An intelligent electromagnetic target recognition method and system based on spatio-temporal joint perception

Abstract

The present invention provides an intelligent electromagnetic target recognition method and system based on spatio-temporal joint perception, belonging to the technical field of electromagnetic signal recognition. It scans the wireless channel to obtain periodically collected signals. For the periodically collected electromagnetic signals, it first converts them into time-frequency diagrams, then realizes frequency point separation by means of models such as YOLO, and then uses the RSSI information of the sampling segments and the spatio-temporal joint positioning algorithm based on radio frequency fingerprints to achieve spatial separation and obtain grid space address coding. Next, by constructing the dataset R and using the spatio-temporal joint attention mechanism, spatio-temporal separation is realized to obtain the electromagnetic recognition target result. By separating overlapping electromagnetic signals in multiple dimensions of "time-space-frequency", accurate positioning of electromagnetic targets can be achieved under low SNR conditions and dynamic scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromagnetic signal recognition, and particularly relates to an intelligent electromagnetic target recognition method and system based on spatio-temporal joint perception. Background Art

[0002] With the rapid development and popularization of radio technologies and devices, the wireless electromagnetic environment has become increasingly complex, and various types of electromagnetic signals may be mixed in the signals at the same time period. Precise detection of suspicious / target signals existing in a certain frequency band from the complex wireless environment and timely control and processing have broad application prospects in many fields.

[0003] Early electromagnetic signal recognition technologies were mainly statistical perception methods (such as energy detection, cyclic stationary feature detection, matched filter detection, covariance matrix detection method, etc.), and their detection accuracy and false alarm rate and other indicators deteriorated significantly under low signal-to-interference ratio conditions. With the development of technology, machine learning-related technologies such as decision trees and support vector machines have been applied in the field of electromagnetic signal recognition. By extracting multi-dimensional features such as the frequency domain and time domain of signals for pattern classification and correlation analysis, the performance of electrical signal recognition has been significantly improved. However, since its recognition of electromagnetic signals depends on the extraction and processing of features, it has the problem of a large workload. In recent years, deep learning technology has gradually become the mainstream research direction of current electromagnetic signal recognition, which benefits from its excellent automatic feature extraction and expression capabilities. By automatically mining the internal laws and deep features in a large number of high-dimensional samples, the performance and efficiency of this technology have been greatly improved compared with traditional technologies.

[0004] At present, there are still many challenges in deep learning-based electromagnetic recognition methods: on the one hand, it comes from the common problems of deep learning-based technologies, such as the need for a large number of signal samples for model training, poor generalization, redundant parameters, lack of samples, etc.; on the other hand, it comes from the complexity and dynamics of electromagnetic signals. The electromagnetic spectrum environment is an open space, and signals from various communication systems, radars, and other electronic devices overlap. How to separate different types of electromagnetic signals mixed in unknown wireless signals is one of the current problems. In addition, due to factors such as spatial selective fading, shadow, hidden terminal, and target device mobility in the wireless channel, the SNR of electromagnetic signals for electronic detection is low, resulting in a sharp decline in the performance of existing electromagnetic recognition methods.

[0005] In addition, existing electromagnetic identification methods often process from a single signal source (such as the main signal of cognitive radio), a single feature dimension (such as energy, correlation feature), and a single node (difficult to eliminate the influence of multipath, etc.), without considering comprehensively from a systematic perspective. For example, most image classification-based methods often convert the electromagnetic signal data at a certain moment into a 2D image, and then use models such as CNN / YOLO for identification, but they do not fully consider the feature information of electromagnetic signals in the time dimension. The time series classification-based methods convert the electromagnetic signal data into a time series, fully mining the time-related feature information of the signal, but they face the problem of insufficient utilization of spatial correlation. In response to these problems, an electromagnetic identification method based on multi-point cooperation has emerged. For example, Chinese Patent ZL202411463040.X we have applied for discloses a collaborative electromagnetic signal intelligent identification method and system, which performs fusion processing on the potential electromagnetic target lists reported by each collaborative sensing node. The method comprehensively considers the states of each collaborative sensing node, the consistency of time-frequency positions, the consistency of graphic forms, etc., and performs weighted processing, improving the accuracy of the identification results; however, it still fails to solve the problem of multi-signal overlap, and has poor identification performance and insufficient adaptability in low SNR conditions and dynamic scenario conditions.

