Unmanned aerial vehicle acoustic signal identification method and system based on meta transfer learning

Abstract

The invention discloses an unmanned aerial vehicle acoustic signal identification method and system based on meta transfer learning, and relates to the signal identification technology, and the method comprises the steps: obtaining sound samples of an unmanned aerial vehicle in different environments, different flight states and distances in advance, and extracting the feature vectors of the sound samples; defining learning tasks of a plurality of meta transfer learning models, wherein each learning task is used for identifying different types of unmanned aerial vehicles or unmanned aerial vehicles under different environmental conditions from the feature vectors of the sound samples; for a newly obtained unmanned aerial vehicle sound signal, performing fine tuning on the pre-trained meta-migration learning model by using part of the annotated data; and performing sound signal identification of the unmanned aerial vehicle by using the fine-tuned meta transfer learning model. According to the invention, learning and adaptation can be carried out faster on a new task, and the accuracy of acoustic signal identification is improved.

Description

Technical Field

[0001] The present application relates to the field of signal recognition technology, and in particular to a method and system for unmanned aerial vehicle acoustic signal recognition based on meta-transfer learning. Background Art

[0002] Drone technology has rapidly emerged and has become a focus of attention in the civil and military fields. It has demonstrated great application value in monitoring, reconnaissance, logistics, agriculture, film and television production and other fields. With the expansion of usage scenarios, it has become particularly important to ensure airspace safety and monitor drones in real time. In this context, acoustic signal recognition technology has become a key means of drone monitoring and management. However, there are many challenges in identifying the acoustic signals of drones, including interference from environmental noise and the diversity and complexity of the acoustic signals themselves. Wind and traffic noise in the environment, as well as the differences in sound produced by different drone models, greatly increase the difficulty of sound recognition. Although there are many acoustic signal processing solutions, they often do not work well when dealing with the complexity of the actual environment. Summary of the invention

[0003] The embodiments of the present application provide a method and system for drone acoustic signal recognition based on meta-transfer learning, which utilizes meta-transfer learning to learn the commonalities between tasks on multiple tasks, thereby better handling the diversity and complexity of acoustic signals, enabling faster learning and adaptation on new tasks, and improving the accuracy of acoustic signal recognition.

[0004] The present application embodiment provides a method for identifying drone acoustic signals based on meta-transfer learning, including:

[0005] Acquire sound samples of the drone in different environments, flight states and distances in advance, and extract feature vectors of the sound samples;

[0006] Define multiple learning tasks of the meta-transfer learning model, each learning task is used to identify drones of different types or under different environmental conditions from the feature vectors of the sound samples;

[0007] For the newly acquired drone sound signals, the pre-trained meta-transfer learning model is fine-tuned using some labeled data;

[0008] Using the fine-tuned meta-transfer learning model, perform drone acoustic signal recognition.

[0009] Optionally, obtaining sound samples of the drone in different environments, flight states and distances in advance, and extracting feature vectors of the sound samples includes:

[0010] Obtain drone sound samples in cities, rural areas, forests, different weather conditions, different flight states, and different flight distances;

[0011] Performing downsampling and pre-emphasis pre-processing on the acquired sound samples;

[0012] For the processed sound samples, a framing operation is performed, wherein each frame of the framing contains N sampling points, and there is an overlapping part with a preset overlapping rate between two adjacent frames;

[0013] Use Hamming window to perform windowing processing on the framed signal;

[0014] The windowed framed signal is subjected to Fourier transform to obtain a frequency domain signal sequence.

[0015] Optionally, pre-acquiring sound samples of the drone in different environments, different flight states and distances, and extracting feature vectors of the sound samples further includes:

[0016] Calculate the power spectral density P(m,k) based on the frequency domain signal sequence obtained by Fourier transform;

[0017] Define the filter bank, which consists of I triangular filters, each with a center frequency of f(i), and the corresponding frequency response function satisfies:

[0018]

[0019] Where k represents the current frequency point, f(i+1), f(i), and f(i-1) represent the center frequency points of the i+1th, ith, and i-1th filters respectively;

[0020] The calculated power spectral density P(m,k) is mapped to the Mel frequency scale using the defined filter bank to obtain the Mel frequency power spectral density (MPSD);

[0021] Perform discrete cosine transform (DCT) on the Mel frequency power spectral density to obtain MFCC coefficients;

[0022] Calculate the first-order difference (ΔMFCC) and the second-order difference (ΔΔMFCC) of the MFCC coefficients;

[0023] The obtained MFCC, ΔMFCC and ΔΔMFCC are used as the input of the feature extraction network.

