Anti-interference method for suppressing flight multi-type vibration by carrying acoustic module on unmanned aerial vehicle

Abstract

The invention discloses an anti-interference method for suppressing flight multi-type vibration of an acoustic module carried by an unmanned aerial vehicle, and the method comprises the following steps: vibration detection and spectrum analysis, active vibration control and digital signal processing, and the vibration detection and spectrum analysis comprises the following steps: data preprocessing, signal conversion and frequency analysis. The active vibration control comprises a backstepping control method and an RBF network adaptive algorithm, and the digital signal processing specifically comprises the step of reconstructing a signal by adopting a deep learning neural network. According to the anti-interference method for suppressing flight multi-type vibration of the acoustic module carried by the unmanned aerial vehicle, the multi-type vibration generated in the flight of the unmanned aerial vehicle can be suppressed in the whole process, the interference of the vibration on the detection data of the acoustic module is effectively reduced, and the accuracy of detecting a loosening signal of a wire clamp bolt is improved; and the operation stability and the signal processing capability of the unmanned aerial vehicle in a complex environment are enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence signal processing, and particularly relates to a method for suppressing multi-type vibrations and anti-interference of an unmanned aerial vehicle (UAV) equipped with an acoustic module. Background Art

[0002] UAVs are widely used in multiple fields. However, during flight, especially in complex airflow and changing environments, they are often affected by various interferences, and vibration interference is one of the main factors affecting the accuracy of sensors carried by UAVs. Mechanical vibration can cause the data accuracy of sensors to decline, thereby affecting the judgment and execution of the flight control system.

[0003] Solving the above problems involves technologies in multiple aspects, such as vibration detection and analysis, active vibration control, digital signal processing, etc. In terms of vibration detection and analysis, a method for detecting UAV attitude vibration based on time-frequency analysis uses the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMD) technique to decompose the vibration signals in the attitude data of an unmanned helicopter in detail. By analyzing the energy entropy of each Intrinsic Mode Function (IMF), the signal is reconstructed, and a time-frequency analysis method combining CEEMD and wavelet analysis is used to explore the vibration characteristics of the UAV during flight in detail. Combining the CEEMD method with energy entropy can effectively eliminate the interference of low peak points on spectrum analysis, thereby optimizing the reconstruction of UAV vibration data. By analyzing the vibration characteristics in three attitude directions of roll, pitch, and yaw, and extracting the natural vibration frequencies in each direction, this method can effectively identify the vibration characteristics of the UAV during flight. It has important research significance for improving the stability of UAVs in the future.

[0004] In terms of active vibration control, a new control method for a four-rotor UAV for power line inspection is designed. This method combines the Active Disturbance Rejection Control (ADRC) algorithm and an improved fuzzy PID controller. By using the ADRC algorithm with excellent anti-interference characteristics to design the inner-loop attitude controller of the UAV, smooth transitions of the yaw, pitch, and roll attitude angles to the initial set values are achieved, and the variation range is controlled within 3 degrees. This not only enables the UAV to quickly return to a stable state but also demonstrates excellent tracking and fast response capabilities. Dynamically adjusting the output of the actuator according to sensor data to actively cancel the detected vibration can fundamentally improve the detection accuracy.

[0005] Although the existing technical solutions are beneficial or improve the anti-interference control performance in different aspects, they do not combine various technologies in applications, thus not meeting the actual needs and further in-depth research is still required. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for suppressing various types of vibrations generated during the flight of an unmanned aerial vehicle (UAV), effectively reducing the interference of these vibrations on the detection data of the acoustic module, improving the accuracy of detecting the loosening signal of the clamp bolt, and enhancing the operation stability and signal processing ability of the UAV in a complex environment.

[0007] The technical solution of the present invention is as follows:

[0008] A method for suppressing multi-type vibrations and anti-interference during the flight of a UAV carrying an acoustic module, comprising the following steps: vibration detection and spectrum analysis, active vibration control, and digital signal processing. The vibration detection and spectrum analysis include the following steps: data preprocessing, signal conversion, and frequency analysis. The active vibration control includes: backstepping control method and RBF network adaptive algorithm. The digital signal processing specifically is: reconstructing the signal using a deep learning neural network;

[0009] The data preprocessing includes the following steps:

[0010] S1. Signal acquisition, where the signal is the signal collected by the acoustic module carried by the UAV;

[0011] S2. Noise removal: removing irrelevant high-frequency noise or environmental interference;

[0012] The signal conversion includes the following steps:

[0013] S3. Converting the time-domain signal into a frequency-domain signal, that is, extracting the amplitude and phase information of different frequency components from the data of the vibration intensity changing with time;

[0014] S4. Identifying the amplitude and phase of each frequency component;

[0015] The frequency analysis includes the following steps:

[0016] S5. Identifying the main vibration frequencies of the UAV and their corresponding amplitude sizes;

[0017] S6. Comparing the obtained spectrum with the data during the stable flight of the UAV to identify faults;

[0018] The backstepping control method includes the following steps:

[0019] S7. Hierarchical control design: decomposing the complex system into an attitude control system and a position control system;

[0020] S8. Virtual control: Design a control objective to make its state variables reach a predetermined stable value, that is: first determine the error between the system state and the desired state, and then design a virtual control law to reduce this error. The virtual control law makes the actual pitch angle approach the desired pitch angle;

[0021] S9. Error function: Design an error function for each error variable to judge the system stability, that is: quantify the "energy" distance between the current state and the desired state of the system, which is the square of the error. Its positive definiteness confirms that the greater the error, the higher the energy state of the system;

[0022] The RBF network adaptive algorithm includes the following steps:

[0023] S10. Initialization: Initialize the RBF network, and the network parameters are randomly selected;

[0024] S11. Input data: Feed the current system state as input data into the RBF network. The current system state includes: the attitude, speed, and position of the UAV;

[0025] S12. Network output: The RBF network calculates the output according to the current input data and network parameters. This output represents an estimate of the external interference and model uncertainty received by the current system;

[0026] S13. Adjust control input: Use the output estimate of the RBF network to adjust the control input of the UAV. By changing the control command, the influence of external interference is offset, thereby maintaining the stability and predetermined tracking performance of the UAV;

[0027] S14. Update network parameters: Update the parameters of the RBF network according to the current input data and system response;

[0028] S15. Evaluate system performance: Monitor the real-time performance of the UAV, including its stability and tracking error;

[0029] S16. Meet the end condition: If the performance of the UAV meets the preset standard, continue normal operation;

[0030] The reconstruction of the signal using a deep learning neural network includes the following steps:

[0031] S17. Feature extraction: Extract according to the spectrum analysis of vibration data, and extract the features that are most useful for predicting vibration;

[0032] S18. Model training and tuning: Use a large amount of historical data to train the model so that it learns the relationship between vibration and noise. Among them, the parameters of the model are carefully adjusted to obtain the best performance;

[0033] S19. Model evaluation condition: Calculate the mean square error or absolute error between the reconstructed signal and the original signal to evaluate the prediction accuracy of the model;

[0034] S20. Signal reconstruction: Use the trained model to identify the noise part in the signal, remove or reduce it, and retain the useful information at the same time;

[0035] S21. Model update: Evaluate the performance of the model by comparing the signals before and after reconstruction, and using the signal-to-noise ratio improvement and prediction accuracy.

