Acoustic feature recognition and trajectory prediction method for low, small and slow target

Abstract

The invention relates to the technical field of acoustic sensing, in particular to an acoustic feature recognition and trajectory prediction method for low, small and slow targets, which comprises the following steps: a low, small and slow target positioning and number solving step: adopting a MUSIC algorithm to cooperate with a DBSCAN clustering algorithm to judge the number of information sources and carry out position solving; and a low, small and slow target tracking and trajectory prediction step: designing a 3D-GAA architecture model based on the GRU network in combination with acceleration, position and speed information, and performing tracking and trajectory prediction through the 3D-GAA architecture model. The method shows the advantages in the aspects of accuracy and robustness, and particularly has higher effectiveness in the aspect of capturing complex dynamic and data modes.

Description

Technical Field

[0001] The present application relates to the field of acoustic sensing technology, and in particular to a method for acoustic feature recognition and trajectory prediction of low, small and slow targets. Background Art

[0002] In recent years, with the rapid development of low, small and slow targets (such as drones) in the military, security and civil fields, the perception and recognition technology of such targets has attracted widespread attention. With their flexibility, low cost and high efficiency, drones are widely used in reconnaissance, monitoring, logistics and emergency rescue. However, low-altitude, slow and small drone targets also bring unprecedented technical challenges, especially in perception, recognition and trajectory prediction.

[0003] To meet this challenge, drone perception technology based on acoustic features has gradually attracted attention in recent years. Compared with radar and optical sensors, acoustic sensing technology can monitor drones without being affected by weather conditions and visual occlusion. For the recognition of drone acoustic features, Anwar et al. (2020) proposed a new machine learning framework, which used multiple machine learning methods to detect drone sounds and significantly improved the recognition accuracy. Uddin et al. (2021) proposed a method to extract and identify drone sounds from interference signals based on the independent component analysis (ICA) algorithm, which effectively solved the problem that drone sound signals are susceptible to interference in complex environments. These studies have provided a solid foundation for the perception of drone acoustic features, but the robustness and accuracy of existing methods are still insufficient in the face of multi-source acoustic signals, complex environmental interference, etc. Summary of the invention

[0004] The present application aims to solve at least one of the above-mentioned technical problems in the prior art to a certain extent. To this end, the embodiment of the present application provides a method for acoustic feature recognition and trajectory prediction of low, small and slow targets. This method has a higher accuracy rate than traditional models in model recognition of low, small and slow targets. The mean absolute percentage error of the 3D-GAA model in the present application is significantly reduced, showing advantages in accuracy and robustness, especially in capturing complex dynamics and data patterns. It has stronger effectiveness.

[0005] A method for acoustic feature recognition and trajectory prediction of a low, small and slow target comprises the following steps: The steps for locating and calculating the number of low, small and slow targets are as follows: the MUSIC algorithm is used in conjunction with the DBSCAN clustering algorithm to determine the number of signal sources and calculate their positions; Tracking and trajectory prediction steps for low, small and slow targets: Based on the GRU network, a 3D-GAA architecture model is designed by combining acceleration, position and velocity information, and tracking and trajectory prediction are performed through the 3D-GAA architecture model.

[0006] The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 1 is characterized in that it also includes the steps of model recognition and load analysis of low, small and slow targets: using EfficientNet and GRU to build a multi-task learning framework model, and combining the gray wolf optimization algorithm to adjust the weight of the loss function.

[0007] In an optional or preferred embodiment, the input of the multi-task learning framework model adopts a multi-image input method, specifically including three images of MFCC, STFT and time domain spectrum. These images are stacked together along a specified dimension to form a multi-channel image input sample.

