Low-slow small unmanned aerial vehicle detection method based on acoustic fusion features

Abstract

The invention discloses a low-slow-small unmanned aerial vehicle detection method based on acoustic fusion features, relates to an unmanned aerial vehicle detection method, and belongs to the field of signal processing technologies and deep learning. The objective of the invention is to solve the problem that the traditional unmanned aerial vehicle detection method mainly depends on radar, vision, infrared and other technologies, but the detection accuracy of a low-speed small unmanned aerial vehicle is low in a complex urban environment. The method comprises the following steps: establishing an unmanned aerial vehicle audio signal data set in an urban environment; constructing an unmanned aerial vehicle detection network model; taking the unmanned aerial vehicle audio signal in the urban environment as the input of the unmanned aerial vehicle detection network model, and taking the detection result as the output of the unmanned aerial vehicle detection network model; the detection result is whether the unmanned aerial vehicle exists; obtaining a trained unmanned aerial vehicle detection network model; and inputting an audio signal in a to-be-detected urban environment into the trained unmanned aerial vehicle detection network model, and outputting a detection result by the trained unmanned aerial vehicle detection network model.

Description

Technical Field

[0001] The present invention relates to a drone detection method, and in particular to a low-speed, slow, and small drone detection method based on acoustic fusion features, which belongs to the field of signal processing technology and deep learning. Background Art

[0002] With the rapid development of drone technology, small drones with low flight altitude, slow flight speed and small radar reflection area have been widely used in agriculture, security, photography, logistics and other fields due to their small size, low cost and easy operation. However, the popularity of these drones has also brought a series of security and privacy issues. In complex urban environments, such as airports, government agencies, large event venues and other sensitive areas, the illegal intrusion and abuse of small drones has become an issue that cannot be ignored. Therefore, how to accurately and timely detect small drones in complex urban environments has become an important issue that needs to be solved urgently.

[0003] The detection of low-speed, slow, and small drones in complex urban environments faces multiple challenges. Due to the presence of a large number of interference sources such as high-rise buildings, vehicles, and crowds in urban environments, the effectiveness of traditional drone detection technology in this environment is severely limited. Traditional drone detection methods mainly rely on radar, vision, and infrared technologies. Radar technology is easily blocked and interfered by tall buildings in complex urban environments, resulting in reduced detection accuracy. The performance of the visual system in complex urban environments is greatly restricted by weather and lighting conditions. Its performance is greatly reduced at night or under severe weather conditions, and it cannot provide reliable detection results. Although infrared technology is effective under certain conditions, its detection accuracy for small targets is not high and it is easily affected by changes in ambient temperature. In contrast, sound signals have significant advantages in the detection of low-speed, slow, and small drones due to their advantages such as easy deployment and not being affected by light. Summary of the invention

[0004] The purpose of this invention is to solve the problem that traditional drone detection methods mainly rely on radar, vision, infrared and other technologies, but have low detection accuracy of low-speed and small drones in complex urban environments, and propose a low-speed and small drone detection method based on acoustic fusion features.

[0005] The specific process of the low-speed and slow small UAV detection method based on acoustic fusion features is as follows:

[0006] Step 1: Establish a drone audio signal dataset in an urban environment;

[0007] Step 2: Build a drone detection network model;

[0008] The drone detection network model includes the Mel-frequency cepstral coefficient MFCC module, the frame-level multi-head attention mechanism module, the time domain feature extractor, the feature fusion module and the detector;

[0009] The drone audio signal in the urban environment is used as the input of the drone detection network model, and the detection result is used as the output of the drone detection network model;

[0010] The detection result is whether a drone exists;

[0011] Train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model;

[0012] Step 3: Input the audio signal in the urban environment to be tested into the trained drone detection network model, and the trained drone detection network model outputs the detection result; if the detection result Result=1, it means that a drone exists, and if the detection result Result=0, it means that no drone exists.

[0013] Preferably, in step 1, a drone audio signal dataset is established in an urban environment; the specific process is:

[0014] Step 1: Obtain the drone audio signal dataset; the specific process is as follows:

[0015] Obtain a drone audio signal dataset from a public dataset, preprocess the drone audio signal to obtain the preprocessed drone audio signal, and obtain the drone audio signal dataset based on the preprocessed drone audio signal;

[0016] The process of preprocessing the drone audio signal is: unifying the sampling frequency, data size, and storage format of each audio signal data in turn;

[0017] Step 1 and 2: Obtain the urban environment audio signal dataset; the specific process is as follows:

[0018] Obtain an urban environment audio signal from a public data set, preprocess the urban environment audio signal to obtain a preprocessed urban environment audio signal, and obtain an urban environment audio signal data set based on the preprocessed urban environment audio signal;

[0019] The process of preprocessing the urban environment audio signal is: unifying the sampling frequency, data size and storage format of each audio signal data in turn;

[0020] Step 13: Combine the drone audio signal dataset obtained in step 11 and the urban environment audio signal dataset obtained in step 12 into one dataset, namely, the drone audio signal dataset in the urban environment.

[0021] Preferably, in step 2, a drone detection network model is constructed;

[0022] The drone detection network model includes the Mel-frequency cepstral coefficient MFCC module, the frame-level multi-head attention mechanism module, the time domain feature extractor, the feature fusion module and the detector;

[0023] The drone audio signal in the urban environment is used as the input of the drone detection network model, and the detection result is used as the output of the drone detection network model;

[0024] The detection result is whether a drone exists;

[0025] Train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model;

[0026] The specific process is:

[0027] Step 21: Input the drone audio signal in the urban environment into the Mel frequency cepstral coefficient MFCC module, and the Mel frequency cepstral coefficient MFCC module extracts the Mel frequency cepstral coefficient MFCC feature matrix F MFCC ;

[0028] Step 2: Convert the Mel frequency cepstral coefficient MFCC feature matrix F MFCC Input the frame-level multi-head attention mechanism module, and the frame-level multi-head attention mechanism module outputs the enhanced MFCC feature F A_MFCC ;

[0029] Step 2: Input the drone audio signal in the urban environment into the time domain feature extractor, and the time domain feature extractor outputs the deep time domain feature F Tim ;

[0030] Step 24: Enhance the MFCC feature F obtained by the frame-level multi-head attention mechanism in step 22 A_MFCC And the deep time domain features F obtained in steps 2 and 3 Tim Input feature fusion module, feature fusion module output fusion feature F Mix ;

[0031] Step 25: The fused features are input into the detector, and the detector outputs the detection result; if the detection result Result=1, it means that a drone exists; if the detection result Result=0, it means that no drone exists;

[0032] Step 26: Repeat steps 21 to 26 to train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model.

