Aero-engine exhaust temperature threshold prediction method based on improved spectrogram convolutional neural network

Abstract

The invention provides an aero-engine exhaust temperature threshold prediction method based on an improved spectrogram convolutional neural network, and aims to solve the problem that a unified exhaust temperature threshold cannot be adopted due to gas path parameter dispersion of engines of the same model caused by factors such as a manufacturing process, assembly tolerance, working conditions and maintenance. Defining and analyzing high-dispersity cluster characteristics of aero-engine gas path parameters; the exhaust temperature threshold value prediction under the variable working condition is regarded as a classification problem; an improved spectrogram convolutional network is proposed to extract features from input data and construct a feature instance graph; a new Involution network is provided, and feature fusion is realized through games of a same seeker and a different seeker; and a fusion loss function of the cross entropy and the multi-kernel-maximum mean value difference is utilized to measure the difference of feature instance graphs formed by different engine data, and the exhaust temperature threshold value of the high-dispersity cluster feature engine is predicted. According to the method, the performance difference between the engines is fully considered, and the exhaust temperature threshold values of the engines in different service periods under variable working conditions can be accurately predicted.

Description

Technical Field

[0001] The present invention proposes a method for predicting the exhaust gas temperature threshold of an aeroengine based on an improved spectral graph convolutional neural network, belonging to the field of intelligent aeroengine health management. Background Art

[0002] The exhaust gas temperature is an important measure of the health status of an aero-gas turbine engine. Monitoring and predicting the exhaust gas temperature and the exhaust gas temperature threshold are important components of aeroengine prediction and health management technology, and play a very important role in the safety and reliability of the entire aircraft.

[0003] Currently, engines of the same model still adopt a unified exhaust gas temperature threshold. For example, the Federal Aviation Administration of the United States specifies a specific exhaust gas temperature threshold for each model. However, when analyzing the outfield flight data of engines of the same model, it is found that there is significant dispersion within a certain range for the same gas path parameters, showing the characteristics of a high-dispersion cluster. The reasons for this difference include: 1) production and assembly tolerances; 2) differences in working hours and states on different aircraft; 3) different degrees of sensor data noise pollution. These factors result in dispersion in the exhaust gas temperature thresholds among individual engines. At the same exhaust gas temperature value, some engines can operate normally for a longer time, while some are close to the threshold. Accurately predicting the exhaust gas temperature threshold of each individual based on gas path parameter data helps to evaluate the engine health status and predict the performance degradation trend. There is currently no literature on realizing the prediction of the exhaust gas temperature threshold for engines with high-dispersion cluster characteristics. Therefore, it is of great significance to study a new prediction method applicable to such aircraft groups.

[0004] The current mainstream prediction algorithms for the exhaust gas temperature threshold of aero-engines mainly fall into three categories: traditional prediction methods based on historical information, prediction methods based on physical models, and data-driven prediction methods. Among them, the data-driven method has received extensive attention because it avoids establishing complex mathematical models. The patent with the publication number CN113297680A discloses a method for analyzing the performance trend of a small bypass ratio aero-gas turbine engine. This method obtains the exhaust gas temperature data within a predetermined operating time and constructs an autoregressive moving average combined model based on the processed data to evaluate the exhaust gas temperature change and predict future performance. However, this method is a prediction of the numerical value of the exhaust gas temperature of an aero-engine for a period of time in the future, rather than the prediction of the exhaust gas temperature threshold. The patent with the publication number CN117168649A discloses a method for predicting the exhaust gas temperature of an aero-engine based on a graph neural network. This method divides graph nodes and edges through the prior knowledge of the aero-engine, constructs the node feature vectors of the sensor monitoring data, and establishes a graph neural network model. It uses a relational graph attention network to update the node features and fuse the component space information, thereby improving the accuracy of the exhaust gas temperature prediction. However, this method does not consider the non-negligible dispersion of the exhaust gas temperature thresholds of the engine fleet. Currently, there is no good method that can take into account the common and individual characteristics of engines of the same type and is applicable to the exhaust gas temperature threshold prediction model of the entire fleet. The proposed method for predicting the exhaust gas temperature threshold of an aero-engine based on an improved spectral graph convolutional neural network in the present invention can effectively solve this problem. Summary of the Invention

[0005] Aiming at the problems existing in the existing data-driven methods, the present invention proposes a method for predicting the exhaust gas temperature threshold of an aero-engine based on an improved spectral graph convolutional neural network, so as to realize the prediction of the exhaust gas temperature threshold of engines with high-dispersion cluster characteristics.

