Radiation source open set identification method and device based on dynamic group equivariant network and generative adversarial network

Abstract

The invention discloses a radiation source open set identification method and device based on a dynamic group isovariant network and a generative adversarial network. The method comprises the following steps: preprocessing a plurality of radiation source individual signals of known categories to obtain a data set for closed set identification; a dynamic group isovariant network is constructed, the network is trained on the obtained known category data set, and optimal model parameters are found; simulating virtual unknown category data on the basis of the known category data by using the generative adversarial network, and retraining the constructed recognition model by combining the unknown category data with the known category data to obtain a model capable of being used for open set recognition; and combining actually acquired unknown category data with known category data to serve as a test data set, extracting features through a retrained model, inputting the features into an OpenMax layer, judging whether the category is a known category or an unknown category according to threshold distribution, and detecting the recognition performance of the retrained network open set. The training efficiency is high in small sample recognition task, the classification recognition performance is good, and the robustness is high in open set recognition task.

Description

Technical Field

[0001] The present application relates to the field of signal recognition, and in particular to a method and device for open set recognition of radiation sources based on a dynamic group equivariant network and a generative adversarial network. Background Art

[0002] With the rapid development of the Internet of Things (IoT), the number of communication devices is expected to surge to 29 billion by 2030. This explosive growth has led to increasingly scarce spectrum resources and more complex electromagnetic environments. Similarly, the severity of communication security threats is also increasing. Malicious actors engage in counterfeiting and illegal intrusions, compete for limited spectrum resources and impersonate device identities, which can lead to significant financial losses. In order to solve these problems safely and efficiently, a specific emitter identification (SEI) method is proposed, which utilizes the inherent damage characteristics of the transmitter component, namely the radio frequency fingerprint (RFF) of the physical layer, to effectively distinguish different devices, thereby enhancing the security of IoT communications.

[0003] Existing SEI methods based on RFF are mainly divided into two categories. One is a feature extraction algorithm manually designed based on signal features, which extracts statistical features sufficient to distinguish different categories of devices and achieves accurate SEI through pre-built classifiers. The other is based on the rapidly developing deep learning network, which uses the powerful feature extraction ability of the network model to replace the manually designed features and achieve SEI efficiently and quickly. However, the research on these two types of methods mainly focuses on closed set recognition tasks, that is, the categories of the test set are consistent with the categories of the training set. The disadvantage of this task is that it cannot identify transmitter devices other than these known individuals. It is difficult to deploy applications in actual scenarios. It is easy to misclassify unknown categories as known category radiation source devices. If this unknown transmitter is a forged entity created by malicious actors, it will not be correctly identified, thereby affecting communication security and causing financial losses. Summary of the invention

[0004] The present application provides a method and device for open-set identification of radiation sources based on a dynamic group equivariant network and a generative adversarial network, which are used to solve the problems of poor robustness, low recognition accuracy and high misjudgment rate of existing specific radiation source identification methods in the task of individual identification of open-set radiation sources.

[0005] In order to solve the above problems, the present application proposes a radiation source open set identification method based on a dynamic group equivariant network and a generative adversarial network, the method comprising:

[0006] S1, preprocessing the received individual signals of radiation sources of known categories to obtain a closed set identification sample set; the sample set includes signal samples of multiple categories, each category has the same number of signal samples, and each signal sample is marked with an identification category label;

[0007] S2, build a dynamic group equivariant network, divide the known category data set obtained into training set, validation set and test set, use the training set to train the model, and use the validation set to verify the model performance. Finally, find the optimal model parameters through the performance on the test set to obtain the closed set recognition model with the best recognition performance;

[0008] S3, build a generative adversarial network, train the known category data through adversarial network training, generate a large number of virtual samples, combine the generated samples with the known category samples, obtain an open set recognition data set close to the real distribution, and retrain the closed set recognition model based on the data set to learn the distribution of unknown category data and obtain a model that can be used for open set recognition;

[0009] S4, treats the signal samples actually received outside the known category as unknown category samples, preprocesses them and combines them with known category data to obtain a test data set, extracts features through the retrained model, inputs them into the constructed OpenMax layer, and judges them as known category samples or unknown category samples based on the classification score and the set threshold, and then detects the open set recognition performance of the constructed model on the overall test sample set.

[0010] A radiation source open set identification device based on a dynamic group equivariant network and a generative adversarial network, comprising:

[0011] The signal sample set acquisition unit is used to perform corresponding preprocessing on the received individual signals of multiple known categories of radiation sources to construct a data set for closed set recognition; the signal sample set includes signal samples of multiple categories, each category has the same number of signal samples and is marked with category labels; in addition, a number of unknown signal samples other than known category samples are received, corresponding preprocessing is completed, N+1 labels are marked, and they are combined with known category samples into a test data set for open set recognition;

[0012] The closed set recognition network training unit is used to construct a dynamic group equivariant network, divide the constructed closed set recognition data set into a training set, a validation set and a test set, input the training set into the constructed network, extract features, obtain the predicted label through the classification layer, compare it with the real label and calculate the loss, and adjust the model parameters to minimize the model training loss and obtain the optimal closed set recognition model;

