed capacity of power generation equipment and annual power generation have surpassed that of the United States, ranking first in the world.

[0003] However, gas turbine units comprise numerous complex sub-equipment components such as combustion chambers, compressors, and turbines. A malfunction in any one of these sub-equipment can affect the overall operation of the gas turbine. Furthermore, the high start-up and shutdown costs of each sub-equipment result in the gas turbine operating continuously under varying and harsh conditions. This complex and challenging environment places extremely high demands on monitoring the safe and stable operation of the entire system and presents significant obstacles to sensor data collection. Sub-components such as the combustion chamber and compressor, being typical internal components, are difficult to inspect visually and therefore rely on sensors. However, sensors are prone to data loss due to high temperatures and pressures. If the system fails to detect faults due to missing data, the gas turbine unit's operating status will remain abnormal, ultimately leading to serious consequences. Specifically, gas turbine unit faults often manifest in changes in the data measured by its sensors. If data is missing, the samples returned by the sensors become invalid, and the system cannot promptly determine the equipment's operating status. Such problems can range from minor issues like wasting resources and disrupting the stable operation of the power grid to more serious problems like threatening the operational lifespan of gas turbine units and even endangering personnel safety. Conversely, if data loss occurs in the early stages of a fault, and the system appropriately fills in the missing data, it can detect the fault in a timely manner and prevent it from escalating irreversibly. Currently, engineers and researchers primarily use the mean method to fill in missing data. However, considering the complex and time-varying operating conditions of gas turbines, their mean and variance are also time-varying and non-stationary. Simply using the mean to fill in the missing data cannot restore the original characteristics of the equipment. This leads to difficulties in accurately restoring data characteristics, making it impossible for the system to monitor and judge these missing data in a timely manner, ultimately failing to detect faults in a timely manner.

[0004] In summary, gas turbine units operate in harsh and variable environments, yet are crucial equipment in my country's energy sector. Therefore, their operational stability is critical, requiring timely condition monitoring and maintenance. However, the complex data characteristics of gas turbines and the potential for missing data make traditional monitoring models ineffective. No effective data completion or corresponding accurate monitoring methods have yet been developed. Current monitoring algorithms for gas turbine faults assume the absence of missing data; otherwise, the method fails. However, this assumption may not hold true for equipment operating in such harsh environments with extremely high accuracy requirements. On the other hand, directly discarding data ignores the valuable information contained in the intact data, hindering effective, timely, and real-time monitoring of gas turbine faults and preventing early detection. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an ordered, segmented data filling and real-time monitoring method for missing data in typical gas turbine equipment. This invention uses natural gas volumetric flow rate as the primary variable of interest and divides the gas turbine data into different operating conditions based on varying natural gas volumetric flow rates. It details the data characteristics of the gas turbine under different natural gas volumetric flow rates, then accurately fills in the missing data based on these characteristics, and establishes monitoring models for different operating conditions based on the filling results, thereby improving the efficiency of gas turbine data utilization.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] An ordered segmented filling and real-time monitoring method for missing data in typical gas turbine equipment includes:

[0008] Acquire real-time variable data of the gas turbine equipment during online operation;

[0009] Determine if there are missing values ​​in the real-time variable data: If so, allocate the real-time variable data to the corresponding operating condition segment based on the indicator variable and use the NLMC method to fill in the data; if the indicator variable also has missing values, first construct a state-space equation based on the Kalman filter principle to predict and fill in the indicator variable, and then allocate the real-time variable data to the corresponding operating condition segment based on the indicator variable and use the NLMC method to fill in the data; wherein, there are S operating condition segments, which are obtained by the following method:

[0010] Acquire historical variable data of gas turbine equipment without missing values ​​and divide it into L operating condition segments, where the data volume of each operating condition segment is the same and the value of the indicator variable varies between 0 and 8.

