A tilt-rotor aircraft aeroelastic stability early warning device and method based on HHT and deep learning
By combining Hilbert-Huang transform and deep learning, the problems of accuracy and real-time performance in predicting the aeroelastic stability of tiltrotor aircraft were solved, achieving high-precision aeroelastic stability early warning, which is applicable to various operating conditions and environmental changes of tiltrotor aircraft.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-05
Smart Images
Figure CN122157442A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of tiltrotor aircraft technology, and in particular to a tiltrotor aircraft aeroelastic stability early warning device and method based on HHT and deep learning. Background Technology
[0002] Tiltrotor aircraft combine the advantages of helicopters and fixed-wing aircraft. The unique rotor, nacelle and wing layout results in a stronger aeroelastic coupling effect than traditional helicopters and fixed-wing aircraft, making them more prone to instability phenomena such as gyroscopic flutter, which seriously threaten flight safety and limit forward speed and flight envelope.
[0003] Therefore, aeroelastic stability prediction is crucial for the design and operation of tiltrotor aircraft. In some cases, aeroelastic stability prediction can be divided into two main methods: physical mechanism modeling and data-driven methods. Pure physical mechanism modeling requires simplified assumptions, makes it difficult to represent the real coupling relationship, and suffers from large errors and poor real-time performance under complex operating conditions. Pure deep learning models do not adequately utilize the dynamic characteristics of non-stationary aeroelastic signals and have weak generalization ability across operating conditions. Therefore, neither of these two methods can achieve high-precision prediction of aeroelastic stability of tiltrotor aircraft and fusion helicopters, let alone high-precision early warning. Summary of the Invention
[0004] The purpose of this application is to provide a tiltrotor aircraft aeroelastic stability early warning device and method based on HHT and deep learning. By combining Hilbert-Huang Transform (HHT) and deep learning for prediction, the high-precision prediction and early warning of tiltrotor aircraft fused with helicopter aeroelastic stability can be improved.
[0005] To achieve the above objectives, this application provides the following solution.
[0006] In a first aspect, this application provides a tiltrotor aircraft aeroelastic stability early warning device based on HHT and deep learning, comprising: A sensing module, installed on the tiltrotor aircraft, is used to collect the raw signals of the tiltrotor aircraft; A data preprocessing module, connected to the sensing module, is used to preprocess the original signal to obtain multiple preprocessed signals; each preprocessed signal includes: aeroelastic response signal, operating condition parameters, and environmental parameters. The feature extraction module, connected to the data preprocessing module, is used to extract features from the aeroelastic response signals in each segment of the preprocessed signal using HHT to obtain a time-frequency feature matrix. The deep learning prediction module is connected to the feature extraction module and the data preprocessing module, respectively, and is used to make predictions based on the time-frequency feature matrix, the operating condition parameters, and the environmental parameters using a deep learning prediction model to obtain prediction results; the prediction results include damping ratio and flutter critical velocity; the deep learning prediction model is obtained by training a CNN-LSTM hybrid model using sample fusion feature vectors and sample prediction results; The result output and early warning module is connected to the deep learning prediction module and is used to display the prediction results, generate a judgment signal based on the prediction results, and issue an early warning based on the judgment signal.
[0007] Secondly, this application provides a tiltrotor aircraft aeroelastic stability early warning method based on HHT and deep learning, including: Acquire the raw signals from the tiltrotor aircraft; The original signal is preprocessed to obtain multiple preprocessed signals; each preprocessed signal includes: aeroelastic response signal, operating condition parameters and environmental parameters; HHT was used to extract features from the aeroelastic response signals in each preprocessed signal segment to obtain the time-frequency feature matrix; Based on the time-frequency feature matrix, the operating condition parameters, and the environmental parameters, a deep learning prediction model is used to make predictions, and the prediction results are obtained. The prediction results include the damping ratio and the flutter critical velocity. The deep learning prediction model is obtained by training a CNN-LSTM hybrid model using sample fusion feature vectors and sample prediction results. A judgment signal is generated based on the prediction result, and an early warning is issued based on the judgment signal.
