Open set recognition method and system for UAV signals based on metric learning

Abstract

The present invention discloses a method and system for open-set identification of drone signals based on metric learning. By dual-channel monitoring of the 2.4G and 5.8G frequency bands, the recognition ability of different types of drone signals is significantly improved. A deep learning network model is used as an encoder to extract feature vectors, and the generalization ability of the model is enhanced through L2 regularization. At the same time, an improved triplet loss function is introduced to simplify the training process and accelerate model convergence. In addition, the present invention proposes an improved KNN unknown category detector with a dynamic update mechanism that can automatically adjust parameters, adapt to new data, and effectively identify samples of unknown categories. The open-set identification application method of the present invention can iteratively update the network model, classifier, and location category detector when a new category of drone signal is detected, ensuring the continuous optimization and adaptability of the system and improving the flexibility and accuracy of drone identification technology.

Description

Technical Field

[0001] This invention belongs to the field of deep learning, specifically relating to a method and system for open set recognition of UAV signals based on metric learning. Background Technology

[0002] As drone technology matures and its applications expand, drones have permeated all aspects of society. However, with the increasing number of drones, effectively identifying and managing them, especially unregistered or illegally used drones, has become a pressing issue.

[0003] Open set identification refers to the use of neural networks where the categories of samples in the test set are not necessarily the same as those in the training set; they may be untrained categories, also known as unknown or unseen categories. This is particularly important for drone identification in real-world conditions, as drone models and signal characteristics may constantly change with technological advancements, rendering traditional identification methods ineffective in recognizing these emerging drones. Summary of the Invention

[0004] To address the shortcomings of existing technologies and achieve the goal of effectively identifying newly emerging drones, this invention adopts the following technical solution:

[0005] The signal recognition system based on metric learning includes a drone signal dataset construction unit, a neural network model training unit based on metric learning, a sample feature vector dataset generation unit, a classifier, and an unknown category detector.

[0006] The UAV signal dataset construction unit acquires UAV signals and constructs a sample dataset of the signals;

[0007] The metric learning-based neural network model training unit classifies UAV signals by constructing and training a neural network model. During training, a triple matrix is ​​constructed for a set of input samples. The distance loss between each pair of samples in each triple is calculated. A mask matrix is ​​created based on the sample labels, and a corresponding mask matrix is ​​set according to the validity of the triples to remove the triple loss values ​​corresponding to invalid mask matrices, thus obtaining the triple loss for that set of samples. Based on the sample dataset, the neural network model is trained using the triple loss. The optimization objective of the triple loss is the distance between samples, bringing the feature vectors of samples of the same class closer together and widening the distance between feature vectors of samples of different classes. The improved triple loss eliminates the need for secondary processing of the dataset, simplifying the training process. Furthermore, each calculation considers all possible triples for all samples in the entire sample batch, resulting in higher sample utilization and faster model convergence. The neural network model also introduces a residual structure to avoid overfitting.

[0008] The sample feature vector dataset generation unit generates the sample feature vector dataset corresponding to the sample dataset through a trained neural network model.

[0009] The classifier and the unknown category detector are trained and constructed using the sample feature vector dataset and are used for the classification of UAV signals and the detection of unknown categories, respectively. The unknown category detector can effectively identify samples of unknown categories while maintaining high accuracy for known categories.

[0010] Traditional recognition networks require pre-determining the number of each category, meaning that each time a new non-classification task is added, the network model needs to be rebuilt and the network structure changed. Metric learning mainly focuses on the distance between sample feature vectors, aiming to reduce or limit the distance between intra-class samples and increase the distance between inter-class samples through training and learning. A typical metric learning consists of two parts: first, an encoder model to encode the original data into feature vectors; second, a similarity detection algorithm (classifier) ​​to compare the similarity of a pair of feature vectors. In this invention, a neural network model is used as the encoder model to extract features and encode the sample data into feature vectors. An independent KNN classifier and an improved KNN unknown category detector are used to perform similarity detection and classification of the feature vectors.

[0011] Furthermore, the UAV signal dataset construction unit simultaneously collects UAV signals from both the 2.4GHz and 5.8GHz frequency bands and combines them to form a sample dataset.