[0011] L working condition segments are sequentially merged into S working condition segments, and the range of change of the indicator variable in each working condition segment is recorded. In each working condition segment, the data of the first working condition segment is randomly discarded, and then it is sequentially merged with the subsequent working condition segments. After each merge, it is determined whether the restoration error restored by the NLMC method has changed. If it has changed, the merging is stopped to form a working condition segment.

[0012] After the data filling is completed, the real-time variable data is input into the monitoring model constructed for the corresponding working condition section, and the monitoring statistics of the real-time data are calculated. It is determined whether the monitoring statistics exceed the control limit. If they exceed the limit, it indicates that the equipment has failed. If they do not exceed the limit, it indicates that the equipment is working normally.

[0013] Furthermore, the variable data includes multiple parameters such as the natural gas volume flow rate of the unit's pre-module, the natural gas mass flow rate of the unit's pre-module, the unit's compressor inlet temperature, the unit's compressor outlet pressure, the unit's compressor outlet temperature, the unit's compressor bearing temperature, the unit's compressor thrust bearing generator end temperature, the unit's compressor thrust bearing gas turbine end temperature, the unit's compressor bearing vibration, the unit's compressor side main shaft vibration, the unit's compressor inlet pressure differential, the unit's compressor anti-icing device inlet electric regulating valve position, the unit's gas turbine exhaust average temperature, the unit's gas turbine side main shaft vibration, the unit's combustion chamber pressure differential, the unit's gas turbine cooling air regulating valve position, the unit's gas turbine humming, the unit's gas turbine speed, the unit's gas turbine second-stage stationary vane ring chamber cooling air pressure, and the unit's gas turbine third-stage stationary vane ring chamber cooling air pressure.

[0014] Furthermore, the indicator variable is the natural gas volumetric flow rate of the unit's front-end module.

[0015] Furthermore, the restoration using the NLMC method specifically involves:

[0016]

[0017] stX i,j =Observed ij

[0018] Where X is the variable matrix to be restored, and the size of X is the same as the data matrix to be restored; K(*,*) is the Gaussian kernel function matrix with * as input; p is the p-norm order; δ l (K(X,X T K(X,X) is the kernel function. T The l-th eigenvalue of ), where n represents the dimension of the X matrix, Observed i,j X represents the data in the i-th row and j-th column of the data matrix to be restored, which is not missing. i,j It is the data in the i-th row and j-th column of X.

[0019] Furthermore, the condition for determining whether the reduction error restored using the NLMC method has changed is:

[0020] E m,g+1 >αE m,1

[0021] Among them, E m,g Let E be the restoration error after adding the (g+1)th working condition piece in the m-th working condition segment, α be the difference coefficient, and E be the error after adding the (g+1)th working condition piece. m,1 This represents the restoration error of the first working condition segment in the m-th working condition segment.

[0022] Furthermore, the monitoring model constructed for the corresponding operating condition segment is obtained by using historical variable data of the gas turbine equipment operation corresponding to the operating condition segment division.

[0023] Furthermore, the monitoring model is the KPCA model.

[0024] Furthermore, the monitoring statistic is T. 2 And one or more of SPE.

[0025] The beneficial effects of this invention are as follows: For gas turbines operating under complex and nonlinear conditions, this invention proposes an ordered segmented filling and real-time monitoring method for missing data in typical gas turbine equipment. It treats the natural gas volumetric flow rate of the unit's upstream module as the main discrete variable and, based on data filling capabilities, orderly divides the gas turbine data, completing the operating condition division based on the natural gas volumetric flow rate of the unit's upstream module as the vertical axis. Then, for each different operating condition, a corresponding data filling and monitoring model is established. This invention innovatively proposes using the natural gas volumetric flow rate of the unit's upstream module as the division basis and employing the NLMC method for clustering to provide monitoring results in the presence of missing data. This enables monitoring of equipment operation in the presence of missing data, solving the problem of gas turbines being unable to perform fault monitoring due to missing data; thus ensuring the safe and reliable operation of the gas turbine while meeting its need for timely fault monitoring. Attached Figure Description

[0026] Figure 1 This is a flowchart of the offline modeling process of the present invention;

[0027] Figure 2 This is a flowchart of the online monitoring process of the method of the present invention;

[0028] Figure 3 This is a schematic diagram of the changes in natural gas volumetric flow rate based on the clustering results of data reconstruction capability during the working condition segment of the method of the present invention.

