A method for dynamic prediction of gas turbine gas path system failure

Abstract

This invention discloses a dynamic prediction method for gas turbine gas path system faults, comprising: acquiring multi-source operating parameters and model output parameters during gas turbine operation; constructing hysteresis features and extracting sliding statistical features from the parameters to form a time-series feature vector; based on an integrated regression prediction model, using the time-series feature vector to predict the state parameters of the compressor, combustion chamber, and turbine online. Using the original baseline efficiency as a reference, the predicted efficiency parameters are normalized to construct a health index reflecting the degree of component degradation; within a sliding time window, a piecewise linear trend analysis of the health index is performed, and the current state is used as an anchor point to extrapolate the future degradation trend; for a preset health threshold, the remaining time for the health index to reach the threshold is calculated, and the uncertainty of the threshold arrival time is quantified using a block bootstrapping method, outputting a prediction interval to achieve early warning of gas turbine degradation status.

Description

Technical Field

[0001] This invention relates to the field of gas turbine technology, and in particular to a method for dynamic prediction of faults in gas turbine gas circuit systems. Background Technology

[0002] As a crucial core component in aerospace power and energy generation, gas turbines operate under complex conditions involving high temperature, high pressure, and high speed. The operational status of the gas turbine's gas path system directly impacts the overall safety, reliability, and economy of the entire unit. With increasing operating time, critical components such as the compressor, combustion chamber, and turbine experience performance degradation over long-term service, leading to decreased efficiency, increased energy consumption, and even a significantly higher risk of failure. Therefore, effective monitoring and prediction of degradation failures in the gas turbine's gas path system is a vital technological foundation for achieving condition-based maintenance and life-cycle management.

[0003] Existing methods for gas turbine condition monitoring and fault prediction mainly include threshold-based empirical methods, physical mechanism model-based methods, and data-driven prediction methods. Among these, empirical threshold methods typically rely on fixed limits for alarms, making it difficult to reflect the continuous evolution of component degradation; physical mechanism model methods require the establishment of high-precision thermodynamic models, which are costly to build in practical engineering applications; while data-driven methods can utilize historical monitoring data for condition prediction, most methods only focus on instantaneous conditions or short-term prediction results, lacking robust modeling of degradation trends.

[0004] Furthermore, some existing data-driven methods, when assessing health status, often rely directly on single prediction results, failing to fully consider uncertainties caused by factors such as sensor noise, operating condition fluctuations, and model errors. This can easily lead to unstable early warning results, thereby affecting the reliability of operation and maintenance decisions. Especially in scenarios involving remaining time or health threshold prediction, the lack of a quantitative description of the confidence interval of the prediction results makes it difficult to provide an effective basis for fault early warning of gas turbine gas circuit systems.

[0005] Therefore, there is an urgent need for a dynamic prediction method for gas turbine gas circuit system failures that integrates multi-source operating data and model output information. This method would enable accurate prediction of the efficiency of key components, robust modeling of degradation trends, and quantitative analysis of the uncertainty of the remaining time of the health threshold, thereby allowing for dynamic assessment and early warning of the degradation state of the gas turbine gas circuit system. Summary of the Invention

[0006] To address the limitations of existing dynamic fault prediction methods for gas turbine gas circuit systems, this invention provides a method for dynamic fault prediction of gas turbine gas circuit systems, comprising the following steps: S1. Obtain multi-source operating parameters and gas turbine simulation model output parameters during gas turbine operation, perform data preprocessing on the multi-source operating parameters and model output parameters, construct hysteresis features and sliding statistical features, and form a time-series feature vector for predicting the health status of the gas turbine gas circuit system; S2. Construct an adaptive multi-model integrated prediction framework, and use the time-series feature vector to predict the efficiency of the compressor, combustion chamber and turbine online to obtain the predicted efficiency parameters; S3. Using the original baseline efficiency of the gas turbine as a reference, the predicted efficiency parameters obtained in S2 are normalized to construct a health index to characterize the degree of degradation of key components; S4. Perform piecewise linear trend analysis on the health index described in S3 within the sliding time window. By performing linear fitting on multiple local time periods within the sliding time window, obtain a robust degradation slope that reflects the current degradation trend, and construct a degradation trend model using the current health index status as the anchor point. S5. Based on the degradation trend model described in S4, extrapolate the future change trend of the health index, and calculate the remaining time for the health index to reach each preset health threshold. S6. The uncertainty of the remaining time mentioned in S5 is quantified by the block bootstrap method to obtain the statistical interval of the remaining time. Based on the remaining time and its statistical interval, the degradation early warning information of the key components of the gas turbine is output to realize the dynamic prediction and early warning of the degradation state of the gas turbine gas circuit system.