[0033] Preferably, in step 21, the audio signal of the drone in the urban environment is input into the Mel frequency cepstral coefficient MFCC module, and the Mel frequency cepstral coefficient MFCC module extracts the Mel frequency cepstral coefficient MFCC feature matrix F MFCC ;

[0034] The specific process is:

[0035] 1) The UAV audio signal in the urban environment is divided into frames. The process is as follows:

[0036] Set the length W of each frame signal and the frame overlap N to obtain a total number of frames M;

[0037] 2) First add a window to each frame signal, and then perform L-point short-time Fourier transform to obtain the time-frequency spectrum of each frame signal;

[0038] 3) Calculate the power spectrum of the time-frequency spectrum of each frame signal;

[0039] 4) Pass the power spectrum of each frame signal through n Mel filters, each filter outputs a Mel frequency bandwidth, and the n Mel frequency bandwidths corresponding to each frame signal constitute the Mel frequency spectrum of each frame signal;

[0040] 5) Perform logarithmic transformation on the Mel frequency spectrum of each frame signal, and perform discrete cosine transform on the Mel frequency spectrum after logarithmic transformation to obtain n Mel frequency cepstrum coefficients MFCC;

[0041] 6) The Mel frequency cepstral coefficients MFCC of M frames are combined into a feature matrix F of size n×M MFCC .

[0042] Preferably, in step 22, the Mel frequency cepstral coefficient MFCC feature matrix F MFCC Input the frame-level multi-head attention mechanism module, and the frame-level multi-head attention mechanism module outputs the enhanced MFCC feature F A_MFCC ; The specific process is:

[0043] Feature matrix F MFCC After the frame-level multi-head attention mechanism module, the enhanced MFCC feature F is obtained. A_MFCC , expressed as:

[0044] F A_MFCC =f MHAttention (F MFCC )

[0045] Among them, f MHAttention (·) represents the frame-level multi-head attention mechanism operation, F A_MFCC Represents the enhanced MFCC features.

[0046] Preferably, the feature matrix F MFCCAfter the frame-level multi-head attention mechanism module, the enhanced MFCC feature F is obtained. A_MFCC ; The specific process is:

[0047] 1) Feature matrix F MFCC Respectively with the weight matrix W Q , W K and W V Multiply them and get the query key Sum

[0048]

[0049] Among them, W Q , W K and W V is the weight matrix of the linear layer;

[0050] 2) Query key Sum After multiple head segmentation, we get num_heads of size d head × Q, K and V of M;

[0051] Among them, num_heads is the number of heads to be segmented, d head is the feature dimension of each head, d head = n / num_heads;

[0052] 3) Use the query Q and key K matrix after multi-head segmentation to calculate the attention score, and apply the normalization function f to the attention score softmax (·) is processed to obtain the attention weight F MHAttention :

[0053]

[0054] Where d is the scaling factor in the multi-head attention mechanism; the superscript T represents the matrix transpose;

[0055] Attention weight F MHAttention Size is d head ×d head ;

[0056] 4) Use attention weight F MHAttention Perform matrix multiplication weighting on the value V after multi-head segmentation to obtain the output of each head;

[0057] 5) The outputs of all heads num_heads are concatenated together through the concatenation layer to restore the matrix of size n×M;

[0058] The matrix restored to size n×M is expanded through a one-dimensional deconvolution layer to obtain the MFCC feature F enhanced by the frame-level multi-head attention mechanism with size n×M1. A_MFCC .

[0059] Preferably, in the steps 2 and 3, the drone audio signal in the urban environment is input into a time domain feature extractor, and the time domain feature extractor outputs a deep time domain feature F Tim ; The specific process is:

[0060] The drone audio signal in the urban environment is represented as a vector y. After the vector y passes through the time domain feature extractor, the deep time domain feature is obtained, which is expressed as:

[0061] F Tim =f Tim (y)

[0062] Among them, f Tim (·) represents the time domain feature extractor operation; F Tim Represents the extracted deep time domain features;

[0063] The process of the time domain feature extractor operation is as follows:

[0064] The time domain feature extractor network includes a first network layer, a second network layer, a third network layer, a fourth network layer, and a fifth network layer in sequence;

[0065] Each of the first network layer, the second network layer, the third network layer, the fourth network layer, and the fifth network layer includes a one-dimensional convolution layer, a normalization layer, an activation layer, and a random inactivation layer in sequence;

[0066] The vector y is input into the first network layer, the second network layer, the third network layer, the fourth network layer, and the fifth network layer in sequence. The fifth network layer outputs a deep time domain feature F of size n1×M2 Tim , where n1 is the channel dimension and M2 is the time step;

[0067] The working process of each network layer is as follows:

[0068] 1) Vector y passes through a one-dimensional convolutional layer, denoted as F Conv1d =f Conv1d (y);

[0069] Among them, f Conv1d (·) represents a one-dimensional convolution operation, F Conv1d Represents the time domain features obtained through the one-dimensional convolution layer;

[0070] 2) Time domain feature F Conv1d After the normalization layer, it is represented as F BatchNorm1d =f BatchNorm1d (FConv1d );

[0071] where f BatchNorm1d (·) represents the normalization operation, F BatchNorm1d Represents the normalized time domain features;

[0072] 3) Time domain feature F BatchNorm1d After the nonlinear activation function ReLu, it is expressed as F ReLu =f ReLu (F BatchNorm1d );

[0073] Among them, f ReLu (·) represents the nonlinear activation operation, F ReLu Represents the time domain features after passing through the nonlinear activation function ReLu layer;

[0074] 4) Time domain feature F ReLu After random dropout layer, denoted as F Dropout =f Dropout (F ReLu );

[0075] Among them, f Dropout (·) represents the random inactivation of some neurons, F Dropout Represents the features after the random dropout layer.

