A method and apparatus for drone noise cancellation for cross-medium direct acoustic communication

Abstract

The application discloses a UAV noise cancellation method and device for cross-medium direct acoustic communication, and the method comprises the following steps: sampling a known probe signal and a received signal, obtaining a frequency domain expression, and calculating a cross-medium sound pressure transmission function to extract an amplitude response, a phase response and a group delay; based on a normalized ratio and a local maximum value detection method, a center frequency set of a UAV propeller noise is identified; a notch depth, a bandwidth, a frequency compensation and a phase compensation are determined in combination with channel characteristic parameters to construct a cascaded notch filter; channel characteristic parameters are input into a deep learning model to output a filter parameter correction amount, dynamic updating of the filter is realized, a filter capable of suppressing noise is obtained, and a communication signal after noise reduction is obtained. The application can realize accurate detection of propeller noise frequency and multi-dimensional adaptive adjustment of filter parameters.

Description

Technical Field

[0001] This invention relates to the field of cross-medium acoustic communication and intelligent signal processing, specifically to a method and apparatus for noise cancellation of unmanned aerial vehicles (UAVs) for cross-medium direct acoustic communication. Background Technology

[0002] In recent years, with the increasingly widespread application of unmanned aerial vehicles (UAVs) in scenarios such as marine monitoring, search and rescue communications, and integrated sea-air operations, direct acoustic communication technology across water and air media has attracted widespread attention. This type of communication typically relies on underwater transducers emitting acoustic signals, which are transmitted through the water-air interface into the air and received by microphones on the UAV. However, due to the presence of propellers, UAVs generate strong broadband mechanical noise and multi-peak modulation interference during flight. This noise is superimposed on the effective signal at the receiving end, significantly reducing the signal-to-noise ratio and reliability of the communication.

[0003] The water-air interface plays a channel mapping role in cross-medium communication. Unlike conventional free-field sound propagation, interface effects lead to amplitude distortion, phase delay, and group delay variations in sound pressure, thus making the statistical characteristics of the received signal more complex. Especially in the low-frequency band, the propeller noise spectrum partially overlaps with the effective communication signal, making it difficult for traditional filtering methods to effectively distinguish between the two.

[0004] Existing noise suppression methods mainly include:

[0005] (1) Fixed notch filter: A notch filter is set at the known propeller fundamental frequency and its harmonics to attenuate noise. However, this method cannot adapt to the noise frequency shift caused by changes in flight attitude, wind speed and load, and is prone to insufficient filtering or excessive suppression of useful signals.

[0006] (2) Adaptive noise cancellation: A noise model is constructed by introducing a reference microphone and then canceled. However, in cross-medium communication scenarios, the reference channel is difficult to obtain, and the nonlinear distortion caused by the water-air interface limits the performance of adaptive noise cancellation.

[0007] (3) Traditional deep learning noise reduction: Directly train end-to-end models to remove noise, but because the cross-medium sound pressure transmission mechanism is not considered, the model is difficult to adapt to dynamic environments and has insufficient generalization ability.

[0008] Therefore, there is an urgent need for a method that can combine cross-medium acoustic pressure transmission modeling with deep learning correction mechanisms to achieve multi-parameter adaptive elimination of UAV propeller noise, so as to improve the reliability of cross-medium communication. Summary of the Invention

[0009] To address the problems of inaccurate noise detection, fixed filter parameters, and lack of physical constraints and online adaptive capabilities in deep learning models in existing technologies, this invention provides a method and apparatus for noise cancellation of unmanned aerial vehicles (UAVs) for cross-medium direct acoustic communication, which can achieve accurate modeling, dynamic adjustment, and closed-loop optimization.

[0010] The objective of this invention is achieved through the following technical solution:

[0011] A method for noise cancellation of unmanned aerial vehicles (UAVs) for cross-medium direct acoustic communication, the method comprising the following steps:

[0012] S1: Sample the known probe and received signals to obtain the frequency domain expression, calculate the transmedium acoustic pressure transmission function, extract the amplitude response, phase response and group delay to form a set of channel characteristic parameters;

[0013] S2: For each received communication signal, based on the normalization ratio calculation and local maximum detection method, determine the center frequency of the UAV noise on its discrete frequency point set, and obtain the noise center frequency set.

