Long-tail radiation source individual identification method and device based on double-branch feature learning

Abstract

The invention relates to the technical field of radiation source individual identification, and provides a dual-branch feature learning-based long-tail radiation source individual identification method and device, and the method comprises the steps: obtaining a sample set which comprises a training set and a test set, the training set comprises a plurality of training long-tail radiation source signals, and the test set comprises a plurality of test long-tail radiation source signals; the training long-tail radiation source signals are any one of preset head type data, preset middle type data and preset tail type data, inputting the plurality of training long-tail radiation source signals into a long-tail radiation source individual identification model for training, and adjusting weight parameters of the model according to a preset loss function until the model converges, and obtaining the trained individual identification model of the long-tail radiation source. And inputting the plurality of test long-tail radiation source signals into the trained long-tail radiation source individual identification model, and obtaining an individual identification result of each test long-tail radiation source signal. The accuracy of individual identification of the long-tail radiation source can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source individual identification, and in particular to a method and device for long-tail radiation source individual identification based on dual-branch feature learning. Background Art

[0002] Specific Emitter Identification (SEI) technology involves analyzing the unique radio frequency characteristics of communications devices when transmitting radio signals to accurately identify the device emitting them. This allows for the authentication of communications devices without relying on traditional encryption and authentication, preventing unauthorized access and disguised communications, and effectively enhancing system security and reliability. Currently, with the advancement of artificial intelligence and deep learning technologies, SEI recognition accuracy is continuously improving, and it will play an increasingly important role in fields such as communications security and electronic countermeasures.

[0003] In the existing technology, on the one hand, the individual identification of radiation sources can be achieved through traditional radiation source individual identification technologies such as the radiation source individual identification method based on signal parameter feature extraction and the radiation source individual identification method based on statistical feature extraction in the transform domain. On the other hand, the individual identification of radiation sources can be achieved through the radiation source individual identification method based on the deep learning model.

[0004] However, using existing technologies, traditional radiation source individual identification technology is sensitive to noise interference and cannot effectively obtain feature information of unknown communication devices. This leads to poor robustness and low accuracy in radiation source individual identification. In contrast, radiation source individual identification methods based on deep learning models, when there is an imbalance in the number of radiation source individual categories, such as in long-tail distribution data scenarios, are more dominated by head-type data during the training process based on deep learning models. As a result, the feature information learned by the model focuses more on the head-type data corresponding to the head-type radiation source individuals, and cannot effectively learn the features of the head-type data corresponding to the tail-type radiation source individuals, resulting in low accuracy in radiation source individual identification. Summary of the Invention

[0005] Based on this, it is necessary to provide a long-tail radiation source individual identification method and device based on dual-branch feature learning to address the above technical problems.

[0006] In a first aspect, an embodiment of the present invention provides a method for identifying individual long-tail radiation sources based on dual-branch feature learning, the method comprising:

[0007] Acquire a sample set, the sample set including a training set and a test set, the training set including multiple training long-tail radiation source signals, the test set including multiple test long-tail radiation source signals, the training long-tail radiation source signals being any one of preset head class data, preset middle class data, and preset tail class data;

[0008] Inputting the plurality of training long-tail radiation source signals into a long-tail radiation source individual recognition model for training, and adjusting the weight parameters of the model according to a preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual recognition model;

[0009] Inputting the plurality of test long-tail radiation source signals into the trained long-tail radiation source individual recognition model to obtain an individual recognition result for each of the test long-tail radiation source signals;

[0010] In which, the long-tail radiation source individual identification model includes: a feature transfer learning branch, a multi-expert nested learning branch, and an individual identification module; the feature transfer learning branch is connected in parallel to the multi-expert nested learning branch, and the feature transfer learning branch is used to obtain a first individual identification result based on the data distribution of the training long-tail radiation source signal, and the data distribution is the first data distribution of the preset head class data, or the second data distribution based on the first data distribution transfer learning, the multi-expert nested learning branch is used to obtain a second individual identification result based on multiple preset features corresponding to different expert features, and the individual identification module is used to obtain the individual identification result based on the first individual identification result and the second individual identification result.

[0011] In a second aspect, an embodiment of the present invention provides a long-tail radiation source individual identification device based on dual-branch feature learning, comprising:

[0012] A sample set acquisition module is used to acquire a sample set, wherein the sample set includes a training set and a test set, wherein the training set includes multiple training long-tail radiation source signals, and the test set includes multiple test long-tail radiation source signals, wherein the training long-tail radiation source signal is any one of preset head class data, preset middle class data, and preset tail class data;

[0013] A training module is used to input the multiple training long-tail radiation source signals into the long-tail radiation source individual recognition model for training, and adjust the weight parameters of the model according to a preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual recognition model;

[0014] A testing module, configured to input the plurality of test long-tail radiation source signals into the trained long-tail radiation source individual recognition model to obtain an individual recognition result of each of the test long-tail radiation source signals;

[0015] In which, the long-tail radiation source individual identification model includes: a feature transfer learning branch, a multi-expert nested learning branch, and an individual identification module; the feature transfer learning branch is connected in parallel to the multi-expert nested learning branch, and the feature transfer learning branch is used to obtain a first individual identification result based on the data distribution of the training long-tail radiation source signal, and the data distribution is the first data distribution of the preset head class data, or the second data distribution based on the first data distribution transfer learning, the multi-expert nested learning branch is used to obtain a second individual identification result based on multiple preset features corresponding to different expert features, and the individual identification module is used to obtain the individual identification result based on the first individual identification result and the second individual identification result.

