Civil aero-engine thrust estimation method and system based on multi-modal convolution model

Abstract

The invention discloses a civil aero-engine thrust estimation method and system based on a multi-modal convolution model, and relates to the technical field of aero-engine control and artificial intelligence, and the method comprises the steps: obtaining engine time sequence data and corresponding thrust label data, and constructing an original data set; preprocessing the original data set to generate a time domain signal, a frequency domain signal and a coupling feature based on an engine physical mechanism; constructing a multi-modal convolutional neural network thrust estimation model, extracting time domain features from the time domain signals according to a time domain convolutional neural network of the model, extracting frequency domain features from the frequency domain signals according to a frequency domain convolutional neural network, and screening coupling features from coupling features according to a coupling feature network; 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; and inputting the multi-modal fusion features into the model, and outputting an engine thrust estimation value. Therefore, the fitting precision of thrust estimation is improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the fields of aero-engine control and artificial intelligence technology, and in particular to a method and system for estimating the thrust of civil aero-engines based on a multimodal convolution model. Background Technology

[0002] Engine thrust is a critical parameter in flight control and engine health management. Directly measuring thrust using traditional dynamometer sensors in ground tests presents challenges such as complex installation, high maintenance costs, and complex test design. Thrust sensors cannot be installed during flight. Therefore, in practical applications, thrust estimators are often used to indirectly obtain thrust feedback information to meet the requirements of advanced control methods such as direct thrust control.

[0003] Existing thrust estimation methods are mainly divided into model-based methods and data-driven methods. Model-based thrust estimation typically uses engine component-level aerodynamic and thermodynamic models or simplified linear models to calculate thrust output. For example, Kalman filtering to fuse linear models can reduce computational complexity and achieve real-time thrust estimation. However, model-based methods suffer from model uncertainty, and simplifying modeling complexity can lead to insufficient thrust estimation accuracy. High-fidelity component-level thermodynamic models that meet accuracy requirements result in increased computational complexity and poor real-time performance, making it difficult to meet airborne thrust estimation needs. Data-driven thrust estimation uses ground test or flight data to train machine learning models to fit the nonlinear relationship between engine state parameters and thrust. For example, training recurrent neural networks (RNNs) with ground test data can achieve real-time fitting of source domain data. However, data-driven methods are highly dependent on the training dataset, requiring the dataset to cover all state information across the entire flight envelope. Furthermore, traditional data-driven methods have poor generalization ability for special operating conditions such as emergency situations. With the development of deep learning, time-series neural networks (such as LSTM long short-term memory neural networks) for modeling the highly time-varying characteristics of aero-engines utilize time-series information to improve the accuracy of traditional feedforward neural networks in stochastic dynamic thrust estimation. However, time-series models have a single data source domain and cannot perceive the parameter coupling relationship of multiple sensors, resulting in insufficient robustness of thrust estimation.

[0004] In summary, the existing technologies have the following shortcomings: (1) They only use information from a single domain and fail to fully characterize the engine state; (2) The model training does not consider the coupling mechanism between aero-engine parameters and the constraints of physical information, and the thrust estimation results are not limited by thermodynamic laws. Summary of the Invention

[0005] This invention addresses the shortcomings of existing data-driven aero-engine thrust estimation methods in terms of accuracy and robustness by providing a civil aero-engine thrust estimation method and system based on a multimodal convolutional model. The method aims to construct temporal, frequency, and parameter coupling features from multiple sensor sources, utilize complementary information from these multiple sources to build an end-to-end deep learning model, and achieve accurate thrust estimation across the entire thrust envelope. By introducing physical constraints based on thermodynamic laws into the model's loss function and implementing a cross-modal attention mechanism in the feature fusion layer, the method ensures real-time performance while enhancing the model's adaptability and robustness.

[0006] In a first aspect, embodiments of the present invention provide a method for estimating the thrust of a civil aircraft engine based on a multimodal convolutional model, comprising: Acquire multi-source sensor time-series data and corresponding thrust label data of the engine across the entire flight envelope to construct the raw dataset; The original dataset is preprocessed to generate time-domain signals, frequency-domain signals, and coupling features based on engine physical mechanisms. A multimodal convolutional neural network thrust estimation model is constructed, which includes a temporal convolutional neural network branch, a frequency domain convolutional neural network branch, and a coupled feature network branch. The time-domain features are extracted from the time-domain signal according to the time-domain convolutional neural network branch, the frequency-domain features are extracted from the frequency-domain signal according to the frequency-domain convolutional neural network branch, and the coupling features are filtered from the coupling features according to the coupling feature network branch. The extracted time-domain features are fused with the frequency-domain features to obtain signal fusion features, and the signal fusion features are fused with the coupling features to obtain multimodal fusion features; The multimodal fusion features are input into the multimodal convolutional neural network thrust estimation model to estimate the engine thrust value, and the engine thrust estimate is output.

[0007] As a preferred implementation, multi-source sensor time-series data and corresponding thrust label data of the engine across the entire flight envelope are acquired to construct an original dataset, including: Based on a thermodynamic component-level model, raw sample data covering multiple operating conditions is generated. The raw sample data includes engine state parameters and corresponding thrust labels. Based on the degree of influence of each state parameter on thrust, time series data of key variables are selected from the original sample data; The time-series data of the selected key variables are windowed to construct the original dataset.

[0008] In a preferred embodiment, the original dataset is preprocessed to generate time-domain signals, frequency-domain signals, and coupling features based on engine physical mechanisms, including: The time-series data in the original dataset are cleaned and standardized to obtain the time-domain signal; The time-domain signal is subjected to a short-time Fourier transform to obtain the corresponding time spectrum, and frequency domain features are extracted from the time spectrum to obtain the frequency domain signal; Based on the engine physics mechanism, an initial feature set containing linear coupling features, nonlinear interaction features, and temporal coupling features is constructed from the original dataset. The initial feature set is then filtered and dimensionality reduced to remove redundancy, resulting in coupling features.

[0009] In a preferred embodiment, the extracted time-domain features are fused with the frequency-domain features to obtain signal fusion features, and the signal fusion features are fused with the coupling features to obtain multimodal fusion features, including: The time-domain features and frequency-domain features are dynamically fused using an attention weight learning mechanism to obtain signal fusion features; The signal fusion features and the coupled features are fused across modes through bidirectional attention interaction and adaptive weight allocation to generate the multimodal fusion features.

