CN-121997459-A - Civil aircraft engine thrust estimation method and system based on multi-modal convolution model

CN121997459ACN 121997459 ACN121997459 ACN 121997459ACN-121997459-A

Abstract

The invention discloses a civil aircraft engine thrust estimation method and system based on a multi-mode convolution model, and relates to the technical field of aircraft engine control and artificial intelligence, wherein the method comprises the steps of acquiring engine time sequence data and corresponding thrust label data, and constructing an original data set; the method comprises the steps of preprocessing an original data set to generate a time domain signal, a frequency domain signal and coupling characteristics based on an engine physical mechanism, constructing a multi-modal convolutional neural network thrust estimation model, extracting the time domain characteristics from the time domain signal according to the time domain convolutional neural network of the model, extracting the frequency domain characteristics from the frequency domain signal according to the frequency domain convolutional neural network, screening the coupling characteristics from the coupling characteristics according to the coupling characteristic network, fusing the extracted time domain characteristics and the frequency domain characteristics to obtain signal fusion characteristics, fusing the signal fusion characteristics and the coupling characteristics to obtain multi-modal fusion characteristics, and inputting the multi-modal fusion characteristics into the model to output an engine thrust estimation value. Thereby improving the fitting accuracy of the thrust estimation.

Inventors

WANG WEI
WANG SHUO
TAO LIQUAN
YANG KUN

Assignees

中国民航大学

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (10)

