Aero-engine performance interval prediction method considering maintenance factors

Abstract

The invention provides an aero-engine performance interval prediction method comprehensively considering maintenance factors. At present, prediction results of most aero-engine performance degradation prediction methods are point prediction, only future performance reference values are provided, and performance recovery caused by maintenance factors is often ignored. On the basis, the invention designs a brand-new QCformer method. According to the method, Pinball loss is adopted as a loss function, and performance interval prediction of the aero-engine is achieved. And the method adopts event coding to model maintenance activities so as to predict possible performance recovery brought by different maintenance strategies in the future. According to the prediction mode, airline companies and maintenance teams can have deeper understanding of the performance change trend of the engine in the service period, reference is provided for making a maintenance strategy, the flight safety and reliability can be ensured, and the maintenance cost is reduced.

Description

Technical Field

[0001] The invention belongs to the field of aircraft engine whole machine performance degradation prediction, and solves the problem of performance degradation trend prediction of aircraft engines during field use by designing an interval prediction method. Background Art

[0002] As the main power source of aircraft, the reliability of aircraft engines is crucial to flight safety. Aircraft engines work in high temperature and high pressure environments for a long time, and need to switch between operating modes continuously to meet the power requirements of different flight missions. Continuous and frequent switching of operating modes accelerates the degradation process and may cause malfunctions or failures. The performance evaluation and prediction of aircraft engines is one of the key issues in PHM. By analyzing and predicting the performance change trend of the engine, more effective maintenance plans can be formulated to improve overall performance and extend engine life.

[0003] Existing prediction methods can be divided into two categories: physics-based methods and data-driven methods. Physics-based methods use thermodynamic calculations and geometry and physics knowledge to build a mathematical model of the equipment to predict the performance degradation behavior of the engine. Although physics-based methods are highly interpretable, the aircraft engine is an extremely complex aerodynamic and thermodynamic rotating mechanical system. The modeling and calculation complexity of building its clear mathematical model is high, and there is an inaccurate modeling problem for some complex physical processes. In recent years, with the development of sensor technology and artificial intelligence, data-driven prediction technology has gradually become an important tool for realizing engine performance prediction. Data-driven methods use machine learning technology to learn system behavior directly from the collected condition monitoring (CM) data, and then use the learned knowledge to predict the degradation trend. Engine performance prediction is an important part of aircraft engine maintenance and health management. Through real-time monitoring and analysis of engine operation data, abnormal conditions of the engine can be discovered in time, and corresponding measures can be taken for maintenance and adjustment to ensure the stable operation of the engine.

[0004] The performance change trend prediction problem can be reduced to a sequence-to-sequence regression problem (long time series prediction problem). Wang et al. realized multi-step prediction of degradation trends through LSTM. Yang et al. proposed a network framework DR-DTPN that integrates data repair and degradation trend prediction, which effectively solved the problem of serious deviation in equipment degradation trend prediction caused by missing monitoring data and distribution changes. Ou et al. designed an AdaDSC to characterize the degradation feature sequence as multiple stages with different distributions, which effectively solved the problem of time distribution differences in performance degradation trend prediction. The prediction results of the above inventions can only give a reference value of the engine performance at a certain moment in the future, and ignore the impact of maintenance activities during the use of the equipment. The main purpose of predicting the performance of an aircraft engine is to be able to determine whether it can work reliably in the future and provide a reference for the formulation of its maintenance strategy. Therefore, it is more desirable to predict the range of performance changes of an aircraft engine to ensure that it can still work effectively in the worst case. In addition, how to model the engine performance recovery brought about by maintenance activities brings challenges to the prediction of engine performance degradation.

[0005] In order to solve the problem of extracting degradation trends and predicting performance change intervals considering maintenance activities, the present invention designs a new network structure QCformer based on the classic Transformer architecture. Specifically, it introduces maintenance events into the network as a special position encoding to achieve modeling of maintenance events. Pinball loss is used as the loss function to achieve the prediction of performance change intervals. In addition, the 1D CNN strategy is introduced in the Encoder learning process to effectively learn local features while reducing the complexity of model training. Summary of the invention

[0006] In order to solve the current problem of performance interval prediction after steady-state data collection of aircraft engines in the field, the present invention proposes an aircraft engine performance interval prediction method (QCformer) taking maintenance factors into consideration, which can effectively solve the problem of difficult performance prediction of aircraft engines after maintenance.

