Multi-type unknown interference double cross attention fusion intelligent reasoning identification method and system

Abstract

The invention discloses a multi-type unknown interference double-cross-attention fusion intelligent reasoning identification method and system, and belongs to the technical field of electromagnetic spectrum and artificial intelligence cross, and the method comprises the steps: obtaining an electromagnetic interference signal; performing feature extraction, feature fusion and feature recognition on the electromagnetic interference signal by using a multi-dimensional feature fusion open set recognition model to obtain an interference recognition classification result; the multi-dimensional feature fusion open set recognition model comprises an interference frequency rule feature extraction module, an interference modulation feature extraction module and a double cross attention fusion module, and the interference frequency rule feature extraction module and the interference modulation feature extraction module extract interference frequency rule features and interference modulation features of electromagnetic interference signals respectively; and the double cross attention fusion module fuses the interference frequency rule characteristics and the interference modulation characteristics, and performs interference identification and classification to obtain an interference identification classification result. According to the method, more accurate interference open set identification can be realized, and the unknown interference identification capability in the electromagnetic spectrum environment is improved.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent reasoning and recognition of multiple types of unknown interference through dual cross-attention fusion, belonging to the field of electromagnetic spectrum and artificial intelligence interdisciplinary technology. Background Technology

[0002] Interference type inference and identification in electromagnetic environments is one of the fundamental technologies for electromagnetic situation control and electromagnetic spectrum management. With the rapid development of deep learning technology, deep learning-based interference identification methods have become a research hotspot in the field of wireless communication. Among these, techniques for identifying interference types by judging the modulation mode of the interference signal have been extensively studied. In addition, existing research also identifies interference types based on the frequency usage patterns of the interference. However, these studies still have certain limitations. For example, most identification methods rely on a single feature, which makes it difficult for the model to comprehensively represent the interference characteristics. The complementarity between features is not effectively utilized, thus limiting the identification accuracy. In fact, combining the modulation mode of the interference signal and the frequency usage patterns of the interference for multi-dimensional feature joint identification is one way to improve interference identification capabilities. Therefore, how to effectively integrate multi-dimensional features such as frequency domain and frequency usage patterns to construct a deep learning model with stronger representational capabilities has become the key to improving interference identification performance. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligent reasoning and identification of multiple types of unknown interference through dual-cross attention fusion. It adopts a dual-branch feature extraction module to process the feature information of the interference signal modulation mode and the interference frequency pattern, respectively. Through the dual-cross attention fusion module, it realizes the deep interaction and adaptive fusion of multi-dimensional features, thereby overcoming the limitation of single features in existing interference identification methods, achieving more accurate open-set interference identification, and improving the ability to identify unknown interference in the electromagnetic spectrum environment.

[0004] To achieve the above objectives / to solve the above technical problems, the present invention is implemented using the following technical solution:

[0005] In a first aspect, the present invention provides a multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method, comprising:

[0006] Acquire electromagnetic interference signals;

[0007] A pre-constructed multi-dimensional feature fusion open set recognition model is used to extract, fuse, and identify features of electromagnetic interference signals, resulting in interference identification and classification results.

[0008] In conjunction with the first aspect, the multi-dimensional feature fusion open set recognition model further includes an interference frequency pattern feature extraction module, an interference modulation feature extraction module, and a dual-cross attention fusion module. The interference frequency pattern feature extraction module is used to extract the interference frequency pattern features of the electromagnetic interference signal. The interference modulation feature extraction module is used to extract the interference modulation features of the electromagnetic interference signal. The dual-cross attention fusion module is used to fuse the interference frequency pattern features and interference modulation features of the electromagnetic interference signal to obtain a fused feature representation. Interference recognition and classification are then performed based on the fused feature representation to obtain the interference recognition and classification results.

[0009] In conjunction with the first aspect, the dual cross-attention fusion module further includes two cross-attention fusion methods: cross-attention fusion based on frequency pattern query and cross-attention fusion based on modulation feature query.

[0010] In conjunction with the first aspect, further, the cross-attention fusion based on usage frequency pattern queries uses interference modulation features as queries and usage frequency pattern features as keys and values, calculating the correlation between the two to generate attention weights, including:

[0011] Interference modulation characteristics Perform a linear transformation to generate a query vector. ,in, This is a trainable weight matrix;

[0012] Characteristics of usage frequency patterns Perform linear transformations to generate key vectors respectively. Sum value vector ,in, and These are the trainable weight matrices corresponding to the key vector and the value vector, respectively;

[0013] Calculate query vector and key vector The inner product, combined with the first scaling factor After normalization, the first attention score is obtained. :

[0014] ;

[0015] in, Represents the query vector Key vector Sum value vector Common dimensions;

[0016] First attention score Perform Softmax normalization to generate the first attention weight matrix. :

[0017] ;

[0018] Using the first attention weight matrix value vector By performing a weighted summation, the first cross-fusion feature is obtained. :

[0019] ;

[0020] The first cross-fusion feature With interference modulation characteristics The features are concatenated to obtain the first fused feature vector. :

[0021] ;

[0022] in, This indicates a feature splicing operation.

