Radiation source information joint identification model based on multi-task learning

Abstract

The invention discloses a radiation source information joint identification model based on multi-task learning, and the model comprises an AMSCN backbone network module which is used for extracting shared features from original signals and providing the shared features for downstream AMC and SEI tasks; the MDHC dual-task classifier module comprises an AMC classifier and an SEI classifier, the MDHC dual-task classifier module maps the shared features provided by the AMSCN backbone network module into category probability vectors of the current task, the subscript of the maximum probability value in the vectors is the output of the classification task, the output vector of the AMC performs mask operation on a probability matrix output by the SEI, and the probability matrix output by the SEI is output by the MDHC dual-task classifier module. According to the result output by the AMC classifier, the group corresponding to the result of the AMC classifier is selected from a plurality of groups of results to serve as final output, the model improves the performance of the SEI task through the characteristics shared by the AMC task, and compared with a traditional single-task model, the average accuracy of the SEI task is improved by more than 2%.

Description

Technical Field

[0001] The present invention relates to the technical field of signal recognition, and in particular to a radiation source information joint recognition model based on multi-task learning. Background Art

[0002] In existing signal feature recognition research, AMC tasks and SEI tasks are often implemented as two independent tasks, that is, each task corresponds to a dedicated model, which not only wastes the computing resources of the monitoring end, but also fails to fully utilize the relationship between the two tasks to improve the recognition performance. As wireless communication systems become more complex, the signals received by the SEI task are no longer generated by a single communication modulation method, but will change over time. At this time, the monitoring system needs to consider the impact of changes in modulation methods on the SEI task when designing it.

[0003] In the field of deep learning, end-to-end model design has become a mainstream method. The advantage of this method is that the model can directly learn the mapping relationship from the original input data to the target output, avoiding the need for staged processing and artificial feature engineering. Multi-task learning is a learning algorithm that incorporates multiple tasks into the same end-to-end model. It allows the model to share information between different tasks, thereby improving the performance of each task. For the AMC and SEI tasks in the field of signal monitoring, both tasks have similarities, whether from the perspective of model structure design or from the perspective of data input form, so it is very suitable to share the same feature extraction network in an end-to-end model. From existing research, it can be seen that both AMC tasks and SEI tasks follow the same model design criteria, and both tasks can use constellation diagrams, IQ data or amplitude phase data as input, so the two tasks have a common model basis. In addition, from the perspective of the signal feature generation mechanism, the AMC task and the SEI task also have similarities. According to the analysis results of the signal model in Chapter 2, the signal modulation feature is an intentional modulation process, while the RF fingerprint feature is an unintentional modulation process. The two change the waveform of the signal to different degrees. The recognition of waveform features can be subdivided into the recognition of signal modulation features and the recognition of signal RF fingerprint features. Therefore, the two tasks have a common feature basis.

[0004] Therefore, dual-task classification has great potential in the field of signal monitoring and needs to be developed urgently. Summary of the invention

[0005] The purpose of the present invention is to provide a radiation source information joint recognition model based on multi-task learning to improve the accuracy of radiation source information joint recognition.

[0006] To achieve the above object, the present invention provides the following technical solution: a radiation source information joint recognition model based on multi-task learning, comprising:

[0007] The AMSCN backbone network module is a shared backbone network based on the CTDNN model, which is used to extract shared features from the original signal and provide them to downstream AMC and SEI tasks;

[0008] MDHC dual-task classifier module: includes an AMC classifier and a SEI classifier. The MDHC dual-task classifier module maps the shared features provided by the AMSCN backbone network module into the category probability vector of the current task. The subscript of the maximum probability in the vector is the output of the classification task. The output vector of AMC performs a mask operation on the probability matrix output by SEI. According to the results of the AMC classifier output, the group corresponding to the AMC classifier results is selected from several groups of results as the final output.

[0009] Preferably, the feature dimension of the input signal in the AMSCN backbone network module is B×C×L, where B represents the number of samples in each batch, C is the number of channels of the signal, the number of channels is 2, and L represents the sample length. The number of feature channels output by each convolutional layer in DenseNet is 16, and the output of the previous convolution layers is cascaded at the output of the last convolutional layer.

