Broadband signal automatic identification method and system based on double-current cross-convergence network

Abstract

The invention provides a broadband signal automatic identification method and system based on a double-current cross-convergence network, the method and system have the characteristics of high automation, high precision and high robustness, the defect that traditional segmentation depends on a fixed threshold and artificial experience is overcome through a grey wolf optimization adaptive frequency spectrum segmentation scheme, and accurate blind segmentation of a broadband power spectrum is realized; local features and time sequence features are decoupled in parallel through a double-current cross-attention network, and deep interaction of heterogeneous features is realized by adopting a bidirectional cross-attention feature fusion mechanism, so that the recognition accuracy under a low signal-to-noise ratio and a complex electromagnetic environment is remarkably improved; meanwhile, according to the whole scheme, the calculation overhead is optimized in the parameter optimization and feature extraction links, the processing efficiency is improved, and the problems that in the prior art, spectrum segmentation is poor in self-adaption, and recognition model robustness is insufficient are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radio communication technology and artificial intelligence technology, in particular to a wideband signal automatic identification method and system based on a double-flow cross-fusion network. BACKGROUND

[0002] With the rapid development of wireless communication technology and the surge in the number of radio equipment, the electromagnetic environment is increasingly complex and crowded, and the management and utilization of spectrum resources face great challenges. In the field of radio spectrum management and monitoring, the wideband signal captured by the receiver usually contains signals of multiple different devices, which are mixed in the time domain and separated in the frequency domain, forming multiple independent "signal islands". This "time domain mixing and frequency domain separation" feature poses a double technical challenge to the automatic detection and identification of signals: on the one hand, it is necessary to accurately segment each signal band, and on the other hand, it is necessary to efficiently identify the signal type and the source device.

[0003] The existing wideband signal detection and identification methods mainly include traditional energy detection method, expert feature-based method and traditional deep learning method. The traditional energy detection method determines the existence of a signal by calculating the power spectrum energy and comparing it with a fixed threshold, which has the advantages of simple implementation and small amount of calculation, but the fixed threshold cannot adapt to the complex and variable spectrum form, and false alarms or missed alarms may occur in a low signal-to-noise ratio environment, and it can only detect the existence of a signal but cannot identify the specific type. The expert feature-based method relies on the complex feature extraction rules designed by human beings, such as cyclic spectrum and high-order cumulants, which are effective for specific signal types but have poor generalization ability, cannot adapt to unknown or variable signals, and are severely dependent on the prior knowledge of domain experts, with low automation level. The traditional deep learning method (such as convolutional neural network) can process signal spectrum graphs, but it is more skilled in extracting local time-frequency features, ignoring the internal correlation of signals in time series, and is difficult to capture long-distance time sequence dependence, resulting in limited recognition accuracy.

[0004] The common problems of the above methods include: the spectrum segmentation relies on human experience or fixed parameters, with poor adaptability; the feature extraction has poor generalization ability and is difficult to cope with complex electromagnetic environment; the model recognition accuracy is low, the robustness is poor, and the computational complexity is high. Therefore, there is an urgent need in the art for a technical solution that can intelligently and adaptively complete spectrum segmentation and achieve high-precision and end-to-end signal recognition, to improve the automation level and recognition reliability of spectrum management. SUMMARY

[0005] The purpose of the present application is to provide a wideband signal automatic identification method and system based on a double-flow cross-fusion network to solve the problems existing in the prior art.

[0006] To achieve the above-mentioned purpose, the present application provides the following solutions: The application provides a wideband signal automatic recognition method based on a double-flow cross-fusion network, comprising: S1. An adaptive spectrum positioning segmentation scheme of GWO-FCME is adopted, the hyperparameters of a forward continuous mean cut algorithm are optimized through a grey wolf optimization algorithm, adaptive blind segmentation of a wideband power spectrum is realized, and multiple independent signal frequency bands are obtained; S2. Digital down-conversion and resampling preprocessing based on the bandwidth of a signal are performed on each segmented signal frequency band, and normalized sub-band IQ data sequences are obtained; S3. The sub-band IQ data sequences are input into a double-flow cross-fusion network, local features of a signal are extracted through a local feature branch, and long-distance time sequence dependent features of the signal are extracted through a time sequence feature branch; S4. A bidirectional cross-attention feature fusion module is adopted to deeply fuse the local features and the long-distance time sequence dependent features, and a signal recognition result is output through a classifier.