[0007] The aircraft engine performance interval prediction method (QCformer) considering maintenance factors includes four parts: data processing, model building, model training and prediction. It uses 1D CNN to enhance the model's sensitivity to data in time series. At the same time, event coding is added so that maintenance factors can be fully considered. Encoders, decoders and pinball loss are used to obtain a prediction model that can output data distribution intervals. Compared with other engine performance prediction methods, QCformer can take into account the impact of event factors when constructing a prediction model, especially for the cleaning and maintenance that are often carried out in the field work of aircraft engines, and effectively constructs a method for predicting engine performance parameters in actual situations.

[0008] The present invention proposes a method for predicting the performance interval of an aircraft engine taking maintenance factors into consideration, comprising the following steps:

[0009] Step 1: Process the data collected from the field aircraft engine to construct a training set;

[0010] The method first performs similarity conversion to standard atmospheric conditions. Then, a sliding window method is used to collect the data set for model training. in represents m input variables of length L, i.e. the sequence to be predicted, represents k output variables of length H, i.e., the prediction sequence.

[0011] Necessary notes for step 1: Common performance parameters include power, instantaneous consumption SFC, exhaust gas temperature EGT and speed Ng, and their corresponding similar conversion formulas are:

[0012]

[0013]

[0014] Step 2: Define the predicted target variable y j Confidence level of

[0015] Necessary note for step 2: Define α as the preset confidence level. Interval prediction requires predicting the target variable y in the data distribution with a given confidence level of 1-α. j Range and satisfy:

[0016]

[0017] The predicted value y of the target variable y corresponding to the qth quantile in the data distribution q , expressed as:

[0018] y q=argmin y {F(y)≥q}

[0019] where F(y) is the cumulative distribution function of the variable y.

[0020] Step 3: Construct an attention mechanism network QCformer containing a CNN layer;

[0021] The method comprises two parts, an encoder and a decoder, wherein the encoder is used for learning sequence features, and the decoder is used for generating a target sequence interval that needs to be predicted.

[0022] Necessary note for step 3: In order to improve the model’s sensitivity to local information of sequence data, a 1D CNN layer is introduced in the encoder. Specifically, the sequence The input is sent to the 1D CNN layer, and a convolution operation with a step size of 1 is performed along the time dimension to extract local information. After being mapped by the convolution layer, it is transformed into a length of L and a dimension of d. m new sequence of Then the location information of the data is introduced through position coding, and the coding method is:

[0023]

[0024] In the formula, pos represents the position, i represents the dimension, and d m represents the dimension of the feature vector, and to ensure that the positional encoding remains symmetric d m The method considers maintenance activities as a special kind of position information. The maintenance activities in the engine life cycle are limited, so the position is encoded as a discrete variable, and each different type of maintenance activity is represented by a unique value, which is mapped to d through the embedding layer. m Maintenance, get the maintenance event code Afterwards, the event encoding is subjected to the same one-dimensional convolution operation to Maintaining time consistency can be expressed as The final vector input to the encoder For sequence information Positional encoding and event code The sum of the three is

[0025] The attention module consists of a self-attention layer, residual connection, and LayerNorm, and its output sequence dimension is the same as the input sequence dimension. The self-attention mechanism includes matrices Q (query), K (key value), and V (value), which are all obtained by linear transformation of the input; the self-attention layer enables the model to dynamically capture the relationship between different positions in the sequence. The calculation formula is as follows:

[0026]

[0027] Among them, d k Indicates the number of columns of K and V matrices, i.e., the vector dimension. As a scaling factor, to prevent excessive input dimensions from affecting the calculation of attention weights. After the self-attention calculation is completed, the input and output of the self-attention layer are summed up through the residual connection and the LayerNorm operation is performed to obtain the final output of the self-attention module.

[0028] After completing the attention calculation, the calculation results of the self-attention module are input into the feedforward network for further feature learning. The feedforward network consists of two fully connected layers, using ReLU as the activation function and performing LayerNorm operation at the same time, and introducing residual connection. The calculation formula is as follows:

[0029] f1=LayerNorm(f+max(0,fW1+b1)W2+b2)

[0030] in, is the input of the feedforward network, is the output of the network, W1 and W2 are the weights of the first fully connected layer and the second fully connected layer, respectively, and b1 and b2 correspond to the biases of the two fully connected layers, respectively.

