A method for reconstructing aero-engine performance models based on DLBR

Abstract

This invention relates to a method for reconstructing aero-engine performance models based on DLBR, specifically focusing on the reconstruction of turbofan engine airflow characteristic parameters. The invention addresses the problem that aero-engines rely on a limited number of sensors to collect characteristic parameters, leading to incomplete parameter sets and ultimately, the inability to build a complete aero-engine performance model. The invention inputs the complete aero-engine airflow characteristic parameters X into a trained autoencoder to perform dimensionality reduction and reconstruction. The reconstructed result is then used as input, with measurable airflow characteristic parameters S as input and the autoencoder's feature space h as output, to construct an LSTM-based mapping model between S and h. Finally, the mapping model is linked to the autoencoder to obtain an aero-engine performance reconstruction model that uses the measurable airflow characteristic parameters S as input and the reconstruction result as output.

Description

Technical Field

[0001] This invention belongs to the technical field of reconstructing the gas path characteristic parameters of turbofan engines. Background Technology

[0002] The aircraft engine is the heart of an aircraft, with a complex structure and harsh operating conditions. During use, the health status of an aircraft engine gradually changes. When its function or performance deteriorates to a certain extent, the engine needs to be disassembled and sent for overhaul to restore its function or performance to a certain degree. To improve engine maintainability, aircraft engines now generally adopt a modular design principle. With the introduction and implementation of the modular design principle in the civil aviation industry, Engine Health Management (EHM) needs to analyze the engine's health status at the modular level to formulate reasonable maintenance plans. The modular maintenance management model significantly saves maintenance time and reduces maintenance costs. Therefore, accurate assessment of modular performance is crucial for reasonably determining the performance recovery status of a modular unit / engine as a whole and for determining the unit's repair level. Currently, there are two methods for modular performance assessment: the first is the modular efficiency calculation method given in the OEM (Original Equipment Manufacturer) engine manual. This method calculates the efficiency of the airflow modular unit based on a highly abstract and empirically derived design model to determine whether the unit has deteriorated. However, overly abstract and simplified models are insufficient to express the degree of performance degradation of unit components under complex service conditions and can only serve as a reference. The second approach is a data-driven method. OEMs collect flight and degradation data from various operating companies, and through big data analysis, provide a mapping space (unit characteristic space) between unit component performance and overall aircraft aerodynamic parameters. Based on data returned by operators, they establish reference unit characteristic models, some of which are shown below. Figure 1 As shown, a degraded low-pressure turbine leads to a decrease in the rotational speed of the low-pressure compressor, thereby reducing the outlet pressure P of the low-pressure compressor. 25 and temperature T 25 Degraded fans, high-pressure compressors, and high-pressure turbines can all cause the low-pressure compressor to operate at higher power to maintain a stable engine pressure ratio, thereby increasing the low-pressure compressor outlet pressure and temperature. However, the OEM also points out that due to differences in engine configuration and operating conditions, the unit characteristic model provided by the OEM is only for operator reference. The actual unit characteristic model needs to be established based on actual fleet data to truly reflect the performance status of the engine unit.

[0003] However, all the aerodynamic characteristic parameters required to build a unit characteristic model are often only available when the engine is sent back to the OEM for testing. During the aircraft's service life or maintenance outside the OEM, it is impossible to obtain all the characteristic parameters needed to construct the unit characteristic model. How to reconstruct complete engine characteristic data based on partially available characteristic parameter data, supporting the establishment of unit characteristic models to achieve accurate unit performance evaluation, and thus providing operational data support for airlines and maintenance companies, is a pressing scientific problem that needs to be studied. Summary of the Invention

[0004] This invention aims to address the problem that only a limited number of sensors on aero engines can collect characteristic parameters, resulting in incomplete characteristic parameters required for constructing aero engine performance models and the inability to build models. It provides an aero engine performance model reconstruction method based on DLBR.

[0005] The DLBR-based method for reconstructing aero-engine performance models includes the following steps:

[0006] Step 1: Input the complete aero-engine gas path characteristic parameters X into the trained autoencoder to perform dimensionality reduction and reconstruction of X, and obtain the reconstruction result.