[0006] To achieve the above object, the concept and technical solution of the present invention are realized as follows:

[0007] The basic concept of the present invention is to define and analyze the high-dispersion cluster characteristics of aero-engine gas path parameters, and respectively explain the reasons why the gas path parameters of engines of the same model simultaneously have high dispersion and clustering; treat the prediction of the exhaust gas temperature threshold under variable working conditions as a classification problem to improve the efficiency, accuracy and easy evaluation of the method; propose an improved spectral graph convolutional network to extract features from input data and construct a feature instance graph; propose a new Involution network to learn and fuse features, and realize the feature fusion of two samples by having the similarity seeker and the difference seeker play a game and compete with each other during the training process; use the fusion loss function of cross entropy and multi-kernel maximum mean discrepancy to measure the differences between the feature instance graphs formed by different engine data, so as to realize the prediction of the exhaust gas temperature threshold of engines with high-dispersion cluster characteristics. The invention can accurately predict the exhaust gas temperature threshold of each engine in the fleet after different service cycles under different working conditions.

[0008] Based on the above basic concept, the technical solution proposed by the present invention is a method for predicting the exhaust gas temperature threshold of an aero-engine based on an improved spectral graph convolutional neural network, including the following steps:

[0009] Step 1: Define and analyze the high-dispersion cluster characteristics of aero-engine gas path parameters;

[0010] Step 2: Regard the prediction of the exhaust gas temperature threshold under variable working conditions as a classification problem;

[0011] Step 3: Propose an improved spectral graph convolutional network to extract features from input data and construct a feature instance graph;

[0012] Step 4: Introduce a new Involution network to realize the feature fusion of two samples during the training process by having the similarity seeker and the difference seeker play a game;

[0013] Step 5: Use the fusion loss function of cross entropy and multi-kernel maximum mean discrepancy to measure the differences between the feature instance graphs formed by different engine data, so as to realize the prediction of the exhaust gas temperature threshold of engines with high-dispersion cluster characteristics.

[0014] Further, in the process of defining and analyzing the highly dispersed cluster characteristics of the aero-engine gas path parameters in Step 1, it includes 3 operating parameters and 14 sensor parameters, specifically: the operating parameters are altitude alt, flight Mach number Mach, and throttle valve rotation angle TRA; the sensor parameters are total temperature at the outlet of the low-pressure compressor T24, total temperature at the outlet of the high-pressure compressor T30, total temperature at the outlet of the low-pressure turbine T50, total pressure at the outlet of the high-pressure compressor P30, physical speed of the fan Nf, physical speed of the core Nc, static pressure at the outlet of the high-pressure compressor Ps30, ratio of fuel flow to Ps30 phi, corrected fan speed NRf, corrected core speed NRc, bypass ratio BPR, flow enthalpy htBleed, coolant flow of the high-pressure turbine W31, and coolant flow of the low-pressure turbine W32.

[0015] Similar to the standard deviation and variance, the coefficient of variation can reflect the degree of data dispersion. It is defined as the ratio of the standard deviation to the mean:

[0016]

[0017] The coefficient of variation is a dimensionless index and can be used to compare two sets of data with different dimensions or different means. The coefficient of variation can be used to effectively measure the dispersion of aero-engines caused by factors such as manufacturing processes.

[0018] When the engine fails or its performance degrades due to a fault, there will be a significant difference between its operating state and the rated state at the time of leaving the factory. In this case, the parameters of each gas path component measured by the sensors will change over time. The premise of gas path analysis and performance trend tracking is to be able to obtain the health status of the engine from these changing parameters. For the research on engine health management, the relative data of the parameter changes during the engine operation are more suitable for evaluating the degree of dispersion of the engine performance, which is defined as:

[0019]

[0020] In the formula, Δx i is the difference between the value of a certain engine parameter at the end of its life and the value when it first entered service, is the mean value of all Δx i .