[0024] Optionally, define multiple learning tasks of the meta-transfer learning model, each learning task for identifying different types of drones or drones under different environmental conditions from the feature vectors of the sound samples, including:

[0025] In the training process of the meta-transfer learning model, the pre-trained network parameters are used, and the scaling and translation operations are used to describe the parameter update process, where scaling and translation correspond to adjusting the model weight W and bias b respectively, and the scaling parameter is defined as φs1 , the translation parameter is defined as φ s2 , then the updated parameters W′ and b′ can be expressed as:

[0026] W′=W⊙φ s1

[0027] b′=b+φ s2

[0028] Among them, ⊙ represents element-wise product;

[0029] Then the output of the feature extraction network of meta-transfer learning satisfies:

[0030]

[0031] Among them, X represents the input features of the feature extraction network, W and b represent the weights and biases of the network, and φ si Represents scaling and translation parameters;

[0032] Update the zoom and translation parameters to satisfy:

[0033]

[0034] Among them, β represents the learning rate of each subtask, represents the loss function of the meta-training stage, T train Represents the training data of the meta-training task, corresponding to the data in the support set of each task, and Θ represents the parameters of the model feature extraction network.

[0035] Optionally, defining multiple learning tasks of the meta-transfer learning model, each learning task for identifying different types of drones or drones under different environmental conditions from the feature vectors of the sound samples also includes:

[0036] During the training process, the parameters θ of the feature extraction network of the model are not updated, but the parameters θ of the classifier are fine-tuned. For each subtask, θ is updated according to the gradient descent algorithm using the training data:

[0037]

[0038] Among them, θ′ represents the updated classifier parameters;

[0039] The model parameters φ are fine-tuned according to the current task support set s ' i and θ′, the classifier θ and the scaling and translation parameters φ si The update for the current task is:

[0040]

[0041] Among them, γ represents the learning rate of meta-learning.

[0042] Optionally, for the newly acquired drone sound signal, fine-tuning the pre-trained meta-transfer learning model using some labeled data includes:

[0043] Load the trained Transformer model and introduce new drone acoustic signal recognition task annotation data so that the Transformer model can perform training based on the pre-training results;

[0044] During the training of the Transformer model based on the results of pre-training, the weight parameters of the backend of the Transformer model are adjusted, and the weight update uses the gradient descent method to meet the following requirements:

[0045]

[0046] Among them, W new is the updated weight, is the gradient of the loss function with respect to the weight;

[0047] The training data set is divided into k subsets, and part of the subset is set aside as the test set each time, and the remaining subset is the training set. Repeat the training k times, and select a different subset as the test set each time. The final model performance evaluation is based on the average of the k test results:

[0048]

[0049] Among them, Performance is the performance evaluation result, Accuracy i is the accuracy of the i-th iteration.

[0050] Optionally, for the newly acquired drone sound signal, fine-tuning the pre-trained meta-transfer learning model using part of the labeled data also includes:

[0051] The architecture of the Transformer model is optimized through model pruning and knowledge distillation to reduce model complexity.

[0052] An embodiment of the present application also proposes a drone acoustic signal recognition system based on meta-transfer learning, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the drone acoustic signal recognition method based on meta-transfer learning as described above are implemented.

[0053] The embodiments of the present application utilize meta-transfer learning to learn the commonalities between tasks on multiple tasks, thereby better handling the diversity and complexity of acoustic signals, being able to learn and adapt faster on new tasks, and improving the accuracy of acoustic signal recognition.