[0023] In conjunction with the first aspect, further, the cross-attention fusion based on modulation feature query will use frequency pattern features as queries and interference modulation features as keys and values, and calculate the correlation between the two to generate attention weights, including:

[0024] Characteristics of usage frequency patterns Perform a linear transformation to generate a query vector. ,in, This is a trainable weight matrix;

[0025] Interference modulation characteristics Perform linear transformations to generate key vectors respectively. Sum value vector ,in, and These are the trainable weight matrices corresponding to the key vector and the value vector, respectively;

[0026] Calculate query vector With key vector The inner product, combined with the second scaling factor Normalization is performed to obtain the second attention score. :

[0027] ;

[0028] in, Represents the query vector Key vector Sum value vector Common dimensions;

[0029] Second attention score Perform Softmax normalization to generate the second attention weight matrix. :

[0030] ;

[0031] Using the second attention weight matrix value vector By performing a weighted summation, the second cross-fusion feature is obtained. :

[0032] ;

[0033] The second cross-fusion feature Frequency patterns and characteristics The two vectors are concatenated to obtain the second fused feature vector. :

[0034] ;

[0035] in, This indicates a feature splicing operation.

[0036] In conjunction with the first aspect, further, the first fused feature vector output by cross-attention fusion based on frequency-based query patterns is obtained through a fusion layer. The second fused feature vector is the output of cross-attention fusion based on modulation feature query. The components are fused together to obtain the final fusion feature F. FC The final fused feature F is processed using a classifier. FC The interference type is obtained, and the interference identification and classification results are output.

[0037] In conjunction with the first aspect, the interference frequency pattern feature extraction module further includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, an average pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence, with a ReLU activation function set after each convolutional layer;

[0038] In the interference frequency pattern feature extraction module, multiple convolutional layers are used to model the local context of frequency pattern data in electromagnetic interference signals and extract local features from the input data; multiple fully connected layers are used to integrate local features into global feature representation and further capture high-level feature interactions through nonlinear transformation, and finally output interference frequency pattern features.

[0039] In conjunction with the first aspect, the interference modulation feature extraction module further includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence, with a ReLU activation function set after each convolutional layer;

[0040] In the interference modulation feature extraction module, multiple convolutional layers are used to extract local features of signal modulation data in electromagnetic interference signals; multiple fully connected layers are used to represent local features and global features, and further capture high-level feature interactions through nonlinear transformations, ultimately outputting interference modulation features.

[0041] Secondly, the present invention provides a multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition system, comprising:

[0042] The signal acquisition module is used to acquire electromagnetic interference signals;

[0043] The interference identification module is used to extract, fuse, and identify features of electromagnetic interference signals using a pre-built multi-dimensional feature fusion open set identification model, and obtain interference identification and classification results.

[0044] In conjunction with the second aspect, the multi-dimensional feature fusion open set recognition model further includes an interference frequency pattern feature extraction module, an interference modulation feature extraction module, and a dual-cross attention fusion module. The interference frequency pattern feature extraction module is used to extract the interference frequency pattern features of the electromagnetic interference signal. The interference modulation feature extraction module is used to extract the interference modulation features of the electromagnetic interference signal. The dual-cross attention fusion module is used to fuse the interference frequency pattern features and interference modulation features of the electromagnetic interference signal to obtain a fused feature representation. Interference recognition and classification are then performed based on the fused feature representation to obtain the interference recognition and classification results.

[0045] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0046] This invention proposes a dual-cross attention fusion intelligent reasoning and identification method and system for multi-type unknown interference. It simultaneously extracts interference frequency pattern features and interference modulation features through a dual-branch structure of an interference frequency pattern feature extraction module and an interference modulation feature extraction module. The dual-cross attention fusion module achieves deep interaction and adaptive fusion of multi-dimensional features, effectively combining the characteristics of interference signal modulation features and interference frequency pattern features for interference identification. This overcomes the shortcomings of single-feature methods and significantly improves the performance of interference signal identification. Under different interference-to-noise ratios, this invention effectively improves the accuracy and robustness of interference signal identification compared to single-feature models. Especially when processing interference signals with similar features, the dual-cross attention fusion strategy exhibits superior performance. This invention provides reliable technical support for solving spectrum situation analysis and interference control in complex electromagnetic environments. This invention can also be extended to fields requiring electromagnetic spectrum analysis support, such as spectrum planning and cyber warfare. Attached Figure Description

[0047] Figure 1 The diagram shown is a step-by-step illustration of a multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method provided by an embodiment of the present invention;

[0048] Figure 2 The diagram shown is a structural schematic of the multi-dimensional feature fusion open set recognition model in an embodiment of the present invention.