[0036] Furthermore, the irrelevant high-frequency noise in step S2 is the vibration frequency of bolt loosening, and the method for removing noise in step S2 includes the following steps:

[0037] (1). Frequency measurement: Identify the vibration frequency range f caused by bolt loosening through fast Fourier transform bolt ;

[0038] (2). Frequency range setting: Determine the frequency range [f1, f2] that the band-stop filter needs to suppress according to the analysis;

[0039] (3). Filter using a second-order band-stop filter, and the expression of the second-order band-stop filter is: y[n] = x[n] - 2cos(2πf bolt / f s )x[n - 1] + x[n - 2] + 2r cos(2πf bolt / f s )y[n - 1] - r 2 y[n - 2],

[0040] where y[n] represents the output signal after filtering, x[n] represents the input signal, fbol t represents the central vibration frequency of bolt loosening, f s represents the sampling frequency, r represents the damping coefficient of the filter, and n represents the time index;

[0041] (4). Implement filtering: Implement band-stop filtering through the filter formula or using digital signal processing tools.

[0042] Furthermore, the method for converting the time-domain signal to the frequency-domain signal in step S3 is to use fast Fourier transform to convert the time-domain signal to the frequency-domain signal, which specifically includes the following steps:

[0043] (1). Obtain the time-domain signal: The collected time-domain signal is represented as a discrete signal, that is:

[0044] x[n], n = 0, 1,..., N - 1,

[0045] Among them, x[n] represents the value of the time-domain signal at the nth sampling point, and N represents the number of sampling points;

[0046] Sampling frequency f s is defined as the number of samples per second. Therefore, the sampling interval is:

[0047]

[0048] The sampling duration of the signal is:

[0049] T = N·Δt,

[0050] where Δt represents the sampling interval;

[0051] (2) Fast Fourier Transform, and its expression is:

[0052] k = 0, 1, …, N - 1,

[0053] where X[k] represents the kth frequency component in the frequency domain, including amplitude and phase information, x[n] represents the value of the time-domain signal at the nth sampling point, N represents the number of sampling points, k represents the frequency index, and the corresponding frequency is: k = 0, 1, ..., N / 2,

[0054] where f s represents the sampling frequency, N represents the number of sampling points, k represents the frequency index, and f k represents the frequency value corresponding to the kth frequency component;

[0055] (3) Calculate the amplitude and phase of the frequency-domain signal. The calculation formula for the amplitude is:

[0056]

[0057] where A[k] represents the amplitude of the kth frequency component, X[k] represents the kth frequency component in the frequency domain, Re(X[k]) represents the real part of X[k], and Im(X[k]) represents the imaginary part of X[k];

[0058] The expression for the phase is:

[0059]

[0060] where φ[k] represents the phase angle of the kth frequency component, X[k] represents the kth frequency component in the frequency domain, Re(X[k]) represents the real part of X[k], and Im(X[k]) represents the imaginary part of X[k].

[0061] Furthermore, the spectrum data comparison in step S6 includes:

[0062] (1), Difference calculation: Calculate the difference between the real-time spectrum and the standard spectrum, i.e.:

[0063] ΔA[k] = |S real [k]| - |S normal [k]|,

[0064] where, |S real [k]| represents the amplitude of the k-th frequency component of the real-time spectrum, |S normal [k]| represents the amplitude of the k-th frequency component of the reference spectrum, and ΔA[k] represents the difference amplitude;

[0065] (2), Phase difference calculation, the calculation formula for the phase difference is:

[0066] Δφ[k] = φ real [k] - φ normal [k],

[0067] where, φ real [k] represents the phase of the k-th frequency component of the real-time spectrum, φ normal [k] represents the phase of the k-th frequency component of the reference spectrum, and Δφ[k] represents the phase difference;

[0068] (3), Define the threshold: Set the difference thresholds ΔA threshold and Δφ threshold , and determine whether the data is abnormal;

[0069] (4), Abnormal frequency detection: If |ΔA[k]| > ΔA threshold or |Δφ[k]| > Δφ threshold , then it is considered that the k-th frequency component is abnormal;

[0070] The identification of faults in step S6 includes the following steps:

[0071] (1), Spectrum feature analysis: According to the position and amplitude of the abnormal frequency component, judge the fault mode, i.e.:

[0072] New high-frequency component: Component looseness or resonance;

[0073] Enhanced low-frequency component: Rotor imbalance or external disturbance;

[0074] Harmonic component change: Motor fault or vibration coupling;

[0075] (2), Match with the fault mode: Use the spectrum data to compare with the existing fault library to judge the specific fault type.

[0076] Furthermore, the determination of the system state and error in step S8 includes the following steps:

[0077] (1) System status: The pitch angle system of the UAV is described by the following state variables: the actual pitch angle θ(t), the angular velocity of the pitch angle the desired pitch angle θ d (t);

[0078] (2) Error definition: Define the system error as:

[0079] e(t) = θ(t) - θ d (t),

[0080] where e(t) represents the error between the current state and the desired state of the system, θ(t) represents the actual pitch angle, and θ d (t) represents the desired pitch angle;

[0081] Then the error dynamics is:

[0082]

[0083] where, represents the rate of change of the state error, represents the angular velocity of the pitch angle, represents the angular velocity of the desired pitch angle;

[0084] The design method of the virtual control law in step S8 is as follows:

[0085] (1) Set the virtual control variable: Let the virtual control variable v be used to guide the system state to approach the desired state, and it is defined as:

[0086]

[0087] where v represents the virtual control input, k1 represents a positive gain coefficient, represents the angular velocity of the desired pitch angle;

[0088] (2) Design the virtual control objective: Design the virtual control objective so that the error e(t) points to zero, that is:

[0089] After substituting the expression of v, we have:

[0090]

[0091] where, represents the angular acceleration of the pitch angle, k1 represents a positive gain coefficient, represents the angular velocity of the desired pitch angle, θ(t) represents the actual pitch angle, and θ d (t) represents the desired pitch angle;

[0092] (3) Control rate design: By designing the actual control input u, the virtual control law v is achieved. The system dynamics are represented by a second-order model:

[0093]

[0094] where J represents the moment of inertia, b represents the damping coefficient, u represents the actual control input, represents the angular acceleration of the pitch angle, represents the angular velocity of the desired pitch angle;

[0095] Substituting v into the control objective, we have:

[0096]

[0097] The final control input u is:

[0098]

[0099] where J represents the moment of inertia, b represents the damping coefficient, k1 represents a positive gain coefficient, θ(t) represents the actual pitch angle, θ d (t) represents the desired pitch angle, represents the angular velocity of the pitch angle, represents the angular velocity of the desired pitch angle.

