Harmonic Radar Target Recognition Method, Device, and Storage Medium Based on Subgraph Fusion

Abstract

The present invention discloses a harmonic radar target recognition method based on subgraph fusion, which relates to the technical fields of data recognition and data representation, and includes the following steps: S1. The harmonic radar emits detection signals in three bands of L, S, and K, respectively performs frequency sweep irradiation on each type of non-linear electronic target, and collects the reflection signals corresponding to the three bands, and obtains a reflection signal set for each type of non-linear electronic target; preprocess each reflection signal set to obtain a characteristic curve; S2. Convert each characteristic curve into a characteristic subgraph, obtain a regularized characteristic subgraph through a graph convolutional network, and then fuse the regularized characteristic subgraphs in the form of subgraph fusion to generate a multi-head large graph; S3. Input the multi-head large graph into a graph neural network, obtain the embedded representation of the multi-head large graph, and splice the embedded representations of the multi-head large graph; S4. Input the spliced embedded representation of the multi-head large graph into a fully connected layer, and output the target category. The present invention also discloses an electronic device and a computer-readable storage medium.

Description

Technical Field

[0001] The present invention relates to the technical field of data recognition and data representation, and in particular, to a harmonic radar target recognition method, device, and storage medium based on subgraph fusion. Background Art

[0002] A harmonic radar is a radar system that detects targets by transmitting a fundamental wave signal, receiving second, third, or even higher-order harmonic signals from non-linear targets, and detecting the targets. It is mainly applied in many fields such as public security technical investigation, security guard, and privacy protection. In recent years, the academic research on harmonic radar has mainly focused on the detection of non-linear nodes. For example, there is a harmonic radar switch terminal detection device that is convenient to carry and use, and there is also a polarization-insensitive third-harmonic radar system and a signal transmission method that achieve polarization insensitivity of fundamental wave excitation and third-harmonic feedback between a radar base station and a terminal, as well as a vehicle driving safety warning system for detecting pedestrians based on a harmonic radar, which detects non-linear nodes on electronic devices carried by pedestrians through the harmonic radar.

[0003] However, using harmonic signals to detect intelligent devices belongs to a relatively new detection technology. The above technologies for detecting intelligent devices, that is, non-linear electronic devices, basically stay at the stage of detecting the presence or absence, and cannot achieve the identification and classification of different non-linear nodes. There are some related technologies for this problem currently: There is a harmonic radar based on a Field Programmable Gate Array (FPGA) and deep learning, which performs pulse compression, pulse accumulation, and identification detection on harmonic signals through a harmonic receiver; there is also a harmonic radar target classification method and system based on machine learning, which transmits a mixing signal, changes the transmission power, calculates the transmission harmonic response power curve characteristics of different second-order intermodulation points as the harmonic radar target classification characteristics, extracts the target characteristics of the harmonic radar by using the harmonic information of multiple second-order intermodulation points, and realizes the classification of target characteristics based on machine learning algorithms. However, these harmonic detection technologies mainly rely on the harmonic energy characteristics of a single band of the target. Due to their low feature dimensions, the detection efficiency and confidence are not good, and it is difficult to distinguish devices with information leakage functions.

[0004] In pattern recognition, it is necessary to extract specific parameters as feature quantities. The feature quantities in harmonic signals are frequency, phase and amplitude. The existing technology basically uses amplitude as the research object and uses artificial statistics to find statistics related to amplitude as feature values. However, due to the lack of uniqueness of physical elements and the limitations of artificial statistics, the signal characteristics of the harmonic signal of the detected object cannot be fully extracted, resulting in low accuracy in identifying the type of the detected object. Re-characterization of signal sequences is one of the effective means to improve the accuracy of signal recognition, which can enable deep learning models to mine richer and more suitable features for the task field. The existing technology proposes an adaptive signal graph representation (Adaptive Visibility Graph, AVG) algorithm that is linked to the deep learning model for training. It can automatically represent the signal sequence as a topological structure of the graph, thereby mining more features that fit the characteristics of the task from the perspective of network science, and can form linkage training with the graph neural network to improve the accuracy of signal recognition. However, due to the insufficient mining and representation of harmonic signal features and the lack of an effective learning framework and model, the accuracy of multi-electronic device target recognition under this technology is still not high. Therefore, it is urgent to propose a harmonic radar target recognition method that can effectively characterize harmonic signals and combine deep learning models. Summary of the invention

[0005] The purpose of the present invention is to propose a harmonic radar target recognition method, device, and storage medium based on sub-graph fusion, which transmits radar waves in the form of multi-band frequency sweeping and receives reflected harmonics, calculates the characteristic curve under the frequency sweeping as the physical characteristics of the target object, and combines the sequence graph network representation method and the graph neural network framework in deep learning to achieve sufficient feature mining, so as to solve the defect that it is difficult to effectively identify the type of detection object in the prior art.