[0014] S3: At the noise center frequency, combined with the channel characteristic parameters, determine the initial values ​​of notch depth, bandwidth, frequency compensation and phase compensation respectively, form the parameter set of the notch filter, and obtain the overall filter with multiple notch units cascaded.

[0015] S4: Input the channel feature parameters obtained in S1 into the deep learning model, output the correction amount of the filter parameters, realize the dynamic update of the filter parameters, obtain a filter that can suppress noise, and then obtain the denoised communication signal.

[0016] Furthermore, a linear frequency modulated signal is used as the known detection signal.

[0017] Furthermore, the frequency domain expression includes the spectrum of the probe signal and the spectrum of the received signal, and the transmedium acoustic pressure transmission function is the ratio of the spectrum of the received signal to the spectrum of the probe signal.

[0018] Furthermore, when using the local maximum detection method, the k-th frequency point is identified as a noise candidate frequency point if it simultaneously meets the following three conditions:

[0019] The normalization ratio of the kth frequency point is greater than the normalization ratio of the (k-1)th frequency point;

[0020] The normalization ratio of the kth frequency point is greater than the normalization ratio of the (k+1)th frequency point;

[0021] The normalization ratio at the k-th frequency point is greater than the noise threshold.

[0022] Furthermore, the obtained noise candidate frequency points are further subjected to time persistence determination, that is, the candidate peak corresponding to the noise candidate frequency point appears at least multiple times in consecutive frames; then the candidate peaks with an interval less than a set threshold are merged, and the fundamental frequency point is retained as the noise center frequency when there is a harmonic relationship, and finally the propeller noise center frequency set is obtained.

[0023] Furthermore, in step S3, when determining the initial values ​​of the filter parameters, the initial parameter values ​​need to be pruned and exponentially smoothed.

[0024] Furthermore, the loss function for training the deep learning model in S4 is the sum of the frequency domain mean square error and the parameter regularization term.

[0025] A noise cancellation device for unmanned aerial vehicles (UAVs) used for cross-medium direct acoustic communication includes:

[0026] The channel characteristic extraction module is used to calculate the cross-medium acoustic pressure transmission function based on the known detection signal and the corresponding received signal, and extract the amplitude response, phase response and group delay to form a set of channel characteristic parameters;

[0027] The noise frequency detection module is used to detect the noise frequency components generated by the UAV in the communication signal. It identifies the center frequency and harmonic components of the noise by normalization ratio calculation and local maximum detection method.

[0028] The parameter mapping module is used to determine the initial values ​​of notch depth, bandwidth, frequency compensation and phase compensation at the noise center frequency, combined with channel characteristic parameters, to form the parameter set of the notch filter.

[0029] The adaptive filter module is used to construct an overall noise suppression filter based on the initial parameters provided by the parameter mapping module, perform real-time filtering processing on the communication signal, and adaptively adjust the parameters of the overall filter based on the feedback from the deep learning inference and training module to output the denoised communication signal.

[0030] The deep learning inference and training module is used to input channel feature parameters into the pre-trained deep learning model and output filter parameter corrections.

[0031] An electronic device, comprising:

[0032] One or more processors;

[0033] A storage device for storing one or more programs that, when executed by the electronic device, enable the electronic device to implement a drone noise cancellation method for cross-media direct acoustic communication.

[0034] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements a method for noise cancellation of unmanned aerial vehicles (UAVs) for direct acoustic communication across media.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) Combining physical modeling with deep learning: Introducing the transmedium sound pressure transmission function to model the water-air interface effect, so that the noise reduction process not only depends on statistical learning, but also has physical constraints and interpretability.

[0037] (2) Multi-parameter collaborative adaptive: The notch depth, bandwidth, frequency compensation and phase compensation parameters can be dynamically adjusted according to the channel characteristics, improving the filter's adaptability to different noise patterns.