[0021] Further, in Step 2, predicting the exhaust temperature threshold under off-design conditions is regarded as a classification problem. The specific steps are as follows:

[0022] Step 2.1: Take the difference ΔEGT between the current exhaust temperature of the engine and the exhaust temperature threshold as the classification label;

[0023] Step 2.2: Taking points at equal intervals with ΔEGT as the axis, samples in the same interval correspond to the same label. Starting from 0, divide until the maximum ΔEGT in the fleet, and name them 0, 1, 2, …, n in sequence;

[0024] Step 2.3: Setting intervals of different sizes helps to improve the prediction accuracy and reliability. Initially, set a larger interval value. With fewer label numbers, the classification accuracy of the model is higher, and the prediction of ΔEGT is more reliable. Subsequently, set a smaller interval value. On the basis of ensuring reliable results, further narrow the range of ΔEGT to improve the prediction accuracy.

[0025] Furthermore, in step 3, an improved spectral graph convolutional network is proposed to extract features from the input data and construct a feature instance graph. Compared with the convolutional neural network, the spectral graph convolutional neural network does not require the premise that the data structure has translational invariance, and can extract the spatial features of irregular data structures, extract effective features to accurately train the model.

[0026] Based on spectral graph convolution, an improved spectral graph convolutional neural network is proposed. Introduce multiple filters of different sizes to learn features of different scales and fuse them together. The formula is as follows:

[0027]

[0028] In the formula, h is the weight, g θ is the filter, and X is the input signal.

[0029] Traditional graph convolutional neural networks can only aggregate information at a fixed scale. Multi-scale graph convolution obtains spatial information from multiple perspectives to obtain a powerful feature representation. The improved multi-scale graph convolution uses variable weights for the model to autonomously learn to assign different contribution degrees to spatial information of different scales, and obtains a better feature representation.

[0030] Furthermore, in step 4, a new Involution network is introduced. By allowing the assimilator and the differentiator to play a game, the fusion of the features of two samples is realized during the training process. Inspired by the Involution concept proposed by Kant in "Critique of Judgment", the Involution network is proposed to learn data features to prevent the trained model from being inaccurate or overfitting. Involution means spiral, so the literal meaning of Involution is to turn inward, giving a feeling of screwing in a screw, getting tighter and tighter until it fits perfectly. And the extended meaning of Involution is "changing inward", regarding oneself as an opponent and improving through self-completion.

[0031] The Involution network contains two adversarial models: the assimilator S and the differentiator D. The assimilator performs structural alignment on two input samples, enabling them to learn from each other and fuse features, and outputs samples x 1 and x 2 . The input of the differentiator is x 1 and x 2 , and it outputs a probability value indicating the probability that the sample comes from x 2 . The larger the value, the higher the probability of coming from x 2 ; the smaller the value, the higher the probability of coming from x 1 . The differentiator is used to distinguish the sample source, while the assimilator blurs the difference between x 1 and x 2 , and through the feedback of the differentiator for structural alignment, the difference gradually decreases. The Involution network realizes the fusion of the features of the sample set through the game training of the assimilator and the differentiator, generates new data blocks with richer features, and filters out abnormal features. The network generalizes the data blocks, thereby helping to train a model with good generalization ability.

[0032] The assimilator and the differentiator can be trained by optimizing the following objective function:

[0033]

[0034] In the formula, S is the assimilator, D is the differentiator, E is the expectation, P is the probability, and x is the input.

[0035] Furthermore, in step 5, the fusion loss function of cross-entropy and multi-kernel - maximum mean discrepancy is used to measure the difference of the feature instance graphs composed of different engine data, so as to realize the prediction of the exhaust temperature threshold of the engines with highly dispersed cluster features.

[0036] Cross-entropy adopts an inter-class competition mechanism and is good at learning information between classes. Cross-entropy can change the learning speed according to the model training results and improve the model learning efficiency. However, cross-entropy also has disadvantages. Cross-entropy focuses on the accuracy of the correct label prediction probability and ignores the differences of other incorrect labels. For the measurement of the differences of incorrect labels, the multi-kernel - maximum mean discrepancy loss function is used. The most important part of the maximum mean discrepancy is the kernel function. It is often difficult to select the optimal kernel function during training. The multi-kernel - maximum mean discrepancy proposes to use multiple kernels to construct a total kernel to solve this problem.

