Abstract: The failure rate of high-temperature components of gas turbines is high, hidden, and destructive. After the failure occurs, the maintenance cost is high and the maintenance difficulty is great. Studying a method for predicting the fault evolution trend is of great significance for maintenance personnel to make timely repairs and make maintenance decisions. This paper introduces a method for predicting the fault evolution trend based on FA-LSTM. Combined with the combustion chamber degradation mechanism, the factor analysis method (FA) is used to construct the health factor (HI) to measure the health status of the combustion chamber of the gas turbine. The unique ability of the long short-term memory neural network (LSTM) to process time series data is used to predict the evolution trend of the combustion chamber fault. A certain type of gas turbine is taken as the research object, and the proposed method is compared with other 6 traditional machine learning methods. The prediction results of this method are the lowest, MAE and RMSE, which can achieve accurate degradation trend prediction and provide an effective basis for short-term maintenance.
Keywords: gas turbine; fault evolution trend; factor analysis; long short-term memory neural network
0 Introduction
With the acceleration of industrialization, the role of key energy equipment is becoming increasingly important. It is crucial to ensure the safe operation of gas turbines and accurately predict their fault development trends. China specifically pointed out the goal of mastering key technologies such as gas turbine design, testing and manufacturing in its “Made in China 2025” plan, and plans to establish a comprehensive R&D system by 2025. The high-temperature components of gas turbines need to face extreme temperature, pressure and speed conditions during operation, and are the core components of the gas turbine. Increasing the initial temperature is the main way to improve the performance of gas turbines, which also makes high-temperature components face greater challenges. Gas turbines operate for a long time in harsh environments, resulting in more than half of the failures being related to the combustion chambers that are subject to extreme temperatures. In the past few years, the operation and maintenance of gas turbines worldwide have accumulated more than trillions of dollars in expenses.
Reasonable planning of the maintenance plan of the gas turbine combustor is crucial to improving the safety of gas turbine operation, avoiding unplanned downtime and reducing maintenance costs. Therefore, the degradation trend prediction technology of the gas turbine combustor has become one of the core technologies to support the safe operation and maintenance of advanced gas turbines.
Usually, the degradation trend prediction mainly includes physics-based and data-driven methods. Due to the great difficulty in achieving high-precision physical modeling, the practicality of the physics-based method is insufficient. With the development of artificial intelligence technology, data-driven methods have received more attention. In order to accurately evaluate the health status of the gas turbine and further clarify the process and degree of its performance degradation, the key is to extract key indicators reflecting the working condition of the gas turbine from the monitoring data. These indicators are collectively referred to as health parameters (HI). Volponi uses fuel flow to evaluate the HI of the gas turbine. At present, most airlines choose exhaust temperature as the performance evaluation indicator of aircraft engines. However, a single physical parameter cannot accurately represent such a complex power system. Many scholars have proposed a variety of HI extraction methods based on principal component analysis, autoencoder, linear regression, etc.
Although scholars have conducted a lot of research on the construction of gas turbine health parameters, there are few studies on how to calculate the future changes in the health parameters of the combustion chamber, that is, the prediction of the combustion chamber degradation trend. This paper takes a certain type of gas turbine as the object and studies the method of predicting the fault evolution trend of the combustion chamber based on the gas turbine model. The engine health parameter HI is constructed by factor analysis combined with mechanism analysis. The prediction of the fault evolution trend based on LSTM is the best compared with other methods. This paper realizes the accurate prediction of the degradation trend of the gas turbine combustion chamber, providing an effective basis for short-term maintenance.
1 Combustion chamber performance degradation simulation and health status evaluation
1.1 Combustion chamber performance degradation mechanism and degradation propagation method
The gas turbine combustion chamber is one of the key components of the gas turbine, mainly composed of flame tube, fuel nozzle, swirler, etc. During operation, flame tube fouling, carbon deposition, cracks, nozzle coking, combustion degradation, uneven fuel distribution, thermal deformation and corrosion may occur, resulting in reduced combustion efficiency.
This paper adopts the more mature damage propagation model of combustion engine component performance degradation proposed by NASA. The data generation process is as follows:
(1) Select the initial degradation (e0, f0): e0∈ [0.99, 1], f0∈ [0.99, 1] (1)
(2) Apply the exponential change rate of flow and efficiency loss to each data set, as described in the formula h(t) = 1-exp{atb}. The negative effects of this fault gradually intensify, and over time, affect the overall health index H(t) = g(e(t), f(t)). The trend and progress of performance degradation depend on the following factors:
fi, ei≤ 1%, ak∈ [0.001, 0.003], bk∈ [1.4, 1.6], k=1, 2 (2)
(3) When the health index reaches the predetermined threshold H(t) = 0, it terminates (i.e., it reaches the failure standard).
