A method for predicting body freedom mode flutter based on transfer rate function and flight condition

Abstract

The application provides a body freedom degree flutter prediction method based on a transfer rate function and a flight working condition, and belongs to the technical field of aeroelasticity. In the method, firstly, a simulation data covering a full flight envelope is generated by using a modified high-confidence aircraft rigid-elastic coupling model, and time-domain acceleration responses under atmospheric turbulence excitation are converted into a transfer rate function; then, a two-channel one-dimensional convolutional neural network architecture including a main channel of the transfer rate function and an auxiliary channel of the flight working condition is constructed to synchronously predict a double stability index; finally, a transfer learning strategy of large-scale simulation training and a small amount of flight test data fine-tuning is adopted, a small amount of real flight test data is used to fine-tune the model, and accurate prediction of a real aircraft flutter stability boundary is realized. The application realizes high-precision and high-reliability flutter prediction in the full flight envelope range, can provide early warning, and has important engineering application value in the aviation field.

Description

Technical Field

[0001] This invention belongs to the field of aeroelasticity technology and relates to a method for predicting body degree-of-freedom flutter based on transfer rate function and flight conditions. Background Technology

[0002] Flutter is a self-excited vibration phenomenon that occurs in aircraft structures under the influence of airflow. When the flight speed reaches the critical flutter velocity, the structure will transition from stable vibration to divergent vibration, leading to structural failure in a very short time. It is one of the most dangerous aeroelastic problems in the aerospace field. Among them, volume-degree-of-freedom flutter is a special form of flutter generated by the coupling of rigid-elastic coupling characteristics between the rigid body's short-period modes and elastic modes in aircraft. Accurately predicting the flutter stability boundary of an aircraft is of vital importance for ensuring flight safety, optimizing aircraft design, and guiding flight testing.

[0003] Traditional flutter prediction methods are mainly divided into two categories: theoretical analysis methods and experimental methods. Theoretical analysis methods include frequency domain methods, time domain methods, and energy methods, which rely on accurate structural dynamics and aerodynamic models. However, real aircraft structures are complex, with numerous nonlinear factors and uncertainties. An unavoidable "model mismatch" problem exists between theoretical models and actual physical systems, leading to deviations between theoretical predictions and actual conditions. Experimental methods mainly include wind tunnel testing and flight testing. Wind tunnel testing can simulate flight conditions in a controlled environment, but it cannot completely replicate the real flight environment and structural state. Flight testing is the most reliable flutter verification method, but it is extremely costly and risky, and can only be conducted under subcritical conditions, unable to directly obtain the critical flutter point.

[0004] In recent years, data-driven machine learning methods have provided new insights into flutter prediction. These methods analyze structural vibration response data to learn the mapping relationship between the system's dynamic characteristics and stability, without relying on precise physical models. Flutter prediction methods that use the transferability function as input to machine learning have attracted widespread attention due to their advantages, such as being unaffected by the form of excitation and being sensitive to changes in structural characteristics. The transferability function reflects the ratio of responses between different measurement points on the structure, effectively characterizing the dynamic transferability of the system.

[0005] However, existing flutter prediction methods based on the transferability function still have the following core technical shortcomings: (a) Limitations of single-channel input: Most models adopt a single-channel network structure, using only the transfer rate function as input, completely ignoring the decisive influence of flight conditions (altitude, speed) on the dynamic characteristics of the system. Aerodynamic stiffness and aerodynamic damping change nonlinearly with flight altitude and speed. A single transfer rate function cannot characterize the system characteristics within the entire flight envelope, resulting in extremely poor model generalization ability, and the model is only effective under specific conditions.

[0006] (b) High-dimensional input is inefficient: Directly using the entire transfer rate function as network input results in data dimensions of hundreds or even thousands of dimensions, with a lot of redundant information, leading to long network training time, low computational efficiency, and a high risk of overfitting, making it difficult to apply in engineering.