[0006] In a first aspect, the present invention provides a harmonic radar target recognition method based on sub-graph fusion, wherein the target includes a type I nonlinear electronic target, where I≥2; the harmonic radar target recognition method comprises the following steps:

[0007] S1. Harmonic radar transmits detection signals of three bands, L, S, and K, and performs frequency sweeping on each type of nonlinear electronic target and collects reflection signals corresponding to the three bands, L, S, and K. At this time, a reflection signal set is obtained for each type of nonlinear electronic target; each reflection signal set is preprocessed to obtain a characteristic curve;

[0008] S2. Convert each feature curve into a feature subgraph, obtain a regularized feature subgraph through a graph convolution network, and then fuse the regularized feature subgraphs into a multi-head large graph in the form of subgraph fusion;

[0009] S3. Input the multi - head large graph into the graph neural network to obtain the embedding representation of the multi - head large graph, and concatenate the embedding representations of the multi - head large graph;

[0010] S4. Input the concatenated embedding representation of the multi - head large graph into the fully - connected layer to output the target category; The fully - connected layer is connected after the graph neural network.

[0011] As a possible implementation, for each type of non - linear electronic target, S1 includes:

[0012] S10. Perform frequency - sweep irradiation within the L, S, and K bands at the fixed transmission power of the harmonic radar to obtain the reflected signals corresponding to the L, S, and K bands, and form a set of reflected signals from the reflected signals corresponding to the L, S, and K bands;

[0013] S11. Pre - process the set of reflected signals to obtain the characteristic curve of the set of reflected signals at this transmission power as the prior feature of the L, S, and K bands.

[0014] As a possible implementation, for each characteristic curve, S2 is specifically: Take each node of the characteristic curve as the graph node of the feature sub - graph; Map the sequence composed of graph nodes to the corresponding adjacency matrix through the AVG algorithm to obtain the feature sub - graphs of the L, S, and K bands; The graph node attribute of the feature sub - graph is the feature of the corresponding node of the characteristic curve, and the graph node relationship of the feature sub - graph is represented by the adjacency matrix, and the graph node relationship is the edge between adjacent graph nodes.

[0015] As a possible implementation, the regularized feature sub - graph is obtained through the graph convolutional network, specifically: Perform hierarchical pooling on the feature sub - graph corresponding to the band through the graph convolutional network to regularize the number of graph nodes of the graph convolutional network and the feature vector of each graph node, and obtain the regularized feature sub - graphs of the L, S, and K bands.

[0016] As a possible implementation, the regularized feature sub - graphs are fused in the form of sub - graph fusion to generate a multi - head large graph, specifically: Extract all graph nodes in the regularized feature sub - graphs of the L, S, and K bands respectively, use the multi - head attention mechanism to traverse all graph nodes, calculate the attention scores between graph nodes to obtain the multi - head structural information, denoted as the multi - head dense adjacency matrix;

[0017] Then sparsify the multi - head dense adjacency matrix, that is, for each graph node, select the top k graph nodes with weights sorted from high to low as neighbors to connect, set the weights of other graph nodes to zero, obtain the sparse adjacency matrix, and then perform weight normalization on the sparse adjacency matrix to obtain the multi - head large graph of the L, S, and K bands.

[0018] As a possible implementation, S3 includes: inputting the multi-headed large graph generated by the fusion of S2 into a graph neural network, obtaining the embedded representation of the multi-headed large graph through message passing and state update of the graph neural network, and splicing the embedded representations of the multi-headed large graph along the last feature dimension.

[0019] As a possible implementation, the harmonic radar target recognition method further includes

[0020] forming a non-linear electronic target data set from the characteristic curves of the L, S, and K band reflection signals of each non-linear electronic target, and randomly dividing the non-linear electronic target data set into a training set and a test set;

[0021] Inputting the multi-headed large graph after fusion of the non-linear electronic targets in the training set into a graph neural network for graph neural network training to obtain a trained graph neural network.

[0022] As a possible implementation, the harmonic radar target recognition method based on subgraph fusion includes, in the training stage, inputting the training set into a graph neural network training model to implement adaptive graph representation and subgraph fusion for parameter update of the graph neural network to obtain a trained graph neural network; in the testing stage, inputting the test set into the trained graph neural network to obtain a classification result and use it to verify the performance of the graph neural network.

[0023] In a second aspect, the present invention provides an electronic device, including a memory and a processor, with a program stored on the memory that runs on the processor, and when the processor runs the program, it executes the steps of the harmonic radar target recognition method based on subgraph fusion provided in the first aspect.

[0024] In a third aspect, the present invention provides a computer-readable storage medium, on which computer instructions are stored, and when the computer instructions run, they execute the steps of the harmonic radar target recognition method based on subgraph fusion provided in the first aspect.

[0025] In a fourth aspect, the present invention provides a harmonic radar target recognition system based on subgraph fusion, which is used to detect and identify non-linear electronic targets and is implemented relying on a harmonic radar.

[0026] As a possible implementation, the harmonic radar includes a signal detection device and a signal processing device. The signal detection device is used to transmit radar signals mainly in three bands of L, S, and K, including but not limited to other bands generated based on spread spectrum technologies such as frequency hopping, and receive the reflection signals of non-linear electronic targets. After that, the received reflection signals are mixed and filtered through a mixing circuit and a filtering circuit to obtain a filtered signal. The filtered signal is converted into a digital signal by a data acquisition module and enters the signal processing device.

