Engine flight thrust estimation method fusing adversarial autoencoding and sparse attention

Abstract

The application provides an engine flight thrust estimation method fusing adversarial auto-encoding and sparse attention, and belongs to the cross field of aerospace engineering and artificial intelligence. The method comprises the following steps: collecting ground test data and flight state data of a turbofan aeroengine, and pre-processing to obtain steady-state sensor data; inputting the steady-state sensor data and corresponding working condition labels into a trained adversarial auto-encoder for feature extraction to obtain a latent feature vector; step S3: inputting the latent feature vector into a sparse attention module to obtain fusion features; inputting the fusion features into a multi-scale time series modeling module, performing time series feature extraction through the multi-scale time series modeling module, and obtaining multi-scale fusion features; inputting the multi-scale fusion features into a trained thrust regression module to obtain an aeroengine flight thrust estimation value. The application can realize accurate estimation of the thrust in the flight state, and has significant practical engineering application value and deployment prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of aerospace engineering and artificial intelligence, and particularly relates to an engine flight thrust estimation method fusing adversarial auto-encoding and sparse attention. BACKGROUND

[0002] As the power core of aircraft, the health state evaluation of aero-engine has always been a research hotspot in academia and industry. Thrust, as a key indicator of engine performance, especially in the field of military engines, has important reference value in health state evaluation. However, engine thrust cannot be directly measured during flight, and usually relies on ground or high-altitude test beds, which requires the cooperation of special equipment and professional personnel, resulting in high testing costs and long testing cycles. Therefore, conducting thrust estimation based on on-board monitoring data not only helps to achieve rapid and low-cost engine health evaluation, but also provides important support for on-condition maintenance.

[0003] In recent years, data-driven methods have been widely concerned in engine thrust estimation. For example, Song Hanqiang et al. built an extreme learning machine model based on ground test bed data, taking the monitorable parameters as input and the ground measured thrust as label, and achieved a relative error of ±3% in the steady-state multi-speed operating condition. Wang Zhao et al. proposed a solid rocket ramjet engine thrust estimation method based on PSO optimized BP neural network, and verified that the SNPSO optimization method can control the relative error of thrust within 5% while ensuring simplicity. Zhou Ting proposed a thrust estimation method based on similarity transformation and deep dynamic neural network, which effectively solved the problem of insufficient data under large flight envelope and variable state, and greatly reduced the thrust estimation error in the non-training area. To overcome the problem of repeated training in multi-condition modeling, Zou Yuhang et al. used simulation data, combined with model fine-tuning and improved domain adversarial neural network, to improve the adaptability of the model in different operating conditions.

[0004] Although the above researches continue to improve in the method level, there are still two core problems as follows: Firstly, the thrust during flight cannot be directly obtained, which leads to the fact that existing methods are heavily dependent on simulation data or ground test data as labels. However, the thrust label in simulation data is based on idealized assumptions and lacks credibility; there are significant differences between ground test conditions and flight conditions, which restricts the generalization ability of supervised learning models. Secondly, existing deep learning methods mostly use sequential or grid structure, which is difficult to effectively capture the complex physical coupling relationship between multiple components inside the engine, such as the aerodynamic-mechanical interaction between low-pressure turbine pressure and rotor speed, limiting the further improvement of estimation accuracy. Therefore, it is urgent to develop a thrust estimation method that is suitable for unlabeled scenarios, has structure perception ability and can fuse multi-scale time series information, in order to realize accurate estimation of engine thrust in real flight environment and further improve the health state evaluation level and engineering feasibility. SUMMARY

[0005] The technical problem to be solved by the present application is that: as the core component of the aircraft, the thrust of the aero-engine is directly related to the flight performance and safety. However, in the flight state, the thrust cannot be directly measured due to the limitations of the on-board sensor. The existing method relies on labeled ground test data or simulation data to train the model, which is difficult to directly migrate to the flight condition. Especially in the face of large differences in working conditions and missing thrust labels, traditional supervised learning and shallow models have poor generalization and poor robustness.