[0008] In an optional or preferred embodiment, the step of weighting the loss function by the Grey Wolf Optimization Algorithm includes: Step 1: Initialize the population, randomly initialize a group of gray wolf individuals, each individual represents a weight W, and calculate the Loss corresponding to each individual reg and Loss cla ; Step 2: Set the maximum number of iterations and minimum error; Step 3: Iterative process, calculate the Loss of each individual gray wolf reg and Loss cla , according to Loss reg and Loss cla Update the position of each individual using the iterative formula W in the gray wolf optimization algorithm new =W old +A×D, where W new is the new weight value, A is a random vector, D is some combination of Alpha, Beta, Delta gray wolf positions, calculate the updated Loss reg and Loss cla .

[0009] Step 4: Check the termination condition. If the termination condition is met (the maximum number of iterations is reached or the minimum error is reached), stop the iteration, otherwise return to step 3.

[0010] Step 5: Result analysis: According to the final iteration results, select the one with the minimum Loss reg and Loss cla individuals to obtain the best weights.

[0011] In an optional or preferred embodiment, the steps of using the MUSIC algorithm in conjunction with the DBSCAN clustering algorithm include: Step 1: Let the received data be only noise data and perform eigenvalue decomposition on the covariance matrix; Step 2: Analyze the noise characteristic values ​​and give the REps and Pmin Pts parameter values ​​of the DBSCAN algorithm; Step 3: Perform eigenvalue decomposition on the covariance matrix containing the echo signal data to obtain M eigenvalues ​​and eigenvectors; Step 4: Perform DBSCAN clustering on the eigenvalues ​​to obtain the number of information sources i.

[0012] In an optional or preferred embodiment, the design of the 3D-GAA architecture model includes: each target is represented as a 3D coordinate (xt, yt, zt) at time t, and the initial sequence of these coordinates is provided by the MUSIC algorithm plus binocular intersection. The target of the study predicts the position of t=To+Tp based on the observed target trajectory from t=1 to t=1+To, that is, through Xo=[(x1,y1,z1),..,(xto,yto,zto)] to predict and generate Xp=[(xto,yto,zto),..,(xtp,ytp,ztp)], from the observed trajectory, once Backward difference can obtain velocity information Vs=[(vx1,vy1,vz1),...,(vxo,vyo,vzo)], and two backward differences are performed to calculate the finite difference in time step to obtain acceleration information As=[(ax1,ay1,az1),...,(axo,ayo,azo)], and finally copy the first velocity (vx1,vy1,vz1) and acceleration (ax1,ay1,az1), so that X, V and A have the same amount of time. The GAA model can learn motion patterns that are independent of position by processing position, velocity and acceleration information in parallel.

[0013] In an optional or preferred embodiment, the GAA model includes an adjustment module, which dynamically allocates weights through an attention mechanism to balance the relationship between position, velocity, and acceleration information to achieve accuracy and consistency of the model when performing trajectory prediction.

[0014] In an optional or preferred embodiment, the adjustment module includes an attention layer, a softmax layer and an adjustment layer.

[0015] In an optional or preferred embodiment, OpenSky's AircraftMetadata Database of OpenSky Network is used to compare prediction results.

[0016] In an optional or preferred embodiment, the aircraft longitude and latitude provided by the database, the geometric height based on GPS are taken as the x, y, and z of the input layer, and the drone speed is taken as the first-order difference training indicator, and KF filtering, double-layer LSTM and 3D-GAA networks are run respectively for comparative analysis.

[0017] Based on the above technical solution, the embodiments of the present application have at least the following beneficial effects: the present application adopts the MUSIC algorithm combined with the DBSCAN clustering algorithm to achieve the precise positioning of low, small and slow targets and the determination of the number of signal sources. Finally, tracking and trajectory prediction are performed through the 3D-GAA architecture model that integrates position, velocity and acceleration information.

