Unmanned aerial vehicle signal open set identification method based on class center learning

Abstract

The invention provides an unmanned aerial vehicle signal open set recognition method based on class center learning for unmanned aerial vehicle signal recognition, the limitation of a traditional closed set recognition framework is broken through, known and unknown unmanned aerial vehicle signals are effectively distinguished through the open set recognition method, and the rejection capability for a novel or camouflage target is remarkably improved; physical characteristics of unmanned aerial vehicle signals and a class center learning mechanism of deep learning are deeply fused, and the characterization capability of the model on the essential attributes of the signals is enhanced; and the class center vector and the adaptive threshold adjustment strategy are dynamically optimized, so that the model can flexibly adapt to the signal distribution change, and the robustness in a complex scene is improved. According to the method, unmanned aerial vehicle signal identification is expanded to an open set scene, the risk of missed judgment caused by closed set hypothesis in the prior art is solved, and the safety and the practical application value are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of UAV signal recognition, and particularly relates to an open-set recognition method for UAV signals based on class center learning. Background Art

[0002] With the rapid development and wide application of UAV technology, the phenomena of "unauthorized flight" and "random flight" of UAVs have become increasingly prominent, posing a great threat to public safety and national security. In the field of UAV control, signal recognition technology is a key link for effective supervision. Although existing UAV signal recognition methods have achieved accurate recognition of known UAV signals from multiple perspectives, their ability to reject unknown UAV signals is still insufficient. Currently, new models of UAVs are launched quickly, and advanced technologies such as dynamic frequency conversion and signal camouflage are iterated rapidly, making it difficult for traditional recognition systems to cope with the threats of new UAVs. Therefore, breaking through the limitations of the closed-set recognition framework and developing an open-set recognition technology that can not only accurately identify known UAV signals but also efficiently distinguish unknown UAV signals has become an urgent need to improve UAV control efficiency and maintain social security.

[0003] In recent years, significant progress has been made in UAV signal recognition technology based on deep learning. Existing UAV signal recognition methods mainly rely on four types of data: audio, vision, radar, and radio frequency. For audio signals, by extracting time-frequency features such as logarithmic mel spectrogram and mel frequency cepstral coefficients, and combining convolutional neural networks for fusion analysis, the ability to identify UAV voiceprints is enhanced; vision recognition methods rely on object detection algorithms and combine image enhancement techniques to optimize the recognition efficiency of UAV optical images in complex scenarios; in the aspect of UAV radar signal processing, the limitations of traditional amplitude analysis are broken through, and the amplitude and phase components of radar echoes are jointly modeled to fully exploit the motion information in micro-Doppler features; while radio frequency signal recognition directly analyzes the radio frequency features of UAV communication or remote control signals through an end-to-end deep learning model, taking into account both high-precision classification and real-time requirements. These methods gradually construct an efficient and robust UAV signal recognition technology system through multi-modal feature co-extraction, deep network structure optimization, and fine modeling of signal physical characteristics.

[0004] However, existing UAV signal recognition methods are mainly based on a closed-set classification framework, with significant technical limitations. Current methods limit the recognition scope to known signals within predefined categories. However, the emergence rate of new UAVs far exceeds the update ability of the closed-set model, making it difficult for existing technologies to effectively recognize new and unknown UAV signals. Due to the lack of an effective discrimination mechanism, when an unknown signal appears, the system is prone to misclassify it as a known category, failing to achieve accurate discrimination between known and unknown signals. This confusion not only significantly reduces the recognition accuracy of known signals but also leads to a continuous increase in the false alarm rate of unknown signals, resulting in a double dilemma of declining recognition performance and rising misjudgment rate. The above technical defects seriously weaken the adaptability and scalability of the UAV control system, making it difficult to cope with new UAV threats and effectively guarantee the safety of people, critical infrastructure, and sensitive areas. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides an open-set recognition method for UAV signals based on class center learning. The technical problems to be solved by the present invention are achieved through the following technical solutions:

[0006] An open-set recognition method for UAV signals based on class center learning includes:

[0007] S100, extracting I and Q channel data from the original UAV signal dataset, and after normalizing the extracted data, obtaining a training set;

[0008] S200, setting a class center vector for each category in the training set;

[0009] S300, training a predetermined residual network using the training set, and modifying the output part of the residual network so that the residual network calculates the distance between the feature vector extracted from the input sample and each class center vector, calculates a mixed loss function based on this distance, and adjusts the class center vector and updates the parameters of the residual network using this mixed loss function to obtain a trained residual network;

[0010] S400, inputting the signal to be measured into the trained residual network and outputting an open-set recognition result.