[0007] Preferably, step S1 comprises: S11. Fast Fourier transform is performed on received wideband time-domain IQ data, and a full-resolution wideband power spectrum is obtained; S12. A grey wolf optimization algorithm is used to search for an optimal hyperparameter combination of a forward continuous mean cut algorithm on a down-sampling power spectrum with the objective of minimizing a cost function; The cost function is defined as: ; Wherein, is the number of signals detected under the current parameters , is a preset target number, is an average frequency band gap, is a weight coefficient, prevents the minimum value of the denominator from being zero; S13. The optimal hyperparameters are used to perform the forward continuous mean cut algorithm on the full-resolution power spectrum, the noise mean is iteratively calculated, and a final detection threshold is determined, and the formula of the detection threshold is: ; Wherein, is the noise mean after convergence of the iterative calculation, is the best false alarm probability obtained by GWO optimization; S14. Threshold detection and frequency band merging are performed on the power spectrum according to the final detection threshold and a merging threshold, and multiple independent signal frequency bands are segmented.

[0008] Preferably, step S2 comprises: ​S21. Perform frequency domain band-pass filtering on each signal band, and reconstruct the time domain IQ sequence through inverse fast Fourier transform; S22. Perform digital down-conversion on the reconstructed signal to shift the frequency spectrum to baseband; S23. Resample the baseband signal according to the target sampling rate set according to the signal's own bandwidth, and the formula of the target sampling rate is: wherein, is the bandwidth of the signal band itself; S24. Generate normalized, low-data-rate sub-band IQ data sequences as inputs of the dual-flow fusion network.

[0009] Preferably, in step S3, the dual-flow fusion network comprises: a local feature branch that extracts multi-scale local features using a multi-scale residual deep separable convolutional network and an embedded convolution block attention module, wherein the attention module embedded after each convolution block applies channel attention and spatial attention in turn to dynamically refine the feature map; a temporal feature branch that extracts long-distance temporal dependence features of the signal using a stacked bidirectional gated recurrent unit network.

[0010] Preferably, in the local feature branch, the refinement process of the convolution block attention module comprises: First, calculate the channel attention map and multiply it with the input feature map element by element to obtain the intermediate feature: wherein, is the channel attention map, is the input feature map; Then calculate the spatial attention map and multiply it with the intermediate feature element by element to obtain the refined feature map: wherein, is the spatial attention map.

[0011] Preferably, step S4 comprises: S41. Perform adaptive average pooling on the local feature sequence and the temporal feature sequence respectively to generate a local feature vector and a temporal feature vector; S42. Perform two multi-head cross-attention operations: The first operation takes the local feature vector as the query vector and the temporal feature sequence as the key and value, aggregates the relevant information through the multi-head attention mechanism, and obtains a first fusion vector; ​​​Second operation, the time sequence feature vector is used as the query vector, the local feature sequence is used as the key and the value, the related information is aggregated through the multi-head attention mechanism, and a second fusion vector is obtained; S43. The first fusion vector and the second fusion vector are spliced in the feature dimension to obtain a deep fusion feature vector; S44. The deep fusion feature vector is input into a multilayer perception classifier, the prediction probability of each category is calculated through a Softmax function, and a final signal recognition result is output according to a preset confidence threshold.

[0012] The application also provides a wideband signal automatic recognition system based on a double-flow cross-fusion network, comprising: A spectrum segmentation module is configured to perform adaptive blind segmentation on a wideband spectrum. A signal preprocessing module is configured to perform digital down-conversion and resampling preprocessing on the segmented signal frequency band. A classification and recognition module is configured to extract features and perform fusion classification through a double-flow cross-fusion network. A data management and interaction module is configured to provide a user interaction interface, receive wideband signal data and parameters input by a user, call the above modules to perform processing, and visually display and store the recognition result and a time-frequency graph.

[0013] Preferably, the spectrum segmentation module comprises: A power spectrum acquisition unit is configured to perform fast Fourier transform on wideband time-domain IQ data to generate a power spectrum. A grey wolf optimization unit is configured to iteratively optimize the hyperparameters of a forward continuous mean cut algorithm on the down-sampled power spectrum. A forward continuous mean cut unit is configured to calculate a detection threshold using the optimal hyperparameters and perform frequency band segmentation and merging.