[0007] Furthermore, the construction of lag features in S1 includes: constructing lag orders for the multi-source operating parameters and model output parameters in chronological order. The lagged terms form an extended feature vector; where any original feature vector... Its lag characteristic set Represented as: In the formula, At the current sampling time, for The values ​​of a certain running parameter and the model output parameter at a given time. The lag order; The lag terms of each feature are concatenated as follows: : In the formula, The vector is concatenated from the lagged terms of each feature. For the first The lagged feature vector is constructed from the original features. This represents the total number of original features.

[0008] Furthermore, the sliding statistical feature extraction in S1 includes: in a length of Within a sliding window, the moving mean and moving standard deviation statistics are calculated for the multi-source operating parameters and model output parameters. Represented as: In the formula, For the sequence index within the sliding window, The length of the sliding window; Sliding standard deviation Represented as: Concatenate into statistical feature vectors : In the formula, For the first The moving average of the original features, For the first The sliding standard deviation of the original features.

[0009] Furthermore, S1 also includes concatenating the original features, hysteresis features, and sliding statistical features to form a time-series feature vector for predicting the health status of the gas turbine gas path system. : In the formula, For the first Feature vectors of the original features ;in In order to be in The original feature vector obtained at time step.

[0010] Furthermore, S2 specifically includes the following steps: S21. Combine the timing feature vector with the efficiency of the compressor, turbine, and combustion chamber to form a multi-output target vector. Constructing the training sample set : In the formula, The sample number. The total number of samples, To output multiple target vectors; In the formula, , , These are the efficiencies of the compressor, combustion chamber, and turbine, respectively. S22. Divide the sample set into a training set and a test set in chronological order, and calculate the mean of the features of the training set dimension by dimension. with standard deviation : In the formula, The number of samples in the training set. This is element-wise multiplication; Perform a normalized mapping between the training and test sets: In the formula, To prevent division by zero of extremely small constants, For standardized mapping functions; S23. Construct a set of base learners and perform multi-output predictions. Simultaneously, establish a model-specific uncertainty evaluation function to characterize the prediction reliability of each base learner. Perform adaptive model subset selection on the base learner set and use regularized weighted fusion to output online efficiency prediction results, obtaining the key component efficiency prediction vector. : In the formula, , , The predicted efficiencies are those of the compressor, combustion chamber, and turbine, respectively.

[0011] Furthermore, the set of base learners includes tree model base learners and kernel regression base learners; For tree model base learners, define the confidence level. Output the reciprocal of the variance for the sub-model: In the formula, For the first Each sub-model output Input the feature vector for the sample, corresponding to the standardized version. ; The number of sub-models, This is the variance calculation function; For kernel regression basis learners, the confidence level is defined. The reciprocal of the posterior prediction variance: In the formula, The posterior variance predicted by kernel regression; For base learners without native uncertainty output, construct a hybrid confidence function. : In the formula, The base learner's number. To predict bias, These are predicted values ​​for the efficiency of key components. The reference model predicts the mean. Local sample density, , As weight and .

[0012] Furthermore, the adaptive model subset selection employs a greedy forward selection strategy: In the formula, For candidate subsets Meta-learner fusion model, Let M be the cross-validation error, and M be the complete set of base learners. The meta-learner fusion model is as follows: In the formula, The fusion weights of the k-th base learner The regularization coefficient is . Let be the prediction result of the k-th base learner for the n-th sample. The total number of samples, It is a fusion weight vector. It is the true target value of the nth sample. For the optimal model subset, .

