Method for quantifying state recognition uncertainty of turbine blade of gas turbine

Abstract

The invention relates to the technical field of state recognition uncertainty quantification, and particularly discloses a gas turbine blade state recognition uncertainty quantification method, which comprises the following steps: acquiring operation data in the operation process of gas turbine equipment through data acquisition equipment; carrying out data preprocessing and defining a classification label, a training set and a test set; training the model and predicting a test set to obtain a test set classification result and probability distribution, and calculating logarithm cross entropy according to the probability distribution and a real category to obtain data uncertainty; fusing the multiple models in a probability level to obtain multi-model prediction probability distribution and a prediction result; the standard deviation of the data uncertainty of the multiple models is solved, and the model uncertainty is obtained; and outputting and visualizing the uncertainty of each prediction category. According to the method, uncertainty is quantified and visualized, so that the credibility of each classification result can be visually helped to be observed, and engineering practice can be better helped to make a better and more accurate decision.

Description

Technical Field

[0001] The invention relates to the technical field of state identification uncertainty quantification, and in particular to a method for quantifying the uncertainty of turbine blade state identification of a gas turbine. Background Art

[0002] As a new type of power equipment, gas turbines have the advantages of compact structure, stable operation, safety and reliability, quick start and load driving, and high thermal efficiency, and are increasingly valued by people. The scope of application is also becoming wider and wider. Gas turbines are unique and irreplaceable power equipment in the field of aerospace; they also occupy an increasingly important position in the fields of navigation and land transportation. Due to their high thermal efficiency and low exhaust pollution, they are also becoming the mainstream products of prime movers in the power and energy sectors. Once they fail or have an accident, they will have a serious impact on production and operation. Therefore, it is increasingly important to maintain the gas turbine in normal operation and avoid or timely identify and handle gas turbine operation failures. Since gas turbine status monitoring and fault status identification can greatly improve the safety and reliability of unit operation, help with early warning, status identification and maintenance, the research on gas turbine fault status identification model has important theoretical significance and high application value.

[0003] The origin and development of condition monitoring and fault state identification technology are closely related to the considerable fault losses and equipment maintenance costs, and the significance of condition monitoring and fault state identification is to effectively curb fault losses and equipment maintenance costs. Specifically, it can be summarized as follows: First, it is to be able to discover the early symptoms of faults more timely, avoid, slow down and reduce the occurrence of major accidents; second, once a fault occurs, it can automatically record the complete information of the fault process, so as to analyze the cause of the fault afterwards and avoid the recurrence of similar accidents; third, it can fully understand the performance of the equipment and provide strong evidence for improving the design, manufacturing and maintenance level; fourth, it provides a more effective basis for subsequent fault warning and fault state identification.

[0004] Therefore, it is crucial to use appropriate machine learning classification methods to identify the state of faults in gas turbine blades. At present, there are many methods for state identification of turbine blade faults, including traditional statistical methods, many machine learning methods or neural network methods. However, after these methods give classification categories, it is difficult for us to determine whether these results are reasonable, that is, how much credibility there is, or how much uncertainty there is. This is why uncertainty is introduced: when processing models, accurate uncertainty estimation will bring greater convenience. Predictions may not be accurate every time, but the judgment of accuracy can better assist decision-making. In actual operations, different decisions should often be assigned to each model prediction based on different accuracy. For example, when the uncertainty is high, you can decide not to trust the prediction results at all; or when the prediction is inaccurate, give specific data segments to people for final decision-making; or when the model accuracy is insufficient, the estimated uncertainty can be used as a basis for redesigning and training the model.

[0005] Simply put, the credibility of classification results with large uncertainty is low, and we cannot trust the classification results too much. Conversely, the classification results with small uncertainty can usually be trusted. At present, uncertainty quantification is widely used in computer vision, deep learning, and neural network fields, but in the field of rotating machinery, especially for the complex turbine blade fault state recognition and uncertainty quantification of gas turbines, there is almost no research. At the same time, the existing gas turbine state recognition inventions are all based on deterministic results, and there is a lack of research on uncertainty analysis of their classification results. When using the classification results of the model without considering uncertainty, we cannot have a clear understanding of this result, that is, we cannot know whether this result is accurately predicted by the model or "guessed", which cannot help us make decisions based on the model results in actual work.