[0027] The reflected signal includes harmonic signals of different orders. The signal detection device includes an antenna unit, a transceiver unit, and a data acquisition module. The antenna unit includes a transmitting antenna and a receiving antenna. The transceiver unit includes a mixing circuit and a filtering circuit. The signal processing device includes a signal feature curve generation unit, a feature sub-graph generation unit, a sub-graph fusion unit, a graph convolutional network, a splicing module, and a fully connected layer unit;

[0028] The signal feature curve generation unit includes an L-band feature curve generation unit, an S-band feature curve generation unit, and a K-band feature curve generation unit. The L-band feature curve generation unit, the S-band feature curve generation unit, and the K-band feature curve generation unit are respectively input with the reflected filtered harmonic signals of the L-band transmitting wave, the reflected filtered harmonic signals of the S-band transmitting wave, and the reflected filtered harmonic signals of the K-band transmitting wave, and respectively generate the feature curves of the corresponding bands according to the L-band features, S-band features, and K-band features.

[0029] The L-band features and S-band features include but are not limited to power, envelope, and first-order to fifth-order statistical features. The K-band features include but are not limited to spectrum, time-frequency transformation features, and decomposition features. The time-frequency transformation features include but are not limited to the features generated by wavelet transform, short-time Fourier transform, Garbo transform, and Hilbert-Huang transform (HHT). The decomposition features include but are not limited to the features generated by orthogonal triangular decomposition (QR), singular value decomposition (SVD), empirical mode decomposition (EMD), and variational mode decomposition (VMD);

[0030] The feature sub-graph generation unit is used to generate feature sub-graphs of each band based on the feature curves of each band, including an L-band sub-graph generation unit, an S-band sub-graph generation unit, and a K-band sub-graph generation unit, which are respectively used to generate L-band feature sub-graphs, S-band feature sub-graphs, and K-band feature sub-graphs. The feature sub-graph includes graph nodes and edges. The graph nodes include node numbers and node values. The node numbers are the sequence position numbers corresponding to the L-band feature curve, S-band feature curve, and K-band feature curve. The node values are the values corresponding to the sequence position numbers of the L-band features, S-band features, and K-band feature curves. The edges are generated by training with the AVG algorithm;

[0031] The sub - graph fusion unit performs pooling operations on the L - band feature sub - graph, S - band feature sub - graph, and K - band feature sub - graph respectively, generates edges and replaces the edges to obtain the fused large graph; generating edges and replacing edges is achieved through the self - attention mechanism; the fused large graph is input into the graph convolutional network to output latent features; the latent features are concatenated by the concatenation module and then output the type of non - linear electronic target through the fully - connected layer unit.

[0032] The feature sub - graph generation unit is connected to the sub - graph fusion unit, the sub - graph fusion unit is connected to the graph convolutional network, the graph convolutional network is connected to the concatenation module, and the concatenation module is connected to the fully - connected layer unit.

[0033] The feature sub - graph generation unit generates feature sub - graphs of each band, the sub - graph fusion unit fuses the feature sub - graphs of each band, outputs the fused large graph, the graph convolutional network receives the fused large graph for information transmission and node state update, and outputs latent features; the latent features are input into the concatenation module for concatenation, and the concatenated latent features are classified by the fully - connected layer unit to output the category of the corresponding non - linear electronic target.

[0034] Compared with the prior art, the beneficial effects produced by the present invention are as follows:

[0035] 1. The harmonic radar target recognition method based on sub - graph fusion provided by the present invention relies on the harmonic radar to extract the harmonic power curve and spectrum features as target feature quantities to classify intelligent devices. By mining and enriching the recognition features through the graphical representation of the swept - frequency power curve and spectrum features and multi - band sub - graph fusion, and then using the graph neural network for classification. During the recognition process, the swept - frequency power curve in the received harmonic and the spectrum features in a specific frequency band are used as physical feature media, combined with the graph network representation method of the sequence and the graph neural network framework in deep learning, which can achieve sufficient feature mining and thus realize the effective detection of intelligent devices.

[0036] 2. The harmonic radar target recognition method based on sub - graph fusion provided by the present invention uses detection signals in three bands of L, S, and K. The microwaves in the L - band and S - band have good penetration ability for obstacles and small path loss, which can significantly improve the detection efficiency of electronic devices. And by analyzing basic features such as the amplitude, frequency, and phase response of the echo signal, the type of electronic device target can be efficiently judged; the electromagnetic wave in the K - band has a short wavelength and can well adapt to small - sized electronic devices. Under the condition of keeping the relative bandwidth constant, the K - band can excite and receive signals with a wider bandwidth. Therefore, through the feature fusion of the three bands, the recognition accuracy of the target is greatly improved.