[0100] Furthermore, the design method of the error function and the system stability discrimination method in step S9 are as follows:

[0101] (1) Error function definition: Define the error function V(e) to represent the "energy" distance of the system error:

[0102] e = θ(t) - θ d (t),

[0103] where V(e) represents the Lyapunov function, e represents the state error, that is, the error between the current state and the desired state of the system, θ(t) represents the actual pitch angle, θ d (t) represents the desired pitch angle;

[0104] The positive definiteness in step S9 means that V(e) ≥ 0 holds for all e, and V(e) = 0 if and only if e = 0, that is, the energy state of the system is related to the error, and the larger the error, the higher the energy state;

[0105] (2) Calculate the derivative of the error function. The calculation formula for the derivative of the error function is:

[0106]

[0107] where, represents the derivative of the Lyapunov function with respect to time, and e represents the state error, that is, the error between the current state of the system and the desired state. represents the rate of change of the error, that is, the derivative of the error with respect to time. k1 represents a positive gain coefficient. The larger k1 is, the faster the error converges, and the larger the error is. When, the energy state of the system is decreasing and the error is converging. When, the system reaches the equilibrium point and the error is zero.

[0108] (3) System stability analysis: Select the Lyapunov function as the error function and verify the sign of the derivative of the error function:

[0109]

[0110]

[0111] When and e≠0, it indicates that the system is asymptotically stable.

[0112] Further, the network parameters in step S10 include:

[0113] Network center c i : Used to define the position of each radial basis function;

[0114] Network width σ i : Used to control the range of action of the radial basis function;

[0115] Weight w i : Used to adjust the weight of the output;

[0116] The network structure of the RBF network in step S10 is:

[0117]

[0118]

[0119] Among them, f(x) represents the output of the network, x represents the input data, that is, the current system state, N represents the number of radial basis functions, w i represents the weight of the i-th radial basis function, and φ i (x) represents the i-th radial basis function, and b represents the bias term;

[0120] The expression of the network output in step S12 is:

[0121]

[0122]

[0123] Among them, y represents the network output, that is, the estimated value of external interference and model uncertainty;

[0124] The method for adjusting the control input of the UAV in step S13 is: according to the output y of the RBF network, correct the control input of the UAV, that is:

[0125] u = u0 + Δu,

[0126] Δu = -ky,

[0127] Among them, u represents the adjusted control input, u0 represents the initial control input, Δu represents the compensation control input, k represents the control gain, and y represents the network output;

[0128] The method for updating the parameters of the RBF network in step S14 is: update the center c of the network according to the error i and the width σ i , and use the gradient descent method to update the weight w i , and the expression of the gradient descent method is:

[0129]

[0130]

[0131] Among them, η represents the learning rate, J represents the error cost function, y represents the network output, y target represents the expected network output value, w i (t) represents the weight of the i-th radial basis function at time t, w i (t + 1) represents the updated weight of the i-th radial basis function at time t + 1, represents the partial derivative of the error cost function J with respect to the weight w i ;

[0132] The expression of the tracking error in step S15 is:

[0133]

[0134] Among them, d represents the dimension of the state variable, that is, the number of parameters included in the system state, x d [i] represents the value of the i-th state variable in the expected state of the system, and x[i] represents the value of the i-th state variable in the actual state of the system;

[0135] The preset criterion in step S16 is: the tracking error is lower than the threshold, that is:

[0136] ‖tracking error‖ 〈 ε,

[0137] Among them, ∈ represents the threshold, that is, the maximum allowable error by which the system state can deviate from the desired state.

[0138] Furthermore, the feature extraction in step S17 includes the following steps:

[0139] (1) Normalize the time-domain signal x(t):

[0140]

[0141]

[0142]

[0143] where x norm (t) represents the signal value after normalizing the time-domain signal x(t), μ represents the mean of the time-domain signal x(t), N represents the number of signal sampling points, and σ represents the standard deviation of the time-domain signal x(t);

[0144] (2) Spectrum analysis: Use the fast Fourier transform to extract the frequency-domain feature X[k];

[0145] (3) Construct a feature matrix: Combine the time-domain and frequency-domain features to form a feature matrix F:

[0146] F = [x norm , X[k], f1, f2,..., f m}

[0147] where f1, f2,..., f m represent statistical features, including peak value, mean value, and variance;

[0148] The model in step S18 is a DNN deep neural network, and the structure of the DNN deep neural network includes:

[0149] Input layer: Feature matrix F, whose dimension is [N, d], where N represents the number of samples and d represents the feature dimension;

[0150] Hidden layer: Introduce non-linearity using an activation function, that is:

[0151] hi = ReLU(W i ·h i-1 + b i ),

[0152] where h i represents the output of the i-th layer, W i represents the weight matrix, and b i represents the bias vector;

[0153] Output layer: Reconstructed signal The same length as the input signal x(t);

[0154] The model parameter adjustment in step S18 includes: adjusting hyperparameters, where the hyperparameters include: the number of network layers, the number of neurons, the learning rate, and the batch size; using an optimizer to update the network parameters, that is:

[0155]

[0156] Among them, W i (t) represents the parameter value at the current moment t, that is, the weight value, and W i (t + 1) represents the parameter value at the next moment t + 1, that is, the updated weight value, η represents the learning rate, and L represents the loss function;

[0157] The expressions for the mean square error and the absolute error in step S19 are:

[0158]

[0159]

[0160] Among them, L MSE represents the mean square error, that is, the average value of the squares of the errors between the predicted signal and the actual signal x i , and L MAE represents the mean absolute error, that is, the average value of the absolute values of the errors between the predicted signal and the actual signal xi, N represents the number of samples, and x i represents the actual value of the original signal at the i-th sampling point, represents the predicted value of the model output signal at the i-th sampling point;

[0161] The expression for the reconstructed signal in step S19 is:

[0162]

[0163] Among them, represents the reconstructed signal, x(t) represents the original signal, and n(t) represents the noise signal;

[0164] The expression for the signal-to-noise ratio improvement in step S21 is:

[0165] ΔSNR = SNR after -SNR before ,

[0166] Among them, ΔSNR represents the amplitude of the improvement in the signal-to-noise ratio of the signal after denoising processing, and SNR after represents the signal-to-noise ratio of the signal after denoising processing, and SNR beforerepresents the quality of the signal before denoising processing;

[0167] The expression of the signal-to-noise ratio is:

[0168]

[0169] where SNR represents the signal-to-noise ratio, and x i represents the actual value of the original signal at the i-th sampling point, represents the predicted value of the model output signal at the i-th sampling point;

[0170] The standard of the prediction accuracy in step S21 is: the error is less than the preset threshold ∈.

[0171] Further, the most useful features in step S17 refer to: the frequency-domain features and time-domain features obtained by spectral analysis, and the useful information in step S20 refers to: the effective signal after denoising.