[0006] For high-temperature turbine components of gas turbines, their operating conditions are very complex, and the large amount of nonlinear and high-dimensional monitoring data generated is difficult to process and use. In addition, there are few studies on uncertainty quantification at present, and it is difficult to find suitable and practical uncertainty quantification indicators, and it is even more difficult to establish a complete process that can output the final uncertainty including data uncertainty and model uncertainty. In addition, how to reasonably visualize the obtained uncertainty is also a major problem, which requires a lot of manpower and material resources. Summary of the invention

[0007] In order to solve the problems existing in the prior art, the present invention provides a method for quantifying the uncertainty of turbine blade state identification of a gas turbine. At present, the state identification of turbine blades of a gas turbine can only give the result of state identification classification, without considering the uncertainty hidden in the data, namely data uncertainty, and the uncertainty in model classification identification, namely model uncertainty. However, in engineering practice, the prediction may not be accurate every time, and the judgment of accuracy can better assist decision-making. In actual operation, different decisions should often be assigned to each model prediction based on different accuracy. For example, when the uncertainty is high, it can be decided not to trust the prediction result at all; or when the prediction is inaccurate, a specific data segment is handed over to a person for the final decision; or when the model accuracy is insufficient, the estimated uncertainty can be used as a basis for redesigning the model and training the model. Therefore, it is very necessary to quantify the uncertainty of the results of gas turbine turbine blade state identification classification. The present invention quantifies and visualizes the uncertainty, which can not only help intuitively observe the credibility of each classification result, but also better help engineering practice make better and more accurate decisions, solving the problems mentioned in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: a method for quantifying uncertainty in turbine blade state identification of a gas turbine, comprising the following steps:

[0009] S1. Obtaining operation data of the gas turbine equipment during operation through data acquisition equipment;

[0010] S2, perform data preprocessing and define classification labels, training sets and test sets;

[0011] S3, model training and prediction on the test set: Use the divided training set to train the classification model, set the automatic optimization of parameters during the training process, select appropriate parameters for the model in the parameter library, and then use the trained classification model to predict the test set;

[0012] S4. Obtain the classification results and probability distribution of the test set: After the model completes the prediction of the test set, output the predicted probability distribution of each sample point;

[0013] S5. Calculate the logarithmic cross entropy based on the probability distribution and the true category: convert the true label into a one-hot distribution, and then calculate the logarithmic cross entropy between the predicted probability distribution and the one-hot distribution of the true label;

[0014] S6. Obtaining data uncertainty: After calculating the entropy result, the data uncertainty of the prediction result of each sample point of the test data is also obtained;

[0015] S7, multiple models are integrated at the probability level to obtain the multi-model prediction probability distribution and prediction results;

[0016] S8, calculating the standard deviation of the data uncertainty of multiple models to obtain the model uncertainty;

[0017] S9. Output and visualize the uncertainty of each predicted category.

[0018] Preferably, in step S2, data preprocessing includes data cleaning, data downsampling, data denoising, data dimension reduction and data standardization; after the data are delineated into faulty and healthy areas, labels are calibrated respectively, and after the labels are delineated, the data set is divided into a training set and a test set in a ratio of 3:1.

[0019] Preferably, in step S6, the calculation formula of the data uncertainty is as follows:

[0020]

[0021] Where C is the number of categories, c j is the true one-hot distribution for each category, p i_j The predicted probability of each class for each model.

[0022] Preferably, in step S7, it specifically includes: applying the Ensembles idea, adding the six machine learning classification methods at the probability distribution level and then calculating the average to complete the fusion; after the fusion, the predicted probability distribution of multiple models based on the Ensembles idea is obtained, and at the same time, the predicted one-hot distribution and predicted label are obtained according to this distribution.

[0023] Preferably, the six machine learning classification methods include K nearest neighbor, decision tree, naive Bayes, logistic regression, support vector machine and random forest.

[0024] Preferably, in step S8, specifically including: applying the Ensembles idea, performing a standard deviation operation on the model uncertainty obtained by multiple models, thereby obtaining the model uncertainty of the prediction result of each sample point of the test data;

[0025] The calculation formula of the model uncertainty is as follows:

[0026]

[0027] Among them, mean(data s ) is the mean of the data uncertainty of multiple models, and data s_i is the data uncertainty of each model mentioned above, model s This is the final model uncertainty, which is also the source of standard deviation in subsequent plotting. m is the number of machine learning classification methods.