[0006] The technical solution adopted by the present application to solve the above technical problem is: fusion of generative adversarial feature extraction, sparse attention mechanism and multi-scale neural network, used for thrust estimation of aero-engine in flight state, aiming to improve the modeling ability of complex working conditions and the robustness of unmeasurable variables.

[0007] To solve the above technical problem, the specific technical solution of the present application is as follows:

[0008] An engine flight thrust estimation method fusing adversarial auto-encoding and sparse attention, the method comprising the following steps:

[0009] Step S1: collecting ground test data and flight state data of a turbofan aero-engine, and preprocessing the collected ground test data and flight state data to obtain steady-state sensor data;

[0010] Step S2: inputting the steady-state sensor data and corresponding working condition labels into the trained adversarial auto-encoder for feature extraction to obtain a latent feature vector;

[0011] Step S3: inputting the latent feature vector Z into the sparse attention module, and performing joint processing of sparse compression by the Linformer sparse attention module and Gumbel-Softmax gating to obtain a fusion feature;

[0012] Step S4: inputting the fusion feature into the multi-scale time series modeling module, and performing time series feature extraction by the multi-scale time series modeling module to obtain a multi-scale fusion feature;

[0013] Step S5: inputting the multi-scale fusion feature into the trained thrust regression module to obtain an aero-engine flight thrust estimation value.

[0014] Further, step S1 comprises the following steps:

[0015] Step S11: collecting ground test data of a turbofan aero-engine, the ground test data including sensor data, working condition labels and thrust labels;

[0016] Step S12: Collecting flight state data of a turbofan aero-engine, the flight state data being composed of sensor data and working condition labels;

[0017] Step S13: Unifying the ground test data and the flight state data into a time sequence format;

[0018] Step S14: Normalizing all collected ground test data and flight state data;

[0019] Step S15: Dividing the normalized sensor data of the ground test data and the flight state data into steady state and transition state to obtain steady state sensor data and transition state sensor data.

[0020] Further, the adversarial autoencoder in step S2 is constructed and trained in the following manner:

[0021] Step S21: Building an adversarial autoencoder, the adversarial autoencoder being built from three sub-modules of an encoder, a decoder and an adversarial discriminator;

[0022] Step S22: Building a loss function, the loss function including a reconstruction loss and an adversarial loss;

[0023] Step S23: Using the loss function to adversarially train the adversarial autoencoder to obtain a trained adversarial autoencoder.

[0024] Further, the sparse attention module in step S3 mainly includes five parts of a linear projection sub-module, a low-rank sparse compression sub-module, an attention calculation sub-module, a differentiable channel selection sub-module and a feature fusion sub-module. The sparse attention module performs the following operations:

[0025] Step S31: The linear projection sub-module uses three sets of linear transformation weights to project the input latent feature vector Z into a query matrix, a key matrix and a value matrix, respectively;

[0026] Step S32: The low-rank sparse compression sub-module uses a Linformer sparse compression mechanism to perform low-rank projection on the key matrix and the value matrix to obtain a low-rank key matrix and a low-rank value matrix;

[0027] Step S33: The attention calculation sub-module uses the low-rank key matrix and the low-rank value matrix to calculate an attention score matrix to obtain an attention representation;

[0028] Step S34: The differentiable channel selection sub-module performs channel-level screening on the input latent feature vector Z through a Gumbel-Softmax gating mechanism, and finally outputs the screened gating feature;

[0029] Step S35: Weighted fusion of the attention representation and the gating feature to obtain a fused feature.

[0030] Further, the multi-scale time sequence modeling module in step S4 comprises a short-time transient module and a long-time transient module, and the following operations are performed:

[0031] Step S41: The short-time transient module adopts a one-dimensional convolutional neural network as a short-time modeler, the fusion feature is input into the short-time transient module, and the short-time output is obtained through convolution operation processing of the short-time transient module;

[0032] Step S42: The long-time transient module is composed of a 4-layer stacked Transformer encoder, and the input is a fusion feature vector; the long-time transient module extracts a global stable trend for up to 60 seconds to obtain a long-time output;

[0033] Step S43: The short-time output and the long-time output are multi-scale feature fusion to obtain multi-scale fusion features.