[0012] Furthermore, the UAV signal dataset construction unit performs short-time Fourier transforms on the real and imaginary parts of the acquired signal, respectively:

[0013]

[0014] Where X[m,k] represents the result of the short-time Fourier transform, which is a two-dimensional matrix, m represents the time index, and k represents the frequency index; x[n] represents the original discrete-time signal; ω[nm] represents the window function, which is a finite-length function centered at m, used to limit the time range of the analysis; N represents the size of the discrete Fourier transform in the short-time Fourier transform; represents the complex exponential term in the Fourier transform, used for weighting in the frequency domain, j is the imaginary unit, and n is an intermediate variable between positive and negative infinity;

[0015] The short-time Fourier transform results X_I[m,k] and X_Q[m,k] of the real and imaginary parts are combined along the frequency axis as sample data. The combined frequency axis is twice the length of the original frequency axis, i.e., X_2.4[m,2k].

[0016] Furthermore, the last layer of the neural network model is a fully connected layer, and the output result is L2 regularized, specifically as follows;

[0017]

[0018] Where output represents the L2 regularization result, and x represents the input feature vector. i Let |x| represent the i-th element in the eigenvector x. i | 2 ∑ represents the square of the absolute value of the i-th element in the eigenvector x. i Describes |x| for all i i | 2 The summation function calculates the sum of squares of elements along a specified dimension of the input x. `max(*, ∈)` represents the maximum value of the expression within the parentheses, and `∈` indicates that small positive numbers with zero denominators should be avoided. It represents the square root.

[0019] This invention applies L2 regularization to the network output encoding results, ensuring that the feature vectors have consistent scale across all dimensions. This helps the model to be less sensitive to scale changes and improves the model's generalization ability. L2 regularization can reduce the distance difference between samples of different classes, making the model more focused on distinguishing different classes rather than relying solely on the scale of features. The distance metric after regularization is the dot product between vectors, which can greatly simplify the calculation process.

[0020] Furthermore, the triplet loss is obtained by creating a three-dimensional matrix for each sample group, where each element of the matrix is ​​the triplet loss L(i,j,k) of the i,j,k-th sample in the sample group; creating a three-dimensional mask matrix mask(i,j,k) based on the labels of the samples in the sample group; if sample (i,j,k) is a valid triplet, then the corresponding position in the mask matrix is ​​marked; using the mask matrix mask(i,j,k) to set the illegal triplet loss in the three-dimensional matrix to 0, and then averaging the losses of all elements in the three-dimensional matrix to obtain the triplet loss of the entire sample group.

[0021] Traditional triplet loss requires selecting samples from the dataset to form triplets, which places high demands on the dataset. Therefore, this invention adopts an improved triplet loss, which directly calculates the triplet loss for each sample batch during the training of the neural network model. The improved triplet loss does not require secondary processing of the dataset, simplifying the training process. Furthermore, each calculation considers all possible triplets for all samples in the entire batch, resulting in higher sample utilization and faster model convergence.

[0022] Furthermore, the triplet loss is as follows:

[0023] L(a,p,n)=max{d(a,p)-d(a,n)+margin,0}

[0024] Among them, triplet<a,p,n> In this context, 'a' represents the baseline sample, 'p' represents a positive sample of the same class as 'a', 'n' represents a negative sample of a different class than the baseline sample, 'L(a,p,n)' represents the loss calculated from a triple, 'd(a,p)' represents the distance between samples a and p, 'd(a,n)' represents the distance between samples a and n, 'margin' represents the distance threshold, and 'max{}' represents the function that takes the maximum of the two values ​​within the parentheses.

[0025] The optimization objective of triplet loss is to reduce the distance between samples, bringing the feature vectors of samples of the same class closer together and increasing the distance between feature vectors of samples of different classes further apart.

[0026] Furthermore, while the traditional KNN algorithm performs well in classification tasks, it has certain limitations in detecting unknown categories. This invention provides an improved KNN unknown category detector. Based on the number of neighbors and a distance threshold, it offers an update mechanism that allows the addition of new training samples and labels, while simultaneously refitting and adjusting the distance threshold. In the prediction phase, the unknown category detector performs nearest neighbor queries on the input samples and calculates the distance between each sample and its nearest neighbor. If the average distance between samples exceeds a set threshold, the corresponding sample is classified as an unknown category. Through this dynamic update mechanism, the detector can adapt to new data and continuously optimize its performance. The automatically adjusted distance threshold reduces the need for manual parameter tuning, improving the model's versatility and ease of use.

[0027] Furthermore, if no distance threshold is specified, the distance threshold is set based on the average nearest neighbor distance multiple of a set of samples in the training data.