[0028] Preferably, in step S9, it specifically includes: converting the predicted one-hot distribution into a predicted label as the "mean", and finally calculating the entropy (model uncertainty) as the "standard deviation", and then outputting the uncertainty of the prediction result for each sample point according to the 2σ criterion, that is, the 95% confidence interval, and visualizing it.

[0029] Preferably, the multi-model average probability p of the predicted label is obtained m_j The calculation formula is as follows:

[0030]

[0031] Among them, p i_j is the predicted probability of each category for each model, p m_j is the multi-model prediction probability of the average superposition of multiple models at the probability level, and m is the number of machine learning classification methods.

[0032] The beneficial effects of the present invention are as follows: the present invention can be used to solve the uncertainty problem in the classification problem of turbine blade state identification of gas turbines, and provides a simple, efficient and intuitive method to solve the problem that it is difficult to clearly define the confidence level of the classification results of turbine blade state identification, and uses simple and universal quantitative indicators and processes to obtain reasonable confidence results to solve this problem. Finally, for six mature and commonly used machine learning classification and identification methods, the Ensembles idea is used to sum and average the prediction results of multiple models at the probability level to obtain the prediction probabilities and results of multiple models, and finally obtain the final uncertainty results that include data uncertainty and model uncertainty, and the uncertainty (confidence) of each classification result is given with two times the standard deviation, that is, a 95% confidence interval, for visualization. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic flow chart of the steps of a method for quantifying uncertainty in turbine blade state identification of a gas turbine according to an embodiment of the present invention;

[0034] Figure 2 A schematic diagram of a machine learning classification method and principle for state recognition according to an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the state identification and classification task process of an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of the state recognition and classification results of the data 1 test set using six methods according to an embodiment of the present invention;

[0037] Figure 5 This is a schematic diagram of the state identification and uncertainty quantification of data 1 health data (1) in an embodiment of the present invention;

[0038] Figure 6 This is a schematic diagram of the state identification and uncertainty quantification of data 1 fault data (0) in an embodiment of the present invention;

[0039] Figure 7 This is a schematic diagram of the state recognition classification results of the data 2 test set using six methods according to an embodiment of the present invention;

[0040] Figure 8 This is a schematic diagram of the state identification and uncertainty quantification of data 2 health data (1) in an embodiment of the present invention;

[0041] Fig. 9 Schematic diagram of data 2 fault data (0) state identification and uncertainty quantification in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] At present, there are many studies on turbine blade state recognition and classification in the field of gas turbines, but there is little work on uncertainty quantification, which makes it impossible to give confidence in the classification results and difficult to verify the rationality of the results. Some scholars have shown that for classification problems in engineering applications, the output can be converted into a one-hot distribution and the corresponding category probability distribution, and then the entropy index can be used to quantify data uncertainty.

[0044] Therefore, the task scenario of the present invention is how to analyze and quantify the uncertainty of the actual running gas turbine data after the turbine fault state is identified, so as to help have a clearer understanding of the results, and can help to better analyze and use the classification results. To complete this task, the present invention is based on the classification and identification of fault states for turbines, and creates a set of processes for quantifying the uncertainty of state identification classification with the Ensembles concept as the core. The methods, processes and indicators are first tested on the CMPASS data set, and then migrated to the gas turbine data after completion. After verification, the methods and processes used have achieved good results. Not only are the results of the multi-model comprehensive output more reasonable, but the classification uncertainty can also be well estimated and quantified to a certain extent.

[0045] See also Figure 1-Figure 3 The embodiment of the present invention provides a technical solution: a method for quantifying uncertainty in turbine blade state identification of a gas turbine, such as Figure 1As shown, the specific steps are as follows:

[0046] Step 1: Obtain the operating data of the gas turbine equipment during operation through data acquisition equipment. The data is multi-dimensional time series data of the gas turbine operation. The data is usually complex and changeable and has some kind of fault data. Because the present invention focuses on the fault status identification of the turbine blades of the gas turbine, the turbine blade related parameters are extracted after obtaining the operating data.