[0016] The technical solution provided by the embodiment of the present invention has the following advantages compared with the existing technology:

[0017] An embodiment of the present invention provides a method for identifying long-tail radiation source individuals based on dual-branch feature learning. The method comprises obtaining a sample set, the sample set comprising a training set and a test set, the training set comprising multiple training long-tail radiation source signals, the test set comprising multiple test long-tail radiation source signals, the training long-tail radiation source signals being any one of preset head class data, preset middle class data, and preset tail class data, inputting the multiple training long-tail radiation source signals into a long-tail radiation source individual identification model for training, and adjusting the weight parameters of the model according to a preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual identification model. Multiple test long-tail radiation source signals are input into the trained long-tail radiation source individual identification model to obtain an individual identification result for each test long-tail radiation source signal. The long-tail radiation source individual identification model includes: a feature transfer learning branch, a multi-expert nested learning branch, and an individual identification module. The feature transfer learning branch is used to obtain a first individual identification result based on the data distribution of the training long-tail radiation source signal. The data distribution is a first data distribution of the preset head class data, or a second data distribution based on the first data distribution transfer learning. This avoids the situation in the prior art where, during the training process based on the deep learning model, the model is more dominated by the preset head class data, resulting in the model being unable to perform feature learning on the preset tail class data. The multi-expert nested learning branch is used to obtain a second individual identification result based on multiple preset features corresponding to different expert features, thereby increasing attention to the preset tail class features. The individual identification module is used to obtain an individual identification result based on the fusion of the first individual identification result and the second individual identification result. Based on this, the individual identification result is obtained through the long-tail radiation source individual identification model, which can improve the accuracy of long-tail radiation source individual identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0020] Figure 1 A schematic diagram of a flow chart of a method for identifying individual long-tail radiation sources based on dual-branch feature learning provided by an embodiment of the present invention;

[0021] Figure 2 A schematic diagram of the structure of a long-tail radiation source individual identification model provided by an embodiment of the present invention;

[0022] Figure 3 A schematic diagram of a model test accuracy provided by an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of a training feature provided by an embodiment of the present invention;

[0024] Figure 5 Schematic diagram of comparison results between different ablation experiments provided in embodiments of the present invention and the present invention;

[0025] Figure 6 A schematic diagram showing the comparison results between the radiation source individual identification method based on the deep learning model provided in an embodiment of the present invention and the present invention;

[0026] Figure 7 A schematic diagram of a long-tail radiation source individual identification device based on dual-branch feature learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the scheme of the present invention will be further described below. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features therein can be combined with each other.

[0028] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present invention, rather than all the embodiments.

[0029] In one embodiment, Figure 1 As shown, Figure 1A schematic flow chart of a method and apparatus for identifying individual long-tail radiation sources based on dual-branch feature learning provided by an embodiment of the present invention includes the following steps:

[0030] S10: Obtain a sample set.

[0031] The sample set includes a training set and a test set. The training set includes multiple training long-tail radiation source signals, and the test set includes multiple test long-tail radiation source signals. The training long-tail radiation source signals are any one of preset head class data, preset middle class data, and preset tail class data. The preset head class data, preset middle class data, and preset tail class data are determined based on multiple discrete sampled data included in each training long-tail radiation source signal in a long-tail distribution data distribution scenario.

[0032] Optionally, based on the above embodiments, in some embodiments of the present invention, a method for determining the preset head class data, preset middle class data and preset tail class data may be: by pre-setting a preset threshold, comparing the number of multiple discrete sampling data included in the training long-tail radiation source signal with the size of the preset threshold, so as to determine that the radiation source signal is one of the preset head class data, preset middle class data and preset tail class data.

[0033] Exemplarily, the preset thresholds are T1 and T2, and the number of multiple discrete sampling data included in the training long-tail radiation source signal is t. When t is greater than T1, the radiation source signal is the preset head class data, when t is greater than T2 and less than T1, the radiation source signal is the preset middle class data, and when t is less than T2, the radiation source signal is the preset tail class data, but is not limited to this. The present invention is not specifically limited, and those skilled in the art can set it according to actual conditions.

[0034] It should be noted that the training long-tail radiation source signal and the test long-tail radiation source signal need to be preprocessed. Specifically, the in-phase and quadrature components of the training long-tail radiation source signal and the test long-tail radiation source signal are obtained, and the in-phase and quadrature components are normalized to facilitate subsequent training and testing of the long-tail radiation source individual recognition model. Exemplarily, the training long-tail radiation source signal x(n) is preprocessed to obtain the in-phase and quadrature components of the training long-tail radiation source signal as follows:

[0035] I(n)=real(x(n))

[0036] Q(n)=imag(x(n))

[0037] Specifically, a sample set for training a long-tail radiation source individual recognition model is obtained, and the sample set includes a training set and a test set. The training set includes multiple training long-tail radiation source signals, and the test set includes multiple test long-tail radiation source signals. The training long-tail radiation source signal is any one of the preset head class data, preset middle class data, and preset tail class data.