[0034] Further, step S43 comprises the following steps:

[0035] Step S431: The short-time output and the long-time output are spliced in the feature dimension to obtain a spliced output;

[0036] Step S432: The spliced output is input into a gated recurrent unit network for fusion to obtain multi-scale fusion features.

[0037] Further, the thrust regression module in step S5 is constructed and trained in the following manner:

[0038] Step S51: A fully connected regression network is constructed, and the fully connected regression network is a 3-layer multilayer perceptron; the input is the multi-scale fusion feature, and the output is the thrust estimate value;

[0039] Step S52: A weighted mean square error is used as a regression loss function, and the trained thrust regression module is obtained through training on thrust label data.

[0040] The present application has the following beneficial effects: 1) an end-to-end, interpretable, and transferable aero-engine thrust estimation system is constructed. 2) The method can realize precise estimation of thrust in flight state by using test data knowledge transfer without flight labels, and has significant practical engineering application value and deployment prospect. 3) It is suitable for engine state evaluation and health monitoring under various flight conditions. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0042] Figure 1 is the flow chart of the aero-engine thrust estimation of the present application.

[0043] Figure 2 is the division chart of steady state and transition state of the present application.

[0044] Figure 3 is the air thrust estimation result chart of the simulation data of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort belong to the protection scope of the present application.

[0046] The present application proposes an engine flight thrust estimation method fusing adversarial auto-encoding and sparse attention, as shown in Figure 1 The method comprises the following steps:

[0047] Step S1: Collect the ground test data and flight state data of a turbofan aero-engine, and pretreat the collected ground test data and flight state data to obtain steady state sensor data.

[0048] Step S11: Collect the ground test data of a turbofan aero-engine, and the ground test data comprises sensor data, working condition label and thrust label.

[0049] The ground test bench can set different control parameters to a certain extent to simulate the cruise process in the flight process. In the present application, by setting the control quantities such as engine inlet temperature, pressure, speed and throttle angle, three typical cruise working conditions are defined as the target domain data source: working condition 1: low-altitude subsonic cruise (high-altitude air density); working condition 2: high-altitude subsonic cruise (standard task); working condition 3: high-altitude supersonic cruise (extreme state). The working condition label is represented by c, c={1, 2, 3}; when the working condition label c is 1, it corresponds to working condition 1; when the working condition label c is 2, it corresponds to working condition 2; and when the working condition label c is 3, it corresponds to working condition 3.

[0050] During ground test, high-precision sensor network is deployed at each cross section of the engine and key components to collect real-time sensor data. Key components refer to high-pressure rotor, low-pressure rotor, combustion chamber, compressor, outer duct, low-pressure turbine, tail nozzle, throttle lever, and thrust frame. Test personnel set the working condition parameters under the corresponding working condition label by adjusting the intake temperature, pressure, speed, and throttle angle. The thrust label can be accurately obtained through the thrust sensor on the adapter frame. The collected sensor data covers multiple typical working stages, and the specific collection parameters are shown in Table 1:

[0051] Table 1 Sensor parameter table

[0052]

[0053] The sensor data, working condition label, and thrust label together constitute the ground test data. Among them, the sensor data is collected by the sensors arranged on the key components of the engine in real time and recorded through the data acquisition module in the test control system; the working condition parameters under the corresponding working condition label are set by the test personnel by adjusting the intake temperature, pressure, speed, and throttle angle, and different combinations of settings correspond to different simulated flight conditions; the thrust label is accurately measured by the thrust sensor installed on the adapter frame of the test platform to obtain the axial net thrust output of the engine under the working condition, which serves as the supervision signal of the regression model.

[0054] Step S12: Collecting flight state data of a turbofan aero-engine, the flight state data being composed of sensor data and working condition label.

[0055] The sensor data of the turbofan aero-engine under the flight task working condition is collected by the on-board monitoring system (EMS). The system integrates multiple sensors and is distributed on the key components of the engine, which can monitor the speed, temperature, pressure, and fuel flow in real time. The types of collected parameters are consistent with the ground test data, as shown in Table 1. The working condition label is dynamically determined according to the flight state information (such as altitude, speed, and ambient temperature) and is recorded synchronously with the sensor data.