[0013] Furthermore, S3 specifically involves calculating the ratio of the predicted efficiency to the corresponding original baseline efficiency to obtain the dimensionless health index. : In the formula, For the first Key components at time Prediction efficiency, This represents the corresponding original baseline efficiency.

[0014] Furthermore, the S4 piecewise linear trend analysis is as follows: S41. Within the sliding time window, according to the preset segment length. With segmented step size The health index sequence is divided into multiple overlapping local time periods. Linear fitting is then performed on the health index sequence within each local time period to obtain the corresponding local degradation slope, forming a set of local degradation slopes. : In the formula, For the first The local degradation slope obtained by linearly fitting the health index sequence within a local time period. This represents the total number of data points within a specific time period. S42. Select the slope closest to the current time from the set of local degradation slopes. The local degradation slopes constitute a subset of the final slopes. : Furthermore, the subsets of slopes at the end of the curve are fused using a mean-based statistical method to obtain a robust degradation slope that reflects the current trend of health status changes. : In the formula, For the last slope subset A local degradation slope.

[0015] S43. Based on the current sampling time Health Index As an anchor point, a degradation trend model is constructed based on the robust degradation slope to achieve linear extrapolation of future health index changes. The degradation trend model is as follows: In the formula, Predicting moments for the future.

[0016] Furthermore, the remaining time for the health threshold in step S5 is calculated as follows: In a robust degradation slope In the case where the health index calculated based on the degradation trend model reaches a preset health threshold... The remaining time corresponding to the time : In a robust degradation slope When the health index is determined to be unlikely to reach the health threshold under the current trend, the corresponding operating status judgment result is output.

[0017] Furthermore, the specific steps of S6 are as follows: S61. The health index sequence is processed according to a preset block length. Multiple samplings are performed, and steps S4 and S5 are repeated for each resampled health index sequence to obtain the corresponding health threshold. Set of remaining time estimates : In the formula, For the first The remaining time for the next sampling calculation This represents the number of resampling attempts. S62. Calculate the statistical quantile interval based on the set of remaining time estimates, as the uncertainty interval of the remaining time for the health threshold. : In the formula, , The lower and upper quantile levels are used to construct the uncertainty interval. For quantile operators; S63. Based on the point estimation results of the remaining time and its uncertainty interval, output the degradation early warning information of key components of the gas turbine.

[0018] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention integrates multi-source operating parameters from the gas turbine operation process with the output information of the gas turbine simulation model to construct a multi-dimensional time-series feature vector containing historical information and statistical characteristics. Based on this vector, it enables simultaneous prediction of the efficiency of multiple key components, such as the compressor, combustion chamber, and turbine. This approach fully utilizes the correlation between multi-source data, effectively improving the accuracy, stability, and consistency of predicting the state of key components under complex operating conditions and long-term operation.

[0019] Based on the efficiency prediction results of key components, this invention introduces a baseline efficiency normalization to construct a dimensionless health index, and combines a piecewise linear trend analysis method under a sliding time window to model the degradation process of the health index. By statistically fusing multiple local degradation slopes, a robust degradation slope is obtained, which can more realistically and reliably reflect the current degradation trend of key gas turbine components and provide a stable basis for reasonable extrapolation prediction of health status.

[0020] This invention also addresses the uncertainty in predicting the remaining time of health thresholds by employing a block bootstrapping method to quantify the uncertainty of the remaining time. While maintaining the correlation of the health index time series, it obtains the statistical distribution characteristics and uncertainty interval of the remaining time. By simultaneously outputting the point estimation results of the remaining time and its statistical interval, it provides more reliable and engineering-applicable decision support for the assessment and graded early warning of the degradation status of gas turbine gas path systems.