[0037] 3. The harmonic radar target recognition method based on subgraph fusion provided by the present invention re-characterizes common serialized data into graph network data by setting a series of mathematical and physical rules. The graph network is composed of multiple nodes and edges, which describes whether there is an edge relationship between nodes and the strength of the edge relationship. Graph-based data structures can usually mine more features related to the correlation between nodes, and then reflect the local correlation and global correlation between sequence sampling points. Through automatic mining of graph features and graph neural network technologies, richer data features can be effectively obtained for target recognition, so as to obtain more accurate target recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0039] Figure 1 It is a flow chart of a harmonic radar target recognition method based on sub-graph fusion in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of harmonic radar signal acquisition in an embodiment of the present invention;

[0041] Figure 3 Schematic diagram of adjacency matrix sparsification based on multi-head attention mechanism in an embodiment of the present invention;

[0042] Figure 4 Schematic diagram of an end-to-end framework of deep learning harmonic radar target recognition based on subgraph fusion in an embodiment of the present invention;

[0043] Figure 5 The figure is a block diagram of a harmonic radar target recognition system based on sub-image fusion in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to clearly describe the technical solutions of the embodiments of the present invention, in the embodiments of the present invention, words such as "first" and "second" are used to distinguish the same items or similar items with basically the same functions and effects. For example, the first threshold and the second threshold are only used to distinguish different thresholds, and their order is not limited. Those skilled in the art can understand that words such as "first" and "second" do not limit the quantity and execution order, and words such as "first" and "second" do not necessarily limit them to be different.

[0045] It should be noted that in the present invention, words such as "exemplary" or "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0046] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship of associated objects and indicates that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. The following at least one (item) or its similar expression refers to any combination of these items, including any combination of single item (item) or plural items (items). For example, at least one (item) of a, b or c can represent: a, b, c, the combination of a and b, the combination of a and c, the combination of b and c, or the combination of a, b and c, where a, b and c can be single or multiple.

[0047] In the prior art, the intelligent device using harmonic signal detection can only detect the presence or absence, and cannot effectively identify the type of the detected object. Moreover, the existing harmonic detection only relies on the harmonic energy characteristics on a single target band and only uses the amplitude as the research object. This feature has a low dimension, poor detection efficiency and confidence, and it is difficult to distinguish devices with information leakage functions. To solve the above problems, the present invention proposes a harmonic radar target recognition method, device and storage medium based on subgraph fusion, which emits radar waves in the form of multi-band frequency sweeping and receives the reflected harmonics, calculates the characteristic curve under frequency sweeping as the physical characteristics of the target object, and combines the graph network representation method of the sequence and the graph neural network framework in deep learning to achieve sufficient feature mining, so as to solve the defect that the type of the detected object cannot be effectively identified in the prior art.

[0048] In a first aspect, the present invention provides a harmonic radar target recognition method based on subgraph fusion, the targets including class I non-linear electronic targets, I≥2; see Figure 1 , the harmonic radar target recognition method includes the following steps:

[0049] S1. The harmonic radar emits detection signals in three bands of L, S and K, respectively performs frequency sweeping irradiation on each type of non-linear electronic target and collects the reflected signals corresponding to the three bands of L, S and K. At this time, a reflected signal set is obtained for each type of non-linear electronic target; preprocess each reflected signal set to obtain a characteristic curve;

[0050] A harmonic radar is a monitoring device based on radio frequency waves, including a harmonic signal acquisition circuit and an antenna. The harmonic signal acquisition circuit includes an acquisition card and an acquisition front end. The acquisition card is connected to the computer terminal. The computer terminal outputs transmission and reception signals, which are converted into transmission and reception commands by the acquisition card and sent to the acquisition front end. After the acquisition front end collects data, it outputs the data to the computer terminal through the acquisition card, and deep learning recognition processing is performed on the computer terminal. The acquisition front end is applicable to the L, S, and K frequency bands and is integrated for both transmission and reception. The harmonic radar is based on the fact that a non-linear target will generate corresponding re-radiated radio frequency waves for radio frequency waves of different frequencies; the re-radiated radio frequency waves of the non-linear target depend on the harmonic characteristics of the target. The harmonic radar is used to detect, track, and classify non-linear targets.

[0051] As the number of classification targets increases, the classification features provided by a single band are not sufficient to achieve ideal recognition accuracy. Therefore, it is necessary to fuse the features of multiple bands to assist each other. The microwaves in the L band and the S band have good penetration ability for obstacles and small path loss. Using electromagnetic waves in this band can significantly improve the detection efficiency of electronic devices. Moreover, by analyzing basic features such as the amplitude, frequency, and phase response of the echo signal, the type of the target can be efficiently judged; the electromagnetic waves in the K band have a short wavelength, and due to the influence of the random antenna effect on the re-radiation effect of harmonics, the millimeter wave wavelength in the K band can well adapt to small electronic devices and is suitable for detecting electronic devices with relatively small feature sizes and high integration levels. In addition, when the relative bandwidth is kept constant, the K band can excite and receive signals with a wider bandwidth.

[0052] As an example, the emission frequency range of the L band is 1 - 2 GHz, such as 1 GHz and 2 GHz; the emission frequency range of the S band is 2 - 4 GHz, such as 2 GHz, 3 GHz, and 4 GHz. The power curve is obtained by using the frequency sweep method to mine features. The emission frequency of the K band is a single frequency point of 24 GHz, and the spectrum curve is used as a priori feature.

[0053] As a possible implementation method, refer to Figure 2 , use a harmonic radar acquisition device to perform frequency sweep irradiation and collection of reflection signals on type I non-linear electronic targets respectively. During the irradiation and collection process, detection signals in the L, S, and K bands are used respectively, including but not limited to other band radar signals generated based on spread spectrum technologies such as frequency hopping, for example, signals such as single carrier, amplitude, frequency, phase modulation, spread spectrum, and frequency hopping. The received and collected signals are mixed and filtered with the mixed frequency signals through a mixing circuit.