[0031] After the encoder feedforward network calculation is completed, the output features are operated with 1D CNN to enhance the local feature learning ability of the model while halving the length of the input sequence. The convolution kernel size of this convolution layer is 3 and the convolution stride is 2. ReLU is used as the activation function and the maximum pooling operation is performed. The final output can be expressed as:

[0032] f2=MaxPool(ReLU(Conv1d(f1)))

[0033] In model construction, multiple encoders are usually stacked to increase the model depth and expressiveness. In each calculation of the encoder, the length of the output sequence is halved compared to the length of the input sequence. e The number of encoders is , compared with the original Transformer model, the computational complexity is reduced from Reduce to The η value is smaller than that of the original Transformer encoder. Therefore, the 1D CNN module in QCformer improves the model's ability to learn local information of sequence features while reducing the model's computational complexity.

[0034] The decoder is used to generate the target sequence. The decoder consists of five parts: mask attention module, position encoding, event encoding, cross attention module and feedforward network. The first layer of the decoder receives input similar to the encoder, which includes sequence features, position encoding and event encoding. The sequence features include the full labeled sequence and placeholder sequences It is expressed as:

[0035]

[0036] Represents the target sequence y j The front length is H token The position is used as the start mark. Different from the encoder using 1DCNN calculation, the decoder maps the features from K dimensions to d through a fully connected layer. m The position encoding and time encoding are implemented in the same way as the encoder, and are extended to d through the embedding layer. m After dimensioning, directly input the encoder.

[0037] After combining sequence features, position encoding, and event encoding, they are input into the masked multi-head attention module, which is used to hide information related to subsequent sequences to avoid future information leakage. In the masked multi-head attention module, masking is implemented through a mask tensor. The position with a value of 1 in this tensor represents accessible information, while the position with a value of "-inf" represents inaccessible information. During the calculation process of the masked self-attention mechanism, the mask tensor is multiplied element-by-element with the query, key, and value to achieve information shielding. The formula is as follows:

[0038] Q * =Q⊙M

[0039] Where M is the mask tensor, Q is the computation query, and ⊙ is an element-wise multiplication so that information about future positions is ignored when computing the attention score.

[0040] The query matrix Q in the cross-attention module is the output of the masked multi-head self-attention module in the decoder, and the key matrix K and value matrix V are the outputs of the encoder. After the cross-attention module is calculated, further feature learning is also performed through the feedforward network. The decoder feedforward network structure is exactly the same as the encoder. d The calculation results after stacking are input into the fully connected layer to obtain the final prediction result:

[0041]

[0042] in is the prediction result of the starting token sequence. It is the prediction result of the target token sequence.

[0043] Step 4: Use Pinball loss as the loss function to predict the performance change range.

[0044] Necessary note for step 4: Pinball loss is used as the model optimization target, which calculates the general conditional probability distribution of the target variable. The function is defined as follows:

[0045]

[0046] Among them, y t is the actual observed value, is the predicted value at the quantile level q∈(0,1); when the predicted quantile exceeds the actual value, the penalty will be adjusted by (1-q); if it is lower than the actual value, the penalty will be adjusted by q.

[0047] Step 5: Input the training set data into the constructed network, update the model parameters through the Adam algorithm to minimize the loss function, and perform network training until convergence.

[0048] Necessary note for step 5: During training, the Adam algorithm calculates each parameter w of the model i and its corresponding gradient g i , and update the first-order moment estimate m t and the second-order moment estimate v t , respectively, the deviation correction is and Update each parameter based on the corrected first-order and second-order moment estimates. The function is defined as follows:

[0049]

[0050] Where η is the learning rate and ∈ is a constant added for numerical stability, which is 10 -8 .