[0008] According to the specific embodiments provided in this application, this application has the following technical effects: This application employs a data preprocessing module to preprocess the raw signals of the tiltrotor aircraft collected by the sensing module, obtaining multiple preprocessed signal segments that provide a precise data foundation for subsequent predictions. Then, a feature extraction module extracts the aeroelastic response signals from each preprocessed signal segment using HHT (Heterogeneous Theory of Time) to obtain a time-frequency feature matrix. This accurately extracts the time-frequency features of the non-stationary aeroelastic signal, ensuring accurate input data for the deep learning prediction model. Furthermore, a deep learning prediction module, based on the time-frequency feature matrix, operating parameters, and environmental parameters, uses a deep learning prediction model to make predictions, obtaining the damping ratio and flutter critical velocity. By ensuring the accuracy of the input data, the prediction accuracy and generalization ability under complex operating conditions are improved. Finally, through the result output and early warning module, a judgment signal is generated based on the prediction results, and an early warning is issued based on the judgment signal. This achieves high-precision prediction and early warning of the aeroelastic stability of the tiltrotor aircraft integrated with helicopter. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0010] Figure 1 A schematic diagram of the functional modules of the tiltrotor aeroelastic stability early warning device based on HHT and deep learning provided in this application.
[0011] Figure 2 This is a schematic diagram of each sensor in the sensing module provided in this application.
[0012] Figure 3 (a) is a schematic diagram of the timing sequence of the Hilbert-Huang transform feature extraction principle provided in this application; (b) is a schematic diagram of the original aeroelastic response time domain signal; (c) is a schematic diagram of each modal function; (d) is a schematic diagram of the analytic signal of each modal function; and (d) is a schematic diagram of the time-frequency feature matrix.
[0013] Figure 4 A schematic diagram of the network structure of the deep learning prediction model provided in this application.
[0014] Figure 5 A flowchart illustrating a tiltrotor aeroelastic stability early warning method based on HHT and deep learning provided in this application.
[0015] Figure reference numerals: 1-Sensing module; 2-Data preprocessing module; 3-Feature extraction module; 4-Deep learning module; 5-Result output and early warning module; 6-Power supply module; 11-Strain sensor; 12-Acceleration sensor; 13-Pressure sensor; 14-Operating condition parameter sensor; 15-Environmental sensor. Detailed Implementation
[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0017] Terminology Explanation: Empirical Mode Decomposition (EMD) is an adaptive time-domain signal processing method suitable for nonlinear and non-stationary signals.
[0018] The Hilbert-Huang Transform (HHT) is a powerful tool specifically designed for analyzing nonlinear and non-stationary signals. This transform comprises two key steps: Empirical Mode Decomposition (EMD) and the Hilbert Transform. EMD decomposes complex signals into intrinsic mode functions (IMFs) with local characteristics, while the Hilbert Transform is used to obtain the instantaneous frequency and amplitude information of the IMFs.
[0019] Convolutional Neural Network (CNN): A type of feedforward neural network.
[0020] Long Short-Term Memory (LSTM): A type of time-recurrent neural network.
[0021] Principal Components Analysis (PCA), also known as principal component analysis technique, aims to use the idea of dimensionality reduction to transform multiple indicators into a few comprehensive indicators.
[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0023] In one exemplary embodiment, such as Figure 1 As shown, a tiltrotor aeroelastic stability early warning device based on HHT and deep learning is provided. The device includes a sensing module 1, a data preprocessing module 2, a feature extraction module 3, a deep learning prediction module 4, and a result output and early warning module 5. Each module is described in detail below.
[0024] (1) Sensing module 1.
[0025] Sensor module 1, installed on the tiltrotor aircraft, is used to collect the raw signals of the tiltrotor aircraft.
[0026] As an feasible approach, the raw signals include: raw aeroelastic response signals, raw operating parameters, and raw environmental parameters; the raw aeroelastic response signals include: raw strain signals, raw acceleration signals, and raw aerodynamic pressure signals; the raw operating parameters include: rotor speed signals, tilt angle signals, and forward speed signals; and the raw environmental parameters include: raw temperature signals and raw humidity signals.
[0027] like Figure 2 As shown, the sensing module 1 includes a strain sensor 11, an acceleration sensor 12, a pressure sensor 13, a working condition parameter sensor 14, and an environmental sensor 15.
[0028] Strain sensor 11 is installed on the wing of the tiltrotor aircraft to collect the raw strain signal of the wing.
[0029] Specifically, strain sensor 11, model BX120-3AA, is installed on the critical section of the tiltrotor wing (mainly the wing tip) to collect the raw strain signals of the wing's bending / torsion.
[0030] Acceleration sensor 12 is installed on the rotor blades of the tiltrotor aircraft to collect the raw acceleration signal of the blades.
[0031] Specifically, the acceleration sensor 12 is model ADXL345, which is installed at the root of the rotor blades of the tiltrotor aircraft to collect the raw acceleration signal of the blade flapping / swaying.
[0032] Pressure sensor 13 is installed on the wing of the tiltrotor aircraft to collect raw aerodynamic pressure signals.
