A UAV classification method and system based on noise perception model

Abstract

The present invention discloses a method and system for drone classification based on a noise perception model, comprising the following steps: acquiring drone audio data and processing it to obtain audio data energy; processing the audio data energy to obtain an enhanced signal; optimizing the enhanced signal to obtain high and low frequency bands of the enhanced signal; and generating a mixed-resolution spectrum based on the high and low frequency bands; processing the high and low frequency bands using a differential Mel filter bank to obtain band boundaries; generating a Mel-spectrogram based on the band boundaries and the mixed-resolution spectrum; inputting the Mel-spectrogram into a convolutional neural network to output the probability of the drone category. The present invention uses a noise perception model to perform noise quantization, noise reduction, and feature enhancement on drone audio data, and combines this with a convolutional neural network to achieve high-precision drone classification.

Description

Technical Field

[0001] The present invention relates to the fields of sound processing and unmanned aerial vehicle (UAV) technology, and in particular to a UAV classification method and system based on a noise perception model. Background Art

[0002] With the rapid development of drone technology, drones are increasingly being used in aerial photography, logistics, agricultural monitoring, disaster relief, and other fields. However, existing drone identification technologies mostly rely on visual or radar signals. These methods have significant limitations in complex environments (such as at night, in fog, haze, rain, and snow) or when vision is obstructed, making it difficult to achieve high-precision drone classification.

[0003] Traditional drone classification methods rely primarily on image recognition technology, but its accuracy and reliability decrease significantly in low-light conditions or when the target is far away. Furthermore, visual recognition technology requires high computing resources and hardware, making it difficult to widely apply in resource-constrained scenarios.

[0004] Drones generate specific sound signals during flight, which are closely related to the type, structure, and operating status of the drone. Compared with visual signals, sound signals have the advantage of not being restricted by light and line of sight and can operate stably in complex environments.

[0005] This paper proposes a drone classification method based on sound processing, combining a noise perception model with a convolutional neural network. This method not only effectively handles classification tasks in complex noisy environments but also reduces reliance on visual recognition technology, providing a new solution for drone monitoring, air traffic management, and security protection. This method has broad application prospects, including military reconnaissance, border surveillance, urban security, and other fields. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide a method and system for drone classification based on a noise perception model, which solves the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for classifying drones based on a noise perception model, comprising the following steps:

[0008] Step S1: Acquire drone audio data, process the drone audio data to obtain mono drone audio data, resample the mono drone audio data to obtain resampled audio data, calculate the resampled audio data, and obtain audio data energy;

[0009] Step S2: Processing the audio data energy to obtain a time domain quantization index and a frequency domain quantization index, processing the time domain quantization index and the frequency domain quantization index to obtain a denoised time domain signal and a denoised frequency domain signal, fusing the time domain quantization index and the frequency domain quantization index with the denoised time domain signal and the denoised frequency domain signal to obtain an enhanced signal;

[0010] Step S3: Optimizing the enhanced signal using Mel spectrum transform to obtain a low-frequency band of the enhanced signal and a high-frequency band of the enhanced signal; analyzing the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal to generate a mixed-resolution spectrum; optimizing the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal using a differential Mel filter group to obtain a frequency band boundary of the differential Mel filter group under the control of a resolution adjustment coefficient; processing the mixed-resolution spectrum and the frequency band boundary of the differential Mel filter group under the control of the resolution adjustment coefficient to generate a Mel spectrum graph;

[0011] Step S4: Input the mel spectrogram into the convolutional neural network for processing and output the probability of the drone category.