[0018] This method improves the accuracy of model recognition of low, small and slow targets by about 30% compared with traditional CNN and SVM models, meeting the engineering requirement of 70%. At the same time, compared with the average absolute percentage error of 248.05 of traditional methods, the average absolute percentage error of the 3D-GAA model in this application is significantly reduced to 10.89, showing advantages in accuracy and robustness, especially in capturing complex dynamics and data patterns. It has stronger effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present application is further described below with reference to the accompanying drawings and embodiments; Figure 1 It is a structural diagram of the basic unit of GRU in this application; Figure 2 is a schematic diagram comparing the classification capabilities of the ESC-50 public model in this application; Figure 3 It is the BDSCAN classification effect chart in this application; Figure 4 This is a graph showing the error in the number of sources determined by the DBSCAN algorithm in this application; Figure 5 It is a schematic diagram of the design of the 3D-GAA trajectory prediction and tracking algorithm in this application. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0021] The following is a further detailed description of the implementation of the present application in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application but cannot be used to limit the scope of the present application.

[0022] In the description of the embodiments of the present application, it should be noted that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the embodiments of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the embodiments of the present application. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0023] In the description of the embodiments of the present application, it should be noted that, unless otherwise clearly specified and limited, the terms "connected" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the embodiments of the present application can be understood according to specific circumstances.

[0024] In the embodiments of the present application, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" and "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0025] In recent years, with the rapid development of low, small and slow targets (such as drones) in the military, security and civil fields, the perception and recognition technology of such targets has attracted widespread attention. With their flexibility, low cost and high efficiency, drones are widely used in reconnaissance, monitoring, logistics and emergency rescue. However, low-altitude, slow and small drone targets also bring unprecedented technical challenges, especially in perception, recognition and trajectory prediction.

[0026] To meet this challenge, drone perception technology based on acoustic features has gradually attracted attention in recent years. Compared with radar and optical sensors, acoustic sensing technology can monitor drones without being affected by weather conditions and visual occlusion. For the recognition of drone acoustic features, Anwar et al. (2020) proposed a new machine learning framework, which used multiple machine learning methods to detect drone sounds and significantly improved the recognition accuracy. Uddin et al. (2021) proposed a method to extract and identify drone sounds from interference signals based on the independent component analysis (ICA) algorithm, which effectively solved the problem that drone sound signals are susceptible to interference in complex environments. These studies have provided a solid foundation for the perception of drone acoustic features, but the robustness and accuracy of existing methods are still insufficient in the face of multi-source acoustic signals, complex environmental interference, etc.

[0027] The following is a detailed explanation using a drone as an example of a low, small and slow target.

[0028] The present application provides a method for acoustic feature recognition and trajectory prediction of low, small and slow targets, comprising the following steps: The steps for locating and calculating the number of low, small and slow targets are as follows: the MUSIC algorithm is used in conjunction with the DBSCAN clustering algorithm to determine the number of signal sources and calculate their positions; Tracking and trajectory prediction steps for low, small and slow targets: Based on the GRU network, a 3D-GAA architecture model is designed by combining acceleration, position and velocity information, and tracking and trajectory prediction are performed through the 3D-GAA architecture model.

[0029] This application adopts the MUSIC algorithm combined with the DBSCAN clustering algorithm to achieve the precise positioning of the drone and the determination of the number of signal sources. Finally, tracking and trajectory prediction are performed through the 3D-GAA architecture model that integrates position, velocity and acceleration information.

[0030] This method improves the accuracy of drone model recognition by about 30% compared with traditional CNN and SVM models, meeting the engineering requirement of 70%. At the same time, compared with the average absolute percentage error of 248.05 of traditional methods, the average absolute percentage error of the 3D-GAA architecture model in this application is significantly reduced to 10.89, showing advantages in accuracy and robustness, especially in capturing complex dynamics and data patterns.

[0031] In some embodiments, the model recognition and load analysis steps of low, small and slow targets are also included: EfficientNet and GRU are used to build a multi-task learning framework model, and the gray wolf optimization algorithm is used to adjust the weight of the loss function. By introducing the gray wolf optimization algorithm to adjust the weight of the loss function, the recognition accuracy of the model is further improved.

[0032] In some embodiments, the input of the multi-task learning framework model adopts a multi-image input method, specifically including three images of MFCC, STFT and time domain spectrum. These images are stacked together along a specified dimension to form a multi-channel image input sample.

