Multi-resolution feature pre-training-based cross-receiver radiation source fingerprint identification method

Abstract

The invention provides a cross-receiver radiation source fingerprint identification method based on multi-resolution feature pre-training, and the method comprises the steps: extracting features under different resolutions through a multi-resolution tree structure when a deep learning network model of radiation source fingerprint identification is trained, and carrying out the recognition of the radiation source fingerprint through an attention mechanism; according to the method, time correlation of a time sequence is modeled in a multi-resolution mode, features under different resolutions are fused through feature fusion on the basis, long-distance correlation under a coarse resolution and short-distance correlation under a fine resolution are combined, and finally multi-level feature representation containing local details and a global trend is obtained. And richer and more comprehensive characteristics are provided. When the training sample is generated, abundant training information is provided by randomly selecting the original data and the converted data of the original data. And finally, a resolution weighting technology is introduced into classification prediction, abnormal points are resisted while noise is suppressed, and generalization performance improvement of the model in a complex time sequence mode is finally realized.

Description

Background Art

[0001] Radiation source "fingerprint" identification or specific emitter identification SEI refers to measuring the characteristics of received electromagnetic signals and determining the individual radiation source that generates the signals based on existing prior information. The characteristics of the radiation source individual are representative and are generated by factors such as machine thermal noise, equipment non-linearity, and harmonic interference. Such characteristics are not easily forged and are unique. Whether it is a mobile phone, a radio station, or a radar system, due to the slight differences in their hardware and software, unique "fingerprints" will be left on the transmitted radio signals. These fingerprints can be reflected in multiple aspects such as the amplitude, phase, frequency, and modulation of the signal. By capturing and analyzing these fingerprint features, accurate identification of the emission source can be achieved.

[0002] However, the electromagnetic environment is intricate, and radiation source data may pass through different scenarios. These scenarios may have different influencing factors such as channel conditions, humidity, and wind speed, which usually bring about a considerable domain change between training and test data. Inevitably, during the acquisition process of radiation source signals, they are affected by various factors such as natural factors like weather, wind speed, and season, as well as the non-linearity of hardware receiving devices. The signals are interfered with, resulting in changes in the signal feature distributions collected under different scenarios. Although collecting more training data can alleviate this problem, it is impractical for the expensive and time-consuming labeling process.

[0003] The core of the radiation source individual identification algorithm is the extraction of radiation source fingerprint features. Traditional methods for extracting radiation source fingerprint features based on expert experience are mainly divided into three categories, including methods based on direct measurement of signal parameters, methods based on mathematical transformation, and methods based on analysis of transmitter mechanisms. However, these three methods are limited by the understanding of signal essence, mathematical tools, and transmitter mechanisms, making it difficult to understand and extract the complex essential features of radiation sources, and they have poor robustness in different environments. With the rapid development of deep learning methods, radiation source individual identification and classification based on convolutional neural network CNN have received extensive attention and achieved excellent performance.

[0004] The implementation of radiation source individual identification and classification based on convolutional neural network can generally be divided into three categories. A large part directly inputs the original signal waveform or transformed information into a relatively simple convolutional neural network for training and learning. A part focuses on improving the neural network or feature extractor to extract more refined and representative features. Another part focuses on the physical characteristics and hardware characteristics of the device or assigns a unique signature to each device.

[0005] In practical applications, the acquisition of radiation source signals will inevitably be affected by various factors such as weather, wind speed, season and other natural factors, as well as hardware non-linearity. The target domain to be evaluated is not certain. However, most classification algorithms based on convolutional neural network (CNN) are limited to the training domain and have poor data classification performance in new scenarios with large differences. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a cross-receiver fingerprint recognition method that can eliminate the influence of receiver hardware differences for the problem that signals of the same transmitter may exhibit different fingerprint characteristics on different receivers due to hardware differences between different receivers.