[0054] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0056] Figure 1 The following is a basic flow chart of a method for identifying drone acoustic signals based on meta-transfer learning according to an embodiment of the present application;

[0057] Figure 2 This is a schematic diagram of the MFCC feature extraction process of an embodiment of the present application;

[0058] Figure 3 The figure shows the meta-transfer learning network architecture of the embodiment of the present application. DETAILED DESCRIPTION

[0059] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0060] The present application embodiment provides a method for identifying drone acoustic signals based on meta-transfer learning, such as Figure 1 As shown, the following steps are included:

[0061] In step S101, sound samples of the drone in different environments, flight states and distances are acquired in advance, and feature vectors of the sound samples are extracted.

[0062] like Figure 2As shown, in some embodiments, pre-acquiring sound samples of the drone in different environments, different flight states and distances, and extracting the feature vectors of the sound samples includes: acquiring sound samples of the drone in cities, rural areas, forests, different weather conditions, different flight states and different flight distances. For example, the sound characteristics of the drone in various environments (including cities, rural areas, forests), various weather conditions, different flight states (take-off, cruising, hovering, landing) and multiple distances (ranging from a few meters to tens of meters) can be captured. The collection work is carried out using high-sensitivity microphones and recording equipment to ensure that the full spectrum from the subtle sound of the propeller to the overall flight noise can be covered. The accuracy of sound source positioning is enhanced by array microphone technology to achieve effective sound capture within a distance of 50 meters, while maintaining the sampling rate of the recording equipment at least 44.1kHz to ensure that the details of the sound data are fully captured.

[0063] The acquired sound samples are pre-processed by downsampling and pre-emphasis.

[0064] like Figure 2 As shown, for the processed sound samples, a framing operation is performed, wherein each frame of the framing contains N sampling points, and there is an overlapping part with a preset overlapping rate between two adjacent frames. Specifically, after the preprocessing is completed, the audio signal of the mth frame is obtained, which is represented by x(n). Next, a framing operation is performed on the preprocessed audio signal x(n), wherein each frame contains N sampling points, that is, the length of each frame is N. An overlapping rate of, for example, 50% is set between two adjacent frames to ensure the continuity of information and reduce information loss. Assuming that the sampling rate of the audio is fs, the actual time length of each frame can be calculated as:

[0065] T=N / f s

[0066] The framed signal is windowed using a Hamming window to reduce the spectrum leakage effect of the discrete signal. Assuming that the window function of the mth frame is w(m, n), the signal after the window function of the mth frame is weighted is:

[0067] XW(m,n)=X(n)× W (m,n)

[0068] Perform Fourier transform on the windowed framed signal to obtain a frequency domain signal sequence. The obtained frequency domain representation is:

[0069] X(m,k)=FFT[xw(m,n)]

[0070] By applying Fourier transform to the signal sequence xw(m,n) of the mth frame, we can convert it from the time domain to the frequency domain, thereby obtaining the frequency domain sequence X(m,k). This sequence describes the strength of different frequency components in the signal.

[0071] In some embodiments, pre-acquiring sound samples of the drone in different environments, different flight states and distances, and extracting feature vectors of the sound samples further includes:

[0072] The power spectral density P(m,k) is calculated based on the frequency domain signal sequence obtained by Fourier transform. The power spectral density (PSD) is a function that describes the change of signal power distribution with frequency. It is the square of Fourier transform and represents the signal power at different frequencies. The power spectral density P(m,k) is calculated as:

[0073]

[0074] Define a filter bank, which consists of I triangular filters, where each triangular filter in a group of triangular filters corresponds to a specific Mel frequency point. This design enables the triangular filter to effectively filter and downsample the signal on the Mel frequency scale, thereby extracting audio features that are more in line with the human ear perception characteristics. The specific center frequency of each filter is f(i), and the corresponding frequency response function satisfies:

[0075]

[0076] Where k represents the current frequency point, f(i+1), f(i), and f(i-1) represent the center frequency points of the i+1th, ith, and i-1th filters respectively;

[0077] The calculated power spectral density P(m,k) is mapped to the Mel frequency scale using the defined filter bank to obtain the Mel frequency power spectral density (MPSD) as:

[0078]

[0079] Perform discrete cosine transform (DCT) on the Mel frequency power spectral density to obtain the MFCC coefficients:

[0080]

[0081] Among them, j represents the dimension of MFCC coefficients, and I represents the number of triangular filters.