[0037] When setting the loss function of the engine mathematical model with high-dispersion cluster characteristics, first use the cross-entropy loss function to continuously narrow the distance between the actual output and the expected output to improve the accuracy of correct label prediction; subsequently, use the multi-kernel maximum mean discrepancy loss function to learn the characteristics of the cluster parameters in the high-dimensional space. Finally, use the fused loss value as the basis for updating the parameters during backpropagation.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] In the present invention, aiming at the problem that the dispersion of flight data of the same type of engine within a certain range brings difficulties to data analysis, a method for predicting the exhaust gas temperature threshold of an aero-engine based on an improved spectral graph convolutional neural network is proposed. Define and analyze the high-dispersion cluster characteristics of aero-engines, and emphasize the influence of the dispersion of gas path parameters of the same model engine on the exhaust gas temperature; regard the prediction of all exhaust gas temperature thresholds under variable working conditions as a classification problem, which can improve the prediction efficiency, the accuracy and easy evaluation of the prediction results; the improved spectral graph convolutional neural network can better mine the characteristics of engine data, make the samples more distinguishable, and thus be easier to classify; the Involution network fuses features through the game between the conformers and the differentiators, improving the generalization ability of the samples; using the fusion of cross-entropy and multi-kernel maximum mean discrepancy as the loss function of the model can not only learn the inter-class information, but also learn the characteristics of the cluster in the high-dimensional space, making up for the shortcoming of insufficient attention to wrong labels. This method can accurately predict the exhaust gas temperature threshold during the full cycle under variable working conditions. Brief Description of the Drawings

[0040] Figure 1 is the flowchart of predicting the exhaust gas temperature threshold of an aero-engine based on an improved spectral graph convolutional neural network provided by the present invention;

[0041] Figure 2 is the schematic diagram of working condition identification of the FD002 dataset provided by the present invention;

[0042] Figure 3 is the schematic diagram of the classification label setting of the exhaust gas temperature threshold prediction model of the engine provided by the present invention;

[0043] Figure 4 is the schematic diagram of feature fusion based on the Involution network provided by the present invention;

[0044] Figure 5 is the schematic diagram of the exhaust gas temperature threshold distribution of the FD001 cluster provided by the present invention;

[0045] Figure 6 is the schematic diagram of the training result (interval is 4) of FD001 provided by the present invention;

[0046] Figure 7 It is a schematic diagram of the FD001 training result (interval is 2) provided by the present invention;

[0047] Figure 8 It is a schematic diagram of the service time distribution of the test samples provided by the present invention;

[0048] Figure 9 It is a schematic diagram of the ΔEGT distribution of the test samples provided by the present invention;

[0049] Figure 10 It is a schematic diagram of the prediction result of the exhaust gas temperature threshold of the test samples provided by the present invention. Specific embodiments

[0050] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0051] To verify the effectiveness of the proposed method, this example uses the publicly available C-MAPSS simulation dataset of NASA for verification. The data used to train and verify the model includes multi-sensor data of 709 engines from factory to retirement in the time series. Among them, 200 are data under a single working condition, and the rest are data under variable working conditions. The ultimate goal of the model is to accurately predict the exhaust gas temperature threshold of the engines in the fleet after different service cycles.

[0052] Please refer to Figure 1 , which shows a flowchart of the prediction of the exhaust gas temperature threshold of an aeroengine based on an improved spectral graph convolutional neural network provided by the present invention, specifically including the following steps:

[0053] Step 1: The process of defining and analyzing the highly dispersed cluster characteristics of the aeroengine gas path parameters, including 3 operating parameters and 14 sensor parameters. Specifically, the operating parameters are altitude alt, flight Mach number Mach, and throttle valve angle TRA; the sensor parameters are total temperature at the outlet of the low-pressure compressor T24, total temperature at the outlet of the high-pressure compressor T30, total temperature at the outlet of the low-pressure turbine T50, total pressure at the outlet of the high-pressure compressor P30, physical speed of the fan Nf, core physical speed Nc, static pressure at the outlet of the high-pressure compressor Ps30, ratio of fuel flow to Ps30 phi, corrected fan speed NRf, corrected core speed NRc, bypass ratio BPR, flow enthalpy htBleed, high-pressure turbine coolant flow W31, and low-pressure turbine coolant flow W32.

[0054] The coefficient of variation, like the standard deviation and variance, can reflect the degree of data dispersion. It is defined as the ratio of the standard deviation to the mean:

[0055]

[0056] The coefficient of variation is a dimensionless index that can be used to compare two sets of data with different dimensions or different means. Through the coefficient of variation, the dispersion caused by factors such as manufacturing processes in aero-engines can be measured. The coefficients of variation of 14 sensor parameters are shown in Table 1.