[0011] Advantageous Effects:

[0012] The present invention proposes an open-set recognition method for UAV signals based on class center learning, which breaks through the limitations of the traditional closed-set recognition framework. By using open-set recognition technology, it can effectively distinguish known and unknown UAV signals, significantly improving the rejection ability for new or camouflaged targets. It deeply integrates the physical characteristics of UAV signals with the class center learning mechanism of deep learning, enhancing the model's ability to represent the essential attributes of signals. It dynamically optimizes the class center vector and the adaptive threshold adjustment strategy, enabling the model to flexibly adapt to signal distribution changes and improving the robustness in complex scenarios. The present invention extends UAV signal recognition to the open-set scenario, solves the risk of missed judgment caused by the closed-set assumption in the prior art, and significantly improves the safety and practical application value.

[0013] The following will further elaborate on the present invention in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0014] Figure 1 is the flowchart of the open-set recognition method for UAV signals based on class center learning of the present invention;

[0015] Figure 2 is the structural diagram of the residual network;

[0016] Figure 3 is the curve of the known class classification accuracy and the unknown class rejection rate of UAV-1 when both the known and unknown signals are UAV signals;

[0017] Figure 4 is the confusion matrix of the open-set recognition result when both the known and unknown signals are UAV signals;

[0018] Figure 5 is the curve of the known class classification accuracy and the unknown class rejection rate of UAV-1 when the known signal is a UAV signal and the unknown signal is a modulation recognition signal;

[0019] Figure 6 is the confusion matrix of the open-set recognition result when the known signal is a UAV signal and the unknown signal is a modulation recognition signal. Detailed Embodiments

[0020] The following further describes the present invention in detail with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0021] As Figure 1 shown, the present invention provides an open-set recognition method for UAV signals based on class center learning, including:

[0022] S100, extracting the I and Q channel data from the original UAV signal dataset, and obtaining the training set after normalizing the extracted data;

[0023] Among them, the original UAV signal dataset is composed of UAV signals, with a total of 4 categories. There are multiple samples under each category, and each sample is a UAV signal. The 4 categories are UAV-1 (Unmanned Aerial Vehicle), UAV-2, UAV-3, and UAV-4, and there are 208 samples under each category. The specific introduction is shown in Table 1.

[0024] Table 1 Known-class UAV Signal Dataset in the Training Phase

[0025] UAV model DJI Phantom 4 RTK Dataset size 17.5MB UAV type 4 Sample length 5400 Total number of samples 832

[0026] This step includes: S110, obtaining the original UAV signal; S120, extracting the I and Q channel data of the original UAV signal, and calculating the mean and standard deviation of the I and Q channel data; S130, using the mean and standard deviation to normalize the extracted data to zero mean and unit variance, and dividing the normalized data into a training set and a test set according to a predetermined ratio.

[0027] The original UAV signal can be expressed as follows:

[0028] s(n) = I(n) + jQ(n)

[0029] Among them, n represents the discrete time index, I(n) is the in-phase component, Q(n) is the quadrature component, and j is the imaginary unit;

[0030] The relationship between the I and Q channel data and the amplitude and phase of the original signal is as follows:

[0031]

[0032] Among them, |s(n)| represents the amplitude of the original signal, and φ(n) represents the phase of the original signal;

[0033] According to the above relationship, the I channel data I(n) and the Q channel data Q(n) can be obtained;

[0034] Calculate the mean and standard deviation of the I and Q channel data respectively, and then normalize the data to zero mean and unit variance as follows:

[0035]

[0036] Among them, μ I and σ I are the mean and standard deviation of the I channel data respectively, and μ Q and σ Q are the mean and standard deviation of the Q channel data respectively;

[0037] Divide the normalized dataset into a training set and a test set according to a ratio of 7:3.