[0051] Beneficial effects:

[0052] QCformer improves the Transformer model by adding an event encoding module and combining it with Pinball loss to build a prediction model for turboprop engines including maintenance activities, so as to better realize the prediction of aircraft engine performance change intervals. At the same time, a one-dimensional convolutional layer neural network is introduced into the encoder to enhance the model's sensitivity to data timing and effectively reduce the encoder's training complexity. The QCformer model is used in the analysis of the performance degradation trend of the entire aircraft engine to complete the performance prediction of the aircraft engine during field use. Through a large number of performance degradation trend prediction experiments on a certain type of turboprop engine field use data, it is shown that compared with traditional data-driven methods and many numerical calculation methods, the QCformer model is a more practical and effective method in the field of aircraft engine performance prediction. Considering the influence of various uncertainties and maintenance factors, it can avoid the inaccuracy of a single precise prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a diagram of the QCformer encoder input information;

[0054] Figure 2 This is a schematic diagram of the mask operation calculation in the masked self-attention calculation;

[0055] Figure 3 It is the comparison of root mean square error of prediction results of different models;

[0056] Figure 4 It is the comparison of the root mean square error of the prediction results of different models under different prediction windows;

[0057] Figure 5 It is the comparison of the prediction results of QCformer and QCformer\E for TP001;

[0058] Figure 6 This is the overall framework diagram of QCformer. DETAILED DESCRIPTION

[0059] This case takes a certain type of airborne turboprop engine as the invention object. This type of engine is widely used in large transport aircraft and some rescue aircraft. Specifically, this type of engine is a fixed turbine single-rotor turboprop engine, which uses an axial flow compressor to provide compressed air for the combustion chamber, and transmits the shaft power output by the engine to the propeller through a reducer. During each flight, multiple steady-state working data are collected from the flight parameters for the performance analysis of the engine as a whole. The sampling interval is 1 hour, and the exhaust temperature is used as the performance indicator of this type of engine.

[0060] The exhaust temperature of the past 40 hours is used as the input training set to predict the range of exhaust temperature variation in the next 10 hours (the maximum single flight time of this engine is 4.96 hours). In order to evaluate the performance of the proposed QCformer, five commonly used and popular methods are selected for comparison. These methods are as follows:

[0061] 1) PF: Particle filtering is a Monte Carlo-based method that uses multiple system state assumptions and observation data to adjust weights to provide comprehensive state estimation and quantile forecasting for time series;

[0062] 2) QLSTM: LSTM is an enhanced recurrent neural network with a gating mechanism and a memory module, which can avoid gradient explosion and enhance long-term memory capacity;

[0063] 3) QTCN: TCN is a time series modeling technique that uses convolutional neural networks to capture long-range dependencies. It has features such as parallel computing and parameter sharing;

[0064] 4) QTransformer: QCformer is based on the encoder and decoder structure of Transformer. Compared with QTransformer, it enhances the network learning ability and efficiency. The two models share similar network parameters. The key difference is that there is no 1dCNN in QTransformer.

[0065] 5) AdaDSC: AdaDSC divides the degraded feature sequence into multiple degradation stages and combines it with the Distribution Difference Evaluation Gated Recurrent Unit (DDE-GRU) module to handle the distribution difference and capture the temporal dependency in the data.

[0066] 6) QCformer\E: By excluding event coding in QCformer, this method aims to demonstrate the effectiveness of event coding in modeling maintenance activities. QCformer performs regression tasks at the 0.05, 0.5, and 0.95 quantile levels and generates prediction points at the 0.5 quantile level. The specific hyperparameter values ​​of QCformer are shown in Table 1.

[0067] Table 1 QCformer hyperparameter values

[0068]

[0069] In this paper, both point prediction and probability interval prediction experiments were conducted. In each case, different metrics were used to evaluate the performance of the model. The specific metrics are as follows:

[0070] 1) Point prediction metrics: The root mean square error (RMSE) is used to evaluate the average level of overall prediction error between the model output and the actual value. In addition, the mean absolute percentage error is used to quantify the deviation between the predicted value and the actual value. The respective calculation formulas are as follows:

[0071]

[0072] Among them, y i is the actual observed value, is the model prediction value, and N is the number of samples in the test dataset.

[0073] 2) Probabilistic prediction metrics: In addition to the average quantile loss (AQL) to evaluate the reliability and clarity of quantile regression, the prediction interval coverage probability (PICP) is also used to evaluate the reliability of the prediction interval. The PICP calculation method is as follows:

[0074]

[0075] Among them, ζ means it falls within the confidence interval [Low i ,Up i ] is the number of H predicted targets in .