[0021] This invention constructs a model-specific uncertainty evaluation function and combines it with a greedy forward selection strategy to achieve adaptive selection and regularized weighted fusion of the base learner set. This avoids redundant model stacking and overfitting problems, improving the stability and generalization ability of multi-output predictions. Compared with fixed model sets or simple averaging fusion methods, this invention comprehensively considers prediction accuracy and reliability, achieving more robust and trustworthy online prediction results. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 Overall flowchart of the dynamic prediction method for gas turbine gas circuit system faults. Detailed Implementation

[0024] The following is in conjunction with the instruction manual appendix. Figure 1 This paper provides a more detailed description of a dynamic fault prediction method for a gas turbine gas circuit system provided by the present invention. The method includes the following steps: S1. Obtain multi-source operating parameters and gas turbine simulation model output parameters during gas turbine operation, perform data preprocessing on the multi-source operating parameters and model output parameters, construct hysteresis features and sliding statistical features, and form a time-series feature vector for predicting the health status of the gas turbine gas circuit system; The specific steps are as follows: S11. First, obtain the multi-source operating parameters and the output parameters of the gas turbine simulation model during the operation of the gas turbine. The multi-source operating parameters include the temperature, pressure, speed and flow operating parameters related to the compressor, combustion chamber and turbine. The model output parameters are the temperature, pressure, speed and flow parameters related to the compressor, combustion chamber and turbine output by the simulation model. exist The original feature vector obtained at time step Represented as: In the formula, Represented as the first The operating parameters and model output parameters at time... The value of .

[0025] S12. Perform data preprocessing on the acquired multi-source operating parameters and model output parameters. The data preprocessing includes missing value imputation, outlier removal, and time alignment to ensure that each parameter has a consistent time reference at the same sampling time. S13. Construct lag orders for the multi-source operating parameters and model output parameters in time sequence. The lagged terms form an extended feature vector; where for any original feature vector Its lag characteristic set Represented as: In the formula, At the current sampling time, for The values ​​of a certain running parameter and the model output parameter at a given time. The lag order; The lag terms of each feature are concatenated as follows: : In the formula, The vector is concatenated from the lagged terms of each feature. For the first The lagged feature vector is constructed from the original features. This represents the total number of original features.

[0026] S14. In a length of Within a sliding window, the moving mean and moving standard deviation statistics are calculated for the multi-source operating parameters and model output parameters. Represented as: In the formula, For the sequence index within the sliding window, The length of the sliding window; Sliding standard deviation Represented as: Concatenate into statistical feature vectors : In the formula, For the first The moving average of the original features, For the first The sliding standard deviation of the original features; S15. Concatenate the original features, lagged features, and statistical features to form a time-series feature vector for health status assessment. : In the formula, For the first Feature vectors of the original features ; The time-series feature vector It also includes current status information, historical evolution information, and statistical information, providing sufficient data support for subsequent efficiency prediction and degradation trend analysis of key components.

[0027] S2. Construct an adaptive multi-model integrated prediction framework, and use the time-series feature vector to predict the state parameters of the compressor, combustion chamber and turbine online to obtain the prediction efficiency parameters; The specific steps are as follows: S21. Combine the timing feature vector with the efficiencies of the compressor, turbine, and combustion chamber to form a multi-output target vector. Constructing the training sample set : In the formula, The sample number. For the total number of samples, For multiple output target vectors: In the formula, , , These are the efficiencies of the compressor, combustion chamber, and turbine, respectively. S22. Divide the sample set into a training set and a test set in chronological order, and calculate the mean of the features of the training set dimension by dimension. with standard deviation : In the formula, The number of samples in the training set. For element-wise multiplication, This is a time-series feature vector; Perform a normalized mapping between the training and test sets: In the formula, To prevent division by zero of extremely small constants, For standardized mapping functions; S23. Construct a set of base learners and perform multi-output predictions. Simultaneously, establish a model-specific uncertainty evaluation function to characterize the prediction reliability of each base learner. Perform adaptive model subset selection on the base learner set and use regularized weighted fusion to output online efficiency prediction results, obtaining the key component efficiency prediction vector. : In the formula, , , The predicted efficiencies are those of the compressor, combustion chamber, and turbine, respectively.