[0054] As a possible implementation method, for each type of non-linear electronic target, S1 includes:

[0055] S10. Under the fixed transmission power of the harmonic radar, perform swept-frequency irradiation within three bands of L, S, and K to obtain the reflected signals corresponding to the three bands of L, S, and K, and form a reflected signal set from the reflected signals corresponding to the three bands of L, S, and K;

[0056] As an example, irradiate type-I nonlinear electronic targets in the form of swept frequency respectively and collect the reflected signals to obtain a cluster of echo signals, then continue to collect to obtain multiple clusters of echo signals, and form a reflected signal set.

[0057] S11. Preprocess the reflected signal set to obtain the characteristic curve of the reflected signal set under this transmission power as the prior feature of the L, S, and K bands.

[0058] As an example of preprocessing, for each target signal, obtain the power curve of the L and S band signals and the spectral curve of the K band signal to form the corresponding nonlinear electronic target data set, and randomly divide it from the data set according to a certain ratio to obtain the training set and the test set.

[0059] The power curve describes the power distribution of the signal at different frequencies or times. This curve can show how the power of the signal changes with frequency or time. For periodic signals, the power spectrum shows the power components of the signal at different frequencies; for non-periodic signals, the power spectrum usually describes the power distribution of the signal at different time periods.

[0060] As an example of obtaining the characteristic curve of the reflected signal set, by preprocessing a cluster of echo signals in the L and S bands obtained by swept frequency under the fixed transmission power of the harmonic radar, the power curve of the echo signal under this power can be obtained as the prior feature of the L and S bands:

[0061] s fk ∈R d ,

[0062] S L =[s f1 ,...s fk ...,s fn ,k∈(1,n),

[0063] P LW =f P (S L )=[p f1 ,...p fk ...,p fn ,k∈(1,n);

[0064] Among them, r is a real number, R d is a real vector with dimension d, d is the length of the echo signal, n represents the swept frequency range, s fkRepresents the frequency f k The echo signal S at L Represents the swept-frequency echo signal P in the L band fk Indicates the power curve P of the k-th L band LW Represents the power curve in the L band. The processing method of the S-band power curve is similar. Replace L in the formula with S to represent the power curve in the S band

[0065] Spectral characteristics refer to the representation of a signal in the frequency domain, which describes the energy or power distribution of the signal at different frequencies. It shows the components of the signal on the frequency axis, that is, which frequencies have how much energy or power. Spectral analysis is achieved by transforming the signal into the frequency domain, usually using Fourier transform or related techniques

[0066] For the echo signal in the K band, obtain the spectral curve of the echo signal through Fourier transform as a prior feature

[0067] Fre K = FFT(s K ),

[0068] where S K is the single-frequency point echo signal in the K band, and Fre K represents the spectrum in the K band

[0069] S2. Convert each feature curve into a feature subgraph, obtain the regularized feature subgraph through the graph convolutional network, and then fuse the regularized feature subgraphs in the form of subgraph fusion to generate a multi-head large graph

[0070] As a possible implementation, the feature curve is drawn by the corresponding curve sequence. For each feature curve, S2 is specifically: use each node of the feature curve as the graph node of the feature subgraph; map the sequence composed of graph nodes to the corresponding adjacency matrix through the AVG algorithm to obtain the feature subgraphs in the L, S, and K bands; the graph node attribute of the feature subgraph is the feature of the corresponding node of the feature curve, and the graph node relationship of the feature subgraph is represented by the adjacency matrix, and the graph node relationship is the edge between adjacent graph nodes

[0071] As an example, for the power curves in the L and S bands, use the sampling points as graph nodes; for the spectral curve in the K band: due to the irregularity of the K-band spectral curve, adopt the idea of segmentation, use the subsequence as the graph node, and map each L, S, and K curve sequence to the corresponding adjacency matrix through the AVG algorithm to obtain the feature subgraph, that is, arrange the feature sequences obtained by convolution at different scales and dimensions along the diagonal parallel direction, which can be mapped into the graph G = <V, E>, where V represents the set of nodes, v i , v j ∈V represents two nodes, E represents the set of connecting edges, (vi , v j ) ∈ E represents v i and v j has a direct dependency. It can be represented by an n×n feature matrix as follows:

[0072]

[0073]

[0074] Among them, φ represents the feature sequence obtained by convolution; Conv s is convolution, which performs convolution operations on all sample signals in the dataset; P 1 is the edge connection weight of the nodes in the graph network; m is the upper limit of the value of the size of the convolution kernel; M represents the adjacency matrix of the feature subgraph obtained by arranging the convolution sequences along the diagonal.

[0075] As a possible implementation, the regularized feature subgraph is obtained through the graph convolutional network. Specifically: the feature subgraphs corresponding to the bands are subjected to hierarchical pooling through the graph convolutional network to regularize the number of graph nodes and the feature vectors of each graph node in the graph convolutional network, obtaining the regularized feature subgraphs of the L, S, and K bands.