[0014] Preferably, the signal preprocessing module comprises: A digital down-conversion unit is configured to shift the signal spectrum to the baseband. A resampling unit is configured to adaptively adjust the sampling rate according to the signal bandwidth. A data normalization unit is configured to generate a normalized sub-band IQ data sequence.

[0015] Preferably, the classification and recognition module comprises: A local feature extraction unit is configured to extract local features through a multi-scale residual deep separable convolutional network and a convolution block attention module. A time sequence feature extraction unit is configured to extract time sequence features through a bidirectional gated recurrent unit network. A feature fusion unit is configured to realize deep interaction between local and time sequence features through a bidirectional cross-attention feature fusion module. The classification output unit is configured to output a signal category and a confidence level by using a multilayer perception classifier.

[0016] The present application has the following beneficial technical effects relative to the prior art: The wideband signal automatic recognition method and system based on a double-flow cross-fusion network have the characteristics of high automation, high precision and high robustness. The adaptive spectrum segmentation scheme combining the grey wolf optimization (GWO) algorithm and the forward continuous mean elimination (FCME) algorithm is used to overcome the defects of the traditional segmentation relying on a fixed threshold and artificial experience, and to realize accurate blind segmentation of a wideband power spectrum. The double-flow cross-fusion network is used to parallelly decouple local features and time sequence features, and a bidirectional cross-attention feature fusion mechanism is used to realize deep interaction of heterogeneous features, thereby significantly improving the recognition accuracy in a low signal-to-noise ratio and a complex electromagnetic environment. Meanwhile, the parameter optimization and feature extraction links are optimized to reduce the calculation overhead and improve the processing efficiency, thereby effectively solving the problems of poor spectrum segmentation adaptability and insufficient robustness of the recognition model in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0018] Figure 1 The figure is a schematic diagram of the overall architecture of the wideband signal automatic recognition method and system based on a double-flow cross-fusion network provided by the present application. Figure 2 The figure is a schematic diagram of the overall architecture of the wideband signal automatic recognition method and system based on a double-flow cross-fusion network provided by the present application. Figure 3 The figure is a detailed structure block diagram of the double-flow cross-fusion network used in the classification and recognition module in the present application. Figure 4 The figure is a schematic diagram of the adaptive segmentation process combining the grey wolf optimization (GWO) algorithm and the forward continuous mean elimination (FCME) algorithm used in step S1 in the present application. Figure 5 The figure is a schematic diagram of the adaptive threshold detection effect on the wideband power spectrum in step S1 in the present application. Figure 6 The figure is a time domain and frequency domain diagram of the first type of device signal (such as LINK-K) after being processed by step S2 in the embodiments of the present application. Figure 7 The figure is a time domain and frequency domain diagram of the second type of device signal (such as TETRA) after being processed by step S2 in the embodiments of the present application. Figure 8Time and frequency domain graph of the third type of device signal (such as ACMI) in the application after step S2 processing; Figure 9 Time and frequency domain graph of the fourth type of device signal (such as Global Hawk) in the application after step S2 processing; Figure 10 Comparison test graph of recognition accuracy of the double-flow cross-fusion network model and the baseline model in the application under different signal-to-noise ratios; Figure 11 Schematic diagram of the frequency spectrum situation analysis result interface generated after the system in the application performs the whole process of S1 to S4; Figure 12 Schematic diagram of the database table structure (E-R diagram) of the data management module in the application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0020] The application aims to provide a wideband signal automatic recognition method and system based on a double-flow cross-fusion network, to solve the problems in the prior art.

[0021] To make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the application will be further described in detail below with reference to the drawings and specific embodiments.

[0022] Embodiment 1 The embodiment of the application provides a wideband signal automatic recognition method based on a double-flow cross-fusion network, which realizes automatic detection and recognition of multiple devices in a wideband signal through adaptive spectrum segmentation and a deep learning network. The scenario processed in this embodiment is that a receiver (such as an SDR device connected to the back-end server shown in the figure) captures a complex wideband IQ signal data. The wideband signal mixes multiple signals, including four known signals (TETRA, ACMI, Global Hawk and LINK) of the system in the application. The purpose of this embodiment is to automatically and accurately segment all the four signals from the wideband signal and accurately recognize the four known signals. Figure 1

[0023] The overall process is as shown in Figure 2 ​As shown, first, the received wideband signal is adaptively spectrum- segmented, then the segmented sub-band signals are pre-processed, and then the features are extracted and fused by a dual-flow cross-fusion network for classification, and finally the recognition result is output.