[0038] S200, set a class center vector for each category in the training set;

[0039] In this step, for the known 4 categories of drone signals in the training set, a class center vector is set for each category, and all vectors are formed into a matrix.

[0040] For the known 4 categories of drone signals, set a class center vector for each category i, and merge these vectors into a matrix. For convenience, set the i-th element of the i-th row vector to m i , and all other elements are 0. Therefore, the class centers of all categories can be represented by a diagonal matrix M:

[0041]

[0042] where the i-th row of the matrix represents the class center vector of the i-th class, and the value of m i is a manually set constant;

[0043] After such settings, the class center vectors of each category are orthogonal, and the distances between the class center vectors of different categories can be adjusted by changing the value of m i . In the experiment, all m i are set to 10.0, that is, the class center matrix is a diagonal matrix with all diagonal elements being 10.0. In addition, the class center matrix M is fine-tuned during the network training process to adapt to the actual distribution of the training samples.

[0044] S300, use the training set to train a predetermined residual network, and modify the output part of the residual network so that the residual network calculates the distance between the feature vector extracted from the input sample and each class center vector, calculates a mixed loss function based on this distance, and adjusts the class center vector and updates the parameters of the residual network using this mixed loss function to obtain a trained residual network;

[0045] Reference Figure 2 , the structure of the residual network is as Figure 2 shown. The predetermined residual network includes a convolutional layer, a batch normalization layer, an activation layer, a pooling layer, a residual block, and a fully connected layer; where the residual block contains a convolutional layer, a batch normalization layer, and an activation layer; based on the residual network, modify the output part of the network so that it outputs the Euclidean distance between the feature vector of the input sample and all class center vectors. In the process of training the predetermined residual network of the present invention, the cosine annealing algorithm is used to dynamically adjust the learning rate of the residual network.

[0046] For any input sample x, extract the feature vector f through the residual network, and calculate its Euclidean distance from each class center in the class center matrix M:

[0047] d i = ∥f - m i ∥₂, i = 1, 2, 3, 4

[0048] where ∥·∥₂ represents the Euclidean distance, f is the feature of the input sample extracted by the network, and m i is the i-th row of the matrix M, representing the class center vector of the i-th class.

[0049] The result output by the final residual network is the vector (d₁, d₂, d₃, d₄).

[0050] During the training process, the present application designs a hybrid loss function, expressed as

[0051]

[0052] In the formula, L cross represents the inter-class loss, and L in represents the intra-class loss. d y is the Euclidean distance between the feature vector of the input sample and its corresponding class center vector, and d j is the Euclidean distance between the feature vector of the input sample and the class center vector of the j-th class. λ is a hyperparameter, and its value is set to 0.1 in the experiment.

[0053] During the training process, in order to reduce the loss, the value of |d y - d j | should be increased, so that the input sample is close to the center of its own class and far from the centers of other classes; at the same time, the value of d y should be reduced, so that the samples of the same class are more compactly centered around the class center vector of that class;

[0054] During training, the batch size is set to 32, the total number of training epochs is 100, the Adam optimizer is used, the weight decay is 0.0001, the initial learning rate is 0.001, and the cosine annealing algorithm is adopted to realize the dynamic adjustment of the learning rate.

[0055] The present application converts the distances (d₁, d₂, d₃, d₄) output by the residual network according to s i = -d i into class scores (s₁, s₂, s₃, s₄), obtaining the probability distribution of the input sample belonging to each class; selects the one with the largest probability as the predicted class of the input sample; calculates the hybrid loss function according to the predicted class of the input sample and the known class. The probability distribution of each class is as follows:

[0056]

[0057] Among them, \(x\) is the input sample, and \(P(y = i|x)\) represents the probability that it belongs to the category with label \(i\).