[0012] Furthermore, the specific process of obtaining the enhanced signal is as follows:

[0013] The energy of the input audio data is divided into frames and then calculated by short-time Fourier transform to obtain the energy of the kth frame. and the noise energy estimate of the kth frame , the energy of the kth frame is tracked using the minimum statistical noise tracking algorithm and the noise energy estimate of the kth frame Perform dynamic updates to obtain time domain quantitative indicators ,express:

[0014] ;

[0015] Where, The total number of frames representing the energy of the audio data; represents the mean energy of the kth frame; represents the standard deviation of the energy of the kth frame; represents the hyperbolic tangent function; represents the adaptive adjustment factor; ; SNR means signal-to-noise ratio; log means logarithmic operation;

[0016] The audio data energy is extracted through the Mel filter bank to obtain the signal spectrum components of the vth frame respectively and the noise spectrum component of the vth frame , the noise floor estimation method is used to calculate the extracted signal spectrum components of the vth frame and the noise spectrum component of the vth frame , obtain frequency domain quantitative indicators ,express:

[0017] ;

[0018] Where, Indicates the total number of channels in the filter bank; Represents the slope parameter of the Sigmoid function; represents the energy threshold; e represents the natural constant;

[0019] Normalize the signal-to-noise ratio value range to [0,1] to obtain the normalized signal-to-noise ratio ; ; Indicates the maximum value of the signal-to-noise ratio; Indicates the minimum value of the signal-to-noise ratio;

[0020] By normalizing the signal-to-noise ratio Calculate the parameters p and q, ; ; p represents the signal amplitude spectrum power adjustment parameter; q represents the dynamic noise suppression coefficient;

[0021] Calculate the frequency based on p and q The spectral attention mask value at , represents:

[0022] ;

[0023] Where, Indicates at frequency point The spectral attention mask value at ; Indicates the signal at the frequency point The amplitude spectrum at ; Indicates dynamic noise in the frequency band The average amplitude spectrum at ; Represents the adaptive scaling index Through frequency domain quantitative indicators adjust; represents the balance factor; where = ;

[0024] At the frequency point The spectral attention mask value at is used for noise reduction calculation to obtain the denoised time domain signal. and the frequency domain signal after noise reduction ; represents the inverse Fourier transform;

[0025] Fusion of denoised time domain signals by dynamic weights and the frequency domain signal after noise reduction and time domain quantitative indicators and frequency domain quantitative indicators , and get enhanced signal ,express:

[0026] .

[0027] Furthermore, the specific process of generating the mel spectrogram is as follows:

[0028] According to the noise environment, based on time domain quantitative indicators Calculate the resolution adjustment coefficient by summing the noise spectrum entropy , based on the resolution adjustment coefficient The time-frequency resolution of the enhanced signal is improved by using Mel spectrum transform Perform dynamic adjustment and optimization to obtain enhanced signal Low frequency band and enhanced signal High frequency band;

[0029] Resolution adjustment factor ,express:

[0030] ;

[0031] In the formula, sigmoid represents the activation function; Hnoise represents the noise spectrum entropy; w1 represents the time domain quantitative index Weight; w2 represents the noise spectrum entropy weight;

[0032] To enhance the signal Low frequency band and enhanced signal The high frequency band of is analyzed to generate a mixed resolution spectrum, which is represented by:

[0033] ;

[0034] Where, represents the mixed-resolution spectrum; Indicates enhanced signal low-band spectrum; Indicates enhanced signal The high frequency spectrum;

[0035] Differentiable Mel filter bank based on resolution adjustment coefficient , dynamic fusion enhanced signal Low frequency band and enhanced signal The high frequency band of , indicating:

[0036] ;

[0037] Where, Indicates the Differentiable Mel filter bank with resolution adjustment coefficient Band boundaries under control; Indicates the Enhanced signal The low-frequency filter boundary of Indicates the Enhanced signal The high-frequency filter boundary of

[0038] Based on the Differentiable Mel filter bank with resolution adjustment coefficient The frequency band boundaries under control define the A Mel filter at the frequency point The weight coefficient at ;

[0039] Based on mixed resolution spectrum and A Mel filter at the frequency point The weight coefficient at , generates a mel spectrum, which is expressed as:

[0040] ;

[0041] Where, represents a mel-spectrogram; Indicates the number of fast Fourier transform FFT points; Indicates the A Mel filter at the frequency point The weight coefficient at ; Indicates a minimum value.