[0056] Step S13: Unifying the ground test data and the flight state data into time series format.

[0057] The ground test data and the flight state data are in time series format, and the flight state data does not have a thrust label. Each piece of data constitutes a complete sample, indicating the running state of the engine in a certain continuous time window. The format of each sample is unified as a triple: (X; c; Y), X represents sensor data, c represents a working condition label, and Y represents a thrust label. The thrust label only exists in the ground test data, and the thrust label of the flight state data is null. Through sample format standardization, the quality and scale of the input data can be unified, and the requirements of the downstream neural network module for the time series structure can be met.

[0058] The data sampling frequency is set to 5Hz, that is, 5 groups of data points are collected per second. The duration of each piece of data is set to 60 seconds, corresponding to a sample sequence length of 300.

[0059] Step S14: Normalizing all collected ground test data and flight state data.

[0060] All collected ground test data and flight state data are normalized to eliminate the influence between different dimensions. The normalization operation is applied to the sensor data X and the thrust label Y (only existing in the ground test data) of each sample. The training stage also performs normalization processing to improve the numerical stability of the regression model; in the inference stage, the actual thrust value is restored through the inverse normalization operation.

[0061] In the present application, x represents a time series sensor sample data (300 time steps, d-dimensional sensor data), and the sensor data X represents a sample set. The sensor data normalization processing is calculated by the following formula:

[0062]

[0063] wherein x * is the normalized sensor data; x is the sensor sample data; x min is the minimum value of the sensor sample data; x max is the maximum value of the sensor sample data.

[0064] y represents a time series thrust sample label (300 time steps), and the thrust label Y represents a sample set. The thrust label normalization processing is calculated by the following formula:

[0065]

[0066] wherein y * is the normalized thrust label; y is the thrust sample label; y min is the minimum value of the thrust sample label; y max is the maximum value of the thrust sample label.

[0067] Step S15: The normalized sensor data of the ground test data and the flight state data are divided into steady state and transition state to obtain steady state sensor data and transition state sensor data.

[0068] The process of changing from one working state to another working state of an aircraft engine is called a transition working process, and the transition working process is called a transition state, and other smooth running times can be regarded as a steady state. The main purpose of the present application is to obtain sensor data in the steady state of the flight process, so it is necessary to divide the sensor data into steady state and transition state. First, the change rate of all normalized sensor data is calculated:

[0069]

[0070] where Δx i is the change rate of the i-th data point (feature vector); is the normalized sensor data of the i-th data point, i.e. the normalized sensor data at the current sampling time; is the normalized sensor data of the i-1-th data point, i.e. the normalized sensor data at the previous sampling time; and Δt is the sampling time interval.

[0071] Based on the change rate, the data is divided using the k-means clustering algorithm. The basic idea is to divide the data set into K different clusters, so that the data points in each cluster are as similar as possible, and the data difference between different clusters is as large as possible. The goal of this algorithm is to minimize the sum of squares of distances from data points to cluster centers, and the formula is as follows:

[0072]

[0073] where J is the loss function, i.e. the sum of squares of intra-cluster errors, which is used to measure the clustering effect; K is the number of clusters; C k represents the k-th cluster; μ k is the center of the k-th cluster. Through the above clustering result, the cluster with small mean value and low fluctuation amplitude is determined as a steady state cluster, and the corresponding sample is used as steady state sensor data X'; the cluster with large mean value and large change is determined as a transition state cluster, and the corresponding sample is used as transition state sensor data. Finally, the sensor samples in the steady state cluster and the working condition labels are taken as the input of the subsequent adversarial autoencoder. Figure 2 The division results of the transition state and the steady state are shown.

[0074] Step S2: The steady state sensor data X' and the corresponding working condition labels are input into the trained adversarial autoencoder for feature extraction to obtain the latent feature vector Z.