[0038] (3) Precise positioning of propeller noise frequency: Based on the dual criteria method of normalization ratio and maximum detection, it can accurately identify the noise center frequency in complex backgrounds and avoid the frequency deviation and missed detection problems of traditional methods.

[0039] (4) Closed-loop online optimization: A deep learning model is used for real-time correction, combined with the regularization constraint of the loss function, to achieve adaptive updates across time domains and ensure robustness in dynamic environments.

[0040] (5) This invention can not only effectively eliminate the interference of UAV propeller noise on cross-medium communication, but also has strong adaptability, good real-time performance and high reliability, and can significantly improve the signal quality and system performance of water-air cross-medium communication. Attached Figure Description

[0041] Figure 1 This is a flowchart of a method for noise cancellation of unmanned aerial vehicles (UAVs) for cross-medium direct acoustic communication according to an embodiment of the present invention.

[0042] Figure 2 This is a schematic diagram of a drone noise cancellation device for cross-medium direct acoustic communication according to another embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram of an electronic device according to another embodiment of the present invention. Detailed Implementation

[0044] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0045] On the one hand, such as Figure 1 As shown, one embodiment of the present invention provides a method for noise cancellation of unmanned aerial vehicles (UAVs) for direct acoustic communication across media, comprising the following steps:

[0046] Step 1: Transmission of cross-medium sounding signal and estimation of transmission function: Before communication, a known sounding signal is transmitted, and the receiver collects the signal after it passes through the water-air interface; the known sounding signal and the received signal are sampled, and the frequency domain expression is obtained by fast Fourier transform. The cross-medium sound pressure transmission function is calculated, and the amplitude response, phase response and group delay are extracted to form a set of channel characteristic parameters.

[0047] In this embodiment, during the detection time slot before communication, the transmitting end transmits a known detection signal. The receiving end simultaneously acquires the corresponding received signal. This embodiment uses a linear frequency modulated (LFM) signal as the probe signal. This signal has good broadband characteristics, which is beneficial for stable estimation of the transmedium transmission function. The expression for the probe signal is:

[0048]

[0049] in, For signal duration, For bandwidth, This is the starting frequency.

[0050] This embodiment uses the sampling frequency for sampling known detection and received signals. 48 kHz, Fast Fourier Transform Points Take equal to the window length And using the Hann window function, frame shift .

[0051] The frequency domain representation obtained through Fast Fourier Transform includes the spectrum of the probe signal. and received signal spectrum In the discrete set of frequency points Calculate the transmedium transmission function at the k-th frequency point. :

[0052]

[0053] To avoid the division by zero problem, a regularization term defined specifically for this purpose can be added to the denominator. ,Right now:

[0054]

[0055]

[0056] Where the coefficient Pick to .

[0057] The transmedium transmission function characterizes the amplitude and phase changes of sound waves as they propagate through the water-air interface.

[0058] Based on the transmedium acoustic pressure transmission function, the amplitude response, phase response, and group delay at each frequency point are calculated; where the amplitude response at the k-th frequency point is... Phase response The group delay at the k-th frequency point is calculated using center difference. ,Right now:

[0059]

[0060] In this embodiment,

[0061]

[0062]

[0063] in, It represents the frequency difference between two adjacent frequency points.

[0064] The phase response needs to be considered during the calculation process. Expand to avoid Jump, and the group delay result Gaussian or median filtering is applied to reduce noise from the differential amplification, thereby mitigating its impact. These channel characteristic parameters will be used in subsequent steps for filter parameter initialization and deep learning correction.

[0065] Step 2: Propeller noise frequency detection: Sample each received communication signal, and determine the center frequency of the propeller noise based on the normalization ratio calculation and local maximum detection method on the discrete frequency point set of the communication signal, thus obtaining the noise center frequency set.