[0010] In a preferred embodiment, the signal fusion features and the coupled features are fused across modally through bidirectional attention interaction and adaptive weight allocation to generate the multimodal fusion features, including: Map the signal fusion features and the coupling features to the same feature space; Bidirectional attention interaction is performed on the signal fusion features and coupling features mapped to the same feature space to obtain the signal fusion features and coupling features enhanced by interaction. Based on the enhanced signal fusion features and coupling features, the multimodal fusion features are generated through an adaptive weight allocation mechanism.

[0011] In a preferred embodiment, the multimodal fusion features are input into the multimodal convolutional neural network thrust estimation model to estimate the engine thrust value, and the engine thrust estimate is output, including: The multimodal fusion features are input into the multilayer perceptron regression output layer of the model; The multimodal fusion features are nonlinearly mapped by at least one fully connected hidden layer in the regression output layer of the multilayer perceptron to output the engine thrust estimate.

[0012] In a preferred embodiment, the structures of the temporal convolutional neural network branch, the frequency domain convolutional neural network branch, and the coupled feature network branch are as follows: The temporal convolutional neural network branch adopts a network structure that includes multi-scale one-dimensional convolution, residual connections, and channel attention mechanisms. The frequency domain convolutional neural network branch adopts a network structure that includes multi-scale two-dimensional convolution, residual connections, and a time-frequency joint attention mechanism. The coupled feature network branch adopts a network structure with a fully connected network as the main body and an integrated feature importance evaluation module.

[0013] In a preferred embodiment, the coupled feature network branch further includes a physical constraint layer: The physical constraint layer is used to transform and weight and fuse physical constraint relationships constructed based on at least one of the laws of thermodynamics, conservation of momentum, conservation of mass, and conservation of energy through a learnable network layer, and to combine the fused constraint enhancement features with the original coupling features.

[0014] In a preferred embodiment, the training process of the multimodal convolutional neural network thrust estimation model includes: The parameters of a multimodal convolutional neural network thrust estimation model are iteratively optimized based on a joint loss function, which includes a basic prediction loss term and a physical constraint loss term. The basic prediction loss term is mean squared error loss, mean absolute error loss, or Huber loss. The physical constraint loss term is constructed based on at least one of the laws of thermodynamics, the law of conservation of momentum, the law of conservation of mass, and the law of conservation of energy.

[0015] Secondly, embodiments of the present invention also provide a thrust estimation system for civil aircraft engines based on a multimodal convolution model, comprising: The data acquisition module is used to acquire multi-source sensor time-series data and corresponding thrust label data of the engine within the entire flight envelope, and to construct the raw dataset; The data preprocessing module is used to preprocess the original dataset to generate time-domain signals, frequency-domain signals, and coupling features based on engine physical mechanisms. The thrust estimation model building module is used to construct a multimodal convolutional neural network thrust estimation model, which includes a temporal convolutional neural network branch, a frequency domain convolutional neural network branch, and a coupled feature network branch. The feature filtering module is used to extract time-domain features from the time-domain signal according to the time-domain convolutional neural network branch, extract frequency-domain features from the frequency-domain signal according to the frequency-domain convolutional neural network branch, and filter coupling features from the coupling features according to the coupling feature network branch; The feature fusion module is used to fuse the extracted time-domain features with the frequency-domain features to obtain signal fusion features, and to fuse the signal fusion features with the coupling features to obtain multimodal fusion features; The engine thrust estimation module is used to input the multimodal fusion features into the multimodal convolutional neural network thrust estimation model to estimate the engine thrust and output the engine thrust estimate.

[0016] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the civil aviation engine thrust estimation method based on a multimodal convolution model as described in any embodiment of the present invention.

[0017] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the civil aviation engine thrust estimation method based on a multimodal convolution model as described in any embodiment of the present invention.

[0018] Compared with existing technologies, the present invention achieves the following beneficial effects: (1) This invention achieves complementary enhancement of information in different dimensions through multimodal joint modeling of time-domain signals, frequency-domain signals, and coupling features based on physical mechanisms, thus avoiding the shortcomings of incomplete representation of information in a single modality. By combining cross-modal attention fusion and adaptive weight allocation, the contribution relationship of each modality to thrust is accurately explored, significantly reducing systematic bias and random error. Experimental data show that the root mean square error (RMSE) of this scheme is reduced by about 53.7% compared with the optimal single-branch model, the mean absolute error (MAE) is reduced by about 26.9% compared with the optimal double-branch model, and the coefficient of determination (R²) reaches 0.999. The fitting accuracy and numerical accuracy of thrust estimation far exceed those of traditional single-modal or simple fusion schemes.

[0019] (2) This invention embeds physical constraints such as thermodynamics, momentum, mass flow rate and energy conservation in the coupled feature network branches. Through learnable transformation and residual fusion, it ensures that the estimation results conform to the core physical laws of the engine and avoids the "physical inconsistency" problem that is easy to occur in pure data-driven methods. On the other hand, the bidirectional attention interaction mechanism effectively suppresses cross-modal information conflict. The adaptive weight allocation can dynamically adjust the contribution of each path according to the working conditions. It can still maintain stable performance in complex scenarios such as strong transient and emergency working conditions. It solves the problem of amplitude deviation and phase lag that is easy to occur in traditional methods under special working conditions and improves the engineering reliability of thrust estimation.

[0020] (3) This invention takes into account the differences in working conditions under different Mach numbers, altitudes, and thrust requirements within the entire envelope by using multi-scale convolution, parameter sensitivity analysis, and coupled feature dimensionality reduction. Compared with traditional data-driven methods that rely on specific datasets, this scheme can achieve working condition transfer through physical constraints and cross-modal complementarity without covering all extreme working condition samples. It has a strong adaptability to unseen transient processes and boundary working conditions, and effectively solves the problems of poor generalization ability and high dependence on training datasets in traditional methods.

[0021] (4) This invention adopts a modular multi-branch network structure: a one-dimensional convolutional neural network is used for temporal feature extraction, and a lightweight two-dimensional CNN is used for frequency domain feature extraction. The computational complexity of the coupled feature network is controlled by a simplified design of fully connected layers and physical constraint layers. At the same time, the end-to-end training mode simplifies the data flow process, and feature selection and dimensionality reduction in the preprocessing stage further reduce the computational load of the model. The above design ensures that the model can meet the real-time requirements of the airborne control system of aero-engines while maintaining accuracy, and solves the pain points of high-fidelity physical models being computationally complex and difficult to deploy on air.