[0075] An adversarial autoencoder is used as a feature extraction front-end module to realize label-free latent space modeling and migration in the flight state. The adversarial autoencoder is trained jointly using reconstruction loss and adversarial loss. Through adversarial optimization training, the ground test data and flight state data can be uniformly distributed in the latent space, thereby improving the label-free generalization ability of the model.

[0076] The adversarial autoencoder is constructed and trained in the following manner:

[0077] Step S21: An adversarial autoencoder is built, which is built from three sub-modules of an encoder, a decoder and an adversarial discriminator.

[0078] The encoder is configured to receive input steady-state sensor data and a working condition label to output a latent feature vector d represents the number of sensors, n represents the number of samples, k represents the number of working conditions, and m represents the dimension size of the latent space.

[0079] The decoder receives the latent feature vector Z and the working condition label c, and attempts to reconstruct the original input steady-state sensor data.

[0080] The discriminator is configured to distinguish whether the input latent feature vector is generated by the encoder (fake sample) or is from a standard normal distribution (real sample).

[0081] Step S22: A loss function is constructed, which includes a reconstruction loss and an adversarial loss.

[0082] The reconstruction loss is represented as follows:

[0083]

[0084] wherein, is the reconstruction loss; E φ (x * ,c) represents the latent feature vector Z output by the encoder according to the normalized sensor data x * and the working condition label c, i.e. Z = E φ (x * ,c); D θ (E φ (x * ,c), c) represents the reconstruction result of the decoder on the latent feature vector Z and the working condition label c i.e. represents the expected (average) value of all training samples.

[0085] The adversarial loss includes a discriminator loss and an encoder deception loss:

[0086]

[0087] where, is the discriminator loss; is the encoder fooling loss; the random variable z follows a standard normal distribution with mean 0 and variance I, z is a real sample; E φ (x * ,c) denotes the latent feature vector Z output by the encoder according to the normalized sensor data x * and the working condition label c; T ψ (·) denotes that the discriminator discriminates whether the input latent feature vector Z is generated by the encoder (fake sample) or the standard normal distribution (real sample); denotes the expectation (average) value over all training samples; denotes the expectation (average) value over the random variable z following the standard normal distribution.

[0088] The total training objective is to jointly minimize the following joint loss function:

[0089]

[0090] where λ is a weight hyperparameter for balancing the reconstruction accuracy and the latent distribution alignment performance.

[0091] Step S23: Perform adversarial training on the adversarial autoencoder to obtain a trained adversarial autoencoder.

[0092] The training of the adversarial autoencoder adopts an alternating optimization strategy, including the following steps:

[0093] Step S231: Update the discriminator while fixing the encoder parameters to minimize the discriminator loss to enhance the discrimination ability.

[0094] Step S231: Update the encoder while fixing the discriminator parameters to minimize the encoder fooling loss so that the fake sample is closer to the real distribution.

[0095] Step S232: Update the encoder + decoder (reconstruction) while minimizing the reconstruction loss to improve the restoration accuracy of the original input.

[0096] Step S233: Jointly optimize the joint loss to achieve a balance between latent space alignment and input information preservation.

[0097] Step S234: Freeze the adversarial autoencoder after training to obtain a trained adversarial autoencoder and use it as a feature extractor.

[0098] Input all steady-state sensor data and corresponding working condition labels into the trained adversarial autoencoder to obtain a latent feature vector Z.

[0099] Step S3: input the latent feature vector Z into the sparse attention module, and perform joint processing of sparse compression and Gumbel-Softmax gating through the Linformer sparse compression module to obtain fused features.

[0100] Based on the sparse attention module, the sparse attention mechanism is introduced, and the latent feature vector is input into the sparse attention module to model the nonlinear dependence relationship between sensor channels. In order to further reduce the interference of redundant input, the sparse attention module further includes a differentiable channel selection module, which uses the Gumbel-Softmax strategy to dynamically filter the channel combination that significantly affects the thrust from the input channels, and shields irrelevant or noise channels.