[0010] Preferably, two layers of Transformer Block are stacked in the AMSCN backbone network module, and the structure of each layer is consistent. The input feature dimension is 162, which is the same as the number of channels of the DenseNet output.

[0011] Preferably, the AMC classifier is composed of two layers of full connection, the number of fully connected neurons in the first layer is 256, and the number of fully connected neurons in the second layer is the same as the total number of modulation categories at the transmitting end.

[0012] Preferably, the number of fully connected neurons in the second layer of the SEI classifier is the product of the total number of modulation categories at the transmitter and the number of different transmitters.

[0013] Preferably, the transmitting end has a total of U modulation modes and V different transmitters. The length of the output vector of the last layer of the SEI classifier is U×V. The final output of the MDHC dual-task classifier module is to select the group corresponding to the AMC classifier results from the U group results of the AMC classifier as the final output. The feature vector output by the SEI classifier is rearranged to form a U×V matrix, which is point-multiplied with the generated mask matrix and then added column by column to convert the matrix into a vector of length V, which is the final output of the SEI classifier.

[0014] Preferably, the radiation source information joint identification model based on multi-task learning is trained using a supervised training method.

[0015] Preferably, the cross entropy loss is used to constrain the two classifiers, and the final loss of the model is the sum of the losses of the two classifiers.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: the present model proposes an improvement on the problem of performance loss caused by the lack of feature fusion between the two signal feature classification tasks existing in the traditional radio frequency signal monitoring system. When both AMC and SEI tasks are required at the same time, a new multi-task classification method AMSCN is proposed to improve detection accuracy and computational efficiency. The model learns the common features of the two tasks at the same time through a high-capacity backbone network, and then learns the unique features of each task separately through a mask-based dual channel. The results of simulation experiments show that AMC and SEI are two feature-related tasks, and the dual-task model can make full use of this correlation to improve the detection accuracy of SEI. In addition, the computational structure of the model is the same under the two tasks, which can simplify the training and deployment of monitoring systems under multiple tasks. Through the research of this application, it is found that multi-task learning has great potential in the field of signal monitoring, that is, the more comprehensive the information obtained by the model during the training process, the more accurate the signal recognition will be. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a model of the present application designed based on the influence of carrier modulation information and radio frequency fingerprint information of the transmitting end on the signal waveform in an embodiment of the present invention;

[0018] Figure 2 Schematic diagram of a signal multi-task recognition scenario in this embodiment;

[0019] Figure 3 Schematic diagram of a signal formation process of a single carrier system transmitting end in this embodiment;

[0020] Figure 4 This is a comparison chart of the RF fingerprint recognition accuracy under different modulation modes in this embodiment;

[0021] Figure 5 Schematic diagram of calculation of mask in MDHC in this embodiment;

[0022] Figure 6 Schematic diagram of the effect of the mask mechanism on the dual-task recognition performance in this embodiment;

[0023] Figure 7 Schematic diagram of the effect of dual-task joint training on recognition performance in this embodiment;

[0024] Figure 8Schematic diagram of accuracy comparison among multiple models in this embodiment. DETAILED DESCRIPTION

[0025] 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.

[0026] like Figure 1 As shown in the figure, since the carrier modulation information and RF fingerprint information of the transmitter have a joint impact on the signal waveform, the radiation source information joint identification model based on multi-task learning of this application is designed, and a mask-based dual-task classifier head (Mask-based Dual-Head Classifier, MDHC) is designed to improve the classification performance of the two tasks. And the multi-task cross entropy loss is used to drive the training of the entire model to ensure the balance of the two tasks during the training process.