1. A civil aircraft engine thrust estimation method based on a multi-modal convolution model is characterized by comprising the following steps: Acquiring time sequence data of a multisource sensor of an engine in a full flight envelope range and corresponding thrust label data, and constructing an original data set; Performing data preprocessing on the original data set to generate a time domain signal, a frequency domain signal and a coupling characteristic based on an engine physical mechanism; Constructing a multi-mode convolutional neural network thrust estimation model, wherein the model comprises a time domain convolutional neural network branch, a frequency domain convolutional neural network branch and a coupling characteristic network branch; Extracting time domain features from the time domain signals according to the time domain convolutional neural network branches, extracting frequency domain features from the frequency domain signals according to the frequency domain convolutional neural network branches, and screening coupling features from the coupling features according to the coupling feature network branches; fusing the extracted time domain features and the frequency domain features to obtain signal fusion features, and fusing the signal fusion features and the coupling features to obtain multi-mode fusion features; And inputting the multi-modal fusion characteristics into the multi-modal convolutional neural network thrust estimation model to estimate an engine thrust value, and outputting the engine thrust estimated value.
2. The method of claim 1, wherein obtaining multi-source sensor timing data and corresponding thrust signature data for an engine over a full flight envelope to construct a raw dataset comprises: Generating original sample data covering multiple working conditions based on a thermodynamic component level model, wherein the original sample data comprises engine state parameters and corresponding thrust labels; Screening time sequence data of key variables from the original sample data based on the influence degree of each state parameter on thrust; and windowing the time sequence data of the screened key variables to construct an original data set.
3. The method of claim 1, wherein data preprocessing the raw dataset to generate a time domain signal, a frequency domain signal, and a coupling feature based on an engine physical mechanism, comprising: cleaning and standardizing the time sequence data in the original data set to obtain a time domain signal; performing short-time Fourier transform on the time domain signal to obtain a corresponding time spectrum, and extracting frequency domain features from the time spectrum to obtain a frequency domain signal; Based on an engine physical mechanism, an initial feature set containing linear coupling features, nonlinear interaction features and time sequence coupling features is constructed from the original data set, and screening and dimension reduction are carried out on the initial feature set to remove redundancy, so that coupling features are obtained.
4. The method of claim 3, wherein fusing the extracted time domain features with the frequency domain features to obtain signal fusion features, and fusing the signal fusion features with the coupling features to obtain multi-modal fusion features, comprises: dynamically fusing the time domain features and the frequency domain features through an attention weight learning mechanism to obtain signal fusion features; And performing cross-modal fusion on the signal fusion characteristic and the coupling characteristic through bidirectional attention interaction and self-adaptive weight distribution to generate the multi-modal fusion characteristic.
5. The method of claim 4, wherein cross-modal fusing the signal fusion feature with the coupling feature by bi-directional attention interactions and adaptive weight distribution to generate the multi-modal fusion feature comprises: Mapping the signal fusion feature and the coupling feature to the same feature space; Performing bidirectional attention interaction on the signal fusion characteristics and the coupling characteristics mapped to the same characteristic space to obtain the signal fusion characteristics and the coupling characteristics after interaction enhancement; and generating the multi-mode fusion characteristic through a self-adaptive weight distribution mechanism according to the signal fusion characteristic and the coupling characteristic after interaction enhancement.
6. The method of claim 1, wherein inputting the multi-modal fusion feature into the multi-modal convolutional neural network thrust estimation model for engine thrust value estimation, outputting an engine thrust estimation value, comprises: inputting the multi-mode fusion characteristics into a multi-layer perceptron regression output layer of the model; And carrying out nonlinear mapping on the multi-mode fusion characteristics through at least one fully-connected hidden layer in the multi-layer perceptron regression output layer, and outputting the engine thrust estimated value.
7. The method of claim 1, wherein the structures of the time domain convolutional neural network branches, the frequency domain convolutional neural network branches, and the coupling feature network branches are respectively: The time domain convolution neural network branch adopts a network structure comprising multi-scale one-dimensional convolution, residual error connection and a channel attention mechanism; the frequency domain convolution neural network branches adopt a network structure comprising multi-scale two-dimensional convolution, residual error connection and a time-frequency joint attention mechanism; The coupling characteristic network branch adopts a network structure which takes a fully connected network as a main body and integrates a characteristic importance evaluation module.
8. The method of claim 7, wherein the coupling feature network branch further comprises a physical constraint layer: The physical constraint layer is used for carrying out transformation and weighted fusion on a physical constraint relation constructed based on at least one of a thermodynamic law, a momentum conservation law, a mass conservation law and an energy conservation law through a learnable network layer, and combining the fused constraint enhancement characteristic with the original coupling characteristic.
9. The method of claim 1, wherein the training process of the multi-modal convolutional neural network thrust estimation model comprises: Iteratively optimizing parameters of a multi-modal convolutional neural network thrust estimation model based on a joint loss function, the joint loss function comprising a base predictive loss term and a physical constraint loss term, wherein, The basic prediction loss term is mean square error loss, average absolute error loss or Huber loss; The physical constraint loss term is constructed based on at least one of a law of thermodynamics, a law of conservation of momentum, a law of conservation of mass, and a law of conservation of energy.
10. A civil aircraft engine thrust estimation system based on a multi-modal convolution model, comprising: The data acquisition module is used for acquiring time sequence data of the multisource sensor of the engine in the full flight envelope range and corresponding thrust label data and constructing an original data set; the data preprocessing module is used for carrying out data preprocessing on the original data set to generate a time domain signal, a frequency domain signal and a coupling characteristic based on an engine physical mechanism; the thrust estimation model construction module is used for constructing a multi-mode convolutional neural network thrust estimation model, and the model comprises a time domain convolutional neural network branch, a frequency domain convolutional neural network branch and a coupling characteristic network branch; the characteristic screening module is used for extracting time domain characteristics from the time domain signals according to the time domain convolutional neural network branches, extracting frequency domain characteristics from the frequency domain signals according to the frequency domain convolutional neural network branches, and screening coupling characteristics from the coupling characteristics according to the coupling characteristic network branches; The feature fusion module is used for fusing the extracted time domain features with the frequency domain features to obtain signal fusion features, and fusing the signal fusion features with the coupling features to obtain multi-mode fusion features; and the engine thrust value estimation module is used for inputting the multi-mode fusion characteristic into the multi-mode convolutional neural network thrust estimation model to estimate the engine thrust value and outputting an engine thrust estimated value.