[0047] Step 2: Perform data preprocessing and define classification labels, training sets and test sets. The overall process is data cleaning, data downsampling, data denoising, data dimension reduction and data standardization, which usually includes operations such as deleting start and stop data, downsampling, vacancy value processing, outlier processing, normalization and wavelet packet denoising. However, depending on the data, some operations are not used, such as whether to delete start and stop data, whether to downsample, whether to reduce dimensions, etc., and use normalization operations with better effects. When defining labels, the gas turbine data has no clear label information, but here it is necessary to roughly divide the fault and healthy areas based on certain criteria for quantitative analysis. The present invention uses the moment of fault warning to divide, and can also use other methods to define the fault and healthy areas, and then calibrate the labels separately. It is worth mentioning that the gas turbine data used in the present invention has only a single type of fault, so the problem here is a two-class problem, but if there are data of multiple fault categories, the method of the present invention is still applicable. After the label is defined, the data set is divided, and the data is divided into a training set and a test set according to a ratio of 75%:25%.

[0048] Step 3: Model training and prediction on the test set. Use the divided training set to train the classification model, set the automatic optimization of parameters during the training process, select the most suitable parameters for the model in the parameter library, and then use the trained classification model to predict the test set. You can get the scores corresponding to the three labels of each sample point, and then convert these scores into probability distributions according to a certain conversion rule. This probability distribution is the basis for obtaining the final prediction result. At the same time, the logarithmic cross entropy can be calculated with the one-hot distribution of the real label. For the sake of convenience, the specific process is as follows Figure 3 It is worth mentioning that entropy is one of the parameters that characterize the state of matter in thermodynamics, represented by the symbol S. Its physical meaning is the measure of the degree of disorder of the system. It has been used in various fields. It has the basic ability to characterize uncertainty. Logarithmic cross entropy is also a commonly used loss function indicator for classification tasks in deep learning. Its principle is relatively simple, its application is relatively wide, and its generalization ability is relatively strong. Therefore, it is selected as the indicator for quantifying uncertainty in the present invention.

[0049] Step 4: Get the classification results and probability distribution of the test set. After the model completes the prediction of the test set, it outputs the predicted probability distribution of each sample point. The predicted probability distribution can be calculated according to Figure 3 The process obtains the predicted one-hot distribution and label.

[0050] Step 5: Calculate the logarithmic cross entropy based on the probability distribution and the true category. Convert the true labels such as 0 and 1 into one-hot distribution, and then calculate the logarithmic cross entropy of the predicted probability distribution and the one-hot distribution of the true label.

[0051] Step 6: Get data uncertainty. After calculating the entropy result, the data uncertainty of the prediction result of each sample point of the test data is obtained. The calculation formula of data uncertainty is as follows:

[0052]

[0053] Where C is the number of categories, c j is the true one-hot distribution for each category, p i_j The predicted probability of each class for each model.

[0054] Step 7: Fusion of multiple models at the probability level. Because data uncertainty and model uncertainty often appear simultaneously during the classification process, separate evaluation is not reasonable. In order to obtain the final uncertainty that includes both uncertainties, the present invention applies the Ensembles idea, adds the six machine learning classification methods at the probability distribution level and then calculates the average to complete the fusion. This is a way to obtain model uncertainty.

[0055] Machine learning classification methods for gas turbine state recognition and classification, such as Figure 2 As shown, they are: K nearest neighbor, decision tree, naive Bayes, logistic regression, support vector machine and random forest, which include classification methods based on tree, distance, classification, hypothesis and regression principles. They are all mature and commonly used machine learning classification methods in various principle categories, which are easy to operate and use. The principle of each method is briefly described as follows:

[0056] K-nearest neighbor (KNN) is a classification method that uses the distance measurement method between different feature values. That is, given a training data set, for a new input instance, find the K instances closest to the instance in the training data set and classify them according to the majority class of these K instances;

[0057] Decision tree (DTC) is a non-parametric supervised learning method, that is, it has no fixed parameters and performs classification or regression learning on data. The goal of decision tree is to learn a set of rules from known data and predict unknown data through simple rule judgment.

[0058] Naive Bayes (NB) is a classification method based on Bayes theorem and the conditional independence assumption of features. It calculates the probability of classification through features and selects the cases with high probability for classification.

[0059] Logistic regression (LR) is a statistical method used to predict the outcome of a dependent variable based on previous observations. It is a type of regression analysis and a common algorithm for solving binary classification problems. The key is to find the minimum point of the loss function;

[0060] Support vector machine (SVM) is a supervised learning method, which is widely used in statistical classification and regression analysis. It is also a very common classifier. The core idea is to separate data by constructing a segmentation surface.