[0038] Exemplarily, the sample set can be collected by a fixed receiving end, such as a USRP B210. Specifically, the fixed receiving end receives airborne radio signals transmitted by 16 similarly configured USRP X310 transmitting end devices at a sampling rate of 5 MHz, wherein all transmitting end devices are consistent in hardware composition, communication protocol, physical address, and MAC address. The airborne radio signals contain a randomly generated payload and a unified address field and are transmitted over the air within the 2.45 GHz radio frequency band. For the signals corresponding to the 16 transmitting end devices, 10 signals are randomly selected as training long-tail radiation source signals, that is, the sample set includes training long-tail radiation source signals of ten individual categories, each category includes 10,000 training long-tail radiation source signals, and the 10,000 training long-tail radiation source signals can be configured according to a preset exponential function. The preset exponential function can be defined according to the following expression:

[0039]

[0040] Among them, n max Indicates the maximum number of samples, n i represents the number of training long-tail radiator signals of the i-th category, N represents the total number of training long-tail radiator signals, and μ represents the long-tail factor, which is used to describe the long-tail distribution of each category of training long-tail radiator signals. For example, μ can be 0.05.

[0041] S11: Input multiple training long-tail radiation source signals into the long-tail radiation source individual recognition model for training, and adjust the weight parameters of the model according to the preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual recognition model.

[0042] Among them, the preset loss function is determined by weighting the first loss function corresponding to the feature transfer learning branch and the second loss function corresponding to the multi-expert nested learning branch.

[0043] Optionally, based on the above embodiment, in some embodiments of the present invention, the first loss function is defined by the following expression:

[0044]

[0045] Among them, N represents the batch size of input training samples, C represents the number of input training sample categories, Represents the label of the nth training long-tail radiation source signal; represents the angle between the feature vector and the category weight vector corresponding to the nth training long-tail radiation source signal; S represents the scaling factor, m c represents the hyperparameters of the first loss function.

[0046] The second loss function is defined by the following expression:

[0047]

[0048] Among them, N h+m+t Represents all training long-tail radiation source signals of the same batch input, N m+t Indicates the training long-tail radiant source signals of the same batch input, which are the preset intermediate class data and the preset tail class data, N t represents the training long-tail radiator signal of the preset tail class data in the training long-tail radiator signal of the same batch input, x represents the training long-tail radiator signal, y represents the label of the training long-tail radiator signal, represents the cross entropy loss function, represents the hyperparameter, Indicates N h+m+t 、N m+t and N t The fusion loss function that performs loss calculation at the same time, Indicates N h+m+t 、N m+t and N t Branch loss functions for separate loss calculations.

[0049] The default loss function is defined by the following expression:

[0050]

[0051] Where α represents the weight coefficient, which is determined by the number of training iterations and is used to change the learning capabilities of the feature transfer learning branch and the expert nested learning branch during model training. In other words, as the number of training iterations increases, the focus of learning can gradually shift from the feature transfer learning branch to the expert nested learning branch, further improving the accuracy of individual long-tail emitter identification. The weight coefficient can be defined according to the following expression:

[0052]

[0053] Among them, T max Represents the total number of training iterations, and T represents the current number of training iterations.

[0054] Specifically, for the initial long-tail radiation source individual recognition model, multiple training long-tail radiation source signals are input into the long-tail radiation source individual recognition model for training. During the training process, the weight parameters of the model are adjusted according to the preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual recognition model.

[0055] Optionally, based on the above embodiments, in some embodiments of the present invention, Figure 2 This figure shows the structure of a long-tail radiator individual identification model provided by an embodiment of the present invention. The long-tail radiator individual identification model includes a feature transfer learning branch 11, a multi-expert nested learning branch 12, and an individual identification module 13. The feature transfer learning branch 11 and the multi-expert nested learning branch 12 are connected in parallel.

[0056] Among them, the feature transfer learning branch 11 is used to obtain the first individual recognition result based on the data distribution of the training long-tail radiation source signal. The data distribution is the first data distribution of the preset head class data, or the second data distribution based on the first data distribution transfer learning. In this way, it can avoid the existing situation in the prior art that in the training process based on the deep learning model, because the model is dominated by the preset head class data, the feature information learned by the model pays more attention to the preset head class data, and cannot perform feature learning on the preset tail class data well, thereby improving the accuracy of individual recognition of the long-tail radiation source.

[0057] The multi-expert nested learning branch 12 is used to obtain the second individual recognition result according to multiple preset features corresponding to different expert features, where the preset feature is the first preset feature f h+m+t , the second preset feature f m+t , and the third preset feature f t Any one of the first preset features includes: head class features, middle class features and tail class features, the second preset features includes: middle class features and tail class features, and the third preset features includes: tail class features. In this way, the preset features can be extracted according to multiple different expert features, thereby increasing the attention to the preset tail class features, thereby improving the accuracy of individual identification of long-tail radiation sources.

[0058] The individual identification module 13 is used to obtain the individual identification result based on the first individual identification result and the second individual identification result. In this way, the first individual identification result and the second individual identification result extracted by the two branches can be fused to obtain the final individual identification result, thereby improving the accuracy of individual identification of long-tail radiation sources.

[0059] Based on this, one implementation of S11 may be:

[0060] S111: Input multiple training long-tail radiation source signals into the feature transfer learning branch, and obtain the first individual recognition result according to the data distribution corresponding to each training long-tail radiation source signal.

[0061] Specifically, multiple training long-tail radiation source signals are input into the feature transfer learning branch, and feature transfer learning is performed on the multiple training long-tail radiation source signals through the feature transfer learning branch to obtain the data distribution corresponding to the long-tail radiation source signals, and the first individual recognition result is obtained according to the data distribution.