[0076] In order to verify the superiority of the QCformer method, three sets of experiments were conducted using the collected data of 12 engines. The training and test data sets of each experiment are set as shown in Table 2. Each experiment was conducted 5 times with different random seeds, and the average value of the statistical test results was obtained as shown in Table 4. In addition, the prediction value corresponding to the 0.5 quantile is used to represent the point prediction.

[0077] The PICP index is the basis for evaluating the reliability of the probabilistic prediction model, and the AQL result provides a comprehensive score of the probabilistic prediction skill. The results show that the accuracy of the exhaust temperature EGT fluctuation range predicted by all models is above 90% (i.e., PICP is greater than 0.9), especially the proposed QCformer model can 100% cover the observed exhaust temperature prediction interval under the confidence interval of [0.05, 0.95] in all three groups of experiments. In addition, our QCformer model obtains a better AQL score than the other two competing models by minimizing the quantile loss. From the perspective of the two evaluation indicators, the probabilistic prediction results of the proposed QCformer model are more informative and competitive, which is convenient for better evaluation of the performance of the engine during future use.

[0078] In addition, the proposed QCformer model outperforms all other baseline models in terms of MAPE and RMSE indicators for point prediction. In three sets of experiments, QCformer improves the MAPE standard by at least 0.8% and the RMSE standard by 3.16 compared with other comparison methods. Figure 3 The RMSE distribution of the prediction results of the QCformer model for the test samples is further shown. It can be seen that the variance of the QCformer prediction results is smaller than that of other comparison models. This shows the advantage of the QCformer model in predicting aircraft engine performance, because it can better learn local features from historical sequences while taking into account the learning of long-distance dependency characteristics. Further comparison of the prediction results of QLSTM, QTCN and QTransformer shows that QLSTM has the worst prediction effect, and QTransformer's prediction results are better than QLSTM but not as good as QTCN. We can conclude that in time series prediction tasks, it is necessary to learn local features and capture long-distance dependencies.

[0079] In order to evaluate the computational cost of the proposed method, we compared the training cost of all methods. Except for PF, all methods are deep learning methods. It can be seen that the number of parameters and computational efficiency of the PF method are much higher than other methods. Compared with other methods, AdaDSC has a large number of domain difference evaluation calculations, so its computational efficiency is the lowest. Although QTransformer and QCformer can improve the computational efficiency of the network through parallel computing, their parameters are more than QLSTM, so more computing resources are required. Compared with QTransformer, QCformer has a higher number of parameters due to the introduction of the 1D CNN mechanism, but its computational efficiency is higher because the length of the sequence is halved each time it passes through the Encoder. Therefore, the proposed QCformer improves the computational efficiency on the basis of improving the prediction accuracy of the model.

[0080] Table 2. Turboprop engine steady-state data set

[0081]

[0082] Table 3 Comparison and evaluation results of all models

[0083]

[0084] Furthermore, in order to analyze the performance changes of the model under different prediction windows, Task 1 is performed under windows 5, 10, 15, 20, and 25. The experimental results are shown in Figure 4 As shown in the figure, it can be seen that the performance of all models decreases with the length of the prediction window. When the prediction window is 20, the RMSE and AQL of the PF prediction results are higher than those of QLSTM. This shows that the deep learning model is more robust in long time series prediction. In addition, in all tasks, the RMSE and AQL of QCformer are the smallest among all methods. This shows that QCformer has excellent performance in long time series prediction and is more suitable for engine performance prediction.

[0085] In addition, aircraft engines often undergo maintenance during use, and engine performance will recover after maintenance. In the present invention, maintenance activities are encoded through event coding. In order to illustrate the effectiveness of event coding in modeling maintenance activities, the prediction results of the model without adding maintenance coding are compared with the prediction results of adding maintenance coding. As can be seen from Table 3, in the three tasks, the QCformer model is superior to QCformer\E in the MAPE and RMSE indicators of point prediction. And the accuracy of QCformer's prediction range of exhaust temperature in the three tasks is 100%, while QCformer\E has a 100% accuracy rate in the prediction range of exhaust temperature only in task 3. Figure 5 The prediction results of QCformer and QCformer\E for TP001 in Task 1 are shown. It can be seen that QCformer\E is not accurate enough in predicting the exhaust temperature before and after the maintenance, while QCformer can more accurately predict the impact of maintenance activities on engine performance changes. This shows that event coding can effectively model maintenance activities and effectively predict the effect of potential maintenance activities on engine performance recovery, providing a reference for the formulation of engine maintenance strategies.