[0007] Step 2: Using the measurable gas path characteristic parameter S as input and the feature space h of the autoencoder as output, construct a mapping model between S and h based on LSTM.

[0008] Step 3: Establish a connection between the mapping model constructed in Step 2 and the autoencoder in Step 1 to obtain the model with the measurable gas path characteristic parameter S as input and the reconstruction result as output. The output is the aero-engine performance reconstruction model.

[0009] Furthermore, the specific process of dimensionality reduction and reconstruction of X in step one above includes:

[0010] Encoding process:

[0011] The complete air path characteristic parameters X of the aero-engine are encoded by the encoding function f. encoder (·) Mapped to the feature space h:

[0012] h = f encoder (X) = g(W*X + b),

[0013] Where g(·) is the encoding activation function, W is the encoding weight matrix, and b is the encoding bias vector;

[0014] Decoding process:

[0015] The feature space h is passed through a nonlinear mapping function f. decoder(·) Reconstruction yields the reconstruction result.

[0016]

[0017] Where W′ is the decoding weight matrix and b′ is the decoding bias vector.

[0018] Furthermore, the above-mentioned encoding activation function g(·) is a ReLU function.

[0019] Furthermore, during the training of the autoencoder, the autoencoder descent algorithm is used to backpropagate the error to adjust the autoencoder parameters {θ,θ′}={W,b;W′,b′}, so that the reconstruction result... Mean square error between the complete gas path characteristic parameter X of the aero-engine Minimum.

[0020] Furthermore, the above reconstruction results Mean square error between the complete gas path characteristic parameter X of the aero-engine The expression is:

[0021]

[0022] This invention proposes a DLBR-based method for reconstructing aero-engine performance models. The fundamental idea is that the nonlinear system constructed from measurable engine airflow characteristic parameters should be consistent with the nonlinear system constructed from complete engine airflow characteristic parameters to achieve system reconstruction. Since engine airflow parameters are time-series data, this invention establishes a CNN-LSTM autoencoder network to learn the feature space of the engine airflow system. Because the autoencoder network can customize the number of spatial dimensions, the number of spatial dimensions can be adjusted according to the dimensional requirements of the airflow characteristic parameters, thereby achieving the requirement of consistency between the two nonlinear spaces of the engine.

[0023] The problem of reconstructing the complete engine airflow characteristic parameter space from the measurable airflow characteristic parameter space is essentially a process of reconstructing a high-dimensional space from low-dimensional data. Analysis of the autoencoder principle reveals that the decoding part of the autoencoder is a mapping process from low-dimensional to high-dimensional space. However, the feature space obtained by the autoencoder is the embedding space of the original data space. To ensure that the embedding expression of the complete engine airflow characteristic parameter space is consistent with the embedding expression of the measurable engine airflow characteristic parameter space, a mapping model between the two needs to be established. This allows for the indirect reconstruction of the complete engine airflow characteristic parameter space from the measurable airflow characteristic parameter space. Figure 3The present invention uses the feature space of the complete air path characteristic parameters of the engine (feature space 2 in the figure) as the learning target, and establishes a mapping model (model 4 in the figure) between feature space 2 and the space of measurable air path characteristic parameters of the engine, thereby indirectly mapping the space of measurable air path characteristic parameters of the engine to the feature space 2 of the complete air path characteristic parameters of the engine.

[0024] Since mapping from mapping model 4 to feature space 2 inevitably introduces errors, these errors, after being reconstructed (by decoder 3), will result in a significant discrepancy between the final air path space and the original complete engine air path space. Analysis reveals that the key to the entire reconstruction process is not ensuring that mapping model 4 accurately maps to feature space 2, but rather the ability to accurately reconstruct the original space using feature space 2 as an intermediate medium within the engine's measurable air path characteristic parameter space. Therefore, after the autoencoder completes the feature space learning of the engine's complete air path characteristic parameters, a transfer learning method is used, employing the learned parameters from decoder 3 as initial learning parameters. A relearning process from mapping model 2 and decoder 3 is designed to further improve the mapping accuracy from the engine's measurable air path characteristic parameters to the engine's complete air path characteristic parameters. Attached Figure Description

[0025] Figure 1 This is a characteristic model diagram on the characteristic space of a unit cell;

[0026] Figure 2 A schematic diagram showing the basic configuration of a civil aviation engine and the location of its sensors.