[0033] like Figure 1 The model input adopts a multi-image input method, which includes three images: MFCC, STFT and time domain spectrum. These images are stacked together along the specified dimension to form a multi-channel image input sample. The purpose of this design is to enrich the model's ability to capture the characteristics of sound signals and thus improve recognition accuracy.

[0034] Specifically, the input of this application adopts a multi-image (MFCC, STFT and time domain spectrum) input method. This design choice is based on the following considerations: Richer information: Image input is rich in information, easy to process, and has strong compatibility, which helps to fully capture the characteristics of sound signals.

[0035] Easy to process: Image data can be processed with the help of image processing techniques, such as convolutional neural networks (CNN). These technologies are very mature in processing image data, and there are a variety of pre-trained models available.

[0036] Generalization ability: Image data usually performs well in the generalization ability of prediction models because image data has certain similarities and structures with other image data.

[0037] The code stacks three images together along the specified dimension to form a multi-channel image and learn at once. Its advantages include: (1) Simplified data processing: By stacking multiple images together as one input sample, the data processing process can be simplified, especially when processing multi-channel data or needing to consider multiple frames of data at the same time.

[0038] (2) Effective use of neural network architecture: Combining multiple images into a multi-channel input can enable the neural network to more effectively learn the correlation between images and improve the generalization and expression capabilities of the model.

[0039] (3) Reduce memory consumption: By stacking images, the memory consumption required to store and process a large number of independent images can be reduced, thereby improving the efficiency of processing large-scale data sets.

[0040] Figure 2The two key components, EfficientNet-v2 and LSTM, are combined to form a deep learning network consisting of a series of convolutional layers, pooling layers, fully connected layers, GRU layers, and fully connected layers for regression and classification. It is designed to perform drone model recognition and load prediction through audio data.

[0041] (1) Selection of Convolutional Neural Network As a relatively lightweight convolutional neural network (CNN), EfficientNet-v2s excels in real-time performance and outperforms larger neural network models. The core concept behind its design includes uniform scaling of model depth, width, and resolution to achieve enhanced computational efficiency and accuracy. This design enables EfficientNet to provide performance comparable to or even exceeding that of larger models with relatively few parameters.

[0042] EfficientNet-V2, represents a breakthrough in convolutional network architectures. This family of models outperforms its predecessors in terms of both training speed and parameter efficiency. The development of EfficientNet-V2 involves an approach that jointly optimizes training-aware neural architecture search and scaling to jointly optimize training speed and parameter efficiency. This family of models, searched from a rich search space that includes innovative designs such as fused MBConv, exhibits significantly faster training times—up to 6.8 times faster than state-of-the-art models. In the context of this work, EfficientNet-v2s extract spatial features from acquired spectrograms, capturing implicit feature information by gradually reducing the dimensionality. The features obtained from this process are then passed to a fully connected layer that helps in further extracting and fusing high-level features. This prepares the input to the Gated Recurrent Unit (GRU) layer, which helps in the overall efficiency and accuracy of the analysis.

[0043] The multi-task learning framework design is based on EfficientNet and GRU, and is mainly used for drone model recognition and load analysis.

[0044] In terms of network structure, EfficientNet-v2, as the core of convolutional neural networks, has demonstrated outstanding performance in real-time performance and parameter efficiency. EfficientNet-v2 extracts spatial features from the spectrum graph and gradually reduces the dimension to capture more implicit feature information, and then passes these features to the fully connected layer to further extract and fuse high-level features. At the same time, the introduction of the GRU layer enables the network to process temporal information, especially to deeply mine the dynamic features in audio data. The output of the GRU layer is processed by an additional fully connected layer to produce a feature representation suitable for regression and classification tasks. The regression output is used to predict the drone load information, and the classification output is processed by the sigmoid function to obtain the classification probability of the drone model.