[0062] Optionally, based on the above embodiments, in some embodiments of the present invention, continue to refer to Figure 2 As shown, the feature transfer learning branch 11 includes: a feature extraction module, a feature transfer learning module and a first individual recognition module. Based on this, one implementation of S111 can be:

[0063] S1111: Input multiple training long-tail radiation source signals into a feature extraction module for feature extraction to obtain initial features corresponding to the training long-tail radiation source signals.

[0064] Specifically, a plurality of training long-tail radiation source signals are input into a feature extraction module, and feature extraction is performed on the plurality of training long-tail radiation source signals by the feature extraction module to obtain initial features corresponding to each training long-tail radiation source signal.

[0065] It should be noted that the feature extraction module includes: one-dimensional convolution layer, batch normalization layer, ReLU layer, and maximum pooling layer.

[0066] S1112: Input the initial features into the feature transfer learning module to perform feature transfer learning to obtain the data distribution of the initial features.

[0067] Specifically, multiple initial features are input into the feature transfer learning module, and the feature transfer learning module performs transfer learning on the initial features to obtain the data distribution corresponding to each initial feature.

[0068] Optionally, based on the above embodiments, in some embodiments of the present invention, continue to refer to Figure 2 As shown, one implementation of S1112 may be:

[0069] S21: Determine the target type data of the initial feature according to the number of discrete sampling data corresponding to the training long-tail radiation source signal and a preset threshold.

[0070] The target type data is any one of preset header type data and preset tail type data.

[0071] Specifically, the number of discrete sampling data corresponding to the training long-tail radiation source signal is compared with a preset threshold to determine the target type data of the initial feature corresponding to the training long-tail radiation source signal.

[0072] Exemplarily, when the number of discrete sampling data is greater than a preset threshold, the target type data of the initial feature is determined to be the preset head type data; otherwise, the target type data of the initial feature is determined to be the preset tail type data.

[0073] S22: Determine data distribution according to the target type data and the first data distribution.

[0074] The first data distribution is determined by performing mean calculation on multiple data distributions corresponding to preset head class data in multiple training long-tail radiation source signals.

[0075] Optionally, based on the above embodiment, in some embodiments of the present invention, since the target type data is any one of the preset header type data and the preset tail type data, based on this, one implementation of S22 may be:

[0076] S221: When it is determined that the target type data is preset header type data, the first data distribution is determined as the data distribution.

[0077] Specifically, after determining that the target type data of the initial feature is the preset header type data, the first data distribution corresponding to the preset header type data is determined as the data distribution.

[0078] Optionally, based on the above embodiment, in some embodiments of the present invention, another implementation of S22 may be:

[0079] S222: When it is determined that the target type data is the preset tail type data, transfer learning is performed according to the first data distribution to obtain the individual data distribution.

[0080] Specifically, after determining that the target type data of the initial feature is the preset tail class data, transfer learning is performed according to the first data distribution corresponding to the preset head class data, and the second data distribution obtained is the data distribution.

[0081] For example, following the above embodiment, we continue to refer to Figure 2 As shown, for the first data distribution For the preset tail class data, calculate and The difference between The data distribution difference corresponding to the distribution angle of the preset tail class data and the preset head class data is Further transfer learning is performed to obtain the second data distribution: That is, the data distribution is

[0082] S1113: Input the data distribution into the first individual recognition module for individual recognition, and obtain a first individual recognition result.

[0083] Specifically, the obtained data distribution is input into a first individual recognition module, and the first individual recognition module performs individual recognition on the data distribution to obtain a first individual recognition result.

[0084] In this way, this embodiment obtains the first individual recognition result corresponding to the training long-tail radiation source signal through the feature transfer learning branch including: a feature extraction module, a feature transfer learning module and a first individual recognition module, so as to avoid the situation in the existing technology where, during the training process based on the deep learning model, the model is dominated by the preset head class data, resulting in the feature information learned by the model paying more attention to the preset head class data and failing to perform feature learning on the preset tail class data, thereby improving the accuracy of individual recognition of the long-tail radiation source.

[0085] S112: Inputting a plurality of training long-tail radiation source signals into a multi-expert nested learning branch, and obtaining a second individual recognition result according to a plurality of preset features corresponding to different expert features of the training long-tail radiation source signals.

[0086] Optionally, based on the above embodiments, in some embodiments of the present invention, continue to refer to Figure 2 As shown, the multi-expert nested learning branch 12 includes: a feature extraction module, multiple parallel-connected expert feature extraction modules, multiple residual fusion modules, and a second individual recognition module. The expert feature extraction modules correspond to the residual fusion modules one-to-one. Based on this, one implementation of S112 may be:

[0087] S1121: Inputting multiple training long-tail radiation source signals into a feature extraction module for feature extraction to obtain initial features corresponding to the training long-tail radiation source signals.

[0088] Specifically, a plurality of training long-tail radiation source signals are input into a feature extraction module, and feature extraction is performed on the plurality of training long-tail radiation source signals by the feature extraction module to obtain initial features corresponding to each training long-tail radiation source signal.

[0089] It should be noted that the feature extraction module includes: one-dimensional convolution layer, batch normalization layer, ReLU layer, and maximum pooling layer.

[0090] S1122: Inputting the initial features into a plurality of parallel-connected expert feature extraction modules to perform expert feature extraction, and obtaining a plurality of different expert features of the initial features.