[0076] Preferably, in step 24, the MFCC feature F enhanced by the frame-level multi-head attention mechanism obtained in step 22 is A_MFCC And the deep time domain features F obtained in steps 2 and 3 Tim Input feature fusion module, feature fusion module output fusion feature F Mix ;

[0077] The specific process is:

[0078] 1) The obtained frame-level multi-head attention mechanism enhanced MFCC feature F A_MFCC and deep temporal features F Tim Perform alignment processing, the process is:

[0079] By interpolation method, the feature F A_MFCC and feature F Tim Aligned in the time dimension, i.e. M1 = M2;

[0080] 2) The feature F A_MFCC and feature F Tim The channel dimension is concatenated into an overall feature matrix F of size (n+n1)×M2, expressed as F=[F A_MFCC ; F Tim ];

[0081] 3) After the overall feature matrix F passes through the feature fusion module, the dynamically weighted fusion feature is obtained, which is expressed as:

[0082] F Mix =f Mix (F)

[0083] Among them, f Mix (·) represents the feature fusion module operation, F Mix Represents the fusion features after dynamic weighting;

[0084] The operation process of the feature fusion module is as follows:

[0085] The overall feature matrix F passes through the average pooling layer, linear layer, ReLu activation layer, linear layer, and Sigmoid activation layer in sequence. The Sigmoid activation layer outputs the dynamically weighted fusion feature F Mix .

[0086] Preferably, the fusion feature F in step 25 Mix Input the detector, and the detector outputs the detection result; if the detection result Result = 1, it means there is a drone, if the detection result Result = 0, it means there is no drone; the specific process is:

[0087] The detector sequentially comprises a first convolution module, a second convolution module, a third convolution module and a linear classification module;

[0088] Each of the first convolution module, the second convolution module, and the third convolution module includes a one-dimensional convolution layer, a normalization layer, an activation layer, and a random inactivation layer in sequence;

[0089] The linear classification module includes a first linear layer, a nonlinear activation function layer, and a second linear layer in sequence;

[0090] It is expressed as:

[0091] Result = f Det (F Mix )

[0092] Among them, f Det (·) represents the operation process of the detector;

[0093] Result indicates the result of the detector detecting the drone; if Result = 1, it means there is a drone, if Result = 0, it means there is no drone.

[0094] Preferably, in step 26, the drone detection network model is trained until the maximum number of iterations is reached to obtain a trained drone detection network model; the specific process is:

[0095] Set the initial learning rate and maximum learning rate;

[0096] Set the learning rate decay function;

[0097] Set the loss function to the cross entropy loss function;

[0098] Set the optimizer to Adam optimizer;

[0099] Set the number of iterations of training epoch;

[0100] Set the amount of training data for each iteration during training;

[0101] Repeat steps 21 to 26 to train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model.

[0102] The beneficial effects of the present invention are:

[0103] The present invention proposes a low-speed, slow and small UAV detection method based on acoustic fusion features and deep learning, aiming to improve the accuracy and robustness of UAV detection in complex urban environments by fusing the time domain and frequency domain features of UAV audio and using deep learning methods for detection.

[0104] The low-flying, slow, and small UAV detection method based on acoustic fusion features described in the present invention detects low-flying, slow, and small UAVs through time-frequency feature fusion and deep learning technology. This method enhances the feature distinction by fusing the different attention levels of audio features, and uses convolutional neural networks for detection, thereby improving the accuracy and robustness of UAV detection in complex urban environments.

[0105] The low-speed, slow, and small UAV detection method based on acoustic fusion features described in the present invention utilizes a multi-head attention mechanism combined with Mel-cephalogram coefficients, and enhances the distinguishability and anti-interference ability of features by adaptively adjusting the coefficient weights, thereby improving the accuracy of UAV detection in complex urban environments.

[0106] The low-speed, slow, and small UAV detection method based on acoustic fusion features described in the present invention utilizes a time delay neural network (TDNN) to extract deep time domain features of UAV audio, effectively capturing timing information and long-term dependencies. Through a delay and feedback mechanism, it is able to identify and utilize cross-time correlations in UAV audio signals, thereby capturing time domain features that are difficult to extract using traditional methods, and improving the accuracy of UAV detection in complex urban environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 This is a principle block diagram of the low-speed, slow, and small UAV detection method based on acoustic fusion features described in the present invention;

[0108] Figure 2 This is a deep time domain feature extraction framework diagram of the low-speed and slow small UAV detection method based on acoustic fusion features described in the present invention;

[0109] Figure 3 This is a classifier framework diagram of the low-speed and slow small UAV detection method based on acoustic fusion features described in the present invention;

[0110] Figure 4 This is an overall framework diagram of the low-speed, slow, and small UAV detection method based on acoustic fusion features described in the present invention;

[0111] Figure 5 This is a graph showing the detection accuracy results of the low-speed, slow, and small UAV detection method based on acoustic fusion features described in the present invention. DETAILED DESCRIPTION

[0112] Specific implementation method 1: The specific process of the low-speed and slow small UAV detection method based on acoustic fusion features in this implementation method is as follows:

[0113] In order to accurately detect the existence of drones in complex urban environments, the present invention aims at the problem of detecting low, slow and small drones in complex urban environments. It uses time domain and frequency domain fusion features and deep learning technology to construct a drone detection network based on acoustic features, aiming to improve the accuracy of low, slow and small drone detection.

[0114] Step 1: Establish a drone audio signal dataset in a complex urban environment;

[0115] Step 2: Build a drone detection network model;

[0116] The drone detection network model includes the Mel-frequency cepstral coefficient MFCC module, the frame-level multi-head attention mechanism module, the time domain feature extractor, the feature fusion module and the detector;

[0117] The drone audio signal in a complex urban environment is used as the input of the drone detection network model, and the detection result is used as the output of the drone detection network model;

[0118] The detection result is whether a drone exists;

[0119] Train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model;

[0120] Step 3: Input the audio signal in the complex urban environment to be tested into the trained drone detection network model, and the trained drone detection network model outputs the detection result; if the detection result Result = 1, it means that a drone exists, and if the detection result Result = 0, it means that no drone exists.