[0007] (c) The model mismatch problem is serious: if only simulation data is used to train the model, the model's predictive performance on actual flight data will often be significantly reduced because the simulation model cannot fully reproduce the manufacturing tolerances, assembly errors and material property differences of the real aircraft, and it cannot be directly used for engineering practice.

[0008] (d) Poor reliability of single-indicator predictions: Most predictions only forecast a single flutter critical dynamic pressure index, lacking characterization of the system's dynamic characteristic evolution process. In complex aeroelastic coupling situations, single-indicator predictions may be misjudged, failing to provide early warning and physical explanation. Summary of the Invention

[0009] The technical problem this invention aims to solve is that existing data-driven flutter prediction methods suffer from fundamental defects such as the inability of single-channel input to characterize the full flight envelope, low computational efficiency of high-dimensional input, severe mismatch between simulation and real system models, and poor reliability of single-index prediction. These deficiencies fail to meet the high-precision, high-reliability, and real-time flutter stability boundary prediction requirements across the entire flight envelope in aeronautical engineering. This invention proposes a flutter prediction method based on the transfer rate function and flight conditions. This method embeds the transfer rate function and flight conditions into a dual-channel neural network. Through three core innovations—a dual-channel feature fusion architecture, joint prediction of dual-dimensional stability indices, and a transfer learning strategy combining simulation pre-training and flight test fine-tuning—it systematically solves the problems of insufficient generalization ability, low computational efficiency, severe model mismatch, and poor prediction reliability inherent in existing technologies.

[0010] The technical solution is as follows: A body-degree-of-freedom flutter prediction method based on transfer rate function and flight conditions includes the following steps: S1: Constructing a full flight envelope simulation dataset: For each flight condition point, atmospheric turbulence excitation is applied based on a high-confidence aircraft rigid-elastic coupling model to perform time-domain dynamic response simulation, and time-domain acceleration response signals of each measurement point in key parts are obtained; the time-domain acceleration response signals are processed to calculate the transfer rate function between two selected measurement points; based on the bistability index of the transfer rate function, a full flight envelope simulation dataset is formed. S2: Dataset stratification: The full flight envelope simulation dataset is randomly shuffled according to flight conditions and divided into training set, validation set and test set to ensure that each dataset contains the same distribution of flight conditions; S3: Establish a dual-channel one-dimensional convolutional neural network model: Establish a dual-channel one-dimensional convolutional neural network model that includes the main channel of the transfer rate function and the auxiliary channel of flight conditions; S4: Perform model training and performance validation: Train the model end-to-end using the training set; evaluate the unbiased performance of the trained model using the test set; S5: Fine-tuning model parameters based on transfer learning: Fine-tuning model parameters using a small amount of flight test data to achieve adaptive modeling from the simulation domain to the real domain; S6: Utilize a finely tuned dual-channel one-dimensional convolutional neural network model for online flutter prediction and multi-level early warning.

[0011] Specifically, S1 is: Based on the modified high-confidence aircraft rigid-elastic coupling model, different altitudes are selected within the entire flight envelope. With speed For each of the combined flight operating conditions, a time-domain dynamic response simulation is performed based on a high-confidence aircraft rigid-elastic coupling model with atmospheric turbulence excitation. This yields the time-domain acceleration response signals at key measurement points. The time-domain acceleration response signals are then processed using a sliding window slicing technique to generate multiple independent time-series data segments. For each time-series data segment, the transfer rate function between two selected measurement points is calculated. Extract the frequency corresponding to the maximum peak value of the transfer rate function. Calculate the current dynamic pressure based on flight conditions. The theoretical critical flutter frequency for each flight condition point was obtained through frequency domain flutter analysis. Compared with theoretical vibration pressure ; Calculate the frequency ratio and dynamic pressure ratio The bistability index was determined, and the bistability index samples were differentiated and standardized to form a full flight envelope simulation dataset.