[0049] Figure 3 The diagram shown is a schematic representation of the network structure of the interference frequency pattern feature extraction module in an embodiment of the present invention.

[0050] Figure 4 The diagram shown is a schematic representation of the network structure of the interference modulation feature extraction module in an embodiment of the present invention.

[0051] Figure 5 The diagram shown is a schematic diagram of the network structure of the dual-cross attention fusion module in an embodiment of the present invention;

[0052] Figure 6 The diagram shown is a schematic representation of some interference signal characteristics under -2 dB conditions in an embodiment of the present invention.

[0053] Figure 7 The diagram shown illustrates the recognition accuracy of the interference frequency pattern feature extraction module on various interference signals in this embodiment of the invention.

[0054] Figure 8 The figure shown is a schematic diagram illustrating the recognition accuracy of the interference modulation feature extraction module on various interference signals in an embodiment of the present invention.

[0055] Figure 9 The figure shown is a schematic diagram illustrating the recognition accuracy of various interference signals based on the feature splicing and fusion strategy in an embodiment of the present invention.

[0056] Figure 10 The figure shown is a schematic diagram illustrating the accuracy of the method of the present invention in identifying various interference signals in an embodiment of the present invention.

[0057] Figure 11 The figure shows the F1-score of different models under different drying ratios in the embodiments of the present invention.

[0058] Figure 12 The diagram shown is a schematic diagram of the confusion matrix of the method of the present invention at -7 dB in an embodiment of the present invention;

[0059] Figure 13 The diagram shown is a schematic diagram of the confusion matrix of the method of the present invention at -2 dB in an embodiment of the present invention;

[0060] Figure 14 The figure shown is a schematic diagram of a multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition system provided in an embodiment of the present invention. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0062] Example 1

[0063] This embodiment introduces a multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method, such as... Figure 1 As shown, the specific steps include the following:

[0064] Step A: Acquire the electromagnetic interference signal and divide it into interference frequency pattern data and interference modulation data, which are used as inputs to the JPRFE and JSMFE modules, respectively.

[0065] Step B: Use a pre-constructed multi-dimensional feature fusion open set recognition model to extract, fuse, and recognize features of electromagnetic interference signals to obtain interference recognition and classification results.

[0066] To achieve effective extraction and fusion of multidimensional interference information, this invention proposes a novel multidimensional feature fusion open set recognition model. This model takes the interference frequency pattern and interference modulation mode as two inputs, processes feature information of different dimensions through a dual-branch feature extraction module, and achieves effective fusion of multidimensional features by combining a dual-cross attention fusion module.

[0067] like Figure 2The multidimensional feature fusion open set recognition model mainly includes a Jamming Frequency Pattern Feature Extraction (JPRFE) module, a Jamming Signal Modulation Feature Extraction (JSMFE) module, and a Dual Cross-Attention Mechanism (DCAM) module. Both the Jamming Frequency Pattern Feature Extraction (JPRFE) and Jamming Signal Modulation Feature Extraction (JSMFE) modules consist of multiple convolutional blocks. The JPRFE module extracts the Jamming Frequency Pattern features of the electromagnetic interference signal, while the JSMFE module extracts the Jamming Signal Modulation features. The Dual Cross-Attention Mechanism (DCAM) module fuses the Jamming Frequency Pattern features and Jamming Signal Modulation features of the electromagnetic interference signal, further learns and integrates them to obtain a fused feature representation, and performs interference recognition and classification based on this fused feature representation to obtain the interference recognition and classification results.

[0068] In this embodiment of the invention, the network structure of the interference frequency pattern feature extraction module is as follows: Figure 3 As shown, the module includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, an average pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence. A ReLU activation function is set after each convolutional layer.

[0069] In the interference frequency pattern feature extraction module, multiple convolutional layers are used to model the local context and extract local features from the input data; the ReLU activation function enhances the feature learning ability by introducing nonlinear transformations; and fully connected layers integrate local features into the global feature representation and further capture high-level feature interactions through nonlinear transformations. Therefore, through multi-layer convolutional operations, the JPRFE module can progressively extract high-level frequency domain features from the input frequency pattern data and output feature representations with rich semantic information.