[0172] Advantages of the present invention:

[0173] 1. Application of spectral analysis technology: Through fast Fourier transform (FFT), key frequency components are efficiently extracted from complex vibration signals, the main vibration sources and frequency characteristics are identified, providing reliable data support for subsequent vibration control and

[0174] fault diagnosis;

[0175] 2. Combination of backstepping method and RBF network adaptive algorithm: The backstepping method ensures the stability of each subsystem through a step-by-step backward design; the RBF network adaptive algorithm enhances the anti-interference performance of the system by online estimating and compensating external disturbances. This not only optimizes the dynamic response of the UAV but also improves its adaptability and stability in a changing environment;

[0176] 3. Advanced signal processing and denoising technology: The deep neural network (DNN) based on the present invention can automatically extract the optimal features in the time domain and frequency domain, remove the noise part, retain the effective signal, and optimize the model using loss functions (MSE and MAE) to improve the denoising performance. At the same time, the denoising effect is evaluated by the signal-to-noise ratio (SNR) to ensure a significant improvement in the quality of the reconstructed signal;

[0177] 4. High reliability and user-friendliness: The algorithm design of the present invention based on Lyapunov stability theory and gradient descent method ensures the reliability of control and optimization. The system performance is quantified through indicators such as tracking error and signal-to-noise ratio improvement, facilitating user evaluation and improvement;

[0178] 5. Integration of the overall technical solution: Through the combination of the above technologies, the present invention can suppress various types of vibrations generated during the flight of the unmanned aerial vehicle (UAV) throughout the process, effectively reducing the interference of these vibrations on the detection data of the acoustic module. The present invention not only improves the accuracy of detecting the signal of the loosening of the clamp bolt, but also enhances the operation stability and signal processing ability of the UAV in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0179] Figure 1 is a flowchart of the process steps of a method for a UAV carrying an acoustic module to suppress multi-type vibrations during flight and anti-interference.

[0180] Figure 2 is a flowchart of the process steps of vibration detection and spectrum analysis of a method for a UAV carrying an acoustic module to suppress multi-type vibrations during flight and anti-interference.

[0181] Figure 3 is a flowchart of the process steps of the backstepping control method of a method for a UAV carrying an acoustic module to suppress multi-type vibrations during flight and anti-interference.

[0182] Figure 4 is a flowchart of the process steps of the RBF network adaptive algorithm of a method for a UAV carrying an acoustic module to suppress multi-type vibrations during flight and anti-interference.

[0183] Figure 5 Flowchart of the digital signal processing method of a method for a UAV carrying an acoustic module to suppress multi-type vibrations during flight and anti-interference. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0184] As Figures 1-5 shown, a method for a UAV carrying an acoustic module to suppress multi-type vibrations during flight and anti-interference includes the following steps: vibration detection and spectrum analysis, active vibration control, and digital signal processing. The vibration detection and spectrum analysis include the following steps: data preprocessing, signal conversion, and frequency analysis. The active vibration control includes: backstepping control method and RBF network adaptive algorithm. The digital signal processing specifically is: using a deep learning neural network to reconstruct the signal;

[0185] The data preprocessing includes the following steps:

[0186] S1. Signal acquisition, where the signal is the signal collected by the acoustic module carried by the UAV;

[0187] S2. Noise removal: Removing irrelevant high-frequency noise or environmental interference;

[0188] The signal conversion includes the following steps:

[0189] S3. Convert the time-domain signal into a frequency-domain signal, that is, extract the amplitude and phase information of different frequency components from the data of the vibration intensity changing with time;

[0190] S4. Identify the amplitude and phase of each frequency component;

[0191] The frequency analysis includes the following steps:

[0192] S5. Identify the main vibration frequencies of the drone and their corresponding amplitude sizes;

[0193] S6. Compare the obtained frequency spectrum with the data during the stable flight of the drone to identify faults;

[0194] The backstepping control method includes the following steps:

[0195] S7. Hierarchical control design: Decompose the complex system into an attitude control system and a position control system;

[0196] S8. Virtual control: Design a control target to make its state variables reach a predetermined stable value, that is: first determine the error between the system state and the desired state, and then design a virtual control law to reduce this error. The virtual control law makes the actual pitch angle approach the desired pitch angle;

[0197] S9. Error function: Design an error function for each error variable to judge the system stability, that is: quantify the "energy" distance between the current state and the desired state of the system, that is, the square of the error. Its positive definiteness confirms that the greater the error, the higher the energy state of the system;

[0198] The RBF network adaptive algorithm includes the following steps:

[0199] S10. Initialization: Initialize the RBF network, and the network parameters are randomly selected;

[0200] S11. Input data: Send the current system state as input data into the RBF network. The current system state includes: the attitude, speed, and position of the drone;

[0201] S12. Network output: The RBF network calculates the output according to the current input data and network parameters. This output represents an estimate of the external interference and model uncertainty received by the current system;

[0202] S13. Adjust the control input: Use the output estimate of the RBF network to adjust the control input of the drone. By changing the control command, offset the influence of external interference, and thus maintain the stability and predetermined tracking performance of the drone;

[0203] S14. Update network parameters: Update the parameters of the RBF network according to the current input data and system response;

[0204] S15. Evaluate the system performance: Monitor the real-time performance of the drone, including its stability and tracking error;

[0205] S16. Meet the end condition: If the performance of the drone meets the preset standard, continue normal operation;

[0206] The signal reconstruction using the deep learning neural network includes the following steps:

[0207] S17. Feature extraction: Extract the vibration data according to the spectrum analysis, and extract the features most useful for predicting vibration from it;

[0208] S18. Model training and tuning: Use a large amount of historical data to train the model so that it learns the relationship between vibration and noise, and carefully adjust the parameters of the model to obtain the best performance;

[0209] S19. Model evaluation condition: Calculate the mean square error or absolute error between the reconstructed signal and the original signal to evaluate the prediction accuracy of the model;

[0210] S20. Signal reconstruction: Use the trained model to identify the noise part in the signal, remove or reduce it, and retain useful information at the same time;

[0211] S21. Model update: Evaluate the performance of the model by comparing the signals before and after reconstruction, and using the signal-to-noise ratio improvement and prediction accuracy.

[0212] Preferably, the non-correlated high-frequency noise in step S2 is the vibration frequency of bolt loosening, and the method for removing noise in step S2 includes the following steps:

[0213] (1). Frequency measurement: Identify the vibration frequency range f caused by bolt loosening through fast Fourier transform bolt ;

[0214] (2). Frequency range setting: Determine the frequency range [f1, f2] that the band-stop filter needs to suppress according to the analysis;

[0215] (3). Filter using a second-order band-stop filter, and the expression of the second-order band-stop filter is:

[0216] y[n] = x[n] - 2cos(2πf bolt / f s )x[n - 1] + x[n - 2] + 2r cos(2πf bolt / f s )y[n - 1] - r 2 y[n - 2],

[0217] Among them, y[n] represents the filtered output signal, x[n] represents the input signal, and f bolt represents the central vibration frequency of bolt loosening, and f s represents the sampling frequency, r represents the damping coefficient of the filter, and n represents the time index;

[0218] (4) Realize filtering: Band-stop filtering is realized through the filter formula or by using digital signal processing tools.