[0022] (5) The present invention uses a thermodynamic component-level model to generate original samples covering multiple working conditions, and combines parameter sensitivity analysis to screen key sensor variables. Model training can be completed without relying on large-scale, high-cost ground test or flight test data, which greatly reduces the cycle and cost of data acquisition. At the same time, the model has a certain tolerance for sensor failure or data loss, which reduces the excessive reliance on the number and accuracy of sensor deployment, and further reduces the hardware and maintenance costs in engineering applications. Attached Figure Description

[0023] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1This is a flowchart of a civil aviation engine thrust estimation method based on a multimodal convolution model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the flight envelope of a certain type of aero-engine provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the thrust fitting effect of each model provided in the embodiments of the present invention; Figure 4 This is a comparative diagram of the performance indicators of various models provided in the embodiments of the present invention; Figure 5 This is a diagram of the multimodal convolutional neural architecture provided in an embodiment of the present invention; Figure 6 This is a flowchart of the aero-engine thrust estimation process provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a civil aviation engine thrust estimation system based on a multimodal convolution model provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0025] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations (or steps) may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0026] Example 1 like Figure 1 The diagram shows a flowchart of a civil aircraft engine thrust estimation method 100 based on a multimodal convolution model provided in Embodiment 1 of the present invention. The method 100 specifically includes the following steps: S110: Acquire multi-source sensor time-series data and corresponding thrust label data of the engine within the entire flight envelope to construct the original dataset.

[0027] As a preferred embodiment, to ensure the effectiveness of model training, the training dataset for the thrust estimator should contain all state parameters of the entire flight envelope, such as... Figure 2 As shown. Although ground testing is accurate, it is costly and time-consuming. Therefore, this embodiment uses a thermodynamic component-level model to obtain raw data. The obtained dataset includes engine state parameters (multiple sensor signals) and thrust measurements under the corresponding states. The sensor signals specifically include parameters such as altitude, Mach number, fuel flow rate, low-pressure rotor speed, high-pressure rotor speed, total pressure ratio, high-pressure compressor outlet pressure, engine exhaust temperature, and turbine inlet temperature.

[0028] Furthermore, based on the original dataset, signals are selected through parameter sensitivity analysis to construct the original training dataset. Specifically, based on the influence of engine state parameters on thrust, Pearson correlation coefficient is used to screen time-series data of key variables from the original sample data. The time-series data of the screened key variables are then windowed to construct the original dataset. Let the i-th sample after windowing be... ,in For multi-channel timing input matrix, If the thrust scalar label is used, then the original dataset can be represented as: (1) in, Indicates the number of sensor channels. Indicates the window length. Indicates the number of samples.

[0029] Pearson correlation coefficient was used to screen candidate variables: (2) in, Indicates the first One variable, It represents the standard deviation.

[0030] S120: Perform data preprocessing on the original dataset to generate time-domain signals, frequency-domain signals, and coupling features based on engine physical mechanisms.

[0031] In a preferred embodiment, the original dataset is preprocessed to obtain three types of inputs: time-domain signal, frequency-domain signal, and coupling features, specifically including: The time-domain signal is subjected to denoising filtering and normalization to eliminate measurement noise and dimensional differences, thus obtaining the time-domain signal. ; Sensor signals for each time series Short-time Fourier Transform (STFT) is employed, and the time spectrum is extracted. Extracting the main frequency components and energy distribution characteristics yields the frequency domain signal. Among them, the sensor signals for each time sequence The Short-Time Fourier Transform (STFT) is used with the following formula: (3) A set of coupled features, c, is constructed based on domain knowledge, including linear combination features, nonlinear interaction features, and time-delay correlation features derived from the original parameters. For example, it includes calculating the ratio / difference of key parameters such as pressure and temperature (linear coupling), the parameters and their squares / products (nonlinear coupling), and introducing lagged terms or historical statistical features of the parameters (time-series coupling). Correlation analysis and principal component analysis are used to perform necessary screening and dimensionality reduction on the coupled features, controlling the feature dimensions and removing redundant information. Specifically, the coupled feature vector c is defined as a window statistic... Based on this, we construct coupled terms such as ratios, differences, products, squares, and lag terms: (4) The result obtained by piecing together: .

[0032] S130: Construct a multimodal convolutional neural network thrust estimation model, which includes a temporal convolutional neural network branch, a frequency domain convolutional neural network branch, and a coupled feature network branch.

[0033] S140: Extract time-domain features from the time-domain signal according to the time-domain convolutional neural network branch, extract frequency-domain features from the frequency-domain signal according to the frequency-domain convolutional neural network branch, and filter coupling features from the coupling features according to the coupling feature network branch.

[0034] In a preferred embodiment, extracting temporal features from the temporal signal based on the temporal convolutional neural network branch includes: The primary function of the temporal CNN branch is to process the preprocessed temporal signal sequence and extract temporal feature patterns reflecting thrust dynamics. The temporal CNN branch receives a preprocessed, standard-sized, multi-channel time series as input. Each channel corresponds to a sequence of measurable parameters for the aero-engine. The temporal branch employs a one-dimensional convolutional neural network structure, integrating residual connections and attention mechanisms to improve feature extraction performance.

[0035] Specifically, the temporal convolutional neural network consists of four concatenated convolutional residual blocks, including multi-scale convolutional layers, nonlinear activation, and skip connections. The first convolutional layer performs independent convolution operations on each sensor channel, using kernel sizes of 3, 5, and 7 to construct the multi-scale convolutional blocks. By using different convolutional kernels in parallel, short-term fluctuations, periodic trends, and long-term variation features of the signal are extracted simultaneously. The outputs of the multi-scale convolutions are concatenated along the channel dimension. After ReLU linear activation and batch normalization, the convolution results are adjusted in dimension by a fully connected layer and superimposed on the original input of the residual branch, achieving residual connections. Residual connections enable the neural network to train more efficiently while preserving the basic information of the original signal. The four convolutional layers in the temporal CNN branch have 64, 128, 256, and 128 convolutional channels, respectively. Adaptive average pooling is applied after each convolutional layer to map the features of each channel to a fixed length to compress the sequence length. This invention introduces a channel attention mechanism into the residual blocks, using the SENet (Squeeze-Excitation) module to weight the features output by each residual block. The output of the residual blocks is globally averaged over time to obtain the statistics for each channel. These statistics are then processed through a two-layer fully connected network (dimensionality reduction and re-enlargement) and activated by the Sigmoid function, generating channel weights between 0 and 1, which are then multiplied by the feature tensor. An attention module adaptively represents the signal channel that contributes the most to the thrust. After the above convolution and attention processing, the temporal CNN branch outputs a 128-dimensional temporal feature vector.