[0042] Furthermore, the convolutional neural network consists of four convolutional blocks connected in sequence, and each convolutional block consists of a convolutional layer, a ReLU activation layer, and a pooling layer connected in sequence; the Mel-spectrogram is input into the convolutional neural network and passes through the four convolutional blocks in sequence to output the category probability of the drone.

[0043] Furthermore, the specific process of obtaining the audio data energy is as follows:

[0044] Convert the acquired drone audio data into a unified format to obtain multi-channel drone audio data in a unified format;

[0045] Convert multi-channel drone audio data to mono drone audio data, which means:

[0046] ;

[0047] Where, Represents mono-channel drone audio data; Indicates the total number of channels for multi-channel drone audio data; Represents the drone audio data of the cth channel;

[0048] Resample the mono drone audio data to obtain resampled audio data, which means:

[0049] ;

[0050] Where, Represents resampled audio data; Represents a resampling operation; Indicates the original sampling rate of mono drone audio data; represents the target sampling rate;

[0051] Calculate the resampled audio data and obtain the audio data energy, which is expressed as:

[0052] ;

[0053] Where, Represents the energy of audio data; Represents the total number of resampling points; Represents the first sampling points.

[0054] A drone classification system based on a noise perception model, applying the drone classification method based on a noise perception model, comprises:

[0055] An acquisition module is used to acquire drone audio data, process the drone audio data, obtain mono-channel drone audio data, resample the mono-channel drone audio data, obtain resampled audio data, calculate the resampled audio data, and obtain audio data energy;

[0056] a processing module for processing the audio data energy to obtain a time domain quantization index and a frequency domain quantization index, processing the time domain quantization index and the frequency domain quantization index to obtain a noise-reduced time domain signal and a noise-reduced frequency domain signal, and fusing the time domain quantization index and the frequency domain quantization index with the noise-reduced time domain signal and the noise-reduced frequency domain signal to obtain an enhanced signal;

[0057] An optimization module is used to optimize the enhanced signal using Mel spectrum transform to obtain a low-frequency band of the enhanced signal and a high-frequency band of the enhanced signal, analyze the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal to generate a mixed-resolution spectrum; optimize the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal through a differential Mel filter group to obtain a frequency band boundary of the differential Mel filter group under the control of a resolution adjustment coefficient, process the frequency band boundary of the mixed-resolution spectrum and the differential Mel filter group under the control of the resolution adjustment coefficient to generate a Mel spectrum map;

[0058] The output module is used to input the Mel spectrogram into the convolutional neural network for processing and output the probability of the drone category.

[0059] An electronic device includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute the drone classification method based on the noise perception model.

[0060] A non-volatile computer storage medium stores computer-executable instructions, wherein the computer-executable instructions refer to a method for classifying drones based on a noise perception model.

[0061] Compared with the existing technology, the present invention has the following beneficial effects:

[0062] (1) The present invention uses a noise perception model to perform noise quantization, noise reduction, and feature enhancement on drone audio data, and combines it with a convolutional neural network to achieve high-precision drone classification. The noise perception model can effectively extract the essential features of drone audio data, significantly improving the accuracy of classification.

[0063] (2) The present invention is based on audio data processing technology and is not restricted by light and line of sight. It can work stably in complex environments such as night, fog, haze, rain and snow, thereby enhancing the environmental adaptability of drone recognition.

[0064] (3) The proposed method is simple and efficient, and is suitable for real-time drone classification tasks. By combining the noise perception model with a convolutional neural network, it can be widely used in resource-constrained scenarios, providing a new solution for drone supervision, air traffic management, and safety protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0066] like Figure 1 As shown, the present invention provides a technical solution: a drone classification method based on a noise perception model, comprising the following steps:

[0067] Step S1: Acquire drone audio data, process the drone audio data to obtain mono drone audio data, resample the mono drone audio data to obtain resampled audio data, calculate the resampled audio data, and obtain audio data energy;

[0068] Step S2: Processing the audio data energy to obtain a time domain quantization index and a frequency domain quantization index, processing the time domain quantization index and the frequency domain quantization index to obtain a denoised time domain signal and a denoised frequency domain signal, fusing the time domain quantization index and the frequency domain quantization index with the denoised time domain signal and the denoised frequency domain signal to obtain an enhanced signal;