[0027] In this embodiment, we first introduce the application scenario of the joint identification of radiation source information, the workflow of the signal identification system, and the signal model of the transmitter, and finally give the effect of the modulation mode on the performance of the RF fingerprint recognition system observed in a semi-simulation experimental environment. Figure 2 As shown, in the signal multi-task recognition scenario, there are multiple similar communication devices at the transmitting end, and only one device is sending signals at a certain time, and the single-carrier modulation methods they use are not fixed but will be constantly changed. The signal monitoring end first captures the signal in the air, and then identifies the radiation source information contained in the signal, where the modulation method information is used for subsequent signal content analysis, and the radio frequency fingerprint information is used to identify whether the information transmitted by the signal is safe and reliable. In the figure, the N-point complex baseband sequence sampled by the receiver is represented by x(n), the real part of the sequence represents the in-phase component of the baseband signal, and the imaginary part of the sequence represents the orthogonal component of the baseband signal. The sequence is sent to the AMSCN model of this application to predict the modulation method of the signal and the individual transmitter to which it belongs. The symbol H represents the assumption made by the model about the current signal, that is, the model believes that the signal uses the i-th modulation method in the modulation method set and is sent by the j-th device. These prediction information can serve the subsequent information cognition module of the monitoring system and strengthen the monitoring system's security management of the entire electromagnetic spectrum space.

[0028] like Figure 3As shown in the figure, the set of modulation methods only includes digital modulation. The modulation process of the baseband signal injects modulation characteristics into the signal, and the non-ideal characteristics of the devices in the up-converter inject RF fingerprints into the signal. Assuming that the baseband signal output by the up-converter is s(t), we let PA[|] represent the equivalent baseband model of a nonlinear power amplifier, fc is the center frequency of the carrier, then the RF signal output by the transmitting antenna can be expressed as: RF(t) = PA[S(t)e 2 πfet, at the receiving end, the output of the baseband signal after passing through an ideal down-converter with a center frequency of fo is: Wherein, h(t) represents the impulse response of the channel, θ represents the initial phase of the local oscillator at the receiving end, and n(t) represents additive white Gaussian noise.

[0029] In the existing deep learning-based communication signal RF fingerprint detection algorithms, most of them assume that the transmitter signal uses a fixed modulation mode, such as detecting a fixed pilot sequence at the transmitter or a fixed modulation method at the transmitter. However, there are few solutions for individual identification of radiators with variable modulation. Therefore, we need to first understand the extent to which changes in the modulation method affect the detection of the signal's RF fingerprint.

[0030] We conducted 10 control experiments. In each experiment, a different digital modulation method was selected at the transmitter. Then, five USRP[150-154]x310 devices were used to simulate different transmitters. USRP-2943R was used to simulate the receiver. Coaxial cables were used between the transceivers to simulate ideal channel propagation. A clock synchronization signal was used between the two USRPs to ensure that the received signal had no frequency deviation. The VGG network

[155] was used to identify the transmitter to which the current signal belonged. The relationship between the network recognition accuracy and the modulation method was observed in these 10 experiments.

[0031] The comparison results of 10 groups of experiments are as follows Figure 4 As shown, it can be seen that under the same recognition model, the accuracy of individual radiator identification will fluctuate in communication systems using different modulation modes, among which BPSK, OQPSK, 8PSK, 16PSK, and 32QAM have lower accuracy, while QAM modulation has higher accuracy, which means that the model has a preference for QAM-type signal identification. However, it should be emphasized that the formation factors and degree of difference of RF fingerprints of different communication systems are not the same, so this control experiment is only to illustrate that under the same set of transmitters, different modulation modes will affect the performance of the model in identifying individual transmitters, rather than to illustrate that a certain modulation mode can better identify individual transmitters.

[0032] Based on the above, the present application provides a design method of an AMSCN model, including the design of a backbone network and the design of a multi-task classifier. AMSCN mainly includes three parts, namely, a DenseNet part in the backbone network, a Transformer part, and a multi-classifier head MDHC. The design of the shared backbone network is based on the CTDNN model. The CTDNN model has high-performance feature extraction capabilities and flexible model scaling capabilities, so it is used as the backbone model of AMSCN. In AMSCN, the shared backbone network is responsible for extracting significant features from the original signal waveform and then providing them to the downstream AMC and SEI tasks. Thanks to the similarity between the AMC task and the SEI task, the shared feature extraction network saves the computing resources of the entire model on the one hand, and enables the large-capacity model to learn richer feature representations on the other hand, and improves the representation ability of the output features. In the downstream classifier design, in order to improve the performance of the two tasks, we designed a separate classifier for each task, which maps the shared features provided by the upstream into the category probability vector of the current task. The subscript of the maximum probability in the vector is the output of the classification task, which represents the index in the transmitter modulation mode set and the transmitter individual set.

