Communication interference signal detection method, system, equipment and medium

Abstract

The invention belongs to the technical field of communication, and discloses a communication interference signal detection method, system and device and a medium, and the method comprises the steps: processing a communication interference source signal through employing adaptive variational mode decomposition, a Grubrum angle field and a cyclic spectrum, and obtaining the intelligent characterization of a time domain, a graph domain and a frequency domain; constructing a depth detection network comprising three branches, and respectively extracting and fusing deep features of each type of representation; constructing detection statistics based on the fusion features, and generating a detection threshold under a corresponding false alarm probability; and inputting the to-be-detected signal into the network, comparing the to-be-detected signal with a threshold, and outputting a detection result. According to the method, multi-modal information can be effectively extracted, the detection robustness and precision are enhanced, and the method has good performance and universality.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a communication interference signal detection method, system, device and medium. Background Art

[0002] Wireless communications continue to play an important role in fields such as satellite internet, the Internet of Things, smart transportation, and military communications. However, the highly open nature of wireless channels makes wireless communications extremely vulnerable to intentional interference, such as electronic warfare attacks and malicious spectrum occupation, which can lead to a serious deterioration in communication quality and even blockage of communication links. There is an urgent need to improve the interference awareness and interference suppression capabilities of wireless communication systems to ensure the efficient and secure transmission of information. As the electromagnetic environment continues to become increasingly complex, the types and number of interference sources continue to increase. Understanding the distribution and activity patterns of interference sources to support the formulation of subsequent targeted countermeasures has gradually become a research focus in the industry. Communication interference source signal detection aims to determine whether interference source signals exist within the electromagnetic environment, providing a foundation for interference awareness and interference suppression. Therefore, exploring intelligent detection methods for communication interference signals in complex electromagnetic environments will lay a solid information security foundation for responding to interference threats, and has important research value and application prospects.

[0003] At present, research on communication interference source signal detection technology has made some progress. The methods proposed by domestic and foreign researchers for signal detection can be mainly divided into detection methods based on improved traditional algorithms and detection methods based on deep learning.

[0004] Regarding the detection method based on the improvement of traditional algorithms, Fan Guangwei et al. proposed an interference detection method based on energy detection, which judges the presence of interference signals by comparing with the threshold (Fan Guangwei, Deng Jiangna, Wang Zhenhua, et al. Research on GNSS interference detection technology based on energy detection [J]. Radio Engineering, 2013, 43(3): 33-35, 39.). Kang et al. combined forward CME with gradient clustering, constructed the maximum value of each cluster obtained by gradient clustering as a new data sample, and then used forward CME detection (Kang Y, Wu H, Zhao Z, et al. Signal detection based on gradient clustering and forward consecutive mean excision [C] / / 2022 IEEE 9th International Symposium on Microwave, Antenna, Propagation and EMC Technologies for Wireless Communications (MAPE). Chengdu, China: IEEE, 2022: 341-346.). The performance of the above detection methods has been improved compared to detection methods based on traditional algorithms. However, these methods rely on the setting of parameters such as thresholds and are sensitive to noise. They are unable to effectively cope with complex electromagnetic environments and perform poorly in the face of complex data situations.

[0005] Regarding deep learning-based detection methods, Gao et al. constructed a signal detection framework called DetectNet, which takes raw signals as input and is easily transferable to any detection task (Gao J, Yi X, Zhong C, et al. Deeplearning for spectrum sensing [J]. IEEE Wireless Communications Letters, 2019, 8(6): 1727-1730.). Li et al. collected four types of interference signals and used their spectrograms as data sets to train four types of networks, including AlexNet. The results showed that compared with manually extracted features, the detection performance of the deep learning method was improved by nearly 10%. (Li Y, Pawlak J, Price J, et al. Jamming detection and classification in OFDM-based UAVs via feature and spectrogram-tailored machine learning [J]. IEEE Access, 2022, 10: 16859-16870.). The above detection methods utilize the feature extraction capabilities of deep learning, but the signal representation is insufficient, the feature fusion is not comprehensive, and the complementarity of multi-domain features of the signal is ignored, which limits the improvement of detection accuracy.