[0029] Figure 4 This is a graph showing the verification results of the method of the present invention on actual normal and abnormal datasets. Detailed Implementation

[0030] This invention innovatively considers slicing data containing missing values ​​using natural gas volumetric flow rate as an indicator variable. Based on this, it proposes a method for segmenting and imputing missing data in high-rank and even full-rank matrices. A nonlinear matrix imputation method is used to evaluate the newly defined data restoration error in each data segment to achieve clustering of similar operating conditions. The clustered data can then be imputed using a holistic nonlinear data imputation method and monitored. Finally, based on the clustering results, online data is classified in real-time, and segmented missing value imputation and monitoring are achieved. This invention provides a refined evaluation of the operating characteristics of gas turbines, solving the problem that mean imputation and other imputation methods are difficult to accurately recover and monitor due to the complex and variable operating conditions of the combustion chamber, compressor, and turbine equipment, and the strong nonlinearity of the data. The invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0031] like Figures 1-2 As shown, this invention is an ordered segmented filling and real-time monitoring method for missing data in typical gas turbine equipment. The implementation of this invention includes the following steps:

[0032] Step 1: Obtain real-time variable data of the gas turbine equipment during online operation;

[0033] Step 2: Determine if there are missing values ​​in the real-time variable data: If so, allocate the real-time variable data to the corresponding operating condition segment based on the indicator variable, and form a matrix with the real-time variable data and the corresponding operating condition segment data. Then, use the Nonlinear Matrix Completion (NLMC) method to fill in the missing values. If the indicator variable also has missing values, first construct a state-space equation based on the Kalman filter principle to predict and fill in the missing values ​​for the indicator variable, and then allocate the real-time variable data to the corresponding operating condition segment based on the indicator variable. Finally, form a matrix with the real-time variable data and the corresponding operating condition segment data and use the NLMC method to fill in the missing values. The operating condition segments include S segments, which are obtained by the following method:

[0034] (2.1) Obtain historical variable data of gas turbine equipment operation without missing values: Obtain I historical data of gas turbine and discard M data with missing values. Each historical data has J measurable variables. The available historical data is described as a two-dimensional matrix X. I ((IM)×J); where IM represents the number of available samples, J represents the number of variables, and both I and J are positive integers, and M is an integer not less than 0.

[0035] (2.2) From the two-dimensional matrix X IExtract the natural gas volume flow rate variable from the unit's front-end module and define it as an indicator variable. Under the premise of ensuring the same data volume for each operating condition segment and a small range for the indicator variable (the indicator variable value varies between 0 and 8), the two-dimensional matrix X... I The system is divided into L working condition segments and then discretized.

[0036] (2.3) Based on the different data complexity and reconstruction capabilities, the entire operation process of the gas turbine is divided into different operating condition segments in an orderly manner, including the following sub-steps:

[0037] (2.3.1) Discard data with a certain probability: Randomly discard the number of data points shown in the following formula for the first working condition segment to obtain the first working condition segment X. I,1 Working condition video after data loss

[0038] num=P×(IM)×J (1)

[0039] Where num is the number of discarded data points, and P is the set probability of missing data.

[0040] (2.3.2) Merging test cases: starting from the test case after the first missing data. Initially, it is merged sequentially with subsequent non-missing work condition segments until the cutoff condition is met, thus forming a work condition segment.