[0036] When the kernel size is kb, the output is: (5) And concatenate the outputs of each branch: .

[0037] In a preferred embodiment, extracting frequency domain features from the frequency domain signal based on the frequency domain convolutional neural network branch includes: The primary function of the frequency-domain CNN branch is to process the spectrum / time-frequency image of each sensor signal and extract frequency-domain features reflecting the periodicity of thrust. The frequency-domain CNN branch employs a two-dimensional convolutional neural network structure for feature extraction from the time-frequency image. Similar to the time-domain CNN branch, the frequency-domain CNN branch consists of multiple layers of two-dimensional convolutional residual blocks. The number of channels in the convolutional layers is 32, 64, 128, and 64. Convolutional kernels of different sizes slide along both the frequency and time dimensions, simultaneously capturing the pattern of the spectrum in the frequency direction and the evolutionary features in the time direction. Texture features at different scales are extracted through multi-scale convolution, and the outputs are concatenated along the channel dimension. Based on the characteristics of spectral data, this invention introduces frequency attention and joint time-frequency attention mechanisms into the frequency-domain CNN. By introducing a frequency attention module, the convolutional features are globally pooled along the time axis to obtain the feature mean at each frequency bin. A 1×1 convolution is used to generate weights for the frequency dimension, representing thrust-related components within a specific frequency range. A joint time-frequency attention module is designed to globally pool convolutional features in both the frequency and time dimensions, generating channel weights that are distributed with frequency and those that vary with time. These two types of weights are then passed through a 1×1 convolutional network to produce weight sequences of the same length as the original number of channels C, and normalized using a sigmoid function. Each channel of the original feature is multiplied by the time-attention weights and frequency-attention weights, achieving feature recalibration along both dimensions. Based on this, the average and maximum values ​​of the features in the channel dimension are calculated, stacked to form a two-channel time-frequency plane, and then spatial attention is extracted using a 7×7 convolution to obtain the importance of specific time-frequency locations. The calculated time-frequency attention weight matrix is ​​multiplied point-by-point with the original features to enhance the expressive power of the convolutional features. Each residual block of the frequency-domain CNN branch is adjusted by the above attention module, then combined with and activated by the residual branch, resulting in an output feature dimension of 128 dimensions, the same as the time-domain CNN branch.

[0038] Multi-scale two-dimensional convolution is used for joint time-frequency feature extraction. The convolution output is: (6) And merge the branch results: .

[0039] In a preferred embodiment, filtering coupling features from the coupling features based on the coupling feature network branches includes: The coupled feature network branch is used to process the parameter combination features obtained from physical mechanism analysis, and to deeply explore the intrinsic relationship between individual state parameters and thrust. The feature vectors are multidimensional and include various types such as linear coupling, nonlinear coupling, and time-delay coupling. Following the construction approach of embedding physical information into neural networks, the main body of the coupling feature network adopts a fully connected neural network, combined with physical constraint fusion, feature importance estimation, and hierarchical feature selection modules, to progressively deepen the representation capability of coupling features on thrust.

[0040] The coupled feature network first performs dimensionality upscaling on the original coupled features through an input projection layer, projecting the input dimension of 50 to the hidden layer dimension of 128. The data stream is then processed using BatchNorm normalization, ReLU activation, and Dropout regularization. The normalized data stream is processed sequentially according to the physical constraint layer, feature importance evaluation, and hierarchical feature extraction.

[0041] Input projection: .

[0042] As a preferred embodiment, in order to ensure that the features extracted by the model and the subsequent training process conform to the laws of engine thermodynamics, the present invention introduces a physical constraint layer in the coupled feature network branch. This physical constraint layer applies transformations and weighted fusions to the input features based on a variety of predefined physical constraint relationships, ensuring that the output features have the representational ability after constraint enhancement.

[0043] The physical constraint layer includes thermodynamic constraints, momentum constraints, mass flow rate constraints, and energy conservation constraints. For each constraint, the physical constraint layer defines a small neural network to implement the constraint transformation function, setting learnable weight parameters to measure the importance of each constraint. For example, thermodynamic constraints use a two-layer fully connected network to compress and then expand the input features, and use a Sigmoid function to normalize the output in the 0-1 range to simulate the normalization constraint of the relationship between variables such as temperature and pressure; momentum constraints use Tanh activation to emphasize the symmetric relationship between velocity and flow rate; mass flow rate constraints use the Softplus activation function to ensure the output is positive, meeting the mass conservation requirement; and energy balance constraints use an identity mapping to ensure the linear relationship of energy conservation. The physical constraint layer multiplies the transformed features of the above constraints with their corresponding weights to obtain the feature vector under each constraint, and then uses an attention mechanism to fuse the outputs of different constraints again. Specifically, multiple constraint outputs are stacked along the newly added dimension, the average value of each constraint output in the feature dimension is calculated, the attention weight coefficients of each constraint are obtained through Softmax normalization, and finally, a weighted sum is obtained to obtain the fused physical constraint features. The fused features are added to the original output features to form a residual, ensuring that the original feature information is preserved while embedding the influence of physical constraints.

[0044] Since there are many types of coupling features, this invention introduces a feature importance module to adaptively adjust the contribution of different coupling features to thrust prediction. This module uses an attention mechanism to measure the importance of each feature and performs weighted processing.