[0028] Furthermore, the unknown category detector, based on a voting confidence threshold, allows the addition of new training samples and labels while simultaneously refitting and adjusting the voting confidence threshold. If a sample's voting confidence is lower than the threshold, the sample is classified as an unknown category. This automatically adjusted voting confidence threshold reduces the need for manual parameter tuning, improving the model's versatility and ease of use.

[0029] The metric learning-based open set identification method for UAV signals involves, according to the metric learning-based signal identification system, taking unknown UAV radio frequency signals through the UAV signal dataset construction unit to obtain unlabeled samples, and inputting them into a neural network model to generate sample feature vectors. These vectors are then classified and detected using the classifier and the unknown category detector. For newly detected category samples, they are mixed with known category samples to construct a new dataset. The neural network model, the classifier, and the location category detector are iteratively updated. For detected known category samples, the classification result of the classifier is adopted, and the neural network, KNN classifier, and improved KNN unknown category detector are not updated.

[0030] The advantages and beneficial effects of this invention are as follows:

[0031] This invention significantly improves the ability to identify different types of UAV signals by monitoring the 2.4G and 5.8G frequency bands through dual channels. It employs a deep learning network model as the encoder to extract feature vectors and enhances the model's generalization ability through L2 regularization. Simultaneously, it introduces an improved triplet loss function to simplify the training process and accelerate model convergence. Furthermore, the improved KNN unknown category detector proposed in this invention has a dynamic update mechanism, which can automatically adjust parameters to adapt to new data and effectively identify unknown category samples. The open-set recognition application method of this invention can iteratively update when new categories are detected, ensuring continuous optimization and adaptability of the model and improving the flexibility and accuracy of UAV recognition technology. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the structure of a signal recognition system based on metric learning according to an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of the neural network model used in this embodiment of the invention.

[0034] Figure 3 This is a flowchart of the open set identification method for UAV radio frequency signals according to an embodiment of the present invention. Detailed Implementation

[0035] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0036] A method for open set recognition of UAV signals based on metric learning, including a signal recognition model training system based on metric learning and a method for open set recognition of UAV radio frequency signals;

[0037] Signal recognition model training system based on metric learning, such as Figure 1 As shown, the system includes a UAV RF signal dataset construction unit, a metric learning-based neural network model training unit, a sample feature vector dataset generation unit, and a KNN classifier and improved KNN detector unit. The UAV RF signal dataset construction unit preprocesses existing raw UAV RF signals and generates a sample dataset; the metric learning-based neural network model training unit receives the preprocessed samples and trains a deep learning network model using triplet loss; the sample feature vector dataset generation unit saves the neural network output of the sample dataset as a sample feature vector dataset; and the KNN classifier and improved KNN detector unit trains a classifier and detector using the sample feature vector dataset.

[0038] Traditional recognition networks require pre-determining the number of each category, meaning that each time a new non-classification task is added, the network model needs to be rebuilt and the network structure changed. Metric learning mainly focuses on the distance between sample feature vectors, aiming to reduce or limit the distance between intra-class samples and increase the distance between inter-class samples through training and learning. A typical metric learning consists of two parts: first, an encoder model to encode the original data into feature vectors; second, a similarity detection algorithm (classifier) ​​to compare the similarity of a pair of feature vectors. In this invention, a neural network model is used as the encoder model to extract features and encode the sample data into feature vectors. An independent KNN classifier and an improved KNN unknown category detector are used to perform similarity detection and classification of the feature vectors.

[0039] The construction process of the signal recognition model training system based on metric learning specifically includes the following steps:

[0040] Step S1: By constructing a deep learning sample dataset of UAV radio frequency signals, a deep learning database of UAV radio frequency signals is obtained.