[0033] As shown in Table 1, the hyperparameter settings of DenseNet in AMSCN, the feature dimension of the network input is B×C×L, where B represents the number of samples in each batch, C is the number of channels of the signal, and in this chapter, the signal is input in the form of a one-dimensional sequence of 2 channels, L represents the sample length, and the number of feature channels output by each convolutional layer in DenseNet is 16, and the output of the previous convolutional layers is cascaded at the output of the last convolutional layer. The backbone network of AMSCN stacks two layers of TransformerBlock, and the structure of each layer is consistent. The feature dimension of the input is 162, the same as the number of channels of DenseNet output.

[0034] Table 1:

[0035]

[0036] After the AMSCN backbone network extracts the shared features, we use an MDHC dual-task classifier head for classification. The AMC classifier head is a typical MLP classifier, which consists of two layers of full connections. The number of fully connected neurons in the first layer is 256, and the number of fully connected neurons in the second layer is the same as the total number of modulation categories at the transmitter. The SEI classifier head is also an MLP classifier, but the number of fully connected neurons in its second layer is the product of the total number of modulation categories at the transmitter and the number of different transmitters. Assuming that the transmitter has a total of U modulation modes and V different transmitters, the length of the output vector of the last layer of the SEI classifier is U×V. We divide this long vector into U groups, each of which has a length of V. From a mathematical point of view, we hope that the SEI classifier gives the conditional probability that the signal belongs to a certain transmitter under the assumption of each modulation mode, so there are a total of U groups, and the final output of the SEI model is selected according to the results of the AMC classifier, that is, from the U group results, the group corresponding to the AMC classifier result is selected as the final output. From the calculation form of the model, the output vector of AMC performs a mask operation on the probability matrix output by SEI. At this time, the SEI classifier can be regarded as an integration of U MLP classifiers with the same structure. Each classifier outputs a vector of V and shares the first layer of fully connected weights. By selecting the output of the SEI classifier, a regularization effect is achieved on the classifier network.

[0037] like Figure 5 , shows the working process of the mask in MDHC. For the convenience of explanation, we set U and V to be equal to 3. The numbers in the vector in the figure are random and are only used for illustration. The dimension of the mask matrix corresponding to AMC is U×V. Each row of the matrix corresponds to an element position in the AMC feature vector. If the element at this position in AMC obtains the maximum value, the corresponding row of the matrix takes all 1s, otherwise it is all zeros. The last element of the AMC vector in the figure is the largest, so the last row of the mask matrix is ​​all 1s. The feature vector output by the SEI classifier also forms a U×V matrix after rearrangement. After point multiplication with the mask matrix generated in the previous step, it is added by column, and the matrix is ​​converted into a vector of length V, which is the final output of the SEI classifier.

[0038] Since it is relatively easy for the model to classify the AMC of the signal, the main waveform features in the signal are determined by the modulation features of the signal. When different transmitters use the same modulation method, the similarity of the signal features is very high. Therefore, in the MDHC classifier, the AMC classifier branch is independent and not affected by SEI. When the modulation method of the signal changes, the signal fingerprint features introduced by the transmitter will shift to a certain extent. Therefore, we need to first determine the modulation features of the current signal, and then eliminate the uncertainty of the modulation method when determining the signal fingerprint features, so as to increase the model's attention to the signal fingerprint features.