[0028] The set of base learners includes tree model base learners and kernel regression base learners; For tree model base learners, define the confidence level. Output the reciprocal of the variance for the sub-model: In the formula, For the first Each sub-model output Input the feature vector for the sample, corresponding to the standardized version. ; The number of sub-models, This is the variance calculation function; For kernel regression basis learners, the confidence level is defined. The reciprocal of the posterior prediction variance: In the formula, The posterior variance predicted by kernel regression; For base learners without native uncertainty output, construct a hybrid confidence function. : In the formula, The base learner's number. To predict bias, These are predicted values ​​for the efficiency of key components. The reference model predicts the mean. Local sample density, , As weight and .

[0029] The adaptive model subset selection employs a greedy forward selection strategy: In the formula, For candidate subsets Meta-learner fusion model, Let M be the cross-validation error, and M be the complete set of base learners. The meta-learner fusion model is as follows: In the formula, The fusion weights of the k-th base learner The regularization coefficient is . Let be the prediction result of the k-th base learner for the n-th sample. The total number of samples, It is a fusion weight vector. It is the true target value of the nth sample. For the optimal model subset, .

[0030] S3. Using the original baseline efficiency of the gas turbine as a reference, the predicted efficiency parameters obtained in S2 are normalized to construct a health index to characterize the degree of degradation of key components; For the first A key component, definition of a dimensionless health index. for: In the formula, For the first Key components at time Prediction efficiency, This corresponds to the original baseline efficiency; S4. Perform piecewise linear trend analysis on the health index described in S3 within the sliding time window. By performing linear fitting on multiple local time periods within the sliding time window, obtain a robust degradation slope that reflects the current degradation trend, and construct a degradation trend model using the current health index status as the anchor point. S4 specifically includes the following steps: S41. Within the sliding time window, according to the preset segment length. With segmented step size The health index sequence is divided into multiple overlapping local time periods. Linear fitting is then performed on the health index sequence within each local time period to obtain the corresponding local degradation slope, forming a set of local degradation slopes. : In the formula, For the first The local degradation slope obtained by linearly fitting the health index sequence within a local time period. This represents the total number of data points within a specific time period.

[0031] S42. Select the slope closest to the current time from the set of local degradation slopes. The local degradation slopes constitute a subset of the final slopes. : Furthermore, the subsets of slopes at the end of the curve are fused using a mean-based statistical method to obtain a robust degradation slope that reflects the current trend of health status changes. : In the formula, For the last slope subset A local degradation slope.

[0032] S43. Based on the current sampling time Health Index As an anchor point, a degradation trend model is constructed based on the robust degradation slope to achieve linear extrapolation of future health index changes. The degradation trend model is as follows: In the formula, Predicting moments for the future.

[0033] S5. Based on the degradation trend model described in S4, extrapolate the future change trend of the health index, and calculate the remaining time for the health index to reach each preset health threshold. In a robust degradation slope In the case where the health index calculated based on the degradation trend model reaches a preset health threshold... The remaining time corresponding to the time : In a robust degradation slope When the health index is determined to be unlikely to reach the health threshold under the current trend, the corresponding operating status judgment result is output.

[0034] S6. The uncertainty of the remaining time mentioned in S5 is quantified by the block bootstrap method to obtain the statistical interval of the remaining time. Based on the remaining time and its statistical interval, the degradation early warning information of the key components of the gas turbine is output to realize the dynamic prediction and early warning of the degradation state of the gas turbine gas circuit system.

[0035] S6 specifically includes the following steps: S61. The health index sequence is processed according to a preset block length. Multiple samplings are performed, and steps S4 and S5 are repeated for each resampled health index sequence to obtain the corresponding health threshold. Set of remaining time estimates : In the formula, For the first The remaining time for the next sampling calculation This represents the number of resampling attempts. S62. Calculate the statistical quantile interval based on the set of remaining time estimates, as the uncertainty interval of the remaining time for the health threshold. : In the formula, , The lower and upper quantile levels are used to construct the uncertainty interval. This is the quantile operator.

[0036] S63. Based on the point estimation results of the remaining time and its uncertainty interval, output the degradation early warning information of key components of the gas turbine.