[0058] After the present application uses the mixed loss function to adjust the parameters of the residual network to obtain a trained residual network, the method for open-set recognition of UAV signals based on class center learning further includes:

[0059] Setting a discrimination threshold according to the Euclidean distance between the input sample and the class center vector, which is expressed as:

[0060]

[0061] In the formula, \(\alpha\) is a hyperparameter with a default value of 1.0, and \(S\) i represents the set of input samples correctly classified as category \(i\).

[0062] The present application preset and dynamically updates the class center vectors of each category, optimizes the feature distribution by using the Euclidean distance between the input sample and the class center, forces the inter-class distance to increase and the intra-class distance to decrease, thereby enhancing the model's ability to distinguish between known and unknown classes. And based on the maximum intra-class distance of the correctly classified samples in the training stage, the judgment threshold of each known class is dynamically determined, and the known class and the unknown class are directly distinguished through threshold logic, realizing robust rejection of unknown class UAV signals.

[0063] In the test stage, the test set consists of two parts: known category signals and unknown category signals. The known category signals are selected from four types of UAVs, namely UAV-1 to UAV-4, but none of them participated in the model training process. The number of samples in each category is consistent with the validation set division standard in the training stage. The unknown category signals have two different sources. One is the RF signals of two newly collected UAVs, UAV-5 and UAV-6. The other is randomly selected from the publicly available radar modulation recognition dataset RML2016.10A and the length is unified to 5400 through interpolation. The number of samples of the two different unknown category signals is the same as the number of UAV-1 mentioned above. To maintain the consistency of the preprocessing process, all data used in the test stage adopts the same normalization method as in the training stage. The specific introduction of the dataset in this stage is shown in Table 2.

[0064] Table 2 Test stage dataset

[0065]

[0066] All experiments of the present invention are completed on a computing platform equipped with an Intel Core i9-11950H processor and an NVIDIA RTX A3000 graphics card. The construction and testing of the deep learning model are implemented based on the PyTorch 2.0 framework.

[0067] In S400, the signal to be measured is input into the trained residual network, and an open-set recognition result is output.

[0068] S400 includes:

[0069] In S410, the signal to be measured passes through the trained residual network, and the Euclidean distance d between the signal to be measured and each class center vector is calculated. i ;

[0070] Among them, the signal to be measured can be a drone signal or other signals.

[0071] In S420, if d i is greater than the threshold θ of class i i , then it is determined that the signal to be measured is an unknown class signal; otherwise, it is determined as a known class signal, and the class with the smallest distance is selected as the predicted class of the signal to be measured, and finally an open-set recognition result is obtained.

[0072] Select UAV-1 to UAV-4 as known class signals and UAV-5 and UAV-6 as unknown class signals. Taking UAV-1 as an example, evaluate the ability of the open-set model to distinguish known classes and unknown classes at different thresholds. The results are as Figure 3 shown. It can be seen that there is a trade-off relationship between the recognition accuracy of known classes and the rejection rate of unknown classes. The threshold obtained according to the training results is 0.31. Near the intersection of the two curves, at this time, both the recognition accuracy of known classes and the rejection rate of unknown classes are at a good level, indicating that the model has good open-set recognition ability for UAV-1. Appropriately adjust the rejection threshold of each class, identify all test data, and finally obtain the confusion matrix as Figure 4 shown. It can be seen that the open-set recognition model can well complete the open-set recognition task of drone signals.

[0073] Select UAV-1 to UAV-4 as known class signals and the modulation recognition data set as unknown class signals. Similarly, taking UAV-1 as an example, evaluate the ability of the open-set model to distinguish known classes and unknown classes at different thresholds. The results are as Figure 5 shown. By comparing with Figure 3 , it can be known that since the unknown class signal is very different from the drone signal, the open-set recognition model can more easily distinguish the unknown class signal. Adjust the rejection threshold of each class, identify the signal to be measured, and obtain the confusion matrix as Figure 6 shown. It can be seen that the model has better open-set recognition ability for such unknown class data.

[0074] It can be proved from the above that the open set recognition model proposed by the present invention exhibits remarkable robustness and generalization ability in two typical test scenarios. It can not only effectively process unknown signals with features similar to known classes (rejecting same-type judgments), but also accurately identify unknown classes with significantly different features (rejecting different-type judgments), demonstrating engineering practical value in the open set recognition task of UAV signals in complex electromagnetic environments.