[0013] Generate adversarial simulation retraining unit, used to build a generative adversarial network, input the constructed known category data set into the constructed network, use the idea of ​​adversarial training to minimize the loss between the simulated virtual samples and the real samples, to imitate unknown samples, and combine them with known category data to obtain a virtual sample data set, retrain the optimal closed set recognition model, let the model learn the distribution of unknown data, and obtain a model that can perform open set recognition;

[0014] The open set recognition performance detection unit is used to extract the features of the test set samples using the retrained closed set recognition model, and input them into the built OpenMax layer. The Weibull element recognition probability is calculated based on the distance between the feature vector of each signal sample and the center point of each category, and the probability estimate of each class is adjusted based on the weight calculated by the maximum activation class. The probability calculation of each sample is summed and normalized to obtain the probability that the sample belongs to each category. When the probability of belonging to the unknown class is the largest, OpenMax will reject these samples. Combined with the maximum a posteriori probability (MAV) and threshold processing technology, open set recognition performance detection is realized.

[0015] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects:

[0016] The scheme of the embodiment of the present application adopts a dynamic group equivariant network as the feature extraction network framework for open set signal recognition, and adopts two stages: closed set recognition training and generating virtual data for retraining; a dynamic group equivariant convolutional network is adopted, and the adaptability of the dynamic convolution kernel and the equivariance of the group equivariant convolution are utilized to enhance the feature extraction capability of communication signals in complex electromagnetic environments; virtual unknown samples are generated by generative adversarial network simulation to retrain the network, so that the model can effectively learn the distribution of unknown category samples and improve the discrimination capability of the scheme for individual category recognition of unknown radiation sources; the present application also combines the reasoning and discrimination capability of the OpenMax layer, calculates the Weibull CDF probability of each category sample, modifies the original sample activation vector, obtains the probability that the sample belongs to each category, and uses the maximum a posteriori probability (MAV) and threshold processing technology to improve the discrimination detection capability of unknown samples; compared with other open set recognition methods such as simple OpenMax, fully connected layer and Softmax layer, the method of the present application has better unknown sample detection and recognition capability and recognition robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of the process of an unknown radiation source individual identification method proposed in an embodiment of the present application is shown;

[0018] Figure 2 A schematic diagram of a method for identifying and classifying unknown radiation sources based on a dynamic group equivariant network and a generative adversarial network according to an embodiment of the present application is shown;

[0019] Figure 3 A schematic diagram of an equivariant convolution kernel structure and a dynamic routing layer structure of an embodiment group of the present application is shown;

[0020] Figure 4 A comparison result diagram of the test accuracy of the group equivariant convolutional network of the present application on the measured ADS-B signal in the closed set recognition scenario is shown;

[0021] Figure 5 A comparison result diagram of the unknown sample recognition and classification performance of the measured ADS-B signal by the open set recognition method combining the dynamic group equivariant network and the generative adversarial network of the present application in the open set recognition scenario is shown;

[0022] Figure 6 A functional structure diagram of an unknown radiation source individual identification device based on a dynamic group equivariant network and a generative adversarial network proposed in an embodiment of the present application is shown;

[0023] Figure 7 A schematic structural diagram of an electronic device proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0025] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] The core of the present application is to extract signal sample features by using a method based on a dynamic group equivariant network, combined with virtual samples generated by generative adversarial network simulation, to improve the model's out-of-domain generalization ability and classify individual signals of unknown radiation sources, so as to solve the problem of poor performance of other network frameworks in the task of unknown radiation source identification.

[0027] Example 1

[0028] See also Figure 1 and Figure 2 As shown, the embodiment of the present application proposes a method for identifying an unknown radiation source individual based on a dynamic group equivariant network and a generative adversarial network, including steps S1 to S5:

[0029] S1, preprocessing the received individual signals of radiation sources of known categories to obtain a closed set identification sample set; the sample set includes signal samples of multiple categories, each category has the same number of signal samples, and each signal sample is marked with an identification category label.

[0030] In some embodiments of the present application, the preprocessing of the received individual signal of the radiation source in step S1 includes:

[0031] The received individual signal of each radiation source is downsampled and discretized, and down-converted to a zero intermediate frequency to obtain a discrete complex baseband signal as a signal sample. The discrete complex baseband signal is expressed as:

[0032] y[n]=y I [n]+j*y Q [n]n∈[0,N p -1]

[0033] Here N p is the number of sampling points, y I and Q are in-phase and quadrature signals respectively, and y[n] is a discrete complex signal.

[0034] Taking the ADS-B signal collected as an example, the frequency of the ADS-B signal is 1090MHz, and the sample length is 6000 points. ADS-B refers to the Automatic Dependent Surveillance-Broadcast system, which is a broadcast system that automatically sends signals containing a variety of aircraft attribute information. Ground personnel can receive the signal through the built receiving equipment. The attribute information includes the ICAO code, longitude, latitude, speed, altitude, heading and other information of the aircraft. After parsing the signal to obtain the above information, the signal can be labeled with the ICAO code. In the constructed signal data set, the signal is ICAO code segment information removed. At this time, the sample length is 4800 points to prevent the network from learning the identity information of the signal during the feature extraction process, so as to have a higher recognition accuracy. This application only uses the ADS-B signal as a specific signal for radiation source identification, which can be extended to other radiation source individual identification tasks.