[0037] The actual operating state of the gas turbine gas circuit system and the output parameters of the simulation model were used as inputs for calculation. The output results of the multi-output regressor are shown in Table 1, and the compressor fault early warning information is shown in Table 2. Analysis results show that the dynamic fault prediction method for gas turbine gas circuit systems adopted in this invention has high state prediction accuracy and fault classification early warning capability. The prediction errors of key thermodynamic parameters are all within 1%, and it can robustly model degradation trends and perform uncertainty analysis on the remaining time of the health threshold, realizing dynamic assessment and early warning of the degradation state of the gas turbine gas circuit system.

[0038] Table 1 Table 2 In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic prediction of faults in a gas turbine gas circuit system, characterized in that, The steps include the following: S1. Obtain multi-source operating parameters and gas turbine simulation model output parameters during gas turbine operation, perform data preprocessing on the multi-source operating parameters and model output parameters, construct hysteresis features and sliding statistical features, and form a time-series feature vector for predicting the health status of the gas turbine gas circuit system; S2. Construct an adaptive multi-model integrated prediction framework, and use the time-series feature vector to predict the efficiency of the compressor, combustion chamber and turbine online to obtain the predicted efficiency parameters; S3. Using the original baseline efficiency of the gas turbine as a reference, the predicted efficiency parameters obtained in S2 are normalized to construct a health index to characterize the degree of degradation of key components; S4. Perform piecewise linear trend analysis on the health index described in S3 within the sliding time window. By performing linear fitting on multiple local time periods within the sliding time window, obtain a robust degradation slope that reflects the current degradation trend, and construct a degradation trend model using the current health index status as the anchor point. S5. Based on the degradation trend model described in S4, extrapolate the future change trend of the health index, and calculate the remaining time for the health index to reach each preset health threshold. S6. The uncertainty of the remaining time mentioned in S5 is quantified by the block bootstrap method to obtain the statistical interval of the remaining time. Based on the remaining time and its statistical interval, the degradation early warning information of the key components of the gas turbine is output to realize the dynamic prediction and early warning of the degradation state of the gas turbine gas circuit system.

2. The method according to claim 1, characterized in that, The lag feature construction in S1 includes: constructing lag orders for the multi-source operating parameters and model output parameters in time sequence. The lagged terms form an extended feature vector; where any original feature vector... Its lag characteristic set Represented as: In the formula, At the current sampling time, for The values ​​of a certain running parameter and the model output parameter at a given time. The lag order; The lag terms of each feature are concatenated as follows: : In the formula, The vector is concatenated from the lagged terms of each feature. For the first The lagged feature vector is constructed from the original features. This represents the total number of original features; S1's sliding statistical feature extraction includes: in a length of... Within a sliding window, the moving mean and moving standard deviation statistics are calculated for the multi-source operating parameters and model output parameters. Represented as: In the formula, For the sequence index within the sliding window, The length of the sliding window; Sliding standard deviation Represented as: Concatenate into statistical feature vectors : In the formula, For the first The moving average of the original features, For the first The sliding standard deviation of the original features.

3. The method according to claim 2, characterized in that, S1 also includes concatenating the original features, hysteresis features, and sliding statistical features to form a time-series feature vector for predicting the health status of the gas turbine gas path system. : In the formula, For the first Feature vectors of the original features ;in In order to be in The original feature vector obtained at time step.

4. The method according to claim 1, characterized in that, S2 specifically includes the following steps: S21. Combine the timing feature vector with the efficiency of the compressor, turbine, and combustion chamber to form a multi-output target vector. Constructing the training sample set : In the formula, The sample number. The total number of samples, To output multiple target vectors; In the formula, , , These are the efficiencies of the compressor, combustion chamber, and turbine, respectively. S22. Divide the sample set into a training set and a test set in chronological order, and calculate the mean of the features of the training set dimension by dimension. with standard deviation : In the formula, The number of samples in the training set. For element-wise multiplication, This is a time-series feature vector; Perform a normalized mapping between the training and test sets: In the formula, To prevent division by zero of extremely small constants, For standardized mapping functions; S23. Construct a set of base learners and perform multi-output predictions. Simultaneously, establish a model-specific uncertainty evaluation function to characterize the prediction reliability of each base learner. Perform adaptive model subset selection on the base learner set and use regularized weighted fusion to output online efficiency prediction results, obtaining the key component efficiency prediction vector. : In the formula, , , The predicted efficiencies are those of the compressor, combustion chamber, and turbine, respectively.