[0075] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, which should all be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for open-set recognition of UAV signals based on class center learning, characterized in that, Including: S100: Extract the I and Q channel data from the original UAV signal dataset, and normalize the extracted data to obtain a training set. S200: Set a class center vector for each category in the training set. S300: Use the training set to train a predetermined residual network, and modify the output part of the residual network so that the residual network calculates the distance between the feature vector extracted from the input sample and each class center vector, calculates a mixed loss function based on this distance, and uses this mixed loss function to adjust the class center vector and update the parameters of the residual network to obtain a trained residual network. S400: Input the signal to be measured into the trained residual network and output an open-set recognition result.

2. The method for open-set recognition of UAV signals based on class center learning according to claim 1, wherein The original UAV signal dataset is composed of UAV signals, with a total of 4 categories. There are multiple samples under each category, and each sample is a UAV signal.

3. The method for open-set recognition of UAV signals based on class center learning according to claim 1, wherein S100 includes: S110: Obtain the original UAV signal. S120: Extract the I and Q channel data of the original UAV signal, and calculate the mean and standard deviation of the I and Q channel data. S130: Use the mean and standard deviation to normalize the extracted data to zero mean and unit variance, and divide the normalized data into a training set and a dataset according to a predetermined ratio.

4. The method for open-set recognition of UAV signals based on class center learning according to claim 2, wherein, S200 includes: For the known 4 categories of UAV signals in the training set, set a class center vector for each category, and form all vectors into a matrix.

5. The method for open-set recognition of UAV signals based on class center learning according to claim 2, characterized in that The predetermined residual network includes a convolutional layer, a batch normalization layer, an activation layer, a pooling layer, a residual block, and a fully connected layer; among them, the residual block contains a convolutional layer, a batch normalization layer, and an activation layer; the output part of the predetermined residual network outputs the Euclidean distance between the feature vector of the input sample and all class center vectors.

6. The method for open-set recognition of UAV signals based on class center learning according to claim 1, wherein The calculating the mixed loss function according to this distance includes: Convert the distances (d1, d2, d3, d4) output by the residual network according to s i = -d i to class scores (s1, s2, s3, s4) to obtain the probability distribution of the input sample belonging to each class; Select the one with the highest probability as the predicted category of the input sample. Calculate the mixed loss function according to the predicted category of the input sample and the known categories.

7. The method for open-set recognition of UAV signals based on class center learning according to claim 6, wherein The mixed loss function includes an inter-class loss and an intra-class loss, and the mixed loss function is expressed by the formula: where L cross represents the inter-class loss, and L in represents the intra-class loss. d y is the Euclidean distance between the feature vector of the input sample and its corresponding class center vector, and d j is the Euclidean distance between the feature vector of the input sample and the class center vector of the j-th class, and λ is a hyperparameter.

8. The open set recognition method for drone signals based on class-centered learning according to claim 1 is characterized in that: During the process of training the predetermined residual network, use the cosine annealing algorithm to dynamically adjust the learning rate of the residual network.

9. The method for open-set recognition of UAV signals based on class-center learning according to claim 1, characterized in that After using this mixed loss function to adjust the class center vector and update the parameters of the residual network to obtain a trained residual network, the open-set recognition method for UAV signals based on class center learning further includes: Set a discrimination threshold according to the Euclidean distance between the input sample and the class center vector, expressed as: where α is a hyperparameter, and S i represents the set of input samples correctly classified as class i.

10. The method for open-set recognition of UAV signals based on class-center learning according to claim 1, wherein, S400 includes: S410, calculate the Euclidean distance d between the signal to be measured and each class center vector by passing the signal to be measured through the trained residual network i ; S420, if d i is greater than the threshold θ of category i i , then it is determined that the signal to be measured is an unknown class signal; otherwise, it is determined as a known class signal, and the category with the smallest distance is selected as the predicted category of the signal to be measured, and finally the open-set recognition result is obtained.

Images