[0041] The sequential merging is as follows: starting from the first working condition segment according to the variable expansion method. Up to the g-th working condition X I,g The working condition images are obtained by merging them in sequence. Each merge adds a new test case to the test case obtained from the previous merge to increase the corresponding number of samples;

[0042] The cutoff condition is: if the merged work condition segments... With the next working condition film X I,g+1 merged into This leads to a restoration error E. m,1 If a change occurs, merging stops, and the first through the gth working condition segments are treated as a single working condition segment, specifically:

[0043] The merged working condition video As a whole, the NLMC method is used to reconstruct the working condition film. Discarded data and calculation of restoration error:

[0044]

[0045]

[0046]

[0047] Where X is the matrix of variables to be restored, and the size of X is... Consistent, K(*,*) is the Gaussian kernel function matrix with * as input, p is the order of the p-norm norm, δ l (K(X,X T K(X,X) is the kernel function. T The l-th eigenvalue of ), where n represents the dimension of the X matrix, Observed i,j for The data in the i-th row and j-th column that is not missing, X i,j It is the data in the i-th row and j-th column of X. express The restored matrix, X I,1～g+1 It is the matrix obtained by merging the original 1st to gth working condition segments, E m,1 The error is the restoration error of the first industrial control chip in the m-th operating condition segment.<X,X> E is the inner product of matrix X. m,g+1 This is the restoration error added to the g+1th working condition segment of the mth working condition segment.

[0048] Determine whether the newly added test case meets the following conditions:

[0049] E m,g+1 >αE m,1 (5)

[0050] Where α is the difference coefficient, which can be any value between 1 and infinity based on the monitoring results. If the newly added working condition segment satisfies the above formula, it means that the newly added working condition segment X... I,g+1 If the data restoration performance of this working condition segment changes, then 1 to g working condition segments will be output as one working condition segment;

[0051] (2.3.3) Cyclic partitioning: Starting from the (g+1)th working condition segment X I,g+1 A new operating condition segment is being created, and this segment is designated as the one where data is randomly lost. Repeat steps (2.3.1)-(2.3.2) to traverse all working condition segments, obtaining S working condition segments; record the range of change of the indicator variable in each working condition segment, where the range of change of the indicator variable in the m-th working condition segment is R. m The real-time variable data can be divided and filled in using the S operating condition segments obtained from the division.

[0052] Step 3: Input the real-time variable data after data filling into the monitoring model constructed for the corresponding operating condition segment, and calculate the monitoring statistics of the real-time data; determine whether the monitoring statistics exceed the control limits. If they exceed the limits, it indicates that the equipment has malfunctioned; if they do not exceed the limits, it indicates that the equipment is operating normally. The monitoring model constructed for the corresponding operating condition segment is obtained by using historical variable data of the gas turbine equipment operating under the segment, as detailed below:

[0053] (3.1) Data imputation: Determine the working condition segment to which the M data with missing values ​​in step (2.1) belong, and use the optimization process shown in equation (2) to imput the data with missing values ​​in conjunction with the corresponding working condition segment, and add the imputed data to the data of the corresponding working condition segment.

[0054] (3.2) Monitoring Model Construction: Using S working conditions as training data, S KPCA models are established respectively. Taking the first working condition as an example, the model construction is as follows:

[0055]

[0056] K(X1,X1 T )=U1Λ1V1 T (7)

[0057]

[0058] Where X1 represents the data for the first operating condition segment. For the normalized data, E(X1) and D(X1) are the mean and variance of X1, respectively. U1, Λ1, and V1 are the variances of K(X1, X1) / K. T The three matrices obtained after singular value decomposition are T1 and T1. T1 represents the mapping feature of the first operating condition segment extracted. Based on the extracted feature T1, two monitoring statistics can be constructed as follows:

[0059] T 2 i =t i T Λ1 -1 t i (9)

[0060] SPE i =N1k(x i ,x i )-t i T t i (10)

[0061] Where N1 represents the number of samples in the first working condition segment, t i Let k(x) be the i-th eigenvector of T1. i ,xi The input is the i-th data sample x within this operating condition segment. i The Gaussian kernel function was then used. Subsequently, the kernel density estimation method was employed to determine T within this operating condition range. 2 i and SPE i The 95% confidence interval value is used to obtain the corresponding confidence interval control limit for the operating condition segment, such as the confidence interval control limit for the m-th segment, and denoted as T. 2 m,limit and SPE m,limit .