[0027] Figure 3 A schematic diagram for reconstructing complete engine air path characteristic parameters based on measurable air path characteristic parameters;

[0028] Figure 4 Reconstruct the model diagram for DLBR;

[0029] Figure 5 This is a schematic diagram of the self-encoder principle;

[0030] Figure 6 Here is a diagram of the LSTM network structure;

[0031] Figure 7 This is a schematic diagram of a multi-component closed-loop simulation system.

[0032] Figure 8 Divide the simulation system parameters into Gantt charts;

[0033] Figure 9 A bar chart showing the experimental results based on simulation data;

[0034] Figure 10 Create Gantt charts for engine parameters;

[0035] Figure 11A bar chart showing the experimental results of the mean MSE based on engine data;

[0036] Figure 12 A bar chart showing the experimental results of MSE variance based on engine data;

[0037] Figure 13 This is a reconstruction error map based on the Scheme3 method. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0039] The parameters required to establish a unit characteristic model are often only available when the engine is sent back to the OEM for testing. During normal flight or maintenance outside the OEM, it is impossible to obtain all the parameters needed to construct a unit characteristic model. Therefore, retrieving the physical parameters of engine units using useful information from measurable data is of great significance for unit characteristic modeling. The turbofan aero-engine studied in this invention mainly consists of six main units: fan, low-pressure compressor (LPC), high-pressure compressor (HPC), combustor (CC), low-pressure turbine (LPT), and high-pressure turbine (HPT). Figure 2 As shown. Figure 2The document also provides a general component composition and a layout scheme for measuring the status parameters of an aero-engine. The engine employs a dual-rotor structure, and its in-flight testing system can measure the airflow conditions at multiple air passage locations. The operating parameters and external system inputs include: inlet total temperature T2, inlet total pressure P2, variable stator vane (VSV) opening AVSV, variable bleed valve (VBV) opening AVBV, low-pressure rotor speed N1, LPC outlet total temperature T25, LPC outlet total pressure P25, combustion chamber total temperature T3, combustion chamber static pressure Ps3, fuel flow rate WFF, engine exhaust temperature TEGT, low-pressure turbine outlet total temperature T5, high-pressure rotor speed N2, and outer bypass duct inlet total pressure P13. Figure 2 In addition to the 14 sensors mentioned above, the system also includes 13 other parameters that are not normally measurable. The parameters measured by these 27 sensors constitute the complete airflow characteristic parameters of the aero-engine. Considering factors such as cost, safety, and maintainability, sensors such as T25, P25, T5, Ps3, and T3 are only installed on a few engines. Most other fleet engines lack these sensors to varying degrees due to cost, safety, maintainability, or other factors. Therefore, the parameters measured by the 14 measurable sensors constitute the measurable airflow characteristic parameters.

[0040] During engine operation, its state parameters and operating condition parameters exhibit strong coupling. The measured parameters lack explicit correlation, making direct solutions difficult. To address these issues, this implementation proposes a physical field reconstruction model based on an autoencoder. The task of the DLBR model is to reconstruct the physical field based on a small amount of measurement data S = [s1, s2, ..., s...]. N To the greatest extent possible, accurately and comprehensively retrieve the global physical parameter information X = [x1, x2, ..., x...] M ],like Figure 4 As shown. The specific implementation method is as follows:

[0041] The DLBR-based aero-engine performance model reconstruction method described in this embodiment includes the following steps:

[0042] Step 1: As a neural network used in unsupervised learning, the input and expected output of an autoencoder are both unlabeled samples, while the output in the feature space is an abstract representation of the sample's features. The training process of an autoencoder includes two stages: encoding and decoding. Figure 5 As shown, the autoencoder first receives the input data, transforms it into an efficient abstract representation, and then outputs a reconstructed version of the original sample. The process from the input layer to the intermediate feature space is the encoding process, while the process from the intermediate feature space to the output layer is the decoding process. If X∈R M Let h ∈ R be the input vector.r Represents the eigenvector. This represents the output layer vector. Where r < M. The specific process is as follows:

[0043] Encoding process:

[0044] The complete air path characteristic parameters X of the aero-engine are encoded by the encoding function f. encoder (·) Mapped to the feature space h:

[0045] h = f encoder (X) = g(W*X + b),

[0046] Where g(·) is the encoding activation function, specifically the ReLU function. W is the encoding weight matrix, and b is the encoding bias vector.

[0047] Decoding process:

[0048] The feature space h is passed through a nonlinear mapping function f. decoder (·) Reconstruction yields the reconstruction result.

[0049]

[0050] Where W′ is the decoding weight matrix and b′ is the decoding bias vector. Same as the X dimension.

[0051] During the training of the autoencoder, the autoencoder descent algorithm is used to backpropagate the error to adjust the autoencoder parameters {θ,θ′}={W,b;W′,b′}, so that the reconstruction result... Mean square error between the complete gas path characteristic parameter X of the aero-engine Minimum. Reconstruction result Mean square error between the complete gas path characteristic parameter X of the aero-engine The expression is:

[0052] In actual training, the mini-batch gradient method is usually used to find the optimal parameters. Specifically, the training set is divided into several mini-batches, which is the number of samples selected before each parameter adjustment. Each time, a mini-batch is input into the model, and the mean squared error of the entire mini-batch is used as the loss function.

[0053] In this step, the complete airflow characteristic parameters X of the engine are used as the model input. An autoencoder model is trained using Mean Square Error (MSE) as the loss function, and Adam is used as the optimizer to extract the feature space h. Through the encoding and decoding process, dimensionality reduction and reconstruction of the input data are achieved.

[0054]

[0055] Among them, f encoder and f decoder They represent Figure 5 The encoder and decoder in the model. Furthermore, in the autoencoder, the feature space h is a highly compressed low-dimensional space, which to some extent achieves dimensionality reduction and denoising of high-dimensional physical fields.

[0056] Step 2: Considering that the engine dataset is typical time series data, the Long Short-Term Memory (LSTM) learning model performs well in processing time series data. Therefore, we consider using LSTM as the mapping model.

[0057] In a standard recurrent neural network (RNN), the hidden layer state at each time step is determined simultaneously by the current input and the states of the previous hidden layers. This results in limited memory capacity and a susceptibility to the vanishing gradient problem. LSTM networks improve upon RNNs by incorporating forget gates, input gates, and output gates. Figure 6 As shown.

[0058] Figure 6 In this context, σ represents the sigmoid activation function, with an output range of 0 to 1, and tanh is a function used to adjust the numerical value, with an output range of -1 to 1. Based on the feature space h obtained in step one, a mapping model is established with the measurable gas path characteristic parameter S as input and the feature space h as output:

[0059]

[0060] Among them, f reg This represents the mapping model, i.e., the LSTM learning model.

[0061] Step 3: X passes through encoder f encoder The obtained feature space h and the measurable gas path characteristic parameters S are mapped through the model f reg The obtained feature space h contains errors, which are amplified when reconstructed using the decoder trained in step one. Therefore, to eliminate this error, the mapping model constructed in step two is linked to the autoencoder in step one to obtain a model that takes the measurable gas path characteristic parameter S as input and the reconstruction result as input. The output is the aero-engine performance reconstruction model:

[0062]

[0063] Where, f′ reg It is in f reg Based on this, perform refined training, f′decoder It is in f decoder Based on this, fine-grained training is performed. Because f encoder and f reg Since training had been conducted before, the learning rate was reduced during the third training session.