[0069] Step S3: Optimizing the enhanced signal using Mel spectrum transform to obtain a low-frequency band of the enhanced signal and a high-frequency band of the enhanced signal; analyzing the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal to generate a mixed-resolution spectrum; optimizing the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal using a differential Mel filter group to obtain a frequency band boundary of the differential Mel filter group under the control of a resolution adjustment coefficient; processing the mixed-resolution spectrum and the frequency band boundary of the differential Mel filter group under the control of the resolution adjustment coefficient to generate a Mel spectrum graph;

[0070] Step S4: Input the mel spectrogram into the convolutional neural network for processing and output the probability of the drone category.

[0071] The convolutional neural network consists of four sequentially connected convolutional blocks, each of which consists of a sequentially connected convolutional layer, a ReLU activation layer, and a pooling layer. The mel spectrogram is input into the convolutional neural network and passes through the four convolutional blocks in sequence to output the category probability of the drone.

[0072] Convolutional layer operation, said:

[0073] ;

[0074] Where, Indicates that the convolutional layer outputs the enhanced feature map at position The value of Indicates the convolution kernel size; Indicates that the convolution kernel is at position The weight value of represents the input Mel spectrogram; represents the bias term;

[0075] ReLU activation layer operation, indicating:

[0076] ;

[0077] Where, Represents the output of the ReLU activation function; represents the function that takes the maximum value; Indicates that the convolutional layer outputs the enhanced feature map at position The value of

[0078] The pooling layer operation downsamples the output of the ReLU activation function through the maximum pooling layer, which means:

[0079] ;

[0080] Where, Indicates that the output feature map of the pooling layer is at position The value of; R represents the range of the pooling window; Represents the output of the ReLU activation function;

[0081] The pooling layer output feature map is at position The value of is flattened into a one-dimensional vector and input into the fully connected layer. The probability of the drone category is calculated by the Softmax function, which is expressed as:

[0082] ;

[0083] Where, represents the probability of the i-th drone category; e represents a natural constant; Represents the output value of the fully connected layer for the i-th drone category; Indicates that the fully connected layer is Output value of each drone category; Indicates the total number of categories output by the fully connected layer.

[0084] The specific process of obtaining audio data energy is as follows:

[0085] Convert the acquired drone audio data into WAV format to obtain multi-channel drone audio data in WAV format;

[0086] Convert multi-channel drone audio data to mono drone audio data, which means:

[0087] ;

[0088] Where, Represents mono-channel drone audio data; Indicates the total number of channels for multi-channel drone audio data; Represents the drone audio data of the cth channel;

[0089] Resample the mono drone audio data to obtain resampled audio data, which means:

[0090] ;

[0091] Where, Represents resampled audio data; Represents a resampling operation; Indicates the original sampling rate of mono drone audio data; represents the target sampling rate;

[0092] Calculate the resampled audio data and obtain the audio data energy, which is expressed as:

[0093] ;

[0094] Where, Represents the energy of audio data; Represents the total number of resampling points; Represents the first sampling points.

[0095] The specific process of obtaining the enhanced signal is as follows:

[0096] Time-frequency dual-stream noise quantization uses an adaptive dual-stream noise perception algorithm (ADNPA) that integrates time and frequency domain features to quantize the audio data energy from both the time domain and frequency domain levels to obtain the time domain quantization index. and frequency domain quantitative indicators ;

[0097] Time domain quantitative indicators The specific process is:

[0098] The energy of the input audio data is framed and then calculated using the short-time Fourier transform (STFT) to obtain the energy of the kth frame. and the noise energy estimate of the kth frame , the minimum statistical noise tracking algorithm (MSNR) is used to track the energy of the kth frame and the noise energy estimate of the kth frame Perform dynamic updates to obtain time domain quantitative indicators ,express:

[0099] ;