[0082] The first-order difference (ΔMFCC) of the MFCC coefficients is calculated as:

[0083]

[0084] Wherein, L is the frame number range of the difference calculation, for example, L is set to 3.

[0085] Compared with the general difference using two adjacent frames for calculation, the embodiment of the present application uses more previous and next frames to calculate ΔMFCC, thereby reducing interference caused by noise and disturbance, and further calculates the second-order difference (ΔΔMFCC) as:

[0086]

[0087] The obtained MFCC, ΔMFCC and ΔΔMFCC are used as the input of the feature extraction network.

[0088] In step S102, a plurality of learning tasks of the meta-transfer learning model are defined, each learning task being used to identify drones of different types or under different environmental conditions from feature vectors of sound samples.

[0089] In step S103, for the newly acquired drone sound signal, the pre-trained meta-transfer learning model is fine-tuned using part of the annotated data. For the new drone acoustic signal recognition task, a small amount of annotated data is used to fine-tune the pre-trained meta-learning model, and the model parameters and structure are adjusted to optimize the recognition performance, including accuracy, response speed, and robustness.

[0090] In step S104, the fine-tuned meta-transfer learning model is used to perform drone sound signal recognition.

[0091] The embodiments of the present application utilize meta-transfer learning to learn the commonalities between tasks on multiple tasks, thereby better handling the diversity and complexity of acoustic signals, being able to learn and adapt faster on new tasks, and improving the accuracy of acoustic signal recognition.

[0092] After the sound data has been accurately collected and features extracted, the training phase of the meta-transfer learning model begins to play a role. The meta-transfer learning strategy combines the core advantages of meta-learning and transfer learning, with the goal of enabling the model to quickly adapt when encountering new categories of data. The core of this method is to generate and learn multiple subtasks. Through iterative learning and application transfer on these subtasks, the model can gain strong adaptability and can efficiently adapt to new tasks with only a few gradient descents. Meta-transfer learning effectively addresses the problem of traditional meta-learning being prone to overfitting when dealing with complex network structures and small sample data. By randomly sampling the subtasks generated, the model is guided to learn weights with high adaptability.

[0093] like Figure 3 As shown, meta-transfer learning is divided into two processes: meta-training and meta-testing. In some embodiments, multiple learning tasks of meta-transfer learning models are defined, and each learning task is used to identify drones of different types or under different environmental conditions from feature vectors of sound samples, including:

[0094] like Figure 3As shown, during the training process of the meta-transfer learning model, the model will perform meta-learning fine-tuning on the feature extraction network. In order to avoid the risk of overfitting caused by directly fine-tuning the entire feature extraction network, the embodiment of the present application uses pre-trained network parameters, and only needs to update a small number of parameters to accelerate the training process of the model. In the specific example, scaling and translation operations are used to describe the parameter update process to avoid the risk of overfitting, where scaling and translation correspond to adjusting the model weight W and bias b respectively, and the scaling parameter is defined as φ s1 , the translation parameter is defined as φ s2 , then the updated parameters W′ and b′ can be expressed as:

[0095] W′=W⊙φ s1

[0096] b′=b+φ s2

[0097] Among them, ⊙ represents element-wise product;

[0098] Then the output of the feature extraction network of meta-transfer learning satisfies:

[0099]

[0100] Among them, X represents the input features of the feature extraction network, W and b represent the weights and biases of the network, and φ si Represents scaling and translation parameters;

[0101] Update the zoom and translation parameters to satisfy:

[0102]

[0103] Among them, β represents the learning rate of each subtask, represents the loss function of the meta-training stage, T train Represents the training data of the meta-training task, corresponding to the data in the support set of each task, and Θ represents the parameters of the model feature extraction network.

