Electromagnetic target identification method based on interpretable multi-task learning, storage medium and equipment

Abstract

The invention discloses an electromagnetic target recognition method based on interpretable multi-task learning, a storage medium and equipment, and the method comprises the steps: firstly constructing a multi-task learning network facing modulation recognition and individual recognition, inputting an original signal, setting a hyper-parameter, and carrying out the training of the multi-task learning network; constructing a generator for outputting a disturbance signal, segmenting an original signal and generating an original mask, generating a disturbance mask through an inversion element, and further generating the disturbance signal; inputting the disturbance signals into the trained multi-task learning network to obtain probability distribution of classification results under different tasks; taking a prediction probability of a multi-task model to a disturbance signal as a supervision label, training a local linear model on an original mask and a disturbance mask, and taking a model weight Wg as a contribution degree of each subsequence to explanation through optimization; and performing normalization processing on the Wg, mapping the weight back to the original signal length, and visually marking a time-frequency domain signal region which plays a key role in model decision based on a constellation diagram.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromagnetic signals, and mainly relates to an electromagnetic target recognition method, storage medium and device based on interpretable multi-task learning. Background Art

[0002] In modern communication systems, specific emitter identification (SEI) and automatic modulation classification (AMC) are two core technologies. They focus on extracting the radio frequency fingerprint (RFF) of individual transmitters from electromagnetic signals and automatically distinguishing different signal modulation schemes, respectively. Together, they support key applications such as dynamic spectrum optimization in cognitive radio, transmitter fingerprint authentication in secure communications, illegal signal monitoring in spectrum management, and signal source localization in electronic countermeasures. In this process, feature distortion caused by noise interference, the coupling between modulation schemes and individual transmitter characteristics, and the inherent black-box uninterpretability of deep learning pose significant challenges in practical engineering. Deep learning remains the mainstream research focus, but it faces core challenges such as strong data dependence (requiring massive amounts of labeled data), insufficient generalization across scenarios and frequency bands, insufficient utilization of physical layer features, and opaque decision-making basis. These issues collectively hinder the development of more robust and interpretable electromagnetic signals.

[0003] In the existing technologies, signal classification and recognition technologies are provided. For example, a Chinese patent with the patent number CN202011184054.X proposes a method for identifying individual communication radiation sources based on a complex deep residual network, which integrates the extraction and recognition processes of radio frequency fingerprint features to improve the recognition accuracy. This method is aimed at steady-state radio frequency baseband signals. When collecting data, noise segment data is removed to obtain radio frequency fingerprints, and then the complex deep residual network is input for recognition. However, for transient radio frequency baseband signals, the carrier frequency and phase deviation of the receiver need to be estimated and compensated before collection, and the patent does not mention the aspect of interpretability. A Chinese patent with the patent number CN202010961558.1 discloses a method for identifying individual communication radiation sources based on a residual neural network. Data is obtained through steps such as receiving communication radiation source signals, calculating the bispectrum of the signals, performing non-parametric indirect estimation of the bispectrum, and obtaining the bispectrum contour map, and then the residual network is trained with the data. Finally, the trained network is used to detect and identify different communication radiation sources. This method can reduce signal noise interference, reduce the computational amount, and improve the recognition accuracy. However, the patent does not mention the recognition accuracy under low signal-to-noise ratio and the aspect of interpretability, lacking certain generalization ability and model transparency. In addition, a Chinese patent with the patent number CN202110008169.1 provides a method for automatically identifying the modulation type of communication signals. The 2FSK and 4FSK signals are identified by extracting the number of spectral peaks of the unknown signal to be identified; the 16QAM, BPSK, and QPSK signals are identified by obtaining the standard deviation of the time-domain envelope; the BPSK and QPSK signals are distinguished by using the instantaneous autocorrelation algorithm to obtain the ratio of the number of zero-crossing points. The recognition performance of each signal is improved through three-level classification. The recognition accuracy of this patent is relatively low under low signal-to-noise ratio. In addition, most of the existing patents still focus on the single-task recognition field, lacking models with high recognition accuracy in the multi-task field, and at the same time, the existing models lack the ability to explain decisions.

