Unmanned aerial vehicle model identification method and equipment based on image transmission signal, and medium
By performing aliasing sampling, density calculation, and data cleaning on the UAV image transmission signal, and combining it with a neural network model, high-precision identification of UAV models was achieved. This solved the problem of complex and low-precision identification in existing technologies, reduced equipment power consumption, and enhanced the identification effect.
Patent Information
- Application Number
- CN202411010486.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-12
AI Technical Summary
Existing methods for identifying drone models are complex and their accuracy is not ideal, and they are greatly affected by the selection of detection thresholds.
By acquiring the initial image transmission signal from the UAV, aliasing sampling and density calculation are performed to generate an energy density spectrum. Data cleaning and outlier filtering are then carried out, and a pre-trained neural network model is used for classification.
It improves the accuracy of drone model identification, reduces equipment power consumption, saves computing resources, enhances the representation of image transmission features, and suppresses interference from WIFI signals.
Smart Images

Figure CN118939979A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of drone management, and in particular to a method, device and medium for drone model identification based on image transmission signals. Background Art
[0002] In recent years, drones have been widely used due to their advantages of miniaturization, lightness and intelligence. As the application scope of drones gradually expands, it is necessary to detect and identify drones, with the focus on identifying the models of drones. In the existing drone model recognition methods, artificial features need to be designed, the recognition process is relatively complicated, and the recognition effect is greatly affected by the selection of detection thresholds. The recognition accuracy of the current drone model recognition methods is not ideal. Summary of the invention
[0003] The purpose of this application is to provide a method, device and medium for drone model identification based on image transmission signals, which can improve the recognition accuracy of drone models.
[0004] To achieve the above objectives, this application provides the following solutions:
[0005] In a first aspect, the present application provides a method for identifying a drone model based on an image transmission signal, comprising:
[0006] Obtain the initial image transmission signal of the drone to be identified;
[0007] Performing aliasing sampling on the initial image transmission signal to obtain an aliased image transmission signal;
[0008] Performing density calculation on the aliased image transmission signal to obtain an energy density spectrum;
[0009] Performing data cleaning and outlier filtering on the energy density spectrum to obtain an image transmission feature vector;
[0010] A pre-trained neural network model is used to classify the image transmission feature vector to determine the model of the drone to be identified; the neural network model is trained using a training sample set, and the training sample set includes multiple sample image transmission feature vectors and the drone model corresponding to each sample image transmission feature vector.
[0011] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned method for identifying drone models based on image transmission signals.
[0012] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for identifying drone models based on image transmission signals.
[0013] According to the specific embodiments provided in this application, this application discloses the following technical effects:
[0014] The present application provides a method, device and medium for identifying the model of a UAV based on an image transmission signal. By performing aliasing sampling on the initial image transmission signal of the UAV to be identified, the power consumption of the device is reduced and computing resources are saved. By performing density calculation on the aliased image transmission signal, the image transmission characteristics can be well reflected. By performing data cleaning and outlier filtering on the energy density spectrum, the WIFI signal can be suppressed and the classification and recognition accuracy of the neural network model can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0016] Figure 1 This is a schematic diagram of the time domain image transmission signal of the DJI FPV 5.8GHz drone;
[0017] Figure 2 This is a schematic diagram of the frequency domain image transmission signal of the DJI FPV 5.8GHz drone;
[0018] Figure 3 This is a schematic diagram of the time domain image transmission signal of the DJI Air2S 5.8GHZ drone;
[0019] Figure 4 This is a schematic diagram of the frequency domain image transmission signal of the DJI Air2S 5.8GHZ drone;
[0020] Figure 5 A schematic diagram of a flow chart of a method for identifying a drone model based on image transmission signals provided in one embodiment of the present application;
[0021] Figure 6 It is a schematic diagram of an image transmission signal before aliasing;
[0022] Figure 7 It is a schematic diagram of the image transmission signal after aliasing;