[0066] Considering that propeller noise typically exhibits comb-like narrow-band spectral peaks, with its energy concentrated at the fundamental frequency and its harmonics and significantly higher than the smooth spectral components of the background and communication signals, this embodiment employs a combined method of normalized spectrum and local peak detection to improve detection reliability. Specifically, in the transmitted communication signal... In each frame of the signal, the normalization ratio is first constructed. :

[0067]

[0068] in, This represents a constant defined to avoid division by zero. This indicates the spectrum of the received communication signal.

[0069] This ratio can reduce the impact of channel attenuation on the spectrum, making the narrowband peak of propeller noise more prominent.

[0070] Then, local maximum detection is used. When the k-th frequency point meets the following condition, it is determined as a noise candidate frequency point:

[0071]

[0072] Wherein, γ is the noise threshold, which is set automatically based on background noise statistics and can measure the mean background noise in an empty field or at a recent non-noise time. and variance ,set up , Take 2.5 to 4.

[0073] All frequency points that meet the conditions constitute the set of propeller noise center frequencies:

[0074]

[0075] To avoid false detections due to transient noise, this embodiment also employs a time persistence determination, meaning that candidate peaks must be within a continuous time frame. At least appearing within the frame Second-rate, and The appropriate frequency can be selected based on the actual situation. Candidate peaks that are too close together are merged, and the fundamental frequency is retained as the noise center frequency when harmonic relationships exist, ultimately yielding the set of propeller noise center frequencies. .

[0076] Step 3: Filter parameter mapping and notch unit construction: At the noise center frequency, combined with channel characteristic parameters, determine the initial values ​​of notch depth, bandwidth, frequency compensation and phase compensation respectively to form the parameter set of the notch filter; construct notch units based on these parameters, and cascade multiple notch units to obtain the overall filter, thereby achieving the suppression of multi-band noise.

[0077] After obtaining the channel characteristic parameters and the propeller noise center frequency set Then, it is necessary to analyze each noise center frequency. The parameter set for initializing the notch filter includes parameters for determining the notch depth. ,bandwidth Frequency compensation and phase compensation The initial values ​​are then used. This process maps physical channel quantities to filter parameters through a mapping function, ensuring that the initial filter is physically interpretable and providing a good starting point for subsequent deep learning corrections.

[0078] in:

[0079]

[0080]

[0081]

[0082]

[0083] in, These are all mapping functions. In this embodiment, the expressions for these mapping functions are:

[0084]

[0085]

[0086]

[0087] Where, α min α max These are the preset minimum and maximum notch depths, respectively. a1 and b1 are constants. b1 is acceptable. Or empirical value; σ min This represents the minimum preset bandwidth; k2 is a preset constant, ranging from 0.05 to 1; k3 is a proportional coefficient, ranging from 0.01 to 0.5.

[0088] To ensure stability, the initial parameter values ​​need to be trimmed and exponentially smoothed to prevent large fluctuations caused by a single abnormal measurement.

[0089] After obtaining the parameters at each noise center frequency, the corresponding frequency domain notch filter units are constructed and cascaded to form the overall filter transfer function.

[0090] Single frequency domain notch unit Defined as:

[0091]

[0092] Overall filter The result is obtained by multiplying M frequency domain notch elements:

[0093] .

[0094] In real-time applications, the preferred method is frame-by-frame multiplication in the frequency domain, that is, directly multiplying the received spectrum in the short-time Fourier transform domain. Multiply by filter The time-domain signal is then recovered through inverse transformation and overlapping, with a delay of approximately one frame. Alternatively, it can be achieved by calculating the time-domain impulse response of the filter and performing convolution. During implementation, it is necessary to prevent... Approaching zero at certain frequencies leads to numerical distortion, therefore it is generally limited. A lower bound is set for the overall transfer function. Additionally, overlapping notch elements should be normalized to avoid excessive suppression of the communication signal bandwidth.

[0095] Step 4: Deep Learning Model Correction: The channel feature parameters obtained in Step 1 are used as input to the deep learning model, which outputs the correction amount for the filter parameters, realizing dynamic updating of the filter parameters and thus enhancing the adaptability to propeller noise in complex environments. In the training sequence, the loss function is calculated with the transmitted communication signal as a reference, and the deep learning model parameters are updated through the backpropagation algorithm, forming a closed loop of "channel characteristics - filter parameters - model update", resulting in a filter that can effectively suppress UAV propeller noise, and thus obtaining the denoised communication signal.