[0024] In a specific implementation, the method first performs step S1: using a GWO-FCME adaptive spectrum positioning segmentation scheme. The system performs a fast Fourier transform (FFT) on the received wideband time-domain IQ data to obtain a full-resolution wideband power spectrum . To achieve adaptive segmentation, the system starts a grey wolf optimization (GWO) algorithm to automatically search for the optimal hyperparameter combination of the forward continuous mean excision (FCME) algorithm on the down-sampled version of the power spectrum , aiming to minimize the cost function . The cost function is defined as: ; Here, is the number of signals detected under the current hyperparameters, is preset to 4 (corresponding to four known signals) in this embodiment; is the average frequency band gap; is the weight of the difference in the number of signals, which is set to 10.0 in one preferred embodiment; is the weight of the gap, which is set to 1.0, is a constant to prevent zero minimum.

[0025] Using the current parameters , the preliminary frequency bands detected are merged to obtain final frequency bands; if , the gap value of all adjacent frequency bands is calculated ; and is set to a maximum value.

[0026] The GWO algorithm converges and outputs the optimal hyperparameters , and using this set of parameters, the final forward continuous mean excision (FCME) algorithm is performed on the full-resolution wideband power spectrum . In one preferred embodiment, the algorithm is performed on the sorted version of (total number of points ): Using the found by GWO optimization, the initial number of noise points is calculated; Iterative calculation: on the sorted , the mean value of the first points is taken as the noise mean value ; a new threshold is calculated according to ; until re-determine ; until converge or reach the maximum number of iterations, the final and

[0027] ; wherein, is the noise mean value after the convergence of the iterative calculation, is the best false alarm probability obtained by GWO optimization.

[0028] In each iteration of the GWO algorithm, the position of each wolf in the population is updated according to the following formula : Calculate the guide coefficient: ; Calculate the distance from the three wolves (Alpha, Beta, Delta): ; ; ; Calculate the next position guided by the three wolves: ; ; ; Update the final position: ; wherein and are both random coefficients in the iteration.

[0029] Finally, the system performs threshold detection on according to , and uses the optimal merging threshold obtained by GWO optimization to merge adjacent frequency bands, and finally successfully separates all four signal frequency bands (including TETRA, ACMI, GlobalHawk, LINK). The flowchart is shown in Figure 4 , and an example of threshold detection effect is shown in Figure 5 .

[0030] Step S2 performs preprocessing on each signal frequency band separated. For frequency band B i (with center frequency f c,i and bandwidth BW i ), the system performs bandpass filtering in the frequency domain to isolate the target signal, and reconstructs the time domain IQ sequence si (t). Subsequently, digital down conversion (DDC) is performed to shift the signal spectrum to baseband:

[0031] Step S3 inputs the pre-processed sub-band IQ data sequence into a dual-stream cross-fusion network (DSCFNet). The network structure is as shown in Figure 3 With regard to the network, the embodiment provides an optimization method for training a wideband signal recognition model, which aims to generate a dual-stream cross-fusion network model (as shown in Figure 3 ) with high robustness and high precision. The method comprises the following steps: Step TA1: training data preparation and enhancement Step TA2: dual-stream cross-fusion network (DSCFNet) construction Step TA3: model optimization and training execution Further, the step TA1 specifically comprises: Load a training data set containing multiple known signal types (for example: TETRA, ACMI, Global Hawk, LINK). To improve the robustness of the model, perform online data enhancement on the input IQ data sequence , preferably using the method of superimposing additive white Gaussian noise (AWGN) with different signal-to-noise ratios (SNRs): ; and perform normalization processing on the enhanced samples.

[0032] Further, the step TA2 aims to construct the dual-stream cross-fusion network (DSCFNet), which specifically comprises: TA2-1: input feature projection 9 inputs the sub-band IQ data sequence into a shared input projection module, uses Conv1d(kernel_size=1) to preliminarily fuse the original 2-channel data and project it to dimension, generating a shared input feature map : ; TA2-2: parallel multi-modal feature extraction inputs the shared feature map into the following two independent feature extraction branches in parallel: 1. Local feature branch (MSC-CBAM) This branch uses a multi-scale residual depth separable convolutional network (MultiScaleResidualDSCNN) to extract fine local features. To enhance the representativeness of the features, a convolution block attention module (CBAM) is embedded after each convolution block in the module. The CBAM module applies channel attention and spatial attention , the output feature map of this layer is dynamically refined: ; ; Finally, the feature maps of all scales are aggregated, and the local features are output : ; Temporal feature branch (Bi-GRU): This branch adopts a standard temporal encoder, whose core is a stacked bidirectional gated recurrent unit (Bi-GRU) network. After the feature map is normalized by layer, it is input into the Bi-GRU network, which processes the sequence in both forward and backward directions to calculate the hidden state of each time step and . Finally, the hidden states of the two directions are concatenated in the feature dimension ; to obtain the temporal features containing context information.