[0006] Through the above analysis, the problems and defects of the existing technology are as follows:

[0007] (1) Detection methods based on improved traditional algorithms require less prior information, have simple principles, and are less complex. However, these methods rely on the setting of parameters such as thresholds, are sensitive to noise, and are difficult to effectively cope with complex electromagnetic environments. They also have difficulty taking advantage of the advantages of neural networks.

[0008] (2) Although the detection method based on deep learning effectively utilizes the advantages of neural networks, the existing methods do not fully represent the signal, the feature fusion is not comprehensive, and the complementarity of the multi-domain features of the signal is ignored, which limits the improvement of the detection accuracy. Summary of the Invention

[0009] In response to the problems existing in the prior art, the present invention provides a communication interference signal detection method, system, device and medium.

[0010] The present invention is implemented as follows: a communication interference signal detection method, system, device and medium comprising:

[0011] In order to overcome the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a method, system, device and medium for intelligent detection of communication interference signals; using adaptive variational mode decomposition, Gram angular field and cyclic spectrum to process communication interference source signals, respectively obtaining time domain, image domain and frequency domain intelligent representations of the communication interference source, and using these three representations as inputs of the communication interference source signal detection network; constructing and training a communication interference source signal detection network comprising three branches to extract the deep features of each intelligent representation respectively, and fuse the deep features; using the fused features to construct a detection statistic, and calculating and obtaining the detection threshold under different false alarm probabilities according to the different false alarm probabilities; the communication interference source signal to be detected is processed by adaptive variational mode decomposition, Gram angular field and cyclic spectrum and input into the trained detection network, and the detection result is obtained by comparing with the detection threshold under the preset false alarm probability. The three signal intelligent representations more fully extract the information contained in the signal; the fusion of the deep features of each intelligent representation realizes the complementarity of multi-domain information, further improving the detection performance; the detection result is obtained by comparing the detection statistic and the detection threshold, which is more in line with the actual detection task. The present invention has better performance and wider versatility.

[0012] In order to achieve the above object, the technical solution adopted by the present invention is:

[0013] A method for intelligently detecting communication interference signals, characterized in that the method comprises:

[0014] Step S1: Using adaptive variational mode decomposition, Gram angle field, and cyclic spectrum to process the communication interference source signal, respectively, obtain intelligent representations of the communication interference source in the time domain, image domain, and frequency domain, and use these three representations as inputs to the communication interference source signal detection network;

[0015] Step S2: construct and train a communication interference source signal detection network containing three branches to extract the deep features of each intelligent representation respectively, and fuse the deep features;

[0016] Step S3: constructing detection statistics using fusion features, and calculating and obtaining detection thresholds under different false alarm probabilities according to different false alarm probabilities;

[0017] Step S4: The communication interference source signal to be detected is processed using adaptive variational mode decomposition, Gram angle field and cyclic spectrum and then input into the trained detection network. The detection result is obtained by comparing it with the detection threshold under the preset false alarm probability.

[0018] The specific method of step S1 is:

[0019] Step S1.1: Process the original communication interference source signal X using the adaptive variational mode decomposition method to obtain its time domain representation vector m, which serves as the time domain intelligent representation of the communication interference source signal;

[0020] The adaptive variational mode decomposition method uses the global search advantage of the PSO algorithm to obtain the variational mode decomposition mode number K and penalty factor α that best suits the signal through continuous iteration of the fitness function. The fitness function is the envelope entropy E of the signal. p , the mathematical expression is:

[0021]

[0022] Where a(j) represents the signal envelope, p j Represents the probability distribution sequence of a(j).

[0023] In order to unify the size of the time domain representation vector, a maximum decomposition number M is set based on the modal number K. If K < M, the missing vectors are filled with 0.