[0006] In view of this, it is necessary to further improve the existing technology. Summary of the Invention

[0007] Aiming at the deficiencies of the existing technology, the present invention provides an intelligent electromagnetic target identification method and system based on spatio-temporal joint perception, aiming to solve the problem of complex spatial electromagnetic signal overlap and separation, and improve the identification performance of target electromagnetic signals under low SNR conditions and the adaptability in dynamic scenarios.

[0008] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0009] An intelligent electromagnetic target identification method based on spatio-temporal joint perception, which is realized by means of an intelligent electromagnetic target identification system. The intelligent electromagnetic target identification system at least includes a joint processing and identification unit capable of information interaction and multiple collaborative sensing nodes. The method includes the following steps:

[0010] Step S1, the collaborative sensing nodes scan the wireless channel, collect electromagnetic signals at time intervals of period T, and obtain their respective periodic acquisition signals;

[0011] Step S2, each collaborative sensing node separately performs frequency point separation on its obtained periodic acquisition signals to form electromagnetic signal information indexed by frequency points;

[0012] Step S3: The collaborative sensing node transmits the electromagnetic signal information indexed by frequency points obtained by frequency separation to the joint processing and recognition unit. Meanwhile, the collaborative sensing node calculates the RSSI information of the time-frequency map sampling segments of each frequency point and uploads it to the joint processing and recognition unit. The joint processing and recognition unit first performs spatial separation on the corresponding electromagnetic signal information indexed by frequency points according to the obtained RSSI information to obtain the positions of electromagnetic target nodes, and then performs grid space address coding on the electromagnetic targets based on the positions of the electromagnetic target nodes. The joint processing and recognition unit aggregates the electromagnetic signal information indexed by frequency points and the grid space address coding information uploaded by each collaborative sensing node to construct a dataset R of a spatio-temporal-frequency 3D index in real time;

[0013] Step S4: The joint processing and recognition unit obtains the final electromagnetic target recognition result through spatio-temporal joint attention recognition based on the dataset R of the spatio-temporal-frequency 3D index constructed in step S3.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects:

[0015] The present invention proposes an intelligent electromagnetic target recognition method based on spatio-temporal joint sensing for the problem of complex spatial electromagnetic signal overlapping separation. For the collected electromagnetic signals, they are first converted into time-frequency maps, then frequency separation is achieved by means of models such as YOLO, and then spatial separation is realized by using the RSSI information of the sampling segments in combination with the spatio-temporal joint positioning algorithm based on radio frequency fingerprints to obtain the grid space address coding. Next, by constructing the dataset R and using the spatio-temporal joint attention mechanism, spatio-temporal separation is achieved to obtain the electromagnetic recognition target result; through the multi-dimensional separation of overlapping electromagnetic signals in "time-space-frequency", the recognition performance under low SNR conditions and the adaptive ability in dynamic scenarios can be improved, and the accurate positioning of moving electromagnetic targets can be realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, form a part of this application, and do not limit the embodiments of the present invention. In the drawings:

[0017] Figure 1 is the overall architecture schematic diagram of the intelligent electromagnetic target recognition system based on spatio-temporal joint sensing of the present invention;

[0018] Figure 2 is the recognition flow chart of the intelligent electromagnetic target recognition method based on spatio-temporal joint sensing of the present invention;

[0019] Figure 3 is the sampling timing schematic diagram of the data acquisition process of the present invention;

[0020] Figure 4 is the flow schematic diagram of frequency separation of the present invention;

[0021] Figure 5 It is the geospatial grid layout diagram constructed by the joint processing and recognition unit of the present invention;

[0022] Figure 6 It is the schematic flow diagram of the multi-channel radio frequency fingerprint joint positioning algorithm used in the present invention;

[0023] Figure 7 It is the schematic structural diagram of the spatio-temporal multi-channel graph neural network model of the present invention. Detailed implementation manners

[0024] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with embodiments. The illustrative embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0025] In view of the characteristics of wireless electromagnetic signal heterogeneity and diversity, multi-dimensional space correlation, diverse sources and incompleteness, etc., the present invention proposes a deep learning technical solution based on time-space-frequency joint perception from a systematic perspective. By separating overlapping electromagnetic signals in multiple dimensions of "time-space-frequency", intelligent recognition of electromagnetic targets is achieved under low SNR conditions and dynamic scenario conditions.