[0041] Existing datasets and research mostly focus on the 2.4 GHz band, lacking research on other frequency bands. To address the potential multi-band signal switching modes of new drones, the dataset described in this invention employs a dual-channel simultaneous monitoring method to effectively record drone communication activities. It includes raw sample radio frequency signals from multiple drones of different types and background environments. Each sample, after preprocessing, contains short-time Fourier transform results for both the 2.4 GHz and 5.8 GHz bands. The sample dataset construction process includes the following steps:

[0042] Step S1.1: Taking the signal containing the real and imaginary parts of the time domain in the 2.4G band as an example, the original UAV radio frequency signal is windowed to increase the number of samples;

[0043] Step S1.2: Perform Short-Time Fourier Transform (STFT) on the real and imaginary parts of the signal respectively. The STFT is expressed as:

[0044]

[0045] Where X[m, k] is the result of STFT, a two-dimensional matrix, where m represents the time index and k represents the frequency index; x[n] is the original discrete-time signal; ω[nm] is the window function, usually a finite-length function centered at m, used to limit the time range of the analysis, such as a Hanning window or a rectangular window; N is the size of the DFT (Discrete Fourier Transform) in STFT; j is the imaginary unit, satisfying j 2 =-1; It is the complex exponential term in the Fourier transform, used for weighting in the frequency domain.

[0046] The real and imaginary STFT results X_I[m,k] and X_Q[m,k] are merged along the frequency axis, and the merged frequency axis is twice the length of the original frequency axis, i.e., X_2.4[m,2k].

[0047] The 5.8 GHz band signal was also windowed and subjected to Fourier transform. The obtained X_5.8[m, 2k] result and X_2.4[m, 2k] were then expanded to obtain the STFT result X[2, m, 2k] for the two bands.

[0048] Step S1.3: Package all results into a standard dataset format to facilitate neural network training;

[0049] Step S2, construct and train the deep learning network model, including the following steps:

[0050] Step S2.1, construct a deep learning network model, such as Figure 2 As shown;

[0051] This network model introduces a residual structure to avoid overfitting; unlike traditional classification-based neural network models, the deep learning neural network constructed in this invention does not have a classification layer, and the last layer is a fully connected layer, the purpose of which is to reduce the dimensionality of the original data into a lower-dimensional feature vector; the output result is L2 regularized in the final output layer, specifically expressed as follows;

[0052]

[0053] Where output is the calculation result, and x is the input vector. i It is the i-th element in vector x, |x i | 2 It is the square of the absolute value of the i-th element in vector x, ∑ i Describes |x| for all i i | 2 The summation function calculates the sum of squares of the elements along a specified dimension of the input x. `max(*, ∈)` represents the maximum value of the expression within the parentheses, where ∈ is a small positive number used to avoid division by zero. To represent the square root;

[0054] This invention applies L2 regularization to the network output encoding results, ensuring that the feature vectors have consistent scale across all dimensions. This helps the model to be less sensitive to scale changes and improves the model's generalization ability. L2 regularization can reduce the distance difference between samples of different classes, making the model more focused on distinguishing different classes rather than relying solely on the scale of features. The distance metric after regularization is the dot product between vectors, which can greatly simplify the calculation process.

[0055] Step S2.2: Select the improved triplet loss function as the loss function for training the neural network;

[0056] A typical Triplet loss consists of a triplet.<a,p,n> The loss is calculated using the following inputs: where 'a' represents the baseline sample, 'p' represents a positive sample of the same class as 'a', and 'n' represents a negative sample of a different class than the baseline sample. The loss is expressed as:

[0057] L(a,p,n)=max{d(a,p)-d(a,n)+margin,0}

[0058] L(a,p,n) is the loss calculated from a triple, d(a,p) represents the distance between samples a and p, d(a,n) represents the distance between a and n, margin is the distance threshold, and max{} is a function that takes the maximum of the two.

[0059] Traditional triplet loss requires selecting samples from the dataset to form triplet pairs, which places high demands on the dataset. Therefore, this invention employs an improved triplet loss method, directly calculating the triplet loss for each batch of samples during the neural network model training process. The specific steps are as follows:

[0060] Step S2.2.1: Create a three-dimensional matrix P(B,B,B) for each batch, where B = batch_size, and each element is equal to the triplet loss L(i,j,k) of the i,j,k samples in the batch.

[0061] Step S2.2.2: Create a three-dimensional mask matrix mask(i,j,k) based on the labels of the samples in the batch. If the sample (i,j,k) is a valid triplet, then the corresponding position of the mask matrix is ​​1.

[0062] In step S2.2.3, the illegal triplet loss in P(B,B,B) is set to 0 using the mask matrix mask(i,j,k), and then the average of all elements of the loss matrix P(B,B,B) is calculated to obtain the triplet loss of the entire batch.

[0063] The optimization objective of triplet loss is to reduce the distance between samples, bringing the feature vectors of samples of the same class closer together and increasing the distance between feature vectors of samples of different classes further apart. The improved triplet loss does not require secondary processing of the dataset, simplifying the training process. Furthermore, each calculation considers all possible triplets of all samples in the entire batch, resulting in higher sample utilization and faster model convergence.