[0104] In some embodiments, defining a plurality of learning tasks of a meta-transfer learning model, each learning task for identifying drones of different types or under different environmental conditions from a feature vector of a sound sample further includes:

[0105] During the training process, the parameters θ of the feature extraction network of the model are not updated, but the parameters θ of the classifier are fine-tuned. For each subtask, θ is updated according to the gradient descent algorithm using the training data:

[0106]

[0107] Among them, θ′ represents the updated classifier parameters, and the φ obtained by fine-tuning each subtask iss ' i Both θ and θ′ are temporary and are not the final optimization results of the current task. After each subtask completes fine-tuning training, the test data, i.e., the query set, is used to evaluate the quality of model fine-tuning.

[0108] The model parameters φ are fine-tuned according to the current task support set s ' i and θ′, the classifier θ and the scaling and translation parameters φ si The update for the current task is:

[0109]

[0110] Among them, γ represents the learning rate of meta-learning.

[0111] Afterwards, for each given meta-training task, the translation, scaling and classifier parameters are optimized and updated continuously to finally complete the training of the meta-training stage.

[0112] In the meta-testing phase, the translation, scaling, and classifier parameters are fine-tuned using the support set of the new task, and the prediction results obtained from the query set of the new task are used to evaluate the recognition performance of the model. In addition, in order to enhance the generalization ability of the model, a difficult category sampling retraining strategy is introduced to specifically select cases where recognition fails for key training. This targeted optimization strategy significantly improves the adaptability and accuracy of the model.

[0113] In some embodiments, for a newly acquired drone sound signal, fine-tuning the pre-trained meta-transfer learning model using part of the labeled data includes:

[0114] Load a trained Transformer model, that is, a model trained on a large-scale dataset. The model has excellent generalization and feature extraction capabilities, and introduces new drone acoustic signal recognition task labeled data, so that the Transformer model can perform training based on the pre-training results.

[0115] During the training process of the Transformer model based on the results of pre-training, adjust the weight parameters of the back-end part of the Transformer model. Specifically, during the training process, focus on adjusting the weight parameters of the back-end part of the Transformer network, which is closer to the task-specific features, especially the last few layers of the network. During fine-tuning, the learning rate can be accurately set to 0.0001 to ensure that the model learning is accelerated while avoiding the loss of pre-training knowledge. Implement data enhancement to increase the diversity of training samples by changing the audio rate and adjusting the pitch to improve the generalization ability of the model. The learning rate adjustment formula is:

[0116]

[0117] Among them, η base is the basic learning rate, set to 0.0001; decay_rate is the decay rate, selected as 0.95; epoch represents the current iteration round; decay_steps is the decay step, set according to the amount of training data, and can be set to decay once every 100 rounds. During the fine-tuning process, the weight update of the Transformer model follows the gradient descent method, satisfying:

[0118]

[0119] Among them, W new is the updated weight, is the gradient of the loss function with respect to the weight;

[0120] Divide the training data set into k subsets, set aside some subsets as test sets each time, and use the remaining subsets as training sets. Repeat the training k times, and select different subsets as test sets each time. For example, use cross-validation and integration of multiple Transformer models to comprehensively improve performance, and divide the data set into k subsets. Set aside one subset each time as the test set, and use the remaining k-1 subsets for training. Repeat this process k times, and select different subsets as test sets each time. The final model performance evaluation is based on the average of the k test results:

[0121]

[0122] Among them, Performance is the performance evaluation result, Accuracy i is the accuracy of the i-th iteration.

[0123] In some embodiments, for a newly acquired drone sound signal, fine-tuning the pre-trained meta-transfer learning model using part of the labeled data further includes:

[0124] The Transformer model architecture is optimized through model pruning and knowledge distillation to reduce model complexity, improve response speed, and achieve efficient operation in real-time applications.

[0125] Enhance the robustness of the model, simulate sound interference for data enhancement, and introduce adversarial training samples to improve the model's adaptability to challenges.