[0096] To ensure the long-term effectiveness of the aforementioned physically prior filters in complex, non-steady-state cross-medium channels, this invention proposes coupling channel characteristics with a deep learning model to online correct filter parameters and fine-tune the projection constraint range. Specifically, a class of parameterizable functions is selected. Its input is the constructed channel feature vector:

[0097]

[0098] Where K can be the number of sampling points in the entire frequency band or selectively extracting points from the subband of interest to reduce the input dimension.

[0099] The network output is a vector of correction values ​​for each filter parameter. It contains The model can adopt a one-dimensional convolutional neural network structure, such as two convolutional layers to extract frequency domain features followed by a multilayer perceptron, with the hyperparameter scale controlled between 200k and 2M to adapt to edge inference.

[0100] The output correction is constrained within a reasonable range by the mapping function, and then superimposed with the initial parameters before being updated exponentially smoothed. For example:

[0101]

[0102] Among them, smoothing factor This allows for real-time fine-tuning of parameters while ensuring stability.

[0103] Training strategies for deep learning models include offline pre-training and online fine-tuning. The offline phase generates a large number of diverse samples on a simulation platform, including various... (Simulating transmission characteristics under different interfaces, angles, weather conditions, and distances), different UAV rotation speeds and harmonic structures, different SNRs, etc., trained using supervised learning methods. During the online phase, when a reference pilot exists in the deployed system, the actual pilot or training sequence is used as the reference. Supervised fine-tuning is performed using a known reference; the loss function with reference can take the form of frequency domain mean square error plus parameter regularization term:

[0104]

[0105] in, For filter output, It is a regularization coefficient used to constrain the notch depth and bandwidth, avoiding excessive notch filtering or overly wide bandwidth suppression of communication frequency bands.

[0106] In the absence of a reference, self-supervised training can be performed using a blind objective (e.g., minimizing output energy in a predefined noise band and maintaining energy in the signal band) or a semi-supervised strategy. Blind loss under no-reference conditions can employ a hybrid objective as follows:

[0107]

[0108] in, The reference spectrum is obtained based on speech activity detection or prior spectrum estimation. and These are the index sets for the main frequency band of noise and the main frequency band of communication signals, respectively.

[0109] On the other hand, such as Figure 2 As shown, the UAV noise cancellation device for cross-medium direct acoustic communication of the present invention includes a signal transmission unit, a receiving and acquisition unit, a channel characteristic extraction module, a noise frequency detection module, a parameter mapping module, an adaptive filter module, and a deep learning inference and training module.

[0110] The signal transmitting unit is used to generate and transmit a known detection signal before cross-medium communication, to establish an acoustic propagation model and obtain cross-medium channel characteristics; and to transmit communication signals during cross-medium communication. The known detection signal can be a linear frequency modulated signal, a pseudo-random sequence, or other reference signal with broadband characteristics, and its amplitude, bandwidth, and duration are controlled by a signal generator.

[0111] The receiving and acquisition unit is used to acquire acoustic signals received by the UAV during cross-media communication. The receiving path includes a preamplifier circuit, an anti-aliasing filter, and an analog-to-digital converter module, used for signal amplification, filtering, analog-to-digital conversion, and digitization. After framing and window function processing, the sampled signal is sent to the channel characteristic extraction module for subsequent frequency domain analysis and channel parameter calculation.

[0112] The channel characteristic extraction module is used to calculate the transmedium acoustic pressure transmission function based on the known probe signal and the corresponding received signal, and extract multi-dimensional parameters reflecting the transmedium acoustic propagation characteristics. Specifically, this includes: calculating the frequency domain transmission function to obtain the amplitude response and phase response, and expanding the phase to obtain the group delay. The obtained amplitude response, phase response, and group delay form the channel feature vector, providing input for the subsequent noise detection and parameter mapping modules.