[0070] Given a As input to the JPRFE module, This represents the frequency pattern data of interference. Let the size of the original input image be... ,in, Indicates the number of channels. , These represent the height and width of the image, respectively. Input image After entering the JPRFE module, the local features of the input image are gradually extracted through multiple convolutional layers, and the output feature maps of each convolutional layer are used for further processing.

[0071] The JPRFE module was built to... The convolutional layer mapped to the embedded features is represented as:

[0072] (1)

[0073] in, This represents the output of the i-th layer in the JPRFE module. It is the size of the space. and output channels The output; This represents the convolution or pooling operation in the i-th layer of the JPRFE module; This represents the output of the i-th layer in the JPRFE module. In this embodiment of the invention, i = 1, 2, 3, 4, 5. When i = 1, 2, 3, 4, it corresponds to the 1st to 4th convolutional layers in the JPRFE module; when i = 5, it corresponds to the average pooling layer in the JPRFE module; when i = 1, ... Original input image .

[0074] The input to each convolutional layer is derived from the output of the preceding convolutional layers. Batch normalization, ReLU activation, and pooling operations are added after each convolutional layer to further enhance the model's nonlinear transformation capabilities and feature representation. Convolutional layers significantly reduce the dimensionality of the original input that may obscure frequency patterns.

[0075] In the JPRFE module, the fully connected layer maps local features to the global feature space through a series of nonlinear transformations, ultimately outputting a high-level semantic feature representation. The operation of the fully connected layer can be represented as follows:

[0076] (2)

[0077] in, This represents the operation of the j-th fully connected layer. This represents the output of the j-th fully connected layer. The feature dimension is The output, This represents the input of the j-th fully connected layer. In this embodiment of the invention, j = 1, 2, 3.

[0078] according to Figure 3 In this embodiment of the invention, the output of the JPRFE module can be expressed as:

[0079] (3)

[0080] (4)

[0081] in, This represents the output of the average pooling layer in the JPRFE module. This represents the output of the third fully connected layer of the JPRFE module, i.e., the output of the JPRFE module.

[0082] In this embodiment of the invention, the network structure of the interference modulation feature extraction module is as follows: Figure 4 As shown, the module includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence. A ReLU activation function is set after each convolutional layer.

[0083] Similar to the JPRFE module, the JSMFE module uses convolutional layers and fully connected layers to extract local and high-dimensional features of the signal modulation data, respectively. The convolutional layers use orthogonal IQ signals... For input, where, The length of the modulation feature data of the input signal. The number of channels is 2 for the IQ signal.

[0084] In the JSMFE module, the process of extracting convolutional layers can be represented by the following formula:

[0085] (5)

[0086] in, This represents the output of the i-th convolutional layer. It is a length of and output channels are The output; Indicates the first Convolution operations of each convolutional layer This represents the output of the (i-1)th convolutional layer. When i=1, For the original input .

[0087] Similar to the frequency pattern feature extraction process, the output of the JSMFE module can be expressed as:

[0088] (6)

[0089] (7)

[0090] in, This represents the output of the third convolutional layer of the JSMFE module. This represents the output of the third fully connected layer of the JSMFE module, i.e., the output of the JSMFE module.

[0091] After acquiring the features of interference frequency patterns and interference modulation patterns, the two sets of features need to interact to achieve feature fusion, thereby completing the identification of interference signals. The cross-attention fusion strategy utilizes an attention mechanism to achieve feature fusion, fully leveraging the interactive information between interference frequency patterns and modulation features to achieve deep feature fusion and enhancement.

[0092] The dual cross-attention fusion module proposed in this invention dynamically calculates and fuses the feature correlations of interference signals in two dimensions: modulation mode and frequency usage pattern by introducing a cross-attention mechanism. The DCAM module can enhance the distinguishability of these similar features through a query-guided mechanism, thereby improving the recognition accuracy of interference signals.

[0093] The network structure of the dual-cross attention fusion module is as follows: Figure 5 As shown, the model consists of two cross-attention fusion methods: cross-attention fusion based on frequency usage pattern query (CAM-M) and cross-attention fusion based on modulation feature query (CAM-P). The two cross-attention fusion modules work together to enable the model to model the frequency usage pattern and modulation characteristics of interference signals at a deeper level, capture subtle differences between different feature dimensions, and achieve more accurate interference signal identification.

