Radiation source individual open set recognition algorithm based on activation recombination and important neurons

Abstract

The invention discloses a radiation source individual open set recognition algorithm based on activation recombination and important neurons. The algorithm comprises a signal feature extraction module and an open set classification module. The feature extraction module uses a ResNet-18 convolutional neural network, and the open set recognition module integrates three methods: activation recombination, pruning based on a Shapley value and an open set classifier based on an energy (Entry) scoring function. Compared with a classical SEI-OSR algorithm, the algorithm provided by the invention can distinguish known radiation source individuals from unknown radiation source individuals with high accuracy under different openness. Particularly, when a known class comprises emitter individuals with the same model as an unknown class, the precision of the proposed method is obviously higher than that of other methods.

Description

Technical Field

[0001] The invention relates to the technical field of radiation source individual identification, and in particular to an open set identification algorithm for radiation source individuals based on activation recombination and important neurons. Background Art

[0002] The increasing complexity of the electromagnetic environment has brought great difficulties to the task of source identification (SEI). New electromagnetic signals may appear in the electromagnetic environment at any time. Since the number of communication radiation sources with known signal characteristics is limited, many communication radiation source signals with unknown signal characteristics will be faced in the actual electromagnetic environment. However, most of the current deep learning-based SEI methods are based on the closed set assumption, that is, the radiation source individuals in the data set are known, and there are no unknown radiation source individuals. This will cause any unknown radiation source individual to be mistakenly classified as a known category during identification, and the accuracy of the SEI model will drop sharply.

[0003] Therefore, SEI in an open environment is an open set recognition problem, that is, unknown radiator individuals will appear in the testing phase. The problem of rejecting unknown classes and classifying known classes is called the open set recognition (OSR) problem. Open set recognition of radiator individuals (SEI-OSR) is the core challenge of safely deploying SEI models in open environments.

[0004] At present, there are few studies on SEI-OSR tasks. Existing studies attempt to apply established OSR techniques to SEI. Hanna proposed a transmitter authorization method based on deep learning open set identification, and conducted comparative experiments using OvA classifiers and OpenMax classifiers based on extreme value theory; Ziyu Lin used a method based on adversarial autoencoders to identify unknown individual transmitters in electromagnetic environments. These OSR methods introduced from other disciplines have filled the gap in open set SEI to a certain extent. However, since the differences between the individual transmitters to be identified are very subtle, and the unknown class samples are very similar to the known class samples, this makes the OSR task of SEI more difficult than other OSR tasks, and the SEI open set identification technology has poor applicability in practical applications. Summary of the invention

[0005] In order to make up for the above shortcomings, the present invention provides an individual open set identification algorithm of radiation sources based on activation recombination and important neurons, aiming to improve the problem in the prior art that operators need to make on-site adjustments and the operation is inconvenient.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: an individual open set identification algorithm of radiation sources based on activation recombination and important neurons, including a signal feature extraction module and an open set identification module.

[0007] The feature extraction module uses the ResNet-18 convolutional neural network.

[0008] The open set recognition module integrates three methods: activation recombination, pruning based on Shapley value and open set classifier based on energy scoring function.

[0009] The signal feature extraction module performs short-time Fourier transform on the collected radiation source signal to obtain a time-frequency feature image of the signal, and inputs the image into the feature extraction module for feature extraction.

[0010] During the training process, the open set recognition module inputs the divided training set into the radiation source feature extraction module for training. The SEI-OSR model uses the SoftMax function as a classifier. Suppose a pre-trained radiation source individual recognition network parameterized by θ is f θ (x), which encodes the input x into a q-dimensional feature space. Let h(x) denote the activation vector of the penultimate layer of the network. W∈R q×K represents the weight matrix corresponding to the activation vector h(x), where K is the number of known radiation source individuals.