[0100] Where, The total number of frames representing the energy of the audio data; represents the mean energy of the kth frame; represents the standard deviation of the energy of the kth frame; represents the hyperbolic tangent function; represents the adaptive adjustment factor; ; SNR means signal-to-noise ratio; log means logarithmic operation;

[0101] Frequency domain quantitative indicators The specific process is:

[0102] The audio data energy is extracted through the Mel filter bank to obtain the signal spectrum components of the vth frame respectively and the noise spectrum component of the vth frame , the noise floor estimation method (NBE) is used to calculate the signal spectrum components of the extracted v-th frame and the noise spectrum component of the vth frame , obtain frequency domain quantitative indicators ,express:

[0103] ;

[0104] Where, Indicates the filter channel index ; Indicates the total number of channels in the filter bank; Represents the slope parameter of the Sigmoid function, which controls the steepness of the transition region; represents the energy threshold; β = 0.8; τ = 0.6;

[0105] Normalize the signal-to-noise ratio value range to [0,1] to obtain the normalized signal-to-noise ratio ; ; Indicates the maximum value of the signal-to-noise ratio; Indicates the minimum value of the signal-to-noise ratio;

[0106] By normalizing the signal-to-noise ratio Calculate the parameters p and q, ; ; p represents the signal amplitude spectrum power adjustment parameter; q represents the dynamic noise suppression coefficient;

[0107] Calculate the frequency based on p and q The spectral attention mask value at , represents:

[0108] ;

[0109] Where, Indicates at frequency point The spectral attention mask value at ; Indicates the signal at the frequency point The amplitude spectrum at ; Indicates dynamic noise in the frequency band The average amplitude spectrum at ; Represents the adaptive scaling index Through frequency domain quantitative indicators adjust; represents the balance factor; where = ;

[0110] At the frequency point The spectral attention mask value at is used for noise reduction calculation to obtain the denoised time domain signal. and the frequency domain signal after noise reduction ; represents the inverse Fourier transform;

[0111] Fusion of denoised time domain signals by dynamic weights and the frequency domain signal after noise reduction and time domain quantitative indicators and frequency domain quantitative indicators , and get enhanced signal ,express:

[0112] .

[0113] The specific process of generating the Mel spectrogram is as follows:

[0114] First, according to the noise environment, based on the time domain quantitative indicators Calculate the resolution adjustment coefficient by summing the noise spectrum entropy , based on the resolution adjustment coefficient The time-frequency resolution of the enhanced signal is improved by using Mel spectrum transform Perform dynamic adjustment and optimization to obtain enhanced signal The low frequency band (harmonic components) and enhanced signal High frequency band (transient noise) to adapt to the feature extraction requirements of different frequency bands;

[0115] Resolution adjustment factor ,express:

[0116] ;

[0117] Where sigmoid represents the activation function; Hnoise represents the noise spectrum entropy (range 0~10), which represents the noise complexity; w1 represents the time domain quantitative index Weight; w2 represents the noise spectrum entropy weight; α∈(0,1); when α→1, the signal is enhanced The low-frequency high-resolution weight of the low-frequency band increases, and the signal is enhanced when α→0 The high-frequency time resolution weight of the high-frequency band is increased;

[0118] Dynamic adaptability: The traditional Mel spectrum transform uses a fixed window, while the Mel spectrum transform of the present invention adjusts the resolution adjustment coefficient in real time through the noise perception model. , for example, in the time domain quantitative indicators When it is low, the high frequency time resolution weight increases; in the time domain quantitative index When high, the low-frequency and high-resolution weights are increased to extract the characteristics of the harmonic components, directly using the time domain quantitative indicators The closed-loop optimization is formed by combining the noise spectrum entropy. Traditional methods do not achieve such cross-module coupling.