Description

Civil aircraft engine thrust estimation method and system based on multi-modal convolution model Technical Field The embodiment of the invention relates to the technical field of aero-engine control and artificial intelligence, in particular to a civil aero-engine thrust estimation method and system based on a multi-mode convolution model. Background Aeroengine thrust is a key parameter in flight control and engine health management. The method for directly measuring the thrust by using the traditional dynamometer sensor in the ground test has the problems of complex installation, high maintenance cost, complex test design and the like. The thrust sensor cannot be installed in flight. Therefore, in practical application, a thrust estimator is often adopted to indirectly obtain thrust feedback information so as to meet the requirements of advanced control methods such as direct thrust control and the like. The existing thrust estimation methods are mainly divided into a model-based method and a data-driven method. Model-based thrust estimation typically utilizes an aerodynamic thermodynamic model or a simplified linear model at the engine component level to infer thrust output. For example, fusing the linear model by Kalman filtering may reduce the amount of computation and enable real-time thrust estimation. However, the model uncertainty exists in the model-based method, and the complexity of simplified modeling can lead to the fact that the thrust estimation accuracy does not meet the requirements. The requirement of meeting the precision of the high-fidelity part-level thermodynamic model can lead to increased computational complexity and poorer real-time performance, and the requirement of estimating the airborne thrust is difficult to realize. Data-driven thrust estimation is used to train a machine learning model using ground test or flight data to fit a nonlinear relationship between engine state parameters and thrust. For example, real-time fitting of source domain data may be achieved by training a Recurrent Neural Network (RNN) with ground test data. But the data-driven based approach has high dependence on the training dataset, requiring that the dataset cover all state information for the full envelope range of the flight. And the traditional data driving method has poor generalization capability on special working conditions such as emergency working conditions. With the development of deep learning, a modeling time sequence neural network (such as an LSTM long-short-term memory neural network) aiming at high time-varying characteristics of an aeroengine utilizes time sequence information, so that the accuracy problem of random dynamic thrust estimation of a traditional feedforward neural network is improved. However, the data source domain of the time sequence model is single, and the parameter coupling relation of the multi-source sensor cannot be perceived, so that the robustness of thrust estimation is insufficient. In summary, the prior art mainly has the following defects that (1) only single domain information is utilized and engine state cannot be comprehensively represented, and (2) model training does not consider the coupling mechanism among aeroengine parameters and the constraint of physical information, and the result of thrust estimation is not limited by thermodynamic rules. Disclosure of Invention Aiming at the defects of the existing data-driven aeroengine thrust estimation method in terms of precision and robustness, the invention provides a multi-mode convolution model-based civil aeroengine thrust estimation method and system, which aim to construct time domain features, frequency domain features and parameter coupling features from a multi-source sensor, construct an end-to-end deep learning model by utilizing complementary information of multi-source parameters, realize accurate thrust estimation in a full envelope range, and realize a cross-mode attention mechanism at a feature fusion layer by introducing physical constraint based on thermodynamic laws into a loss function of the model, thereby ensuring real-time performance and improving self-adaption capability and robustness of the model. In a first aspect, an embodiment of the present invention provides a method for estimating thrust of a civil aircraft engine based on a multi-modal convolution model, including: Acquiring time sequence data of a multisource sensor of an engine in a full flight envelope range and corresponding thrust label data, and constructing an original data set; Performing data preprocessing on the original data set to generate a time domain signal, a frequency domain signal and a coupling characteristic based on an engine physical mechanism; Constructing a multi-mode convolutional neural network thrust estimation model, wherein the model comprises a time domain convolutional neural network branch, a frequency domain convolutional neural network branch and a coupling characteristic network branch