[0057] Table 1 Coefficients of variation of sensor parameters

[0058]

[0059] When the engine fails or its performance degrades due to a fault, its operating state will be significantly different from the rated state at the time of factory shipment. The parameters of each gas path component measured by the sensors will change over time. The prerequisite for gas path analysis and performance trend tracking is to be able to obtain the health status of the engine from these changing parameters. In the research on engine health management, the relative data of the parameters changing during operation is more suitable for evaluating the degree of dispersion of engine performance, which is defined as:

[0060]

[0061] In the formula, Δx i is the difference between the value of a certain engine parameter at the end of its life and its value when it first entered service, is the mean value of all Δx i .

[0062] The c v,engine of the engine parameters is shown in Table 2. It can be seen that the maximum value of the c v,engine of these parameters is 1.33 and the minimum value is 0.26, indicating a relatively large degree of dispersion.

[0063] Table 2 c of sensor parameters v,engine

[0064]

[0065]

[0066] The main basis for gas path analysis and performance trend based on sensor parameters is the changes in parameter data over service time, and these change amounts are also the focus of the research. Compared with the change amount Δx, the degree of dispersion of the initial values of the parameters and the degree of dispersion of the subsequent time series data are very large; while for the parameters themselves, it can be seen from the coefficient of variation that their numerical distributions are relatively concentrated. Therefore, the engine sensor data exhibits high-dispersion cluster characteristics.

[0067] Step 2: Consider the prediction of the exhaust gas temperature threshold under off-design conditions as a classification problem. The specific steps are as follows:

[0068] Step 2.1: Take the difference ΔEGT between the current exhaust gas temperature of the engine and the exhaust gas temperature threshold as the classification label. To eliminate the influence of the dimension of features, perform min-max normalization. The maximum and minimum values of the two single-condition datasets FD001 and FD003 are obtained from the same parameters of the entire fleet. The FD002 and FD004 data contain six operating conditions and belong to the data under the off-design conditions of the engine. For these two datasets, perform operating condition normalization. Take FD002 as an example. Please refer to Figure 2 , and use the k-means clustering algorithm to identify the six operating conditions of the engine. Group the data under the same operating condition into one group, with a total of 6 groups. Perform min-max normalization on these 6 groups of time series data respectively, and then merge the 6 groups of processed data. To facilitate obtaining ΔEGT, the normalized exhaust gas temperature is inverse-normalized based on the maximum and minimum values of one of the operating conditions;

[0069] Step 2.2: Take ΔEGT as the axis, take points at equal intervals. Samples in the same interval correspond to the same label, starting from 0 and dividing until the maximum ΔEGT in the fleet, and naming them 0, 1, 2, …, n in sequence. Please refer to Figure 3 , which shows the label distribution with an interval size of 2. The size of the interval used for division will depend on the specific data situation, such as the degree of noise pollution of the data;

[0070] Step 2.3: Setting intervals of different sizes helps to improve the prediction accuracy and reliability. Initially set a larger interval value, with fewer labels, higher model classification accuracy, and more reliable prediction of ΔEGT. Subsequently, set a smaller interval value to further narrow the range of ΔEGT on the basis of ensuring the reliability of the results to improve the prediction accuracy.

[0071] Step 3: Propose an improved spectral graph convolutional network to extract features from the input data and construct a feature instance graph. The fully connected layer of the improved spectral graph convolutional neural network has 336 units. The Chebyshev polynomials are expanded to the 1st, 2nd, and 3rd orders respectively. To enhance the robustness of the model, after the improved spectral graph convolutional neural network, use a Dropout layer with a random packet loss rate of 0.5 for processing. The initial learning rate is 0.001. Set the weight decay to 0.0001 and dynamically adjust the learning rate.

[0072] Step 4: Propose a new Involution network to realize the fusion of the features of two samples during the training process by having the identifier and the differentiator play a game.