[0033] TA2-3: Bidirectional cross-attention feature fusion To realize the deep fusion of heterogeneous features extracted by the two parallel branches, a bidirectional cross-attention pooling module (CrossAttentionFusion) is used to fuse the features of the two branches: 1. Feature vectorization: First, adaptive average pooling is applied to the sequence feature maps and of the two branches in the time dimension to generate feature vectors representing the global summaries of each branch and .

[0034] 2. Bidirectional attention query: Perform two multi-head cross-attention (MultiheadAttention) operations, setting the internal feature dimension of the fusion module to .

[0035] 3. Direction a (MSC query GRU): Use as the query vector (Query), and use the complete sequence of as the key (Key) and value (Value): ; ; This step aims to filter and aggregate the most relevant information from the temporal features to the local summary.

[0036] b. Direction Two (GRU query MSC): Using As a query vector, used The complete sequence serves as the key and value: ; ; This step is intended to... In local features, filter and aggregate with The most relevant information in the timeline summary.

[0037] c. Final vector concatenation: Pay attention to the vectors in both directions and Concatenate along the feature dimension: ; TA2-4: Classification output 45 is the fully fused feature vector generated in S3-3. The input is fed into a multilayer perceptron (MLP) classifier. This classifier consists of several fully connected layers, ReLU activation functions, and Dropout layers stacked together. Finally, the predicted probabilities for each category are obtained through the Softmax function. .

[0038] ; Furthermore, step TA3 is model optimization and training execution, which specifically includes: TA3-1: Model Optimization To achieve optimal recognition performance for the model described in TA2, the total loss needs to be calculated. . Label smoothing cross entropy is preferred. Computation to improve generalization ability: ; Here, It is the total number of categories. It is a model for categories The predicted probability, It is a real tag index. It is a smoothing factor. It is the Kronecker delta function.

[0039] TA3-2: Training Execution: Using the Adam optimizer to optimize the total loss Backpropagation is performed, and a learning rate scheduling strategy is used to dynamically adjust the learning rate. After each training round, the model performance is evaluated on an independent validation set and the validation accuracy is calculated. If the model does not improve in consecutive validation rounds, an early stopping mechanism is triggered, training is terminated, and the best model (best_model.pth) is saved.

[0040] Step S4 achieves deep feature fusion through a bidirectional cross-attention feature fusion module. This step is performed by the classification and recognition module during inference and aims to process the parallel features (spatial features) output from step S3. and time series characteristics This process enables efficient feature fusion and classification decisions. It is consistent with the feature fusion and classification network structures (TA2-3 and TA2-4) described in the training phase (TA1-TA3).

[0041] Example 2: In terms of system implementation, the spectrum segmentation module integrates GWO and FCME algorithms to automatically complete broadband spectrum segmentation; the signal preprocessing module generates standardized IQ sequences through DDC and resampling units; the classification and recognition module loads a pre-trained dual-stream cross-fusion network model (best_model.pth) for feature extraction and fusion; and the data management and interaction module provides a user interface to receive input data and visualize results (such as...). Figure 11 (Interface example), recognition results are categorized by Figure 12 The database table structure shown is used for storage. During the training phase, the model employs label-smooth cross-entropy loss and the Adam optimizer, and saves the optimal model through an early stopping mechanism. Experiments show that this method achieves significantly higher recognition accuracy than traditional CNN models (e.g., compared to...) in low signal-to-noise ratio environments. Figure 10 As shown in the figure, it effectively improves the robustness and automation level of identification in complex electromagnetic environments.