[0101] The sparse attention module mainly includes a linear projection submodule, a low-rank sparse compression submodule, an attention calculation submodule, a differentiable channel selection submodule, and a feature fusion submodule. The sparse attention module performs the following operations:

[0102] Step S31: the linear projection submodule uses three sets of linear transformation weights to project the input latent feature vector Z into a query matrix (Query), a key matrix (Key), and a value matrix (Value), as shown below:

[0103] Q=ZW Q ,K=ZW K ,V=ZW V

[0104] wherein Z is the input latent feature vector; W Q , W K , and W V are learnable parameters, and Q, K, and V are the query matrix, the key matrix, and the value matrix, respectively.

[0105] Step S32: the low-rank sparse compression submodule uses the Linformer sparse compression mechanism to perform low-rank projection on the key matrix K and the value matrix V to obtain a low-rank key matrix K' and a low-rank value matrix V'.

[0106] The Linformer sparse compression mechanism can reduce the attention calculation complexity from O(n 2 ) to O(n), significantly improving the calculation efficiency.

[0107] Step S33: the attention calculation submodule calculates the attention score matrix using the low-rank key matrix and the low-rank value matrix to obtain the attention representation, and the attention score matrix calculation formula is as follows:

[0108]

[0109] wherein Z att is the attention representation, d k is the feature dimension size of the low-rank bond matrix K', is used for the scaling factor to prevent gradient explosion; represents the attention representation of each channel to other channels, N represents the final attention matrix dimension; T represents transposition; softmax represents an activation function, which converts an arbitrary real number vector into a probability distribution vector with a value range of [0, 1] and a sum of 1.

[0110] Step S34: The differentiable channel selection submodule performs channel-level screening on the input latent feature vector Z through the Gumbel-Softmax gating mechanism, and finally outputs the screened gating feature Z gate .

[0111] The differentiable channel selection submodule is the core component of the sparse attention module, which performs channel-level screening on the input latent feature vector Z through the Gumbel-Softmax gating mechanism. Its working logic is: using the differentiable Gumbel-Softmax sampling strategy, dynamic weights are generated for each sensor channel, automatically identifying the channel combination that significantly affects the thrust estimation, suppressing the interference of redundant or noisy channels, and finally outputting the screened gating feature Z gate , realizing the strengthening of key information and the filtering of irrelevant information.

[0112] Step S35: The attention representation Z att is weighted and fused with the gating feature Z gate to obtain the fusion feature. The fusion method is represented as follows:

[0113] Z fused = a · Z att +(1-a)·Z gate

[0114] wherein a∈[0, 1] is a learnable fusion weight; the final output fusion feature N represents the number of samples, and d' represents the fusion feature dimension.

[0115] Step S4: The fusion feature is input into the multi-scale time sequence modeling module, and after time sequence feature extraction by the multi-scale time sequence modeling module, a multi-scale fusion feature is obtained.

[0116] To model the short-time transient response and long-time stable trend in the sensor sequence of an aero-engine simultaneously, a multi-scale time series modeling module is adopted to complete time series feature extraction. Through the multi-scale time series modeling module, transient nonlinear disturbance and long-time stable trend can be perceived simultaneously, so as to realize more comprehensive engine time series modeling and improve thrust estimation accuracy.

[0117] The multi-scale time series modeling module includes a short-time transient module and a long-time transient module, which performs the following operations:

[0118] Step S41: The short-time transient module adopts a one-dimensional convolutional neural network (1D CNN) as a short-time modeler, fuses the feature input into the short-time transient module, and obtains a short-time output through convolution operation processing of the short-time transient module.

[0119] The network depth of the one-dimensional convolutional neural network of the short-time transient module is set to 5 layers, each layer adopts a causal convolution with a sliding window; the time step of each input sample is set to 300, corresponding to a 60-second time series window (the sampling frequency is 5 Hz); the length of each convolution kernel is set to 10, and the step is 1, which ensures covering the local gradient features corresponding to typical transient phenomena; the output of the short-time transient module is a short-time output , which represents the transient response features of each sensor channel in the short term; d s , which represents the short-time hidden dimension of each layer.

[0120] Step S42: The long-time transient module is composed of a Transformer encoder with 4 layers stacked, and the input is a fusion feature vector. The long-time transient module extracts a global stable trend for up to 60 seconds to obtain a long-time output.