[0219] Preferably, the method of converting the time-domain signal into a frequency-domain signal in step S3 is to convert the time-domain signal into a frequency-domain signal by using the fast Fourier transform, which specifically includes the following steps:

[0220] (1) Obtain the time-domain signal: The collected time-domain signal is represented as a discrete signal, that is:

[0221] x[n], n = 0, 1,..., N - 1,

[0222] where x[n] represents the value of the time-domain signal at the nth sampling point, and N represents the number of sampling points;

[0223] The sampling frequency f s is defined as the number of samples per second. Therefore, the sampling interval is:

[0224]

[0225] The sampling duration of the signal is:

[0226] T = N·Δt,

[0227] where Δt represents the sampling interval;

[0228] (2) Fast Fourier transform, and its expression is:

[0229] k = 0, 1,..., N - 1,

[0230] where X[k] represents the kth frequency component in the frequency domain, including amplitude and phase information, x[n] represents the value of the time-domain signal at the nth sampling point, N represents the number of sampling points, k represents the frequency index, and the corresponding frequency is: k = 0, 1,..., N / 2,

[0231] where f s represents the sampling frequency, N represents the number of sampling points, k represents the frequency index, and f k represents the frequency value corresponding to the kth frequency component;

[0232] (3) Calculate the amplitude and phase of the frequency-domain signal. The calculation formula for the amplitude is:

[0233]

[0234] Among them, A[k] represents the amplitude of the k-th frequency component, X[k] represents the k-th frequency component in the frequency domain, Re(X[k]) represents the real part of X[k], and Im(X[k]) represents the imaginary part of X[k];

[0235] The expression for the phase is:

[0236]

[0237] Among them, φ[k] represents the phase angle of the k-th frequency component, X[k] represents the k-th frequency component in the frequency domain, Re(X[k]) represents the real part of X[k], and Im(X[k]) represents the imaginary part of X[k].

[0238] Preferably, the spectrum data comparison in step S6 includes:

[0239] (1) Difference calculation: Calculate the difference between the real-time spectrum and the standard spectrum, that is:

[0240] ΔA[k] = |S real [k]| - |S normal [k]|,

[0241] Among them, |S real [k]| represents the amplitude of the k-th frequency component of the real-time spectrum, |S normal [k]| represents the amplitude of the k-th frequency component of the reference spectrum, and ΔA[k] represents the difference amplitude;

[0242] (2) Phase difference calculation. The calculation formula for the phase difference is:

[0243] Δφ[k] = φ real [k] - φ normal [k],

[0244] Among them, φ real [k] represents the phase of the k-th frequency component of the real-time spectrum, φ normal [k] represents the phase of the k-th frequency component of the reference spectrum, and Δφ[k] represents the phase difference;

[0245] (3) Define a threshold: Set the difference thresholds ΔA threshold and Δφ threshold , and determine whether the data is abnormal;

[0246] (4) Abnormal frequency detection: If |ΔA[k]| > ΔA threshold or |Δφ[k]| > Δφ threshold , it is considered that the k-th frequency component is abnormal;

[0247] The identification of faults in step S6 includes the following steps:

[0248] (1) Spectrum feature analysis: Based on the position and amplitude of abnormal frequency components, judge the fault mode, that is:

[0249] New high-frequency components: Component looseness or resonance;

[0250] Enhanced low-frequency components: Rotor imbalance or external disturbance;

[0251] Change in harmonic components: Motor fault or vibration coupling;

[0252] (2) Matching with the fault mode: Use the spectrum data to compare with the existing fault library to judge the specific fault type.

[0253] Preferably, the determination of the system state and error in step S8 includes the following steps:

[0254] (1) System state: The pitch angle system of the unmanned aerial vehicle is described by the following state variables: The actual pitch angle θ(t), the angular velocity of the pitch angle The desired pitch angle θ d (t);

[0255] (2) Error definition: Define the system error as:

[0256] e(t) = θ(t) - θ d (t),

[0257] where e(t) represents the error between the current state and the desired state of the system, θ(t) represents the actual pitch angle, and θ d (t) represents the desired pitch angle;

[0258] Then the error dynamics is:

[0259]

[0260] where, represents the change rate of the state error, represents the angular velocity of the pitch angle, represents the angular velocity of the desired pitch angle;

[0261] The design method of the virtual control rate in step S8 is:

[0262] (1) Set the virtual control variable: Set the virtual control variable v to guide the system state to approach the desired state, defined as:

[0263]

[0264] where, \(v\) represents the virtual control input, and \(k_1\) represents a positive gain coefficient, represents the angular velocity of the desired pitch angle;

[0265] (2) Design the virtual control objective: Design the virtual control objective such that the error \(e(t)\) points to zero, i.e.:

[0266]

[0267] After substituting the expression of \(v\), we have:

[0268]

[0269] where, represents the angular acceleration of the pitch angle, \(k_1\) represents a positive gain coefficient, represents the angular velocity of the desired pitch angle, \(\theta(t)\) represents the actual pitch angle, and \(\theta\) d (t) represents the desired pitch angle;

[0270] (3) Control law design: By designing the actual control input \(u\), the virtual control law \(v\) is implemented. The system dynamics are represented by a second-order model:

[0271]

[0272] where, \(J\) represents the moment of inertia, \(b\) represents the damping coefficient, \(u\) represents the actual control input, represents the angular acceleration of the pitch angle, represents the angular velocity of the desired pitch angle;

[0273] Substitute \(v\) into the control objective, we have:

[0274]

[0275] The final control input \(u\) is:

[0276]

[0277] where, \(J\) represents the moment of inertia, \(b\) represents the damping coefficient, \(k_1\) represents a positive gain coefficient, \(\theta(t)\) represents the actual pitch angle, and \(\theta\) d (t) represents the desired pitch angle, represents the angular velocity of the pitch angle, represents the angular velocity of the desired pitch angle.

[0278] Preferably, the design method of the error function and the system stability discrimination method in step S9 are:

[0279] (1) Error function definition: Define the error function \(V(e)\) to represent the "energy" distance of the system error:

[0280] e = θ(t) - θ d (t),

[0281] where V(e) represents the Lyapunov function, e represents the state error, i.e., the error between the current state of the system and the desired state, θ(t) represents the actual pitch angle, and θ d (t) represents the desired pitch angle;

[0282] The positive definiteness in step S9 means that V(e) ≥ 0 holds for all e, and V(e) = 0 if and only if e = 0, i.e., the energy state of the system is related to the error, and the greater the error, the higher the energy state;

[0283] (2) Calculate the derivative of the error function. The calculation formula for the derivative of the error function is:

[0284]

[0285] where represents the derivative of the Lyapunov function with respect to time, e represents the state error, i.e., the error between the current state of the system and the desired state, represents the rate of change of the error, i.e., the derivative of the error with respect to time, k1 represents a positive gain coefficient, and the larger k1 is, the faster the error converges, and the greater the error, when, the energy state of the system is decreasing and the error is converging, when, the system reaches the equilibrium point and the error is zero;

[0286] (3) System stability analysis: Select the Lyapunov function as the error function and verify the sign of the derivative of the error function:

[0287]

[0288]

[0289] When and e ≠ 0, it indicates that the system is asymptotically stable.

[0290] Preferably, the network parameters in step S10 include:

[0291] Network center c i : Used to define the position of each radial basis function;

[0292] Network width σ i : Used to control the range of action of the radial basis function;

[0293] Weight w i: Used to adjust the weight of the output;

[0294] The network structure of the RBF network in step S10 is:

[0295]

[0296]

[0297] Among them, f(x) represents the output of the network, x represents the input data, that is, the current system state, N represents the number of radial basis functions, w i represents the weight of the i-th radial basis function, φ i (x) represents the i-th radial basis function, and b represents the bias term;

[0298] The expression of the network output in step S12 is:

[0299]

[0300]

[0301] Among them, y represents the network output, that is, the estimated value of external interference and model uncertainty;

[0302] The method for adjusting the control input of the UAV in step S13 is: According to the output y of the RBF network, correct the control input of the UAV, that is:

[0303] u = u0 + Δu,

[0304] Δu = -ky,

[0305] Among them, u represents the adjusted control input, u0 represents the initial control input, Δu represents the compensation control input, k represents the control gain, and y represents the network output;

[0306] The method for updating the parameters of the RBF network in step S14 is: Update the center c of the network according to the error i and the width σ i , and use the gradient descent method to update the weight w i , the expression of the gradient descent method is:

[0307]

[0308]

[0309] Among them, η represents the learning rate, J represents the error cost function, y represents the network output, y target represents the expected network output value, w i (t) represents the weight of the i-th radial basis function at time t, w i(t + 1) represents the updated weight value of the i-th radial basis function at time t + 1. represents the partial derivative of the error cost function J with respect to the weight w i .