[0039] Regarding the offline training of the radiation source information joint recognition model based on multi-task learning, the model is regarded as a conditional probability distribution P(ylx;θ), where y represents the label, x represents the input data, and θ is the model parameter. The goal of MLE is to find a set of parameters θ that maximizes the probability of observing the target data y when the input data x is given. There are two classification tasks in AMSCN. We let the probability vector output by the AMC classifier be: and, The scalar he|M=i(x12) represents the probability that the current signal belongs to the i-th modulation mode. In the vector h(x12), the scalar subscript with the largest probability value represents the modulation mode category predicted by the model. This calculation process can be expressed as: u represents the modulation index of the model output. For the SEI task, the generation of the probability vector output by the classifier depends on the prediction result of the AMC task. The original vector group output by the first stage of the SEI classifier is expressed as: Among them, H represents the set of U group vectors, which represents the probability vector of the signal belonging to each transmitter when the modulation type is u, and the final probability vector output by the SEI classifier is selected by the AMC classification result. The goal of AMSCN model training is to optimize two tasks at the same time to ensure that the probability that the result predicted by each classifier is the same as the label is maximized. We use cross entropy loss to constrain each classifier. The final loss of the model is the sum of the losses of the two classifiers, that is: L(θ)=LM(θ)+LF(θ), where LM(θ) and LE(θ) represent the cross entropy losses of the AMC task and SEI task respectively, and their calculation process is: Generally speaking, multiple tasks need to consider the balance problem during the training process, but for the joint task of AMC and SEI, its essence is to extract the waveform characteristics of the signal and perform classification in the downstream, so the optimization direction of the two tasks is very similar. In MDHC, the optimization of the AMC task is independent, while the optimization of the SEI task depends on the convergence of the AMC task. During the training process, we also found that no additional weight adjustment is required between the two losses, and the same weight can be used. During the training process, the AMC task converges first, and then the SEI task begins to converge. In the final stage, the two tasks jointly tend to the best convergence state, and the model uses the Adam optimizer with default parameters to update the weights of neurons.

[0040] In the existing public research, there is no available dataset for both AMC and SEI tasks, so we apply to generate a simulation dataset containing both tasks based on the signal model. Figure 3As shown, this application studies signal recognition in a single-carrier communication scenario. In terms of the selection of modulation types, we selected five common digital modulation methods, namely BPSK, QPSK, 8PSK, 16QAM, and 32QAM, and used a common raised cosine roll-off filter for shaping (roll-off coefficient 0.25), where the number of sampling points per symbol is 8; in transmitter modeling, we set five transmitters with different parameter models. The model parameters of each transmitter are shown in Table 2:

[0041]

[0042] The model coefficients in the parameter list are derived from measurements of the power amplifier model in public literature. Signal limitation refers to the amplitude limitation that needs to be done before the signal is input to the power amplifier so that it is within the reasonable working range of the power amplifier [158,159]. E{|r(n)2|} represents the power requirement of the input signal, and max(|r(n)|) indicates that the maximum value of the IQ signal after scaling cannot exceed 1.

[0043] The channel of the single-carrier communication system adopts a relatively simple Gaussian channel environment, the purpose is to simplify the interference factors in the system, and focus on the fusion of signal modulation characteristics and RF fingerprint characteristics. At the receiving end, the signal-to-noise ratio range of the data set is between [-20dB, 20dB], and 11 signals with different signal-to-noise ratios are sampled at intervals of 4dB. 625 groups of signals are generated under each modulation mode, each transmitter, and each signal-to-noise ratio. The signal length is 256, divided into I and Q dual channels and input into the model. During training, it is divided into training set and test set in a ratio of 6:4, and the accuracy index of the model in the test set is used as the model evaluation index.

[0044] Ablation experiment on the mask mechanism in MDHC: The MDHC classifier in the AMSCN model is mainly designed to strengthen the information interaction between the AMC task and the SEI task. Considering that in the signal feature classification task, the change of one feature will interfere with the recognition of another feature, we hope that the model can comprehensively consider the global information. In order to compare the impact of the use of the mask mechanism in MDHC on the performance of the two tasks, this application designed a set of control experiments, in which the mask mechanism was removed from the control group, and the SEI classifier was made exactly the same as the AMC classifier, directly outputting a probability vector of the same length as the number of categories. Figure 6 The curves of the classification accuracy of the two models as the signal-to-noise ratio changes are shown. We use “No Mask” to represent the control group and “Mask” to represent the complete AMSCN model.

[0045] from Figure 6It can be seen that the mask mechanism has a more obvious improvement on the AMC task, while the improvement on the SEI task is relatively weak. From the perspective of model design, the premise for the convergence of the SEI task is that the AMC task can make accurate classifications. Therefore, the design of the mask can be regarded as adding model constraints on the AMC task through the SEI task branch in addition to the constraints of the objective function of the AMC branch itself. In order to quantitatively describe the difference in the performance curves in the figure, we use the global average accuracy indicator to measure the performance difference between the two models. The comparison is shown in Table 3. The AMSCN model using the mask mechanism has an accuracy improvement of 1.73% on the AMC task and 0.59% on the SEI task. There are some improvements on both tasks, which proves the effectiveness of this design method. In addition, from the comparison of the accuracy of the AMC task and the SEI task, the difficulty of the SEI task is higher than that of the AMC. This is mainly because the signal-to-noise ratio range required by the SEI task is higher than that of the AMC. Figure 6 It can be seen that under the condition of 0dB, the accuracy of the AMC task is about 70%, while the accuracy of the SEI task is only about 40%.