[0121] The long-time transient module is used to extract a global stable trend for up to 60 seconds, and position encoding is added in the time dimension to preserve the order of working condition changes; the number of attention heads of the Transformer encoder is set to 4, and the output is a long-time output l , which represents the long-time hidden dimension of each layer.

[0122] Step S43: The short-time output and the long-time output are multi-scale feature fused to obtain a multi-scale fusion feature.

[0123] Step S431: The short-time output Z short is concatenated with the long-time output Z long in the feature dimension to obtain a concatenated output Z fused , Z fused = [Z short , Z long ].

[0124] Step S432: The concatenated output Z fused ​The input gated recurrent unit (GRU) network is fused to obtain multi-scale fusion features Z mulit .

[0125] The gated recurrent unit models the time dynamics and unifies the feature dimension output; the final output is the multi-scale fusion features Z mulit , i.e., the multi-scale time modeling representation after fusion, as the input of the subsequent thrust estimation module.

[0126] Step S5: input the multi-scale fusion features into the trained thrust regression module to obtain the aero-engine flight thrust estimate value.

[0127] The thrust regression module is used to map the multi-scale fusion features Z mulit into the aero-engine flight thrust estimate value. Through the thrust regression module, high-precision thrust prediction can be realized, which helps to ensure the reliability of critical flight missions.

[0128] The thrust regression module is constructed and trained in the following way:

[0129] Step S51: construct a fully connected regression network, which is a 3-layer multi-layer perceptron (MLP); the input is the multi-scale fusion features Z mulit , and the output is the thrust estimate value

[0130] The network structure of the fully connected network is as follows:

[0131] First layer MLP: the number of neurons is set to 2d merge , and the activation function uses GeLU (Gaussian Error Linear Unit), whose formula is:

[0132] GeLU(x) = x·Φ(x)

[0133] where x represents the input vector of the hidden layer in the fully connected regression network, and Φ(x) is the cumulative distribution function of the standard normal distribution. This activation function can better capture nonlinear features and alleviate the "dead neuron" problem of ReLU.

[0134] Second layer MLP: the number of neurons is reduced to d merge , and the activation function uses ReLU (Rectified Linear Unit), whose formula is:

[0135] ReLU(x) = max(0, x)

[0136] where x represents the input vector of the hidden layer in the fully connected regression network.

[0137] Output layer: linear layer (without activation function), neuron number is 1, directly output normalized thrust estimation value Subsequently, the actual thrust value is restored through denormalization.

[0138] Step S52: training on labeled data (ground test) using weighted mean square error as regression loss function, obtaining trained thrust regression module.

[0139] The regression loss function is expressed as follows:

[0140]

[0141] Wherein, is the thrust estimation value; y * is the normalized true thrust label (only exists in ground test data); w(x * ) is a weight function adjusted according to the working condition parameters, used to balance the error contribution of high thrust and low thrust stages; represents the expected (average) value of all training samples; The final regression loss function is used to train the thrust regression module to obtain the trained thrust regression module.

[0142] The aviation engine flight thrust estimation module is composed of the adversarial autoencoder, the sparse attention module, the multi-scale time sequence modeling module and the thrust regression module.

[0143] The simulation experiment of the present application is based on F-119 engine test and flight data, and the in-flight thrust estimation result of the simulation data is shown in Figure 3 , wherein the solid line is the true engine reasoning value, and the gray line is the thrust value predicted by the present method. The results show that the engine flight thrust estimation method proposed in the present application which fuses adversarial autoencoder and sparse attention still has good estimation accuracy and generalization ability under unsupervised flight state, 95% error falls within ±3kN interval, R 2 index is more than 0.75 on average, showing high practical application potential.

[0144] The above only describes the preferred embodiments of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application are included in the protection scope of the present application.