[0064] Considering the structural variability of multi-component closed-loop systems used in industrial applications, testing the adaptability of DLBR-based reconstruction models across all system types is extremely difficult. To ensure the data experiments do not lack generality, this implementation method constructs a representative multi-component closed-loop simulation system, the structure of which is as follows: Figure 7 As shown. In a multi-component closed-loop simulation system, the local models of the four components are interconnected by information interspersed within them, forming a system closed loop. Observing these information connections, it can be found that the system contains a parallel structure, two series structures, a connection junction, and a loop. This loop contains a set of equality constraints E, specifying μ′1, The numerical correlation between μ′4 provides a feedback loop to make the closed-loop system controllable. This simulation system includes all the basic types of correlations between variables and contains all the elements found in industrial system modeling problems, such as control variables, unmeasurable variables, and measurable variables. Therefore, it can mathematically represent most multi-component closed-loop systems. The transfer functions of Ω2, Ω3, and Ω4 are defined in this implementation as follows:

[0065]

[0066] In the formula, ε i 1% Gaussian noise; c i This is an adjustable coefficient.

[0067] Correspondingly, unlike Ω2, Ω3, and Ω4, the transfer function Ω1 is implicitly defined by the constraint set E:

[0068]

[0069] The sampling data for each simulation system is generated as follows: First, 500 sample points are randomly selected in the vector space [μ′1,μ′4] and E is solved to obtain... and The value, and then and The inputs are fed into Ω2 and Ω3 respectively to obtain y2 and y3. Finally, μ′4, y2, and y3 are output into Ω4 to obtain y4, and then the result is hidden. And y2 are the unmeasurable parameters in the simulation system. The parameter partitioning of the simulation system is as follows: Figure 8 As shown.

[0070] In this experiment, to ensure the generality of the dataset, 10 sets of adjustable coefficients were randomly selected (as shown in Table 1), and 10 simulation systems were generated using these 10 sets of adjustable coefficients. For each simulation system, 500 working point data points labeled with the system input vector [μ′1,μ′4] were selected from its workspace as the initial dataset, and then the other parameter values ​​were obtained by solving. The simulation dataset is shown in Table 2, and the complete dataset can be obtained from the Mendeley database.

[0071] Table 1. Ten adjustable coefficients of the simulation system

[0072]

[0073]

[0074] Table 2 Simulation Dataset

[0075]

[0076] Considering that the engine dataset is typical time-series data, and that both One-dimensional Convolutional Neural Network (1DCNN) and Long Short-Term Memory (LSTM) learning models perform well in processing time-series data, although the simulation dataset is not time-series data, 1DCNN and LSTM also have good nonlinear feature extraction capabilities. Therefore, this implementation chooses these two algorithms as autoencoders or mapping models. For the DLBR model, this implementation combines the two learning models to design three different gas path characteristic parameter reconstruction models, named Scheme1, Scheme2, and Scheme3, as shown in Table 3. Since algorithms such as 1DCNN, RBFNN, SVR, BPNN, and LSTM have good regression capabilities, and in order to compare the reconstruction results with the above reconstruction models, 1DCNN, RBFNN, SVR, BPNN, and LSTM are used as comparative experiments. That is, μ′1, μ′4, y3, and y4 are input into the above five algorithms, and the output is... The value of y2.

[0077] Table 3 Reconstruction Models of Three Gas Path Characteristic Parameters

[0078]

[0079] Each method was run 10 times on randomly partitioned training and test datasets. The ratio of training to test set data size was 3:2. All algorithms were run a total of 100 times on 10 simulation models. The mean and variance of the MSE obtained from these 100 reconstructions were recorded as performance metrics for each algorithm. The performance metrics for each method in the reconstruction experiments for each parameter are given in Table 4. The bold text in the table indicates the optimal performance metric among all methods. The experimental results are shown in Table 4, and the corresponding bar chart is shown below. Figure 9 As shown.