[0023] Figure 8 The effect of using the energy density method to remove jitter from the image transmission signal of the DJI Phantom4 Pro drone;
[0024] Fig. 9 The effect of using the energy density method to remove jitter from the image transmission signal of the DJI MINI3 drone;
[0025] Fig.10 The effect of using the energy density method to remove jitter from the image transmission signal of the EV0 PRO V3 drone;
[0026] Fig.11 The effect of using the energy density method to remove jitter from the image transmission signal of the DJI Air2S drone;
[0027] Fig.12 This is the energy spectrum of the DJIMINI3 drone;
[0028] Fig.13 The effect of removing jitter using the energy density method (window density is 10) for the image transmission signal of the DJI Air2S drone;
[0029] Fig.14 The effect of removing jitter from the image transmission signal of DJI Air2S UAV using the energy density method (window density is 100);
[0030] Fig.15 The effect of removing jitter from the image transmission signal of DJAir2S drone using the energy density method (window density is 1000);
[0031] Fig.16 The effect of removing jitter from the image transmission signal of DJAir2S drone using the energy density method (window density is 3000);
[0032] Fig.17 The flowchart of data cleaning is shown below;
[0033] Fig.18 It is a performance diagram of the binary classification model;
[0034] Fig.19 Schematic diagram of the performance degradation of the neural network model;
[0035] Fig. 20 It is a schematic diagram of a strong WIFI signal;
[0036] Fig.21 This is a schematic diagram of the captured DJIMIMI3 pro drone outliers;
[0037] Fig. 22 To remove Fig.21 Schematic diagram after the outlier;
[0038] Fig.23 A schematic diagram of capturing the outliers of the DJIAVATA drone;
[0039] Fig.24 To remove Fig.23 Schematic diagram after the outlier;
[0040] Fig.25 A schematic diagram to capture the outliers of a certain drone;
[0041] Fig.26 To remove Fig.25 Schematic diagram after the outlier;
[0042] Fig. 27 This is a schematic diagram of the outliers that are not captured;
[0043] Fig.28 This is a schematic diagram after filtering without outliers;
[0044] Fig.29 The figure is a schematic diagram showing the discrimination of four types of aircraft by the average interval between the peak functions of the energy density spectrum;
[0045] Fig.30 This is a schematic diagram of the discrimination between the average width of the peak function of the energy density spectrum and the four models;
[0046] Fig.31 It is a schematic diagram of the discrimination degree of the average energy density of the energy density spectrum for the four models;
[0047] Fig.32 This is a schematic diagram of the discrimination between the average variance of the peak function of the energy density spectrum and the four models;
[0048] Fig.33 It is a schematic diagram of the discrimination degree of the average third-order moment of the peak function of the energy density spectrum for four types of aircraft;
[0049] Fig.34 It is a schematic diagram of the discrimination degree of the four models by the average 4th order cumulative amount of the peak function of the energy density spectrum;
[0050] Fig.35 It is a schematic diagram of the Pearson coefficient matrix;
[0051] Fig.36 This is a diagram showing the effect of evaluating the clustering model based on the mean deviation method;
[0052] Fig.37 This is a diagram to evaluate the effect of clustering model based on the sum of squared errors within clusters;
[0053] Fig.38 This is a diagram to evaluate the effect of clustering model based on silhouette coefficient;
[0054] Fig.39 Schematic diagram of clustering effect;
[0055] Fig.40 Schematic diagram of the training performance of a multi-layer perceptron. DETAILED DESCRIPTION
[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions 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 this field without creative work are within the scope of protection of this application.
[0057] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0058] UAV image transmission technology can be divided into analog image transmission and digital image transmission. Analog image transmission refers to the source and channel processing of analog image signals with continuously changing time and amplitude, and the process of analog channel transmission. Its transmitter is equivalent to broadcasting. As long as the frequency of the receiving end is consistent with that of the transmitting end, the video signal can be received. Its advantages are low price and low latency; its disadvantages are poor image quality, low resolution, and the signal is easily interfered, resulting in serious snow and horizontal ripples on the screen. Digital image transmission refers to the process of transmitting digitized image signals through digital channels after source coding and channel coding. The transmitting end of digital image transmission needs to encode and the receiving end needs to decode. Due to the different digital signal processing technologies of various manufacturers, the sending end and receiving end of digital image transmission need to be used in conjunction, resulting in poor versatility. The principle of digital image transmission determines that it has strong anti-interference ability and clearer image quality, but the delay will be slightly higher.