[0091] The expert feature extraction module refers to the different features obtained based on the experience of different experts. The expert feature extraction module can perform feature learning through a shared and publicly available backbone network, so that the long-tail radiator individual identification model can be trained and learned on different expert feature extraction modules. Because each different expert feature extraction module focuses on different feature information of the long-tail radiator signal, it can obtain more rich feature information, improve the robustness of the long-tail radiator individual identification model and improve the accuracy of long-tail radiator individual identification.

[0092] Specifically, after the initial features are obtained, the initial features are respectively input into different expert feature extraction modules, and the initial features are subjected to feature extraction by the multiple expert feature extraction modules to obtain multiple different expert features of the initial features.

[0093] S1123: Inputting the multiple different expert features into the corresponding residual fusion module respectively to extract multiple preset features, obtain multiple preset features and perform fusion processing to obtain residual fusion features.

[0094] Among them, the preset feature is any one of the first preset feature, the second preset feature, and the third preset feature. The first preset feature includes: head class feature, middle class feature and tail class feature. The second preset feature includes: middle class feature and tail class feature. The third preset feature includes: tail class feature.

[0095] Specifically, the obtained multiple different expert features are respectively input into the corresponding residual fusion modules, and multiple preset features are extracted from the input expert features through each residual fusion module to obtain multiple preset features corresponding to each expert feature, and the multiple preset features are fused to obtain the residual fusion feature corresponding to each expert feature.

[0096] S1124: Input the residual fusion feature into the second individual recognition module for individual recognition to obtain a second individual recognition result.

[0097] Specifically, the obtained residual fusion features are input into the second individual recognition module, and the second individual recognition module performs individual recognition on the input residual fusion features to obtain a second individual recognition result.

[0098] For example, for multiple preset features such as out1, out2, out3...out n , multiple multi-level features are fused to obtain residual fusion features such as out a , score the fusion features through the activation function Softmax to obtain the second individual recognition result.

[0099] In this way, this embodiment can extract multiple preset features based on multiple different expert features through the multi-expert nested learning branch, thereby increasing the focus on the preset tail class features, thereby improving the accuracy of individual identification of long-tail radiation sources.

[0100] S113: Input the first individual recognition result and the second individual recognition result into the individual recognition module to obtain the predicted individual recognition result corresponding to the training long-tail radiation source signal, adjust the weight parameters of the model according to the predicted individual recognition result, individual label, and preset loss function until the model converges, and obtain the trained long-tail radiation source individual recognition model.

[0101] Specifically, the first individual recognition result and the second individual recognition result are input into the individual recognition module, and the predicted individual recognition result corresponding to the training long-tail radiation source signal is obtained through the individual recognition module. The weight parameters of the model are further adjusted according to the predicted individual recognition result, individual label, and preset loss function until the model converges, and the trained long-tail radiation source individual recognition model is obtained.

[0102] Optionally, based on the above embodiments, in some embodiments of the present invention, continue to refer to Figure 2 As shown, the first individual recognition result and the second individual recognition result are input into the individual recognition module to obtain the predicted individual recognition result corresponding to the training long-tail radiation source signal. One implementation method may be:

[0103] S1131: Perform weighted calculation on the first individual recognition result and the second individual recognition result to obtain a predicted individual recognition result.

[0104] Specifically, the first individual recognition result and the second individual recognition result are input into the individual recognition module, and the individual recognition module performs weighted calculation on the first individual recognition result and the second individual recognition result using a preset weighting formula to obtain a predicted individual recognition result.

[0105] Optionally, the preset weighting formula may be defined by the following expression:

[0106] y=∝y1+(1-∝)y2

[0107] Among them, ∝ represents the weight coefficient.

[0108] S12: Inputting multiple test long-tail radiation source signals into the trained long-tail radiation source individual recognition model to obtain an individual recognition result of each test long-tail radiation source signal.

[0109] Among them, the individual identification result refers to the individual communication device that emits the test long-tail radiation source signal.

[0110] Specifically, after obtaining the trained long-tail radiation source individual recognition model, multiple test long-tail radiation source signals are used for testing, and the multiple test long-tail radiation source signals are input into the trained long-tail radiation source individual recognition model to obtain the individual recognition result of each test long-tail radiation source signal.

[0111] Thus, this embodiment obtains a sample set, which includes a training set and a test set. The training set includes multiple training long-tail radiation source signals, and the test set includes multiple test long-tail radiation source signals. The training long-tail radiation source signals are any one of preset head class data, preset middle class data, and preset tail class data. The multiple training long-tail radiation source signals are input into a long-tail radiation source individual recognition model for training, and the weight parameters of the model are adjusted according to a preset loss function until the model converges. The trained long-tail radiation source individual recognition model is obtained. Multiple test long-tail radiation source signals are input into the trained long-tail radiation source individual recognition model to obtain an individual recognition result for each test long-tail radiation source signal. The long-tail radiation source individual identification model includes: a feature transfer learning branch, a multi-expert nested learning branch, and an individual identification module. The feature transfer learning branch is used to obtain a first individual identification result based on the data distribution of the training long-tail radiation source signal. The data distribution is a first data distribution of the preset head class data, or a second data distribution based on the first data distribution transfer learning. This avoids the situation in the prior art where, during the training process based on the deep learning model, the model is dominated by the preset head class data, resulting in the model being unable to perform feature learning on the preset tail class data. The multi-expert nested learning branch is used to obtain a second individual identification result based on multiple preset features corresponding to different expert features, thereby increasing attention to the preset tail class features. The individual identification module is used to obtain an individual identification result based on the fusion of the first individual identification result and the second individual identification result. Based on this, the individual identification result is obtained through the long-tail radiation source individual identification model, which can improve the accuracy of long-tail radiation source individual identification.