[0062] Online status monitoring can be performed based on the constructed detection model:

[0063] The real-time data obtained after step 2 is put into the monitoring model constructed for the corresponding working condition segment, and the monitoring statistics of the real-time data are calculated. It is determined whether the monitoring statistics exceed the control limit. If they exceed the limit, it indicates that the equipment has failed. If they do not exceed the limit, it indicates that the equipment is working normally. The monitoring statistics of the sample are calculated using the KPCA model: the corresponding feature is calculated using the m0th KPCA model according to equation (8), and then the two monitoring statistics are calculated using the feature according to equations (9) and (10) and denoted as: T 2 real-time SPE real-time And based on the calculated monitoring statistics and the confidence interval T obtained from the offline assessment... 2 m,limit SPE m,limit Compare the segments. Taking the m-th segment as an example:

[0064] SPE real-time >SPE m,limit (11)

[0065] T 2 real-time >T 2 m,limit (12)

[0066] If any one of the inequalities is true, the device is considered to be malfunctioning. Monitoring is achieved through the above logic.

[0067] This embodiment uses a real gas turbine dataset with 60,000 available samples and 32 variables as an example, and includes the following steps:

[0068] (1) Establish the monitoring model offline. This step is mainly achieved by the following sub-steps:

[0069] (1.1) Acquiring Historical Data: Acquire 5000 available historical data points for the gas turbine, each with 32 measurable variables. Measurable variables include: natural gas volumetric flow rate from the unit's pre-amplifier module, natural gas mass flow rate from the unit's pre-amplifier module, unit compressor inlet temperature (representing ambient temperature), unit compressor outlet pressure 1, unit compressor outlet pressure 2, unit compressor outlet temperature, unit compressor bearing temperature 1, unit compressor bearing temperature 2, unit compressor bearing temperature 3, unit compressor thrust bearing generator end temperature 1, unit compressor thrust bearing generator end temperature 2, unit compressor thrust bearing generator end temperature 3, unit compressor thrust bearing gas turbine end temperature 1, unit compressor thrust bearing gas turbine end temperature 2, unit compressor thrust bearing gas turbine end temperature 3, and unit compressor... The following data points are considered: 1. Compressor bearing vibration; 2. Compressor bearing vibration; 3. Compressor side shaft vibration; 4. Compressor intake pressure differential; 5. Compressor anti-icing device intake electric regulating valve position; 6. Average exhaust temperature of the gas turbine; 7. Gas turbine side shaft vibration; 8. Combustion chamber pressure differential; 9. Gas turbine cooling air regulating valve position; 10. Gas turbine cooling air regulating valve position; 21. Gas turbine humming; 22. Gas turbine speed; 13. Gas turbine stage 2 stationary vane ring chamber cooling air pressure; 24. Gas turbine stage 3 stationary vane ring chamber cooling air pressure; 25. The 5000 historical data points are described as a two-dimensional matrix X. I (5000×32).

[0070] (1.2) Transform the two-dimensional matrix X I Data with missing values ​​in (5000×32) are filtered out and discretized, specifically: the two-dimensional matrix X is... I The natural gas volumetric flow rate variable in (5000×32) is extracted separately, and the data is divided into operating condition segments based on 50 samples. To ensure that the number of samples in each operating condition segment is the same, the original data is divided into 100 operating condition segments, thus discretizing the data; each initial operating condition segment contains I... g There are (g = 1, 2, ..., 50) samples and 32 measurable variables, where the subscript g indicates that the sample belongs to the g-th working condition segment, and has The discretized g-th working condition segment is denoted as X. I,g .

[0071] (1.3) The merging is achieved sequentially through the following sub-steps:

[0072] (1.3.1) Discard data with a certain probability: For the first working condition segment X I,1 Randomly discard the number of data points shown in the following formula to obtain the first working condition segment X. I,1 Working condition video after data loss

[0073] num = P × (IM) × J

[0074] Where P is set to 0.4, J to 32, and IM to 50, then num is 640.