[0076] As a possible implementation, the regularized feature subgraphs are fused in the form of subgraph fusion to generate a multi-head large graph. Specifically: all graph nodes in the regularized feature subgraphs of the L, S, and K bands are extracted respectively, and the multi-head attention mechanism is used to traverse all graph nodes to calculate the attention scores between the graph nodes to obtain the multi-head structure information, denoted as the multi-head dense adjacency matrix;

[0077] A s = ReLu(Attention s (S)),

[0078] where S ∈ R N×F represents a real matrix belonging to a vector that is N×F, R is a real number, A ∈ R N×N represents the structure information, s ∈ [1, S] represents the attention head, and S represents the number of attention heads.

[0079] As a possible implementation, the multi-head dense adjacency matrix is further sparsified. That is, for each graph node, the top k graph nodes with the highest weights are selected as neighbor connections, and the weights of other graph nodes are set to zero to obtain the sparsified adjacency matrix. Then, the weights of the sparsified adjacency matrix are normalized to obtain the multi-head large graphs of the L, S, and K bands.

[0080] As an example of adjacency matrix sparsification, see Figure 3, perform adjacency matrix sparsification to reduce the computational cost of subgraph convolution. For each node, select the top k graph nodes with the highest weights in descending order as neighbor connections. For example, the top 3:

[0081]

[0082] Among them, idx is the index, which is the counting unit of the sequence. Those beyond the range of k are not included in the index, indicating taking the top k data with larger weights. The i-th number traverses all nodes, and : indicates traversing to the last number.

[0083] Set the weights of non-connected nodes to zero, and keep other weights unchanged. Since the sum of the weights of all neighbor nodes should be equal to 1, perform weight normalization on the sparse adjacency matrix. Through normalization, synthesize all the graph network structures of the L, S, and K band signals into a large graph with a multi-head adjacency matrix. The formula is as follows:

[0084]

[0085] Among them, -idx represents the elements excluding the idx index; Normalize represents the normalization function.

[0086] S3. Input the multi-head large graph into the graph neural network to obtain the embedding representation of the multi-head large graph, and concatenate the embedding representations of the multi-head large graph;

[0087] As a possible implementation, S3 includes: Input the multi-head large graph generated by the fusion in S2 into the graph neural network, obtain the embedding representation of the multi-head large graph through message passing and state update of the graph neural network, and concatenate the embedding representations of the multi-head large graph along the last feature dimension.

[0088] S4. Input the concatenated embedding representation of the multi-head large graph into the fully connected layer and output the target category; The fully connected layer is connected after the graph neural network.

[0089] As a possible implementation, the harmonic radar target recognition method further includes

[0090] Form a non-linear electronic target data set from the characteristic curves of the L, S, and K band reflection signals of each non-linear electronic target, and randomly divide the non-linear electronic target data set into a training set and a test set;

[0091] As an example, the ratio of the training set to the test set is 7:3.

[0092] Input the fused multi-head large graph of the non-linear electronic targets in the training set into the graph neural network, perform the training of the graph neural network, and obtain the trained graph neural network.

[0093] As an example, the optimization process of the end-to-end framework training based on the graph neural network is specifically as follows:

[0094] Input the fused multi-head large graph into the Graph Convolutional Networks (GCN), and denote the output result of the multi-head attention graph as h s , concatenate the multi-head outputs as the input of the two-layer fully connected network, and then the classification output of the graph neural network can be obtained using the Softmax activation function:

[0095]

[0096] Among them, || represents the vector concatenation operation, connecting the individual output results of the multi-head attention graph together; H is the concatenated feature vector; Label represents the classification result; max is the maximum value label function; Softmax represents the activation function; FC2 represents two fully connected layers.

[0097] As a possible implementation, the harmonic radar target recognition method based on subgraph fusion includes, in the training stage, inputting the training set into the graph neural network training model to achieve adaptive graph representation and subgraph fusion for parameter update of the graph neural network, obtaining the trained graph neural network; in the testing stage, inputting the test set into the trained graph neural network to obtain the classification result and use it to verify the performance of the graph neural network.

[0098] As an example, see Figure 4 , establish a GCN graph neural network, use the fused large graph as the input of the graph neural network, use the target category as the output of the graph neural network, and connect it with the preprocessing in step S11, graph representation and subgraph fusion in S2, etc., to form a training and testing framework from the input end to the output end. In the training stage, input the training set into the graph neural network training model to achieve adaptive graph representation and subgraph fusion for parameter update of the graph neural network, obtaining the trained graph neural network; in the testing stage, input the test set into the trained graph neural network to obtain the classification result and use it to verify the performance of the graph neural network.

[0099] In a second aspect, the present invention provides an electronic device, including a memory and a processor, with a program stored on the memory and running on the processor. When the processor runs the program, it executes the steps of the harmonic radar target recognition method based on subgraph fusion provided in the first aspect.

[0100] In a third aspect, the present invention provides a computer-readable storage medium, on which computer instructions are stored. When the computer instructions run, they execute the steps of the harmonic radar target recognition method based on subgraph fusion provided in the first aspect.

[0101] Fourthly, the present invention provides a harmonic radar target recognition system based on subgraph fusion, see Figure 5 , which is used to detect and identify non-linear electronic targets and is realized based on a harmonic radar.