[0121] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that in step 1, a drone audio signal dataset is established in a complex urban environment; the specific process is as follows:

[0122] Step 1: Obtain the drone audio signal dataset; the specific process is as follows:

[0123] Obtain a drone audio signal dataset from a public dataset, preprocess the drone audio signal to obtain the preprocessed drone audio signal, and obtain the drone audio signal dataset based on the preprocessed drone audio signal;

[0124] The process of preprocessing the drone audio signal is as follows: unifying the sampling frequency (16000Hz resampling), data size (3s cutting of each audio), and storage format (wav) of each audio signal data in turn;

[0125] Step 1 and 2: Obtain a complex urban environment audio signal dataset; the specific process is as follows:

[0126] Acquire complex urban environment audio signals from a public data set, preprocess the complex urban environment audio signals to obtain preprocessed complex urban environment audio signals, and obtain a complex urban environment audio signal data set based on the preprocessed complex urban environment audio signals;

[0127] The process of preprocessing the audio signal in a complex urban environment is as follows: unifying the sampling frequency (16000Hz resampling), data size (cutting each audio by 3s), and storage format (wav) of each audio signal data in turn;

[0128] Step 13: Combine the drone audio signal dataset obtained in step 11 and the complex urban environment audio signal dataset obtained in step 12 into one dataset, namely, the drone audio signal dataset in a complex urban environment.

[0129] The drone audio signal dataset is obtained by collecting or using existing public datasets. In order to enhance the wide applicability and representativeness of the data, the drone audio signal is preprocessed, including unifying the sampling frequency, data size, and storage format of each audio data, to complete the construction of the drone audio signal dataset. In order to realize the detection of drones in complex urban environments, a complex urban environment audio signal dataset is also established. This dataset is obtained by collecting or using existing public datasets, and the same preprocessing is performed as the drone audio signal dataset. Then the two datasets are combined into one dataset, which is the drone audio signal dataset in a complex urban environment.

[0130] The experiment used an open source drone sound dataset from the Internet. The drone type was the Bebop produced by Parrot. The complex urban environment audio dataset used the Urban Sound 8K public dataset, which covers sound data in a variety of urban scenes, such as car horns, children playing, dogs barking, street music, etc., capturing background noise and interference in different environments, ensuring the diversity and representativeness of the data.

[0131] The experiment used 8 types of audio data, including drones, car horns, street music, bird calls, children playing, dog barking, street music, and air conditioner sounds. There were 300 groups of each type of audio data, for a total of 2,400 groups of audio signals. The file was stored in wav format. The audio signal was resampled to 16,000 Hz, and then each audio was cut for 3 seconds. The length of each audio data was 48,000 sampling points.

[0132] The other steps and parameters are the same as those in the first embodiment.

[0133] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that a drone detection network model is constructed in step two;

[0134] The drone detection network model includes the Mel-frequency cepstral coefficient MFCC module, the frame-level multi-head attention mechanism module, the time domain feature extractor, the feature fusion module and the detector;

[0135] The drone audio signal in a complex urban environment is used as the input of the drone detection network model, and the detection result is used as the output of the drone detection network model;

[0136] The detection result is whether a drone exists;

[0137] Train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model;

[0138] The specific process is:

[0139] Step 21: Input the drone audio signal in a complex urban environment into the Mel-Frequency Cepstral Coefficients (MFCC) module, and extract the Mel-Frequency Cepstral Coefficients (MFCC) feature matrix F MFCC ;

[0140] Step 2: Convert the Mel frequency cepstral coefficient MFCC feature matrix F MFCC Input the frame-level multi-head attention mechanism module, and the frame-level multi-head attention mechanism module outputs the enhanced MFCC feature F A_MFCC ;

[0141] Step 2 and 3: Input the drone audio signal in a complex urban environment into the time domain feature extractor, and the time domain feature extractor outputs the deep time domain feature F Tim ;

[0142] Step 24: Enhance the MFCC feature F obtained by the frame-level multi-head attention mechanism in step 22 A_MFCC And the deep time domain features F obtained in steps 2 and 3 Tim Input feature fusion module, feature fusion module output fusion feature F Mix ;

[0143] Step 25: The fused features are input into the detector, and the detector outputs the detection result; if the detection result Result=1, it means that a drone exists; if the detection result Result=0, it means that no drone exists;

[0144] Step 26: Repeat steps 21 to 26 to train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model.

[0145] The other steps and parameters are the same as those in the first or second embodiment.

[0146] Specific implementation method 4: This implementation method is different from the specific implementation methods 1 to 3 in that in step 21, the drone audio signal in a complex urban environment is input into the Mel-Frequency Cepstral Coefficients MFCC (Mel-Frequency Cepstral Coefficients) module, and the Mel-Frequency Cepstral Coefficients MFCC module extracts the Mel-Frequency Cepstral Coefficients MFCC feature matrix F MFCC ;

[0147] The specific process is:

[0148] 1) The UAV audio signal in a complex urban environment is divided into frames. The process is as follows:

[0149] Set the length W of each frame signal and the frame overlap N to obtain a total number of frames M;

[0150] 2) First add a window to each frame signal, and then perform L-point short-time Fourier transform to obtain the time-frequency spectrum of each frame signal;

[0151] 3) Calculate the power spectrum of the time-frequency spectrum of each frame signal;

[0152] 4) Pass the power spectrum of each frame signal through n Mel filters, each filter outputs a Mel frequency bandwidth, and the n Mel frequency bandwidths corresponding to each frame signal constitute the Mel frequency spectrum of each frame signal;

[0153] 5) Perform logarithmic transformation on the Mel frequency spectrum of each frame signal, and perform discrete cosine transform on the Mel frequency spectrum after logarithmic transformation to obtain n Mel frequency cepstrum coefficients MFCC;

[0154] 6) The Mel frequency cepstral coefficients MFCC of M frames are combined into a feature matrix F of size n×M MFCC .

[0155] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0156] Specific implementation method 5: This implementation method is different from the specific implementation methods 1 to 4 in that in step 22, the Mel frequency cepstral coefficient MFCC feature matrix F MFCC Input the frame-level multi-head attention mechanism module, and the frame-level multi-head attention mechanism module outputs the enhanced MFCC feature F A_MFCC ; The specific process is:

[0157] In order to enhance the representation ability of MFCC features, a frame-level multi-head attention mechanism (Multi-Head Attention, MHA) is introduced. This attention mechanism dynamically and locally weights the MFCC features of each frame and can adaptively adjust the coefficient weights to effectively capture the local frequency domain features of the audio signal while ensuring that important feature information is retained and enhanced, thereby enhancing the distinguishability and anti-interference ability of the features.