[0024] Step S1.2: Process the original communication interference source signal s(t) using the Gram angular field algorithm to obtain an image n of the processed signal as an image domain intelligent representation of the communication interference source signal;

[0025] The Gram angular field algorithm is an algorithm that encodes a time series into an image, which can maintain the complete time series characteristics while mapping. If the number of sampling points of the communication interference source signal X is n, X is first normalized to [-1, 1] or [0, 1] to reduce the influence of the inner product on the maximum value. The mathematical expression is:

[0026]

[0027] in, Represents the signal value normalized to [-1,1], Indicates the signal value normalized to [0,1]. The normalized signal is obtained at this time. Then, Mapped into polar coordinates, Encoded as angle Time t is encoded as radius r, and its mathematical expression is:

[0028]

[0029] Among them, t i is the timestamp.

[0030] Based on the different trigonometric functions calculated, the Gram angle field can be divided into the Gram angle difference field and the Gram angle sum field. The Gram angle difference field uses the sine difference between angles, while the Gram angle sum field uses the cosine sum of angles. For an interference signal with n sampling points, a Gram angle field image with a maximum size of n×n can be converted. This method allows the communication interference source signal to be converted into a two-dimensional image while preserving its time dependence, facilitating feature extraction by deep learning networks.

[0031] Step S1.3: Process the original communication interference source signal s(t) using a cyclic spectrum algorithm to obtain an image q of the processed signal as a graph-domain intelligent representation of the communication interference source signal;

[0032] The cyclic spectrum algorithm can simultaneously reveal the frequency domain distribution and implicit modulation characteristics of a signal, significantly enhancing the noise immunity of the resulting representation compared to traditional frequency domain analysis methods. The cyclic spectrum can effectively distinguish communication interference signals from noise. This is because Gaussian white noise is completely uncorrelated in the time domain and lacks any periodicity or cyclic characteristics. Therefore, its cyclic spectrum only appears as a horizontal line when the cyclic frequency α = 0. When α ≠ 0, it theoretically has no components.

[0033] To match the structure of the neural network, the present invention normalizes the cyclic spectrum to the range of [0, 1] on the Z axis, and then extracts a two-dimensional image of the cyclic spectrum in the XY plane, thereby converting the three-dimensional cyclic spectrum into a two-dimensional spectrum graph.

[0034] The specific method of step S2 is:

[0035] Step S2.1: For the temporal representation vector obtained in step S1, a temporal convolutional network with a spatial-channel coupled attention mechanism is constructed as branch network 1 to extract features;

[0036] The spatial-channel coupled attention mechanism module in the branch network 1 is composed of a shared multi-semantic spatial attention mechanism and a progressive channel self-attention mechanism. The feature map is first decomposed into multiple groups of independent sub-features according to height and width, and then each group of sub-features is subjected to a multi-receptive field shared depth one-dimensional convolution of different scales to capture the multi-semantic spatial structure. Then, the groups of features are spliced and each group of features is independently normalized to retain the semantic independence between the sub-features and avoid feature dilution. Next, the output enhanced spatial features are input into the progressive channel self-attention mechanism module, which successively performs progressive compression, channel self-attention mechanism calculation, global average pooling and channel weighted output on the spatial features, and finally obtains the feature map enhanced by the attention mechanism. The synergy of SCSA is reflected in the multi-semantic information of the shared multi-semantic spatial attention mechanism helping the progressive channel self-attention mechanism to more accurately allocate channel weights, and the global channel interaction of the progressive channel self-attention mechanism compensates for the local deviation of the shared multi-semantic spatial attention mechanism.

[0037] The temporal convolutional network in the branch network 1 uses convolution instead of recursive operation to process time series data. While inheriting the advantages of recurrent neural networks, it avoids the problems of gradient vanishing and low computational efficiency. Compared with traditional networks such as LSTM and GRU, it has the characteristics of being able to efficiently process long-term dependencies, flexible channel number setting, and low model complexity, and is very suitable for extracting multi-channel time series features.