[0061] Random Forest (RF) is a classifier that contains multiple decision trees, and its output category is determined by the majority of the categories output by individual trees. Random Forest can solve the problem of weak generalization ability of decision trees.

[0062] Step 8: Get the multi-model prediction probability distribution and prediction results. After fusion, the prediction probability distribution of the multi-model based on the Ensembles idea can be obtained, and the predicted one-hot distribution and prediction label can be obtained based on this distribution.

[0063] Step 9: Calculate the standard deviation of the model uncertainty obtained by multiple models. Calculate the standard deviation of the data uncertainty obtained by the above six models to obtain the final logarithmic cross entropy, that is, model uncertainty.

[0064] Step 10: Obtain model uncertainty. After calculating the entropy result at this time, the model uncertainty of the prediction result of each sample point of the test data is obtained, which includes the data uncertainty.

[0065] The calculation formula of model uncertainty is as follows:

[0066]

[0067] Among them, mean(data s ) is the mean of the data uncertainty of multiple models, and data s_i is the data uncertainty of each model mentioned above, model s This is the final model uncertainty, which is also the source of standard deviation in subsequent plotting. m is the number of machine learning classification methods.

[0068] Step 11: Output and visualize the uncertainty of each prediction category. Convert the predicted one-hot distribution into a predicted label as the "mean", and use the final calculated entropy (model uncertainty) as the "standard deviation". Then, according to the 2σ criterion, i.e., the 95% confidence interval, output the uncertainty of the prediction result for each sample point and visualize it. The multi-model average probability p of the predicted label is obtained. m_j The calculation formula is as follows:

[0069]

[0070] Among them, p i_j is the predicted probability of each category for each model, p m_j is the multi-model prediction probability of the average superposition of multiple models at the probability level, and m is the number of machine learning classification methods.

[0071] Experimental verification

[0072] Taking a turbine data and a gas engine data as examples, we give specific results. At the same time, before conducting case analysis, based on preliminary research and empirical reasoning, we summarize two expected conclusions to help verify and analyze the results:

[0073] The lower the uncertainty of a correct classification, the better; the higher the uncertainty of an incorrect classification, the better;

[0074] Often, areas of high uncertainty occur at the boundaries where classification situations change.

[0075] Verification Data 1

[0076] The data comes from the CMAPSS data set, which is simulation data. This is because the structure of aircraft engines is complex, and their gas path changes are complex and changeable; and the operating data of aircraft engines are usually confidential data of various airlines and are generally not easy to obtain. Therefore, NASA used the Commercial Modular Aero-Propulsion System Simulation software to generate this set of data sets. Its purpose is to test the performance of different models in combination with the operating characteristics of the engine. There are four sub-data sets, and each subclass has a different number of operating conditions and fault states. The present invention takes FD001 data as an example to illustrate, and uses clustering methods for pre-classification according to the fault category. The data situation is as follows:

[0077] Table 1 Pre-classification labels for fault status recognition of CMAPSS data

[0078]

[0079] Using the divided and standardized data sets and labels, the data is input into the model for training and prediction. The state recognition classification results of the six method test sets are as follows: Figure 4 As shown in Table 2, the accuracy and prediction time (including uncertainty quantification process) of the test set of these six methods are shown:

[0080] Table 2 Accuracy and prediction time of the test set of six methods

[0081]

[0082] From the confusion matrix and accuracy table, we can see that on the test set, SVM, RF and LR performed well, followed by DTC and KNN, and NB was still relatively poor, less than 90%, but overall it can be found that the prediction is still very accurate. At the same time, the prediction time is given as 1m3.6s, including the time for prediction, uncertainty quantification, and giving results and images. It can be seen that the model runs and predicts very quickly, which is convenient for use and method migration.