[0119] To enhance the signal Low frequency band and enhanced signal The high frequency band of is analyzed to generate a mixed resolution spectrum, which is represented by:

[0120] ;

[0121] Where, represents the mixed-resolution spectrum; Indicates enhanced signal The low-frequency band (0-2kHz) spectrum is obtained using a long-window short-time Fourier transform (STFT) (50ms window length, 1024 fast Fourier transform (FFT) points, and 10Hz / bin frequency resolution). Indicates enhanced signal The high frequency band (greater than 2kHz) spectrum of the short time window is obtained by short time Fourier transform (STFT) (10ms window length, 256 fast Fourier transform (FFT) points, and time resolution of 1ms / bin); kHz is kilohertz; Hz / bin is the frequency resolution; ms / bin is the time resolution;

[0122] The low-frequency harmonics of the drone rotor require low-frequency high resolution, while the high-frequency noise of the motor requires high-frequency time resolution. The traditional Mel spectrogram fixed window (such as 25ms) cannot take both into account. Low frequency band and enhanced signal By analyzing the high frequency band of the Low frequency band and enhanced signal The ability to express features in high-frequency bands;

[0123] Differentiable Mel filter bank based on resolution adjustment coefficient , dynamic fusion enhanced signal Low frequency band and enhanced signal The high frequency band of , indicating:

[0124] ;

[0125] Where, Indicates the Differentiable Mel filter bank with resolution adjustment coefficient Band boundaries under control; Indicates the Enhanced signal The low-frequency band (0-2kHz) filter boundaries are designed to be densely spaced (one filter every 50Hz); Indicates the Enhanced signal The high-frequency (greater than 2kHz) filter boundaries are designed to be sparsely spaced (one filter every 200Hz);

[0126] Based on the Differentiable Mel filter bank with resolution adjustment coefficient The frequency band boundaries under control define the A Mel filter at the frequency point The weight coefficient at ;

[0127] The traditional Mel filter bank has a fixed frequency band (e.g., 40 channels evenly distributed), while the adaptive multi-resolution time-frequency encoder AMR-TFE is designed to be differentiable, allowing dynamic adjustment of the enhanced signal during training. Low frequency band and enhanced signal The high-frequency band filter boundary is adapted to the physical characteristics of drone audio data;

[0128] Based on mixed resolution spectrum and A Mel filter at the frequency point The weight coefficient at , generates a mel spectrum, which is expressed as:

[0129] ;

[0130] Where, represents a mel-spectrogram; Indicates the number of fast Fourier transform FFT points; Indicates the A Mel filter at the frequency point The weight coefficient at ; Indicates a minimum value.

[0131] A drone classification system based on a noise perception model, applied to the aforementioned drone classification method based on a noise perception model, comprises:

[0132] An acquisition module is used to acquire drone audio data, process the drone audio data, obtain mono-channel drone audio data, resample the mono-channel drone audio data, obtain resampled audio data, calculate the resampled audio data, and obtain audio data energy;

[0133] a processing module for processing the audio data energy to obtain a time domain quantization index and a frequency domain quantization index, processing the time domain quantization index and the frequency domain quantization index to obtain a noise-reduced time domain signal and a noise-reduced frequency domain signal, and fusing the time domain quantization index and the frequency domain quantization index with the noise-reduced time domain signal and the noise-reduced frequency domain signal to obtain an enhanced signal;

[0134] An optimization module is used to optimize the enhanced signal using Mel spectrum transform to obtain a low-frequency band of the enhanced signal and a high-frequency band of the enhanced signal, analyze the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal to generate a mixed-resolution spectrum; optimize the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal through a differential Mel filter group to obtain a frequency band boundary of the differential Mel filter group under the control of a resolution adjustment coefficient, process the frequency band boundary of the mixed-resolution spectrum and the differential Mel filter group under the control of the resolution adjustment coefficient to generate a Mel spectrum map;

[0135] The output module is used to input the Mel spectrogram into the convolutional neural network for processing and output the probability of the drone category.

[0136] An electronic device includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute the drone classification method based on the noise perception model.

[0137] A non-volatile computer storage medium stores computer-executable instructions, and the computer can execute the method for classifying unmanned aerial vehicles based on a noise perception model.