[0012] Specifically, S3 establishes a dual-channel one-dimensional convolutional neural network model that includes a main channel for the transfer rate function and an auxiliary channel for flight conditions, as follows: The feature vectors output from the main channel of the transfer rate function and the auxiliary channel of flight conditions are concatenated and fused along the channel dimension. The predicted results of the frequency ratio and dynamic pressure ratio are then simultaneously output through a fully connected output layer, resulting in a dual-channel one-dimensional convolutional neural network model. The main channel of the transfer rate function consists of multiple alternating stacked one-dimensional convolutional layers and max-pooling layers, used for adaptively extracting multi-scale frequency domain features from the transfer rate function. The auxiliary channel of flight conditions consists of fully connected layers, used for altitude... With speed Operating condition information is mapped to a high-dimensional feature space.

[0013] Specifically, S4 involves: using the Adam optimizer and mean squared error loss function, performing end-to-end training on the model using the training set, introducing an early stopping mechanism to prevent overfitting; and using the test set to perform unbiased performance evaluation on the trained model. If the preset acceptance threshold is not met, the network structure and hyperparameters are adjusted and the model is retrained.

[0014] Specifically, S5 involves fine-tuning the model parameters based on transfer learning: collecting flight test data of real aircraft under subcritical conditions, preprocessing the flight test data into a standard input format according to the same process as the full flight envelope simulation data; loading the pre-trained model and freezing the feature extraction layer of the main channel of the transfer rate function; and fine-tuning the fully connected layer at the end of the network with a small amount of flight test data at an extremely low learning rate to achieve model adaptation from the simulation domain to the real domain.

[0015] Specifically, S6 is: The system collects time-domain acceleration response signals and flight condition information during aircraft flight in real time, preprocesses them, and inputs them into a fine-tuned model. It outputs real-time predicted values ​​of frequency ratio and dynamic pressure ratio. Stability is determined based on dual-index thresholds, and a corresponding level of warning signal is issued when either index reaches the warning threshold.

[0016] The sliding window slicing process in step S1 is as follows: along the time domain response with a fixed window length Step size is Sliding slicing is used to generate multiple independent time-series data segments to achieve data augmentation; The differential standardization process in step S1 is as follows: for dynamic pressure ratio Min-Max normalization was applied to scale it to the [0,1] interval; the frequency ratio was adjusted accordingly. Z-score standardization is used, which is performed based on the mean and variance of the training set.

[0017] Specifically, the multi-level early warning in step S6 is as follows: when the dynamic pressure ratio is ≥0.85 or the frequency ratio is ≥0.9, a first early warning signal is issued; when the dynamic pressure ratio is ≥0.95 or the frequency ratio is ≥0.98, a second early warning signal is issued.

[0018] The beneficial effects of this invention are at least as follows: (a) Generalization capability across the entire flight envelope: Through a dual-channel feature fusion architecture, the model utilizes both the structural dynamics characterized by the transfer rate function and the aerodynamic environment information characterized by the flight conditions, enabling the model to learn the independent and coupled effects of altitude and velocity on the system's dynamic characteristics. This allows for high-precision flutter prediction across the entire flight envelope, with a generalization capability superior to that of a single-channel model.

[0019] (b) High-efficiency computing and engineering capabilities: A one-dimensional convolutional neural network is used to automatically extract key frequency domain features from the transfer rate function, avoiding the subjectivity and limitations of manual feature extraction. By reducing the dimensionality of the input data and training time through convolution and pooling operations, real-time requirements can be met, making it suitable for engineering applications.

[0020] (c) Solving the model mismatch problem: A transfer learning paradigm of "large-scale simulation pre-training + fine-tuning with a small amount of flight test data" was established. The general feature extraction capability of the model is trained by using a large amount of low-cost simulation data, and then the high-level features of the model are fine-tuned with a small amount of expensive flight test data. This effectively solves the model mismatch problem between the simulation model and the real system, while significantly reducing the demand for flight test data.