[0310] The expression for the tracking error in step S15 is:

[0311]

[0312] where d represents the dimension of the state variable, that is, the number of parameters included in the system state, and x d [i] represents the value of the i-th state variable in the system desired state, and x[i] represents the value of the i-th state variable in the system actual state;

[0313] The preset criterion in step S16 is: the tracking error is lower than the threshold, that is:

[0314] ||tracking error|| < ∈,

[0315] where ∈ represents the threshold, that is, the maximum allowable error by which the system state can deviate from the desired state.

[0316] Preferably, the feature extraction in step S17 includes the following steps:

[0317] (1) Normalize the time-domain signal x(t):

[0318]

[0319]

[0320]

[0321] where x norm (t) represents the signal value after normalizing the time-domain signal x(t), μ represents the mean of the time-domain signal x(t), N represents the number of signal sampling points, and σ represents the standard deviation of the time-domain signal x(t);

[0322] (2) Spectrum analysis: Use the fast Fourier transform to extract the frequency-domain feature X[k];

[0323] (3) Construct a feature matrix: Combine the time-domain and frequency-domain features to form a feature matrix F:

[0324] F = [x norm , X[k], f1, f2,..., f m ,

[0325] where f1, f2,..., f m represent statistical features, including peak value, mean value, and variance;

[0326] The model in step S18 is a DNN deep neural network, and the structure of the DNN deep neural network includes:

[0327] Input layer: Feature matrix F, whose dimension is [N, d], where N represents the number of samples and d represents the feature dimension;

[0328] Hidden layer: Nonlinearity is introduced using an activation function, that is:

[0329] h i = ReLU(W i ·h i-1 + b i ),

[0330] where h i represents the output of the i-th layer, W i represents the weight matrix, and b i represents the bias vector;

[0331] Output layer: The reconstructed signal is of the same length as the input signal x(t);

[0332] The adjustment of the model parameters in step S18 includes: adjusting hyperparameters, and the hyperparameters include: the number of network layers, the number of neurons, the learning rate, and the batch size; using an optimizer to update the network parameters, that is:

[0333]

[0334] where W i (t) represents the parameter value at the current time t, that is, the weight value, and W i (t + 1) represents the parameter value at the next time t + 1, that is, the weight value after update, η represents the learning rate, and L represents the loss function;

[0335] The expressions for the mean square error and the mean absolute error in step S19 are:

[0336]

[0337]

[0338] where L MSE represents the mean square error, that is, the average of the squares of the errors between the predicted signal and the actual signal x i , and L MAE represents the mean absolute error, that is, the average of the absolute values of the errors between the predicted signal and the actual signal xi, N represents the number of samples, and x i represents the actual value of the original signal at the i-th sampling point, Denotes the predicted value of the model output signal at the \(i\)-th sampling point;

[0339] The expression of the reconstructed signal in step S19 is:

[0340]

[0341] Where, denotes the reconstructed signal, \(x(t)\) denotes the original signal, and \(n(t)\) denotes the noise signal;

[0342] The expression of the SNR improvement in step S21 is:

[0343] \(\Delta SNR = SNR\) after \(- SNR\) before ,

[0344] Where, \(\Delta SNR\) represents the amplitude of the improvement in the signal-to-noise ratio of the signal after denoising processing, \(SNR\) after denotes the signal-to-noise ratio of the signal after denoising processing, \(SNR\) before denotes the quality of the signal before denoising processing;

[0345] The expression of the signal-to-noise ratio is:

[0346]

[0347] Where, \(SNR\) represents the signal-to-noise ratio, \(x\) i denotes the actual value of the original signal at the \(i\)-th sampling point, denotes the predicted value of the model output signal at the \(i\)-th sampling point;

[0348] The criterion for the prediction accuracy in step S21 is: the error is less than the preset threshold \(\epsilon\).

[0349] Preferably, the most useful features in step S17 refer to: the frequency-domain features and time-domain features obtained by spectrum analysis, and the useful information in step S20 refers to: the effective signal after denoising.

[0350] According to the implementation of the above technical solution, this patent can play a role in suppressing multi-type vibration anti-interference of the acoustic module carried by the UAV throughout the process. In the initial stage, spectrum analysis technology is used to detect the vibrations from the UAV or the external environment, and identify the specific amplitudes and frequencies of these vibrations. In the second stage, the RBF network adaptive algorithm based on backstepping control is used to implement active vibration control for the UAV. This algorithm can dynamically adjust the output of the actuator according to the data feedback by the sensor, so that the actuator can generate the necessary force or displacement to offset the vibration effect after receiving the control instruction. In the last stage, by applying machine learning algorithms and learning the correlation between vibrations and noises based on historical data, intelligent reconstruction is performed on the signals interfered by vibrations, aiming to improve the data quality of the signals.

[0351] The specific embodiments described above have further elaborated on the purpose, technical solution and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An anti-interference method for suppressing various types of vibrations during the flight of an unmanned aerial vehicle (UAV) equipped with an acoustic module, characterized in that It includes the following steps: Vibration detection and spectrum analysis, active vibration control, and digital signal processing. The vibration detection and spectrum analysis include the following steps: data preprocessing, signal conversion, and frequency analysis. The active vibration control includes: backstepping control method and RBF network adaptive algorithm. The digital signal processing specifically is: using a deep learning neural network to reconstruct the signal; The data preprocessing includes the following steps: S1. Signal acquisition, where the signal is the signal collected by the acoustic module carried by the drone; S2. Noise removal: removing irrelevant high-frequency noise or environmental interference; The signal conversion includes the following steps: S3. Converting the time-domain signal to a frequency-domain signal, that is, extracting the amplitude and phase information of different frequency components from the data where the vibration intensity changes with time; S4. Identifying the amplitude and phase of each frequency component; The frequency analysis includes the following steps: S5. Identifying the main vibration frequencies of the drone and their corresponding amplitude sizes; S6. Comparing the obtained spectrum with the data during the steady flight of the drone to identify faults; The backstepping control method includes the following steps: S7. Hierarchical control design: decomposing the complex system into an attitude control system and a position control system; S8. Virtual control: designing a control target to make its state variables reach a predetermined stable value, that is: first determining the error between the system state and the desired state, and then designing a virtual control law to reduce this error. The virtual control law makes the actual pitch angle approach the desired pitch angle; S9. Error function: designing an error function for each error variable to judge the system stability, that is: quantifying the "energy" distance between the current state and the desired state of the system, that is, the square of the error. Its positive definiteness confirms that the greater the error, the higher the energy state of the system; The RBF network adaptive algorithm includes the following steps: S10. Initialization: initializing the RBF network, and the network parameters are randomly selected; S11. Input data: sending the current system state as input data into the RBF network. The current system state includes: the attitude, speed, and position of the drone; S12. Network output: The RBF network calculates the output according to the current input data and network parameters. This output represents an estimate of the external interference and model uncertainty received by the current system; S13. Adjusting the control input: using the output estimate of the RBF network to adjust the control input of the drone, and canceling the influence of external interference by changing the control command, thereby maintaining the stability and predetermined tracking performance of the drone; S14. Updating network parameters: updating the parameters of the RBF network according to the current input data and system response; S15. Evaluating system performance: monitoring the real-time performance of the drone, including its stability and tracking error; S16. Meeting the end condition: If the performance of the drone meets the preset standard, continue normal operation; The using of a deep learning neural network to reconstruct the signal includes the following steps: S17. Feature extraction: extracting according to the spectrum analysis of the vibration data, and extracting the features that are most useful for predicting vibrations; S18. Model Training and Tuning: Use a large amount of historical data to train the model so that it learns the relationship between vibration and noise, and carefully adjust the parameters of the model to obtain the best performance; S19. Model Evaluation Conditions: Calculate the mean square error or absolute error between the reconstructed signal and the original signal to evaluate the prediction accuracy of the model; S20. Signal Reconstruction: Use the trained model to identify the noise part in the signal, remove or reduce it, and retain useful information at the same time; S21. Model Update: Evaluate the performance of the model by comparing the signals before and after reconstruction, as well as using the signal-to-noise ratio improvement and prediction accuracy.