[0083] Next, we will quantify the uncertainty of CMAPSS data according to the uncertainty quantification technology route mentioned above to complete the case verification. The visualization images of the uncertainty quantification results of healthy data and fault data are shown in the figure below. Figure 5 and Figure 6 As shown. In the figure, the red dots are the true labels, all 1 or 0. When there are a large number of them, they will be blocked and therefore not very obvious in the figure. It can be seen from the figure that: (1) There are fewer misjudgments of normal data and fault data. At the same time, the uncertainty is generally low at the correct judgment, and vice versa, indicating that most judgments, especially correct judgments, have a high degree of credibility, which is consistent with the first expected conclusion; (2) At the same time, each misclassified point can be regarded as the boundary of the classification, so the uncertainty at the boundary is high, which is also consistent with the second expected conclusion; (3) At the same time, the recognition accuracy of the multi-model output is 99.6%, which improves the stability of the results to a certain extent. Therefore, after verification, the uncertainty quantification of FD001 fault state recognition verifies the rationality and correctness of the method and process.

[0084] Verification Data 2

[0085] The data comes from the 12-day operation data of a heavy-duty gas turbine unit from May 10, 2005 to May 22, 2005. The sampling interval is 1 minute, and the unit was shut down on May 22, 2005 due to a compressor blade failure. First, according to the variable meaning interpretation, the turbine blade related variables such as temperature, speed and turbine-related bearing variables are screened and extracted as turbine-related variables. The data has 40 dimensions in total, and 8 dimensions after extracting the turbine data. The data preprocessing used is vacancy value infinite value processing, data cleaning, data denoising and data standardization. Due to the small amount of data, no downsampling operation was performed. The original data shape is 7748*40, and the data shape after extracting variables and data preprocessing is 7727*8. Then, based on the preprocessed data, the fault area is divided, the label is calibrated, and then the data set is divided. The specific situation is shown in Table 3:

[0086] Table 3 Pre-classification labels for gas turbine blade fault status recognition

[0087]

[0088] Using the divided and standardized data sets and labels, the data is input into the model for training and prediction. The recognition results of the test sets of the six methods are as follows: Figure 7 As shown in Table 4, the accuracy and prediction time (including uncertainty quantification process) of the test set of these six methods are shown:

[0089] Table 4 Accuracy and prediction time of the test set of six methods

[0090]

[0091]

[0092] From the confusion matrix and accuracy table, we can see that on the test set, RF and KNN performed better, followed by DTC and SVM, and NB and LR were less than 80%, which was poor. Of course, this was also related to the small amount of data, complex data features, data quality, etc. The prediction time was 33.2s, including the time for prediction, uncertainty quantification, and giving results and images. It can be seen that the model runs and predicts very quickly, which is convenient for use and method migration.

[0093] Next, we will quantify the uncertainty of gas turbine blade data according to the uncertainty quantification technology route mentioned above to complete the case verification. The visualization images of the uncertainty quantification results of healthy data and fault data are as follows: Figure 8 and Fig. 9 shown.

[0094] It can be seen from the figure that: (1) There are relatively few misjudgments of normal data and fault data. At the same time, the uncertainty is generally low at the correct judgment, and vice versa, which means that the credibility of most judgments, especially correct judgments, is high, which is in line with the first expected conclusion; (2) The uncertainty of this data is sometimes high at the correct judgment of normal data, and there are also slightly more misjudgments, which may be related to data quality and data characteristics, but the uncertainty at the wrong judgment must be high. For example, in the identification and uncertainty quantification of fault data, almost all places with high uncertainty appear at the wrong judgment, which is basically consistent with the second expected conclusion, that is, each misclassified point can be regarded as the boundary of the classification, so the uncertainty at the boundary is high; (3) At the same time, the recognition accuracy of the multi-model output is 91.6%, which improves the stability of the results to a certain extent, and integrates the results of the multi-model to give a better result. Therefore, after verification, the uncertainty quantification of the fault state identification of this set of gas turbine turbine blade data meets the basic expectations and can well complete the uncertainty quantification of the classification results.

[0095] The present invention proposes a method and process for quantifying uncertainty in gas turbine blade state identification, namely, first converting the model output into a one-hot distribution and a corresponding category probability distribution, and then introducing the logarithmic cross entropy indicator to perform logarithmic cross entropy on the predicted probability distribution and the one-hot distribution of the true label to characterize data uncertainty.