[0021] (d) High-Reliability Dual-Indicator Prediction: Innovatively, it simultaneously predicts two complementary stability indices: frequency ratio and dynamic pressure ratio. The dynamic pressure ratio reflects the system's stability margin from an energy perspective, while the frequency ratio reflects the system's evolutionary trend from a dynamic characteristics perspective. The combination of the two provides dual evidence for flutter prediction.

[0022] (e) Strong physical interpretability: The model structure and prediction indices of this invention have clear physical meanings. The transferability function reflects the dynamic transfer characteristics of the structure, the flight condition reflects the aerodynamic environment, and the frequency ratio and dynamic pressure ratio correspond to the dynamic characteristic evolution and energy change of the system, respectively. This allows the model's prediction results to be physically interpreted, enhancing engineers' confidence in the model.

[0023] The method of this invention has a clear principle, high computational efficiency, accurate prediction, and high reliability. It provides a brand-new technical approach for predicting the boundary of aircraft flutter stability and has extremely important engineering application value and broad prospects for promotion in the aviation field. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is an overall flowchart of the body degree-of-freedom flutter prediction method based on the transfer rate function and flight conditions of this invention; Figure 2 This is a schematic diagram of the structure of the dual-channel one-dimensional convolutional neural network model of the present invention; Figure 3 This is a schematic diagram of the sliding window slicing method for the time-domain response signal of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] The features and illustrative embodiments of various aspects of the present invention will now be described in detail. Numerous specific details are set forth in the following detailed description to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without requiring some of these specific details. The following description of embodiments is merely intended to provide a better understanding of the invention by illustrating examples of the invention. The invention is by no means limited to any specific setups and methods set forth below, but covers any improvements, substitutions, and modifications to structures, methods, and devices without departing from the spirit of the invention. Well-known structures and techniques are not shown in the drawings and the following description to avoid unnecessarily obscuring the invention.

[0028] One embodiment of the present invention provides a body degree-of-freedom flutter prediction method based on transfer rate function and flight conditions, comprising the following contents: (a) Dual-channel feature fusion architecture: Construct a dual-channel one-dimensional convolutional neural network model with a main channel of transfer rate function and an auxiliary channel of flight conditions to achieve deep fusion of structural dynamic characteristics and flight conditions information. This enables the model to learn the independent and coupled effects of altitude and speed on the dynamic characteristics of the system, and achieve high-precision flutter prediction across the entire flight envelope.

[0029] (b) High-efficiency computing and engineering capabilities: A one-dimensional convolutional neural network is used to automatically extract key frequency domain features from the transfer rate function, avoiding the subjectivity and limitations of manual feature extraction. By reducing the dimensionality of the input data and training time through convolution and pooling operations, real-time requirements can be met, making it suitable for engineering applications.

[0030] (c) Establish a transfer learning paradigm of simulation pre-training + flight test fine-tuning: use a large amount of low-cost simulation data to train the model’s general feature extraction capabilities, and then use a small amount of expensive flight test data to fine-tune the model’s high-level features, effectively solving the model mismatch problem and significantly reducing the dependence on flight test data.

[0031] (d) A dual-dimensional stability index joint prediction mechanism is proposed: the frequency ratio (characterizing the migration trend of the system's dynamic characteristics) and the dynamic pressure ratio (characterizing the energy stability margin of the system) are predicted simultaneously, providing dual evidence for flutter prediction, significantly improving prediction reliability and robustness, and providing early warning.