[0033] Specifically, the pressure sensor 13 is model MPX5700DP, which is arranged in an array on the wing surface of the tiltrotor aircraft to collect raw aerodynamic pressure signals.
[0034] Operating condition parameter sensor 14 is installed on the tiltrotor aircraft to collect the original operating condition parameters of the tiltrotor aircraft.
[0035] Specifically, the operating condition parameter sensor 14 includes a Hall effect speed sensor and an encoder-type tilt angle sensor, which are used to collect the raw operating condition parameters of the tiltrotor aircraft (such as rotor speed, tilt angle and forward speed signals).
[0036] Environmental sensor 15 is installed on the tiltrotor aircraft to collect raw environmental parameters.
[0037] Specifically, the environmental sensor 15 is a DHT11 temperature and humidity sensor, used to collect raw temperature and raw humidity signals.
[0038] Among them, strain sensor 11 and acceleration sensor 12 use differential signal output, and pressure sensor 13 uses analog voltage signal output. The signal output terminals of the above sensors are connected to the signal input terminal of the data preprocessing module through shielded cables.
[0039] (2) Data preprocessing module 2.
[0040] The data preprocessing module 2, connected to the sensing module 1, is used to preprocess the original signal to obtain multiple preprocessed signals; each preprocessed signal includes: aeroelastic response signal, operating condition parameters and environmental parameters.
[0041] Specifically, preprocessing mainly includes: noise reduction, outlier removal, time synchronization, normalization alignment, and segmentation.
[0042] As one feasible approach, data preprocessing module 2 includes: The ADC converter, connected to the sensing module 1, is used to convert the original signal from analog to digital to obtain the original digital signal.
[0043] The microcontroller, connected to the ADC converter, is used to: filter out the original digital aeroelastic response signal from the original digital signal using an adaptive wavelet threshold denoising algorithm to obtain the filtered aeroelastic response signal; remove outliers from the original digital operating parameters and original digital environmental parameters in the original digital signal using the 3σ criterion to obtain the removed operating parameters and removed environmental parameters; and sequentially perform time synchronization, normalization, and segmentation processing on the filtered aeroelastic response signal, the removed operating parameters, and the removed environmental parameters to obtain multiple preprocessed signals.
[0044] Specifically, the core chip of data preprocessing module 2 is a microcontroller (e.g., STM32H743), accompanied by a signal conditioning chip (e.g., AD8232) and an ADC converter (e.g., ADS1256). The signal conditioning chip first amplifies the original strain signal and original acceleration signal from the differential signal input, with an adjustable amplification factor ranging from 100 to 1000 times, while simultaneously performing low-pass filtering with a cutoff frequency of 500Hz. The ADC converter converts the amplified original strain signal, amplified original acceleration signal, original aerodynamic pressure signal, original operating condition parameters, and original environmental parameters into 24-bit digital signals with a maximum conversion rate of 1000 SPS. The microcontroller executes the data preprocessing algorithm to obtain multiple preprocessed signals, which are then transmitted to the feature extraction module via the SPI interface.
[0045] The preprocessing algorithm involves the following steps: An adaptive wavelet threshold denoising algorithm is used, selecting the db4 wavelet basis and 5-level decomposition to filter out high-frequency noise in the original aeroelastic response signal; the 3σ criterion is used to identify and remove outliers in the original operating parameters and original environmental parameters; signals collected by different sensors are synchronized and aligned based on timestamps to unify the time reference; all preprocessed signals are normalized to the [-1,1] interval, with the normalization formula being: x_norm=(x-x_min) / (x_max-x_min)×2-1 (where x is the original signal, and x_min and x_max are the minimum and maximum values of the original signal, respectively); the synchronized signal is segmented into multiple preprocessed signals according to a 1s time window.
[0046] (3) Feature extraction module 3.
[0047] Feature extraction module 3, connected to data preprocessing module 2, is used to extract features from the aeroelastic response signals in each segment of preprocessed signals using HHT, and obtain time-frequency feature matrix.
[0048] Specifically, feature extraction module 3 uses the Xilinx XC7Z020 FPGA chip, which integrates EMD decomposition and Hilbert transform to achieve hardware acceleration of Hilbert-Huang transform calculation. The extracted time-frequency feature matrix is transmitted to the deep learning prediction module through the PCIe interface.
[0049] As one feasible approach, feature extraction module 3 includes: The decomposition unit, connected to the microcontroller, is used to process the aeroelastic response signal in each preprocessed signal segment (see...). Figure 3 EMD decomposition was performed on (a) to obtain multiple mode functions and a residual component.