[0045] The fused feature vector is passed through a two-layer fully connected network, and the sigmoid function is used to output the weight coefficient of each feature. Each feature is then multiplied by its corresponding weight. Based on the weight magnitude, unimportant features are suppressed while important features are highlighted. Since the coupled features themselves can be divided into several groups based on their characteristics, such as linear and nonlinear classes, this module first performs group-level feature importance evaluation before proceeding with the aforementioned fused feature evaluation process. Therefore, this module ultimately outputs a reweighted feature vector. Through the feature importance module, the network can automatically learn the relative importance of each coupled feature under different operating conditions, improving the model's robustness to different states.

[0046] After the aforementioned physical constraints and importance assessment process, the coupled features enter the backbone processing network for further extraction of higher-order combined features. The backbone network consists of multiple fully connected residual blocks connected in series, with each residual block mapping features from the previous layer to a new hidden space. The hidden layer's dimensionality is designed using a bottleneck structure that first increases and then decreases, progressively filtering the higher-order interaction patterns of the coupled features.

[0047] Before the final output layer, this invention introduces a hierarchical feature extractor to further fuse features from different levels and enhance their expressive power. The hierarchical feature extractor defines the feature vector of the last layer of the backbone network as the highest-level feature, while also including features from several lower levels. The extractor reduces the dimensionality of features at each level through fully connected layers and uses a hierarchical attention fusion mechanism to calculate the weights of each level. This mechanism enables the model to comprehensively utilize information from various levels, from single-parameter to multi-parameter combinations, while considering the impact of features of different complexities on thrust.

[0048] The output dimension of the coupled feature network is 64-dimensional, which prepares for subsequent fusion with the time domain and frequency domain branches.

[0049] Physical constraint fusion: for each type of constraint Establish transformation and weighted fusion: (7) Attention weights: ; Fusion results: ; Feature importance evaluation: Generate feature weights and apply them: (8) Hierarchical feature extraction: Constructing features layer by layer and performing hierarchical attention fusion: (9) In summary, this invention jointly models the original time-domain sensor sequence information, frequency-domain time-spectrum information, and multi-parameter coupled features, and achieves information complementarity through a cross-modal attention fusion mechanism. Table 1 presents the ablation experiment results of this multimodal CNN model architecture.

[0050] Table 1 Among them, the root mean square error (RMSE) is: ; Mean Absolute Error (MAE): ; Coefficient of determination (R²): .

[0051] Analysis of the results shows that the single-time-domain branch has a strong fitting ability (RMSE=123), indicating a significant correlation between thrust and sensor time-series dynamics. The single-time-domain branch can track the main dynamic trends, but exhibits systematic deviations in strongly coupled conditions and nonlinear segments. The single-frequency-domain branch has a significantly larger error (RMSE=459), reflecting that relying solely on spectral energy distribution is insufficient to fully characterize the non-stationary transient process of engine thrust. The single-coupled branch outperforms the single-frequency-domain branch but still performs poorly (RMSE=382), indicating that physically derived features can provide effective priors but cannot replace the original fine-grained time-series information. The significantly larger RMSEs of the single-frequency-domain branch and the single-coupled branch suggest that amplitude deviation and phase lag are prone to occur in dynamically changing segments. The pairwise fusion architecture can reduce some errors, but it exhibits instability in different combinations (e.g., coupling + temporal path RMSE=206), indicating that simple splicing or linear combination is difficult to stably model cross-modal conflict and complementarity relationships; the multimodal CNN of this invention reduces the RMSE by about 53.7% relative to the best single branch (123) and the MAE by about 26.9% relative to the best double branch (52), and the R² is improved to 0.999, indicating that cross-modal attention fusion can significantly reduce systematic bias and improve fitting consistency.

[0052] This invention introduces a coupling feature module, which on the one hand expresses key state relationships with features such as ratios, products, and powers with clear physical meanings; on the other hand, it expresses statistical coupling between multiple sensors with interactive features such as the upper triangular correlation matrix, thereby introducing prior knowledge.

[0053] The cross-modal fusion module of this invention performs bidirectional attention interaction within a unified feature space and can dynamically adjust the contribution of the signal path (time domain + frequency domain) and the coupling path through an adaptive weight learning mechanism, thereby achieving more robust thrust estimation under different operating conditions.

[0054] S150: The extracted time-domain features are fused with the frequency-domain features to obtain signal fusion features, and the signal fusion features are fused with the coupling features to obtain multimodal fusion features.

[0055] In a preferred embodiment, after acquiring the time-domain features, frequency-domain features, and coupling features respectively, a dual-path cross-modal fusion module is designed to fuse the original signal path (time domain + frequency domain) with the coupling feature path to generate a unified thrust estimation feature. The fusion module mainly includes two parts: signal path internal fusion and cross-modal fusion.

[0056] To ensure effective fusion of time-domain and frequency-domain features, this invention designs three fusion methods: concat, add, and attention fusion, to obtain fused signal features. Taking attention fusion as an example, firstly, the time-domain feature vector and the frequency-domain feature vector are concatenated along their feature dimensions to construct a long vector. The concatenated vector is then input into a two-layer attention weight network, which outputs two normalized weights, corresponding to the time-domain path and the frequency-domain path, respectively. The time-domain and frequency-domain features are multiplied by their corresponding weight coefficients and reconnected to obtain the weighted fused feature representation. Compared to the concatenation and add methods, the attention fusion method can adaptively adjust the proportion of the two signal features according to the operating conditions. For example, when a specific frequency mode has a more significant impact on thrust, the frequency-domain path will receive a larger attention weight and account for a larger proportion in the fused features. The fusion module within the signal path outputs a 256-dimensional feature vector as a comprehensive representation of the original signal path.

[0057] Cross-modal fusion further integrates the signal fusion features with the output features of the coupled feature network to obtain the final multimodal features used for thrust estimation. This invention introduces cross-modal attention and adaptive weight adjustment mechanisms into the dual-path fusion, making full use of the complementary information of the two types of features and dynamically balancing their respective contributions.