[0059] In recent years, new drones are mainly based on digital image transmission technology. Take DJI drones as an example. Their drones mainly use OcuSync technology for communication. OcuSync technology provides dual-frequency working mode, and the frequency band can be selected manually or automatically. Each frequency band provides two bandwidth signals to choose from. 10MHz or 20MHz can be selected on the remote control. The image transmission signal with 10MHz bandwidth has better anti-interference performance, while the image transmission signal with 20MHz bandwidth can transmit more data and display clearer pictures. The image transmission signal form of OcuSync technology is as follows: Figures 1 to 4 As shown, it can be seen that the image transmission signal of OcuSync technology does not have a fixed duration or transmission interval in the time domain, but the duty cycle of the image transmission signal is relatively high over a long period of time. In the frequency domain, the image transmission signal strength is significantly higher than other signals, and its bandwidth is about 9MHz or 18MHz.
[0060] By collecting the indoor drone image transmission signal at a fixed frequency and then performing time-frequency spectrum analysis, it can be seen that the electromagnetic environment is very complex in the indoor environment, and there are many irrelevant signals with very strong intensity. Therefore, such signals should be removed, and the bandwidth of the image transmission is relatively large compared to WIFI and other signals, so this information should be utilized to some extent.
[0061] In an exemplary embodiment, Figure 5 As shown, a method for identifying the model of a drone based on an image transmission signal is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiment of the present application, the method for identifying the model of a drone based on an image transmission signal includes the following steps 501 to 505. Among them:
[0062] Step 501: Acquire the initial image transmission signal of the drone to be identified.
[0063] Step 502: perform aliasing sampling on the initial image transmission signal to obtain an aliased image transmission signal.
[0064] Specifically, according to the bandwidth and frequency domain duty cycle of the drone to be identified, the initial image transmission signal is extracted without filtering, so that the image transmission signal of the drone to be identified covers the entire frequency spectrum to obtain an aliased image transmission signal.
[0065] Aliasing is generally a spectrum aliasing error caused by insufficient sampling rate, which is an error that should be avoided as much as possible in signal processing. However, in some cases, intentional signal aliasing can achieve the function of spectrum accumulation. This application takes the Zynq7070+AD9361 development board as an acquisition device as an example. The maximum acquisition bandwidth of the AD9361 chip is 56M, the sampling rate is 61.44M, and the bandwidth of the drone is 10M. Its duty cycle in the frequency domain is 10 / 61.44=0.16. If the acquired signal is not filtered and extracted 8 times, the sampling rate becomes 7.68M, of which 2.5M of the image transmission signal is aliased to 7.68M, which can enhance the relative strength of the signal. Figure 6 and Figure 7 As shown, the image transmission signal will cover the entire spectrum, and then the accumulation algorithm can be used to utilize the bandwidth of the image transmission signal.
[0066] In addition, reducing the sampling rate from 61.44M to 7.68M is helpful to reduce power consumption and save resources.
[0067] Step 503: perform density calculation on the aliased image transmission signal to obtain an energy density spectrum.
[0068] Specifically, the energy of the aliased image transmission signal is first calculated to obtain a preliminary energy spectrum. Then, the density of the preliminary energy spectrum is calculated to obtain an energy density spectrum. The energy spectrum features of different models are different, which helps to identify the model by using this feature. Figures 8 to 11 The energy spectra of four types of drone image transmission signals are listed.
[0069] Among them, the energy operation is to find the norm, that is, to take the modulus of the complex number. The window function for density calculation is 512, and the average energy sum of 512 points is used to replace these 512 points. In practical applications, the energy sum of 512 points is used to replace the average energy. In actual tests, it is found that the window function of 512 points can already meet the needs. In addition, its energy density can be replaced in a simpler way. For example, the accumulated value of 512 points can be used to represent the total energy in the time domain. In this case, the fluctuation can be better removed, and each 512 points can be mapped to one point, for example, 8192*20 points can be mapped to 320 points, and 8192*40 points can be mapped to 640 points.
[0070] In an exemplary embodiment, the energy operation is first performed on the 7.68M data. At this time, the time domain signal has many burrs and cannot well reflect the characteristics of the image transmission. Fig.12 As shown. Then the result of the energy operation is used for density calculation to highlight the characteristics of the model. According to experiments, the window function of density calculation is 1000, which has the best cost performance. Figures 13 to 16 shown.
[0071] Step 504: perform data cleaning and outlier filtering on the energy density spectrum to obtain an image transmission feature vector.
[0072] When training a neural network model, you can first use a smaller data set to train a binary classification model, then use the binary classification model to annotate a small-scale data set, and then train a multi-machine model after manual review. Then use the model prototype to build a larger-scale data set, and then manually review and iterate the model until a data set that meets the requirements is generated, such as Fig.17 shown.