[0075] (1.3.2) Reconstructing the working condition film using the NLMC method The discarded data and the calculated restoration error are as follows:

[0076]

[0077] stX i,j =Observed ij

[0078]

[0079] Where X is the matrix of variables to be restored, K(*,*) is the Gaussian kernel function matrix with * as input, the hyperparameters are set to 1 and 10, p is the p-norm order set to 0.9, and δ l (K(X,X T K(X,X) is the kernel function. T The l-th eigenvalue of ) is observed. i,j for The data in the i-th row and j-th column that is not missing, E m,1 To account for restoration error,<X,X> The inner product of matrix X; here, the reduction error E is calculated. m,1 It is 0.286.

[0080] (1.3.3) Merging test cases: Starting from the first test case Initially, it is merged sequentially with the subsequent original working condition segments until the cutoff condition is met, thus forming a working condition segment.

[0081] The sequential merging is as follows: starting from the first working condition segment according to the variable expansion method. Up to the g-th working condition X I,g The working condition images are obtained by merging them in sequence. Each merge involves adding a new test case to the test case obtained from the previous merge to increase the number of samples accordingly.

[0082] The cutoff condition is: if the merged work condition segments... With the next working condition film X I,g+1 merged into This leads to a restoration error E. m,1 If a change occurs, the merging process stops, and the first through the gth working condition segments are treated as a single working condition segment, specifically:

[0083] The merged working condition video As a whole, perform the optimization steps shown in equation (2) of step (1.3.2); and calculate according to equation (3). The restoration error is calculated each time a new working condition chip is added, and it is determined whether any newly added working condition chip meets the following conditions:

[0084] E m,g+1 >αE m,1

[0085] Where α is set to 1. If the newly added working condition segment satisfies the above formula, it means that the newly added working condition segment X I,g+1 If the data restoration performance of this working condition segment changes, then 1 to g working condition segments are output as one working condition segment;

[0086] (1.3.4) Cyclic partitioning: A new working condition segment is formed starting from the (g+1)th working condition segment, and this working condition segment is designated as the working condition segment where data is randomly lost. Repeat step (1.3.3) to traverse all working condition segments, obtaining S working condition segments; record the range of change of the indicator variable in each working condition segment, where the range of change of the indicator variable in the m-th working condition segment is R. m This example contains four operating conditions, and the range of their indicator variables is shown in Table 1 and... Figure 3 As shown.

[0087] Table 1: Range of Natural Gas Volume Flow Rate Variables in Operating Conditions

[0088]

[0089] (1.3.5) Data imputation: Determine the working condition segment to which the 627 data with missing values ​​in step (1.1) belong, and use the optimization process shown in equation (2) to imput the data with missing values ​​in conjunction with the corresponding working condition segment, and add the imputed data to the data of the corresponding working condition segment.

[0090] (1.3.6) Monitoring Model Construction: Using five working conditions as training data, five KPCA models were built respectively. Taking the first working condition as an example, the model construction is as follows:

[0091]

[0092] K(X1,X1 T )=U1Λ1V1 T

[0093]

[0094] Where X1 represents the data for the first operating condition segment. For the normalized data, E(X1) and D(X1) are the mean and variance of X1, respectively. U1, Λ1, and V1 are the variances of K(X1, X1) / K. T The three matrices obtained after singular value decomposition are given. T1 represents the final extracted mapping features. Based on the extracted features T1, two monitoring statistics can be constructed as follows:

[0095] T 2 i =t i T Λ1 -1 t i

[0096] SPE i =N1k(x i ,x i )-t i T t i

[0097] Among them, t i Let k(x) be the i-th eigenvector of T1. i ,x i ) is the input x i The Gaussian kernel function was then used. Subsequently, the kernel density estimation method was employed to determine T for each segment. 2 i and SPE i The 95% confidence interval value is used to obtain the confidence interval control limits for the corresponding operating condition segment. The confidence interval control limits for the m-th segment are denoted as T. 2 m,limit and SPE m,limit .