[0042] To verify the parallel decoupling architecture described in S3 / S4 of this invention ( Figure 3 To demonstrate the superiority of this approach, this embodiment uses a dataset containing four signal types: TETRA, ACMI, Global Hawk, and LINK, to test the model trained in Example 1 (TA1-TA3) and compare it with a baseline model (such as a traditional CNN). See also... Figure 10 The figure shows a comparison of the recognition accuracy of the dual-branch attention network of the present invention and the baseline model (traditional CNN). Experimental results show that under all test signal-to-noise ratio (SNR) conditions, the dual-stream cross-fusion network of the present invention (such as...) Figure 10 The recognition accuracy of the proposed method is significantly higher than that of the traditional baseline model. This strongly demonstrates that the technical solution of the present invention, which extracts spatial and temporal features in parallel (S3) and performs efficient fusion (S4), has higher recognition accuracy and robustness.

[0043] Full-process system automatic verification: in order to verify the automation and effectiveness of the full process S1 to S4 of the present application, this embodiment inputs a wideband IQ file containing four known signals into the system shown in Figure 1 Referring to Figure 11 , the figure is the interface screenshot of the final result of the full-process analysis of the system of the present application. The system automatically completes the following tasks: S1 (detection): the GWO-FCME algorithm successfully detects and locates all four signal bands in the wideband spectrum (as shown in the right list, with center frequencies of 0.40, 1.10, 1.50, 2.10 MHz). Figure 11

[0044] S2 (preprocessing): the system automatically completes the reconstruction of the four signals (the effect is as shown in Figures 7 to 9 ).

[0045] S3 / S4 (recognition): Figure 3 The dual-branch network (which has been trained with four types of signals such as TETRA, ACMI, etc. in embodiment 1) successfully classifies the reconstructed signals and accurately identifies "LINK-K", "Global Hawk", "ACMI", "TETRA" (as shown in the list Figure 11 ).

[0046] Data management verification: after the analysis process shown in Figure 8 , the recognition result (successfully written into the database shown in Figure 9 ) verifies the effectiveness of the data management module of the present application, and realizes the complete closed loop from analysis to storage.

[0047] In summary, the wideband signal analysis method and system provided by the present application, which fuses adaptive blind detection and dual-branch attention network, realizes adaptive segmentation of S1 through GWO-FCME, realizes high-precision recognition of S3 / S4 through the parallel CNN+GRU dual-stream cross-fusion network, guarantees the training stability of S4 through label smoothing and early stopping mechanism, and finally realizes the automation and data management of the full process through B / S architecture and database, which has extremely high industrial application value.

[0048] The principles and implementation modes of the present application are described by applying specific examples, and the above description of the embodiments is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.​

Claims

1. A method for automatic recognition of wideband signals based on a dual-flow cross-fusion network, characterized in that, The method comprises the following steps: S1. An adaptive spectrum positioning segmentation scheme of GWO-FCME is adopted, the hyperparameters of the forward continuous mean cut algorithm are optimized by the grey wolf optimization algorithm, adaptive blind segmentation of the wideband power spectrum is realized, and multiple independent signal frequency bands are obtained; S2. Digital down-conversion and resampling preprocessing based on the bandwidth of the signal itself are performed on each segmented signal frequency band to obtain a normalized sub-band IQ data sequence; S3. The sub-band IQ data sequence is input into a double-flow cross-fusion network, local features of the signal are extracted through a local feature branch, and long-distance time sequence dependent features of the signal are extracted through a time sequence feature branch; S4. A bidirectional cross-attention feature fusion module is used to deeply fuse the local features and the long-distance time sequence dependent features, and a signal recognition result is output through a classifier.

2. The method of claim 1, wherein the method is a method of automatic recognition of a broadband signal based on a dual-flow cross-fusion network. Step S1 comprises: S11. Fast Fourier transform is performed on the received wideband time domain IQ data to obtain a full-resolution wideband power spectrum; S12. Using the Grey Wolf optimization algorithm, with the objective of minimizing the cost function on the downsampled power spectrum, automatically search for the optimal hyperparameter combination of the forward continuous mean cut algorithm. The cost function is defined as follows: ； wherein is the current parameter is the number of signals detected below is the preset target number is the average frequency band gap is the weight coefficient prevents the denominator from being zero S13. The optimal hyperparameters are used to perform the forward continuous mean cut algorithm on the full-resolution power spectrum, iteratively calculate the noise mean value, and determine the final detection threshold, the formula of the detection threshold being: ； wherein, is the noise mean value after the iteration calculation converges, is the best false alarm probability obtained by GWO optimization; S14. Threshold detection and frequency band merging are performed on the power spectrum according to the final detection threshold and the merging threshold, and multiple independent signal frequency bands are segmented. 3.The method of claim 1, wherein, Step S2 comprises: S21. Frequency domain band-pass filtering is performed on each signal frequency band, and an inverse fast Fourier transform is performed to reconstruct a time domain IQ sequence; S22. Digital down-conversion is performed on the reconstructed signal to shift the frequency spectrum to the baseband; S23. The target sampling rate is set according to the bandwidth of the signal itself, and the baseband signal is resampled, the formula of the target sampling rate being: ； wherein B is the bandwidth of the signal band itself; S24. A normalized, low-data-rate sub-band IQ data sequence is generated as the input of the double-flow cross-fusion network.