[0073] Further, step 504 includes:
[0074] (1) A binary classification model is used to determine the noise signal and the image transmission signal in the energy density spectrum, and the noise signal in the energy density spectrum is removed to obtain an image transmission signal spectrum.
[0075] In the training process of the neural network model, the better the data validity, the better the performance of the neural network model. Therefore, it is necessary to clean the dirty data. First, it is necessary to make a preliminary classification of the sample energy density spectrum by model, and then analyze the vector data transmitted by each acquisition device to mark the noise signal and the image transmission signal. Therefore, it is necessary to train a preliminary model to identify the image transmission signal.
[0076] This step requires a binary classification task to be completed. The purpose is to label the noise signal and the image transmission signal in the energy density spectrum and complete the preliminary cleaning of the data. Here, the softmax model is selected to complete the binary classification task. For this binary classification model, it does not need to have high accuracy, nor does it need to care about whether it is over-trained, as long as it can reduce the burden of labeling.
[0077] First, the signal type is identified by combining the time-frequency spectrum with the energy spectrum, and then manually labeled. Because it is only a preliminary screening of the data, you can choose to manually label a very small data set to train a binary classification model. Then use the softmax model for training, such as Fig.18 As shown, it can be seen that the accuracy of the two-classification model can approach 100% after only 200 epochs.
[0078] After training the binary classification model, the binary classification model can be used to perform binary classification on the energy density spectrum vector collected by the lower computer. Because the training set is extremely small, overfitting will inevitably occur during the training process, so there are still many inaccurately labeled data in the initial classification. At this time, the labeling results are manually reviewed, and the workload is much less than the workload of directly labeling the data.
[0079] With the help of the preliminary model, we can complete the preliminary labeling of each type of drone, and then manually label it, which can easily complete a small-scale data set, and use the binary classification model to complete the prototype of an iterative recognition model for multi-model recognition. Then use the model to complete the labeling of a larger data set, manually review it, and iterate the training until the data set size meets the requirements (the current data set size is 12,000).
[0080] (2) Perform outlier filtering on the image transmission signal spectrum to suppress the WIFI signal in the image transmission signal spectrum and obtain an image transmission feature vector.
[0081] During the iteration process of the neural network model, with the increase of data volume and the increase of model types, the neural network model has severe jitter, such as Fig.19By retrospectively analyzing the results, we found that the reason is that the strength of the WIFI signal is much higher than the strength of the drone image transmission signal, which attracts the attention of the neural network model to the WIFI signal, resulting in the weak ability of the neural network model to capture the image transmission features. Different models may have WIFI signals, which reduces the differentiation of different models and causes the neural network model to not work well. Fig. 20 Therefore, it is necessary to suppress the WIFI signal so that the attention of the training neural network model is forced to shift to the image transmission features.
[0082] In an exemplary embodiment, filtering outliers on the image transmission signal spectrogram specifically includes:
[0083] 1) Calculate the energy vector mean of the image transmission signal spectrogram.
[0084] 2) Determine the peaks in the image transmission signal spectrum whose peak values are greater than the mean value of the energy vector, and merge adjacent peaks to obtain peaks to be filtered. The peaks whose distance is less than 10 are defined as adjacent.
[0085] 3) Calculate the peak mean and maximum peak of all peaks to be filtered in the image transmission signal spectrum.
[0086] 4) According to the peak mean and the maximum peak, determine whether there is an abnormal peak in the image transmission signal spectrum. Specifically, if the maximum peak is greater than 1.25 times the peak mean, it is determined that there is an abnormal peak in the image transmission signal spectrum, otherwise it is determined that there is no abnormal peak in the image transmission signal spectrum.
[0087] 5) If there is an abnormal peak, the peak mean is used as the screening threshold.
[0088] 6) The peaks to be filtered whose average peak height in the image transmission signal spectrum is greater than the screening threshold are removed, and normalized to obtain an image transmission feature vector.
[0089] 7) If there is no peak, the image transmission signal spectrum is normalized to obtain an image transmission feature vector.
[0090] After filtering outliers, the data needs to be normalized. If normalization is not performed, the neural network model is easily affected by the signal strength. Figures 21 to 28 shown.