[0094] For CAM-P, the features of interference modulation are used as the query (Q), and the features of frequency patterns are used as the key (K) and value (V). The correlation between the two is calculated to generate attention weights. The specific process is as follows:

[0095] First, a linear transformation is performed on the interference modulation features to generate a query vector. ,in, For a trainable weight matrix, The interference modulation characteristics are used; simultaneously, a linear transformation is performed on the frequency usage pattern characteristics to generate key vectors. Sum value vector ,in, and These are the trainable weight matrices corresponding to the key vector and value vector, respectively. This represents the frequency usage pattern.

[0096] Secondly, through calculation and The inner product, combined with the first scaling factor After normalization, the first attention score is obtained. , can be represented as:

[0097] (8)

[0098] in, Represents the query vector Key vector Sum value vector The common dimensions.

[0099] Furthermore, in order to more accurately characterize the interference modulation features Characteristics of usage frequency patterns The degree of attention is considered, and the first attention score is normalized using Softmax to generate the first attention weight matrix. The formula is as follows:

[0100] (9)

[0101] Next, the value vectors are weighted and summed using the first attention weight matrix to obtain the first cross-fusion feature. The formula is as follows:

[0102] (10)

[0103] Finally, the first cross-fusion feature is concatenated with the interference modulation feature to form a new first fusion feature vector. , means as follows:

[0104] (11)

[0105] in, This indicates a feature splicing operation.

[0106] For CAM-M, frequency-based features are used as queries, and interference modulation features are used as keys and values. The correlation between the two is calculated to generate attention weights. The specific process is as follows:

[0107] First, a linear transformation is performed on the frequency pattern characteristics to generate a query vector. Simultaneously, a linear transformation is performed on the interference modulation features to generate key vectors. Sum value vector .

[0108] Secondly, through calculation and The inner product, combined with the second scaling factor Normalization is performed to obtain the second attention score. , can be represented as:

[0109] (12)

[0110] in, Represents the query vector Key vector Sum value vector The common dimensions.

[0111] Furthermore, in order to more accurately characterize right The degree of attention is considered, and the second attention score is normalized using Softmax to generate the second attention weight matrix. :

[0112] (13)

[0113] Next, the value vectors are weighted and summed using the second attention weight matrix to obtain the second cross-fusion feature. :

[0114] (14)

[0115] Finally, the second cross-fusion feature is concatenated with the frequency pattern feature to form a new second fusion feature vector. , means as follows:

[0116] (15)

[0117] The first fused feature vector output by CAM-P is processed through the fusion layer. and the second fused feature vector output by CAM-M The components are fused together to obtain the final fusion feature F. FC .

[0118] The final fused feature F is processed using a classifier. FC The interference type is obtained, and the interference identification and classification results are output.

[0119] In this embodiment of the invention, the training process of the multi-dimensional feature fusion open set recognition model is as follows:

[0120] Step 1: Obtain training samples and test samples, and set the number of training iterations. and the number of batches in each loop iteration .

[0121] Step 2: Input the training samples into the multi-dimensional feature fusion open set recognition model, obtain the frequency pattern features and interference modulation features through the JPRFE branch and JSMFE branch, and then perform feature fusion through the dual cross-attention fusion module to obtain the predicted interference type.

[0122] Step 3: Compare the predicted interference type with the actual interference type and calculate the model loss.

[0123] Step 4: Based on the model loss value, specifically, update the model parameters end-to-end using cross-entropy, center loss, and reconstruction loss.

[0124] Step 5: Repeat steps 2-4 until the maximum number of iterations is met, or the model loss is less than the preset threshold, to obtain the trained multi-dimensional feature fusion open set recognition model.

[0125] In this embodiment of the invention, an interference signal dataset is constructed for model training and interference identification experiments. The interference signal dataset mainly includes two layers of features: Jamming Frequency Pattern (JPR) and Jamming Signal Modulation (JSM, also known as interference modulation features). The JPR data includes 10 types, such as single-sweep interference, dual-sweep interference, nonlinear sweep interference, and combined interference; the JSM data includes 10 types, such as 4FSK, 16QAM, CPFSK, PAM4, and QPSK. Under an interference-to-noise ratio (INR) of -2 dB, the time-frequency plots and spectrum diagrams of some interference types are shown below. Figure 6 As shown, the time-frequency plot of the frequency characteristics depicts the pattern of frequency variation of the interference signal over time, while the spectrogram of the interference modulation characteristics describes the microscopic features of the signal spectrum. Together, they constitute a complete feature description of the interference signal. By fusing these two features, the network can learn different feature information, thereby providing more reliable recognition results.