[0058] First, the feature alignment module maps the signal fusion features and coupled features to the same dimensional space. When the initial dimensions are different, the feature alignment module increases the dimensionality of the low-dimensional features. The signal fusion features (256 dimensions) and coupled features (64 dimensions) are converted to a uniform length of 256 dimensions. After feature alignment, a two-path cross-modal bidirectional attention interaction is performed. On one hand, self-attention is calculated on the signal fusion features, utilizing the correlation between the acquired signals to characterize the internal patterns of the signal features. On the other hand, self-attention is also calculated on the coupled features to extract their internal correlations. By defining the signal features as queries and the coupled features as key / value pairs, cross-modal information acquisition is achieved. Furthermore, a multi-head attention mechanism is employed, linearly transforming the input features into several sets of query, key, and value vectors, calculating the attention weight matrix, and weighted summing to obtain the "attended" feature representation. Specifically, the signal features are enhanced through self-attention, and the coupled features are enhanced through self-attention; simultaneously, cross-modal attention is applied to the coupled features by the signal features to obtain coupled-enhanced signal features. These three pieces of information are fused together. The original signal features are added to its self-attention output and the cross-modal enhancement component to obtain new signal features. The original coupling features are added to its self-attention output to obtain new coupling features. Then, layer normalization is applied to obtain enhanced signal features and enhanced coupling features, respectively. Through the above process, the signal path features already contain certain information about the coupling path, and the coupling path features are also self-reinforced.

[0059] An adaptive weight learning module is used to perform weighted fusion of the two feature paths. This module automatically calculates the fusion weights of the signal and coupling features based on the joint representation of the current features using an attention network. Through adaptive weight learning, the model can automatically adjust the contributions of the signal and coupling paths to thrust estimation according to the current engine state, fully leveraging the advantages of multimodal fusion. The final feature vector obtained after fusion is the multimodal integrated feature from input to output layer.

[0060] Cross-modal attention interaction (taking signal query coupling as an example): (10) Adaptive weight learning: (11) S160: Input the multimodal fusion features into the multimodal convolutional neural network thrust estimation model to estimate the engine thrust value, and output the engine thrust estimate.

[0061] In a preferred embodiment, the fused multimodal features are input into the thrust estimation output layer to predict the engine's thrust value. The output layer is essentially a regression model used to map high-dimensional features to a scalar thrust prediction. This invention employs a multilayer perceptron (MLP) as the output layer to achieve nonlinear regression. Specifically, after fusing the feature vectors, several fully connected layers are connected to progressively reduce the dimensionality and finally output a single predicted value. The output layer includes two hidden layers with neuron sizes of 128 and 64, respectively. Each layer is followed by BatchNorm and ReLU activation, as well as Dropout to prevent overfitting. The final layer is a linear regression unit that maps the output of the previous layer to a single neuron, taking its value as the estimated thrust. Therefore, the overall mapping relationship of the output layer can be described as follows: (12) in, For multimodal fusion features, This represents the composite mapping function of the output layer. For example, when using a two-layer MLP, it can be represented as: (13) in, For weight matrix / vector, For bias, The activation function is ReLU. The parameters of the output layer will be learned during the training phase by minimizing the error between the predicted value and the actual thrust.

[0062] Based on the above embodiments, the present invention achieves the following beneficial effects: (1) This invention achieves complementary enhancement of information in different dimensions through multimodal joint modeling of time-domain signals, frequency-domain signals, and coupling features based on physical mechanisms, thus avoiding the shortcomings of incomplete representation of information in a single modality. By combining cross-modal attention fusion and adaptive weight allocation, the contribution relationship of each modality to thrust is accurately explored, significantly reducing systematic bias and random error. Experimental data show that the root mean square error (RMSE) of this scheme is reduced by about 53.7% compared with the optimal single-branch model, the mean absolute error (MAE) is reduced by about 26.9% compared with the optimal double-branch model, and the coefficient of determination (R²) reaches 0.999. The fitting accuracy and numerical accuracy of thrust estimation far exceed those of traditional single-modal or simple fusion schemes.

[0063] (2) This invention embeds physical constraints such as thermodynamics, momentum, mass flow rate and energy conservation in the coupled feature network branches. Through learnable transformation and residual fusion, it ensures that the estimation results conform to the core physical laws of the engine and avoids the "physical inconsistency" problem that is easy to occur in pure data-driven methods. On the other hand, the bidirectional attention interaction mechanism effectively suppresses cross-modal information conflict. The adaptive weight allocation can dynamically adjust the contribution of each path according to the working conditions. It can still maintain stable performance in complex scenarios such as strong transient and emergency working conditions. It solves the problem of amplitude deviation and phase lag that is easy to occur in traditional methods under special working conditions and improves the engineering reliability of thrust estimation.

[0064] (3) This invention takes into account the differences in working conditions under different Mach numbers, altitudes, and thrust requirements within the entire envelope by using multi-scale convolution, parameter sensitivity analysis, and coupled feature dimensionality reduction. Compared with traditional data-driven methods that rely on specific datasets, this scheme can achieve working condition transfer through physical constraints and cross-modal complementarity without covering all extreme working condition samples. It has a strong adaptability to unseen transient processes and boundary working conditions, and effectively solves the problems of poor generalization ability and high dependence on training datasets in traditional methods.

[0065] (4) This invention adopts a modular multi-branch network structure: a one-dimensional convolutional neural network is used for temporal feature extraction, and a lightweight two-dimensional CNN is used for frequency domain feature extraction. The computational complexity of the coupled feature network is controlled by a simplified design of fully connected layers and physical constraint layers. At the same time, the end-to-end training mode simplifies the data flow process, and feature selection and dimensionality reduction in the preprocessing stage further reduce the computational load of the model. The above design ensures that the model can meet the real-time requirements of the airborne control system of aero-engines while maintaining accuracy, and solves the pain points of high-fidelity physical models being computationally complex and difficult to deploy on air.

[0066] (5) The present invention uses a thermodynamic component-level model to generate original samples covering multiple working conditions, and combines parameter sensitivity analysis to screen key sensor variables. Model training can be completed without relying on large-scale, high-cost ground test or flight test data, which greatly reduces the cycle and cost of data acquisition. At the same time, the model has a certain tolerance for sensor failure or data loss, which reduces the excessive reliance on the number and accuracy of sensor deployment, and further reduces the hardware and maintenance costs in engineering applications.

[0067] Example 2 The multimodal convolutional neural network thrust estimation model proposed in this invention includes a temporal CNN branch (outputting 128-dimensional features), a frequency domain CNN branch (outputting 128-dimensional features), a coupled feature network (outputting 64-dimensional features), a dual-path cross-modal feature fusion module (fusing two types of features into 256-dimensional features), and a thrust regression output layer (outputting thrust scalar). The parameters of the entire model are trained and optimized through a unified loss function, enabling it to learn the mapping relationship of aero-engine thrust from multi-source data.