[0091] Step 505: Classify the image transmission feature vector using a pre-trained neural network model to determine the model of the drone to be identified. The neural network model is trained using a training sample set, and the training sample set includes multiple sample image transmission feature vectors and the drone model corresponding to each sample image transmission feature vector.
[0092] The present application first generates multiple characteristic parameters based on the energy density spectrum to form a characteristic vector of image transmission. Among them, the multiple characteristic parameters are: the average interval between the peak functions of the energy density spectrum, the average width of the peak function of the energy density spectrum, the average energy density of the energy density spectrum, the average variance of the peak function of the energy density spectrum, the average third-order moment of the peak function of the energy density spectrum, and the average fourth-order cumulative amount of the peak function of the energy density spectrum.
[0093] The following are the characteristic parameters:
[0094] (1) Average interval between peak functions of energy density spectrogram: The duration and transmission interval of image transmission signals of different models are different. This characteristic parameter will provide timing information for the decision tree. The calculation formula is as follows:
[0095]
[0096] Where interval is the average interval between peak functions of the energy density spectrum, n is the number of peaks in a sample, and position is i is the starting position of the i-th peak, position i-1 is the starting position of the i-1th peak.
[0097] The degree of differentiation of the characteristic parameters for the four models is as follows: Fig.29 shown.
[0098] (2) Average width of the peak function of the energy density spectrum: The transmission rate, duration and correlation of the synchronization signal ZC used by different models of image transmission signals are different. Therefore, in theory, the average width of the peak should be able to reflect the characteristics of the drone image transmission signal. The calculation formula is as follows:
[0099]
[0100] Where peak_len is the average width of the peak function of the energy density spectrum, length i is the width of the ith peak (the peak width calculated starting from the 0.8 times mean position).
[0101] The degree of differentiation of the characteristic parameters for the four models is as follows: Fig.30 shown.
[0102] (3) Average energy density of energy density spectrum: The intensity and correlation of synchronization signal ZC used by different models of image transmission signals are different. Therefore, theoretically, the average energy density of the peak should be able to reflect the characteristics of the drone image transmission signal. The calculation formula is as follows:
[0103]
[0104] Among them, energy_density is the average energy density of the energy density spectrum, amplitude ij is the energy density of the jth point of the ith peak. The energy density of each point is the value of that point.
[0105] The degree of differentiation of the characteristic parameters for the four models is as follows: Fig.31 shown.
[0106] (4) Average variance of the peak function of the energy density spectrum: Different models of image transmission signals have different ZC transmitters, so the variances of different transmitters should also be different. Therefore, theoretically, the average energy density of the peak should be able to reflect the characteristics of the drone image transmission signal. The calculation formula is as follows:
[0107]
[0108] Among them, var is the average variance of the peak function of the energy density spectrum, mean i is the average peak energy density of the ith peak.
[0109] The degree of differentiation of the characteristic parameters for the four models is as follows: Fig.32 shown.
[0110] (5) Average 3rd-order moment of the peak function of the energy density spectrum: High-order moments provide more information about the data distribution pattern, which can help identify whether the data is symmetrical, skewed or has a peak, and further understand the characteristics of the data. In the fields of image processing and pattern recognition, high-order moments are used to describe image texture, shape, edge and other features, which helps to distinguish different image patterns. Therefore, high-order moments can represent certain characteristics of drone image transmission signals to a certain extent.
[0111] This application uses the third-order moment as the characteristic parameter, and its calculation formula is:
[0112]
[0113] Wherein, high_moment is the average third-order moment of the peak function of the energy density spectrum.
[0114] The degree of differentiation of the characteristic parameters for the four models is as follows: Fig.33 shown.
[0115] (6) Average 4th-order cumulative quantity of the peak function of the energy density spectrum: High-order cumulative quantity is another statistic that describes the data distribution, similar to high-order moments, but it has some unique properties and uses. High-order cumulative quantities are usually used to analyze the characteristics of data distribution in more depth. High-order cumulative quantities can help identify non-Gaussian characteristics of data distribution. Unlike standardized moments, the definition of high-order cumulative quantities is calculated starting from the original data, so they more directly reflect the characteristics of data distribution, especially for non-Gaussian distribution. Similar to moments, high-order cumulative quantities can be used to detect the skewness and kurtosis of data distribution. By analyzing the value of high-order cumulative quantities, it can be determined whether the data distribution is symmetrical or skewed, and whether the data distribution has a peak or flat feature. In the field of signal processing, high-order cumulative quantities can be used to analyze the statistical characteristics of signals, such as the distribution of noise, the stability and periodicity of signals, etc.