5. The method according to claim 4, characterized in that, The set of base learners includes tree model base learners and kernel regression base learners; For tree model base learners, define the confidence level. Output the reciprocal of the variance for the sub-model: In the formula, For the first Each sub-model output Input the feature vector for the sample, corresponding to the standardized version. ; The number of sub-models, This is the variance calculation function; For kernel regression basis learners, the confidence level is defined. The reciprocal of the posterior prediction variance: In the formula, The posterior variance predicted by kernel regression; For base learners without native uncertainty output, construct a hybrid confidence function. : In the formula, The base learner's number. To predict bias, These are predicted values ​​for the efficiency of key components. The reference model predicts the mean. Local sample density, , As weight and .

6. The method according to claim 5, characterized in that, The adaptive model subset selection adopts a greedy forward selection strategy: In the formula, For candidate subsets Meta-learner fusion model, Let M be the cross-validation error, and M be the complete set of base learners. The meta-learner fusion model is as follows: In the formula, The fusion weights of the k-th base learner The regularization coefficient is . Let be the prediction result of the k-th base learner for the n-th sample. The total number of samples, It is a fusion weight vector. It is the true target value of the nth sample. For the optimal model subset, .

7. The method according to claim 1, characterized in that, S3 specifically involves calculating the ratio of the predicted efficiency to the corresponding original baseline efficiency to obtain the dimensionless health index. : In the formula, For the first Key components at time Prediction efficiency, This represents the corresponding original baseline efficiency.

8. The method according to claim 1, characterized in that, The S4 piecewise linear trend analysis is as follows: S41. Within the sliding time window, according to the preset segment length. With segmented step size The health index sequence is divided into multiple overlapping local time periods. Linear fitting is then performed on the health index sequence within each local time period to obtain the corresponding local degradation slope, forming a set of local degradation slopes. : In the formula, For the first The local degradation slope obtained by linearly fitting the health index sequence within a local time period. This represents the total number of data points within a specific time period. S42. Select the slope closest to the current time from the set of local degradation slopes. The local degradation slopes constitute a subset of the final slopes. : Furthermore, the subsets of slopes at the end of the curve are fused using a mean-based statistical method to obtain a robust degradation slope that reflects the current trend of health status changes. : In the formula, For the last slope subset A local degradation slope. S43. Based on the current sampling time Health Index As an anchor point, a degradation trend model is constructed based on the robust degradation slope to achieve linear extrapolation of future health index changes. The degradation trend model is as follows: In the formula, Predicting moments for the future.

9. The method according to claim 8, characterized in that, The remaining time for the health threshold in step S5 is calculated as follows: In a robust degradation slope In the case where the health index calculated based on the degradation trend model reaches a preset health threshold... The remaining time corresponding to the time : In a robust degradation slope When the health index is determined to be unlikely to reach the health threshold under the current trend, the corresponding operating status judgment result is output.

10. The method according to claim 1, characterized in that, The specific steps for S6 are as follows: S61. The health index sequence is processed according to a preset block length. Multiple samplings are performed, and steps S4 and S5 are repeated for each resampled health index sequence to obtain the corresponding health threshold. Set of remaining time estimates : In the formula, For the first The remaining time for the next sampling calculation This represents the number of resampling attempts. S62. Calculate the statistical quantile interval based on the set of remaining time estimates, as the uncertainty interval of the remaining time for the health threshold. : In the formula, , The lower and upper quantile levels are used to construct the uncertainty interval. For quantile operators; S63. Based on the point estimation results of the remaining time and its uncertainty interval, output the degradation early warning information of key components of the gas turbine.

Images