[0011] During the test, the network θ The activation reorganization method is applied to the feature vector of the penultimate layer of (x). First, the activation value of the p1th percentile is calculated to determine the threshold t, and the activation values ​​below the value t are set to 0. Then, the activation value of the p2th percentile is calculated to determine the threshold c, and for each activation value above the threshold c, its amplitude is limited to c. Finally, the reorganized activation values ​​are filled into the penultimate layer of the network. For individuals with known and unknown radiation sources, their activation values ​​in the penultimate layer of the SEI feature extraction network will show different distributions. The activation reorganization module uses this feature to increase the difference between known and unknown classes. After activation reorganization, the activation vector h(x) of the penultimate layer of the model becomes:

[0012]

[0013] The output of the model after activation and reorganization can be expressed as:

[0014]

[0015] Furthermore, the activation reorganization module performs Shapley value-based pruning on the neurons after activation reorganization. The key idea of ​​Shapley value-based pruning is to sort the neuron activations and weights according to their contribution metrics, and selectively utilize a subset of important neuron activations and weights. This module can be divided into Shapley value-based activation pruning (AP) and Shapley value-based weight pruning (WP).

[0016] We use Shapley values ​​to quantify the contribution of neuron activations and weights. The Shapley value is defined as the average marginal contribution of a feature over all possible combinations of other features in the input, where the marginal contribution is the change in the output after removing the feature. Therefore, the Shapley value is determined by the effect of removing a neuron. For the individual sample of the lth type of radiation source, the i-th neuron a in the second-to-last layer of the neural network i The Shapley value of can be expressed as:

[0017]

[0018] Activation pruning based on Shapley values ​​selectively prunes the activations of neurons. Based on all samples in the training set, the average contribution of each neuron in the penultimate layer is calculated, and the contribution matrix is ​​defined The (i,j)th entry of the contribution matrix is ​​defined as:

[0019]

[0020] In the formula, n represents the number of training samples of the known radiation source individual category l. For each column of C, find its p3 percentile value, and set the activation value in the column below the percentile value to 0, and the activation value not below the percentile value to 1, so as to obtain the activation mask matrix After activation pruning based on Shapley values, the model output predicted as class l is:

[0021]

[0022] in represents the lth column of the activation mask matrix M, and the operator ⊙ represents element-wise multiplication.

[0023] Furthermore, the weight pruning based on Shapley value selectively prunes the weights of neurons. Based on the contribution matrix C, the weight contribution matrix of the lth class is defined as:

[0024]

[0025] where c lRepresents the lth column of the contribution matrix C. For the weight contribution matrix Find its p4th percentile value and set the values ​​below it to 0 and the rest to 1. In this way, the weight mask matrix for the lth class is obtained After weight pruning based on Shapley values, the model output predicted as class l is:

[0026]

[0027] After activation reorganization and Shapley value-based pruning, the output of the model is:

[0028]

[0029] The open set classifier based on the energy scoring function is mainly divided into two steps. First, the energy scoring function is used to distinguish the known radiation source individuals from the unknown radiation source individuals, and then the closed set classifier SoftMax is used to identify the distinguished known radiation source individuals.

[0030] Since the commonly used scoring function SoftMax often gives a high confidence level to the predicted category, which is not conducive to distinguishing between known radiation source individuals and unknown radiation source individuals, an energy scoring function is used to enhance the algorithm's ability to distinguish between known and unknown categories. The definition of energy scoring is as follows:

[0031]

[0032] In the model inference stage, a threshold mechanism is used to distinguish known classes from unknown classes. A suitable threshold λ is set so that most of the known class individuals can be correctly distinguished. In this algorithm, the threshold λ is set to the energy score value when 95% of the data in the training set is identified as known classes, so as to strike a balance between the accuracy of known classes and unknown classes. At this time, when the energy score value finally output by the model is less than the threshold λ, we classify it as an unknown radiation source individual, otherwise, it is classified as a known radiation source individual.

[0033]

[0034] The present invention has the following beneficial effects:

[0035] Compared with the classic SEI-OSR algorithm, the algorithm proposed in this invention can distinguish known radiation source individuals from unknown radiation source individuals with high accuracy at different openness. In particular, when the known class contains emitter individuals of the same model as the unknown class, the accuracy of the proposed method is significantly higher than other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1This is a flow chart of the open set identification algorithm for radiation sources of the present invention;

[0037] Figure 2 Activate the implementation of the recombination module for the present invention;

[0038] Figure 3 The invention is an implementation of an activation pruning module based on Shapley value;

[0039] Figure 4 The invention provides an implementation of a weight pruning module based on Shapley values;

[0040] Figure 5 Displaying a track map for the terminal of the present invention; DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The SEI-OSR model proposed in this invention includes two main components: signal feature extraction module and open set classification module. The feature extraction module uses ResNet-18 convolutional neural network, while the open set recognition module integrates three methods: activation recombination, Shapley value-based pruning, and open set classifier based on energy scoring function. The algorithm flow chart of the SEI-OSR model is shown in Figure 1 shown.