2. The anti-interference method for suppressing various types of vibrations during flight by an acoustic module carried by a drone according to claim 1, wherein The irrelevant high-frequency noise in step S2 is the vibration frequency of bolt loosening. The method for removing noise in step S2 includes the following steps: (1) Frequency measurement: Identify the vibration frequency range fbol caused by bolt loosening through fast Fourier transform t ; (2). Frequency Range Setting: According to the analysis, determine the frequency range [f1, f2] that the band-stop filter needs to suppress; (3). Filter with a second-order band-stop filter. The expression of the second-order band-stop filter is: y[n] = x[n] - 2cos(2mf bolt / f s )x[n - 1] + x[n - 2] + 2rcos(2mf bolt / f s )y[n - 1] - r 2 y[n - 2], where y[n] represents the filtered output signal, x[n] represents the input signal, and f bolt represents the central vibration frequency of bolt loosening, and f s represents the sampling frequency, r represents the damping coefficient of the filter, and n represents the time index; (4). Implement Filtering: Implement band-stop filtering through the filter formula or using digital signal processing tools.

3. The anti-interference method for suppressing various types of vibrations during flight by an acoustic module carried by a drone according to claim 1, characterized in that, The method in step S3 for converting the time-domain signal to the frequency-domain signal is to use the fast Fourier transform to convert the time-domain signal to the frequency-domain signal, which specifically includes the following steps: (1). Obtain the Time-Domain Signal: The collected time-domain signal is represented as a discrete signal, that is: r[n], n = 0, 1,..., N - 1, where x[n] represents the value of the time-domain signal at the nth sampling point, and N represents the number of sampling points; Sampling frequency f s is defined as the number of samples per second, so the sampling interval is: The sampling duration of the signal is: T = N·Δt, where Δt represents the sampling interval; (2). Fast Fourier Transform, and its expression is: where X[k] represents the kth frequency component in the frequency domain, including amplitude and phase information, x[n] represents the value of the time-domain signal at the nth sampling point, N represents the number of sampling points, k represents the frequency index, and the corresponding frequency is: where, f s represents the sampling frequency, N represents the number of sampling points, k represents the frequency index, and f k represents the frequency value corresponding to the k-th frequency component; (3). Calculate the Amplitude and Phase of the Frequency-Domain Signal. The calculation formula for the amplitude is: where A[k] represents the amplitude of the kth frequency component, X[k] represents the kth frequency component in the frequency domain, Re(X[k]) represents the real part of X[k], and Im(X[k]) represents the imaginary part of X[k]; The expression for the phase is: where φ[k] represents the phase angle of the kth frequency component, X[k] represents the kth frequency component in the frequency domain, Re(X[k]) represents the real part of X[k], and Im(X[k]) represents the imaginary part of X[k].

4. A method for suppressing multi-type vibrations and anti-interference during the flight of an unmanned aerial vehicle carrying an acoustic module according to claim 1, characterized in that, The spectrum data comparison in step S6 includes: (1). Difference Calculation: Calculate the difference between the real-time spectrum and the standard spectrum, that is: ΔA[k] = |S real [k]| - |S normal [k]|, where, |S real [k]| represents the amplitude of the k-th frequency component of the real-time spectrum, and |S normal [k]| represents the amplitude of the k-th frequency component of the reference spectrum, and ΔA[k] represents the difference amplitude; (2). Phase Difference Calculation. The calculation formula for the phase difference is: Δφ[k] = φ real [k] - φ normal [k], Among them, φ real [k] represents the phase of the k-th frequency component of the real-time spectrum, and φ normal [k] represents the phase of the k-th frequency component of the reference spectrum, and Δφ[k] represents the phase difference; (3) Define thresholds: Set the difference thresholds ΔA threshold and Δφ threshold to determine whether the data is abnormal; (4), Abnormal frequency detection: If |ΔA[k]| > ΔA threshold or |Δφ[k]| > Δφ threshold , it is considered that the k-th frequency component is abnormal; The identification of faults in step S6 includes the following steps: (1). Spectrum Feature Analysis: According to the position and amplitude of the abnormal frequency components, judge the fault mode, that is: New high-frequency components: Component loosening or resonance; Enhanced low-frequency components: Rotor imbalance or external disturbance; Change in harmonic components: Motor fault or vibration coupling; (2). Match with the Fault Mode: Compare the spectrum data with the existing fault library to judge the specific fault type.

5. The anti-interference method for suppressing multiple types of vibrations during flight by an acoustic module carried by a drone according to claim 1, characterized in that, The determination of the system state and error in step S8 includes the following steps: (1) System status: The pitch angle system of the UAV is described by the following state variables: the actual pitch angle θ(t), the angular velocity of the pitch angle the desired pitch angle θd(t); (2) Error definition: Define the system error as: e(t) = θ(t) - θ d (t), Among them, e(t) represents the error between the current state and the desired state of the system, θ(t) represents the actual pitch angle, and θ d (t) represents the desired pitch angle; Then the error dynamics is: Among them, represents the change rate of the state error, represents the angular velocity of the pitch angle, represents the angular velocity of the desired pitch angle; The design method of the virtual control law in step S8 is as follows: (1) Set the virtual control variable: Let the virtual control variable v be used to guide the system state to approach the desired state, and it is defined as: where, v represents the virtual control input, and k1 represents a positive gain coefficient, represents the angular velocity of the desired pitch angle; (2) Design the virtual control objective: Design the virtual control objective so that the error e(t) points to zero, that is: After substituting the expression of v, we have: Among them, represents the angular acceleration of the pitch angle, k1 represents a positive gain coefficient, represents the angular velocity of the desired pitch angle, θ(t) represents the actual pitch angle, θ d (t) represents the desired pitch angle; (3) Control law design: By designing the actual control input u, the virtual control law v is realized. The system dynamics is represented by a second-order model: where J represents the moment of inertia, b represents the damping coefficient, u represents the actual control input, represents the angular acceleration of the pitch angle, represents the angular velocity of the desired pitch angle; Substitute v into the control objective, we have: The final control input u is: where J represents the moment of inertia, b represents the damping coefficient, k1 represents a positive gain coefficient, θ(t) represents the actual pitch angle, and θd ( t) represents the desired pitch angle, represents the angular velocity of the pitch angle, represents the angular velocity of the desired pitch angle.