Claims

1. A method for engine flight thrust estimation by fusing adversarial autoencoding and sparse attention, characterized in that, The method comprises the following steps: Step S1: Collecting the ground test data and flight state data of the turbofan aero-engine, and preprocessing the collected ground test data and flight state data to obtain steady-state sensor data; Step S2: Inputting the steady-state sensor data and the corresponding working condition label into the trained adversarial autoencoder for feature extraction to obtain a latent feature vector; Step S3: Inputting the latent feature vector Z into the sparse attention module, and performing joint processing of sparse compression and Gumbel-Softmax gating through the Linformer sparse compression of the sparse attention module to obtain a fusion feature; Step S4: Inputting the fusion feature into the multi-scale time series modeling module, and performing time series feature extraction through the multi-scale time series modeling module to obtain a multi-scale fusion feature; Step S5: Inputting the multi-scale fusion feature into the trained thrust regression module to obtain an aero-engine flight thrust estimate value; The sparse attention module in step S3 comprises a linear projection submodule, a low-rank sparse compression submodule, an attention calculation submodule, a differentiable channel selection submodule, and a feature fusion submodule; The sparse attention module performs the following operations: Step S31: The linear projection submodule uses three sets of linear transformation weights to project the input latent feature vector Z into a query matrix, a key matrix, and a value matrix, respectively; Step S32: The low-rank sparse compression submodule uses the Linformer sparse compression mechanism to perform low-rank projection on the key matrix and the value matrix to obtain a low-rank key matrix and a low-rank value matrix; Step S33: The attention calculation submodule calculates an attention score matrix using the low-rank key matrix and the low-rank value matrix to obtain an attention representation; Step S34: The differentiable channel selection submodule performs channel-level screening on the input latent feature vector Z through the Gumbel-Softmax gating mechanism, and finally outputs the screened gated feature; Step S35: The attention representation and the gated feature are weighted and fused to obtain a fusion feature.

2. The method of claim 1, wherein the method is implemented by a processor. Step S1 comprises the following steps: Step S11: Collecting the ground test data of a turbofan aero-engine, the ground test data comprising sensor data, working condition labels, and thrust labels; Step S12: Collecting the flight state data of a turbofan aero-engine, the flight state data being composed of sensor data and working condition labels; Step S13: Unifying the ground test data and the flight state data into a time series format; Step S14: Normalizing all collected ground test data and flight state data; Step S15: Dividing the normalized sensor data of the ground test data and the flight state data into steady-state and transition-state to obtain steady-state sensor data and transition-state sensor data.

3. The method of claim 1, wherein the method is implemented by a processor. The adversarial autoencoder in step S2 is constructed and trained in the following manner: Step S21: Building an adversarial autoencoder, the adversarial autoencoder being built from an encoder, a decoder, and an adversarial discriminator submodule; Step S22: Constructing a loss function, the loss function comprising a reconstruction loss and an adversarial loss; Step S23: Training the adversarial autoencoder using the loss function to obtain a trained adversarial autoencoder.

4. The method of claim 1, wherein the method is implemented by a processor. The multi-scale time sequence modeling module in step S4 includes a short-time transient module and a long-time transient module, and performs the following operations: Step S41: The short-time transient module adopts a one-dimensional convolutional neural network as a short-time modeler, fuses the feature input into the short-time transient module, and obtains a short-time output through convolution operation processing of the short-time transient module; Step S42: The long-time transient module is composed of a Transformer encoder composed of 4 layers of stacks, and the input is a fusion feature vector. The long-time transient module extracts a global stable trend lasting 60 seconds to obtain a long-time output; Step S43: The short-time output and the long-time output are fused in a multi-scale feature fusion manner to obtain a multi-scale fusion feature.

5. The method of claim 4, wherein the method is implemented by a processor. Step S43 includes the following steps: Step S431: The short-time output and the long-time output are spliced in the feature dimension to obtain a spliced output; Step S432: The spliced output is input into a gated recurrent unit network for fusion to obtain a multi-scale fusion feature.

6. The method of claim 1, wherein the method is implemented by a processor. The thrust regression module in step S5 is constructed and trained in the following manner: Step S51: A fully connected regression network is constructed, and the fully connected regression network is a 3-layer multilayer perceptron; the input is a multi-scale fusion feature, and the output is a thrust estimate value; Step S52: A weighted mean square error is used as a regression loss function, and the trained thrust regression module is obtained through training on thrust label data.

Images