[0035] The signal sent by the device on the aircraft will be affected by the channel in the propagation path and reach the ground receiving device. The received analog signal can be expressed as Where h(t) is the time-varying impulse response of the channel, n(t) is the additive white Gaussian noise, r(t) is the complex baseband signal, and the received signal is down-converted to a zero intermediate frequency signal to obtain y[n] = y I [n]+j*y Q [n], where n∈[0,N p -1],N p is the number of sampling points, y I and Q are in-phase and quadrature signals respectively, y[n] is a discrete complex signal, and the received signal is preprocessed into IQ form for subsequent feature extraction and recognition and classification.

[0036] S2, build a dynamic group equivariant network, divide the obtained known category data set into training set, validation set and test set, use the training set to train the model, and use the validation set to verify the model performance. Finally, find the optimal model parameters through the performance on the test set to obtain the closed set recognition model with the best recognition performance.

[0037] In the scheme of the present application, it is necessary to first obtain a model with good closed-set recognition performance that can extract the deep-level features of the signal. Therefore, a dynamic group equivariant network is built, which is a generalization of the traditional convolutional network. That is, more equivariance is introduced in the convolution operation of the traditional convolutional neural network, which effectively improves the expressive power of the model and the ability of the convolution kernel to extract signal representation.

[0038] See also Figure 3As shown in Figure 1, group equivariant convolution is achieved by directly multiplying the group equivariant convolution kernel with a transformation group with different transformation matrices. During the reverse extraction and propagation process, the update gradient flows from the output transformation group into the convolution kernel and flows out from the input transformation group to the next module.

[0039] The entire dynamic routing layer consists of two parts: one is to generate routing coefficients according to the input data, and the other is to transform the generated coefficients through the circulant matrix. The routing layer is implemented through global average pooling and the fully connected layer W fc1 , followed by Sigmoid as the activation function σ(·) to calculate the routing coefficient

[0040] in, In mathematics, It can be considered as being calculated by a sample-dependent nonlinear routing function.

[0041] Compared with the static convolution with fixed convolution kernel, the convolution kernel in the dynamic convolution layer is not a fixed convolution kernel for all samples, but is parameterized as a dynamic combination of multiple convolution kernels, specifically:

[0042] Among them, W i represents a convolution kernel of fixed size, is the coefficient generated by the above dynamic routing layer, and σ is the activation function.

[0043] Afterwards, the constructed known data set is divided into training set, validation set and test set according to a certain ratio. The training set is used to train the model, and the validation set is used to verify the model performance. Finally, the optimal model parameters are found through the performance of the model on the test set to obtain the closed-set recognition model with the best recognition performance. The training set usually has the most samples, and the number of samples in the validation set and the test set are usually the same.

[0044] In some implementations of the present application, the data set is divided in step S2 in the following manner:

[0045] The dataset is randomly divided into three subsets from all samples of the constructed dataset in a ratio of 8:1:1. Assuming that the total number of samples in the dataset is 1000, the sizes of the three subsets are 800 samples for the training set, 100 samples for the validation set and 100 samples for the test set, covering all radiation source categories. Random sampling and division are performed for each training to avoid the phenomenon that the recognition accuracy is inflated due to the same training data each time.

[0046] S3, build a generative adversarial network, train the known category data through adversarial network training, generate a large number of virtual samples, combine the generated samples with the known category samples, obtain an open set recognition data set close to the real distribution, and retrain the closed set recognition model based on this data set to learn the distribution of unknown category data and obtain a model that can be used for open set recognition.

[0047] Figure 2 The specific structure of the generative adversarial network is demonstrated. The generator part of the network contains 5 convolutional layers, and the discriminator part contains 2 convolutional layers. The generator part forges unknown category data through the input signal data, and minimizes the loss between it and the real data through the discriminator. The data features are similar but the data distribution is different. It acts as unknown radiation source category data and is combined with known data to form a data set containing known category samples and virtual category samples. The closed set recognition model is retrained to allow the model to effectively learn the generated unknown sample distribution, so that it has a certain discrimination ability when facing real unknown samples.

[0048] Generative adversarial networks use a competitive approach similar to that in game scenarios to improve the quality of generated data. The loss function is expressed as follows:

[0049]

[0050] Where D(x) is the probability that the discriminator classifies the real data, D(G(z)) is the probability that the discriminator classifies the generated data, and p data (x) represents the distribution of real data, p z (z) is the distribution of the noise vector input to the generator. Through this adversarial training process based on the loss function, both the generator and the discriminator improve their generation and discrimination capabilities, and the fake samples finally generated are very similar to the real data.

[0051] This application uses generative adversarial network simulation to generate virtual samples, and combines them with known samples to retrain the closed set recognition model, so that the model can learn the distribution of new samples. This method has the following beneficial effects:

[0052] 1) Generate diverse and high-quality data: By constructing a generative adversarial model and using adversarial training of the generator and the discriminator, high-quality and diverse fake samples can be generated. These samples can help the model better understand the data distribution and thus improve the recognition ability of unknown samples.