[0113] The noise frequency detection module is used to detect noise frequency components generated by drone rotors, motors, and other parts of the communication signal. It identifies the center frequency and harmonic components of the noise through normalized ratio calculation and local maximum detection. The method involves calculating the energy ratio in the communication signal spectrum and detecting local maxima. Stable noise frequencies and their harmonic components are then identified based on time duration and threshold values. The detection results include the drone noise center frequency, amplitude, and confidence level information, which serve as the basis for subsequent filter parameter calculations.

[0114] The parameter mapping module is used to determine the initial values ​​of notch depth, bandwidth, frequency compensation, and phase compensation at the noise center frequency, combined with channel characteristic parameters, thus forming the parameter set of the notch filter. By analyzing the noise frequency distribution and channel transmission characteristics, the center frequency, bandwidth, and depth of each notch filter are determined, and phase compensation information is calculated based on the channel phase response. The output filtering parameters are used for the initialization of the adaptive filter module.

[0115] The adaptive filter module is used to construct an overall noise suppression filter based on the initial parameters provided by the parameter mapping module, and to perform real-time filtering processing on the communication signal. During operation, this module adaptively adjusts the filter parameters according to the output of the deep learning inference and training module, achieving adaptive optimization of notch depth, bandwidth, and center frequency, thereby maintaining a stable noise suppression effect when the channel changes, and finally outputting the denoised communication signal.

[0116] The deep learning inference and training module takes the channel feature parameters output by the channel feature extraction module as input, calculates the filter parameter correction amount through the deep learning model, and dynamically updates the filter parameters. The model can employ a one-dimensional convolutional neural network or a multilayer sensing structure to extract the relationship between channel frequency domain features and UAV noise distribution. The model operates through an "offline pre-training + online fine-tuning" approach: in the offline phase, it is trained on diverse simulation samples to obtain initial mapping capabilities; in the online phase, during system operation, the loss function is calculated based on the transmitted communication signal as a reference or a self-supervised target, and the model parameters are updated in real time through the backpropagation algorithm. The updated correction amount is superimposed with the initial filter parameters and smoothed to achieve closed-loop adjustment of "channel characteristics—filter parameters—model update," thereby enhancing the system's adaptability to complex cross-medium sound fields and unsteady UAV noise.

[0117] The embodiments of the UAV noise cancellation device for cross-medium direct acoustic communication of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 3 The diagram shown is a hardware structure diagram of any data processing-capable device in which the UAV noise cancellation device for cross-medium direct acoustic communication of the present invention is located. In addition to the processor, memory, network interface and non-volatile memory shown in the diagram, any data processing-capable device in the embodiment may also include other hardware depending on the actual function of the data processing-capable device, which will not be described in detail here.

[0118] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the UAV noise cancellation method for cross-medium direct acoustic communication described in the above embodiments.

[0119] The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0120] In summary, this invention achieves accurate identification of the propeller noise center frequency by estimating and extracting the cross-medium acoustic pressure transmission function, combined with a local maximum detection method. It also utilizes a physical mapping function to obtain initial values ​​for filter parameters, and then dynamically updates these parameters through online correction using a deep learning model and loss function optimization. This effectively suppresses UAV propeller noise during communication. This method not only ensures accurate noise detection, avoiding the false positives and false negatives common in traditional methods, but also maintains high signal fidelity and robustness in complex, time-varying cross-medium channel environments.