[0116] This application uses the 4th order cumulative amount as the characteristic parameter, and its calculation formula is:
[0117]
[0118] Among them, high_cumulant is the average 4th order cumulative amount of the peak function of the energy density spectrum, moment k is the k-order center of the peak, x i is the energy density of the ith peak, and μ is the mean energy density of all peaks.
[0119] The degree of differentiation of the characteristic parameters for the four models is as follows: Fig.34 shown.
[0120] Based on the above models and characteristic parameters, the summary is shown in Table 1 and Table 2.
[0121] Table 1 Average values of characteristic parameters
[0122] Features parameters\Model MINI4 MINI3 EVO2 NANO interval 2285 6347 2592 4050 peak_len 9200 3732 3255 6823 energy_density 25941 19160 25132 18772 var 6722829 39913409 23913409 16616305 high_moment -0.8951 -0.5296 -1.0145 -0.4339 high_cumulant 9.1558 1.8013 1.0083 -1.5420
[0123] Table 2. Variance of characteristic parameters
[0124] Features parameters\Model MINI4 MINI3 EVO2 NANO interval 296021 1799744 128029 854679 peak_len 9032018 298103 205778 13582161 energy_density 484651 865483 1199465 6473158 var 3189699645195 12213124930873 17773236558619 13950628630005 high_moment 0.2769 0.0042 0.0198 0.0369 high_cumulant 1.3985 1.4086 1.2459 7.6844
[0125] Further, multiple feature parameters are combined into an image transmission feature vector: feature_vector = [interval, peak_len, energy_density, var, high_moment, high_cumulant].
[0126] The requirements and meanings of each characteristic parameter are as follows:
[0127] 1) It is stipulated that each sample (8192*10 points) corresponds to an image transmission feature vector feature_vector.
[0128] 2) Use the label model to mark the image transmission protocol.
[0129] 3) Use the category tag to mark the specific model.
[0130] 4) Use file_name to save the storage location of the data.
[0131] 5) Use img_name to label the corresponding signal image.
[0132] Save the training sample set in CSV format. Further perform ten-fold cross validation on the training sample set. Cross validation can solve the problem of insufficient data volume and parameter tuning. In order to solve the shortcomings of simple cross validation, this application uses k-fold cross validation:
[0133] (1) First, all samples are divided into m sample subsets of equal size.
[0134] (2) Traverse the m sample subsets in turn, each time using the current sample subset as the validation set and all other samples as the training set to train and evaluate the neural network model.
[0135] (3) Finally, the average of the m evaluation indicators is used as the final evaluation indicator. In actual experiments, m is usually set to 10.
[0136] Furthermore, this application uses K-MEANS clustering analysis and TSNE dimensionality reduction algorithm to evaluate the image transmission feature vector.
[0137] K-MEANS cluster analysis: First, analyze the similarity between feature parameters. This part uses the Pearson coefficient to describe the similarity between each dimension of feature parameters. The Pearson coefficient matrix is provided as follows: Fig.35 As shown, it can be analyzed that the correlation between the dimensions is weak, and there is no parameter that needs to be specially optimized in terms of correlation.
[0138] K-MEANS is a method for clustering K types of data. When K is inappropriate, the clustering algorithm will show a relatively low clustering effect. The quantified clustering effect can be used to infer whether the optimal K value of the feature vector is consistent with the model category.
[0139] In practical applications, the sum of squared errors (SSE), the average deviation method and the silhouette coefficient (SC) are often used together to evaluate the effect of the clustering model. The smaller the SSE value, the better the clustering effect; the smaller the average deviation value based on the average deviation method, the better the clustering effect; the larger the silhouette coefficient, the better the clustering effect. In this application, the image transmission feature vector corresponds to 12 models, so the best K is 12, such as Figure 36 to Figure 38 shown.
[0140] t-distributed stochastic neighbor embedding (TSNE) dimensionality reduction algorithm: It is a machine learning algorithm for dimensionality reduction, proposed by Laurens van der Maaten and Geoffrey Hinton in 2008. In addition, TSNE is a nonlinear dimensionality reduction algorithm, which is very suitable for reducing high-dimensional data to 2D or 3D for visualization.