[0064] Step S2.3: Divide the dataset generated in step S1 into a training set and a validation set, and train the neural network.

[0065] The trained neural network model is obtained.

[0066] In this invention, the MindSpore framework is used to train and evaluate the neural network model in a cyclical iterative manner until the model converges.

[0067] Step S3: Construct a UAV feature vector dataset; using the UAV radio frequency signal deep learning sample dataset from step S1, and the output of the neural network model obtained in step S2.3, generate a sample feature vector dataset.

[0068] Step S4: Construct and train the classifier and detector; using the sample feature vector dataset from Step S3, construct and train the classifier and detector; the classifier is used to classify the feature vectors of the UAV signal samples, and the detector is used to detect and label the feature vectors of the UAV signal samples. Both can be iteratively updated. The specific steps are as follows:

[0069] Step S4.1: Divide the feature vector dataset into a training set and a test set. Use the training set to train the KNN classification model to obtain the trained KNN classification model. Use the test set to test the trained KNN classification model. If the test is successful, obtain the KNN classification model for classification.

[0070] Step S4.2: The traditional KNN algorithm performs well in classification tasks, but it has certain limitations in detecting unknown categories. This invention provides an improved KNN unknown category detector that identifies samples of unknown categories by combining a distance threshold and voting confidence. The detection process of this detector includes the following key steps:

[0071] Step S4.2.1, Initialization: The detector receives the number of neighbors n_neighbors, the distance threshold distance_threshold, and the voting confidence threshold as parameters. If no distance threshold is specified, it is automatically set to twice the average distance of the nearest neighbors in the training data.

[0072] Step S4.2.2, Fitting: The detector learns features of known categories by fitting the training dataset X and the corresponding labels y. During this process, the detector records the set of known categories and automatically adjusts the distance threshold as needed.

[0073] Step S4.2.3, Update: In order to adapt to the dynamically changing data environment, the detector provides an update mechanism that allows new training samples and labels to be added to the original model, while refitting the model and adjusting the distance threshold.

[0074] Step S4.2.4, Prediction: In the prediction phase, the detector performs a nearest neighbor query on the input sample X and calculates the distance between each sample and its nearest neighbor. If the average distance of a sample exceeds a set threshold, or its voting confidence score is lower than the voting confidence score threshold, the sample is labeled as an unknown category.

[0075] The improved KNN unknown category detector of this invention can effectively identify samples of unknown categories while maintaining high accuracy for known categories. Through a dynamic update mechanism, the detector can adapt to new data and continuously optimize its performance. Furthermore, automatically adjusted distance and voting confidence thresholds reduce the need for manual parameter tuning, improving the model's versatility and ease of use.

[0076] like Figure 3As shown, after preprocessing the collected unknown UAV radio frequency signals in step S1, a batch of unlabeled samples is obtained. These samples are then input into the neural network model described in step S2.3 to obtain sample feature vectors. The KNN classifier and the improved KNN unknown category detector described in step S4 are used for classification and detection. When the probability of detecting an unknown category sample is greater than the detection threshold, the sample is labeled as a new category and mixed with some known category samples according to a certain weight to construct a new dataset. The neural network, KNN classifier, and improved KNN unknown category detector are iteratively updated. When a known category UAV is detected, the classification result of the KNN classifier is adopted, and the neural network, KNN classifier, and improved KNN unknown category detector are not updated.