[0121] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for noise cancellation of unmanned aerial vehicles (UAVs) for cross-medium direct acoustic communication, characterized in that, The method includes the following steps: S1: Sample the known probe signal and received signal to obtain the frequency domain expression, calculate the transmedium acoustic pressure transmission function, extract the amplitude response, phase response and group delay to form a set of channel characteristic parameters; the frequency domain expression includes the probe signal spectrum and the received signal spectrum, and the transmedium acoustic pressure transmission function is the ratio of the received signal spectrum to the probe signal spectrum; S2: For each received communication signal, based on the normalization ratio calculation and local maximum detection method, determine the center frequency of the UAV noise on its discrete frequency point set, and obtain the noise center frequency set. The normalization ratio The calculation formula is as follows: ； in, This represents a constant defined to avoid division by zero. This represents the spectrum of the received communication signal. Represents the transmedium transmission function at the k-th frequency point: S3: At the noise center frequency, combined with the channel characteristic parameters, determine the initial values ​​of notch depth, bandwidth, frequency compensation and phase compensation respectively, form the parameter set of the notch filter, and obtain the overall filter with multiple notch units cascaded. S4: Input the channel feature parameters obtained in S1 into the deep learning model, output the correction amount of the filter parameters, realize the dynamic update of the filter parameters, obtain a filter that can suppress noise, and then obtain the denoised communication signal.

2. The UAV noise cancellation method for cross-medium direct acoustic communication according to claim 1, characterized in that, A linear frequency modulated signal is used as the known detection signal.

3. The UAV noise cancellation method for cross-medium direct acoustic communication according to claim 1, characterized in that, When using the local maximum detection method, a frequency point is identified as a noise candidate frequency point if it simultaneously meets the following three conditions: The normalization ratio of the kth frequency point is greater than the normalization ratio of the (k-1)th frequency point; The normalization ratio of the kth frequency point is greater than the normalization ratio of the (k+1)th frequency point; The normalization ratio at the k-th frequency point is greater than the noise threshold.

4. The UAV noise cancellation method for cross-medium direct acoustic communication according to claim 3, characterized in that, For the obtained noise candidate frequency points, the time persistence determination is further performed, that is, the candidate peak corresponding to the noise candidate frequency point appears at least multiple times in consecutive frames; then the candidate peaks with an interval less than a set threshold are merged, and the fundamental frequency point is retained as the noise center frequency when there is a harmonic relationship, and finally the propeller noise center frequency set is obtained.

5. The UAV noise cancellation method for cross-medium direct acoustic communication according to claim 3, characterized in that, In step S3, when determining the initial values ​​of the filter parameters, the initial parameter values ​​need to be pruned and exponentially smoothed.

6. The UAV noise cancellation method for cross-medium direct acoustic communication according to claim 3, characterized in that, The loss function for training the deep learning model in S4 is the sum of the frequency domain mean square error and the parameter regularization term.

7. A noise cancellation device for unmanned aerial vehicles (UAVs) used for direct acoustic communication across media, characterized in that, include: The channel characteristic extraction module is used to sample known probe and received signals to obtain frequency domain representation; The transmedium acoustic pressure transmission function is calculated based on the known probe signal and the corresponding received signal, and the amplitude response, phase response and group delay are extracted to form a set of channel characteristic parameters; the frequency domain expression includes the probe signal spectrum and the received signal spectrum, and the transmedium acoustic pressure transmission function is the ratio of the received signal spectrum to the probe signal spectrum; The noise frequency detection module is used to detect noise frequency components generated by the UAV in the communication signal. It identifies the center frequency and harmonic components of the noise through normalization ratio calculation and local maximum detection methods. The formula for calculating the normalization ratio is as follows: ； in, This represents a constant defined to avoid division by zero. This represents the spectrum of the received communication signal. Represents the transmedium transmission function at the k-th frequency point: The parameter mapping module is used to determine the initial values ​​of notch depth, bandwidth, frequency compensation and phase compensation at the noise center frequency, combined with channel characteristic parameters, to form the parameter set of the notch filter. The adaptive filter module is used to construct an overall noise suppression filter based on the initial parameters provided by the parameter mapping module, perform real-time filtering processing on the communication signal, and adaptively adjust the parameters of the overall filter based on the feedback from the deep learning inference and training module to output the denoised communication signal. The deep learning inference and training module is used to input channel feature parameters into the pre-trained deep learning model and output filter parameter corrections.

8. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by the electronic device, cause the electronic device to implement the UAV noise cancellation method for cross-medium direct acoustic communication as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the UAV noise cancellation method for cross-medium direct acoustic communication as described in any one of claims 1 to 6.

Images