[0102] As a possible implementation, the harmonic radar includes a signal detection device and a signal processing device. The signal detection device is used to transmit radar signals mainly in three bands of L, S, and K, including but not limited to other band radar signals generated based on spread spectrum technologies such as frequency hopping, and receive the reflected signals of non-linear electronic targets. Then, the received reflected signals are mixed and filtered through a mixing circuit and a filtering circuit to obtain filtered signals. The filtered signals are transformed into digital signals by a data acquisition module and enter the signal processing device.

[0103] As an example, the reflected signals include harmonic signals of different orders. The signal detection device includes an antenna unit, a transceiver unit, and a data acquisition module; the antenna unit includes a transmitting antenna and a receiving antenna; the transceiver unit includes a mixing circuit and a filtering circuit. The signal processing device includes a signal feature curve generation unit, a feature subgraph generation unit, a subgraph fusion unit, a graph convolutional network, a splicing module, and a fully connected layer unit;

[0104] The signal feature curve generation unit includes an L-band feature curve generation unit, an S-band feature curve generation unit, and a K-band feature curve generation unit; the L-band feature curve generation unit, the S-band feature curve generation unit, and the K-band feature curve generation unit are respectively input with the reflected filtered harmonic signals of the L-band transmitted wave, the reflected filtered harmonic signals of the S-band transmitted wave, and the reflected filtered harmonic signals of the K-band transmitted wave, and respectively generate feature curves of the corresponding bands according to the L-band features, S-band features, and K-band features.

[0105] As an example, the L-band features and S-band features include but are not limited to power, envelope, first-order to fifth-order statistical features; the K-band features include but are not limited to spectrum, time-frequency transformation features, and decomposition features. The time-frequency transformation features include but are not limited to features generated by wavelet transform, short-time Fourier transform, Garbo transform, and Hilbert-Huang transform (HHT); the decomposition features include but are not limited to features generated by orthogonal triangular decomposition (QR), singular value decomposition (SVD), empirical mode decomposition (EMD), and variational mode decomposition (VMD);

[0106] The feature sub - graph generation unit is used to generate feature sub - graphs for each band based on the feature curves of each band, including the L - band sub - graph generation unit, the S - band sub - graph generation unit, and the K - band sub - graph generation unit, which are respectively used to generate the L - band feature sub - graph, the S - band feature sub - graph, and the K - band feature sub - graph; the feature sub - graph includes graph nodes and edges; the graph nodes include node numbers and node values, the node number is the sequence position number corresponding to the L - band feature curve, the S - band feature curve, and the K - band feature curve; the node value is the value corresponding to the sequence position number of the L - band feature, the S - band feature, and the K - band feature curve; the edges are generated through the AVG algorithm training;

[0107] The sub - graph fusion unit performs pooling processing on the L - band feature sub - graph, the S - band feature sub - graph, and the K - band feature sub - graph respectively, generates edges and replaces the edges to obtain the fused large graph; generating edges and replacing the edges are realized through the self - attention mechanism; the fused large graph is input into the graph convolutional network to output latent features; the latent features are spliced through the splicing module and then output the type of the non - linear electronic target after passing through the fully - connected layer unit;

[0108] The feature sub - graph generation unit is connected to the sub - graph fusion unit, the sub - graph fusion unit is connected to the graph convolutional network, the graph convolutional network is connected to the splicing module, and the splicing module is connected to the fully - connected layer unit.

[0109] The feature sub - graph generation unit generates feature sub - graphs for each band, the sub - graph fusion unit fuses the feature sub - graphs for each band, outputs the fused large graph, the graph convolutional network receives the fused large graph for information transmission and node state update, and outputs latent features; the latent features are input into the splicing module for splicing, and the spliced latent features are classified by the fully - connected layer unit to output the category of the corresponding non - linear electronic target.

[0110] Compared with the prior art, the present invention has the following technical effects:

[0111] 1. The harmonic radar target recognition method based on sub - graph fusion provided by the present invention relies on the harmonic radar to extract the harmonic power curve and spectrum features as target feature quantities to classify intelligent devices. By mining and enriching the recognition features through the graphical representation of the swept - frequency power curve and spectrum features and multi - band sub - graph fusion, and then using the graph neural network for classification. During the recognition process, taking the swept - frequency power curve and spectrum features in the received harmonics as physical feature media, combining the graph network representation method of sequences and the graph neural network framework in deep learning, can achieve sufficient feature mining, and thus realize the effective detection of intelligent devices.

[0112] 2. The harmonic radar target recognition method based on sub-graph fusion provided by the present invention uses detection signals of three bands: L, S, and K. Microwaves of L and S bands have good penetration ability on obstructions and small path loss, which can significantly improve the detection efficiency of electronic devices. Moreover, by analyzing the basic characteristics of the echo signal amplitude, frequency, phase response, etc., the type of electronic device target can be efficiently judged; the electromagnetic wave of K band is short in wavelength and can be well adapted to small electronic devices. While maintaining a certain relative bandwidth, K band can excite and receive signals with wider bandwidth. Therefore, the recognition accuracy of the target is greatly improved by the feature fusion of the three bands.