[0007] The technical solution adopted by the present invention to solve the above technical problem is a cross-receiver radiation source fingerprint recognition method based on multi-resolution feature pre-training, which is characterized in that the deep learning network model for radiation source fingerprint recognition is trained according to the following steps:

[0008] Given that a radiation source emits an original time series, electromagnetic signals are received by different receivers as received samples, and the received samples are enhanced to obtain training samples;

[0009] The training samples are represented by multi-resolution tree structure features; in the multi-resolution tree structure, the lower-resolution features as the parent nodes are feature-connected with the features of their child nodes to reflect the cross-scale correlation between coarser and finer resolution features; adjacent nodes within each resolution are feature-connected to reflect the temporal correlation of features at the corresponding resolution;

[0010] The attention mechanism is used to extract the inter-scale and intra-scale temporal correlation features in the multi-resolution tree structure feature representation;

[0011] For other node features except the features of the bottom layer nodes in the multi-resolution tree structure, the features generated by upsampling the lower-resolution features are concatenated with the current first-level resolution features to obtain the current first-level fused features; the fused features of the bottom layer are the bottom layer resolution features themselves;

[0012] Each level of fused features of the training samples is used for sequence prediction and classification prediction of radiation source fingerprints. The training target for sequence prediction is that the predicted sequence is close to the original time series, and the training target for classification prediction is that the predicted classification result is close to the true radiation source type.

[0013] Preferably, during the training process of sequence prediction, the sum of the MSE losses calculated by comparing the predicted sequences at different resolutions with the original time series is used as the loss of sequence prediction to constrain the training process. When calculating the sum of the MSE losses at different resolutions, corresponding learnable weighting coefficients are set for different resolutions to ensure the robustness of the multi-scale representation of electromagnetic radiation signal features during the cross-scenario migration process, and finally improve the generalization performance of the radiation source recognition model in a complex electromagnetic environment.

[0014] The present invention uses a multi-resolution tree structure to extract features at different resolutions; then, through an attention mechanism, the temporal correlation of the time series is modeled in a multi-resolution manner. On this basis, the features at different resolutions are fused through feature fusion, combining the long-distance correlation at a coarser resolution with the short-distance correlation at a finer resolution, and finally obtaining a multi-level feature representation containing local details and global trends, providing richer and more comprehensive features. When generating training samples, rich training information is provided by randomly selecting the original data and the data after transformation of the original data, further enhancing the generalization performance of the model.

[0015] Furthermore, finally, a resolution weighting technique is introduced in the classification prediction to suppress noise and resist outliers at the same time.

[0016] In existing radiation source fingerprint recognition schemes, many methods ignore the long-distance temporal correlation or choose complex long-distance correlation calculation methods, which will lead to a lack of algorithm generalization ability or complex calculation and difficulty in training. The present invention adopts a more concise method to capture the long-distance correlation at a coarser resolution. Under the condition of a small amount of calculation, it takes into account the feature correlation of short-distance and long-distance time series signals, improving the radiation source recognition accuracy.

[0017] The beneficial effects of the present invention are as follows: Feature fusion is carried out at different resolutions, combining the long-distance correlation at a coarser resolution with the short-distance correlation at a finer resolution, and finally obtaining features that combine long and short distance correlations at a finer resolution, which can provide richer and more comprehensive features than existing methods. Robust feature extraction can achieve good cross-domain recognition accuracy for radiation source signals. In addition, by using the resolution weighting technique to suppress noise and resist outliers at the same time, the generalization performance of the radiation source recognition model in a complex electromagnetic environment is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flowchart of the present invention;

[0019] Figure 2 is a schematic diagram of the multi-resolution tree structure feature representation. DETAILED DESCRIPTION OF THE INVENTION

[0020] In practical applications, the acquisition of signals will inevitably be affected by various factors such as weather, wind speed, seasons and other natural factors, as well as hardware non-linearity. That is, in reality, the target domain to be evaluated is not certain. However, most classification algorithms based on the convolutional neural network (CNN) are limited to the training domain and have poor performance in classifying data in new scenarios with large differences. Aiming at the problem of poor performance of traditional methods and existing deep learning methods in the cross-receiver scenario, the present invention models this problem from the perspective of unsupervised domain generalization and trains a deep learning model capable of cross-receiver radiation source fingerprint recognition.

[0021] This method aims to alleviate the problem of performance degradation of the deep learning model when the data distribution changes (Domain Shift), and particularly focuses on the reduction of the radiation source recognition accuracy of the model in the cross-receiver scenario. At the same time, the proposed solution can also handle other domain shift situations including channel changes.