[0126] To verify the effectiveness of the method of this invention, a series of simulation experiments were conducted. The types of interference signals involved in the simulation experiments are shown in Table 1, totaling 10 types. To comprehensively evaluate the model's performance under different interference conditions, for each interference type, 200 samples were generated in 1 dB steps within the interference-to-noise ratio range of -7 to -2 dB. These samples were divided into known and unknown classes. 80% of the samples in the known class constituted the training set, and the remaining 20% ​​of the known class and 20% of the samples in the unknown class constituted the test set, used to evaluate the model's classification performance. All network models were implemented in Python, and the Adam optimizer was used during training with a learning rate of 0.001, 300 iterations, and 128 sets of data per training iteration.

[0127] Table 1

[0128]

[0129] To better illustrate the effectiveness of the multi-dimensional feature fusion open set recognition model proposed in this invention in open set scenarios, the above 10 interference signals are divided into known and unknown categories, of which 9 interferences are known categories and 1 interference is an unknown category, i.e., openness=2.67%.

[0130] The performance analysis of the multidimensional feature fusion open set recognition model will be introduced from two aspects: branch network and double cross fusion network. First, the recognition performance of the two branch networks (i.e., JPRFE module and JSMFE module) will be analyzed.

[0131] Figure 7 and Figure 8 The accuracy rates of the JPRFE and JSMFE modules for identifying various interference signals are shown, where J-Un represents the label of unknown interference signals. As can be seen from the figures, under -2 dB conditions, the JPRFE module performs well overall in identifying the nine known interference signals, all reaching 80% or higher. However, the accuracy rates for J1 and J2 are relatively low, mainly because their frequency bandwidths are similar, resulting in insufficient feature discrimination. Under -2 dB conditions, the JSMFE module's average identification rate for the nine known interference signals is only 71%. The identification accuracy rates for interference signals J4 and J5, J7 and J8, and J9 and J10 are particularly low because their frequency domain characteristics are similar, leading to a lack of discriminative power in the extracted features. This indicates that single feature extraction methods have limitations when processing interference signals with high feature similarity, making effective differentiation difficult.

[0132] The following is a simulation analysis of the dual-cross attention fusion module. To further verify the performance of dual-cross attention fusion, this invention uses a joint multi-domain residual network model based on Feature Concatenation Fusion (FCF) as a comparison model. FCF can fuse features from different modalities by concatenation.

[0133] Different fusion strategies may have varying impacts on model performance. To better verify the effectiveness of the proposed fusion strategy, this invention designed an ablation experiment to compare the impact of different fusion methods on model performance under a jitter-to-noise ratio (JNR) of -6 dB. The results are shown in Table 2, where CAM-P represents cross-attention fusion using only modulation feature queries, and CAM-M represents cross-attention fusion using only frequency-based queries. Comparing the data in Table 2, it can be seen that compared with FCF, CAM-P, and CAM-M, the F1-score of the proposed method is improved by 3%, 3%, and 6%, respectively, and the AUROC reaches 0.84. This indicates that the proposed method can achieve better feature fusion results, better balance the model's performance on known and unknown class samples, and enhance the model's robustness.

[0134] Definition of F1 score: Used to measure the performance of a deep network recognition model. Defined as ,in, express It is used to measure the proportion of true positives among samples predicted as positive by the model. ; Recall rate is used to measure the proportion of samples that are actually positive but are correctly predicted by the model. , This represents the number of true samples. This represents the number of false positives. This represents the number of false negatives.

[0135] Table 2

[0136]

[0137] Figure 9 and Figure 10 The accuracy rates of FCF and the proposed method for identifying various interference signals are shown in the figures. As can be seen from the figures, the proposed method achieves a higher accuracy rate than the FCF fusion algorithm compared to the single-branch identification method. Under -2dB conditions, the proposed method achieves an accuracy rate of 95% or higher for all nine known interference signals. For interference signals J1 and J2, whose characteristics include similar frequency usage patterns and different signal modulation characteristics, the average accuracy rates under different JNRs improved by 8.3% and 11.7% respectively compared to FCF after feature fusion. For interference signals J7 and J8, whose characteristics include different frequency usage patterns and similar signal modulation characteristics, the average accuracy rates under different JNRs improved by 5% and 4.5% respectively compared to FCF after feature fusion. The analysis process for other interference signals such as J4, J5, J9, and J10 is similar to that for the aforementioned interference signals. Simulation results show that the dual-cross attention fusion strategy proposed in this invention can effectively integrate the features of signal modulation mode and frequency usage pattern. When dealing with interference signal identification tasks with the same modulation mode or similar frequency usage pattern, it significantly improves the identification accuracy. At the same time, the method of this invention shows good performance under different JNR conditions, overcoming the limitations of single feature methods when dealing with interference signals with similar features.