[0045] In some embodiments, the step of weighting the loss function by the gray wolf optimization algorithm includes: Step 1: Initialize the population, randomly initialize a group of gray wolf individuals, each individual represents a weight W, and calculate the Loss corresponding to each individual reg and Loss cla ; Step 2: Set the maximum number of iterations and minimum error; Step 3: Iterative process, calculate the Loss of each individual gray wolf reg and Loss cla , according to Loss reg and Loss cla Update the position of each individual using the iterative formula W in the gray wolf optimization algorithm new =W old +A×D, where W new is the new weight value, A is a random vector, D is some combination of Alpha, Beta, Delta gray wolf positions, calculate the updated Loss reg and Loss cla .

[0046] Step 4: Check the termination condition. If the termination condition is met (the maximum number of iterations is reached or the minimum error is reached), stop the iteration, otherwise return to step 3.

[0047] Step 5: Result analysis: According to the final iteration results, select the one with the minimum Loss reg and Loss cla individuals to obtain the best weights.

[0048] Among them, the gray wolf optimization algorithm is an optimization search method developed by the gray wolf hunting. The gray wolf has a strict social hierarchy system, from high to low, namely Alpha, Beta, Delta, and the hunting process is completed under the leadership of Alpha. The hunting of gray wolves includes tracking prey, surrounding prey and attacking prey.

[0049] Loss reg is the absolute error value of the load of the small and slow target, Loss cla The model misjudgment rate of small and slow targets is low.

[0050] The number determination and location solution of information sources mainly use high-resolution spectrum estimation algorithm, conventional beamforming algorithm and arrival time difference algorithm. The MUSIC algorithm is widely used in the number determination and location solution of information sources. By analyzing the orthogonality of the signal subspace and the noise subspace, it can accurately estimate the number and direction of signal sources. In this application, the steps of using the MUSIC algorithm in conjunction with the DBSCAN clustering algorithm include: Step 1: Let the received data be only noise data and perform eigenvalue decomposition on the covariance matrix; Step 2: Analyze the noise characteristic values ​​and give the REps and Pmin Pts parameter values ​​of the DBSCAN algorithm; Step 3: Perform eigenvalue decomposition on the covariance matrix containing the echo signal data to obtain M eigenvalues ​​and eigenvectors; Step 4: Perform DBSCAN clustering on the eigenvalues ​​to obtain the number of information sources i.

[0051] The DBSCAN clustering algorithm plays a crucial role in determining the number of signal sources. In the experiment, by clustering the characteristic values ​​of 1 to 15 signal sources, the DBSCAN algorithm shows extremely high accuracy when the number of signal sources is small.

[0052] like Figure 4 As shown in the data, as the number of actual sources increases, the average error of the DBSCAN algorithm gradually increases, especially when the number of sources exceeds four, the error increases significantly. This shows that the accuracy of the DBSCAN algorithm has decreased when dealing with multi-source clustering. This is because the interference and complexity between signals increase, which makes eigenvalue clustering more difficult.

[0053] Figure 3 and Figure 4 The experimental results in show that when the number of signal sources is between 1 and 4, the signal source judgment error of the combination of MUSIC algorithm and DBSCAN is small and the positioning accuracy is high. Especially in the scenario where multiple types of signal sources coexist, this combined method shows strong adaptability and accuracy, providing reliable data support for subsequent drone positioning and tracking.

[0054] Reference Figure 5 In some embodiments, the design of the 3D-GAA architecture model includes: each target is represented as a 3D coordinate (xt, yt, zt) at time t, and the initial sequence of these coordinates is provided by the MUSIC algorithm plus binocular intersection. The target of the study predicts the position of t=To+Tp based on the observed target trajectory from t=1 to t=1+To, that is, through Xo=[(x1, y1, z1), .., (xto, yto, zto)] to predict and generate Xp=[(xto, yto, zto), .., (xtp, ytp, ztp)], from the observed trajectory, one backward The difference can get the velocity information Vs=[(vx1,vy1,vz1),...,(vxo,vyo,vzo)], and the backward difference is performed twice to calculate the finite difference in the time step to obtain the acceleration information As=[(ax1,ay1,az1),...,(axo,ayo,azo)], and finally copy the first velocity (vx1,vy1,vz1) and acceleration (ax1,ay1,az1) to get the same amount of time for X, V and A. The GAA model can learn motion patterns that are independent of position by processing position, velocity and acceleration information in parallel.