[0158] Feature matrix F MFCC After the frame-level multi-head attention mechanism module, the enhanced MFCC feature F is obtained. A_MFCC , expressed as:

[0159] F A_MFCC =f MHAttention (F MFCC )

[0160] Among them, f MHAttention (·) represents the frame-level multi-head attention mechanism operation, F A_MFCC Represents enhanced MFCC features;

[0161] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0162] Specific implementation method 6: This implementation method is different from the specific implementation methods 1 to 5 in that the characteristic matrix F MFCC After the frame-level multi-head attention mechanism module, the enhanced MFCC feature F is obtained. A_MFCC ; The specific process is:

[0163] Through the linear layer qkv_linear, the input feature matrix F MFCCPerform a linear transformation, and then split it into different matrices according to the same number of heads num_heads. The essence of this operation is to divide the feature space so that different attention heads can learn features from different subspaces;

[0164] 1) Feature matrix F MFCC Respectively with the weight matrix W Q , W K and W V Multiply them and get the query (queries), keys (keys) and values (values):

[0165]

[0166] Among them, W Q , W K and W V is the weight matrix of the linear layer, W Q , W K and W V The size of is n×n, and the input features are converted into query, key and value respectively;

[0167] 2) Query key Sum After multiple head segmentation, we get num_heads of size d head × Q, K and V of M;

[0168] Among them, num_heads is the number of heads to be segmented, d head is the feature dimension of each head, d head = n / num_heads;

[0169] 3) Use the query Q and key K matrix after multi-head segmentation to calculate the attention score, and apply the normalization function f to the attention score softmax (·) is processed to obtain the attention weight F MHAttention :

[0170]

[0171] Where d is the scaling factor in the multi-head attention mechanism, which is used to scale to prevent the dot product result from being too large; the superscript T represents the matrix transpose;

[0172] Attention weight F MHAttention Size is d head ×d head ;

[0173] Q, K and V are determined by the input features FMFCC The back propagation algorithm in the training process is used to update the optimization;

[0174] 4) Use attention weight F MHAttention Perform matrix multiplication weighting on the value V after multi-head segmentation to obtain the output of each head;

[0175] 5) Concatenate the outputs of all heads num_heads together through the concatenation layer (Concat) to restore the matrix of size n×M;

[0176] The matrix restored to size n×M is expanded through a one-dimensional deconvolution layer (ConvTranspose1d) to obtain the MFCC feature F enhanced by the frame-level multi-head attention mechanism with size n×M1. A_MFCC .

[0177] The number of training data for each batch during the frame-level multi-head attention mechanism training process is batch_size;

[0178] The MHA configuration is shown in Table 1;

[0179] Table 1. Frame-level multi-head attention mechanism network configuration table

[0180] Modules Network Layer quantity MHA qkv_linear+Sofmax+Concat+ConvTransposeld 1

[0181] In the experiment, the window function is a Hamming window, W=400, N=200, L=400, n=40, M=241 is determined by the length of the audio signal, num_heads=8, M1=462, d=5, batch_size=64.

[0182] The other steps and parameters are the same as those in Specific Implementations 1 to 5.

[0183] Specific implementation method 7: This implementation method is different from the specific implementation methods 1 to 6 in that in the steps 2 and 3, the drone audio signal in the complex urban environment is input into the time domain feature extractor, and the time domain feature extractor outputs the deep time domain feature F Tim ; The specific process is:

[0184] like Figure 2 As shown, a time-delay neural network is used to construct a time-domain feature extractor, which consists of multiple network layers (TD_Layer), each of which includes a one-dimensional convolution layer (Conv1d), a normalization layer (BatchNorm1d), an activation layer (ReLu), and a random inactivation layer (Dropout);

[0185] The drone audio signal in a complex urban environment is represented as a vector y. After the vector y passes through the time domain feature extractor, the deep time domain feature is obtained, which is expressed as:

[0186] F Tim =f Tim (y)

[0187] Among them, f Tim (·) represents the time domain feature extractor operation; F Tim Represents the extracted deep time domain features;

[0188] The process of the time domain feature extractor operation is as follows:

[0189] The time domain feature extractor network includes the first network layer (TD_Layer), the second network layer (TD_Layer), the third network layer (TD_Layer), the fourth network layer (TD_Layer), and the fifth network layer (TD_Layer) in sequence;

[0190] Each of the first network layer, the second network layer, the third network layer, the fourth network layer, and the fifth network layer includes a one-dimensional convolution layer (Conv1d), a normalization layer (BatchNorm1d), an activation layer (ReLu), and a random inactivation layer (Dropout) in sequence;

[0191] The vector y is input into the first network layer, the second network layer, the third network layer, the fourth network layer, and the fifth network layer in sequence. The fifth network layer outputs a deep time domain feature F of size n1×M2 Tim , where n1 is the channel dimension and M2 is the time step;

[0192] The working process of each network layer is as follows:

[0193] 1) Vector y passes through a one-dimensional convolutional layer (Conv1d), represented by F Conv1d =f Conv1d (y);

[0194] Among them, f Conv1d (·) represents a one-dimensional convolution operation, F Conv1d Represents the time domain features obtained through the one-dimensional convolution layer. The function of the convolution layer is to filter the input signal through a sliding window to extract representative time domain features;

[0195] 2) Time domain feature F Conv1d After the normalization layer (BatchNorm1d), it is represented as F BatchNorm1d =f BatchNorm1d (F Conv1d );

[0196] where f BatchNorm1d(·) represents the normalization operation, F BatchNorm1d Represents the normalized time domain features; the normalization layer standardizes the output of the convolution layer, reduces internal covariate shift, accelerates network training, and improves the stability and generalization ability of the model;

[0197] 3) After the normalization layer, the temporal feature F BatchNorm1d After the nonlinear activation function ReLu, it is expressed as F ReLu =f ReLu (F BatchNorm1d );

[0198] Among them, f ReLu (·) represents the nonlinear activation operation, F ReLu Represents the time domain features after passing through the nonlinear activation function ReLu layer; after applying the activation function, the network can learn and represent more complex time domain features; next, in order to prevent overfitting;

[0199] 4) After the activation layer, the temporal feature F ReLu After random dropout layer, denoted as F Dropout =f Dropout (F ReLu );

[0200] Among them, f Dropout (·) represents the random inactivation of some neurons, F Dropout Represents the features after the random dropout layer; the random dropout layer enhances the robustness and generalization ability of the model by randomly setting the output of some neurons to zero, and prevents overfitting during model training;

[0201] Then, repeat the above network layer (TD_Layer) to extract the time domain features of the audio signal layer by layer. As the number of network layers (TD_Layer) increases, the network can capture more and more high-order and abstract time domain features. The final network layer (TD_Layer) outputs a deep time domain feature F of size n1×M2. Tim , where n1 is the channel dimension and M2 is the time step;

[0202] In the experiment, the number of network layers (TD_Layer) is 5, the number of training data for each batch during the training process of the time domain feature extractor network is 64, n1=512, M2=462.