[0068] In this embodiment, the system input is a multi-source sensor time series with 8 channels, and the set of optional parameters is as follows: To ensure repeatability and fairness of the comparison, the experiment used the same dataset, dividing it into training, validation, and test datasets in a 0.7:0.15:0.15 ratio. Based on this, six representative thrust estimation models were selected as the control group: Temporal Convolutional Attention Network (AttentionTCN), Graph ThrustNet, Wavelet ThrustNet, Federated ThrustNet, Transfer ThrustNet, and the Multimodal CNN of this invention. The training configurations are shown in Table 2.

[0069] Table 2 During the training phase, a joint loss function is constructed based on the thrust prediction error and physical constraints, and the network parameters are iteratively optimized.

[0070] Basic predicted loss (using Huber as an example): (14) Physical constraint loss (example: smoothness vs. boundary constraints): (15) Joint losses: (16) Depend on Figure 2Analysis reveals that the overall thrust of the test set exhibits a non-linear upward trend from low to high, with a significant rapid increase at the end of the sequence. In this data pattern with strong transient characteristics, if the model only learns the global trend and cannot accurately depict the steep slope changes and local curvature at the end, it will produce significant deviations in the high-thrust segment. The prediction curves of MultimodalCNN, AttentionTCN, and TransferThrustNet in this invention are closer to the original thrust curve in overall shape, especially maintaining good slope tracking in the rapid upward segment at the end, indicating a more thorough depiction of key transient structures. In contrast, while GraphThrustNet and WaveletThrustNet can reproduce the overall upward trend, they show more significant deviation accumulation in the middle and later stages. This deviation is reflected in the systematic increase of RMSE and MAE in pointwise error statistics. Although the curve of FederatedThrustNet also shows an upward trend, it deviates more significantly from the original thrust in multiple intervals, indicating that the predicted trend is correct but the amplitude calibration is insufficient, which is consistent with its significantly deteriorated error indicators.

[0071] Figure 3 The training and inference results of each model are presented quantitatively. The performance metrics of MultimodalCNN, AttentionTCN, and TransferThrustNet gradually deteriorate, indicating that models relying solely on strong temporal modeling and attention mechanisms can already fit the main dynamic patterns of this task well, but some residual error still exists. GraphThrustNet and WaveletThrustNet show significantly amplified errors, demonstrating that modeling methods relying solely on graph structure or transform domain features are less effective than end-to-end multimodal representations in characterizing the key nonlinear segments of thrust under these data conditions. The error is reflected not only in the average bias but also in the peak bias between the high-thrust and transient segments. FederatedThrustNet has the worst performance, indicating a significant systematic inconsistency between its predictions and the actual thrust. Compared to AttentionTCN, this invention reduces MAE by 5.77% and RMSE by 4.17%; compared to TransferThrustNet, MAE decreases by 10.91% and RMSE by 8.00%; compared to GraphThrustNet, RMSE decreases by 74.16%; compared to WaveletThrustNet, RMSE decreases by 71.49%; and compared to FederatedThrustNet, RMSE decreases by 93.83%. These differences indicate that this invention simultaneously improves both prediction accuracy and prediction robustness.

[0072] Based on the above embodiments, the present invention achieves the following beneficial effects: This invention achieves complementary enhancement of information from different dimensions through multimodal joint modeling of time-domain signals, frequency-domain signals, and coupling features based on physical mechanisms, avoiding the shortcomings of incomplete representation of single-modal information. By combining cross-modal attention fusion and adaptive weight allocation, the contribution relationship of each mode to thrust is accurately mined, significantly reducing systematic bias and random errors. Experimental data show that the root mean square error (RMSE) of this scheme is reduced by approximately 53.7% compared to the optimal single-branch model, the mean absolute error (MAE) is reduced by approximately 26.9% compared to the optimal two-branch model, and the coefficient of determination (R²) reaches 0.999. The fitting accuracy and numerical accuracy of thrust estimation far exceed those of traditional single-modal or simple fusion schemes.

[0073] Example 3 Figure 7 This is a schematic diagram of the structure of a civil aviation engine thrust estimation system based on a multimodal convolution model provided in Embodiment 3 of the present invention, as shown below. Figure 7 As shown, the device includes: The data acquisition module 710 is used to acquire multi-source sensor time-series data and corresponding thrust label data of the engine within the entire flight envelope, and to construct the raw dataset. The data preprocessing module 720 is used to preprocess the original dataset to generate time-domain signals, frequency-domain signals, and coupling features based on engine physical mechanisms. The thrust estimation model building module 730 is used to build a multimodal convolutional neural network thrust estimation model, which includes a temporal convolutional neural network branch, a frequency domain convolutional neural network branch, and a coupled feature network branch. The feature filtering module 740 is used to extract time-domain features from the time-domain signal according to the time-domain convolutional neural network branch, extract frequency-domain features from the frequency-domain signal according to the frequency-domain convolutional neural network branch, and filter coupling features from the coupling features according to the coupling feature network branch; The feature fusion module 750 is used to fuse the extracted time-domain features with the frequency-domain features to obtain signal fusion features, and to fuse the signal fusion features with the coupling features to obtain multimodal fusion features; The engine thrust estimation module 760 is used to input the multimodal fusion features into the multimodal convolutional neural network thrust estimation model to estimate the engine thrust and output the engine thrust estimate.

[0074] The civil aviation engine thrust estimation system based on a multimodal convolution model provided in this embodiment of the invention can execute the civil aviation engine thrust estimation method based on a multimodal convolution model provided in any of the embodiments of the invention described above. It has the corresponding functions and beneficial effects of executing the civil aviation engine thrust estimation method based on a multimodal convolution model. For detailed process, please refer to the relevant operations of the civil aviation engine thrust estimation method based on a multimodal convolution model in the foregoing embodiments.

[0075] Example 4 Figure 8 This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0076] like Figure 8 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from the storage module 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0077] Multiple components in electronic device 10 are connected to I / O interface 15, including: input module 16, such as keyboard, mouse, etc.; output module 17, such as various types of displays, speakers, etc.; storage module 18, such as disk, optical disk, etc.; and communication module 19, such as network card, modem, wireless transceiver, etc. Communication module 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0078] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the civil aviation engine thrust estimation method based on a multimodal convolutional model described above.