[0032] This invention provides a method for predicting volume degree-of-freedom flutter based on transfer rate function and flight conditions. See [link to relevant documentation]. Figure 1 The specific steps are as follows: Step 1: Construct a full flight envelope simulation dataset (a) Establish a high-confidence rigid-elastic coupling model of an aircraft after ground vibration test correction; the high-confidence rigid-elastic coupling model of an aircraft after ground vibration test correction can more accurately reflect the actual characteristics of the aircraft structure. (b) Selecting different altitudes evenly across the entire flight envelope With speed The combination of these parameters is used as the flight operating point. For each flight operating point, atmospheric turbulence excitation conforming to Von Karman PSD (Von Karman power spectral density) is applied based on a high-confidence aircraft rigid-elastic coupling model to perform time-domain dynamic response simulation and obtain the time-domain acceleration response signals of each measurement point in key parts such as the wing, fuselage, and tail. (c) Sliding window slicing of the time-domain acceleration response signal: See Figure 3 The response along the time domain with a fixed window length Step size is Sliding slicing is used to generate multiple independent time-series data segments, achieving data augmentation. For each time-series data segment, the Welch method (Welch average periodogram) is used to calculate the transfer rate function between two selected measurement points. ; (d) The frequency corresponding to the maximum peak value of the transfer function is automatically identified using a peak search algorithm. According to flight altitude With speed Calculate the current dynamic pressure ,in This refers to the standard atmospheric density at the corresponding altitude. (e) Frequency domain flutter analysis was performed on each flight condition point using the PK method to obtain the theoretical critical flutter frequency. and theoretical vibration pressure Calculate two complementarity stability indices: frequency ratio. (The migration trend characterizing the dynamic properties of the system) and dynamic pressure ratio (Characterizing the energy stability margin of the system); (f) For all samples obtained in step (e) (i.e., frequency ratios) and dynamic pressure ratio Standardize the differentiated data: for dynamic pressure ratio Min-Max normalization was applied to scale it to the [0,1] interval; the frequency ratio was adjusted accordingly. Z-score (standard score) standardization is adopted, which is performed based on the mean and variance of the training set. (g) Construct a full flight envelope simulation dataset that includes structural dynamic characteristics, flight conditions, and stability indices.

[0033] Step 2: Divide the dataset into hierarchical layers The full flight envelope simulation dataset was randomly shuffled according to flight conditions and then divided into training, validation, and test sets in an 8:1:1 ratio. This ensures that each dataset contains flight conditions with the same distribution, avoiding model bias caused by uneven data distribution. The training set is used for parameter learning of the dual-channel one-dimensional convolutional neural network model, the validation set is used for hyperparameter tuning and early stopping, and the test set is used for unbiased evaluation of the final performance of the dual-channel one-dimensional convolutional neural network model.

[0034] Step 3: Establish a dual-channel one-dimensional convolutional neural network model See Figure 2 A dual-channel one-dimensional convolutional neural network model containing a main channel of transfer rate function and an auxiliary channel of flight conditions was constructed to achieve deep fusion of multi-source heterogeneous data.

[0035] (a) Main channel of the transfer rate function: The input is a one-dimensional transfer rate function vector, which consists of three one-dimensional convolutional layers and three max pooling layers stacked alternately. The number of convolutional kernels increases from 32 to 128 layer by layer, and the kernel size decreases from 6 to 3 layer by layer. All convolutional layers use ReLU (rectified linear function) as the activation function. The convolutional layers automatically extract local frequency domain features in the transfer rate function (such as peak position, peak amplitude, valley position, slope change, etc.), and the pooling layers reduce the feature dimension, reduce the amount of computation, and prevent overfitting. (b) Flight Operation Auxiliary Channel: Input is flight altitude With speed Two scalar parameters, consisting of two fully connected layers with 64 and 128 neurons respectively, using ReLU as the activation function. Flight condition information is converted into a high-dimensional feature vector through nonlinear mapping, enabling effective fusion with the frequency domain features extracted from the main channel. (c) Feature fusion layer: The frequency domain feature vector output by the main channel and the operating condition feature vector output by the auxiliary channel are concatenated in the channel dimension to form a fused feature vector, thereby realizing the deep fusion of structural dynamic characteristics and flight operating condition information. (d) Fully Connected Output Layer: Consists of two fully connected layers, one Dropout layer, and one output layer. The fully connected layers further process the fused features. The Dropout layer has a dropout rate of 0.1 to prevent overfitting. The output layer ultimately outputs two stability metrics: frequency ratio. and dynamic pressure ratio .