[0141] TSNE is already available in sklearn and can be directly called. The visualization results of TSNE can be used to determine the feature expression capabilities of the image transmission feature vector. The clustering effect of the current parameters is as follows: Fig.39 shown.
[0142] Through the visualization analysis of TSNE, it can be determined that some categories cannot be clustered clearly, and there are overlapping parts between clusters, which means that the image transmission feature vector cannot represent the image transmission characteristics of the drone well.
[0143] This application considers tree models and multi-layer perceptrons when selecting neural network models. Tree models have good classification capabilities, including decision trees and random forests.
[0144] Decision trees play a key role in the field of machine learning, and their performance is greatly affected by the selected feature vectors. Feature vectors consist of a series of characteristic parameters, which are the basis for the model to perform classification or regression.
[0145] It should be noted that if the selected feature vector cannot fully capture the essential characteristics of the sample data, the generated decision tree may overfit the training data. Overfitting means that the model relies too much on the noise or specific instances of the training data, while ignoring the real patterns in the data, resulting in poor generalization performance on unknown data. Even when trying to limit the size of the decision tree, the classification accuracy on new data will be significantly reduced due to the complexity of the tree model, usually approaching the inverse of the classification category. The decision tree in this case lacks generalization ability and cannot effectively adapt to new data.
[0146] Therefore, selecting a feature vector with good performance is crucial to building an efficient decision tree. An excellent feature vector should be highly representative, able to reflect the characteristics of the data set to the greatest extent, and avoid introducing unnecessary complexity. To this end, feature selection methods can be used to select the most relevant subset from a large number of features by evaluating indicators such as the degree of feature difference.
[0147] In the process of building a decision tree, feature selection plays a key role. By reasonably selecting features and effectively evaluating the generalization ability of the model, the prediction performance of the decision tree on unknown data can be improved, and more reliable and accurate classification results can be achieved.
[0148] The decision tree performs well in 4-class model recognition, with an accuracy of 94%, but performs poorly in 12-class model classification, with an accuracy of 61%. The random forest performs well in 4-class model recognition, with an accuracy of 97%, but performs poorly in 12-class model classification, with an accuracy of 67%. Therefore, this application uses the multi-layer perceptron model for classification and recognition.
[0149] Specifically, a two-layer perceptron model is selected, which includes an input layer, a hidden layer, and an output layer. The length of the input layer is the length of the image transmission feature vector, the length of the hidden layer is 256, and the length of the output layer is the classification category.
[0150] In order to make the neural network focus on the comprehensive relationship between features rather than relying on a specific feature, the output values of 10% of hidden neurons are randomly discarded during training and recognition, so that the neural network model pays more attention to the relationship between features. Using this neural network model training, 95% test set accuracy can be achieved, such as Fig.40 shown.
[0151] In an exemplary embodiment, the above-mentioned drone model identification method based on image transmission signals is deployed in a Zynq7020+AD9361 development board.
[0152] The development board is used to collect the image transmission signal of the drone, and each collected signal is mapped into a vector of length 640, and then written into the Double Data Rate (DDR) through the AXI-LITE interface. The Advanced RISC Machine (ARM) chip of the board is responsible for reading the data stored in the DDR and transmitting it to the host computer through the User Datagram Protocol (UDP). The host computer feeds the received signal to the neural network model, and the neural network model gives the probability of each classification. The category with the highest probability is the predicted model.
[0153] This application further provides the following two test examples:
[0154] Turn on the MavicAir 2s drone and place it 20m away from the room with a wall between them. Count the prediction results within one minute. It can be seen that the neural network model performs well indoors, with a fixed image transmission accuracy of 93%. The cumulative probability can achieve a 100% recognition rate, and the cumulative probability is calculated by the number of predicted models within every 2 seconds.
[0155] Turn on the MavicAir 2s drone and place it 20m away from the room with a wall between them. Count the prediction results within one minute. It can be seen that the neural network model performs well indoors, and the accuracy of frequency hopping image transmission is 53%. The cumulative probability can achieve a 100% recognition rate, and the cumulative probability is calculated by the number of predicted models within every 2 seconds.