[0079] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0080] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A thrust estimation method for civil aircraft engines based on a multimodal convolution model, characterized in that, include: Acquire multi-source sensor time-series data and corresponding thrust label data of the engine across the entire flight envelope to construct the raw dataset; The original dataset is preprocessed to generate time-domain signals, frequency-domain signals, and coupling features based on engine physical mechanisms. A multimodal convolutional neural network thrust estimation model is constructed, which includes a temporal convolutional neural network branch, a frequency domain convolutional neural network branch, and a coupled feature network branch. The time-domain features are extracted from the time-domain signal according to the time-domain convolutional neural network branch, the frequency-domain features are extracted from the frequency-domain signal according to the frequency-domain convolutional neural network branch, and the coupling features are filtered from the coupling features according to the coupling feature network branch. The extracted time-domain features are fused with the frequency-domain features to obtain signal fusion features, and the signal fusion features are fused with the coupling features to obtain multimodal fusion features; The multimodal fusion features are input into the multimodal convolutional neural network thrust estimation model to estimate the engine thrust value, and the engine thrust estimate is output.

2. The method according to claim 1, characterized in that, Acquire multi-source sensor time-series data and corresponding thrust label data of the engine across the entire flight envelope to construct the raw dataset, including: Based on a thermodynamic component-level model, raw sample data covering multiple operating conditions is generated. The raw sample data includes engine state parameters and corresponding thrust labels. Based on the degree of influence of each state parameter on thrust, time series data of key variables are selected from the original sample data; The time-series data of the selected key variables are windowed to construct the original dataset.

3. The method according to claim 1, characterized in that, The original dataset is preprocessed to generate time-domain signals, frequency-domain signals, and coupled features based on engine physics mechanisms, including: The time-series data in the original dataset are cleaned and standardized to obtain the time-domain signal; The time-domain signal is subjected to a short-time Fourier transform to obtain the corresponding time spectrum, and frequency domain features are extracted from the time spectrum to obtain the frequency domain signal; Based on the engine physics mechanism, an initial feature set containing linear coupling features, nonlinear interaction features, and temporal coupling features is constructed from the original dataset. The initial feature set is then filtered and dimensionality reduced to remove redundancy, resulting in coupling features.

4. The method according to claim 3, characterized in that, The extracted time-domain features are fused with the frequency-domain features to obtain signal fusion features, and the signal fusion features are fused with the coupling features to obtain multimodal fusion features, including: The time-domain features and frequency-domain features are dynamically fused using an attention weight learning mechanism to obtain signal fusion features; The signal fusion features and the coupled features are fused across modes through bidirectional attention interaction and adaptive weight allocation to generate the multimodal fusion features.

5. The method according to claim 4, characterized in that, The signal fusion features and the coupled features are fused across modes through bidirectional attention interaction and adaptive weight allocation to generate the multimodal fusion features, including: Map the signal fusion features and the coupling features to the same feature space; Bidirectional attention interaction is performed on the signal fusion features and coupling features mapped to the same feature space to obtain the signal fusion features and coupling features enhanced by interaction. Based on the enhanced signal fusion features and coupling features, the multimodal fusion features are generated through an adaptive weight allocation mechanism.

6. The method according to claim 1, characterized in that, The multimodal fusion features are input into the multimodal convolutional neural network thrust estimation model to estimate engine thrust values, and the engine thrust estimate is output, including: The multimodal fusion features are input into the multilayer perceptron regression output layer of the model; The multimodal fusion features are nonlinearly mapped by at least one fully connected hidden layer in the regression output layer of the multilayer perceptron to output the engine thrust estimate.

7. The method according to claim 1, characterized in that, The structures of the temporal convolutional neural network branch, the frequency domain convolutional neural network branch, and the coupled feature network branch are as follows: The temporal convolutional neural network branch adopts a network structure that includes multi-scale one-dimensional convolution, residual connections, and channel attention mechanisms. The frequency domain convolutional neural network branch adopts a network structure that includes multi-scale two-dimensional convolution, residual connections, and a time-frequency joint attention mechanism. The coupled feature network branch adopts a network structure with a fully connected network as the main body and an integrated feature importance evaluation module.

8. The method according to claim 7, characterized in that, The coupled feature network branch also includes a physical constraint layer: The physical constraint layer is used to transform and weight and fuse physical constraint relationships constructed based on at least one of the laws of thermodynamics, conservation of momentum, conservation of mass, and conservation of energy through a learnable network layer, and to combine the fused constraint enhancement features with the original coupling features.

9. The method according to claim 1, characterized in that, The training process of the multimodal convolutional neural network thrust estimation model includes: The parameters of a multimodal convolutional neural network thrust estimation model are iteratively optimized based on a joint loss function, which includes a basic prediction loss term and a physical constraint loss term. The basic prediction loss term is mean squared error loss, mean absolute error loss, or Huber loss. The physical constraint loss term is constructed based on at least one of the laws of thermodynamics, the law of conservation of momentum, the law of conservation of mass, and the law of conservation of energy.

10. A thrust estimation system for civil aircraft engines based on a multimodal convolution model, characterized in that, include: The data acquisition module is used to acquire multi-source sensor time-series data and corresponding thrust label data of the engine within the entire flight envelope, and to construct the raw dataset; The data preprocessing module is used to preprocess the original dataset to generate time-domain signals, frequency-domain signals, and coupling features based on engine physical mechanisms. The thrust estimation model building module is used to construct a multimodal convolutional neural network thrust estimation model, which includes a temporal convolutional neural network branch, a frequency domain convolutional neural network branch, and a coupled feature network branch. The feature filtering module is used to extract time-domain features from the time-domain signal according to the time-domain convolutional neural network branch, extract frequency-domain features from the frequency-domain signal according to the frequency-domain convolutional neural network branch, and filter coupling features from the coupling features according to the coupling feature network branch; The feature fusion module is used to fuse the extracted time-domain features with the frequency-domain features to obtain signal fusion features, and to fuse the signal fusion features with the coupling features to obtain multimodal fusion features; The engine thrust estimation module is used to input the multimodal fusion features into the multimodal convolutional neural network thrust estimation model to estimate the engine thrust and output the engine thrust estimate.

Images