[0053] 2) Ability to recognize unknown categories: The OpenMax layer is an improved softmax layer that can process samples of unknown categories more effectively. By combining with the generative adversarial network, the OpenMax layer can use the generated high-quality samples to further improve the ability to recognize and process samples of unknown categories.

[0054] 3) Reduce the risk of misclassification: In traditional classification tasks, the model is prone to misclassify samples of unknown categories as known categories, while the samples generated by the generative adversarial network can help the model identify unknown samples. The OpenMax layer further adjusts the classification boundaries based on these generated samples, thereby reducing the risk of misclassification.

[0055] 4) Improve the robustness of the model: The samples generated by the generative adversarial network cover more data distribution situations, allowing the model to be exposed to more marginal distribution samples during training. After combining with the OpenMax layer, the model can more robustly cope with various complex input situations and improve its performance in practical applications.

[0056] 5) Adapt to dynamically changing environments: In practical applications, data distribution may change over time. Generative adversarial networks can continuously generate new samples to help the model adapt to new data distributions, while the OpenMax layer can dynamically adjust the classification strategy, so that the model can still maintain good unknown category recognition performance in the face of a changing data environment.

[0057] S4, treats the signal samples actually received outside the known category as unknown category samples, preprocesses them and combines them with known category data to obtain a test data set, extracts features through the retrained model, inputs them into the constructed OpenMax layer, and judges them as known category samples or unknown category samples based on the classification score and the set threshold, and then detects the open set recognition performance of the constructed model on the overall test sample set.

[0058] In some embodiments of the present application, see Figure 2 , this step S4 specifically includes:

[0059] The signal samples actually received outside the known category are regarded as unknown category samples, which are preprocessed and combined with the known category data to obtain the test data set. The features are extracted through the retrained model. The features are specifically expressed as: F = DEGCONV OSR (x),

[0060] Where DEGCONV OSR (·) represents the model for open set recognition obtained after retraining, and F is the sample feature extracted by the network; the feature is then input into the built OpenMax layer, which uses the sample and μ i The distance between the two is used to calculate the Weibull element recognition probability, where μ iIt is the center point calculated using the i-th class related samples. Then, the first α largest activation classes are selected. These classes are most likely to be the correct classes and the activation values ​​of each class are calculated. The weights are used to calculate the weights to adjust the probability estimates of each class. Based on the calculated activation values ​​and Weibull distribution, the cumulative distribution function (CDF) probability of each class is calculated, which helps to distinguish known classes from unknown classes. The original activation vector is then modified according to the probability. The new activation vector reflects the probability that the sample belongs to a known class. The calculated probability is:

[0061] in It represents the modified activation value of the input sample at category j. The exponential sum of all categories is used as the denominator to ensure that the sum of the probabilities of all categories is 1, and the probability of the sample belonging to each category is obtained. When the probability of belonging to the unknown class is the largest, OpenMax will reject these samples. Finally, the maximum a posteriori probability (MAV) and threshold processing techniques are used to further enhance the detection of unknown samples, reject unknown samples in the presence of unknown samples, and identify the correct samples.

[0062] In some embodiments of the present application, the evaluation method is different from closed set recognition. Closed set recognition only requires the model to classify each type of data to obtain the accuracy index. Since unknown radiation source sample recognition introduces unknown classes, it cannot simply focus on accuracy during evaluation. A macro F1 score is also introduced as an evaluation indicator, which calculates the average F1 score of the model in all known categories and gives equal importance to the performance of each category. The F1 score is the harmonic mean of precision and recall. The classification model provides a comprehensive evaluation indicator, specifically:

[0063]

[0064] Among them, Precision indicates how many true examples there are in the samples predicted by the model as positive examples, and Recall indicates the proportion of all true examples successfully predicted by the model. The macro F1 score can be specifically expressed as follows:

[0065] The above formula calculates the average F1 score of all known categories, which can more comprehensively evaluate the performance of the model in multi-category classification problems, especially when facing open domain data. Still using the above specific training scenario example to illustrate, the known category data contains 100 categories, the unknown category data contains 20 categories, and each category has 500 samples. When training the closed set recognition model, 40,000 samples are randomly selected from the 50,000 samples of known category data as the training set, and 5,000 samples are selected from the remaining 10,000 samples as the validation set, and the remaining 5,000 samples are used as the test set. During the closed set recognition training process, the dynamic convolution kernel will calculate different weights for the input samples to adapt to different samples. The group equivariant convolution kernel will extract group features from the samples that are fused after three Hilbert transforms of the signal samples, and input the features into the fully connected layer. The SoftMax function is used to generate the features according to the distance. The probability of each signal sample corresponding to each category is calculated. The category corresponding to the maximum probability can be considered to belong to this category. For example, the probabilities of a signal sample corresponding to 5 categories A to E are: 10% for A, 18% for B, 22% for C, 30% for D, and 20% for E. Then the sample is considered to belong to category D. Finally, the predicted label of each sample is compared with its true label. If the two labels are consistent, it is considered that the recognition is correct. The final recognition accuracy is counted. For example, for 100 samples, 95 sample labels are correctly recognized and 5 sample labels are incorrectly recognized. Then the recognition accuracy is 95%. By adjusting the model parameters, the recognition accuracy of the closed set recognition model can reach more than 90%.