[0046] Table 3:

[0047]

[0048] Ablation experiment on dual-task joint training in AMSCN: The core idea of ​​the AMSCN model is to integrate two signal feature recognition tasks to improve the performance of each task. As a comparison, this chapter gives the training effect of the AMSCN model under the single-task model. In the AMC single-task mode, we remove the SEI branch in AMSCN and then complete the training with the same hyperparameter settings. Similarly, in the SEI task, this application allows SEI to directly output the probability vector corresponding to the individual transmitter. Figure 7 A performance curve comparison of AMSCN and single-task training is given. "Fusion" represents the joint training model, and "Single" represents the model trained separately for each task. It should be noted that the dataset used for single-task training is consistent with the dataset for joint training, that is, in the AMC single-task mode, the signal still contains RF fingerprint features, and the same is true for the dataset of the SEI single-task model.

[0049] from Figure 7It can be seen that in the AMC task, the performance curves of joint training and single-task training are relatively close, while in the SEI task, the joint training has a very obvious improvement on performance. This shows that the joint training method adopted by the AMSCN model has greatly improved the performance of SEI without damaging the performance of the AMC task, reflecting the performance advantage of multi-task learning. Table 4 shows the accuracy comparison of the two training methods under different tasks. It can be seen that the accuracy of the AMC task is basically the same, while the accuracy of the SEI task has increased by 3.3%. In addition, the comparison Figure 7 In the dual-task training mode, if the mask mechanism is not added, the model will degrade under the AMC task, which shows that simple multi-task joint training may lead to performance degradation under two tasks, so it is necessary to add some feature fusion channels such as mask mechanisms.

[0050] Table 4:

[0051] Model AMC Mission SEI Mission AMSCN single-task model 0.651 0.5142 AMSCN dual-task model 0.65 0.5472

[0052] Performance comparison of multiple models: In order to objectively evaluate the performance of the AMSCN model, this application selected 8 signal feature classification models for comparison, some of which achieved the best classification results on open source datasets. Figure 8 The comparison of the performance curves of different models is given, where ResNet12 is a shallow version of ResNet18, and DenseNet refers to the convolutional network part of the backbone model of AMSCN. Except for AMSCN, the other models are trained in single-task mode, and the average accuracy of the models is shown in Table 5. It can be seen from the figure that on the AMC task, the AMSCN model maintains the same recognition performance as other models, while on the SEI task, AMSCN has achieved a significant performance lead.

[0053] Table 5:

[0054] Model AMC Mission SEI Mission AMSCN(ours) 0.650 0.547 HCGDNN

[160] 0.649 0.2 ResNet12 0.642 0.519 PET-CGDNN

[161] 0.631 0.520 CoBoNet

[162] 0.631 0.522 DenseNet(backbone) 0.625 0.513 MCLDNN

[163] 0.606 0.518 DSCLDNN[105 0.584 0.521 LSTM2[164 0.545 0.372

[0055] As can be seen from Table 5, the performance of the same model may vary greatly under AMC and SEI tasks. A model that achieves good performance in one task may have poor performance or fail to converge in another task, such as the HCGDNN model. This shows that this model is not compatible with the differences between AMC and SEI tasks. In contrast, the AMSCN model achieved the best results in both tasks, proving that in our design, the backbone network has the ability to extract multi-scale features and can extract shared features with stronger representation capabilities from waveform inputs.

[0056] Table 6 shows the comparison of computational complexity and inference latency among models, where the AMSCN model shows the computational complexity and inference latency under dual tasks, and the other models show the results under a single task. Since AMSCN infers two tasks at the same time, its computational complexity and inference latency are not significantly inferior to other models, so AMSCN also has certain advantages in computational efficiency.