(4) The noise during measurement is added to the output data.
1.2 Performance degradation simulation model
After determining the combustion chamber degradation mechanism and propagation method, the combustion chamber fault simulation model is built and degradation simulation is performed. This study takes a certain type of gas turbine as the object. The combustion chamber of the unit is equipped with 8 burners, and 17 thermocouples are used to measure the exhaust temperature at the turbine outlet.
In order to more accurately reflect the performance degradation process of the gas turbine combustion chamber, this paper combines the actual needs and establishes a gas turbine model with multiple flame tubes, multiple turbine channels, and considering the gas rotation mixing effect. The specific modeling method is explained in the literature, and the established gas turbine model is shown in Figure 1.
The efficiency degradation coefficient is injected into each combustion tube. At the same time, considering that the degradation speed of different flame tubes in actual conditions is not completely consistent with the degradation process, a proportional factor is applied to the combustion efficiency. Assuming that the combustion chamber efficiency degradation coefficient drops from 100% to 95% when it is completely faulty, the exhaust temperature data of the degradation process of one of the combustion chambers of the gas turbine is finally obtained as shown in Figure 2. Among them, one operating cycle is one day, and the number of operating cycles is determined according to the design life of the components.
This paper conducts 100 sets of simulations on the performance degradation process of gas turbine combustors. Each gas turbine has a different life span, which is used for the training and testing of the subsequent gas turbine performance degradation trend prediction model.
2 Evaluation of combustion chamber health status
2.1 Factor analysis (FA)
In order to integrate the information of multiple performance parameters and more comprehensively map the performance degradation level of the gas turbine, this section applies factor analysis (FA) to extract common factors between different sensors as indicators for evaluating the performance degradation of the gas turbine.
The common factors in factor analysis refer to the common elements that cannot be directly observed but actually affect each variable. Each variable can be regarded as a linear combination of common factors plus a specific factor. This method aims to discover the correlation between the original variable set and explore the underlying structure of these data, namely: Xi = ai1F1 + ai2F2 + … + aimFm + εi (i = 1, 2, …, p) (3)
The matrix form of the model is as follows: X = AF + ε (4)
The factor analysis technology is used to extract key health parameters from the numerous sensor data of the gas turbine to evaluate the health status of the gas turbine. The specific steps are as follows:
(1) Before performing factor analysis, it is necessary to confirm whether the selected variables are suitable for this analysis. To this end, the KMO (Kaiser-Meyer-Olkin) measurement and Bartlett sphericity test are first performed. If the test result rejects the null hypothesis, it indicates that there is a certain correlation between the variables, which is suitable for factor analysis; on the contrary, if the null hypothesis is not rejected, it means that these variables may be independent and provide different information, which is not suitable for factor analysis.
(2) Extract common factors to construct health parameters HI: If the null hypothesis H0: m = 1 is accepted in the model test at the significance level of 0.05, it shows that it is appropriate to use a factor model containing a common factor to fit the original data. Therefore, a factor can be extracted and, after normalization, used as a health index (HI) that fully reflects the status of the gas turbine.
2.2 Factor analysis test results of degradation data
This paper selects all 17 exhaust temperatures as measurement results for factor analysis. After calculation, the results of the degradation data KMO test and Bartlett sphericity test are shown in Figures 3 and 4.
As shown in Figure 3, the test values of 100 gas turbines in this data are all greater than 0.75, indicating that this data is suitable for factor analysis. As shown in Figure 4, all gas turbine levels are significant, and the p values are all less than 0.05, rejecting the assumption that the Bartlett correlation coefficient matrix is a unit matrix, which is very suitable for factor analysis.
Therefore, the factor analysis method is used to construct the health parameters of the 100 gas turbine exhaust temperature data obtained by simulation, and a gas turbine is randomly selected for display, as shown in Figure 5.
As shown in Figure 5, the combustion chamber of the gas turbine gradually declines from complete health to complete failure, and shows an exponential decline trend, which is in line with the original definition of HI.