[0066] In summary, the radiation source open set identification method based on dynamic group equivariant network and generative adversarial network proposed in the embodiment of the present application first pre-processes the received individual signals of multiple known categories of radiation sources to obtain a known category signal data set; then the known category signal data set is divided into a proportion to obtain a training set, a validation set and a test set; then a dynamic group equivariant network is constructed, and based on the dynamic convolution kernel and the group equivariant convolution kernel, the deep-level features of the signal are extracted, and the signal sample prediction and classification are realized through the fully connected layer, and the model performance is tested on the test set. By adjusting the model parameters, the test recognition accuracy reaches more than 90%, and a closed set recognition model with the best recognition effect is obtained; then, the adversarial training idea of ​​the generative adversarial network is used to generate a batch of virtual samples close to the real samples, which act as unknown category samples and are compared with the known The category samples are combined and the closed set recognition model is retrained to make it fully learn the distribution of unknown category samples and obtain a model that can perform open set recognition; finally, the unknown category samples collected are combined with the known category samples to obtain the open set recognition test set, which is input into the model, and the Weibull element recognition probability is calculated through the OpenMax layer. Then, the first α largest activation classes are selected. These classes are most likely to be the correct classes and the activation value of each class is calculated. The weight is used to calculate the probability estimate of each class. Based on the calculated activation value and Weibull distribution, the cumulative distribution function (CDF) probability of each class is calculated, and then the original activation vector is modified according to the probability to obtain the probability that the sample belongs to the known class. If the probability of belonging to the unknown class is large, it is judged as a position class. It can be seen that the scheme of the embodiment of the present application adopts a dynamic group equivariant network as the network framework for closed-set signal recognition, based on dynamic convolution kernels and group equivariant convolution kernels, extracts deep-level features of the signal, fully characterizes the signal, and breaks through the limitations that the traditional convolution network can only extract local features and the convolution kernel has weak feature extraction ability after the data distribution changes; Generate virtual samples by simulation of the generative adversarial network, retrain the model, which can enhance the generalization ability of the model, reduce the risk of model misclassification, improve the robustness of the model, and adapt to dynamically changing environments; Use the distance metric and activation vector modification of the OpenMax layer to obtain the probability that the sample belongs to each category, and use the maximum a posteriori probability (MAV) and threshold processing technology to further enhance the detection of unknown samples, reject unknown samples in the presence of unknown samples, and identify the correct samples. Compared with other network models such as ResNet, ComplexCNN, Transformer, etc., the scheme of this application has better closed-set recognition performance. Compared with other open-set recognition methods such as Softmax, fully connected layers and OpenMax, it has good classification and recognition performance in the unknown radiation source recognition task, and strong generalization and robustness.

[0067] Example 2

[0068] Figure 4 The figure shows the comparison result of the recognition accuracy of the measured ADS-B signal under closed set recognition conditions with other model structures. The recognition task is 100 categories of ADS-B signals, 500 samples in each category, and the signal-to-noise ratio is -20dB to 20dB. The compared CNN model is a recognition model containing 13 layers of convolution, ResNet is an 18-layer ResNet model, the ComplexCNN model is a recognition model with 12 convolution layers plus one complex convolution layer, and the Transformer model is a 12-layer convolution layer plus one Transformer encoding layer. The parameters of each recognition model are adjusted to achieve the optimal recognition accuracy. The dynamic group equivariant network in this application utilizes the weight distribution of dynamic convolution and the equivariance given by equivariant convolution to fully characterize the signal samples and achieve the optimal closed set recognition accuracy.

[0069] Figure 5 The figure shows the comparison of recognition performance between the proposed method framework and other methods under open set recognition conditions. The test set samples include 100 categories of known samples and 20 categories of unknown samples. The recognition performance of different algorithms under different signal-to-noise ratios is compared. The evaluation indicator at this time is the Macro F1 score. The Macro F1 score of the proposed solution in this application can exceed 0.8 under 0dB conditions. Compared with other methods, the average improvement is 0.1, and it reaches a significant 0.88 under 20dB conditions, which proves the excellent performance of the proposed solution in this application.