[0026] Specifically, as Figures 1 to 7 shown, the present invention provides an intelligent electromagnetic target recognition method based on spatio-temporal joint perception, which is implemented by means of an intelligent electromagnetic target recognition system. The intelligent electromagnetic target recognition system at least includes a joint processing and recognition unit capable of information interaction and multiple cooperative sensing nodes (see Figure 1 , the number of cooperative sensing nodes is at least two, Figure 1 and three are exemplarily shown), and the method specifically includes the following steps:

[0027] Step S1: The cooperative sensing nodes scan the wireless channel, collect electromagnetic signals at time intervals of period T, and obtain their respective periodic acquisition signals;

[0028] Specifically, in this step S1, the periodic acquisition signals are embodied in the form of digital sequences. The sampling duration of each period T is Td, and Td < T. The length of the digital sequence corresponding to each sampling duration Td is p, and the formed digital sequence can be expressed as:

[0029] ;

[0030] In the formula, k is the number of periodic samplings. Each sampling duration Td corresponds to p digital information. For example, It is indicated that p pieces of corresponding digital information are obtained in the first sampling period T1, and the p-th digital information among them is , It represents the k-th sampling period, and both p and k are positive integers;

[0031] Preferably, the start time of each sampling duration Td is the start time of the corresponding sampling period T (see Figure 3 ), so as to ensure the synchronization of the sampling periods of each collaborative perception node;

[0032] In addition, in step S1, the data acquisition process includes but is not limited to operations such as filtering, down-conversion, and A / D conversion. The sampling timing schematic diagram is as Figure 3 shown.

[0033] Step S2: Each collaborative perception node separately performs frequency point separation on the periodically acquired signals it obtains to form electromagnetic signal information indexed by frequency points;

[0034] Specifically, for the obtained periodically acquired signal, the corresponding collaborative perception node first converts it into a time-frequency diagram by using the short-time Fourier transform. Preferably, the conversion is performed using the following formula:

[0035] ;

[0036] In the formula, is the DFT (Discrete Fourier Transform) of the data after windowing at time m, is the n-th digital information in the formed digital sequence, n ∈ {1, 2,..., p}, is the window function, f is the frequency, j is the imaginary unit, and p is the length of the digital sequence corresponding to the sampling duration Td;

[0037] Then, the corresponding collaborative perception node uses a target detection network model such as YOLO to analyze the time-frequency diagram, identifies potential electromagnetic target information from the time-frequency diagram, and based on the identified potential electromagnetic target information, divides the time-frequency diagram into a series of time-frequency diagram sampling segments indexed by frequency. For each time-frequency diagram sampling segment of each frequency point, the inverse short-time Fourier transform is performed one by one to form electromagnetic signal information indexed by frequency points, as Figure 4 shown.

[0038] Step S3: The cooperative perception node transmits the electromagnetic signal information indexed by frequency points obtained by frequency separation to the joint processing and recognition unit. At the same time, the cooperative perception node calculates the RSSI information of the time-frequency map sampling segments of each frequency point and uploads it to the joint processing and recognition unit. The joint processing and recognition unit serves as the data fusion center. First, it performs spatial separation on the corresponding electromagnetic signal information indexed by frequency points according to the obtained RSSI information to obtain the positions of electromagnetic target nodes, and then performs grid space address encoding on the electromagnetic targets based on the positions of the electromagnetic target nodes. The joint processing and recognition unit aggregates the electromagnetic signal information indexed by frequency points and the grid space address encoding information uploaded by each cooperative perception node to construct a dataset R with a spatio-temporal-frequency 3D index in real time;