[0004] Although the existing methods have made certain progress in the field of electromagnetic signal recognition, they still face the following core challenges: (1) Insufficient model interpretability. Existing methods (such as Grad-CAM, LIME) have poor adaptability in the signal field and cannot generate explanations that conform to the characteristics of electromagnetic signals (such as constellation diagram correlation). (2) Noise and sparsity challenges. Sensor noise, signal attenuation, and multipath effects easily lead to sparsity or distortion of electromagnetic signals, and existing feature extraction methods are difficult to effectively extract features. (3) Poor adaptability to dynamic environments. The generalization ability of traditional models significantly decreases in complex electromagnetic environments. (4) Insufficient coordination among branches in a multi-task environment. The coupling between branches affects the recognition accuracy. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention proposes an electromagnetic target recognition method, storage medium and device based on interpretable multi-task learning. First, a multi-task learning network for modulation recognition and individual recognition is constructed, the original signal is input into the multi-task learning network, hyperparameters are set, and the multi-task learning network is trained; a generator for outputting a perturbation signal is constructed, the original signal is segmented to generate an original mask, a perturbation mask is generated by inverting elements, and then a perturbation signal is generated; the perturbation signal is input into the trained multi-task learning network to obtain the probability distribution of classification results under different tasks; taking the prediction probability of the multi-task model for the perturbation signal as a supervision label, a local linear model is trained on the original mask and the perturbation mask, and the model weight W g is used as the contribution degree of each subsequence to the explanation; for W g carry out normalization processing, and map the weights back to the original signal length, and visualize and annotate the time-frequency domain signal region that plays a key role in the model decision based on the constellation diagram.

[0006] To achieve the above object, the technical solution adopted by the present invention is: an electromagnetic target recognition method based on interpretable multi-task learning, including the following steps:

[0007] S1: Construct a multi-task learning network for modulation recognition and individual recognition. The backbone network of the multi-task learning network is DRSN, and an encoding network is built by adopting a hard sharing method and introducing a cross-stitch module; and a double decoding branch is set, and a residual shrinkage module, an average pooling layer and a fully connected layer corresponding to the number of categories are adopted;

[0008] S2: Input the original signal into the multi-task learning network constructed in step S1, set hyperparameters, and train the multi-task learning network; the hyperparameters at least include signal specifications, batch size and number of training epochs;

[0009] S3: Construct a generator for outputting a perturbation signal, segment the original signal to be explained, the generator generates a perturbation mask by randomly inverting elements, and combines the perturbation mask with the original signal to obtain a perturbation signal;

[0010] S4: Input the perturbation signal into the multi-task learning network trained in step S2 to obtain the probability distribution of classification results under different tasks;

[0011] S5: Input the similarity between the perturbation mask and the perturbation mask into a linear model, and perform model fitting in combination with the probability distribution to obtain the explanation weight W g of each subsequence;

[0012] S6: Perform normalization processing on W g and map the weights back to the original signal length, and visualize and annotate the time-frequency domain signal region that plays a key role in the model decision based on the constellation diagram.

[0013] As an improvement of the present invention, the cross-stitching structure in the step S1 cross-stitch module is specifically as follows: Assume that the output of the i-th parallel structure is outA i and outB i , then the inputs inputA i+1 and inputB i+1 of the (i + 1)-th parallel structure are:

[0014]

[0015] where are trainable parameters.

[0016] In the encoding network, the input signal first passes through a convolutional layer, a batch normalization layer, and a Relu activation layer and then enters the cross-stitch module, which is composed of two identical residual shrinkage blocks and a cross-stitching structure. After passing through multiple repeated cross-stitch modules, shared features are extracted.

[0017] As an improvement of the present invention, a soft threshold is provided in the residual shrinkage module in the step S1. Features with a soft threshold close to zero are set to zero, while negative features are retained, and the derivative of the features is 0 or 1; the soft threshold function is defined as:

[0018]

[0019] where x is the input feature, y is the output feature, and τ is the threshold parameter.

[0020] In the dual decoding branches, each branch first passes through multiple residual shrinkage modules and then through an average pooling layer. The output of the average pooling layer is flattened and sent to a fully connected layer with the corresponding number of categories to obtain the prediction result.

[0021] As another improvement of the present invention, among the hyperparameters in the step S2, the original signal sample dimension adopts I / Q signals with a specification of (2, 1024), the batch size is 32, and the number of training rounds is initially set to 30 and adjusted according to the loss change situation during training.

[0022] As another improvement of the present invention, in the step S3, the original signal is segmented into d subsequences by using the ruptures library in the signal generator to generate the original mask x ∈ 1 d . The perturbation mask x' is generated by randomly inverting the elements of x, x' ∈ {0, 1} d . The replacement noise is sampled from the Gaussian distribution where μ i and σ iis the channel mean and standard deviation estimated from the dataset at each time step i. The signal generator synthesizes a perturbation signal based on x'. Specifically, the subsequences of 0s in x' are replaced with Gaussian noise, while the subsequences of 1s are retained.