[0096] In order to obtain model uncertainty, the present invention adopts the Ensembles idea to sum and average the prediction result probabilities of multiple models at the probability level for six mature and commonly used machine learning classification and recognition methods, obtains the prediction probability and results of multiple models, and finally obtains the model uncertainty results containing data uncertainty, and visualizes the uncertainty (confidence) of each classification result with two times the standard deviation, i.e., 95% confidence interval. In the model training process, the model parameter automatic optimization process is added to help the model obtain the optimal parameter configuration, and the methods used are mature and commonly used machine learning methods, which have the advantages of easy operation, easy migration, and fast prediction time. At the same time, replacing the single model results with multi-model results can effectively improve the stability, accuracy and reliability of model prediction classification.

[0097] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for quantifying uncertainty in turbine blade state identification of a gas turbine, characterized in that: The steps include: S1. Obtaining operation data of the gas turbine equipment during operation through data acquisition equipment; S2, perform data preprocessing and define classification labels, training sets and test sets; S3, model training and prediction on the test set: Use the divided training set to train the classification model, set the automatic optimization of parameters during the training process, select appropriate parameters for the model in the parameter library, and then use the trained classification model to predict the test set; S4. Obtain the classification results and probability distribution of the test set: After the model completes the prediction of the test set, output the predicted probability distribution of each sample point; S5. Calculate the logarithmic cross entropy based on the probability distribution and the true category: convert the true label into a one-hot distribution, and then calculate the logarithmic cross entropy between the predicted probability distribution and the one-hot distribution of the true label; S6. Obtaining data uncertainty: After calculating the entropy result, the data uncertainty of the prediction result of each sample point of the test data is also obtained; S7, multiple models are integrated at the probability level to obtain the multi-model prediction probability distribution and prediction results; S8, calculating the standard deviation of the data uncertainty of multiple models to obtain the model uncertainty; S9. Output and visualize the uncertainty of each predicted category.

2. The method for quantifying uncertainty in turbine blade state identification of a gas turbine according to claim 1, characterized in that: In step S2, data preprocessing includes data cleaning, data downsampling, data denoising, data dimension reduction and data standardization; after the data are delineated into faulty and healthy areas, labels are calibrated respectively, and after the labels are delineated, the data set is divided into a training set and a test set in a ratio of 3:

1.

3. The method for quantifying uncertainty in turbine blade state identification of a gas turbine according to claim 1, characterized in that: In step S6, the calculation formula of the data uncertainty is as follows: Where C is the number of categories, c j is the true one-hot distribution for each category, p i_j The predicted probability of each class for each model.

4. The method for quantifying uncertainty in turbine blade state identification of a gas turbine according to claim 1, characterized in that: In step S7, specifically, it includes: applying the Ensembles idea, adding the six machine learning classification methods at the probability distribution level and calculating the average to complete the fusion; after the fusion, the predicted probability distribution of multiple models based on the Ensembles idea is obtained, and at the same time, the predicted one-hot distribution and predicted label are obtained according to this distribution.

5. The method for quantifying uncertainty in turbine blade state identification of a gas turbine according to claim 4, characterized in that: The six machine learning classification methods include K nearest neighbor, decision tree, naive Bayes, logistic regression, support vector machine and random forest.

6. The method for quantifying uncertainty in turbine blade state identification of a gas turbine according to claim 1, characterized in that: In step S8, it specifically includes: applying the Ensembles idea, calculating the standard deviation of the model uncertainty obtained by multiple models, and thus obtaining the model uncertainty of the prediction result of each sample point of the test data; The calculation formula of the model uncertainty is as follows: Among them, mean(data s ) is the mean of the data uncertainty of multiple models, and data s_i is the data uncertainty of each model, model s This is the final model uncertainty, which is also the source of standard deviation in subsequent plotting. m is the number of machine learning classification methods.

7. The method for quantifying uncertainty in turbine blade state identification of a gas turbine according to claim 1, characterized in that: In step S9, it specifically includes: converting the predicted one-hot distribution into a predicted label as the "mean", and finally calculating the entropy, i.e., the model uncertainty, as the "standard deviation", and then outputting the uncertainty of the prediction result for each sample point according to the 2σ criterion, i.e., the 95% confidence interval, and visualizing it.

8. The method for quantifying uncertainty in turbine blade state identification of a gas turbine according to claim 7, characterized in that: Get the multi-model average probability p of the predicted label m_j The calculation formula is as follows: Among them, p i_j is the predicted probability of each category for each model, p m_j is the multi-model prediction probability of the average superposition of multiple models at the probability level, and m is the number of machine learning classification methods.

Images