[0022] Unsupervised domain generalization aims to address the following challenges: training a robust model that can generalize across multiple different but related domains without any labeled data from the target domain. This setting is more difficult than unsupervised domain adaptation because no data from the target domain is accessible.

[0023] The main idea of unsupervised domain generalization is to capture task-related cross-domain invariant features from the source domain data while suppressing the interference of domain-related task-irrelevant features on the model's decision-making. Doing so can enable the trained model to generalize well in new and unseen domains.

[0024] The present invention obtains features at different resolutions based on multi-resolution feature extraction and multi-resolution feature fusion from the perspective of domain generalization, and combines the long-distance correlation at a coarser resolution with the short-distance correlation at a finer resolution to obtain a multi-level feature representation containing local details and global trends, providing richer and more comprehensive features. The specific process is shown in Figure 1 .

[0025] Step1. Data augmentation

[0026] Data augmentation is to generate more different but related training samples by performing a series of transformation operations on the existing training data to improve the performance and generalization ability of the model.

[0027] Given the original time series sequence X emitted by a known radiation source, these electromagnetic signals are received using different receivers as original samples, and transformation operations are applied to the original samples, including adding white noise, random masking, random translation, random linear stretching, etc., to generate more samples with increased data, and the expanded samples are used as the training set, providing more training information to the model, which is beneficial for the model to learn richer feature patterns. At the same time, data augmentation is equivalent to introducing some meaningful noise and perturbations into the training set, guiding the model to focus on the essential feature representations of cross-scene invariance, while reducing non-robust associations that are sensitive to local perturbations, thereby improving the model's generalization ability to new data and reducing the risk of overfitting. At the same time, randomness is introduced in data selection, and in each training iteration, the original samples or augmented samples are randomly selected to participate in the training, effectively enhancing the traversal diversity of the training data while maintaining semantic consistency.

[0028] Data augmentation provides richer training information for the deep learning model, enhances the generalization ability, simulates the real scenario, alleviates the data distribution problem, improves the data utilization rate, and thus effectively improves the performance of the deep learning model.

[0029] Step2. Multi-resolution tree structure feature representation

[0030] After the training samples are first embedded, the feature nodes are initialized at different resolutions to facilitate the subsequent exchange of information between these nodes. Specifically, through L convolutional layers, features F at different resolutions s from fine scale to coarse scale are obtained from top to bottom i (s) , F i (s) represents the i-th feature node with resolution s. The connections between features at different resolutions construct a multi-resolution multi-tree, where each parent node P i (s) has multiple child nodes such as Figure 2 . In the multi-resolution tree structure, the feature connections between the parent node and its child nodes reflect the features of the original time series across resolutions, and the connections between adjacent features within the same resolution simulate the temporal correlation of the features at that resolution. By connecting features within a coarser scale resolution, it is easier to capture the long-range dependence of features at the coarser scale.

[0031] Step3. Attention mechanism feature extraction

[0032] This module uses the attention mechanism to extract the inter-scale and intra-scale temporal correlation features of the time series in a multi-resolution manner. It can be divided into two parts:

[0033] 1. Inter-scale connection

[0034] The above part mentions that connections between features of different resolutions form a multi-scale tree, where each parent node has multiple child nodes. The parent node has a coarser resolution relative to its child nodes, and the child nodes have a finer resolution relative to their parent node. Different resolutions from fine to coarse can be analogized to time from a finer scale to a coarser scale, such as milliseconds, seconds, minutes to hours. The time correlation of features at different resolutions is captured through the attention mechanism to obtain the connections between features of different resolutions. The set of inter-scale neighborhood at feature node i is ∪ represents the union.

[0035] 2. Intra-scale connections

[0036] By connecting adjacent features of the same resolution, for the connections between feature nodes of finer resolutions, short-range correlations at finer resolutions can be obtained, and for the connections between feature nodes of coarser resolutions, long-range correlations at coarser resolutions can be obtained. Capturing long-range correlations at coarser resolutions through the attention mechanism is much more concise than single long-range correlations at finer resolutions. Assume that the set of same-scale neighborhoods of the i-th feature node among A adjacent feature nodes of the same resolution is

[0037] Combining inter-scale and intra-scale feature connections, such as Figure 2 , the set of relevant feature nodes at node i can be obtained

[0038]

[0039] where,

[0040] s ∈ {1, 2,..., S} represents the scale level serial number, s = 1 is the finest granularity, S is the highest scale level, C is the branching factor between scales, that is, the number of feature nodes of finer resolutions related to each feature node, A is the node window length related to node i within the scale, and I is the length of the original time series.