[0023] As another improvement of the present invention, in step S4, the disturbance signal is input into the trained multi-task learning network, the classification results are obtained through different task branches of the network, and the model predicts the probability of each category after processing by the Softmax function.

[0024] As a further improvement of the present invention, in step S5, the predicted probability of the disturbance signal by the multi-task model is used as the supervision label, and a local linear model is trained on the original mask and the disturbance mask. g As the contribution of each subsequence to the explanation, the optimization objective function is:

[0025]

[0026] Among them, F represents the Frobenius distance, f model () represents the prediction of the local linear model, and T is the transposition operation.

[0027] As a further improvement of the present invention, the electromagnetic target recognition method based on interpretable multi-task learning according to claim 7 is characterized in that: in the step S6, the interpretation weight W g The normalization process is as follows:

[0028]

[0029] In order to achieve the above-mentioned purpose, the present invention also adopts the following technical solution: a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the electromagnetic target recognition method based on explainable multi-task learning as described in any one of claims 1 to 8.

[0030] In order to achieve the above object, the present invention also adopts a technical solution: a computer device comprising:

[0031] a memory for storing instructions;

[0032] A processor is configured to execute the instructions so that the computer device performs the operations of the electromagnetic target recognition method based on explainable multi-task learning as described in any one of claims 1 to 8.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] (1) High synergy between multiple tasks: After synergistically optimizing the tasks of radiation source individual identification (SEI) and automatic modulation recognition (AMC), the recognition accuracy of the present invention is effectively improved compared with that of the single-task model;

[0035] (2) Strong anti-noise ability: The residual shrinkage network reduces noise through soft thresholding, significantly enhancing the robustness of the model in low signal-to-noise ratio (SNR) scenarios. Especially in the AMC (automatic modulation classification) and SEI (radiation source individual identification) tasks, it can more accurately extract weak discriminative features masked by noise;

[0036] (3) High interpretability: It supports constellation diagram visualization, allowing users to intuitively understand the decision-making basis of the model and meeting the requirements for model transparency in military and civilian fields;

[0037] (4) Wide application range: The present invention is applicable to fields such as electromagnetic spectrum monitoring, wireless device identification, and communication security, providing technical support for signal analysis and trustworthy decision-making in complex electromagnetic environments. Brief Description of the Drawings

[0038] Figure 1 It is a flowchart of the steps of the electromagnetic target recognition method based on interpretable multi-task learning of the present invention;

[0039] Figure 2 It is a structural framework diagram of the electromagnetic target recognition method based on interpretable multi-task learning of the present invention.

[0040] Figure 3 It is an experimental comparison diagram of the test example of the present invention. Detailed Embodiments

[0041] The following further clarifies the present invention in conjunction with the drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

[0042] Embodiment 1

[0043] The electromagnetic target recognition method based on interpretable multi-task learning, as Figure 1 shown, is applicable within the framework as Figure 2 shown, and specifically includes the following steps:

[0044] Step S1: Construction of a multi-task model.

[0045] Construct a multi-task learning network for modulation recognition and individual recognition. The backbone network selects DRSN, and a cross-stitch module is introduced in a hard-sharing manner to build an encoding network. The cross-stitching unit in the cross-stitch module is:

[0046]

[0047] Among them, the output of the $i$-th parallel structure is and the input of the $(i + 1)$-th parallel structure is and are trainable parameters that control the feature sharing intensity. Through this structure, the model can adaptively learn the robust representation of low signal-to-noise ratio signals.

[0048] The Cross-Stitch Unit dynamically fuses multi-task features and combines with the Shrinkage module to adaptively suppress noise. The Shrinkage module suppresses noise through a soft threshold function, and the soft threshold function is:

[0049]

[0050] where $x$ is the input feature, $y$ is the output feature, and $\tau$ is the threshold parameter. The threshold $\tau$ is adaptively calculated from the feature amplitude, and its derivative is piecewise constant (0 or 1), which can effectively avoid gradient anomalies. The formula is as follows:

[0051]

[0052] In the encoding network, the input signal first passes through a convolutional layer, a batch normalization layer, and a Relu activation layer and then enters the Cross-Stitch Unit. This module consists of two identical Shrinkage blocks and a cross-stitching structure. After passing through multiple repeated Cross-Stitch Units, shared features are extracted. The shared features are sent to a dual-decoding branch. Each branch first inputs the shared features into multiple Shrinkage blocks, and the output results are subjected to average pooling. After processing, the prediction results are output through a fully connected layer corresponding to the category output.

[0053] Step S2: The original I / Q signal is input into the multi-task learning network for training.