[0203] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0204] Specific implementation eight: This implementation differs from any one of the specific implementations one to seven in that in step 24, the MFCC feature F enhanced by the frame-level multi-head attention mechanism obtained in step 22 is A_MFCC And the deep time domain features F obtained in steps 2 and 3 Tim Input feature fusion module, feature fusion module output fusion feature F Mix ;

[0205] The specific process is:

[0206] In the processing of drone audio signals, time domain features and frequency domain features each provide important information about the drone. In order to make full use of these two types of features, a feature fusion method based on the attention mechanism is adopted to construct a feature fusion unit based on the attention mechanism.

[0207] 1) The obtained frame-level multi-head attention mechanism enhanced MFCC feature F A_MFCC and deep temporal features F Tim Perform alignment processing, the process is:

[0208] By interpolation method, the feature F A_MFCC and feature F Tim Aligned in the time dimension, i.e., M1=M2, to ensure consistent time steps;

[0209] 2) The feature F A_MFCC and feature F Tim The channel dimension is concatenated into an overall feature matrix F of size (n+n1)×M2, expressed as F=[F A_MFCC ; F Tim ], the spliced ​​features contain comprehensive information in the time domain and frequency domain, which can more comprehensively characterize the drone audio signal;

[0210] 3) After the overall feature matrix F passes through the feature fusion module, the dynamically weighted fusion feature is obtained, which is expressed as:

[0211] F Mix =f Mix (F)

[0212] Among them, f Mix (·) represents the feature fusion module operation, F Mix Represents the fusion features after dynamic weighting;

[0213] The operation process of the feature fusion module is as follows:

[0214] The overall feature matrix F passes through the average pooling layer, linear layer, ReLu activation layer, linear layer, and Sigmoid activation layer in sequence. The Sigmoid activation layer outputs the dynamically weighted fusion feature F Mix .

[0215] The feature fusion module introduces the global channel attention mechanism Squeeze and Excitation (SE) module, which calculates the weight of each feature and performs weighted processing on all features to obtain the dynamically weighted fusion feature F. Mix ; This feature contains comprehensive information in the time domain and frequency domain after being weighted by the global channel attention mechanism, and the expression ability of important features is enhanced through the attention mechanism;

[0216] The SE attention mechanism configuration is shown in Table 2, including an average pooling layer, two linear layers, an activation layer, and an S-type activation layer.

[0217] Table 2 Attention mechanism network configuration table

[0218] Modules Network Layer quantity SE Average pooling+Linear+ReLu+Linear+Sigmoid 1

[0219] In the experiment, the number of training data in each batch during SE training is 64, and the size of the feature matrix F is 552×462;

[0220] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0221] Specific implementation method 9: This implementation method is different from the specific implementation methods 1 to 8 in that the fusion feature F in step 25 Mix Input the detector, and the detector outputs the detection result; if the detection result Result = 1, it means there is a drone, if the detection result Result = 0, it means there is no drone; the specific process is:

[0222] The detector sequentially comprises a first convolution module, a second convolution module, a third convolution module and a linear classification module;

[0223] Each of the first convolution module, the second convolution module, and the third convolution module includes a one-dimensional convolution layer (Conv1d), a normalization layer (BatchNorm1d), an activation layer (ReLu), and a random inactivation layer (Dropout) in sequence;

[0224] The linear classification module includes a first linear layer, a nonlinear activation function layer, and a second linear layer in sequence.

[0225] It is expressed as:

[0226] Result = f Det (F Mix )

[0227] Among them, fDet (·) represents the calculation process of the detector, and Result represents the result of the detector detecting the drone; if Result = 1, it means there is a drone, and if Result = 0, it means there is no drone.

[0228] like Figure 3 As shown in the figure, a convolutional neural network is used to build a drone detector. The detector consists of three convolution modules and a linear classification module. The convolution module consists of a one-dimensional convolution layer (Conv1d), a normalization layer (BatchNorm1d), an activation layer (ReLu), and a random dropout layer (Dropout). The linear classification module consists of two linear layers (Linear) and a nonlinear activation function. The nonlinear activation function is applied after the first linear layer. By linearly transforming the features, the feature dimension is gradually reduced, thereby improving the expression ability of the model.

[0229] The fusion feature F Mix After passing through the detector, the detection result is obtained, which is expressed as:

[0230] Result = f Det (F Mix )

[0231] Among them, f Det (·) represents the calculation process of the detector, and Result represents the result of the detector detecting the drone. If Result = 1, it means there is a drone, and if Result = 0, it means there is no drone. The drone detector structure makes full use of the comprehensive feature information of the drone audio signal in the time domain and frequency domain, and realizes the detection of the drone audio signal through reasonable feature processing and linear classification.

[0232] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0233] Specific implementation method ten: This implementation method is different from the specific implementation methods one to nine in that in step 26, the drone detection network model is trained until the maximum number of iterations is reached to obtain a trained drone detection network model; the specific process is:

[0234] Set the initial learning rate and maximum learning rate;

[0235] Set the learning rate decay function;

[0236] Set the loss function to the cross entropy loss function;

[0237] Set the optimizer to Adam optimizer;

[0238] Set the number of iterations of training epoch;

[0239] Set the number of training data for each iteration (each batch) during training;

[0240] Repeat steps 21 to 26 to train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model.

[0241] The overall model framework diagram of the drone detection method of the present invention is as follows: Figure 4 As shown, firstly, step 2 is used to extract the fusion feature F from the training set data. Mix , and then use the detector to get the detection result Result; then the loss function calculates the loss value Loss through the detection result Result and the true label; next, the optimizer optimizes the overall parameters according to the loss function value Loss, calculates the parameter gradient of each layer through back propagation, and updates the network parameters according to the adaptive gradient optimization;

[0242] In the experiment, the training set includes 1920 groups of drone audio signals and interference audio signals, and the test set includes 480 groups of drone audio signals and interference audio signals. The initial learning rate is 0.00001, and WarmupCosineSchedulerLR is used as the learning rate decay function. The maximum learning rate is set to 0.001; the number of iterative training epochs is set to 100, and the number of training data for each batch during training is 64; the model training uses the cross entropy loss function and the Adam optimizer.