[0138] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A drone classification method based on a noise perception model, characterized in that: The steps include: Step S1: Acquire drone audio data, process the drone audio data to obtain mono drone audio data, resample the mono drone audio data to obtain resampled audio data, calculate the resampled audio data, and obtain audio data energy; Step S2: Processing the audio data energy to obtain a time domain quantization index and a frequency domain quantization index, processing the time domain quantization index and the frequency domain quantization index to obtain a denoised time domain signal and a denoised frequency domain signal, fusing the time domain quantization index and the frequency domain quantization index with the denoised time domain signal and the denoised frequency domain signal to obtain an enhanced signal; Step S3: Optimizing the enhanced signal using Mel spectrum transform to obtain a low-frequency band of the enhanced signal and a high-frequency band of the enhanced signal; analyzing the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal to generate a mixed-resolution spectrum; optimizing the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal using a differential Mel filter group to obtain a frequency band boundary of the differential Mel filter group under the control of a resolution adjustment coefficient; processing the mixed-resolution spectrum and the frequency band boundary of the differential Mel filter group under the control of the resolution adjustment coefficient to generate a Mel spectrum graph; Step S4: Input the mel spectrogram into the convolutional neural network for processing and output the probability of the drone category.

2. The method for classifying drones based on a noise perception model according to claim 1, characterized in that: The specific process of obtaining the enhanced signal is: The energy of the input audio data is divided into frames and then calculated by short-time Fourier transform to obtain the energy of the kth frame. and the noise energy estimate of the kth frame , the energy of the kth frame is tracked using the minimum statistical noise tracking algorithm and the noise energy estimate of the kth frame Perform dynamic updates to obtain time domain quantitative indicators ,express: ； Where, The total number of frames representing the energy of the audio data; represents the mean energy of the kth frame; represents the standard deviation of the energy of the kth frame; represents the hyperbolic tangent function; represents the adaptive adjustment factor; ; SNR means signal-to-noise ratio; log means logarithmic operation; The audio data energy is extracted through the Mel filter bank to obtain the signal spectrum components of the vth frame respectively and the noise spectrum component of the vth frame , the noise floor estimation method is used to calculate the extracted signal spectrum components of the vth frame and the noise spectrum component of the vth frame , obtain frequency domain quantitative indicators ,express: ； Where, Indicates the total number of channels in the filter bank; Represents the slope parameter of the Sigmoid function; represents the energy threshold; e represents the natural constant; Normalize the signal-to-noise ratio value range to [0,1] to obtain the normalized signal-to-noise ratio ; ; Indicates the maximum value of the signal-to-noise ratio; Indicates the minimum value of the signal-to-noise ratio; By normalizing the signal-to-noise ratio Calculate the parameters p and q, ; ; p represents the signal amplitude spectrum power adjustment parameter; q represents the dynamic noise suppression coefficient; Calculate the frequency based on p and q The spectral attention mask value at , represents: ； Where, Indicates at frequency point The spectral attention mask value at ; Indicates the signal at the frequency point The amplitude spectrum at ; Indicates dynamic noise in the frequency band The average amplitude spectrum at ; Represents the adaptive scaling index Through frequency domain quantitative indicators adjust; represents the balance factor; where = ; At the frequency point The spectral attention mask value at is used for noise reduction calculation to obtain the denoised time domain signal. and the frequency domain signal after noise reduction ; represents the inverse Fourier transform; Fusion of denoised time domain signals by dynamic weights and the frequency domain signal after noise reduction and time domain quantitative indicators and frequency domain quantitative indicators , and get enhanced signal ,express: 。 3. The method for classifying drones based on a noise perception model according to claim 2, characterized in that: The specific process of generating a Mel spectrogram is as follows: According to the noise environment, based on time domain quantitative indicators Calculate the resolution adjustment coefficient by summing the noise spectrum entropy , based on the resolution adjustment coefficient The time-frequency resolution of the enhanced signal is improved by using Mel spectrum transform Perform dynamic adjustment and optimization to obtain enhanced signal Low frequency band and enhanced signal High frequency band; Resolution adjustment factor ,express: ； In the formula, sigmoid represents the activation function; Hnoise represents the noise spectrum entropy; w1 represents the time domain quantitative index Weight; w2 represents the noise spectrum entropy weight; To enhance the signal Low frequency band and enhanced signal The high frequency band of is analyzed to generate a mixed resolution spectrum, which is represented by: ； Where, represents the mixed-resolution spectrum; Indicates enhanced signal low-band spectrum; Indicates enhanced signal The high frequency spectrum; Differentiable Mel filter bank based on resolution adjustment coefficient , dynamic fusion enhanced signal Low frequency band and enhanced signal The high frequency band of , indicating: ； Where, Indicates the Differentiable Mel filter bank with resolution adjustment coefficient Band boundaries under control; Indicates the Enhanced signal The low-frequency filter boundary of Indicates the Enhanced signal The high-frequency filter boundary of Based on the Differentiable Mel filter bank with resolution adjustment coefficient The frequency band boundaries under control define the A Mel filter at the frequency point The weight coefficient at ; Based on mixed resolution spectrum and A Mel filter at the frequency point The weight coefficient at , generates a mel spectrum, which is expressed as: ； Where, represents a mel-spectrogram; Indicates the number of fast Fourier transform FFT points; Indicates the A Mel filter at the frequency point The weight coefficient at ; Indicates a minimum value.