[0112] Optionally, based on the above embodiment, in some embodiments of the present invention, the first data distribution is determined by calculating the mean of multiple data distributions corresponding to preset head-type data in multiple training long-tail radiation source signals. One implementation method may be:

[0113] S31: Determine a plurality of preset head class data from a plurality of training long-tail radiation source signals.

[0114] S32: Obtain distribution angles corresponding to a plurality of discrete sampling data included in each preset header type data.

[0115] Specifically, for a plurality of training long-tail radiation source signals, a plurality of preset head class data included in the plurality of training long-tail radiation source signals are determined, and distribution angles corresponding to a plurality of discrete sampling data included in each preset head class data are obtained.

[0116] Optionally, the distribution angle may be obtained according to a preset angle calculation formula, which may be defined by the following expression:

[0117]

[0118] Among them, c i represents the center point of the i-th type of training long-tail radiation source signal, f i k Represents the features of the kth discrete sampling data of the i-th type of training long-tail radiation source signal.

[0119] It should be noted that for each type of training long-tail radiation source signal, the center point of the current iterative training can be updated using a preset update formula based on the center point of the previous iterative training during the model training process. The preset update formula can be defined by the following expression:

[0120]

[0121] Among them, l represents the current l-th iteration training, and γ represents the update factor.

[0122] S33: Perform mean calculation and variance calculation on multiple distribution angles to obtain the initial mean and initial variance of each preset head class data.

[0123] Specifically, after obtaining multiple distribution angles of each preset head class data, mean calculation and variance calculation are performed on the multiple distribution angles to obtain an initial mean and an initial variance.

[0124] It should be noted that, by pre-setting an angle memory, the distribution angles corresponding to the plurality of discrete sampling data included in each preset head type data are stored.

[0125] S34: performing mean calculation on initial means and initial variances corresponding to a plurality of preset head class data to obtain a target mean and a target variance.

[0126] S35: Determine a first data distribution based on the target mean and the target variance.

[0127] Specifically, after obtaining initial means and initial variances corresponding to a plurality of preset head class data, the plurality of initial means and initial variances are averaged to obtain a target mean and a target variance, and the first data distribution is determined according to the target mean and the target variance.

[0128] For example, the initial mean and initial variance of the distribution angle of the i-th type of preset head data are obtained as μ i 、 Calculate the mean of multiple initial means and initial variances to obtain the target mean and target variance μ h 、 Then the first data distribution obeys

[0129] Specifically, after training the target type data of the initial features corresponding to the long-tail radiation source signal, the data distribution is determined according to the target type data and the first data distribution.

[0130] Optionally, based on the above embodiments, in some embodiments of the present invention, in order to verify that the present invention can improve the accuracy of individual identification of long-tail radiation sources, Figure 3 A schematic diagram of the model test accuracy provided for an embodiment of the present invention shows that with the increase in the number of training iterations, the model training accuracy is basically stable after 40 rounds of training, indicating that the long-tail radiation source individual recognition model training has been completed and has good fitting ability. The long-tail radiation source individual recognition model can accurately obtain individual recognition results. Figure 4 A schematic diagram of a training feature provided for an embodiment of the present invention shows that the feature distributions of individual radiation sources of different types have clear feature boundaries, and the feature distribution of the preset tail class data can be well distinguished from the feature distribution of the preset head class data.

[0131] Optionally, based on the above embodiments, in some embodiments of the present invention, in order to verify that the present invention can improve the accuracy of individual identification of long-tail radiation sources, different ablation experiments are set up for comparison with the present invention. The ablation experiment settings are shown in Table 1 below. The experimental results are shown in Table 1 below. Figure 5 As shown, under the same conditions, the present invention can obtain more accurate individual recognition results than different ablation experiments.

[0132] Table 1 Ablation experiment settings

[0133]

[0134] Optionally, based on the above embodiments, in some embodiments of the present invention, in order to further demonstrate the effect of the present invention, the present invention is compared with existing radiation source individual identification methods based on deep learning models such as ResNet, BBN, FocalLoss, LDAM Loss, and CB Loss. Figure 6 As shown, compared with other long-tail radiation source individual identification methods based on deep learning models, the present invention can improve the accuracy of long-tail radiation source individual identification.

[0135] It should be understood that although Figures 1 to 6 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figures 1 to 6 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The order of execution of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0136] In one embodiment, Figure 7 The figure shows a schematic diagram of a long-tail radiator individual identification device based on dual-branch feature learning, comprising a sample set acquisition module, a training module, and a test module. The sample set acquisition module is configured to acquire a sample set, which includes a training set and a test set. The training set includes multiple training long-tail radiator signals, and the test set includes multiple test long-tail radiator signals. The training long-tail radiator signals are any of preset head-type data, preset middle-type data, and preset tail-type data.

[0137] The training module is used to input multiple training long-tail radiation source signals into the long-tail radiation source individual recognition model for training, and adjust the weight parameters of the model according to the preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual recognition model.