[0243] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0244] Testing the Drone Detector

[0245] The experiment simulates a complex urban environment and detects low-speed, slow, and small drones, that is, detects the presence of drones in a complex urban environment. The accuracy of drone detection by the method of the present invention is as follows: Figure 5 As shown. Figure 5 It can be seen that as the number of iterative training times of the network model increases, the accuracy of the model gradually increases. When epoch is 91, the accuracy of the drone detection network is the highest, with the highest accuracy of 99.13%. The experimental results show that the method of the present invention can accurately detect drones in complex environments.

[0246] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A low-speed, slow, and small UAV detection method based on acoustic fusion features, characterized by: The specific process of the method is: Step 1: Establish a drone audio signal dataset in an urban environment; Step 2: Build a drone detection network model; The drone detection network model includes the Mel-frequency cepstral coefficient MFCC module, the frame-level multi-head attention mechanism module, the time domain feature extractor, the feature fusion module and the detector; The drone audio signal in the urban environment is used as the input of the drone detection network model, and the detection result is used as the output of the drone detection network model; The detection result is whether a drone exists; Train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model; Step 3: Input the audio signal in the urban environment to be tested into the trained drone detection network model, and the trained drone detection network model outputs the detection result; if the detection result Result=1, it means that a drone exists, and if the detection result Result=0, it means that no drone exists.

2. The low-flying, slow-moving, small UAV detection method based on acoustic fusion features according to claim 1 is characterized by: In step 1, a drone audio signal dataset is established in an urban environment; the specific process is as follows: Step 1: Obtain the drone audio signal dataset; the specific process is as follows: Obtain a drone audio signal dataset from a public dataset, preprocess the drone audio signal to obtain the preprocessed drone audio signal, and obtain the drone audio signal dataset based on the preprocessed drone audio signal; The process of preprocessing the drone audio signal is: unifying the sampling frequency, data size, and storage format of each audio signal data in turn; Step 12: Obtain an urban environment audio signal dataset; The specific process is: Obtain an urban environment audio signal from a public data set, preprocess the urban environment audio signal to obtain a preprocessed urban environment audio signal, and obtain an urban environment audio signal data set based on the preprocessed urban environment audio signal; The process of preprocessing the urban environment audio signal is: unifying the sampling frequency, data size and storage format of each audio signal data in turn; Step 13: Combine the drone audio signal dataset obtained in step 11 and the urban environment audio signal dataset obtained in step 12 into one dataset, namely, the drone audio signal dataset in the urban environment.

3. The low-flying, slow-moving, small UAV detection method based on acoustic fusion features according to claim 2 is characterized by: In the step 2, a drone detection network model is constructed; The drone detection network model includes the Mel-frequency cepstral coefficient MFCC module, the frame-level multi-head attention mechanism module, the time domain feature extractor, the feature fusion module and the detector; The drone audio signal in the urban environment is used as the input of the drone detection network model, and the detection result is used as the output of the drone detection network model; The detection result is whether a drone exists; Train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model; The specific process is: Step 21: Input the drone audio signal in the urban environment into the Mel frequency cepstral coefficient MFCC module, and the Mel frequency cepstral coefficient MFCC module extracts the Mel frequency cepstral coefficient MFCC feature matrix F MFCC ; Step 2: Convert the Mel frequency cepstral coefficient MFCC feature matrix F MFCC Input the frame-level multi-head attention mechanism module, and the frame-level multi-head attention mechanism module outputs the enhanced MFCC feature F A_MFCC ; Step 2: Input the drone audio signal in the urban environment into the time domain feature extractor, and the time domain feature extractor outputs the deep time domain feature F Tim ; Step 24: Enhance the MFCC feature F obtained by the frame-level multi-head attention mechanism in step 22 A_MFCC And the deep time domain features F obtained in steps 2 and 3 Tim Input feature fusion module, feature fusion module output fusion feature F Mix ; Step 25: The fused features are input into the detector, and the detector outputs the detection result; if the detection result Result=1, it means that a drone exists; if the detection result Result=0, it means that no drone exists; Step 26: Repeat steps 21 to 26 to train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model.

4. The low-flying, slow-moving, small UAV detection method based on acoustic fusion features according to claim 3 is characterized by: In the step 21, the audio signal of the drone in the urban environment is input into the Mel-frequency cepstral coefficient MFCC module, and the Mel-frequency cepstral coefficient MFCC module extracts the Mel-frequency cepstral coefficient MFCC feature matrix F MFCC ; The specific process is: 1) The UAV audio signal in the urban environment is divided into frames. The process is as follows: Set the length W of each frame signal and the frame overlap N to obtain a total number of frames M; 2) First add a window to each frame signal, and then perform L-point short-time Fourier transform to obtain the time-frequency spectrum of each frame signal; 3) Calculate the power spectrum of the time-frequency spectrum of each frame signal; 4) Pass the power spectrum of each frame signal through n Mel filters, each filter outputs a Mel frequency bandwidth, and the n Mel frequency bandwidths corresponding to each frame signal constitute the Mel frequency spectrum of each frame signal; 5) Perform logarithmic transformation on the Mel frequency spectrum of each frame signal, and perform discrete cosine transform on the Mel frequency spectrum after logarithmic transformation to obtain n Mel frequency cepstrum coefficients MFCC; 6) The Mel frequency cepstral coefficients MFCC of M frames are combined into a feature matrix F of size n×M MFCC .

5. The method for detecting low-flying, slow-moving, small UAVs based on acoustic fusion features according to claim 4 is characterized in that: In step 22, the Mel frequency cepstral coefficient MFCC feature matrix F MFCC Input the frame-level multi-head attention mechanism module, and the frame-level multi-head attention mechanism module outputs the enhanced MFCC feature F A_MFCC ; The specific process is: Feature matrix F MFCC After the frame-level multi-head attention mechanism module, the enhanced MFCC feature F is obtained. A_MFCC , expressed as: F A_MFCC =f MHAttention (F MFCC ) Among them, f MHAttention (·) represents the frame-level multi-head attention mechanism operation, F A_MFCC Represents the enhanced MFCC features.