[0054] Hyperparameters are set to train the multi-task model. The sample dimension uses I / Q signals with a specification of (2, 1024); the batch size is set to 32, and the number of training epochs is initially set to 30 and adjusted according to the loss change during training. In this embodiment, the number of training epochs is determined to be 75, and it runs in the environment of PyTorch 1.8.1 with Python 3.8.19. The optimizer uses Adam, and the learning rate is set to 1e-3. The computing platform includes an NVIDIA GeForce RTX 3080 GPU.

[0055] In addition, the parameter $\lambda$ used to balance the multi-task learning objective function in the following formula (joint optimization function) takes a value of 0.5.

[0056]

[0057] where x i is the received I / Q signal sample, is the modulation type label of AMC, is the transmitter label of SEI, h(·) is the shared feature extractor, f m (·) and f e (·) are the task-specific classifiers of AMC and SEI respectively, L m and L e are the classification cross-entropy loss functions, and λ ∈ [0, 1] is the adaptive task weight parameter.

[0058] Step S3: Construct a generator for outputting the perturbation signal.

[0059] The original signal is segmented into d subsequences in the signal generator using the ruptures library, generating the original mask x ∈ 1 d . The perturbation mask x' is generated by randomly inverting the elements of x, x' ∈ {0, 1} d . The replacement noise is sampled from a Gaussian distribution where μ i and σ i are the channel mean and standard deviation estimated based on the dataset at each time step i. The signal generator synthesizes the perturbation signal according to x', specifically, replacing the subsequences corresponding to 0 in x' with Gaussian distribution noise and retaining the subsequences corresponding to 1.

[0060] Step S4: Obtain the probability distribution.

[0061] The perturbation signal is input into the trained multi-task learning network, and the classification results are obtained through different task branches of the network. Then, the Softmax function is used to convert these results into predicted class probabilities. The Softmax function used is: for a real-valued vector z = [z1, z2, …, z K of length K, the output of the Softmax function is a vector σ(z) = [σ(z1), σ(z2), …, σ(z K )] of the same length K, where each element σ(z i ) is calculated as follows:

[0062]

[0063] Step S5: Train the linear interpretation model.

[0064] Using the predicted probability of the perturbation signal by the multi-task model as the supervision label, train the local linear model on the original mask and the perturbation mask, and optimize the contribution degree of each subsequence to the interpretation with the model weight W g as.

[0065] Taking the original mask x and the perturbation mask x' as input features, the probability distribution of the perturbation signal obtained in the multi-task model as the supervision label, and training the linear model to solve the explanation weight W of the subsequence g The optimization objective function is:

[0066]

[0067] Among them, F represents the Frobenius distance, f model () represents the prediction of the local linear model, and T is the transposition operation. The training samples are weighted and the linear model is optimized to obtain the classification decision contribution coefficient W reflecting each subsequence. g . The proximity function assigns higher weights to perturbed masks that are more similar to the original mask.

[0068] Step S6: Normalize the weights and visualize using a constellation diagram.

[0069] To W g Normalization is performed and the weights are mapped back to the original signal length. The time-frequency domain signal areas that play a key role in model decision-making are visually annotated based on the constellation diagram.

[0070] Explanation weight W g Perform normalization:

[0071]

[0072] The normalized weights are mapped back to the original signal length. Specifically, the time steps in each subsequence are assigned the same weight to generate the corresponding constellation diagram. The constellation diagram is first drawn using the original IQ signal. The constellation diagram then highlights the time steps where the normalized weights mapped back to the original signal length exceed a certain threshold. The highlighted regions in the constellation diagram represent signal segments that are critical to the model's decision-making.

[0073] Test Case

[0074] To obtain a training dataset for the multi-task model, we selected 23 modulation schemes from the open-source Torchsig dataset, including QAM, PSK, and PAM. For this dataset, seven HackRFOnes were used for signal transmission and one USRPB210 for reception. The experiments used a carrier frequency of 1 MHz, a sampling rate of 16 MHz, a signal-to-noise ratio of -15 dB, and a received signal length of 1024.

[0075] The training model selects a multi-task model and a single-modulation task model (the structure of which is the same as that of the multi-task model except for the non-radiation source individual recognition branch). During training, the sample dimension uses I / Q signals with a specification of (2, 1024). The training set and the test set are split at a ratio of 8:2. The batch size is set to 32, and the number of training epochs is determined to be 75. It runs in the environment of PyTorch 1.8.1 with Python 3.8.19. The optimizer selects Adam, and the learning rate is set to 1e-3. The computing platform includes an NVIDIA GeForce RTX 3080 GPU.