[0039] In a preferred embodiment, in step S3, in order to obtain the node positions of the electromagnetic targets, the joint processing and recognition unit uses a spatio-temporal joint positioning method based on radio frequency fingerprints to achieve this. Specifically, the spatio-temporal joint positioning method based on radio frequency fingerprints includes the following steps:

[0040] S3.1: The joint processing and recognition unit takes its own location F 0 as the center and constructs a radio frequency map using a nine-square grid or a similar form with 2D grid encoding (as shown in Figure 5 , which can also be called a geospatial grid layout map. Note: The number of grids and the size of each grid can be adjusted according to the actual scenario);

[0041] S3.2: Based on the constructed radio frequency map, obtain the basic positioning model through offline training and get the final position result:

[0042] S3.21: Traverse each grid position of the radio frequency map by moving the test device (transmitting target) used to obtain the radio frequency fingerprints; for traversal;

[0043] S3.22: The cooperative perception node ( m is the cooperative perception node index, 0, 1, 2,...; Exemplarily, the first cooperative perception node is represented by S 0 ) uses the digital sequence information (from the test device) received in multiple T periods (each period Td) to calculate the corresponding RSS data and AOA data, and constructs a dataset , where , and i is the time period index;

[0044] S3.23: Send the dataset corresponding to each cooperative perception node into a time series neural network such as RNN / LSTM to extract the features of multi-channel signals , this feature is in the plural form and includes a real part and an imaginary part;

[0045] S3.24. Combine the imaginary part and the real part of each feature and stack the features of multiple collaborative perception nodes (here, the combination and stacking can be in the form of multiplication, or simply data stacking, or other methods) to construct a multi-channel two-dimensional feature map (for example, see Figure 6 );

[0046] S3.25. Pass the multi-channel two-dimensional feature map through the SE channel attention mechanism, and perform Squeeze and Excitation operations in sequence to obtain the weights of each channel (essentially the received information of each collaborative perception node), and then perform weighted calculation (such as multiplying channel by channel) on these weights and the multi-channel two-dimensional feature map to obtain the final weighted feature representation;

[0047] S3.26. Use a fully connected neural network for prediction and output the initial recognition result of the position where the grid is located ;

[0048] S3.27. By comparing the initial recognition result of the position where the grid is located obtained by prediction and each grid position obtained by traversal (that is, comparing the predicted value and the true value), use loss functions such as the LSME criterion or cross-entropy combined with optimization algorithms such as stochastic gradient descent (SGD) to train the neural network model (the model here can be a deep learning neural network model such as CNN, RNN, Transformer, etc.) to obtain a basic positioning model (its output is the position result);

[0049] S3.28. Further combine the output result of the basic positioning model with a Kalman filter through the Adaboost mechanism to obtain the final position result output. Here, the Adaboost mechanism and the Kalman filter are used to eliminate uncertain influences such as data acquisition noise using historical determined position information;

[0050] S3.3. Online and real-time obtain the node positions of electromagnetic targets:

[0051] S3.31. The collaborative perception node online and real-time collects wireless signal data from electromagnetic targets (for example, it can be a moving electromagnetic target, Figure 1 as shown by radio target one or two), and then uses the digital sequence information (from the electromagnetic target) received in multiple T periods (each period Td) to calculate the corresponding RSS data and AOA data, and splices them into the data set to obtain a spliced data set ;

[0052] S3.32. Feed the spliced dataset into the basic positioning model for processing (refer to S3.28) to obtain the monitoring target position, which is the node position of the electromagnetic target.

[0053] Step S4. The joint processing recognition unit obtains the final electromagnetic target recognition result through spatio-temporal joint attention recognition based on the spatio-temporal-frequency 3D indexed dataset R constructed in step S3;

[0054] Specifically, in step S4, the joint processing recognition unit performs spatio-temporal joint attention recognition by constructing a spatio-temporal multi-channel graph neural network model. Among them, the spatio-temporal multi-channel graph neural network model adopts a hierarchical network structure. The construction of the bottom layer graph adopts a dynamic graph method, and a graph attention mechanism neural network is used to extract complex spatial dependence relationships based on information such as position and input them into the upper-layer time-domain sequence attention model in real time for extracting feature information in the time dimension. Finally, the judgment result is output through a fully connected layer, as Figure 7 shown.