[0043] First, the collected radiation source signal is subjected to short-time Fourier transform to obtain the signal's time-frequency feature image, which is then input into the feature extraction module for feature extraction. During the training process, the divided training set is input into the radiation source feature extraction module for training. The SEI model uses the SoftMax function as a classifier. Suppose a pre-trained radiation source individual recognition network parameterized by θ is f θ (x), which encodes the input x into a q-dimensional feature space. Let h(x) denote the activation vector of the penultimate layer of the network. represents the weight matrix corresponding to the activation vector h(x), where K is the number of known radiation source individuals.

[0044] During the test, the network θThe activation reorganization method is applied to the feature vector of the penultimate layer of (x). First, the activation value of the p1th percentile is calculated to determine the threshold t, and the activation values ​​below the value t are set to 0. Then, the activation value of the p2th percentile is calculated to determine the threshold c, and for each activation value above the threshold c, its amplitude is limited to c. Finally, the reorganized activation values ​​are filled into the penultimate layer of the network. For individuals with known and unknown radiation sources, their activation values ​​in the penultimate layer of the SEI feature extraction network will show different distributions. The activation reorganization module uses this feature to increase the difference between known and unknown classes. After activation reorganization, the activation vector h(x) of the penultimate layer of the model becomes:

[0045]

[0046] The output of the model after activation and reorganization can be expressed as:

[0047]

[0048] The activation of the reorganization module is implemented by Figure 2 As shown in the figure, each circle represents a neuron activation, and the darker the color of the circle, the larger the activation value.

[0049] The neurons after activation reorganization are pruned based on Shapley values. The key idea of ​​Shapley value-based pruning is to sort the neuron activations and weights according to their contribution metrics and selectively utilize a subset of important neuron activations and weights. This module can be divided into Shapley value-based activation pruning (AP) and Shapley value-based weight pruning (WP).

[0050] We use Shapley values ​​to quantify the contribution of neuron activations and weights. The Shapley value is defined as the average marginal contribution of a feature over all possible combinations of other features in the input, where the marginal contribution is the change in the output after removing the feature. Therefore, the Shapley value is determined by the effect of removing a neuron. For the individual sample of the lth type of radiation source, the i-th neuron a in the second-to-last layer of the neural network i The Shapley value of can be expressed as:

[0051]

[0052] Activation pruning based on Shapley values ​​selectively prunes the activations of neurons. Based on all samples in the training set, the average contribution of each neuron in the penultimate layer is calculated, and the contribution matrix is ​​defined The (i,j)th entry of the contribution matrix is ​​defined as:

[0053]

[0054] In the formula, n represents the number of training samples of the known radiation source individual category l. For each column of C, find its p3 percentile value, and set the activation value in the column below the percentile value to 0, and the activation value not below the percentile value to 1, so as to obtain the activation mask matrix After activation pruning based on Shapley values, the model output predicted as class l is:

[0055]

[0056] in represents the lth column of the activation mask matrix M, and the operator ⊙ represents element-wise multiplication. The flowchart of the activation pruning module based on Shapley value is as follows: Figure 3 shown.

[0057] Weight pruning based on Shapley value selectively prunes the weights of neurons. Based on the contribution matrix C, the weight contribution matrix of the lth class is defined as:

[0058]

[0059] where c l Represents the lth column of the contribution matrix C. For the weight contribution matrix Find its p4th percentile value and set the values ​​below it to 0 and the rest to 1. In this way, the weight mask matrix for the lth class is obtained After weight pruning based on Shapley values, the model output predicted as class l is:

[0060]

[0061] The flowchart of the weight pruning module based on Shapley value is as follows Figure 4 shown.

[0062] After activation reorganization and Shapley value-based pruning, the output of the model is:

[0063]

[0064] The open set classifier based on the energy scoring function is mainly divided into two steps. First, the energy scoring function is used to distinguish the known radiation source individuals from the unknown radiation source individuals, and then the closed set classifier SoftMax is used to identify the distinguished known radiation source individuals.