[0126] Through the above steps, the drone acoustic signal recognition system can effectively cope with new recognition tasks. The fine-tuning and optimization of the Transformer classification network not only improves the system accuracy and response speed, but also ensures reliability in complex environments by enhancing robustness, promoting the development of acoustic signal recognition technology.

[0127] Further step S104 uses the fine-tuned meta-transfer learning model to perform drone sound signal recognition including:

[0128] First, the model is evaluated on an independent test set that includes unseen drone types and new environmental conditions. This test set is specially designed to cover a variety of possible operating scenarios, such as urban and forest environments, distances ranging from 5 meters to 100 meters, and different weather conditions such as sunny and rainy days. Evaluation metrics include accuracy, recall, precision, and F1 score, where the target for accuracy is set to at least 95% to ensure high performance of the model. The response time of the model is critical for real-time recognition, and the goal is to control the recognition time within 500 milliseconds.

[0129] When integrating a well-trained model into a drone monitoring system, model optimization measures are taken to meet the needs of real-time processing. Model optimization includes model quantization and pruning techniques, which aim to reduce the storage footprint of the model to less than 50% of the original size without significantly affecting the accuracy of the model. The optimized model is able to process at least 20 sound signals per second on standard hardware, ensuring the real-time recognition capability of the system.

[0130] After the model is deployed, continuous performance monitoring will be implemented to capture and resolve operational issues in real time. The system will also receive regular model updates, at least every six months, to introduce new drone acoustic data and adapt to the latest advances in acoustic signal processing technology, keeping the system's recognition accuracy above 95%.

[0131] According to the specific evaluation and deployment scheme, the drone acoustic signal recognition method of this application can work accurately and efficiently in various practical application scenarios, providing strong technical support for the safety monitoring and management of drones. This application can not only respond to the drone's acoustic signals in real time, but also adapt to changes in technological progress and application needs through continuous monitoring and regular model updates.

[0132] This application can quickly adapt to new drone acoustic signal recognition tasks, including identifying unknown drone types and coping with new environmental conditions, ensuring its continued effectiveness in the evolving field of drone technology. By comprehensively utilizing cross-validation and model integration methods, this system not only demonstrates high accuracy and strong generalization capabilities in various environments, but also achieves real-time or near-real-time acoustic signal processing capabilities through model optimization, meeting the high efficiency requirements in practical applications.

[0133] An embodiment of the present application also proposes a drone acoustic signal recognition system based on meta-transfer learning, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the drone acoustic signal recognition method based on meta-transfer learning as described above are implemented.

[0134] It should be noted that in the various embodiments of the present application, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0135] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0136] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0137] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A method for identifying drone acoustic signals based on meta-transfer learning, characterized in that: include: Acquire sound samples of the drone in different environments, flight states and distances in advance, and extract feature vectors of the sound samples; Define multiple learning tasks of the meta-transfer learning model, each learning task is used to identify drones of different types or under different environmental conditions from the feature vectors of the sound samples; For the newly acquired drone sound signals, the pre-trained meta-transfer learning model is fine-tuned using some labeled data; Using the fine-tuned meta-transfer learning model, perform drone acoustic signal recognition.

2. The method for identifying drone acoustic signals based on meta-transfer learning according to claim 1, characterized in that: Acquiring the sound samples of the drone in different environments, different flight states and distances in advance, and extracting the feature vectors of the sound samples includes: Obtain sound samples of drones in cities, rural areas, forests, different weather conditions, different flight states, and different flight distances; Performing downsampling and pre-emphasis pre-processing on the acquired sound samples; For the processed sound samples, a framing operation is performed, wherein each frame of the framing contains N sampling points, and there is an overlapping part with a preset overlapping rate between two adjacent frames; Use Hamming window to perform windowing processing on the framed signal; The windowed framed signal is subjected to Fourier transform to obtain a frequency domain signal sequence.