[0050] The transformation unit, connected to the decomposition unit, is used to process each modal function (see...) Figure 3 Perform a Hilbert transform on (b) to obtain the analytic signals of each mode function (see [reference]). Figure 3 (c)); the analytic signal includes instantaneous amplitude and instantaneous phase.
[0051] The calculation unit, connected to the transformation unit, is used to calculate the instantaneous frequency and instantaneous damping ratio of each mode function based on the analytical signal of each mode function.
[0052] The feature extraction unit, connected to the transformation unit, is used to extract features from each modal function to obtain the statistical features of each modal function.
[0053] The integration unit, connected to the transformation unit, calculation unit, and feature extraction unit respectively, is used to integrate the instantaneous amplitude, instantaneous frequency, instantaneous damping ratio, and statistical characteristics of each modal function to obtain the time-frequency feature matrix (see...). Figure 3 (d)
[0054] Specifically, the aeroelastic response signal in each preprocessed signal is decomposed using EMD to obtain 4-6 intrinsic mode functions (IMFs) and 1 residual component. The decomposition terminates when the standard deviation of the IMF component is less than 0.2. A Hilbert transform is then performed on each IMF to obtain the analytic signal z(t) = a(t)e^(j (t) (where z(t) is the analytic signal, a(t) is the instantaneous amplitude, (t) represents the instantaneous phase); the instantaneous frequency and instantaneous damping ratio of each IMF are calculated based on the analytical signal of each IMF; the formula for calculating the instantaneous frequency is: f(t) = d (t) / (2πdt), where f(t) is the instantaneous frequency; the instantaneous damping ratio is calculated through the instantaneous amplitude attenuation characteristics, and the peak value, root mean square, centroid frequency and other statistical features of each IMF are extracted; the instantaneous frequency, instantaneous amplitude, instantaneous damping ratio and statistical features of each IMF are integrated to form a time-frequency feature matrix with a dimension of 64×T (T is the number of sampling points in the time window), and the obtained time-frequency feature matrix is transmitted to the deep learning prediction module through the PCIe interface.
[0055] (4) Deep learning prediction module 4.
[0056] The deep learning prediction module 4 is connected to the feature extraction module 3 and the data preprocessing module 2 respectively. It is used to make predictions based on the time-frequency feature matrix, operating parameters and environmental parameters using a deep learning prediction model to obtain prediction results. The prediction results include the damping ratio and flutter critical velocity. The deep learning prediction model is obtained by training the CNN-LSTM hybrid model using sample fusion feature vectors and sample prediction results.
[0057] Specifically, the core of the deep learning prediction module 4 is the NVIDIA Jetson Nano deep learning acceleration chip, which integrates a GPU core for parallel computation of the CNN-LSTM hybrid model, pre-stores the trained model parameters, receives the time-frequency feature matrix through the PCIe interface, performs feature fusion and prediction calculation, and transmits the prediction results to the result output and warning module through the UART interface.
[0058] As an implementable approach, deep learning prediction module 4 includes: The dimensionality reduction module unit, connected to the feature extraction module, is used to reduce the dimensionality of the time-frequency feature matrix using principal component analysis to obtain the time-frequency feature vector.
[0059] The splicing unit is connected to the dimensionality reduction module unit and the data preprocessing module 2, respectively, and is used to splice the time-frequency feature vector, operating condition parameters and environmental parameters to obtain the fused feature vector.
[0060] The prediction unit, connected to the concatenation unit, is used to input the fused feature vector into the deep learning prediction model to obtain the prediction result.
[0061] Specifically, principal component analysis (PCA) is used to reduce the dimensionality of the time-frequency feature matrix, retaining 95% of the feature information, to obtain a time-frequency feature vector (64-dimensional). The time-frequency feature vector, operating parameters (8-dimensional), and environmental parameters (2-dimensional) are then concatenated to form a 74-dimensional fused feature vector. The 74-dimensional fused feature vector is then input into a deep learning prediction model, which outputs prediction results (damping ratio and flutter critical velocity). The prediction results are transmitted to the result output module via a UART interface.
[0062] As an implementable approach, the CNN-LSTM hybrid model comprises a first convolutional layer, a second convolutional layer, an LSTM layer, a first fully connected layer, and a second fully connected layer connected in sequence.