[0073] Please refer toFigure 4 The Involution network contains two adversarial models: the similarity seeker S and the dissimilarity seeker D. The similarity seeker performs structural alignment on two input samples, enabling them to learn from each other and fuse features, and outputs the sample x 1 and x 2 . The input of the dissimilarity seeker is x 1 and x 2 , and it outputs a probability value indicating the probability that the sample comes from x 2 . The larger the value, the higher the probability of coming from x 2 ; the smaller the value, the higher the probability of coming from x 1 . The dissimilarity seeker is used to distinguish the sample sources, while the similarity seeker blurs the differences between x 1 and x 2 , and through the feedback of the dissimilarity seeker for structural alignment, the differences gradually decrease. The Involution network realizes the fusion of the features of the sample set through the game training of the similarity seeker and the dissimilarity seeker. The features of the generated new data blocks are more abundant, and abnormal features are filtered out. The network generalizes the data blocks, thus helping to train a model with good generalization ability.

[0074] The similarity seeker and the dissimilarity seeker can be trained by optimizing the following objective function:

[0075]

[0076] In the formula, S is the similarity seeker, D is the dissimilarity seeker, E is the expectation, P is the probability, and x is the input.

[0077] Step 5: Use the fusion loss function of cross-entropy and multi-kernel - maximum mean discrepancy to measure the differences in the feature instance graphs composed of different engine data, so as to realize the prediction of the exhaust temperature threshold of engines with highly dispersed cluster features.

[0078] Take the FD001 dataset as an example. Refer to Figure 5 , which shows the distribution of the exhaust temperature thresholds of the engines. After rounding and pooling all ΔEGT values, the minimum value is 0 and the maximum value is 32 (the unit is °R, and the numerical unit of ΔEGT involved below is all °R). Taking ΔEGT as the axis, points are taken at equal intervals on the axis, and all samples between two points correspond to the same label. Starting from 0 on the axis, it is equally divided into intervals until the largest ΔEGT in the entire fleet. Each interval is named 0, 1, 2,..., n in turn and used as the classification label.

[0079] When setting the interval size, the fluctuations in the exhaust temperature data collected by the sensor due to noise and other reasons should be taken into account. This fluctuation is approximately between 2 and 4. Therefore, the interval sizes are set to 4 and 2 respectively. Refer to Figure 6 and Figure 7, which respectively shows the training results for intervals of 4 and 2. The curve in the figure is the accuracy of the training set, and the horizontal line is the accuracy of the validation set. The model accuracy is based on the accuracy of the validation set. When the interval is set to 4, the classification accuracy of the FD001 dataset is as high as 97.98%. When the interval is set to 2, the accuracy of the dataset drops slightly to 95.55%.

[0080] The predicted value of the exhaust temperature threshold of an engine in service in the FD001 test set can be obtained based on the classification results of the model. The predicted results of some selected test samples are shown. The selected test samples should be representative and can fully prove that the exhaust temperature thresholds of the engines in the fleet can be accurately predicted after different service cycles. Please refer to Figure 8 , which shows the distribution of the service time and remaining useful life of the test samples. It can be seen that the life distribution of the selected test samples is relatively wide and covers various periods such as the early-middle, middle, and late stages of the engine's service life. Please refer to Figure 9 , which shows the ΔEGT distribution of the test samples, covering almost all values within the ΔEGT range. Please refer to Figure 10 , which shows the predicted results of the exhaust temperature thresholds of the test samples. The positive error is the error by which the predicted value of the exhaust temperature threshold is more than the actual value, and the negative error is the error by which the predicted value of the exhaust temperature threshold is less than the actual value. The root mean square error of the predicted values of all samples in the figure is 0.60, which can reasonably and accurately predict the exhaust temperature thresholds of the samples.

[0081] To prove the superiority of the method proposed in the present invention, the FD001 dataset of the CMAPSS dataset is used as the experimental object, and it is compared with existing common deep learning methods such as long short-term memory neural network (LSTM), deep convolutional neural network (DCNN), convolutional neural-long short-term memory neural network (CNN-LSTM), gated convolutional unit enhanced transformer encoder (GCU-Transformer), and deep reinforcement learning (DRL). It is further illustrated that this model can effectively predict the remaining useful life and improve the prediction accuracy. The average index results of the comparison are shown in Table 3.