[0065] Since the commonly used scoring function SoftMax often gives a high confidence level to the predicted category, which is not conducive to distinguishing between known radiation source individuals and unknown radiation source individuals, an energy scoring function is used to enhance the algorithm's ability to distinguish between known and unknown categories. The definition of energy scoring is as follows:

[0066]

[0067] In the model inference stage, a threshold mechanism is used to distinguish known classes from unknown classes. A suitable threshold λ is set so that most of the known class individuals can be correctly distinguished. In this algorithm, the threshold λ is set to the energy score value when 95% of the data in the training set is identified as known classes, so as to strike a balance between the accuracy of known classes and unknown classes. At this time, when the energy score value finally output by the model is less than the threshold λ, we classify it as an unknown radiation source individual, otherwise, it is classified as a known radiation source individual.

[0068]

[0069] Working principle: This patent mainly compares the distance-based OpenMax open set recognition algorithm, the confidence-based Sigmoid-threshold open set recognition algorithm, and the reconstruction loss-based CROSR open set recognition algorithm. Openness is used to measure the openness of the experiment. Openness quantifies the difficulty of the SEI-OSR task as a function of the number of known radiation source individuals and unknown radiation source individual devices. Openness is defined as:

[0070]

[0071] In the formula, C TR represents the number of radiation source individuals in training, C TE Indicates the number of radiation source individuals that appear during the test. The known class recognition accuracy KA, unknown class recognition accuracy UA, and normalized accuracy NA are used to measure the algorithm performance. The normalized accuracy NA is defined as:

[0072] NA=(KA+UA) / 2.

[0073] The experiment used the communication signals sent by 6 maritime walkie-talkies as the individual signals of the radiation source. Among them, 3 models were IC-M510 and the other 3 models were IC-M25. The center frequency was 156.8MHz, and the signal receiving device was a USRPx310.

[0074] An open set identification experiment of radiation source individuals was conducted according to the openness. Three IC-M510 maritime walkie-talkies were used as known radiation source individuals, and three IC-M25 maritime walkie-talkies were used as unknown radiation source individuals. The ratios of known and unknown classes were 3:0, 3:1, 3:2, 3:3, and 2:3, respectively, and the corresponding openness was 0, 0.074, 0.134, 0.184, and 0.244, respectively. The experimental results are shown in Figure 2. Figure 5 shown.

[0075] As can be seen from the figure, the proposed method outperforms other methods in all cases except the case of openness 0.244. When the openness is 0.244, although the normalized accuracy of the proposed method is slightly lower than that of the CROSR method, the difference between the two is small.

[0076] Three IC-M510 maritime intercoms and two IC-M25 maritime intercoms are used as known classes, and one IC-M25 maritime intercom is used as an unknown class. At this time, the ratio of known classes to unknown classes is 5:1, and the corresponding openness is 0.047. From the perspective of openness, this is a relatively simple open set recognition task. However, since the known class contains radiation source individuals of the same model as the unknown class, the differences between the transmitter individuals are more subtle. Therefore, this is a more difficult radiation source individual open set recognition scenario. Therefore, for this scenario, the four SEI-OSR methods are tested separately. The test results are shown in Table 1. The second line is the recognition accuracy of the four methods for known radiation source individuals, the third line is the recognition accuracy of the four methods for unknown radiation source individuals, and the fourth line is the normalized recognition accuracy of the four methods. By comparing the accuracy of the four algorithms, it can be found that the SEI-OSR algorithm proposed in the present invention achieves the optimal normalized accuracy, which is much higher than the other three methods, and achieves the optimal trade-off between the accuracy of the known class and the accuracy of the unknown class.

[0077] Table 1. Comparison of CFAR module running time

[0078]

[0079] The above experimental results show that compared with the classic SEI-OSR algorithm, the algorithm proposed in this invention can distinguish known radiation source individuals from unknown radiation source individuals with high accuracy under different openness. In particular, when the known class contains emitter individuals of the same model as the unknown class, the accuracy of the proposed method is significantly higher than other methods.