[0036] Step 4: Perform model training and performance validation (a) Configure training hyperparameters: Use the Adam optimizer with an adaptive learning rate and an initial learning rate of 1e. -4 The loss function is the mean squared error function, the batch size is 32, and the number of training epochs is 400. (b) Set an early stopping mechanism: When the validation set loss no longer decreases for 20 consecutive training rounds, stop training early, save the model weights with the minimum validation set loss, and prevent overfitting. (c) Train the model end-to-end using the training set and update all weights and bias parameters of the network using the backpropagation algorithm; (d) Use the test set to perform unbiased performance evaluation on the trained model, and calculate metrics such as mean squared error (MSE) and coefficient of determination (R²). If the model performance does not meet the acceptance threshold (MSE≤0.001, R²≥0.95), adjust the network structure and hyperparameters and retrain until the requirements are met.

[0037] Step 5: Fine-tune the model parameters based on transfer learning. (a) Collect flight test data of real aircraft under subcritical conditions, including time-domain acceleration response signals at each measurement point and corresponding flight condition information (altitude, speed). (b) Process the flight test data into a standard input format, i.e., the transfer rate function and the corresponding bistability index, following the same processing procedure as the simulation data. (c) Load the model trained in step four, freeze all convolutional and pooling layers in the main channel of the transfer rate function (to preserve general feature extraction capabilities), and only enable the last two fully connected layers for training. Configure the fine-tuning parameters to the optimizer Adam, with a learning rate of 1e. -8 The batch size is 8, and the training rounds are 20. (d) Fine-tune the model using a small amount of flight test data, making only minor adjustments to the model weights to adapt it to the characteristics of a real aircraft; (e) Save the fine-tuned model for actual flight flutter prediction.

[0038] Step Six: Real-time Flutter Prediction and Multi-level Early Warning (a) Real-time acquisition of time-domain acceleration response signals and flight condition information at various measurement points during aircraft flight; (b) Following the same processing procedure as the simulation data, calculate the transfer rate function and standardize it; (c) Input the preprocessed data into the fine-tuned model and output the prediction results of frequency ratio and dynamic pressure ratio in real time; (d) Determine the current stability state of the aircraft based on the prediction results and set multi-level early warning thresholds: when the dynamic pressure ratio ≥0.85 or frequency ratio A yellow alert is issued when the value is ≥0.9; when ≥0.95 or A red alert is issued when the value is ≥0.98. (e) Push the forecast results and early warning information to the flight monitoring system and pilots in real time in a specified format to provide decision support for flight safety.

Claims

1. A method for predicting body degree-of-freedom flutter based on transfer rate function and flight conditions, characterized in that, Includes the following steps: S1: Constructing a full flight envelope simulation dataset: For each flight condition point, atmospheric turbulence excitation is applied based on a high-confidence aircraft rigid-elastic coupling model to perform time-domain dynamic response simulation, and time-domain acceleration response signals of each measurement point in key parts are obtained; the time-domain acceleration response signals are processed to calculate the transfer rate function between two selected measurement points; based on the bistability index of the transfer rate function, a full flight envelope simulation dataset is formed. S2: Dataset stratification: The full flight envelope simulation dataset is randomly shuffled according to flight conditions and divided into training set, validation set and test set to ensure that each dataset contains the same distribution of flight conditions; S3: Establish a dual-channel one-dimensional convolutional neural network model: Establish a dual-channel one-dimensional convolutional neural network model that includes the main channel of the transfer rate function and the auxiliary channel of flight conditions; S4: Perform model training and performance validation: Train the model end-to-end using the training set; evaluate the unbiased performance of the trained model using the test set; S5: Fine-tuning model parameters based on transfer learning: Fine-tuning model parameters using a small amount of flight test data to achieve adaptive modeling from the simulation domain to the real domain; S6: Utilize a finely tuned dual-channel one-dimensional convolutional neural network model for online flutter prediction and multi-level early warning.