[0082] Table 3 Method Comparison

[0083] Method Root Mean Square Error LSTM 0.94 DCNN 0.91 CNN-LSTM 0.78 GCU-Transformer 0.82 DRL 0.86 The present invention 0.60

[0084] The results of the remaining three datasets are as follows. When the interval is set to 4, the classification accuracy of the single-condition dataset FD003 is 97.07%. FD002 and FD004 are multi-condition datasets, and the situation is more complex. The accuracy will be lower than that of the single-condition data, and the difference between the validation set accuracy and the test set accuracy will also be a little larger than that of the single-condition data. The classification accuracies of FD002 and FD004 are 95.22% and 94.55% respectively. When the interval is set to 2, the accuracies of the three datasets decrease a little, to 93.20%, 96.15% and 92.84% respectively. When predicting the exhaust gas temperature threshold, the interval can be gradually set from large to small, which not only ensures the prediction accuracy but also gradually narrows down the range of the exhaust gas temperature threshold.

[0085] The present invention is not limited to the above embodiments. Based on the technical solutions disclosed in the present invention, those skilled in the art can make some simple modifications, equivalent changes and modifications to some of the technical features without creative labor, and all of them fall within the scope of the technical solutions of the present invention.

Claims

1. A method for predicting the exhaust temperature threshold of an aero-engine based on an improved spectral convolutional neural network, characterized in that: The specific steps include: Step 1: Define and analyze the high-dispersion cluster characteristics of aircraft engine gas path parameters; Step 2: The prediction of the exhaust temperature threshold under variable conditions is regarded as a classification problem; Step 3: Propose an improved spectral graph convolutional network to extract features from input data and construct a feature instance graph; Step 4: Propose a new network Involution network to learn and fuse features. By letting the similarity seeker and the difference seeker compete with each other during the training process, the feature fusion of the two samples is achieved. Step 5: Use the fusion loss function of cross entropy and multi-core-maximum mean difference to measure the difference in feature instance graphs composed of different engine data to achieve the exhaust temperature threshold prediction of engines with high dispersion cluster features.

2. The method for predicting the exhaust temperature threshold of an aircraft engine based on an improved spectral graph convolutional neural network according to claim 1 is characterized in that: The step 1 explains why the gas path parameters of the same type of engine have both high dispersion and clustering. In statistics, the coefficient of dispersion is a statistic used to characterize the degree of dispersion of a probability distribution, defined as the standard deviation σ and the mean value. The ratio: The discrete coefficient can be used to measure the clustering of gas path parameters of engines of the same model, thereby reflecting the manufacturing process level of the engine. Considering that the evaluation of engine health is based on the change of parameters of various gas path components over time, the relative data of parameter changes during engine operation should be used to evaluate the discreteness of the parameters rather than the absolute values ​​of the parameters: In the formula, Δx i It is the difference between the value of a parameter of the engine at the end of its service life and the value when it is just put into service. For all Δx i The average value of .

3. The method for predicting the exhaust temperature threshold of an aircraft engine based on an improved spectral graph convolutional neural network according to claim 1, characterized in that: The step 2 treats the prediction of the exhaust temperature threshold under variable operating conditions as a classification problem to improve the efficiency, accuracy and ease of evaluation of the method.

4. The method for predicting the exhaust temperature threshold of an aircraft engine based on an improved spectral graph convolutional neural network according to claim 1, characterized in that: The step 3 proposes to improve the spectral convolutional network to extract features from the input data and construct a feature instance graph. The improved spectral convolutional neural network introduces multiple filters of different sizes, learns features of different scales, and fuses them together. The formula is: In the formula, h is the weight, g is θ is the filter and X is the input signal.

5. The method for predicting the exhaust temperature threshold of an aircraft engine based on an improved spectral graph convolutional neural network according to claim 1, characterized in that: Step 4 proposes a new network Involution network to learn and fuse features, by letting the similarity seeker and the difference seeker play a game and compete with each other during the training process to achieve feature fusion of the two samples. The similarity seeker and the difference seeker can be trained by optimizing the following objective function: In the formula, S is the similarity finder, D is the difference finder, E is the expectation, P is the probability, and x is the input.

6. The method for predicting the exhaust temperature threshold of an aircraft engine based on an improved spectral graph convolutional neural network according to claim 1, characterized in that: Step 5 uses the fusion loss function of cross entropy and multi-core-maximum mean difference to measure the difference in feature instance graphs composed of different engine data. The cross entropy loss function is used to continuously shorten the distance between the actual output and the expected output, thereby improving the accuracy of the correct label prediction probability; and the multi-core-maximum mean difference loss function is then used to learn the characteristics of the cluster parameters in high-dimensional space. The final fusion loss value is used as the basis for updating parameters during back propagation to achieve the exhaust temperature threshold prediction of high-dispersion cluster feature engines.

Images