6. The low-flying, slow-moving, small UAV detection method based on acoustic fusion features according to claim 5 is characterized by: The feature matrix F MFCC After the frame-level multi-head attention mechanism module, the enhanced MFCC feature F is obtained. A_MFCC ; The specific process is: 1) Feature matrix F MFCC Respectively with the weight matrix W Q , W K and W V Multiply them and get the query key Sum Among them, W Q , W K and W V is the weight matrix of the linear layer; 2) Query key Sum After multiple head segmentation, we get num_heads of size d head × Q, K and V of M; Among them, num_heads is the number of heads to be segmented, d head is the feature dimension of each head, d head = n / num_heads; 3) Use the query Q and key K matrix after multi-head segmentation to calculate the attention score, and apply the normalization function f to the attention score softmax (·) is processed to obtain the attention weight F MHAttention : Where d is the scaling factor in the multi-head attention mechanism; the superscript T represents the matrix transpose; Attention weight F MHAttention Size is d head ×d head ; 4) Use attention weight F MHAttention Perform matrix multiplication weighting on the value V after multi-head segmentation to obtain the output of each head; 5) The outputs of all heads num_heads are concatenated together through the concatenation layer to restore the matrix of size n×M; The matrix restored to size n×M is expanded through a one-dimensional deconvolution layer to obtain the MFCC feature F enhanced by the frame-level multi-head attention mechanism with size n×M1. A_MFCC .

7. The method for detecting low-flying, slow-moving, small UAVs based on acoustic fusion features according to claim 6 is characterized in that: In the steps 2 and 3, the drone audio signal in the urban environment is input into the time domain feature extractor, and the time domain feature extractor outputs the deep time domain feature F Tim ; The specific process is: The drone audio signal in the urban environment is represented as a vector y. After the vector y passes through the time domain feature extractor, the deep time domain feature is obtained, which is expressed as: F Tim =f Tim (y) Among them, f Tim (·) represents the time domain feature extractor operation; F Tim Represents the extracted deep time domain features; The process of the time domain feature extractor operation is as follows: The time domain feature extractor network includes a first network layer, a second network layer, a third network layer, a fourth network layer, and a fifth network layer in sequence; Each of the first network layer, the second network layer, the third network layer, the fourth network layer, and the fifth network layer includes a one-dimensional convolution layer, a normalization layer, an activation layer, and a random inactivation layer in sequence; The vector y is input into the first network layer, the second network layer, the third network layer, the fourth network layer, and the fifth network layer in sequence. The fifth network layer outputs a deep time domain feature F of size n1×M2 Tim , where n1 is the channel dimension and M2 is the time step; The working process of each network layer is as follows: 1) Vector y passes through a one-dimensional convolutional layer, denoted as F Conv1d =f Conv1d (y); Among them, f Conv1d (·) represents a one-dimensional convolution operation, F Conv1d Represents the time domain features obtained through the one-dimensional convolution layer; 2) Time domain feature F Conv1d After the normalization layer, it is represented as F BatchNorm1d =f BatchNorm1d (F Conv1d ); where f BatchNorm1d (·) represents the normalization operation, F BatchNorm1d Represents the normalized time domain features; 3) Time domain feature F BatchNorm1d After the nonlinear activation function ReLu, it is expressed as F ReLu =f ReLu (F BatchNorm1d ); Among them, f ReLu (·) represents the nonlinear activation operation, F ReLu Represents the time domain features after passing through the nonlinear activation function ReLu layer; 4) Time domain feature F ReLu After random dropout layer, denoted as F Dropout =f Dropout (F ReLu ); Among them, f Dropout (·) represents the random inactivation of some neurons, F Dropout Represents the features after the random dropout layer.

8. The method for detecting low-flying, slow-moving, small UAVs based on acoustic fusion features according to claim 7 is characterized in that: In step 24, the MFCC feature F enhanced by the frame-level multi-head attention mechanism obtained in step 22 is A_MFCC And the deep time domain features F obtained in steps 2 and 3 Tim Input feature fusion module, feature fusion module output fusion feature F Mix ; The specific process is: 1) The obtained frame-level multi-head attention mechanism enhanced MFCC feature F A_MFCC and deep temporal features F Tim Perform alignment processing, the process is: By interpolation method, the feature F A_MFCC and feature F Tim Aligned in the time dimension, i.e. M1 = M2; 2) The feature F A_MFCC and feature F Tim The channel dimension is concatenated into an overall feature matrix F of size (n+n1)×M2, expressed as F=[F A_MFCC ; F Tim ]; 3) After the overall feature matrix F passes through the feature fusion module, the dynamically weighted fusion feature is obtained, which is expressed as: F Mix =f Mix (F) Among them, f Mix (·) represents the feature fusion module operation, F Mix Represents the fusion features after dynamic weighting; The operation process of the feature fusion module is as follows: The overall feature matrix F passes through the average pooling layer, linear layer, ReLu activation layer, linear layer, and Sigmoid activation layer in sequence. The Sigmoid activation layer outputs the dynamically weighted fusion feature F Mix .

9. The method for detecting low-flying, slow-moving, small UAVs based on acoustic fusion features according to claim 8 is characterized in that: In step 25, the fusion feature F Mix Input the detector, and the detector outputs the detection result; if the detection result Result = 1, it means there is a drone, if the detection result Result = 0, it means there is no drone; the specific process is: The detector sequentially comprises a first convolution module, a second convolution module, a third convolution module and a linear classification module; Each of the first convolution module, the second convolution module, and the third convolution module includes a one-dimensional convolution layer, a normalization layer, an activation layer, and a random inactivation layer in sequence; The linear classification module includes a first linear layer, a nonlinear activation function layer, and a second linear layer in sequence; It is expressed as: Result=f Det (F Mix ) Among them, f Det (·) represents the operation process of the detector; Result indicates the result of the detector detecting the drone; if Result = 1, it means there is a drone, if Result = 0, it means there is no drone.

10. The method for detecting low-flying, slow-moving, small UAVs based on acoustic fusion features according to claim 9 is characterized in that: In step 26, the drone detection network model is trained until the maximum number of iterations is reached to obtain a trained drone detection network model; the specific process is as follows: Set the initial learning rate and maximum learning rate; Set the learning rate decay function; Set the loss function to the cross entropy loss function; Set the optimizer to Adam optimizer; Set the number of iterations of training epoch; Set the amount of training data for each iteration during training; Repeat steps 21 to 26 to train the drone detection network model until the maximum number of iterations is reached to obtain a trained drone detection network model.

Images