[0138] The simulation experiments of this invention also analyzed the F1-score of FCF, JPRFE, JSMFE and the method of this invention under different JNRs, and the results are as follows: Figure 11As shown, under different JNRs, the F1-score of the JPRFE branch network is higher than that of the JSMFE branch network. This is because the frequency pattern features extracted by the JPRFE branch network contain richer feature information and perform more stably in the recognition task. It is worth noting that the F1-score of the FCF is close to that of the JPRFE branch network, and even slightly lower at higher JNRs. This is because the frequency pattern features extracted by the JPRFE branch network are already rich enough at high JNRs to achieve high recognition performance independently, while the FCF introduces additional complexity during feature fusion, resulting in poorer performance. In contrast, the F1-score of the method in this invention is higher than the other three network models, with an average F1-score of 90.3%, which is 1.8% and 4.8% higher than the JPRFE and JSMFE branch networks, respectively. This indicates that it can more effectively integrate the feature information of the two branch networks, thereby achieving a better feature fusion effect and significantly improving network performance.

[0139] Finally, to more clearly demonstrate the performance of the method of the present invention under different noise conditions, Figure 12 and Figure 13 The confusion matrices of the method of the present invention are shown under the conditions of interference-to-noise ratio (JNR) of -7 dB and -2 dB, respectively. It can be seen that when the JNR is -7 dB, some interference signals are confused to a certain extent, especially the identification results between some interference signals with similar characteristics are not completely accurate. However, when the JNR is -2 dB, all nine known interference signals and one unknown interference signal can be correctly identified, which fully verifies the effectiveness of the method of the present invention in the task of interference signal identification.

[0140] Example 2

[0141] Based on the same inventive concept as Embodiment 1, this embodiment introduces a multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition system, such as... Figure 14 As shown, it mainly includes a signal acquisition module and an interference identification module.

[0142] The signal acquisition module is used to acquire electromagnetic interference signals; the interference identification module is used to extract, fuse and identify features of electromagnetic interference signals using a pre-built multi-dimensional feature fusion open set identification model, and obtain interference identification and classification results.

[0143] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0144] In summary, this invention addresses the problem that existing interference signal identification methods often struggle to effectively distinguish interference signals with similar characteristics using a single feature. It proposes an open-set interference identification framework based on dual-cross-attention fusion. This framework includes a interference frequency pattern feature extraction module, an interference modulation feature extraction module, and a dual-cross-attention fusion module. By combining the two-branch feature extraction and fusion modules, it effectively integrates interference signal modulation features and interference frequency pattern features, overcoming the shortcomings of single-feature methods and significantly improving the performance of interference signal identification. Under different interference-to-noise ratios, this invention effectively improves the accuracy and robustness of interference signal identification compared to single-feature models. Particularly when processing interference signals with similar characteristics, the dual-cross-attention fusion strategy exhibits superior performance. This invention provides reliable technical support for solving spectrum situation analysis and interference control in complex electromagnetic environments. Furthermore, this invention can be extended to fields requiring electromagnetic spectrum analysis support, such as spectrum planning and cyber warfare.

[0145] This invention places particular emphasis on simultaneously extracting feature information from both the modulation scheme and frequency usage pattern of the interference signal, and achieves deep interaction and adaptive fusion of multi-dimensional features through a dual-cross-attention fusion module. This invention is the first to propose a dual-cross-attention fusion strategy, introducing a cross-attention mechanism to dynamically calculate and fuse the feature correlations of interference in both modulation scheme and frequency usage pattern dimensions, thereby more fully exploring the complementarity of multi-dimensional features.

[0146] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0147] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0148] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0149] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0150] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for intelligent reasoning and recognition of multiple types of unknown interference through dual-cross-attention fusion, characterized in that, include: Acquire electromagnetic interference signals; A pre-constructed multi-dimensional feature fusion open set recognition model is used to extract, fuse, and identify features of electromagnetic interference signals, resulting in interference identification and classification results.

2. The multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method according to claim 1, characterized in that, The multidimensional feature fusion open set recognition model includes an interference frequency pattern feature extraction module, an interference modulation feature extraction module, and a dual-cross attention fusion module. The interference frequency pattern feature extraction module is used to extract the interference frequency pattern features of the electromagnetic interference signal. The interference modulation feature extraction module is used to extract the interference modulation features of the electromagnetic interference signal. The dual-cross attention fusion module is used to fuse the interference frequency pattern features and interference modulation features of the electromagnetic interference signal to obtain a fused feature representation. Interference recognition and classification are then performed based on the fused feature representation to obtain the interference recognition and classification results.