[0055] The GAA model can learn position-independent motion patterns by processing position, velocity, and acceleration information in parallel, and further improves the accuracy and robustness of trajectory prediction by combining scene-related features.

[0056] Furthermore, the GAA model includes an adjustment module, which dynamically assigns weights through an attention mechanism to balance the relationship between position, velocity, and acceleration information, thereby ensuring the accuracy and consistency of the model when making trajectory predictions.

[0057] The adjustment module consists of an attention layer, a softmax layer and an adjustment layer. This design enables the network to reasonably allocate the weights between various physical quantities based on the input time series data, thereby improving the accuracy of the prediction results.

[0058] In some embodiments, OpenSky's Aircraft Metadata Database of OpenSky Network is used to compare prediction results.

[0059] In some embodiments, the aircraft longitude and latitude provided by the database, the geometric height based on GPS are taken as the x, y, and z of the input layer, and the drone speed is taken as the first-order difference training indicator, and KF filtering, double-layer LSTM and 3D-GAA networks are run respectively for comparative analysis.

[0060] Specifically, the OpenSky Network's OpenSky's Aircraft Metadata Database is used. Its weekly 24-hour state vector dataset provides comprehensive and detailed flight data, which can be effectively used to verify and improve the accuracy of UAV tracking algorithms. These data not only enrich the model's training set, but also provide the necessary verification data to evaluate the performance of the algorithm in practical applications.

[0061] Here, the aircraft longitude, latitude, and GPS-based geometric height provided by the database are taken as x, y, and z of the input layer, respectively, and the drone speed is taken as the first-order difference training indicator. KF filtering, double-layer LSTM, and 3D-GAA networks are run respectively for comparative analysis.

[0062] The results of KF filtering are shown in Table 1. The results show that the RMSE and MAE performances are relatively low. KF filtering performs well in processing linear dynamic systems and noisy data, and can better capture the dynamic changes of the system. It also shows that KF can provide more accurate predictions in some cases. However, the MAPE value is high, indicating that the model performs poorly at the inflection point of some data points and may not be able to effectively capture nonlinear and complex dynamics in some cases.

[0063] Table 1: The results of the two-layer LSTM are shown in Table 2 Table 2: The results of 3D-GAA are shown in Table 3 Table 3: It can be seen from Table 3 that the 3D-GAA architecture model performs better than the KF filter in all error indicators, especially in MAPE. The relative error of the custom model is significantly lower, indicating that it is more effective in capturing data patterns. The lower RMSE and MAE indicate that the custom model is more accurate and stable in predicting data.

[0064] This application uses the drone audio data set as a signal and a random background noise in ESC-50 at a ratio of 20:1 as a new audio. All load parameters are set to 1. During training, the classification and regression weight ratio is adjusted to 5000:1 through the adjustment of the above loss function, so that the network almost only classifies and ignores the regression error. The classification ability of the ESC-50 public model is compared. Figure 2 shown.

[0065] Depend on Figure 2 It can be seen that the classification accuracy of the model is about 30% higher than that of ordinary CNN, SVM and other native models, and the accuracy meets the 70% requirement given by the project.

[0066] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiments of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0067] The embodiments of the present application are described in detail above in conjunction with the accompanying drawings, but the present application is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in the technical field without departing from the purpose of the present application.

Claims

1. A method for acoustic feature recognition and trajectory prediction of low, small and slow targets, characterized in that: The following steps are involved: The steps for locating and calculating the number of low, small and slow targets are as follows: the MUSIC algorithm is used in conjunction with the DBSCAN clustering algorithm to determine the number of signal sources and calculate their positions; Tracking and trajectory prediction steps for low, small and slow targets: Based on the GRU network, a 3D-GAA architecture model is designed by combining acceleration, position and velocity information, and tracking and trajectory prediction are performed through the 3D-GAA architecture model.

2. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 1 is characterized by: It also includes the model identification and load analysis steps of low, small and slow targets: using EfficientNet and GRU to build a multi-task learning framework model, and combining the Gray Wolf Optimization Algorithm to adjust the weight of the loss function.

3. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 2 is characterized by: The input of the multi-task learning framework model adopts a multi-image input method, specifically including three images of MFCC, STFT and time domain spectrum. These images are stacked together along a specified dimension to form a multi-channel image input sample.

4. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 2 is characterized by: The steps of the Gray Wolf Optimization Algorithm to adjust the weight of the loss function include: Step 1: Initialize the population, randomly initialize a group of gray wolf individuals, each individual represents a weight W, and calculate the Loss corresponding to each individual reg and Loss cla ; Step 2: Set the maximum number of iterations and minimum error; Step 3: Iterative process, calculate the Loss of each individual gray wolf reg and Loss cla , according to Loss reg and Loss cla Update the position of each individual using the iterative formula W in the gray wolf optimization algorithm new =W old +A×D, where W new is the new weight value, A is a random vector, D is some combination of Alpha, Beta, Delta gray wolf positions, calculate the updated Loss reg and Loss cla ; Step 4: Check the termination condition. If the termination condition is met, stop the iteration, otherwise return to step 3; Step 5: Result analysis: According to the final iteration results, select the one with the minimum Loss reg and Loss cla individuals to obtain the best weights.

5. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 1 is characterized by: The steps of using the MUSIC algorithm in conjunction with the DBSCAN clustering algorithm include: Step 1: Let the received data be only noise data and perform eigenvalue decomposition on the covariance matrix; Step 2: Analyze the noise characteristic values ​​and give the REps and Pmin Pts parameter values ​​of the DBSCAN algorithm; Step 3: Perform eigenvalue decomposition on the covariance matrix containing the echo signal data to obtain M eigenvalues ​​and eigenvectors; Step 4: Perform DBSCAN clustering on the eigenvalues ​​to obtain the number of information sources i.

6. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 1 is characterized by: The design of the 3D-GAA architecture model includes: each target is represented as a 3D coordinate (xt, yt, zt) at time t. The initial sequence of these coordinates is provided by the MUSIC algorithm plus binocular intersection. The goal of the study is to predict the position of t=To+Tp based on the observed target trajectory from t=1 to t=1+To, that is, to predict through Xo=[(x1,y1,z1),..,(xto,yto,zto)], and generate Xp=[(xto,yto,zto),..,(xtp,ytp,ztp)], from the observed trajectory, a backward difference can be Get the velocity information Vs=[(vx1,vy1,vz1),...,(vxo,vyo,vzo)], perform two backward differences, calculate the finite difference in the time step to get the acceleration information As=[(ax1,ay1,az1),...,(axo,ayo,azo)], and finally copy the first velocity (vx1,vy1,vz1) and acceleration (ax1,ay1,az1), so that X, V and A have the same amount of time. The GAA model can learn motion patterns that are independent of position by processing position, velocity and acceleration information in parallel.

7. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 6 is characterized by: The GAA model includes an adjustment module, which dynamically assigns weights through an attention mechanism to balance the relationship between position, velocity, and acceleration information to achieve accuracy and consistency in trajectory prediction.

8. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 7 is characterized by: The adjustment module includes an attention layer, a softmax layer and an adjustment layer.

9. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 1, characterized in that: The prediction results are compared using OpenSky's Aircraft Metadata Database from OpenSky Network.

10. The method for acoustic feature recognition and trajectory prediction of low, small and slow targets according to claim 7, characterized in that: The aircraft longitude and latitude provided by the database are taken, and the geometric height based on GPS is used as the x, y, and z of the input layer. The drone speed is used as the first-order difference training indicator. KF filtering, double-layer LSTM and 3D-GAA networks are run respectively for comparative analysis.

Images