[0055] More specifically, the spatio-temporal multi-channel graph neural network model mainly includes a data construction layer, a spatial domain graph attention layer, a time-domain sequence attention layer, and a fully connected layer. Among them, the following settings are adopted for each layer:

[0056] (1) Data construction layer

[0057] The data construction layer uses the input data to construct spatial graphs for each time period in chronological order , , where is the vertex of the graph in the t time period, is the corresponding edge; the wireless signal data collected by the cooperative perception node in the cycle i is expressed as:

[0058] ,

[0059] Then the vertex feature data of the spatial graph is

[0060] ;

[0061] To reflect the mutual relationship of signals in space, the adjacency matrix does not adopt the topological connection or spatial distance between nodes, but is characterized by the similarity of the path vectors of signal transceiver nodes. Specifically, let the spatial positions of the cooperative perception node and the target node be and , then the transceiver node path vector in its plural form is ; Similarly, for the collaborative perception node , the transceiver node path vector in its plural form is ;

[0062] Define the adjacency matrix A of the network nodes and and (vertices) on the spatial graph as the similarity between the vector and the vector . Then the new adjacency matrix can be characterized as the product form of vector correlation and vector length difference, resulting in:

[0063] ;

[0064] Among them, on the right side of the equal sign "=" in the formula, the whole before the symbol "*" is the first term, and the whole after the symbol "*" is the second term. The first term represents the angular correlation between the vector and the vector , and the second term represents the relative magnitude of the amplitudes of the vector and the vector ; If the vector and the vector are in the same direction, the first term reaches the maximum value of 1; If and the vector have the same amplitude, the second term reaches the maximum value of 1;

[0065] (2) Spatial graph attention layer

[0066] Compared with the existing graph attention mechanism, the present invention adopts a new adjacency matrix to replace the traditional adjacency matrix, and then combines it with the GCN mechanism to form a new graph attention mechanism. This mechanism not only saves the computational complexity of the existing graph attention mechanism, but also introduces the correlation between vertex features into the GCN model. Correspondingly, the feature update can be expressed as:

[0067] ,

[0068] In the formula, H is a parameter, ko is the number of updates, representing the ko-th update. Each calculation of the parameter H needs to be updated. represents the degree matrix of the collaborative perception node; σ is a non-linear activation function; A is the adjacency matrix, composed of elements; W is a weight parameter; Since the perception degree matrices of all collaborative perception nodes are the same under the same fusion center, can be directly omitted, and the expression is simplified to ;

[0069] (3) Time-domain sequence attention layer

[0070] The input time series at different times is correlated, and the correlation at different times is also different. Therefore, this layer uses a time attention mechanism to calculate the temporal dependence relationship and extracts features using the following formula :

[0071] ;

[0072] In the formula, is the query vector, is the query weight matrix, is the input vector; is the key vector, is the key weight matrix, is the value vector, is the value weight matrix; is the dimension of the vector;

[0073] (4) Fully connected layer

[0074] Using the feature information (complex number) obtained by spatio-temporal joint processing, the decision category is obtained through the fully connected layer and the softmax function, and the decision result is output.

[0075] The joint processing recognition unit performs spatio-temporal joint attention recognition on the dataset R of the spatio-temporal-frequency 3D index constructed in step S3 with the help of the above spatio-temporal multi-channel graph neural network model, so as to identify the final electromagnetic target recognition result.

[0076] It should be understood that the present invention also relates to an electromagnetic target intelligent recognition system based on spatio-temporal joint perception. The electromagnetic target intelligent recognition system at least includes a joint processing recognition unit capable of information interaction and a plurality of cooperative perception nodes. Among them, the cooperative perception nodes are mainly responsible for the acquisition of electromagnetic signals and frequency point separation; the joint processing recognition unit is responsible for the spatial separation of electromagnetic signals and realizes spatio-temporal separation based on the spatio-temporal multi-channel graph neural network model, so as to finally realize the separation and recognition of complex spatial electromagnetic signals.