[0063] Specifically, such as Figure 4 As shown, the CNN-LSTM hybrid model consists of two convolutional layers, one LSTM layer, and two fully connected layers. The first convolutional layer has 32 kernels (3×3 kernel size), uses ReLU activation, and employs max pooling (2×2 kernel size). The second convolutional layer has 64 kernels (3×3 kernel size), also uses ReLU activation, and employs max pooling. The LSTM layer has 128 hidden units and uses tanh activation. The first fully connected layer has an output dimension of 64 and uses ReLU activation, while the second fully connected layer has an output dimension of 2, corresponding to the "damping ratio" and "flutter critical velocity," respectively. "Two core stability metrics; the model uses the adaptive momentum estimation (Adam) optimizer with an initial learning rate of 0.001 and a learning rate decay strategy (decaying to 0.9 every 100 epochs). The loss function is mean squared error (MSE). The training data is a mixed dataset of tiltrotor wind tunnel test data and flight test data (containing 5000 samples), which is divided into a training set (sample fusion feature vector and sample prediction results), a validation set, and a test set in a 7:2:1 ratio. During training, an early stopping strategy is adopted (training stops if the validation set loss does not decrease for 20 consecutive epochs) to prevent overfitting."
[0064] (5) Result output and early warning module 5.
[0065] The result output and early warning module 5 is connected to the deep learning prediction module 4. It is used to display the prediction results, generate a judgment signal based on the prediction results, and issue an early warning based on the judgment signal.
[0066] As an implementable approach, the results output and early warning module 5 includes: The Bluetooth communication unit is connected to the deep learning prediction module 4 and is used for wireless transmission of prediction results.
[0067] The display screen, connected to the Bluetooth communication unit, is used to display the prediction results.
[0068] The determination unit, connected to the Bluetooth communication unit, is used to generate determination signals based on the prediction results; the determination signals include: stable signals, metastable signals, and unstable signals.
[0069] An alarm buzzer, connected to the judgment unit, is used to issue an early warning when the judgment signal is unstable.
[0070] Specifically, the result output module 5 consists of a Bluetooth communication unit (e.g., HC-08), a display screen (e.g., a 128×64 resolution LCD display), a judgment unit, and an alarm buzzer. The Bluetooth communication module wirelessly transmits the prediction results to the ground monitoring system and the airborne flight control system. The LCD display screen displays the current judgment signal, damping ratio, and flutter critical speed in real time. When the judgment signal is unstable, the alarm buzzer emits a continuous buzzing at a frequency of 1kHz and simultaneously outputs a high-level signal to the airborne warning light.
[0071] As an feasible approach, such as Figure 1 As shown, the tiltrotor aeroelastic stability early warning device based on HHT and deep learning also includes: a power supply module 6, used to power the sensing module 1, data preprocessing module 2, feature extraction module 3, deep learning prediction module 4, and result output and early warning module 5. The power supply module 6 is a DC regulated power supply module with an input voltage of 28V airborne standard power and multiple stable output voltages: 3.3V for the microcontroller and sensing module 1, 5V for the display screen and Bluetooth communication unit, and 12V for the FPGA chip of the feature extraction module 3 and the deep learning acceleration chip of the deep learning prediction module 4; the module has built-in overvoltage and overcurrent protection circuits to ensure stable operation of the device.
[0072] The beneficial effects of the tiltrotor aeroelastic stability early warning device based on HHT and deep learning proposed in this application are mainly reflected in the following aspects: 1. Higher prediction accuracy: This application uses Hilbert-Huang transform to extract time-frequency features. Compared with traditional deep learning methods that only use time-domain signals, it can more accurately capture the instantaneous dynamic changes of non-stationary aeroelastic signals, providing more comprehensive and effective feature inputs for the prediction model. This solves the problem of insufficient utilization of signal dynamic laws by traditional deep learning models and achieves complementary advantages at the feature level. At the same time, it integrates operating condition parameters and environmental parameters, enabling the model to learn the coupling law of multiple factors and reduce the errors caused by the simplification of physical modeling.
[0073] 2. Enhanced Generalization Ability: The aeroelastic signal characteristics of tiltrotor aircraft vary significantly across different flight modes, and traditional models are prone to feature distribution shifts due to changes in operating conditions. This application utilizes a time-frequency feature matrix extracted by a feature extraction module, which exhibits better domain invariance and effectively characterizes the essential characteristics of the aeroelastic system under different operating conditions. Furthermore, training with multimodal data (sample fusion feature vectors and sample prediction results) enables the deep learning prediction model in the deep learning prediction module to learn more general stability laws, rather than being limited to specific operating conditions. Therefore, the deep learning prediction model of this application maintains high prediction accuracy across various operating conditions of tiltrotor aircraft, including helicopter mode, transition mode, fixed-wing mode, and extreme environments. It generates fused feature vectors from the input time-frequency feature matrix, operating condition parameters, and environmental parameters, and then uses these fused feature vectors for prediction.