[0041] The set of cross-scale parent nodes P i (s) represents the feature aggregation path from fine granularity to coarse granularity. When and only when the current scale is not the top layer (s ≤ S - 1), each feature node F i (s) determines its corresponding parent node at the previous scale s + 1 through operation, denotes rounding up. This operation is equivalent to pooling and compressing the features of C consecutive time windows. For example, when C = 4, the feature node F at the second-level scale s = 2 in seconds 3 (2) associates with the feature node F at the third-level scale s = 3 in minutes 1(3) 。

[0042] Cross-scale sub-node set Realize the decomposition and traceability of features from coarse-grained to fine-grained. When at a non-bottom layer scale (s≥2), each feature node F i (s) Can be decomposed into C sub-nodes at the next scale s-1, and its index range follows the interval ((i-1)C, iC]. For example, the node of the hourly feature (s = 3, C = 4) Will correspond to the To Four sub-nodes at the minute-level scale (s = 2).

[0043] Same-scale neighborhood set Construct a local time context window by defining the neighborhood radius (A-1) / 2. In particular, when dealing with boundary nodes (such as i = 1 or i = I / C s-1 ), the set range will be automatically truncated to the valid index interval. The parity design of the parameter A ensures a symmetric neighborhood structure. For example, when A = 5, the attention range of each node includes the two time steps before and after and the current node.

[0044] Set of relevant feature nodes Obtain the feature of feature node i through the attention mechanism

[0045]

[0046] Among them, Is the set of relevant feature nodes, and the exponential term Represents the query vector q of the current node i i And the key vector k of the neighborhood node j j Similarity measure. The scaling factor Is used to suppress the numerical inflation of the dot product operation in the high-dimensional space. The value vector v j Is the feature projection of the neighborhood node j, and its weight is dynamically modulated by the exponential term, reflecting the contribution intensity of node j to the current node i. The normalization term Sums the similarities of all nodes within the relevant feature node To ensure the probability normalization of the attention weights.

[0047] Step4. Feature fusion

[0048] Multiple preliminary effective feature layers with different resolutions can be obtained through the previous attention mechanism feature extraction. In this step, the preliminary effective feature layers are mainly used for feature fusion, and the way of feature fusion is to upsample and stack the feature layers. Specifically, except for the feature of the bottom layer node in the tree structure, the features of other nodes are all fused using the lower (coarser) resolution feature F ′(s+1) The feature generated by upsampling UpSample (restoring the time dimension resolution) and the current first-level resolution feature (N is the total number of feature nodes at resolution s), are concatenated through the concat operation to achieve multi-scale feature fusion under the guidance of coarse-grained semantics; when s = S, the lowest resolution feature is directly retained. Upsampling can make the lower resolution feature containing low long-distance time correlation retain the long-distance time correlation and become a higher resolution at the same time, and then perform the concat operation with the higher resolution feature with short-distance time correlation, so that the long-range temporal dependence carried by the low resolution feature and the local detail information of the high resolution feature complement each other, forming a composite feature with multi-scale representation ability, ensuring that the temporal patterns of different time granularities are co-represented in the feature space.

[0049] The fused feature F of each s layer ′(s) is expressed as:

[0050]

[0051] Step5. Sequence prediction

[0052] Take the fused feature F ′(s) as the radiation source fingerprint, and send the radiation source fingerprint into the linear layer Linear to obtain the prediction result Y of the original time series at different resolutions (s) = w (s) F ′(s) + b (s) , where, w (s) represents the prediction weight at resolution s, and b (s) represents the bias at resolution s.

[0053] During the training process of the deep learning network model, the MSE loss is calculated by comparing the prediction results at different resolutions with the original time series X to constrain the training process, so that the deep learning network model can extract the essential features of the radiation source across receivers. The present invention proposes a resolution-weighted method, adding S (the total number of resolutions) learnable parameters ω ′s to the model, that is, when calculating the MSE loss L, it is weighted according to the coarseness of its resolution to enhance the multi-scale feature representation ability of the model for the radiation source signal, resist noise and outliers, and improve the generalization ability:

[0054]

[0055] Among them, x i is the input data point i, and ω ′s is the MSE loss weight of the feature layer with a resolution of s.