Claims

1. A method for predicting aircraft engine performance intervals taking maintenance factors into consideration, characterized in that: The steps include: Step 1: Process the data collected from the field aircraft engine to construct a data set, which includes input variables and output variables. The input variables are engine performance parameters, and the output variables are prediction target variables that characterize the engine performance range. Step 2: Construct an attention mechanism network QCformer containing a CNN layer, which includes an encoding part and a decoding part. The encoding part includes a one-dimensional convolutional network, a position encoding, an event encoding, and an encoder, and the decoding part includes a fully connected layer, a position encoding, an event encoding, and a decoder. The event encoding refers to encoding the maintenance event corresponding to the input variable. The input of the encoder is the sum of the input variable, the position encoding, and the event encoding after being processed by the one-dimensional convolutional network. The input of the decoder is the sum of the output variable, the position encoding, and the event encoding after being processed by the fully connected layer. Step 3: Use the dataset to train the QCformer model; Step 4: Input the current aircraft engine performance parameters into the trained QCformer model and output the prediction of the performance change range.

2. The method for predicting aircraft engine performance intervals considering maintenance factors according to claim 1, characterized in that: Step 1 specifically includes: converting the engine performance parameters collected under different working conditions to standard atmospheric conditions through similarity conversion, dividing the converted data into n segments using a sliding window method, and each segment of data in represents m input variables of length L, represents k predicted target variables of length H.

3. The method for predicting aircraft engine performance intervals considering maintenance factors according to claim 2, characterized in that: Also includes the steps: define the predicted target variable y j The confidence level includes: defining α as the preset confidence level, and interval prediction requires predicting the target variable y in the data distribution with a given confidence level of 1-α j Range and satisfy: The predicted value y of the target variable y corresponding to the qth quantile in the data distribution q , expressed as: y q =argmin y {F(y)≥q} where F(y) is the cumulative distribution function of the variable y.

4. The method for predicting aircraft engine performance intervals considering maintenance factors according to claim 1, characterized in that: The encoder includes an attention module and a feedforward network, the output of the attention module is connected to the input of the feedforward network; the attention module is composed of a self-attention layer, a residual connection and a LayerNorm. After the self-attention calculation is completed, the input and output of the self-attention layer are added through the residual connection, and the LayerNorm operation is performed to obtain the final output of the attention module; the feedforward network is composed of two fully connected layers, ReLU is used as the activation function, the LayerNorm operation is performed at the same time, and the residual connection is introduced.

5. The method for predicting aircraft engine performance intervals considering maintenance factors according to claim 4, characterized in that: After the encoder feedforward network calculation is completed, a one-dimensional convolutional network operation is performed on the output features to enhance the local feature learning ability of the model while halving the length of the input sequence; the one-dimensional convolutional network has a convolution kernel size of 3, a convolution stride of 2, uses ReLU as the activation function, and performs a maximum pooling operation.

6. The method for predicting aircraft engine performance intervals considering maintenance factors according to claim 1, characterized in that: The QCformer model stacks multiple layers of encoders to increase the model depth and enhance the expressiveness, and after each iteration, the output sequence length is halved compared to the input sequence length.

7. The method for predicting aircraft engine performance intervals considering maintenance factors according to claim 1, characterized in that: The decoder includes a mask attention module, a cross attention module and a feedforward network; the query matrix Q in the cross attention module is the output of the mask attention module in the decoder, and the key matrix K and the value matrix V are the output of the encoder; the output of the cross attention module is input into the feedforward network for further feature learning; The calculation results of the decoders after multiple stacking are input into the fully connected layer to obtain the final prediction result.

8. The method for predicting aircraft engine performance intervals considering maintenance factors according to claim 1, characterized in that: The QCformer model uses Pinball loss as the model optimization objective, which calculates the general conditional probability distribution of the target variable.

9. The method for predicting aircraft engine performance intervals considering maintenance factors according to claim 1, characterized in that: Step 3 specifically includes: inputting the data set into the constructed network, updating the model parameters through the Adam algorithm to minimize the loss function, and training the network until convergence.

Images