2. The method according to claim 1, characterized in that, S1 specifically refers to: Based on the modified high-confidence aircraft rigid-elastic coupling model, different altitudes are selected within the entire flight envelope. With speed Combined flight operating points, for each flight operating point, atmospheric turbulence excitation is applied based on a high-confidence aircraft rigid-elastic coupling model to perform time-domain dynamic response simulation, and time-domain acceleration response signals of each measurement point in key parts are obtained; The time-domain acceleration response signal is sliced ​​using a sliding window to generate multiple independent time-series data segments. For each time-series data segment, the transfer rate function between two selected measurement points is calculated. Extract the frequency corresponding to the maximum peak value of the transfer rate function. Calculate the current dynamic pressure based on flight conditions. The theoretical critical flutter frequency for each flight condition point was obtained through frequency domain flutter analysis. Compared with theoretical vibration pressure ; Calculate the frequency ratio and dynamic pressure ratio The bistability index was determined, and the bistability index samples were differentiated and standardized to form a full flight envelope simulation dataset.

3. The method according to claim 1, characterized in that, In S3, a dual-channel one-dimensional convolutional neural network model is established, comprising a main channel for the transfer rate function and an auxiliary channel for flight conditions. Specifically: The feature vectors output from the main channel of the transfer rate function and the auxiliary channel of flight conditions are concatenated and fused along the channel dimension. The predicted results of the frequency ratio and dynamic pressure ratio are then simultaneously output through a fully connected output layer, resulting in a dual-channel one-dimensional convolutional neural network model. The main channel of the transfer rate function consists of multiple alternating stacked one-dimensional convolutional layers and max-pooling layers, used for adaptively extracting multi-scale frequency domain features from the transfer rate function. The auxiliary channel of flight conditions consists of fully connected layers, used for altitude... With speed Operating condition information is mapped to a high-dimensional feature space.

4. The method according to claim 1, characterized in that, S4 specifically involves: using the Adam optimizer and mean squared error loss function, performing end-to-end training on the model using the training set, introducing an early stopping mechanism to prevent overfitting; using the test set to perform unbiased performance evaluation on the trained model, and adjusting the network structure and hyperparameters to retrain if the preset acceptance threshold is not met.

5. The method according to claim 1, characterized in that, S5 specifically involves: collecting flight test data of real aircraft under subcritical conditions, preprocessing the flight test data into a standard input format according to the same process as the full flight envelope simulation data; loading the pre-trained model and freezing the feature extraction layer of the main channel of the transfer rate function; and fine-tuning the fully connected layer at the end of the network with a very low learning rate using a small amount of flight test data to achieve model adaptation from the simulation domain to the real domain.

6. The method according to claim 1, characterized in that, S6 specifically refers to: The system collects time-domain acceleration response signals and flight condition information during aircraft flight in real time, preprocesses them, and inputs them into a fine-tuned model. It outputs real-time predicted values ​​of frequency ratio and dynamic pressure ratio. Stability is determined based on dual-index thresholds, and a corresponding level of warning signal is issued when either index reaches the warning threshold.

7. The method according to claim 1, characterized in that, The sliding window slicing process in step S1 is as follows: along the time domain response with a fixed window length Step size is Sliding slices are used to generate multiple independent time-series data segments to achieve data augmentation.

8. The method according to claim 1, characterized in that, The differential standardization process mentioned in step S1 is as follows: for dynamic pressure ratio Min-Max normalization was applied to scale it to the [0,1] interval; the frequency ratio was adjusted accordingly. Z-score standardization is used, which is performed based on the mean and variance of the training set.

9. The method according to claim 1, characterized in that, The multi-level early warning in step S6 is as follows: when the dynamic pressure ratio is ≥0.85 or the frequency ratio is ≥0.9, a first early warning signal is issued; when the dynamic pressure ratio is ≥0.95 or the frequency ratio is ≥0.98, a second early warning signal is issued.

Images