[0077] In order to better achieve the purpose of the present invention, the joint processing recognition unit also applies the final electromagnetic target recognition result to spatial separation to optimize the accuracy of target positioning in subsequent links.

[0078] It should be noted that the specific embodiments described above can enable those skilled in the art to understand the present invention more comprehensively, but do not limit the present invention in any way. Therefore, although the present specification and embodiments have described the present invention in detail, those skilled in the art should understand that the present invention can still be modified or equivalently replaced; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered by the protection scope of the patent of the present invention. Any reference signs in the claims should not be construed as limiting the claimed invention.

Claims

1. An electromagnetic target intelligent recognition method based on spatiotemporal joint perception, the method is implemented with the aid of an electromagnetic target intelligent recognition system, the electromagnetic target intelligent recognition system at least comprising a joint processing recognition unit capable of information exchange and a plurality of collaborative sensing nodes, characterized in that: The method includes the following steps: Step S1: The cooperative perception nodes scan the wireless channel, collect electromagnetic signals at time intervals of period T, and obtain their respective periodic acquisition signals; Step S2: Each cooperative perception node separately performs frequency point separation on its obtained periodic acquisition signal to form electromagnetic signal information indexed by frequency points; Step S3: The cooperative perception nodes transmit the electromagnetic signal information indexed by frequency points obtained through frequency point separation to the joint processing and recognition unit. At the same time, the cooperative perception nodes calculate the RSSI information of the time-frequency map sampling segments of each frequency point and upload it to the joint processing and recognition unit. The joint processing and recognition unit first performs spatial separation on the corresponding electromagnetic signal information indexed by frequency points according to the obtained RSSI information to obtain the positions of electromagnetic target nodes, and then performs grid space address coding on the electromagnetic targets based on the positions of the electromagnetic target nodes. The joint processing and recognition unit aggregates the electromagnetic signal information indexed by frequency points and the grid space address coding information uploaded by each cooperative perception node to construct a data set R of a spatio-temporal-frequency 3D index in real time; Step S4: The joint processing and recognition unit obtains the final electromagnetic target recognition result through spatio-temporal joint attention recognition according to the data set R of the spatio-temporal-frequency 3D index constructed in step S3; In step S3, in order to obtain the node positions of the electromagnetic targets, the joint processing and recognition unit uses a spatio-temporal joint positioning method based on radio frequency fingerprints to achieve this. Among them, the spatio-temporal joint positioning method based on radio frequency fingerprints includes the following steps: S3.1: The joint processing and recognition unit takes its own location F0 as the center and constructs a radio frequency map with 2D grid coding; S3.2: Based on the constructed radio frequency map, obtain the basic positioning model through offline training and obtain the final position result; S3.3: Obtain the node positions of electromagnetic targets online in real time.

2. The electromagnetic target intelligent recognition method based on time-space joint perception according to claim 1 is characterized in that: In step S1, the periodic acquisition signal is embodied in the form of a digital sequence. The sampling duration of each period T is Td, and Td < T. The length of the digital sequence corresponding to each sampling duration Td is p, and the formed digital sequence is expressed as: ； In the formula, k is the number of periodic sampling, and each sampling time Td corresponds to p digital information, among which the pth digital information is expressed as , represents the kth sampling period, where p and k are both positive integers.

3. The electromagnetic target intelligent recognition method based on time-space joint perception as claimed in claim 1 is characterized in that: In step S2, for the obtained periodic acquisition signal, the corresponding cooperative perception node first converts it into a time-frequency map by using the short-time Fourier transform. Then, the corresponding cooperative perception node uses the target detection network model to analyze the time-frequency map, identifies potential electromagnetic target information from the time-frequency map, and based on the identified potential electromagnetic target information, divides the time-frequency map into a series of time-frequency map sampling segments indexed by frequency. For each time-frequency map sampling segment of each frequency point, the inverse short-time Fourier transform is performed one by one to form electromagnetic signal information indexed by frequency points.