[0080] Table 4 Simulation Data and Experimental Results

[0081]

[0082]

[0083] As shown in Table 4, BPNN, LSTM, RBFNN, Scheme1, Scheme2, and Scheme3 all exhibit high reconstruction accuracy for all parameters, with almost all MSE values ​​below 10. -3 Of these six algorithms, Scheme3 is the best. and The reconstruction accuracy is the highest, and in and The robustness demonstrated by Scheme3 in the reconstruction experiments is better than other algorithms. For the reconstruction parameter y2, LSTM achieves the highest reconstruction accuracy and exhibits better robustness than other algorithms. Analysis shows this is because the output gate of LSTM is obtained by multiplying the sigmoid function and the tanh function. The sigmoid function contains an exponential function, consistent with the transfer function Ω2, which improves the reconstruction accuracy of parameter y2. Table 4 shows the running time, including training and testing times. Scheme3 takes the longest, requiring 64.7 seconds, which is still within an acceptable range. The testing time from inputting the test set to outputting the reconstruction result is only 0.22 seconds. Based on the analysis in Table 4 and the above explanation, in terms of reconstruction accuracy and robustness, Scheme3 outperforms other algorithms.

[0084] To further verify the adaptability of the proposed model to unknown nonlinear black-box systems, an engine service dataset will be used for further testing. Engine data experimental verification:

[0085] This implementation uses the gas path characteristic data of a well-measured engine in the fleet, covering 500 cycles. Some data are shown below; the complete sensor dataset can be obtained from the Mendeley database, as shown in Table 5. The training process was executed 10 times on randomly partitioned training and test sets, with a training set to test set size ratio of 3:2. The mean, variance, and runtime of the obtained MSEs were used to evaluate the algorithm's performance. The experimental results based on the engine data are shown in Table 6, and the corresponding bar chart is shown below. Figure 11 and Figure 12 As shown. The division between measurable engine parameters and reconstructable parameters is as follows. Figure 10 As shown.

[0086] Table 5. Partial sensor data for an engine with complete measurement conditions.

[0087]

[0088] Table 6 Engine Data Experiment Results

[0089]

[0090]

[0091] As shown in Table 6, Scheme3 provides the lowest reconstruction accuracy among these eight methods for reconstruction parameters P25, T25, PS3, and T3, with an average MSE of approximately 10. -4 While its accuracy is on the same order of magnitude as other algorithms, it represents the best possible result. In particular, T25's mean MSE is 10. -5 The accuracy is on the order of magnitude higher than other algorithms, whose best-in-class accuracy is on the order of 10⁻⁴. Regarding MSE variance, Scheme3 demonstrates significantly better robustness than other algorithms in the P25, T25, and T3 reconstruction experiments, with an MSE variance of 10⁻⁴. -8 The order of magnitude, while other algorithms are all on the order of 10. -7 Order of magnitude. The MSE variance obtained by reconstructing the Scheme3 parameter PS3 is 2.93 × 10⁻⁶. -7Although the results obtained are not optimal, they are on the same order of magnitude as the optimal results and are very close. When reconstructing parameter T5, the mean and variance of the MSE obtained by Scheme3 are on the same order of magnitude as the optimal results. Since the engine air path system is not a nonlinear system composed of basic log, Exp, and other basis functions like the simulation system, but a black-box complex nonlinear system, the reconstruction accuracy of LSTM does not have an advantage over other algorithms. By observing Tables 4 and 5, it is not difficult to analyze that compared with the simulation data samples, the parameters that Scheme3 performs best in the reconstruction results of engine data samples are not only more numerous, but also have an order-of-magnitude improvement in accuracy. The accuracy of individual parameters that are not optimal is on the same order of magnitude as the optimal ones and is very close. By analyzing the running time of Tables 4 and 5, it can be seen that the engine data samples require longer time than the simulation data samples. This is because, with the same number of input samples, the larger the dimension of the sample, the greater the computation time required. In the table, the maximum time required by Scheme3 is only 80.1s, which includes training and testing time and is still within an acceptable range. Based on the above analysis and explanation, Scheme3 has advantages in both accuracy and robustness compared to other algorithms. Scheme3 also demonstrates a superior ability to reconstruct engine gas path characteristic parameters.