[0098] (2) Online status monitoring, including the following steps:

[0099] (2.1) Obtaining data and determining the corresponding operating condition: Obtaining real-time data of the gas turbine's online operation is as follows: Extracting natural gas volumetric flow rate variables According to the range of indicator variable R recorded in Table 1 for each operating condition, m Based on the analysis, the natural gas volumetric flow rate in this sample is 54.3386, therefore, this sample belongs to the 4th operating condition segment. From this, the corresponding operating condition segment and model can be obtained.

[0100] (2.2) Determine if there are any missing values ​​in this real-time data. If there are no missing values, no processing is required. If there are missing values, further determine if the indicator variable is missing. If the indicator variable is missing, construct a state-space equation based on the Kalman filter principle to predict and fill in the missing values. If the indicator variable is not missing, allocate the real-time data to the corresponding operating condition segment based on the indicator variable to fill in the missing values.

[0101] (2.3) The real-time data obtained after processing in step (2.2) is put into the monitoring model constructed for the corresponding working condition segment, and the monitoring statistics of the real-time data are calculated; it is determined whether the monitoring statistics exceed the control limit. If they exceed the limit, it indicates that the equipment has failed. If they do not exceed the limit, it indicates that the equipment is working normally. The monitoring statistics of the sample are calculated using the KPCA model: the corresponding feature is calculated using the m0th KPCA model according to equation (8), and then the two monitoring statistics are calculated using the feature according to equations (9) and (10) and denoted as: T 2 real-time SPE real-time And based on the calculated monitoring statistics and the confidence interval T obtained from the offline assessment... 2 i,limit SPE i,limit A comparison is made. This corresponds to the fourth working condition segment, namely:

[0102] SPE real-time >SPE 4,limit

[0103] T 2 real-time >T 2 4,limit

[0104] If any one of the inequalities is true, the equipment is considered to be malfunctioning. The monitoring statistic SPE calculated for this sample is... real-time and T 2 real-time The values ​​were 0.9962 and 0.0013 respectively, while the corresponding monitoring statistics for the fourth operating condition were 0.9987 and 0.0329 respectively. Since none of the corresponding statistics exceeded the control limits, it can be concluded that this sample is under normal conditions.

[0105] Here is a simple demonstration using a fault sample, such as... Figure 4 As shown, it can be clearly seen that at the point corresponding to the fault, i.e. before sample number 150, this method had already detected the anomaly and issued an alarm in a timely manner.