[0070] Example 3

[0071] like Figure 6 As shown, the embodiment of the present application also proposes a radiation source open set identification device based on a dynamic group equivariant network and a generative adversarial network, the device comprising:

[0072] The signal sample set acquisition unit 61 is used to perform corresponding preprocessing on the received individual signals of multiple known categories of radiation sources to construct a data set for closed set recognition; the signal sample set includes signal samples of multiple categories, each category has the same number of signal samples and is marked with category labels; in addition, many unknown signal samples other than known category samples are received, corresponding preprocessing is performed, N+1 labels are marked, and they are combined with known category samples into a test data set for open set recognition;

[0073] The closed set recognition network training unit 62 is used to construct a dynamic group equivariant network, divide the constructed closed set recognition data set into a training set, a validation set and a test set, input the training set into the constructed network, extract features, obtain predicted labels through the classification layer, compare with the real labels to calculate the loss, and adjust the model parameters to minimize the model training loss to obtain the optimal closed set recognition model;

[0074] Generate adversarial simulation retraining unit 63, used to construct a generative adversarial network, input the constructed known category data set into the constructed network, use the idea of ​​adversarial training to minimize the loss between the simulated virtual samples and the real samples, so as to imitate the unknown samples, and combine them with the known category data to obtain the virtual sample data set, retrain the optimal closed set recognition model, let the model learn the distribution of unknown data, and obtain a model that can perform open set recognition;

[0075] The open set recognition performance detection unit 64 is used to extract the features of the test set samples by retraining the closed set recognition model, and input them into the constructed OpenMax layer. The Weibull element recognition probability is calculated based on the distance between the feature vector of each signal sample and the center point of each category, and the probability estimate of each category is adjusted based on the weight calculated by the maximum activation class. The probability calculation of each sample is summed and normalized to obtain the probability that the sample belongs to each category. When the probability of belonging to the unknown class is the largest, OpenMax will reject these samples, and combine the maximum a posteriori probability (MAV) and threshold processing technology to realize open set recognition performance detection.

[0076] In some embodiments of the present application, see Figure 6 The closed set recognition network training unit 62 includes a signal sample extraction module 620, which is used to extract signal samples from a known sample set and an unknown sample set during each training, and is specifically used to:

[0077] In the stage of constructing a test data set for open set recognition performance testing, the number of known sample categories and the number of unknown sample categories during testing are determined according to the recognition task, where the number of samples in each category is the same. A set number of categories are randomly selected from the known sample set and the unknown sample set according to the openness determined during recognition to construct the test data set, and the number of known sample categories will not be less than the number of unknown sample categories.

[0078] In some embodiments of the present application, the signal sample set acquisition unit 61 preprocesses the received individual signal of the radiation source including:

[0079] The received individual signal of each radiation source is downsampled and discretized, and down-converted to a zero intermediate frequency to obtain a discrete complex baseband signal as a signal sample. The discrete complex baseband signal is expressed as:

[0080] y[n]=y I [n]+j*y Q [n]n∈[0,N p -1]

[0081] Here N p is the number of sampling points, y I and Qare in-phase and quadrature signals respectively, y[n] is a discrete complex signal, and then the signals of each individual are combined to construct a data set for closed set recognition. The data set is in the form of [x n ,y n ], where x represents the signal sample, y represents the corresponding label, n is the individual category of the radiation source, and each category contains the same number of samples.

[0082] In some embodiments of the present application, the closed set recognition network training unit 62 is specifically used to combine the dynamic convolution mechanism with the group equivariant convolution network structure to construct a dynamic group equivariant network. The dynamic convolution can increase the model capacity while maintaining efficient reasoning. First, the dynamic routing coefficients of different convolution kernels are calculated, and then combined into a dynamic convolution kernel with low computational overhead, which is specifically expressed as the following formula:

[0083]

[0084] In formula (1), represents the fully connected layer, is the input vector, σ is the activation function; W in formula (2) i represents a convolution kernel of fixed size, It is the dynamic routing coefficient obtained by the attention module in formula (1). Then, the known category data set obtained is divided into training set, validation set and test set. The training set is used to train the model, and the validation set is used to verify the model performance. Finally, the optimal model parameters are found through the recognition performance on the test set to obtain the closed set recognition model with the best recognition performance.

[0085] In some embodiments of the present application, the generative adversarial simulation retraining unit 63 is specifically used to build a generative adversarial network, simulate and generate a large number of virtual samples by adversarial training, and the loss function is set to

[0086]

[0087] where D(x) is the probability that the discriminator classifies the real data, and D(G(z)) is the probability that the discriminator classifies the generated data. data (x) represents the distribution of real data, p z (z) is the distribution of the noise vector input to the generator. The generated samples are combined with samples of known categories to obtain an open set recognition data set close to the true distribution. The generated samples are uniformly classified as unknown category samples and labeled with N+1 labels, where N is the number of known categories. The closed set recognition model is retrained based on the data set to learn the distribution of unknown category data and obtain a model that can be used for open set recognition.

[0088] In some embodiments of the present application, the open set recognition performance detection unit 64 is specifically used to treat the signal samples actually received outside the known category as unknown category samples, pre-process them and combine them with the known category data to obtain a test data set, and extract features through the retrained model.