3. The method for identifying drone acoustic signals based on meta-transfer learning according to claim 1, characterized in that: Pre-acquiring sound samples of the drone in different environments, different flight states and distances, and extracting feature vectors of the sound samples also includes: Calculate the power spectral density P(m,k) based on the frequency domain signal sequence obtained by Fourier transform; Define the filter bank, which consists of I triangular filters, each with a center frequency of f(i), and the corresponding frequency response function satisfies: Where k represents the current frequency point, f(i+1), f(i), and f(i-1) represent the center frequency points of the i+1th, ith, and i-1th filters respectively; The calculated power spectral density P(m,k) is mapped to the Mel frequency scale using the defined filter bank to obtain the Mel frequency power spectral density (MPSD); Perform discrete cosine transform (DCT) on the Mel frequency power spectral density to obtain MFCC coefficients; Calculate the first-order difference (ΔMFCC) and the second-order difference (ΔΔMFCC) of the MFCC coefficients; The obtained MFCC, ΔMFCC and ΔΔMFCC are used as the input of the feature extraction network.

4. The method for identifying drone acoustic signals based on meta-transfer learning according to claim 1, characterized in that: Define multiple learning tasks of meta-transfer learning models, each of which is used to identify drones of different types or in different environmental conditions from the feature vectors of sound samples, including: In the training process of the meta-transfer learning model, the pre-trained network parameters are used, and the scaling and translation operations are used to describe the parameter update process, where scaling and translation correspond to adjusting the model weight W and bias b respectively, and the scaling parameter is defined as φ s1 , the translation parameter is defined as φ s2 , then the updated parameter W ′ and b ′ It can be expressed as: IN ′ =W⊙φ s1 b ′ =b+φ s2 Among them, ⊙ represents element-wise product; Then the output of the feature extraction network of meta-transfer learning satisfies: Among them, X represents the input features of the feature extraction network, W and b represent the weights and biases of the network, and φ si Represents scaling and translation parameters; Update the zoom and translation parameters to satisfy: Among them, β represents the learning rate of each subtask, represents the loss function of the meta-training stage, T train Represents the training data of the meta-training task, corresponding to the data in the support set of each task, and Θ represents the parameters of the model feature extraction network.

5. The method for identifying drone acoustic signals based on meta-transfer learning according to claim 4, characterized in that: Defining multiple learning tasks of meta-transfer learning models, each of which is used to identify drones of different types or in different environmental conditions from the feature vectors of sound samples, also includes: During the training process, the parameters θ of the feature extraction network of the model are not updated, but the parameters θ of the classifier are fine-tuned. For each subtask, θ is updated according to the gradient descent algorithm using the training data: Among them, θ′ represents the updated classifier parameters; The model parameters φ are fine-tuned according to the current task support set s ' i and θ′, the classifier θ and the scaling and translation parameters φ si The update for the current task is: Among them, γ represents the learning rate of meta-learning.

6. The method for identifying drone acoustic signals based on meta-transfer learning according to claim 5, characterized in that: For the newly acquired drone sound signal, the pre-trained meta-transfer learning model is fine-tuned using some labeled data, including: Load the trained Transformer model and introduce new drone acoustic signal recognition task annotation data so that the Transformer model can perform training based on the pre-training results; During the training of the Transformer model based on the results of pre-training, the weight parameters of the backend of the Transformer model are adjusted, and the weight update uses the gradient descent method to meet the following requirements: Among them, W new is the updated weight, is the gradient of the loss function with respect to the weight; The training data set is divided into k subsets, and part of the subset is set aside as the test set each time, and the remaining subset is the training set. Repeat the training k times, and select a different subset as the test set each time. The final model performance evaluation is based on the average of the k test results: Among them, Performance is the performance evaluation result, Accuracy i is the accuracy of the i-th iteration.

7. The method for identifying drone acoustic signals based on meta-transfer learning according to claim 6, characterized in that: For the newly acquired drone sound signals, fine-tuning the pre-trained meta-transfer learning model using some labeled data also includes: The architecture of the Transformer model is optimized through model pruning and knowledge distillation to reduce model complexity.

8. A UAV acoustic signal recognition system based on meta-transfer learning, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method for recognizing unmanned aerial vehicle acoustic signals based on meta-transfer learning as described in any one of claims 1 to 7 are implemented.

Images