6. The anti-interference method for suppressing multiple types of vibrations during flight of an unmanned aerial vehicle carrying an acoustic module according to claim 1, wherein The design method of the error function and the method for judging the system stability in step S9 are as follows: (1) Error function definition: Define the error function V(e) to represent the "energy" distance of the system error: e = θ(t) - θ d (t), Among them, V(e) represents the Lyapunov function, e represents the state error, that is, the error between the current state of the system and the desired state, θ(t) represents the actual pitch angle, and θ d (t) represents the desired pitch angle; The positive definiteness in step S9 means that V(e)≥0 holds for all e, and V(e)=0 if and only if e = 0, that is, the energy state of the system is related to the error, and the larger the error, the higher the energy state; (2) Calculate the derivative of the error function. The calculation formula for the derivative of the error function is: Among them, represents the derivative of the Lyapunov function with respect to time, and e represents the state error, that is, the error between the current state of the system and the desired state. represents the rate of change of the error, that is, the derivative of the error with respect to time. k1 represents a positive gain coefficient. The larger k1 is, the faster the error converges, and the larger the error is. When, the energy state of the system is decreasing and the error is converging. When, the system reaches the equilibrium point and the error is zero. (3) System stability analysis: Select the Lyapunov function as the error function and verify the sign of the derivative of the error function: When and e≠0, it indicates that the system is asymptotically stable.

7. A method for suppressing multi-type vibrations and anti-interference during flight of an unmanned aerial vehicle carrying an acoustic module according to claim 1, characterized in that The network parameters in step S10 include: Network center c i : used to define the position of each radial basis function; Network width σi: Used to control the action range of the radial basis function; Weight w i : The weight used to adjust the output; The network structure of the RBF network in step S10 is: Among them, f(x) represents the output of the network, x represents the input data, that is, the current system state, N represents the number of radial basis functions, wi represents the weight of the i-th radial basis function, φi(x) represents the i-th radial basis function, and b represents the bias term; The expression of the network output in step S12 is: Among them, y represents the network output, that is, the estimated value of external disturbance and model uncertainty; The method for adjusting the control input of the UAV in step S13 is: According to the output y of the RBF network, correct the control input of the UAV, that is: u = u0 + Δu, Δu = -ky, where, u represents the adjusted control input, u0 represents the initial control input, Δu represents the compensation control input, k represents the control gain, and y represents the network output; The method for updating the parameters of the RBF network in the step S14 is as follows: update the center c of the network according to the error i and the width σ i , and update the weight w using the gradient descent method i , and the expression of the gradient descent method is: Among them, η represents the learning rate, J represents the error cost function, y represents the network output, and y target represents the expected network output value, and w i (t) represents the weight of the i-th radial basis function at time t, and w i (t + 1) represents the updated weight of the i-th radial basis function at time t + 1, represents the partial derivative of the error cost function J with respect to the weight w i ; The expression of the tracking error in step S15 is: where d represents the dimension of the state variables, i.e., the number of parameters included in the overall state, and x d [i] represents the value of the i-th state variable in the desired state of the system, and x[i] represents the value of the i-th state variable in the actual state of the system; The preset criterion in step S16 is: The tracking error is lower than the threshold, that is: ||tracking error|| < ∈, where, ∈ represents the threshold, that is, the maximum allowable error that the system state can deviate from the desired state.

8. A method for suppressing multi-type vibrations and anti-interference during flight of an unmanned aerial vehicle carrying an acoustic module according to claim 1, wherein The feature extraction in step S17 includes the following steps: (1) Normalize the time-domain signal x(t): where x norm (t) represents the signal value after normalizing the time-domain signal x(t), μ represents the mean value of the time-domain signal x(t), N represents the number of signal sampling points, and σ represents the standard deviation of the time-domain signal x(t); (2) Spectrum analysis: Use the fast Fourier transform to extract the frequency-domain feature X[k]; (3) Construct the feature matrix: Combine the time-domain and frequency-domain features to form the feature matrix F: F = [x norm , X[k], f1, f2,...., f m , Among them, f1, f2, ..., f m represent statistical features, including peak value, mean value and variance; The model in step S18 is a DNN deep neural network. The structure of the DNN deep neural network includes: Input layer: The feature matrix F, whose dimension is [N, d], where, N represents the number of samples, and d represents the feature dimension; Hidden layer: Nonlinearity is introduced using an activation function, that is: h i = ReLU(W i ·h i-1 + b i ) where h i represents the output of the i-th layer, W i represents the weight matrix, and b i represents the bias vector; Output layer: Reconstructed signal The same length as the input signal x(t); The model parameter adjustment in step S18 includes: adjusting hyperparameters, where the hyperparameters include: the number of network layers, the number of neurons, the learning rate, and the batch size; updating network parameters using an optimizer, that is: Among them, W i (t) represents the parameter value at the current moment t, that is, the weight value, W i (t + 1) represents the parameter value at the next moment t + 1, that is, the updated weight value, η represents the learning rate, and L represents the loss function; The expressions for the mean squared error and the absolute error in step S19 are: Among them, L MSE represents the mean square error, that is, the predicted signal and the actual signal x i The mean value of the square of the error between them, L MAE represents the mean absolute error, that is, the predicted signal The mean value of the absolute value of the error between and the actual signal xi, N represents the number of samples, x i represents the actual value of the original signal at the i-th sampling point, represents the predicted value of the model output signal at the i-th sampling point; The expression for the reconstructed signal in step S19 is: Among them, represents the reconstructed signal, \(x(t)\) represents the original signal, and \(n(t)\) represents the noise signal; The expression for the signal-to-noise ratio improvement in step S21 is: ΔSNR = SNR after -SNR before , where ΔSNR represents the increased amplitude of the signal-to-noise ratio of the signal after denoising processing, and SNR after represents the signal-to-noise ratio of the signal after denoising processing, and SNR before represents the quality of the signal before denoising processing; The expression for the signal-to-noise ratio is: where SNR represents the signal-to-noise ratio, and x i represents the actual value of the original signal at the i-th sampling point, and represents the predicted value of the model output signal at the i-th sampling point; The criterion for the prediction accuracy in step S21 is: the error is less than a preset threshold ∈.

9. A method for suppressing multi-type vibrations and anti-interference during the flight of an unmanned aerial vehicle equipped with an acoustic module, according to claim 1, characterized in that The most useful features in step S17 refer to: frequency-domain features and time-domain features obtained through spectral analysis, and the useful information in step S20 refers to: the effective signal after denoising.

Images