[0074] 3. Superior Real-Time Performance: This application employs a hardware-algorithm co-optimization strategy: hardware acceleration of the Hilbert-Huang transform is achieved through FPGA, resulting in over 10 times higher efficiency compared to software implementation; a lightweight CNN-LSTM hybrid model is used, reducing the number of model parameters by over 30% through parameter pruning and quantization optimization, optimizing the number of network parameters while maintaining prediction accuracy, thus balancing prediction accuracy and real-time performance; data preprocessing steps are executed in parallel within the microcontroller, avoiding the computational bottleneck of a single chip. In summary, the device in this application achieves efficient and real-time operation, with prediction latency controllable within 50ms, meeting the requirements for real-time monitoring during flight.
[0075] 4. Enhanced Engineering Practicality: The device in this application adopts a modular design, with each module having independent functions and standardized interfaces, facilitating installation, maintenance, and future upgrades. The result output module supports multiple methods, including wireless communication, local display, and audible and visual alarms. It can simultaneously interface with ground monitoring systems and airborne flight control systems, providing multi-dimensional support for flight safety decision-making. The device's power supply is compatible with the standard 28V airborne power supply and has built-in protection circuitry to ensure stable operation in harsh flight environments. Therefore, this invention better meets the practical needs of engineering applications.
[0076] 5. Enhanced Multiphysics Coupling Adaptability: Existing technologies often neglect the influence of environmental factors such as temperature and humidity. This application collects temperature and humidity data through environmental sensors and integrates them into a fusion feature vector, enabling the model to automatically learn the coupling relationship between environmental parameters and aeroelastic stability during training. Therefore, in extreme environments such as high temperature and high humidity, the decrease in prediction accuracy of this application is less than that of existing technologies, resulting in higher prediction accuracy and better ensuring flight safety in all environments.
[0077] Based on the same inventive concept, this application also provides a tiltrotor aeroelastic stability early warning method based on HHT and deep learning, which is applied to the aforementioned tiltrotor aeroelastic stability early warning device based on HHT and deep learning. Figure 5 As shown, a tiltrotor aeroelastic stability early warning method based on HHT and deep learning includes steps S1 to S5: Step S1: Obtain the raw signals from the tiltrotor aircraft.
[0078] Specifically, the sensor modules collect raw signals from the tiltrotor aircraft, including: raw strain signals of vertical bending / chord bending / torsion of the wing collected by strain sensors (sampling frequency 1000Hz); raw acceleration signals of blade flapping / flapping collected by acceleration sensors (sampling frequency 2000Hz); raw aerodynamic pressure signals collected by a pressure sensor array (sampling frequency 500Hz); operating condition signals such as rotor speed, tilt angle, and forward speed collected by operating condition parameter sensors; and environmental signals such as raw temperature and humidity collected by environmental sensors.
[0079] Step S2: Preprocess the original signal to obtain multiple preprocessed signals; each preprocessed signal includes: aeroelastic response signal, operating parameters and environmental parameters.
[0080] Step S3: Use HHT to extract features from the aeroelastic response signals in each preprocessed signal segment to obtain the time-frequency feature matrix.
[0081] As one feasible approach, step S3 specifically includes steps S31 to S35: Step S31: Perform EMD decomposition on the aeroelastic response signal in each preprocessed signal segment to obtain multiple mode functions and a residual component.
[0082] Step S32: Perform Hilbert transform on each mode function to obtain the analytic signal of each mode function; the analytic signal includes instantaneous amplitude and instantaneous phase.
[0083] Step S33: Based on the analytical signals of each mode function, calculate the instantaneous frequency and instantaneous damping ratio of each mode function.
[0084] Step S34: Extract features from each modal function to obtain the statistical features of each modal function.
[0085] Step S35: Integrate the instantaneous amplitude, instantaneous frequency, instantaneous damping ratio, and statistical characteristics of each modal function to obtain the time-frequency characteristic matrix.
[0086] Step S4: Based on the time-frequency feature matrix, operating parameters, and environmental parameters, a deep learning prediction model is used to make predictions and obtain prediction results. The prediction results include the damping ratio and flutter critical velocity. The deep learning prediction model is obtained by training a CNN-LSTM hybrid model using sample fusion feature vectors and sample prediction results. As an implementable approach, step S4 specifically includes: using principal component analysis to reduce the dimensionality of the time-frequency feature matrix to obtain the time-frequency feature vector; concatenating the time-frequency feature vector, operating parameters, and environmental parameters to obtain the fused feature vector; and inputting the fused feature vector into a deep learning prediction model to obtain the prediction result.