[0080] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An open set identification algorithm for radiation source individuals based on activation recombination and important neurons, characterized in that: Including signal feature extraction module and open set recognition module, The feature extraction module uses the ResNet-18 convolutional neural network. The open set recognition module integrates three methods: activation recombination, pruning based on Shapley value and open set classifier based on energy scoring function. The signal feature extraction module performs short-time Fourier transform on the collected radiation source signal to obtain the time-frequency feature image of the signal, and inputs it into the feature extraction module for feature extraction. During the training process, the open set recognition module inputs the divided training set into the radiation source feature extraction module for training. The SEI-OSR model uses the SoftMax function as a classifier. Let a pre-trained radiation source individual recognition network parameterized by θ be f θ (x), which encodes the input x into a q-dimensional feature space. Let h(x) represent the activation vector of the penultimate layer of the network. represents the weight matrix corresponding to the activation vector h(x), where K is the number of known radiation source individuals, During the test, the network θ The activation reorganization method is applied to the feature vector of the penultimate layer of (x). First, the activation value of the p1th percentile is calculated to determine the threshold t, and the activation value below the t value is set to 0. Then, the activation value of the p2th percentile is calculated to determine the threshold c. For each activation value above the threshold c, its amplitude is limited to c. Finally, the reorganized activation value is filled into the penultimate layer of the network. For individuals with known and unknown radiation sources, their activation values ​​in the penultimate layer of the SEI feature extraction network will show different distributions. The activation reorganization module uses this feature to increase the difference between known and unknown classes. After activation reorganization, the activation vector h(x) of the penultimate layer of the model becomes: The output of the model after activation and reorganization can be expressed as:

2. According to claim 1, an open set identification algorithm for radiation source individuals based on activation recombination and important neurons is characterized in that: The activation reorganization module performs Shapley value-based pruning on the neurons after activation reorganization. The key idea of ​​Shapley value-based pruning is to sort the neurons according to their contribution metrics of activation and weight, and selectively use a subset of important neuron activations and weights. This module can be divided into Shapley value-based activation pruning (AP) and Shapley value-based weight pruning (WP). We use Shapley values ​​to quantify the contribution of neuron activations and weights. Shapley values ​​are defined as the average marginal contribution of a feature over all possible combinations of other features in the input, where the marginal contribution is the change in the output after removing the feature. Therefore, the Shapley value is determined by the effect of removing a neuron. For the individual sample of the lth type of radiation source, the i-th neuron a in the second-to-last layer of the neural network is i The Shapley value of can be expressed as: Activation pruning based on Shapley values ​​selectively prunes the activation of neurons. Based on all samples in the training set, the average contribution of each neuron in the penultimate layer is calculated, and the contribution matrix is ​​defined The (i,j)th entry of the contribution matrix is ​​defined as: Where n represents the number of training samples of the known radiation source individual category l. For each column of C, find its p3 percentile value, and set the activation value in the column below the percentile value to 0, and the activation value not below the percentile value to 1, so that the activation mask matrix is ​​obtained. After activation pruning based on Shapley values, the model output predicted as class l is: in represents the lth column of the activation mask matrix M, and the operator ⊙ represents element-wise multiplication.

3. The open set identification algorithm of radiation source individuals based on activation recombination and important neurons according to claim 1 is characterized in that: The weight pruning based on Shapley value selectively prunes the weights of neurons. Based on the contribution matrix C, the weight contribution matrix of the lth class is defined as: where c l Represents the lth column of the contribution matrix C, for the weight contribution matrix Find its p4th percentile value and set the values ​​below it to 0 and the rest to 1. In this way, we get the weight mask matrix for the lth class. After weight pruning based on Shapley values, the model output predicted as class l is: After activation reorganization and Shapley value-based pruning, the output of the model is: The open set classifier based on the energy scoring function is mainly divided into two steps. First, the energy scoring function is used to distinguish the known radiation source individuals from the unknown radiation source individuals. Then, the closed set classifier SoftMax is used to identify the distinguished known radiation source individuals. Since the commonly used scoring function SoftMax often gives a high confidence level to the predicted category, which is not conducive to distinguishing between known radiation source individuals and unknown radiation source individuals, an energy scoring function is used to enhance the algorithm's ability to distinguish between known and unknown categories. The definition of energy scoring is as follows: In the model inference stage, a threshold mechanism is used to distinguish known classes from unknown classes. A suitable threshold λ is set so that most known class individuals can be correctly distinguished. In this algorithm, the threshold λ is set to the energy score value when 95% of the data in the training set is identified as known classes, so as to strike a balance between the accuracy of known classes and unknown classes. At this time, when the energy score value finally output by the model is less than the threshold λ, we classify it as an unknown radiation source individual, otherwise, it is classified as a known radiation source individual.

Images