4. The electromagnetic target intelligent recognition method based on time-space joint perception according to claim 1 is characterized in that: Step S3.2 specifically includes: S3.

21. Move the test device used to obtain the RF fingerprint to each grid position of the RF map. Conduct traversal; S3.

22. Collaborative sensing nodes Using the digital sequence information received in multiple T periods, the corresponding RSS data and AOA data are calculated and a data set is constructed. ,in, , i is the time period index; S3.

23. The data set corresponding to each collaborative sensing node Send it to the time series neural network to extract the characteristics of multi-channel signals , this feature is a complex number, including real and imaginary parts; S3.

24. Each feature The imaginary and real parts of are combined, and the features of multiple collaborative sensing nodes are superimposed to construct a multi-channel two-dimensional feature map; S3.25: Pass the multi-channel two-dimensional feature map through the SE channel attention mechanism, perform compression and excitation operations in sequence to obtain the weights of each channel, and then perform weighted calculation on these weights and the multi-channel two-dimensional feature map to obtain the weighted final feature representation; S3.

26. Use a fully connected neural network to predict and output the initial recognition result of the grid location ; S3.

27. Obtain the initial recognition result of the grid location by comparing and predicting And each grid position obtained by traversal , the basic positioning model is obtained by training the neural network model using the LSME criterion or the cross entropy loss function combined with the optimization algorithm; S3.

28. The output result of the basic positioning model is further combined with the Kalman filter through the Adaboost mechanism to obtain the final position result output.

5. The electromagnetic target intelligent recognition method based on time-space joint perception as claimed in claim 4 is characterized in that: Step S3.3 specifically includes: S3.

31. Collaborative sensing nodes Online real-time collection of wireless signal data from electromagnetic targets, and then use the digital sequence information received in multiple T cycles to calculate the corresponding RSS data and AOA data, and splice them into the data set , get the spliced ​​dataset ; S3.

32. Concatenate the datasets The basic positioning model is sent to the basic positioning model, and its output result is further combined with the Kalman filter through the Adaboot mechanism to obtain the final position result output, which is the monitoring target position of the electromagnetic target, that is, the node position of the electromagnetic target.

6. A method for intelligent electromagnetic target recognition based on spatiotemporal joint perception as claimed in any one of claims 1 to 5, characterized in that: In step S4, the joint processing recognition unit performs spatiotemporal joint attention recognition by constructing a spatiotemporal multi-channel graph neural network model. The spatiotemporal multi-channel graph neural network model includes a data construction layer, a spatial domain graph attention layer, a temporal sequence attention layer and a fully connected layer.

7. The electromagnetic target intelligent recognition method based on time-space joint perception according to claim 6 is characterized in that: In the data construction layer, the spatial graph of each time period is constructed in chronological order. In order to reflect the mutuality of signals in space, an adjacency matrix is ​​constructed. , the adjacency matrix The similarity of the signal sending and receiving node path vectors is used for characterization. in, Set up collaborative sensing nodes and the target node The spatial positions are and , then the path vector of the sending and receiving nodes is The plural form of ; For collaborative sensing nodes , the path vector of the sending and receiving nodes The plural form of ; Then, the adjacency matrix Characterized by: ； Among them, on the right side of the equal sign "=", the whole before the symbol "*" is the first term, and the whole after the symbol "*" is the second term. The first term represents the vector With vector The angle correlation of With vector The relative size of the amplitude.

8. An electromagnetic target intelligent recognition system based on time-space joint perception, which is used to implement an electromagnetic target intelligent recognition method based on time-space joint perception as described in any one of claims 6-7, characterized in that: The collaborative sensing node is responsible for the collection and frequency separation of electromagnetic signals; the joint processing and identification unit is responsible for the spatial separation of electromagnetic signals and realizes spatiotemporal separation based on the spatiotemporal multi-channel graph neural network model, thereby ultimately realizing the separation and identification of complex spatial electromagnetic signals.

9. The electromagnetic target intelligent recognition system based on time-space joint perception as claimed in claim 8, characterized in that: The joint processing and identification unit also applies the final electromagnetic target recognition results to spatial separation to optimize the accuracy of target positioning in subsequent links.

Images