[0056] Step6. Classification prediction

[0057] Send all the fused features of the training samples into the multi-layer perceptron MLP module of the deep learning network model for multi-classification prediction of the radiation source to obtain the prediction result. For example, all the fused features of sample i After being sent into the multi-layer perceptron MLP module, the prediction result of sample i is output Use the prediction result and the true label Calculate the cross-entropy loss to prompt the deep learning network model to learn the classification task well:

[0058]

[0059] Among them, N is the number of batch samples, M represents the total number of radiation source categories, is the true label (one-hot encoding) of sample i in category c, is the predicted probability normalized by Softmax. The gradient characteristic of the loss function can effectively guide the multi-scale features to be optimized in the direction of category discriminability: when deviates from the true probability, a large gradient update amount will be generated in the backpropagation process, forcing the network to enhance the category discriminability of cross-resolution features.

[0060] The multi-resolution feature extraction and feature fusion method based on Transformer takes into account the feature correlation of short-distance and long-distance time-series signals under the condition of a small amount of computation, making the long-range time-series dependence carried by the low-resolution features complementary to the local detail information of the high-resolution features, and forming composite features with multi-scale representation ability. During the training process of the deep learning network model, by constraining the sequence prediction and classification prediction, the trained deep learning network model has a high cross-domain radiation source recognition accuracy. In subsequent actual application scenarios, when the received radiation source signal is input into the trained deep learning network model, the deep learning network model can output the recognition result of the radiation source.

Claims

1. A cross-receiver emitter fingerprint recognition method based on multi-resolution feature pre-training, characterized in that: The deep learning network model for radiation source fingerprint recognition is trained according to the following steps: It is known that the radiation source emits an original time series, and different receivers are used to receive electromagnetic signals as received samples, and data enhancement is performed on the received samples to obtain training samples; Represent the training samples with multi-resolution tree structure features; The lower resolution feature of the parent node in the multi-resolution tree structure is feature-connected with the higher resolution feature of its child node to reflect the cross-scale correlation of the lower and higher resolution features; the neighboring nodes in each resolution are feature-connected to reflect the temporal correlation of the features at the corresponding resolution; The attention mechanism is used to extract the inter-scale and intra-scale temporal correlation features in the multi-resolution tree structure feature representation; Except for the node features at the bottom layer in the multi-resolution tree structure, all other node features are generated by upsampling the lower resolution features and concatenating them with the current first-level resolution features to obtain the current first-level fusion features; the bottom-level fusion features are the bottom-level resolution features themselves; Each level of fusion features of the training samples is used as the radiation source fingerprint for sequence prediction and classification prediction. The training goal of sequence prediction is to make the predicted sequence close to the original time series sequence, and the training goal of classification prediction is to make the predicted classification result close to the actual radiation source type.

2. The method according to claim 1, characterized in that: In the training process of sequence prediction, the training process is constrained by calculating the sum of the MSE losses by comparing the predicted sequences at different resolutions with the original time series as the loss of sequence prediction; When calculating the sum of the MSE losses at different resolutions, different resolutions are set with corresponding learnable weighting coefficients to enhance the model's ability to characterize the multi-scale features of radiation source signals.

3. The method according to claim 1 or 2, characterized in that: The sequence prediction is completed by the linear layer in the deep learning network model, and the predicted sequence output by the linear layer is used as the prediction result of the original time series sequence.

4. The method according to claim 1, characterized in that: The classification prediction is completed by the multi-layer perceptron MLP in the deep learning network model, and the multi-layer perceptron MLP outputs the prediction probability normalized by Softmax as the prediction classification result.

5. The method according to claim 4, characterized in that: During the training process of classification prediction, the training process is constrained by calculating the cross entropy loss between the predicted classification results and the true labels.

6. The method according to claim 1, characterized in that: Data augmentation includes adding white noise, random masking, random translation, and random linear stretching.

7. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

8. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method of claim 1.

Images