[0156] The present application also provides an application scenario, which applies the above-mentioned drone model identification method based on image transmission signals. Specifically: The drone model identification method based on image transmission signals provided in this embodiment can be applied in the drone trapping process. The drone trapping scenario includes a model identification link, a trapping signal generation link, and a trapping link; the drone model identification method provided in this embodiment is applied in the model identification link. After identifying the model of the drone, a trapping signal is generated based on the relevant parameters of this model of drone to trap the drone.
[0157] In an exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the initial image transmission signal of the drone to be identified. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for identifying a drone model based on an image transmission signal is implemented.
[0158] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0159] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0160] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.
[0161] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.
[0162] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0163] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A method for identifying drone models based on image transmission signals. It is characterized in that The method for identifying the model of a drone based on image transmission signals includes: Obtain the initial image transmission signal of the drone to be identified; Performing aliasing sampling on the initial image transmission signal to obtain an aliased image transmission signal; Performing density calculation on the aliased image transmission signal to obtain an energy density spectrum; Performing data cleaning and outlier filtering on the energy density spectrum to obtain an image transmission feature vector; A pre-trained neural network model is used to classify the image transmission feature vector to determine the model of the drone to be identified; the neural network model is trained using a training sample set, and the training sample set includes multiple sample image transmission feature vectors and the drone model corresponding to each sample image transmission feature vector.
2. According to the method for identifying the type of drone based on the image transmission signal according to claim 1, It is characterized in that Performing aliasing sampling on the initial image transmission signal to obtain an aliased image transmission signal specifically includes: According to the bandwidth and frequency domain duty cycle of the unmanned aerial vehicle to be identified, the initial image transmission signal is extracted without filtering, so that the image transmission signal of the unmanned aerial vehicle to be identified covers the entire frequency spectrum to obtain an aliased image transmission signal.
3. The method for identifying drone models based on image transmission signals according to claim 1, It is characterized in that Density calculation is performed on the aliased image transmission signal to obtain an energy density spectrum, which specifically includes: Performing an energy calculation on the aliased image transmission signal to obtain a preliminary energy spectrum; Density calculation is performed on the preliminary energy spectrum to obtain an energy density spectrum.
4. The method for identifying drone models based on image transmission signals according to claim 3, It is characterized in that The window function of the density calculation is 1000.
5. The method for identifying drone models based on image transmission signals according to claim 1, It is characterized in that The energy density spectrum is cleaned and outlier filtered to obtain a feature vector for image transmission, which specifically includes: A binary classification model is used to determine the noise signal and the image transmission signal in the energy density spectrum, and the noise signal in the energy density spectrum is removed to obtain an image transmission signal spectrum; An outlier filtering is performed on the image transmission signal spectrogram to suppress the WIFI signal in the image transmission signal spectrogram to obtain an image transmission feature vector.
6. The method for identifying drone models based on image transmission signals according to claim 5, It is characterized in that Performing outlier filtering on the image transmission signal spectrogram to suppress the WIFI signal in the image transmission signal spectrogram to obtain an image transmission feature vector specifically includes: Calculating the energy vector mean of the image transmission signal spectrogram; Determine a peak in the image transmission signal spectrum whose peak value is greater than the mean value of the energy vector, and merge adjacent peaks to obtain a peak to be filtered; Calculate the peak value mean and maximum peak value of all peaks to be filtered in the image transmission signal spectrum; According to the peak mean and the maximum peak, determining whether there is an abnormal peak in the image transmission signal spectrum; If there is an abnormal peak, the peak mean is used as the screening threshold; The peaks to be filtered whose average peak height is greater than the screening threshold in the image transmission signal spectrum are removed, and normalized to obtain an image transmission feature vector; If there is no peak, the image transmission signal spectrogram is normalized to obtain an image transmission feature vector.
7. The method for identifying drone models based on image transmission signals according to claim 6, It is characterized in that If the maximum peak value is greater than 1.25 times the peak mean value, it is determined that there is an abnormal peak in the image transmission signal spectrum; otherwise, it is determined that there is no abnormal peak in the image transmission signal spectrum.
8. The method for identifying drone models based on image transmission signals according to claim 1, It is characterized in that The neural network model is a multi-layer perceptron.
9. A computer device, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for identifying the model of a drone based on image transmission signals as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the method for identifying the model of a drone based on image transmission signals described in any one of claims 1 to 8 is implemented.
Citation Information
Cited By
Unmanned aerial vehicle detection method and system based on pilot frequency synchronization sequence matching
CN119807861A