[0087] Step S5: Generate a judgment signal based on the prediction results, and issue an early warning based on the judgment signal.
[0088] Specifically, the output module compares the damping ratio and flutter critical speed output by the deep learning prediction model with preset thresholds. Specifically, if the damping ratio is greater than 0 and the flutter critical speed is greater than the current flight speed, the current state is determined to be stable, and the evaluation result of the stable signal and specific indicators are output. If the damping ratio decreases to close to 0 or the flutter critical speed is close to the current flight speed, the current state is determined to be metastable, and the evaluation result of the metastable signal and the first warning signal of the limit cycle oscillation are output. If the damping ratio is less than or equal to 0 or the flutter critical speed is less than or equal to the current flight speed, the current state is determined to be unstable, and the evaluation result of the unstable signal and the second warning signal are output. At the same time, the prediction results are fed back to the flight control system to provide a basis for adjusting the control strategy.
[0089] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0090] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A tiltrotor aeroelastic stability early warning device based on HHT and deep learning, characterized in that, The tiltrotor aeroelastic stability early warning device based on HHT and deep learning includes: A sensing module, installed on the tiltrotor aircraft, is used to collect the raw signals of the tiltrotor aircraft; A data preprocessing module, connected to the sensing module, is used to preprocess the original signal to obtain multiple preprocessed signals; each preprocessed signal includes: aeroelastic response signal, operating condition parameters, and environmental parameters. The feature extraction module, connected to the data preprocessing module, is used to extract features from the aeroelastic response signals in each segment of the preprocessed signal using HHT to obtain a time-frequency feature matrix. The deep learning prediction module is connected to the feature extraction module and the data preprocessing module, respectively, and is used to make predictions based on the time-frequency feature matrix, the operating condition parameters, and the environmental parameters using a deep learning prediction model to obtain prediction results; the prediction results include damping ratio and flutter critical velocity; the deep learning prediction model is obtained by training a CNN-LSTM hybrid model using sample fusion feature vectors and sample prediction results; The result output and early warning module is connected to the deep learning prediction module and is used to display the prediction results, generate a judgment signal based on the prediction results, and issue an early warning based on the judgment signal.
2. The tiltrotor aeroelastic stability early warning device based on HHT and deep learning according to claim 1, characterized in that, The raw signals include: raw aeroelastic response signals, raw operating condition parameters, and raw environmental parameters; the raw aeroelastic response signals include: raw strain signals, raw acceleration signals, and raw aerodynamic pressure signals; the raw operating condition parameters include: rotor speed signals, tilt angle signals, and forward velocity signals; the raw environmental parameters include raw temperature signals and raw humidity signals; the sensing module includes: A strain sensor is installed on the wing of the tiltrotor aircraft to collect the raw strain signal of the wing; An acceleration sensor is installed on the rotor blades of the tiltrotor aircraft to collect the raw acceleration signal of the blades; A pressure sensor is installed on the wing of the tiltrotor aircraft to collect raw aerodynamic pressure signals; Operating condition parameter sensors are installed on the tiltrotor aircraft to collect the raw operating condition parameters of the tiltrotor aircraft; An environmental sensor is installed on the tiltrotor aircraft to collect raw environmental parameters.
3. The tiltrotor aeroelastic stability early warning device based on HHT and deep learning according to claim 2, characterized in that, The data preprocessing module includes: An ADC converter, connected to the sensing module, is used to perform analog-to-digital conversion on the original signal to obtain the original digital signal; The microcontroller, connected to the ADC converter, is used to: filter the original digital aeroelastic response signal from the original digital signal using an adaptive wavelet threshold denoising algorithm to obtain a filtered aeroelastic response signal; remove outliers from the original digital operating parameters and original digital environmental parameters in the original digital signal using the 3σ criterion to obtain removed operating parameters and removed environmental parameters; and sequentially perform time synchronization, normalization, and segmentation processing on the filtered aeroelastic response signal, removed operating parameters, and removed environmental parameters to obtain multiple preprocessed signals.