[0092] To further quantify and evaluate the reconstruction results obtained from Scheme3, the first 300 cycles were used as the training set, and the last 200 cycles were used as the test set. The reconstruction results of the engine airflow characteristic parameters are shown in Tables 6 and 6, respectively. Figure 13 As shown.

[0093] Table 6. Relative Errors in Engine Gas Path Characteristic Parameter Reconstruction

[0094]

[0095] Figure 13 In this context, R represents the correlation coefficient, indicating the degree of correlation between the reconstructed value and the original value. The closer R is to 1, the stronger the positive correlation between the two.

[0096] Table 6 shows that the average relative error of parameters P25, T25, PS3, and T3 is less than 1%, and the maximum relative error is less than 4%, indicating that using Scheme3 as the base model for the DLBR deep network model yields high-precision reconstruction results. However, the reconstruction result for T5 shows a significant decrease in accuracy compared to the other four parameters. Observing the reconstruction results in the table, it can be found that the reconstruction error of other regression algorithms for T5 is also much higher than that for the other four parameters. The above phenomena fully illustrate that T5 itself has the fundamental property of being difficult to reconstruct, while the average relative error obtained by the method in this paper is close to 1%, which is still within an acceptable range. A credible explanation for the difficulty in reconstructing T5 is that the outlet interface of the low-pressure turbine of the aero-engine is relatively wide and the airflow is relatively turbulent, making it difficult for the sensor readings of T5 to represent the average reading of the overall airflow cross section. This also makes the sensor measurement results of T5 fluctuate greatly.

[0097] This implementation analyzes sensors that measure airflow states at multiple locations in an engine's airflow system, classifying measurable parameters and parameters to be reconstructed to determine the parameters requiring reconstruction. Validation was performed using simulation data and engine service datasets. Results show that compared to regression models BPNN, SVR, LSTM, 1DCNN, RBFNN, and Scheme3, this implementation demonstrates better reconstruction accuracy and robustness. It also illustrates that the novel learning model built in this implementation, based on a learning model as the foundation and employing deep learning as the technical model, the DLBR (deep learning-based reconstruction) network model, can establish a mapping relationship between sensor measurement data and high-dimensional physical field data, possessing the capability to reconstruct high-dimensional physical fields.

[0098] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for aeroengine performance model reconstruction based on DLBR, characterized in that, The method comprises the following steps: Step 1: Obtain complete airflow characteristic parameters for the aero-engine Input into the trained autoencoder Perform dimensionality reduction and reconstruction to obtain the reconstruction result. , ， wherein and respectively represent an encoder and a decoder; Step two: with measurable gas path characteristic parameters As input, the feature space of the autoencoder As output, a mapping model based on LSTM is constructed Between and ， wherein, denotes a mapping model, i.e. an LSTM learning model; Step three: link the mapping model built in step two with the autoencoder in step one to obtain an aero-engine performance reconstruction model taking measurable aero-path characteristic parameters as input, and the reconstruction result as output ， wherein is fine-tuned on top of is fine-tuned on top of is fine-tuned on top of is fine-tuned on top of In step one, The specific process of dimensionality reduction and reconstruction includes: Encoding process: Aeroengine complete gas path characteristic parameters Through an encoding function Mapping to feature space : ， wherein, is an encoding activation function, is an encoding weight matrix, is an encoding bias vector; Decoding process: characteristic space through a non-linear mapping function reconstruction to obtain a reconstruction result : ， wherein, is a decoded weight matrix, is a decoded bias vector.

2. The DLBR-based aeroengine performance model reconstruction method of claim 1, wherein, Encoding activation function is a ReLU function.

3. The DLBR-based aeroengine performance model reconstruction method of claim 1, wherein, In training the autoencoder, an autoencoder descent algorithm is used to backpropagate errors to adjust the autoencoder parameters such that the mean square error between the reconstruction result and the complete gas path characteristic parameters of the aero-engine is minimized .

4. The DLBR-based aeroengine performance model reconstruction method of claim 3, wherein, reconstruction result between the aeroengine complete gas path characteristic parameters the mean square error is 。

Images