[0089] F = DEGCONV OSR (x)(5)

[0090] DEGCONV in formula (5) OSR (·) represents the model for open set recognition obtained after retraining, and F is the sample feature extracted by the network;

[0091] The features are then fed into the OpenMax layer that uses samples and μ i The distance between the two is used to calculate the Weibull element recognition probability, where μ i It is the center point calculated using the relevant samples of the i-th class, and then the first α largest activation classes are selected. These classes are most likely to be the correct classes and the activation values ​​of each class are calculated to calculate the weights to adjust the probability estimate of each class. Based on the calculated activation values ​​and Weibull distribution, the cumulative distribution function (CDF) probability of each class is calculated, which helps to distinguish known classes from unknown classes. The original activation vector is then modified according to the probability. The new activation vector reflects the probability that the sample belongs to a known class, as shown in the following formula:

[0092]

[0093] In formula (6) It represents the modified activation value of the input sample at category j. The exponential sum of all categories is used as the denominator to ensure that the sum of the probabilities of all categories is 1, and the probability of the sample belonging to each category is obtained. When the probability of belonging to the unknown class is the largest, OpenMax will reject these samples. Finally, the maximum a posteriori probability (MAV) and threshold processing techniques are used to further enhance the detection of unknown samples, reject unknown samples in the presence of unknown samples, and identify the correct samples.

[0094] Can understand, Figure 6 The radiation source open set identification device based on dynamic group equivariant network and generative adversarial network shown in the figure can implement each step in the method of the aforementioned embodiment 1, and the relevant explanations of the method of embodiment 1 are applicable to Figure 6 The various units or modules in the device shown are not described in detail here. In addition, it should be noted that the names of the units or modules involved in the embodiments of the present application do not constitute limitations on the units or modules themselves in some cases.

[0095] Example 4

[0096] An embodiment of the present application also proposes an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method of the present application.

[0097] Reference below Figure 7 , which shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present application. The electronic devices in the embodiments of the present application may include but are not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 7 The electronic device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0098] like Figure 7 As shown, the electronic device may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage device into a random access memory (RAM). In the RAM, various programs and data required for the operation of the electronic device are also stored. The processing device, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0099] Typically, the following systems can be connected to the I / O interface: input devices including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices including, for example, a magnetic tape, a hard disk, etc.; and communication devices. The communication device can allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows an electronic device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or have alternatively.

[0100] In particular, according to an embodiment of the present application, the processes described in the flowcharts above can be implemented as computer software programs. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, the above-mentioned functions defined in the method of the embodiment of the present application are executed.

[0101] Finally, it is necessary to explain:

[0102] The serial numbers of the embodiments of the present application are for description only and do not represent the advantages and disadvantages of the embodiments. In the above embodiments of the present application, the descriptions of the various embodiments have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The embodiments of the present application can be implemented by hardware, software, firmware or a combination thereof.

[0103] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the system embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, units or modules, which can be electrical or other forms.

[0104] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0105] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the method, system and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0106] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. Moreover, if the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application.

Claims

1. A radiation source open set identification method based on dynamic group equivariant network and generative adversarial network, characterized in that: The steps include: S1, preprocessing the received individual signals of radiation sources of known categories to obtain a closed set identification sample set; the sample set includes signal samples of multiple categories, each category has the same number of signal samples, and each signal sample is marked with an identification category label; S2, build a dynamic group equivariant network, divide the known category data set obtained into training set, validation set and test set, use the training set to train the model, and use the validation set to verify the model performance. Finally, find the optimal model parameters through the performance on the test set to obtain the closed set recognition model with the best recognition performance; S3, build a generative adversarial network, train the known category data through adversarial network training, generate a large number of virtual samples, combine the generated samples with the known category samples, obtain an open set recognition data set close to the real distribution, and retrain the closed set recognition model based on the data set to learn the distribution of unknown category data and obtain a model that can be used for open set recognition; S4, treats the signal samples actually received outside the known category as unknown category samples, preprocesses them and combines them with known category data to obtain a test data set, extracts features through the retrained model, inputs them into the constructed OpenMax layer, and judges them as known category samples or unknown category samples based on the classification score and the set threshold, and then detects the open set recognition performance of the constructed model on the overall test sample set.

2. The radiation source open set identification method based on dynamic group equivariant network and generative adversarial network according to claim 1 is characterized in that: The steps of preprocessing the received individual radiation source signals in step S1 are as follows: The received individual signal of each radiation source is downsampled and discretized, and down-converted to a zero intermediate frequency to obtain a discrete complex baseband signal as a signal sample. The discrete complex baseband signal is expressed as: y[n]=y I [n]+j*y Q [n]n∈[0,N p -1] Here N p is the number of sampling points, y I and Q are in-phase and quadrature signals respectively, y[n] is a discrete complex signal, and then the signals of each individual are combined to construct a data set for closed set recognition. The data set is in the form of [x n ,y n ], where x represents the signal sample, y represents the corresponding label, n is the individual category of the radiation source, and each category contains the same number of samples.

3. The radiation source open set identification method based on dynamic group equivariant network and generative adversarial network according to claim 1 is characterized in that: In step S2, the optimal closed set recognition model is obtained by using the following method: Combining the dynamic convolution mechanism with the group equivariant convolution network structure to build a dynamic group equivariant network, dynamic convolution can increase the model capacity while maintaining efficient reasoning. First, the dynamic routing coefficients of different convolution kernels are calculated, and then combined into a dynamic convolution kernel with low computational overhead, which is specifically expressed as follows: In formula (1), represents the fully connected layer, is the input vector, σ is the activation function; W in formula (2) i represents a convolution kernel of fixed size, It is the dynamic routing coefficient obtained by the attention module in formula (1). Then, the known category data set obtained is divided into training set, validation set and test set. The training set is used to train the model, and the validation set is used to verify the model performance. Finally, the optimal model parameters are found through the recognition performance on the test set to obtain the closed set recognition model with the best recognition performance.