4. The method of claim 1, wherein the method is a method of automatic recognition of a broadband signal based on a dual-flow cross-fusion network. In step S3, the double-flow cross-fusion network comprises: A local feature branch adopts a multi-scale residual deep separable convolutional network and an embedded convolution block attention module to extract multi-scale local features, wherein the attention module embedded after each convolution block applies channel attention and spatial attention in turn to dynamically refine the feature map; A time sequence feature branch adopts a stacked bidirectional gated recurrent unit network to extract long-distance time sequence dependent features of the signal.

5. The method of claim 4, wherein the method further comprises: In the local feature branch, the refinement process of the convolution block attention module comprises: First, the channel attention map is calculated and multiplied with the input feature map element by element to obtain an intermediate feature: ； wherein, is a channel attention map, is an input feature map; Then, the spatial attention map is calculated and multiplied with the intermediate feature element by element to obtain a refined feature map: ； wherein, is the spatial attention map.

6. The method of claim 1, wherein the method further comprises: Step S4 comprises: S41. Adaptive average pooling is performed on the local feature sequence and the time sequence feature sequence respectively to generate a local feature vector and a time sequence feature vector; S42. Two multi-head cross-attention operations are performed: In the first operation, the local feature vector is taken as a query vector, the time sequence feature sequence is taken as a key and a value, relevant information is aggregated through a multi-head attention mechanism, and a first fusion vector is obtained; In the second operation, the time sequence feature vector is taken as a query vector, the local feature sequence is taken as a key and a value, relevant information is aggregated through a multi-head attention mechanism, and a second fusion vector is obtained; S43. Splice the first fusion vector and the second fusion vector in the feature dimension to obtain a deep fusion feature vector; S44. Input the deep fusion feature vector into a multilayer perception classifier, calculate the prediction probability of each category through a Softmax function, and output the final signal recognition result according to a pre-set confidence threshold.

7. A wideband signal automatic identification system based on a dual-flow cross-fusion network, characterized in that, Comprise: a spectrum segmentation module for adaptive blind segmentation of a wideband spectrum; a signal preprocessing module for digital down-conversion and resampling preprocessing of the segmented signal frequency band; a classification and identification module for extracting features and performing fusion classification through a dual-flow cross-fusion network; a data management and interaction module for providing a user interaction interface, receiving wideband signal data and parameters input by a user, calling the above modules to perform processing, and visualizing and storing the recognition result and time-frequency graph.

8. The dual-flow cross fusion network-based broadband signal automatic identification system according to claim 7, wherein, The spectrum segmentation module comprises: a power spectrum acquisition unit for performing fast Fourier transform on wideband time-domain IQ data to generate a power spectrum; a gray wolf optimization unit for iteratively optimizing the hyperparameters of a forward continuous mean cut algorithm on the down-sampled power spectrum; a forward continuous mean cut unit for calculating a detection threshold using the optimal hyperparameters and performing frequency band segmentation and merging.

9. The dual-flow cross fusion network-based broadband signal automatic identification system according to claim 7, wherein, The signal preprocessing module comprises: a digital down-conversion unit for shifting the signal spectrum to the baseband; a resampling unit for adaptively adjusting the sampling rate according to the signal bandwidth; a data normalization unit for generating normalized sub-band IQ data sequences.

10. The dual-flow cross fusion network-based broadband signal automatic identification system according to claim 7, wherein, The classification and identification module comprises: a local feature extraction unit for extracting local features through a multi-scale residual deep separable convolution network and a convolution block attention module; a time series feature extraction unit for extracting time series features through a bidirectional gated recurrent unit network; a feature fusion unit for realizing deep interaction of local and time series features through a bidirectional cross-attention feature fusion module; a classification output unit for outputting signal categories and confidence through a multilayer perception classifier.

Images