3 Prediction of high-temperature component failure evolution trend based on LSTM
3.1 Analysis of performance degradation trend prediction problems
When analyzing the correlation of health parameter sequences, compared with autocorrelation analysis, partial autocorrelation analysis can effectively eliminate the interference of other variables in the time series on the previous and subsequent orders. Therefore, this paper chooses to use the partial autocorrelation method to study the time dependence between health factor data.
As shown in Figure 6, the partial autocorrelation analysis results of the established health factors reveal the time dependence in the data, which shows that the time series has predictable characteristics. In view of this, this study selected the long short-term memory network (LSTM), which is a learning method suitable for processing sequence data and can deal with data without long-term time dependence. The LSTM model is very suitable for solving time series prediction problems.
The process of neural network prediction of short-term gas turbine performance degradation time series is called “sliding window rolling”, as shown in Figure 7. The historical cycle monitoring data before the current operation cycle of the gas turbine is used as the training sample set to train the neural network model, and the prediction target is the performance index in the short term after the current cycle.
3.2 Long Short-Term Memory Neural Network (LSTM)
Long Short-Term Memory Network (LSTM) is an efficient recurrent neural network that is particularly good at processing sequence learning tasks. LSTM finely regulates the flow of information through its input gate, forget gate and output gate. This structure makes it very suitable for predicting the performance degradation trend of gas turbines.
As shown in Figure 8, there are 3 LSTM units. The input and output are 3 adjacent different moments in the time series. The input is xt-1, xt, xt+1, and the output is the hidden state ht-1, ht, ht+1.
3.3 Performance degradation trend prediction results
First, 100 sets of gas turbine degradation process data are divided into training sets and test sets in a ratio of 8:2, and then the data is used for model training and testing. After the model optimization experiment, the final LSTM model structure is shown in Table 1.
The sample is constructed by sliding window, and the first 80 moments are used to predict the next moment. The rolling prediction method is used to predict the health parameters of the gas turbine for the next 80 cycles, and the degradation trend of the gas turbine is predicted. The prediction results at a certain moment are shown in Figures 9 and 10.
It can be seen that the prediction results of the degradation trend of the gas turbine based on LSTM are highly accurate and monotonic, indicating that the model also has the function of eliminating dynamic noise.
3.4 Comparison of prediction results of different methods
In order to reflect the effectiveness of the method proposed in this paper, six other traditional data-driven methods are compared, including: feedforward neural network (FNN), convolutional neural network (CNNN), radial basis function (RBF), autoregressive model (AR), extreme learning machine (ELM) and gradient boosting decision tree (GBDT). The parameter settings of the six traditional methods used for comparison are shown in Table 2.
The degradation trend of the next 80 steps is predicted at the 25%, 50%, and 75% life cycle positions and the last 80 cycles of a gas turbine. The prediction results of different methods are shown in Figures 11 to 14.
From the comparison of prediction results, it can be clearly seen that LSTM has better prediction accuracy and linearity than other methods, and can better reflect the degradation trend of the combustion chamber. In order to evaluate the prediction performance and generalization ability of the model, this study uses the root mean square error (RMSE) and the mean absolute error (MAE) as quantitative indicators to measure the rolling prediction effect. Calculate the error value of the prediction result of the next n steps at each moment of the whole life cycle, and calculate the error when the number of prediction steps is n: see formula (5) in the figure
The error comparison of different models is shown in Figures 15 and 16.
It can be seen that due to the consideration of time series information, compared with the other six methods, LSTM has higher accuracy throughout the entire prediction period, with the lowest RMSE and MAE. At the same time, the error grows more slowly during the rolling prediction process, and the prediction results are more reliable.
4 Conclusion
Considering that a single physical parameter cannot accurately describe a complex gas turbine system, and there are few existing studies on the prediction of combustion chamber fault evolution trends. This paper takes the model gas turbine as the research object and establishes a fault evolution trend prediction method based on FA-LSTM. The engine health parameter HI is constructed by factor analysis, and the LSTM neural network’s unique ability to process time series data is used to accurately predict the fault evolution trend HI. The main conclusions are as follows:
(1) The mechanism and impact of combustion chamber performance degradation are analyzed, and the propagation process of degradation is explored. The study quantitatively analyzed the health status of the gas turbine during the performance degradation period and proposed a health status characterization method based on factor analysis, thereby constructing a comprehensive health index (HI) that reflects the status of the gas turbine.
(2) The LSTM-based method can accurately predict the performance of the gas turbine. The results show that it has good robustness and accuracy. Compared with the other six traditional machine learning methods, the prediction results of this method have the lowest MAE and RMSE, which can achieve accurate degradation trend prediction.