[0138] A testing module is used to input multiple test long-tail radiation source signals into a trained long-tail radiation source individual identification model to obtain individual identification results for each test long-tail radiation source signal; wherein, the long-tail radiation source individual identification model includes: a feature transfer learning branch, a multi-expert nested learning branch, and an individual identification module; the feature transfer learning branch is connected in parallel with the multi-expert nested learning branch, the feature transfer learning branch is used to obtain a first individual identification result based on the data distribution of the training long-tail radiation source signal, the data distribution is a first data distribution of a preset head class data, or a second data distribution based on transfer learning of the first data distribution, the multi-expert nested learning branch is used to obtain a second individual identification result based on multiple preset features corresponding to different expert features, and the individual identification module is used to obtain an individual identification result based on the first individual identification result and the second individual identification result.

[0139] Thus, this embodiment uses a training module to input multiple training long-tail radiation source signals into a long-tail radiation source individual recognition model for training, and adjusts the model's weight parameters according to a preset loss function until the model converges, thereby obtaining a trained long-tail radiation source individual recognition model. A testing module then inputs multiple test long-tail radiation source signals into the trained long-tail radiation source individual recognition model to obtain an individual recognition result for each test long-tail radiation source signal. The long-tail radiation source individual recognition model includes a feature transfer learning branch, a multi-expert nested learning branch, and an individual recognition module. The feature transfer learning branch is used to obtain a first individual recognition result based on the data distribution of the training long-tail radiation source signals. The data distribution is a first data distribution of a preset head-class data, or a second data distribution learned based on the first data distribution. This avoids the situation in the prior art where, during the training process of deep learning models, the model is dominated by the preset head-class data, resulting in the model being unable to effectively learn features for the preset tail-class data. The multi-expert nested learning branch is used to obtain a second individual recognition result based on multiple preset features corresponding to different expert features, thereby increasing the focus on the preset tail-class features. The individual identification module is used to obtain the individual identification result by fusing the first individual identification result and the second individual identification result. Based on this, the individual identification result is obtained through the long-tail radiation source individual identification model, which can improve the accuracy of long-tail radiation source individual identification.

[0140] Regarding the specific limitations of the long-tail radiation source individual identification device based on dual-branch feature learning, please refer to the limitations of the long-tail radiation source individual identification method based on dual-branch feature learning above, which will not be repeated here. Each module in the above-mentioned server can be implemented in whole or in part by software, hardware, and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.

[0141] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static random access memory (SRAM) and dynamic random access memory (DRAM).

[0142] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0143] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A method for identifying individual long-tail radiation sources based on dual-branch feature learning, characterized in that: include: Acquire a sample set, the sample set including a training set and a test set, the training set including multiple training long-tail radiation source signals, the test set including multiple test long-tail radiation source signals, the training long-tail radiation source signals being any one of preset head class data, preset middle class data, and preset tail class data; Inputting the plurality of training long-tail radiation source signals into a long-tail radiation source individual recognition model for training, and adjusting the weight parameters of the model according to a preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual recognition model; Inputting the plurality of test long-tail radiation source signals into the trained long-tail radiation source individual recognition model to obtain an individual recognition result for each of the test long-tail radiation source signals; In which, the long-tail radiation source individual identification model includes: a feature transfer learning branch, a multi-expert nested learning branch, and an individual identification module; the feature transfer learning branch is connected in parallel to the multi-expert nested learning branch, and the feature transfer learning branch is used to obtain a first individual identification result based on the data distribution of the training long-tail radiation source signal, and the data distribution is the first data distribution of the preset head class data, or the second data distribution based on the first data distribution transfer learning, the multi-expert nested learning branch is used to obtain a second individual identification result based on multiple preset features corresponding to different expert features, and the individual identification module is used to obtain the individual identification result based on the first individual identification result and the second individual identification result.

2. The method according to claim 1, characterized in that The preset loss function is determined by weighting the first loss function corresponding to the feature transfer learning branch and the second loss function corresponding to the multi-expert nested learning branch; The first loss function is defined by the following expression: Among them, N represents the batch size of input training samples, C represents the number of input training sample categories, Represents the label of the nth training long-tail radiation source signal; represents the angle between the feature vector and the category weight vector corresponding to the nth training long-tail radiation source signal; S represents the scaling factor, m c represents the hyperparameters of the first loss function; The second loss function is defined by the following expression: Among them, N h+m+t Represents all training long-tail radiation source signals of the same batch input, N m+t Indicates the training long-tail radiant source signals of the same batch input, which are the preset intermediate class data and the preset tail class data, N t represents the training long-tail radiator signal of the preset tail class data in the training long-tail radiator signal of the same batch input, x represents the training long-tail radiator signal, y represents the label of the training long-tail radiator signal, represents the cross entropy loss function, represents the hyperparameter, Indicates N h+m+t 、N m+t and N t The fusion loss function that performs loss calculation at the same time, Indicates N h+m+t 、N m+t and N t Branch loss functions for separate loss calculations; The preset loss function is defined by the following expression: Here, α represents a training factor, which is determined according to the number of training iterations.

3. The method according to claim 2, characterized in that The training set also includes individual labels corresponding to the training long-tail radiation source signals. The plurality of training long-tail radiation source signals are input into a long-tail radiation source individual recognition model for training, and the weight parameters of the model are adjusted according to a preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual recognition model, including: Inputting the plurality of training long-tail radiation source signals into a feature transfer learning branch, and obtaining a first individual recognition result according to the data distribution corresponding to each of the training long-tail radiation source signals; Inputting the plurality of training long-tail radiation source signals into a multi-expert nested learning branch, and obtaining a second individual recognition result according to a plurality of preset features corresponding to different expert features of the training long-tail radiation source signals; The first individual identification result and the second individual identification result are input into the individual identification module to obtain the predicted individual identification result corresponding to the training long-tail radiation source signal. The weight parameters of the model are adjusted according to the predicted individual identification result, the individual label, and the preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual identification model.