[0076] The performance of the two models on the test set is as Figure 3 shown. Among them, the multi-task model has improved accuracy in multiple modulation methods, and the average accuracy has increased by about 4%. Finally, it can be concluded that there is a strong synergy between the two tasks, and the accuracy is improved through joint training.

[0077] In summary, the electromagnetic signal recognition method, storage medium, and device based on interpretable multi-task learning disclosed in the present invention can effectively enhance the electromagnetic signal recognition ability while having a certain model interpretability.

[0078] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements all fall within the protection scope of the claims of the present invention.

Claims

1. An electromagnetic target recognition method based on interpretable multi-task learning, characterized in that It includes the following steps: S1: Construct a multi-task learning network for modulation recognition and individual recognition. The backbone network of the multi-task learning network is DRSN. An encoding network is built using the hard sharing method and introducing a cross-stitch module. And a dual decoding branch is set up, using a residual shrinkage module, an average pooling layer, and fully connected layers corresponding to the number of categories. S2: Input the original signal into the multi-task learning network constructed in step S1, set hyperparameters, and conduct training of the multi-task learning network. The hyperparameters at least include signal specifications, batch size, and number of training epochs. S3: Construct a generator for outputting perturbed signals, segment the original signal to be explained. The generator generates a perturbation mask by randomly inverting elements, combines the perturbation mask with the original signal to obtain a perturbed signal. S4: Input the perturbed signal into the multi-task learning network trained in step S2 to obtain the probability distribution of classification results under different tasks. S5: Input the perturbation mask and the similarity of the perturbation masks into a linear model, and perform model fitting in combination with the probability distribution to obtain the interpretation weight W of each subsequence g ; S6: Normalize W g Perform normalization processing and map the weights back to the original signal length, and visualize the time-frequency domain signal region that plays a key role in the model decision based on the constellation diagram annotation.

2. The electromagnetic target recognition method based on interpretable multi-task learning according to claim 1, wherein: The cross-stitching structure in the step S1 cross-stitch module is specifically as follows: Assume that the output of the i-th parallel structure is outA i and outB i , then the inputs inputA i+1 and inputB i+1 of the (i + 1)-th parallel structure are: wherein are trainable parameters.

3. The electromagnetic target recognition method based on interpretable multi-task learning according to claim 2, wherein: In the residual shrinkage module in step S1, there is a soft threshold. Features with soft thresholds close to zero are set to zero, while negative features are retained. The derivative of the feature is 0 or 1. The soft threshold function is defined as: where x is the input feature, y is the output feature, and τ is the threshold parameter.

4. The electromagnetic target recognition method based on interpretable multi-task learning according to claim 1, characterized in that: Among the hyperparameters in step S2, the original signal sample dimension uses I / Q signals, with specifications of (2, 1024), the batch size is 32, and the initial number of training epochs is set to 30, which is adjusted according to the loss change during training.

5. The electromagnetic target recognition method based on interpretable multi-task learning according to claim 1, characterized in that: In the step S3, the original signal is segmented into d sub-sequences, generating a binary mask x ∈ {0, 1} d , generating a perturbation mask x' by randomly inverting the elements of x, and the replacement noise is sampled from a Gaussian distribution , where μ i and σ i are the channel mean and standard deviation estimated based on the dataset at each time step i, and the perturbation signal is synthesized according to x'.

6. The electromagnetic target recognition method based on interpretable multi-task learning according to claim 5, characterized in that: In step S4, the perturbed signal is input into the trained multi-task learning network, and classification results are obtained through different task branches of the network. The Softmax function is used to convert the results into a probability distribution.

7. The electromagnetic target recognition method based on interpretable multi-task learning according to claim 5, characterized in that: In step S5, the predicted probability of the perturbation signal by the multi-task model is used as the supervision label, and a local linear model is trained on the original mask and the perturbation mask. By optimizing the model weights W g As the contribution degree of each subsequence pair to the explanation, the optimization objective function is: where F represents the Frobenius distance, and f model () represents the prediction of the local linear model, and T represents the transpose operation.

8. The electromagnetic target recognition method based on interpretable multi-task learning according to claim 7, wherein: In the said step S6, for the interpretation weight W g The normalization process specifically is as follows:

9. A computer-readable storage medium, characterized in that: It stores a computer program, which when executed by a processor, implements the electromagnetic target recognition method based on interpretable multi-task learning as described in any one of claims 1-8.

10. A computer device, characterized in that: It includes a memory for storing instructions; a processor for executing the instructions, enabling the computer device to perform the operations of the electromagnetic target recognition method based on interpretable multi-task learning as described in any one of claims 1-8.

Images