[0113] 3. The harmonic radar target recognition method based on subgraph fusion provided by the present invention re-characterizes common serialized data into graph network data by setting a series of mathematical and physical rules. The graph network is composed of multiple nodes and edges, which describes whether there is an edge relationship between nodes and the strength of the edge relationship. Graph-based data structures can usually mine more features related to the correlation between nodes, and then reflect the local correlation and global correlation between sequence sampling points. Through automatic mining of graph features and graph neural network technologies, richer data features can be effectively obtained for target recognition, so as to obtain more accurate target recognition results.

[0114] Those skilled in the art can understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A harmonic radar target recognition method based on sub-graph fusion for identifying type I nonlinear electronic targets, I≥2, characterized in that: The steps include: S1. The harmonic radar transmits detection signals of L, S, and K bands, performs frequency sweeping on each type of the nonlinear electronic target and collects reflection signals corresponding to the L, S, and K bands. At this time, a reflection signal set is obtained for each type of the nonlinear electronic target; each reflection signal set is preprocessed to obtain a characteristic curve; S2. Convert each feature curve into a feature subgraph, obtain a regularized feature subgraph through a graph convolution network, and then fuse the regularized feature subgraphs into a multi-head large graph in the form of subgraph fusion; S3, inputting the multi-head large image into the graph neural network, obtaining an embedded representation of the multi-head large image, and splicing the embedded representation of the multi-head large image; S4, input the embedded representation of the spliced ​​multi-head large image into the fully connected layer, and output the target category; the fully connected layer is connected after the graph neural network; The method of fusing the regularized feature subgraphs to generate a multi-head large graph in the form of subgraph fusion is as follows: all graph nodes in the regularized feature subgraphs of the L, S, and K bands are extracted respectively, and a multi-head attention mechanism is used to traverse all graph nodes, and the attention scores between the graph nodes are calculated to obtain multi-head structural information, which is recorded as a multi-head dense adjacency matrix; The multi-head dense adjacency matrix is ​​then sparsed, that is, for each graph node, the first k graph nodes with the highest to lowest weights are selected as neighbor connections, and the weights of other graph nodes are set to zero to obtain a sparse adjacency matrix, and then the sparse adjacency matrix is ​​weight-normalized to obtain a large multi-head graph for the L, S, and K bands.

2. The harmonic radar target recognition method based on sub-graph fusion according to claim 1 is characterized in that: For each type of nonlinear electronic target, S1 includes: S10, performing frequency sweeping illumination in the L, S, and K bands at a fixed transmission power of the harmonic radar to obtain reflection signals corresponding to the L, S, and K bands, and forming a reflection signal set by the reflection signals corresponding to the L, S, and K bands; S11. Preprocess the reflected signal set to obtain a characteristic curve of the reflected signal set at the transmission power as a priori characteristics of the L, S, and K bands.

3. The harmonic radar target recognition method based on sub-graph fusion according to claim 1 is characterized in that: For each characteristic curve, S2 is specifically: each node of the characteristic curve is used as a graph node of the characteristic subgraph; the sequence composed of the graph nodes is mapped to the corresponding adjacency matrix through the AVG algorithm to obtain the characteristic subgraphs of the L, S, and K bands; the graph node attributes of the characteristic subgraph are the characteristics of the nodes corresponding to the characteristic curve, the graph node relationship of the characteristic subgraph is represented by the adjacency matrix, and the graph node relationship is the edge between adjacent graph nodes.

4. The harmonic radar target recognition method based on sub-graph fusion according to claim 3 is characterized in that: The regularized feature subgraph is obtained by the graph convolution network, specifically: the feature subgraph of the corresponding band is layered and pooled through the graph convolution network, the number of graph nodes of the graph convolution network and the feature vector of each graph node are regularized, and the regularized feature subgraphs of the L, S, and K bands are obtained.

5. The harmonic radar target recognition method based on sub-graph fusion according to claim 1 is characterized in that: The S3 includes: inputting the multi-head large graph generated by S2 fusion into the graph neural network, obtaining the embedded representation of the multi-head large graph through message passing and state update of the graph neural network, and splicing the embedded representation of the multi-head large graph along the last feature dimension.

6. The harmonic radar target recognition method based on sub-graph fusion according to claim 1 is characterized in that: The harmonic radar target recognition method also includes, A nonlinear electronic target data set is formed from characteristic curves of L, S, and K band reflection signals of each nonlinear electronic target, and a training set and a test set are randomly divided from the nonlinear electronic target data set; The multi-head large image after the nonlinear electronic targets in the training set are fused is input into the graph neural network for graph neural network training to obtain a trained graph neural network.

7. The harmonic radar target recognition method based on sub-graph fusion according to claim 6 is characterized in that: Including, in the training phase, inputting the training set into the graph neural network training model to realize adaptive graph representation, sub-graph fusion to update the parameters of the graph neural network, and obtaining a trained graph neural network; During the testing phase, the test set is input into the trained graph neural network to obtain the classification results and use them to verify the performance of the graph neural network.

8. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory stores a program running on the processor, and the processor executes the steps of the harmonic radar target recognition method based on sub-image fusion as described in any one of claims 1 to 7 when running the program.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed, the steps of the harmonic radar target recognition method based on sub-image fusion according to any one of claims 1 to 7 are executed.

Images