[0106] This invention is not limited to the gas turbine equipment described in the above examples. Anyone skilled in the art can make equivalent modifications or substitutions without departing from this invention, and all such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for ordered piecewise imputation and real-time monitoring of typical equipment data missing of a gas turbine, characterized in that, The method comprises the following steps: acquiring real-time variable data of the gas turbine equipment in online operation; judging whether the real-time variable data has missing values: if yes, distributing the real-time variable data to corresponding working condition segments according to the indicator variable and filling in the data by using the NLMC method; if the indicator variable also has missing values, first constructing a state space equation according to the Kalman filter principle to predict and fill in the indicator variable, and then distributing the real-time variable data to corresponding working condition segments according to the indicator variable and filling in the data by using the NLMC method; wherein the working condition segments include S, which are obtained by the following method: (2.1) Obtain historical variable data of the operation of the gas turbine equipment without missing values: obtain historical data of the gas turbine equipment, discard data with missing values, each historical data has M measurable variables, and the available historical data is described as a two-dimensional matrix; wherein, N represents the number of available samples, J represents the number of variables, and I and J are positive integers, and M is an integer not less than 0. ​​​​​​ (2.2) Extract a variable from the two-dimensional matrix and define it as an indication variable, divide the two-dimensional matrix into condition pieces under the premise of ensuring the same amount of data for each condition piece and the value range of the indication variable between 0-8, and perform discretization processing; (2.3) According to the different data complexity and restoration ability, the whole operation process of the gas turbine is sequentially divided into different working condition segments, including the following sub-steps: (2.3.1) Discarding data with a certain probability: randomly discard the amount of data shown in the following formula from the first condition piece to obtain the 1st condition piece Condition piece after data loss : wherein is the number of discarded data, is the set data loss probability; (2.3.2) Merge the working condition pieces: from the first data missing working condition piece Start, and then sequentially order it with the subsequent non-missing working condition pieces, stop merging until the stop condition is met, and form a working condition section; The sequential and orderly merging is: according to the variable expansion mode, merging the first working condition piece to the gth working condition piece to obtain the working condition piece , each time, a new working condition piece is added to the working condition piece obtained by the last merging to increase the corresponding sample quantity. The cutoff condition is: if the merged work condition segments With the next working condition film merged into This leads to restoration errors. If changes occur, merging stops, and the first to g working condition segments are treated as a single working condition segment, specifically: The merged condition patch As a whole, the NLMC method restores the condition patch The discarded data and calculate the restoration error: （2） （3） （4） wherein, is a to-be-recovered variable matrix, is of the same size as is a Gaussian kernel function matrix with is a norm order, is an lth eigenvalue of a kernel function , and n represents a dimension of the X matrix, is data of an ith row and a jth column in is data of an ith row and a jth column in is a restored matrix, is a matrix obtained by merging original 1st to gth condition pieces, is a restoration error of a first condition piece of an mth condition section, is an inner product of a matrix , and is a restoration error after adding a g+1th condition piece to the mth condition section.​​​​​​ judging whether the newly added working condition segment meets the following conditions: （5） wherein, is a difference coefficient, which can be any value between 1 and infinity according to the monitoring result; if the newly added working condition piece satisfies the above formula, it indicates that the newly added working condition piece so that the data restoration performance of the working condition section is changed, and the output is 1~g working condition pieces for a working condition section; (2.3.3) Cycle division: from the g+1th condition piece Start to form a new condition section, set the condition piece as the data random loss condition piece , and repeat steps (2.3.1)-(2.3.2) to traverse all condition pieces, and obtain condition sections; record the change range of the indicating variable in each condition section, and the change range of the indicating variable in the mth condition section is ; the S condition sections obtained by division can be used for corresponding division and filling of real-time variable data; inputting the real-time variable data after the data filling into the monitoring model constructed for the corresponding working condition segment, and calculating the monitoring statistics of the real-time data; judging whether the monitoring statistics exceeds the control limit, if yes, indicating that the equipment has a fault, if not, indicating that the equipment is working normally.

2. The method of claim 1, wherein, The variable data includes multiple types of the following: natural gas volume flow of a unit pre-module, natural gas mass flow of the unit pre-module, unit compressor inlet temperature, unit compressor outlet pressure, unit compressor outlet temperature, unit compressor bearing temperature, unit compressor thrust pad bearing generator end temperature, unit compressor thrust pad bearing turbine end temperature, unit compressor bearing vibration, unit compressor side large shaft vibration, unit compressor inlet duct pressure difference, unit compressor anti-icing device inlet electric regulating valve position, unit turbine exhaust average temperature, unit turbine side large shaft vibration, unit combustion chamber pressure difference, unit turbine cooling air regulating valve position, unit turbine humming, unit turbine speed, unit turbine 2nd stage static blade holding ring cavity cooling air pressure, and unit turbine 3rd stage static blade holding ring cavity cooling air pressure.

3. The method of claim 1, wherein, The indicator variable is the natural gas volume flow of the unit pre-module.

4. The method of claim 1, wherein, The monitoring model constructed for the corresponding working condition segment is constructed by using historical variable data of the gas turbine equipment corresponding to the working condition segment division.

5. The method of claim 1, wherein, The monitoring model is a KPCA model.

6. The method of claim 1, wherein, The monitoring statistics are one or more of and one or more of

Images