4. The method for classifying drones based on a noise perception model according to claim 1, characterized in that: The convolutional neural network consists of four convolutional blocks connected in sequence, each of which consists of a convolutional layer, a ReLU activation layer, and a pooling layer connected in sequence. The mel spectrogram is input into the convolutional neural network and passes through the four convolutional blocks in sequence to output the category probability of the drone.

5. The method for classifying drones based on a noise perception model according to claim 3, characterized in that: The specific process of obtaining audio data energy is as follows: Convert the acquired drone audio data into a unified format to obtain multi-channel drone audio data in a unified format; Convert multi-channel drone audio data to mono drone audio data, which means: ； Where, Represents mono-channel drone audio data; Indicates the total number of channels for multi-channel drone audio data; Represents the drone audio data of the cth channel; Resample the mono drone audio data to obtain resampled audio data, which means: ； Where, Represents resampled audio data; Represents a resampling operation; Indicates the original sampling rate of mono drone audio data; represents the target sampling rate; Calculate the resampled audio data and obtain the audio data energy, which is expressed as: ； Where, Represents the energy of audio data; Represents the total number of resampling points; Represents the first sampling points.

6. A drone classification system based on a noise perception model, applied to the drone classification method based on a noise perception model described in any one of 1-5, characterized in that: include: An acquisition module is used to acquire drone audio data, process the drone audio data, obtain mono-channel drone audio data, resample the mono-channel drone audio data, obtain resampled audio data, calculate the resampled audio data, and obtain audio data energy; a processing module for processing the audio data energy to obtain a time domain quantization index and a frequency domain quantization index, processing the time domain quantization index and the frequency domain quantization index to obtain a noise-reduced time domain signal and a noise-reduced frequency domain signal, and fusing the time domain quantization index and the frequency domain quantization index with the noise-reduced time domain signal and the noise-reduced frequency domain signal to obtain an enhanced signal; An optimization module is used to optimize the enhanced signal using Mel spectrum transform to obtain a low-frequency band of the enhanced signal and a high-frequency band of the enhanced signal, analyze the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal to generate a mixed-resolution spectrum; optimize the low-frequency band of the enhanced signal and the high-frequency band of the enhanced signal through a differential Mel filter group to obtain a frequency band boundary of the differential Mel filter group under the control of a resolution adjustment coefficient, process the frequency band boundary of the mixed-resolution spectrum and the differential Mel filter group under the control of the resolution adjustment coefficient to generate a Mel spectrum map; The output module is used to input the Mel spectrogram into the convolutional neural network for processing and output the probability of the drone category.

7. An electronic device, characterized in that: The method comprises a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a drone classification method based on a noise perception model according to any one of claims 1 to 5.

8. A non-volatile computer storage medium storing computer-executable instructions, characterized in that: The computer-executable instructions execute the drone classification method based on the noise perception model described in any one of claims 1-5.

Images