4. The method according to claim 3, characterized in that The feature transfer learning branch includes: a feature extraction module, a feature transfer learning module and a first individual recognition module. The inputting of the plurality of training long-tail radiation source signals into the feature transfer learning branch and obtaining a first individual recognition result according to the data distribution corresponding to each of the training long-tail radiation source signals include: Inputting the plurality of training long-tail radiation source signals into a feature extraction module for feature extraction to obtain initial features corresponding to the training long-tail radiation source signals; Inputting the initial features into a feature transfer learning module to perform feature transfer learning to obtain data distribution of the initial features; The data distribution is input into a first individual recognition module for individual recognition to obtain the first individual recognition result.

5. The method according to claim 4, characterized in that Inputting the initial features into a feature transfer learning module for feature transfer learning to obtain data distribution of the initial features includes: Determining target type data of the initial feature according to the number of discrete sampling data corresponding to the training long-tail radiation source signal and a preset threshold, wherein the target type data is any one of preset head type data and preset tail type data; Determining the data distribution according to the target type data and the first data distribution; The first data distribution is determined by performing mean calculation on multiple data distributions corresponding to the preset head class data in multiple training long-tail radiation source signals.

6. The method according to claim 5, characterized in that The determining the data distribution according to the target type data and the first data distribution includes: When it is determined that the target type data is preset header type data, the first data distribution is determined to be the data distribution; or When it is determined that the target type data is preset tail type data, transfer learning is performed according to the first data distribution to obtain the data distribution.

7. The method according to claim 6, characterized in that The first data distribution is determined by performing mean calculation on multiple data distributions corresponding to the preset head class data in multiple training long-tail radiation source signals, including: Determining a plurality of preset head class data from a plurality of training long-tail radiation source signals; Obtaining distribution angles corresponding to a plurality of discrete sampling data included in each of the preset head type data; Performing mean and variance calculations on multiple distribution angles to obtain an initial mean and initial variance of each of the preset head class data; Calculating the mean of initial means and initial variances corresponding to the plurality of preset head class data to obtain a target mean and a target variance; The first data distribution is determined based on the target mean and the target variance.

8. The method according to claim 3, characterized in that The multi-expert nested learning branch includes: a feature extraction module, multiple parallel-connected expert feature extraction modules, multiple residual fusion modules, and a second individual recognition module, wherein the residual fusion modules correspond one-to-one to the expert feature extraction modules. The multiple training long-tail radiation source signals are input into the multi-expert nested learning branch, and a second individual recognition result is obtained according to multiple preset features corresponding to different expert features of the training long-tail radiation source signals, including: Inputting the plurality of training long-tail radiation source signals into a feature extraction module for feature extraction to obtain initial features corresponding to the training long-tail radiation source signals; Inputting the initial features into a plurality of parallel connected expert feature extraction modules to perform expert feature extraction, thereby obtaining a plurality of different expert features of the initial features; Inputting the multiple different expert features into the corresponding residual fusion module to extract multiple preset features, obtain multiple preset features, and perform fusion processing to obtain residual fusion features, wherein the preset features are any one of the first preset features, the second preset features, and the third preset features, the first preset features include: head class features, middle class features, and tail class features, the second preset features include: middle class features and tail class features, and the third preset features include: tail class features; The residual fusion feature is input into the second individual recognition module for individual recognition to obtain the second individual recognition result.

9. The method according to claim 3, characterized in that Inputting the first individual recognition result and the second individual recognition result into an individual recognition module to obtain a predicted individual recognition result corresponding to the training long-tail radiation source signal includes: A weighted calculation is performed on the first individual recognition result and the second individual recognition result to obtain the predicted individual recognition result.

10. A long-tail radiation source individual identification device based on dual-branch feature learning, characterized in that: include: A sample set acquisition module is used to acquire a sample set, wherein the sample set includes a training set and a test set, wherein the training set includes multiple training long-tail radiation source signals, and the test set includes multiple test long-tail radiation source signals, wherein the training long-tail radiation source signal is any one of preset head class data, preset middle class data, and preset tail class data; A training module is used to input the multiple training long-tail radiation source signals into the long-tail radiation source individual recognition model for training, and adjust the weight parameters of the model according to a preset loss function until the model converges, thereby obtaining the trained long-tail radiation source individual recognition model; A testing module, configured to input the plurality of test long-tail radiation source signals into the trained long-tail radiation source individual recognition model to obtain an individual recognition result of each of the test long-tail radiation source signals; In which, the long-tail radiation source individual identification model includes: a feature transfer learning branch, a multi-expert nested learning branch, and an individual identification module; the feature transfer learning branch is connected in parallel to the multi-expert nested learning branch, and the feature transfer learning branch is used to obtain a first individual identification result based on the data distribution of the training long-tail radiation source signal, and the data distribution is the first data distribution of the preset head class data, or the second data distribution based on the first data distribution transfer learning, the multi-expert nested learning branch is used to obtain a second individual identification result based on multiple preset features corresponding to different expert features, and the individual identification module is used to obtain the individual identification result based on the first individual identification result and the second individual identification result.

Images