3. The multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method according to claim 2, characterized in that, The dual cross-attention fusion module includes two cross-attention fusion methods: cross-attention fusion based on frequency pattern query and cross-attention fusion based on modulation feature query.

4. The multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method according to claim 3, characterized in that, The cross-attention fusion based on usage frequency pattern queries uses interference modulation features as queries and usage frequency pattern features as keys and values, calculating the correlation between the two to generate attention weights, including: Interference modulation characteristics Perform a linear transformation to generate a query vector. ,in, is a trainable weight matrix; Characteristics of usage frequency patterns Perform linear transformations to generate key vectors respectively. Sum value vector ,in, and These are the trainable weight matrices corresponding to the key vector and the value vector, respectively; Calculate query vector and key vector The inner product, combined with the first scaling factor After normalization, the first attention score is obtained. : ； in, Represents the query vector Key vector Sum value vector Common dimensions; First attention score Perform Softmax normalization to generate the first attention weight matrix. : ； Using the first attention weight matrix value vector By performing a weighted summation, the first cross-fusion feature is obtained. : ； The first cross-fusion feature With interference modulation characteristics The features are concatenated to obtain the first fused feature vector. : ； in, This indicates a feature splicing operation.

5. The multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method according to claim 3, characterized in that, The cross-attention fusion based on modulation feature queries uses frequency regularity features as queries and interference modulation features as keys and values, calculating the correlation between them to generate attention weights, including: Characteristics of usage frequency patterns Perform a linear transformation to generate a query vector. ,in, This is a trainable weight matrix; Interference modulation characteristics Perform linear transformations to generate key vectors respectively. Sum value vector ,in, and These are the trainable weight matrices corresponding to the key vector and the value vector, respectively; Calculate query vector With key vector The inner product, combined with the second scaling factor Normalization is performed to obtain the second attention score. : ； in, Represents the query vector Key vector Sum value vector Common dimensions; Second attention score Perform Softmax normalization to generate the second attention weight matrix. : ； Using the second attention weight matrix value vector By performing a weighted summation, the second cross-fusion feature is obtained. : ； The second cross-fusion feature Frequency patterns and characteristics The two vectors are concatenated to obtain the second fused feature vector. : ； in, This indicates a feature splicing operation.

6. The multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method according to any one of claims 3 or 4, characterized in that, The first fused feature vector is output by the cross-attention fusion based on frequency-based queries through the fusion layer. The second fused feature vector is the output of cross-attention fusion based on modulation feature query. The components are fused together to obtain the final fusion feature F. FC The final fused feature F is processed using a classifier. FC The interference type is obtained, and the interference identification and classification results are output.

7. The multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method according to claim 2, characterized in that, The interference frequency pattern feature extraction module includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, an average pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence, with a ReLU activation function set after each convolutional layer; In the interference frequency pattern feature extraction module, multiple convolutional layers are used to model the local context of frequency pattern data in electromagnetic interference signals and extract local features from the input data; multiple fully connected layers are used to integrate local features into global feature representation and further capture high-level feature interactions through nonlinear transformation, and finally output interference frequency pattern features.

8. The multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition method according to claim 2, characterized in that, The interference modulation feature extraction module includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence, with a ReLU activation function set after each convolutional layer; In the interference modulation feature extraction module, multiple convolutional layers are used to extract local features of signal modulation data in electromagnetic interference signals; multiple fully connected layers are used to represent local features and global features, and further capture high-level feature interactions through nonlinear transformations, ultimately outputting interference modulation features.

9. A multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition system, characterized in that, include: The signal acquisition module is used to acquire electromagnetic interference signals; The interference identification module is used to extract, fuse, and identify features of electromagnetic interference signals using a pre-built multi-dimensional feature fusion open set identification model, and obtain interference identification and classification results.

10. The multi-type unknown interference dual-cross attention fusion intelligent reasoning and recognition system according to claim 9, characterized in that, The multidimensional feature fusion open set recognition model includes an interference frequency pattern feature extraction module, an interference modulation feature extraction module, and a dual-cross attention fusion module. The interference frequency pattern feature extraction module is used to extract the interference frequency pattern features of the electromagnetic interference signal. The interference modulation feature extraction module is used to extract the interference modulation features of the electromagnetic interference signal. The dual-cross attention fusion module is used to fuse the interference frequency pattern features and interference modulation features of the electromagnetic interference signal to obtain a fused feature representation. Interference recognition and classification are then performed based on the fused feature representation to obtain the interference recognition and classification results.

Images