[0057] Table 6:

[0058] Model Model parameter quantity Floating point calculation Delay AMSCN(ours) 1.57M 259.34M 0.007s HCGDNN 0.457M 92.53M 0.012s ResNet12 13.85M 1549.2M 0.006s PET-CGDNN 0.841M 206.3M 0.009s CoBoNet 0.67M 135.3M 0.035s DenseNet(backbone) 0.174M 33.9M 0.004s MCLDNN 3.53M 872.6M 0.007s DSCLDNN 1.13M 293.2M 0.014s LSTM2 3.16M 811.08M 0.006s

[0059] In summary, this application proposes an improvement on the problem of performance loss caused by the lack of feature fusion between two signal feature classification tasks in traditional RF signal monitoring systems. Specifically, when both AMC and SEI tasks are required, we propose a new multi-task classification method AMSCN to improve detection accuracy and computational efficiency. AMSCN learns the common features of the two tasks simultaneously through a high-capacity backbone network, and then learns the unique features of each task separately through a mask-based dual channel. Simulation experimental results show that AMC and SEI are two feature-related tasks, and the dual-task model can make full use of this correlation to improve the detection accuracy of SEI. In addition, the computational structure of AMSCN is the same under both tasks, which can simplify the training and deployment of monitoring systems under multiple tasks. Through the research in this chapter, we believe that multi-task learning has great potential in the field of signal monitoring, that is, the more comprehensive the information obtained by the model during training, the more accurate the signal recognition will be.

[0060] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A radiation source information joint identification model based on multi-task learning, characterized by: include: The AMSCN backbone network module is a shared backbone network based on the CTDNN model, which is used to extract shared features from the original signal and provide them to downstream AMC and SEI tasks; MDHC dual-task classifier module: includes an AMC classifier and a SEI classifier. The MDHC dual-task classifier module maps the shared features provided by the AMSCN backbone network module into the category probability vector of the current task. The subscript of the maximum probability in the vector is the output of the classification task. The output vector of AMC performs a mask operation on the probability matrix output by SEI. According to the results of the AMC classifier output, the group corresponding to the AMC classifier results is selected from several groups of results as the final output.

2. According to claim 1, a radiation source information joint identification model based on multi-task learning is characterized in that: The feature dimension of the input signal in the AMSCN backbone network module is B×C×L, where B represents the number of samples in each batch, C is the number of channels of the signal, which is 2, and L represents the sample length. The number of feature channels output by each convolutional layer in DenseNet is 16, and the output of the previous convolutional layers is cascaded at the output of the last convolutional layer.

3. The radiation source information joint identification model based on multi-task learning according to claim 2 is characterized in that: Two layers of Transformer Block are stacked in the AMSCN backbone network module. The structure of each layer is consistent, and the input feature dimension is 162, which is the same as the number of channels of DenseNet output.

4. The radiation source information joint identification model based on multi-task learning according to claim 1 is characterized in that: The AMC classifier consists of two fully connected layers. The number of fully connected neurons in the first layer is 256, and the number of fully connected neurons in the second layer is the same as the total number of modulation categories at the transmitter.

5. The radiation source information joint identification model based on multi-task learning according to claim 1 or 4, characterized in that: The number of fully connected neurons in the second layer of the SEI classifier is the product of the total number of modulation categories at the transmitter and the number of different transmitters.

6. The radiation source information joint identification model based on multi-task learning according to claim 1 is characterized in that: There are U modulation modes and V different transmitters at the transmitting end. The length of the output vector of the last layer of the SEI classifier is U×V. The final output of the MDHC dual-task classifier module is to select the group corresponding to the AMC classifier results from the U group results of the AMC classifier as the final output. The feature vector output by the SEI classifier is rearranged to form a U×V matrix. After point multiplication with the generated mask matrix, it is added column by column to convert the matrix into a vector of length V, which is the final output of the SEI classifier.

7. The radiation source information joint identification model based on multi-task learning according to claim 1, characterized in that: The radiation source information joint recognition model based on multi-task learning is trained in a supervised training manner.

8. The radiation source information joint identification model based on multi-task learning according to claim 7 is characterized in that: The cross entropy loss is used to constrain the two classifiers, and the final loss of the model is the sum of the losses of the two classifiers.

Images