4. The radiation source open set identification method based on dynamic group equivariant network and generative adversarial network according to claim 3 is characterized in that: The step S3 comprises: Build a generative adversarial network and use adversarial training to simulate and generate a large number of virtual samples. The loss function is set to where D(x) is the probability that the discriminator classifies the real data, and D(G(z)) is the probability that the discriminator classifies the generated data; p data (x) represents the distribution of real data, p z (z) is the distribution of the noise vector input to the generator. The generated samples are combined with samples of known categories to obtain an open set recognition data set that is close to the true distribution. The generated samples are uniformly classified as unknown category samples and marked with N+1 labels, where N is the number of known categories. The closed set recognition model is retrained based on the data set to learn the distribution of unknown category data and obtain a model that can be used for open set recognition.

5. The radiation source open set identification method based on dynamic group equivariant network and generative adversarial network according to claim 4 is characterized in that: The step S4 comprises: The signal samples actually received outside the known category are regarded as samples of unknown category, which are preprocessed and combined with the known category data to obtain the test data set. The features are extracted through the retrained model. F=DEGCONV OSR (x)(5) DEGCONV in formula (5) OSR (·) represents the model for open set recognition obtained after retraining, and F is the sample feature extracted by the network; The above sample features are input to the built OpenMax layer, which uses samples and μ i The distance between the two is used to calculate the Weibull element recognition probability, where μ i It is the center point calculated using the i-th class related samples. Then the first α largest activation classes are selected and the activation value of each class is calculated to calculate the weight to adjust the probability estimate of each class. Based on the calculated activation value and Weibull distribution, the cumulative distribution function probability of each class is calculated to help distinguish known classes from unknown classes. The original activation vector is then modified according to the probability. The new activation vector reflects the probability that the sample belongs to a known class, as shown in the following formula: In formula (6) It represents the modified activation value of the input sample at category j. The exponential sum of all categories is used as the denominator to ensure that the sum of the probabilities of all categories is 1, and the probability of the sample belonging to each category is obtained. When the probability of belonging to the unknown class is the largest, OpenMax will reject these samples. Finally, the maximum a posteriori probability and threshold processing techniques are used to further enhance the detection of unknown samples, reject unknown samples in the presence of unknown samples, and identify the correct samples.

6. A radiation source open set identification device based on dynamic group equivariant network and generative adversarial network, characterized in that: The device comprises: The signal sample set acquisition unit is used to perform corresponding preprocessing on the received individual signals of multiple known categories of radiation sources to construct a data set for closed set recognition; the signal sample set includes signal samples of multiple categories, each category has the same number of signal samples and is marked with category labels; in addition, a number of unknown signal samples other than known category samples are received, corresponding preprocessing is completed, N+1 labels are marked, and they are combined with known category samples into a test data set for open set recognition; The closed set recognition network training unit is used to construct a dynamic group equivariant network, divide the constructed closed set recognition data set into a training set, a validation set and a test set, input the training set into the constructed network, extract features, obtain the predicted label through the classification layer, compare it with the real label and calculate the loss, and adjust the model parameters to minimize the model training loss and obtain the optimal closed set recognition model; Generate adversarial simulation retraining unit, used to build a generative adversarial network, input the constructed known category data set into the constructed network, use the idea of ​​adversarial training to minimize the loss between the simulated virtual samples and the real samples, to imitate unknown samples, and combine them with known category data to obtain a virtual sample data set, retrain the optimal closed set recognition model, let the model learn the distribution of unknown data, and obtain a model that can perform open set recognition; The open set recognition performance detection unit is used to extract the features of the test set samples using the retrained closed set recognition model, and input them into the built OpenMax layer. The Weibull element recognition probability is calculated based on the distance between the feature vector of each signal sample and the center point of each category, and the probability estimate of each class is adjusted based on the weight calculated by the maximum activation class. The probability calculation of each sample is summed and normalized to obtain the probability that the sample belongs to each category. When the probability of belonging to the unknown class is the largest, OpenMax will reject these samples. The maximum a posteriori probability and threshold processing technology are combined to realize open set recognition performance detection.

7. The radiation source open set identification device based on dynamic group equivariant network and generative adversarial network according to claim 6 is characterized in that: The signal sample set acquisition unit includes a signal sample extraction module, which is used to extract signal samples from a known sample set and an unknown sample set during each training, and is specifically used to: In the stage of constructing a test data set for open set recognition performance testing, the number of known sample categories and the number of unknown sample categories during testing are determined according to the recognition task, where the number of samples in each category is the same. A set number of categories are randomly selected from the known sample set and the unknown sample set according to the openness determined during recognition to construct the test data set, and the number of known sample categories will not be less than the number of unknown sample categories.

Images