[0077] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A signal recognition system based on metric learning, comprising a UAV signal dataset construction unit, a neural network model training unit based on metric learning, a sample feature vector dataset generation unit, a classifier, and an unknown category detector, characterized in that: The UAV signal dataset construction unit acquires UAV signals and constructs a sample dataset of the signals; specifically, it simultaneously acquires UAV signals from both the 2.4GHz and 5.8GHz frequency bands and combines them as the sample dataset; it then performs short-time Fourier transforms on the real and imaginary parts of the acquired signals respectively. Where X[m,k] represents the result of the short-time Fourier transform, which is a two-dimensional matrix, m represents the time index, and k represents the frequency index; x[n] represents the original discrete-time signal; ω[nm] represents the window function, which is a finite-length function centered at m, used to limit the time range of the analysis; N represents the size of the discrete Fourier transform in the short-time Fourier transform; represents the complex exponential term in the Fourier transform, used for weighting in the frequency domain, j is the imaginary unit, and n is an intermediate variable between positive and negative infinity; The short-time Fourier transform results of the real and imaginary parts are combined along the frequency axis, and the results of the 2.4G and 5.8G channels are combined as sample data. The metric learning-based neural network model training unit classifies UAV signals by constructing and training a neural network model. During training, a triplet matrix is ​​constructed for a set of input samples. The distance loss between each pair of samples in each triplet is calculated. A mask matrix is ​​created based on the sample labels, and a corresponding mask matrix is ​​set according to the validity of the triplet to remove the triplet loss values ​​corresponding to invalid mask matrices, thus obtaining the triplet loss for that set of samples. Based on the sample dataset, the neural network model is trained using the triplet loss. The last layer of the neural network model is a fully connected layer, and the output result is L2 regularized, specifically as follows: Where output represents the L2 regularization result, and x represents the input feature vector. i Let |x| represent the i-th element in the eigenvector x. i | 2 ∑ represents the square of the absolute value of the i-th element in the eigenvector x. i Describes |x| for all i i | 2 The summation function calculates the sum of squares of elements along a specified dimension of the input x. `max(*, ∈)` represents the maximum value of the expression within the parentheses, and `∈` indicates that small positive numbers with zero denominators should be avoided. To represent the square root; The sample feature vector dataset generation unit generates the sample feature vector dataset corresponding to the sample dataset through a trained neural network model. The classifier and the unknown category detector are trained and constructed using the sample feature vector dataset and are used for the classification of UAV signals and the detection of unknown categories, respectively.

2. The signal recognition system based on metric learning according to claim 1, characterized in that: The triplet loss is obtained by creating a three-dimensional matrix for each sample group, where each element of the matrix is ​​the triplet loss L(i,j,k) of the i, j,k samples in the sample group; and creating a three-dimensional mask matrix mask(i,j,k) based on the labels of the samples in the sample group. If the sample (i,j,k) is a valid triplet, then the corresponding position in the mask matrix is ​​marked. The loss of illegal triples in the three-dimensional matrix is ​​set to 0 using the mask matrix mask(i,j,k), and then the loss of all elements in the three-dimensional matrix is ​​averaged to obtain the triple loss of the entire sample group.

3. The signal recognition system based on metric learning according to claim 1 or 2, characterized in that: The triplet loss is as follows: L(a,p,n)=max{d(a,p)-d(a,n)+margin,0} Among them, triplet<a,p,n> In this context, 'a' represents the baseline sample, 'p' represents a positive sample of the same class as 'a', 'n' represents a negative sample of a different class than the baseline sample, 'L(a,p,n)' represents the loss calculated from a triple, 'd(a,p)' represents the distance between samples a and p, 'd(a,n)' represents the distance between samples a and n, 'margin' represents the distance threshold, and 'max{}' represents the function that takes the maximum of the two values ​​within the parentheses.

4. The signal recognition system based on metric learning according to claim 1, characterized in that: The unknown category detector provides an update mechanism based on the number of neighbors and a distance threshold, allowing the addition of new training samples and labels while refitting and adjusting the distance threshold. During the prediction phase, the unknown category detector performs nearest neighbor lookup on the input samples and calculates the distance between each sample and its nearest neighbor. If the average distance between samples exceeds a set threshold, the corresponding sample is classified as an unknown category.

5. The signal recognition system based on metric learning according to claim 4, characterized in that: If no distance threshold is specified, the distance threshold is set based on the average nearest neighbor multiple of a set of samples in the training data.

6. The signal recognition system based on metric learning according to claim 1, characterized in that: The unknown category detector, based on a voting confidence threshold, allows the addition of new training samples and labels, while refitting and adjusting the voting confidence threshold. If the voting confidence of a sample is lower than the voting confidence threshold, the sample is treated as an unknown category.

7. A method for open set recognition of UAV signals based on metric learning, characterized in that: According to claim 1, the signal recognition system based on metric learning obtains unlabeled samples from unknown UAV signals through the UAV signal dataset construction unit, and inputs them into a neural network model to generate sample feature vectors. The system then classifies and detects these samples using the classifier and the unknown category detector. For newly detected category samples, the system mixes them with known category samples to construct a new dataset. The system iteratively updates the neural network model, the classifier, and the unknown category detector. For detected known category samples, the system adopts the classification result of the classifier.

Images