4. The tiltrotor aeroelastic stability early warning device based on HHT and deep learning according to claim 3, characterized in that, The feature extraction module includes: The decomposition unit, connected to the microcontroller, is used to perform EMD decomposition on the aeroelastic response signal in each segment of preprocessed signal to obtain multiple mode functions and a residual component. A transformation unit, connected to the decomposition unit, is used to perform Hilbert transform on each mode function to obtain the analytic signal of each mode function; the analytic signal includes instantaneous amplitude and instantaneous phase; The calculation unit, connected to the transformation unit, is used to calculate the instantaneous frequency and instantaneous damping ratio of each mode function based on the analytical signal of each mode function; The feature extraction unit, connected to the transformation unit, is used to extract features from each mode function to obtain the statistical features of each mode function; The integration unit is connected to the transformation unit, the calculation unit and the feature extraction unit respectively, and is used to integrate the instantaneous amplitude, instantaneous frequency and instantaneous damping ratio and statistical characteristics of each modal function to obtain the time-frequency feature matrix.
5. The tiltrotor aeroelastic stability early warning device based on HHT and deep learning according to claim 1, characterized in that, The deep learning prediction module includes: The dimension reduction module unit, connected to the feature extraction module, is used to reduce the dimension of the time-frequency feature matrix using principal component analysis to obtain the time-frequency feature vector. The splicing unit is connected to the dimensionality reduction module and the data preprocessing module respectively, and is used to splice the time-frequency feature vector, the operating condition parameters and the environmental parameters to obtain the fused feature vector; The prediction unit, connected to the splicing unit, is used to input the fused feature vector into the deep learning prediction model to obtain the prediction result.
6. The tiltrotor aeroelastic stability early warning device based on HHT and deep learning according to claim 1, characterized in that, The CNN-LSTM hybrid model comprises a first convolutional layer, a second convolutional layer, an LSTM layer, a first fully connected layer, and a second fully connected layer connected in sequence.
7. The tiltrotor aeroelastic stability early warning device based on HHT and deep learning according to claim 1, characterized in that, The result output and early warning module includes: A Bluetooth communication unit, connected to the deep learning prediction module, is used for wireless transmission of the prediction results; A display screen, connected to the Bluetooth communication unit, is used to display the prediction results; A determination unit, connected to the Bluetooth communication unit, is used to generate a determination signal based on the prediction result; the determination signal includes: a stable signal, a metastable signal, and an unstable signal. An alarm buzzer, connected to the determination unit, is used to issue an early warning when the determination signal is an unstable signal.
8. A tiltrotor aeroelastic stability early warning method based on HHT and deep learning, applied to the tiltrotor aeroelastic stability early warning device based on HHT and deep learning as described in any one of claims 1-7, characterized in that, The tiltrotor aeroelastic stability early warning method based on HHT and deep learning includes: Acquire the raw signals from the tiltrotor aircraft; The original signal is preprocessed to obtain multiple preprocessed signals; each preprocessed signal includes: aeroelastic response signal, operating condition parameters and environmental parameters; HHT was used to extract features from the aeroelastic response signals in each preprocessed signal segment to obtain the time-frequency feature matrix; Based on the time-frequency feature matrix, the operating condition parameters, and the environmental parameters, a deep learning prediction model is used to make predictions, and the prediction results are obtained. The prediction results include the damping ratio and the flutter critical velocity. The deep learning prediction model is obtained by training a CNN-LSTM hybrid model using sample fusion feature vectors and sample prediction results. A judgment signal is generated based on the prediction result, and an early warning is issued based on the judgment signal.
9. The tiltrotor aeroelastic stability prediction method based on HHT and deep learning according to claim 8, characterized in that, HHT was used to extract features from the aeroelastic response signals in each preprocessed signal segment to obtain the time-frequency feature matrix, which specifically includes: EMD decomposition was performed on the aeroelastic response signals in each preprocessed signal segment to obtain multiple mode functions and a residual component. The Hilbert transform is applied to each mode function to obtain the analytic signal of each mode function; the analytic signal includes instantaneous amplitude and instantaneous phase. Based on the analytical signals of each mode function, calculate the instantaneous frequency and instantaneous damping ratio of each mode function; Feature extraction is performed on each mode function to obtain its statistical characteristics; The instantaneous amplitude, instantaneous frequency, instantaneous damping ratio, and statistical characteristics of each modal function are integrated to obtain the time-frequency characteristic matrix.
10. The tiltrotor aeroelastic stability prediction method based on HHT and deep learning according to claim 8, characterized in that, Based on the time-frequency feature matrix, the operating condition parameters, and the environmental parameters, a deep learning prediction model is used to make predictions, and the prediction results are obtained, specifically including: Principal component analysis is used to reduce the dimensionality of the time-frequency feature matrix to obtain the time-frequency feature vector. The time-frequency feature vector, the operating condition parameters, and the environmental parameters are concatenated to obtain a fused feature vector. The fused feature vector is input into the deep learning prediction model to obtain the prediction result.