[0038] Step S2.2: Based on the graph domain intelligent representation and frequency domain intelligent representation obtained in step S1, a lightweight ConvNeXt network is constructed as branch network 2 and branch network 3 respectively to extract features;

[0039] The network structures of branch network 2 and branch network 3 are exactly the same, and ConvNeXt is used as the basic network. This network draws on the basic structure of the ResNet network and the design concept of Swin Transformer (Swin-T). While surpassing the Transformer network in multiple tasks such as image classification, it has faster reasoning speed and smaller number of parameters. The network is further improved in two aspects. First, for the stacking ratio of 3:3:9:3 between the blocks of the original ConvNeXt-Tiny layers, the lightweight ConvNeXt is modified to 1:1:3:1 while maintaining the relative ratio, significantly reducing the number of parameters while reducing the degree of network fragmentation; secondly, the Drop Path of the third layer of convolution in the ConvNeXt block is removed to further reduce the computational complexity of the network;

[0040] Step S2.3: Combine the branch networks 1 to 3 obtained in step S2.2 into a communication interference source signal detection network, and use the representation obtained in step S1 to train the network to achieve deep feature fusion.

[0041] During the training of the communication interference source signal detection network, binary cross entropy is used as the loss function. Binary cross entropy can measure the difference between the probability predicted by the model and the true label. The mathematical expression is:

[0042]

[0043] Branch networks 1 to 3 are trained in parallel, each outputting corresponding deep features, which are then fused through splicing. Training ends when the number of training rounds meets the required number of iterations or the loss function falls below a threshold, resulting in a detection model with optimal parameters.

[0044] The specific method of step S3 is:

[0045] Step S3.1: Introduce the selection vector e i , the mathematical expression is:

[0046]

[0047] After the test set is input into the network, the output is They represent the probability of detecting the absence of a signal and the probability of detecting the presence of a signal. and All probability values are within (0,1), so we can refer to the construction of the test statistic. The mathematical expression of the test statistic is:

[0048]

[0049] Step S3.2: Extract the sample data with only noise in the training set and input it into the network. The output is The mathematical expression of the false alarm probability at this time is:

[0050]

[0051] The K T noise The values are sorted in descending order, and the A set of sets represented by

[0052]

[0053] The detection threshold is expressed as:

[0054]

[0055] in, Indicates rounding down to the nearest integer.

[0056] The specific method of step S4 is:

[0057] After the test set is input into the network, the obtained detection statistic T and the detection threshold η are used to make a detection decision according to the following formula:

[0058]

[0059] When T>η, it is assumed that H1 is true, and it is determined that an interference signal exists; when T<η, it is assumed that H0 is true, and it is determined that no interference signal exists.

[0060] A system for an intelligent detection method for communication interference signals, characterized by comprising:

[0061] An intelligent characterization module, configured to intelligently characterize the communication interference source signal in step S1 as an input to the detection network model;

[0062] The communication interference source signal detection network, which includes a temporal convolutional network with a spatial-channel coupled attention mechanism and a lightweight ConvNeXt network, is used to extract the deep features of each intelligent representation in step S2 and achieve feature fusion;

[0063] A detection statistic construction method is used to construct the detection statistic in step S3 to obtain threshold values under different false alarm probabilities;

[0064] A device for an intelligent detection method for communication interference signals includes: a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it can implement any of the above-mentioned intelligent detection methods for communication interference signals.

[0065] A computer storage medium for receiving a program input by a user. When the computer program stored in the storage medium is executed by a processor, communication interference source signal detection can be performed based on the communication interference signal intelligent detection method described in any one of the above technical solutions.

[0066] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0067] First, in view of the problems in the above-mentioned prior art of communication interference signal detection methods based on improved traditional algorithms, such as strong parameter sensitivity, susceptibility to noise interference, and low detection accuracy in complex electromagnetic environments, as well as the technical difficulties such as insufficient single feature representation and insufficient multi-domain information fusion based on deep learning methods that limit further improvement of detection performance, the present invention closely combines the proposed technical solutions with the experimental results and data obtained during the research and development process, and analyzes in detail how the present invention effectively solves the above-mentioned technical problems and brings about obvious creative technical effects. The specific description is as follows:

[0068] The present invention proposes a method for intelligent characterization of communication interference source signals, obtains intelligent characterization in each domain, effectively extracts multi-domain information of the signal, and thus more fully characterizes the information of the interference source signal.

[0069] The present invention proposes to train a communication interference source signal detection network containing three branches. The network selects a suitable network according to the modal characteristics of each representation, fully considers the unique characteristics of the modal features, more deeply mines the signal features, and realizes the complementarity between multi-domain features, thereby improving the detection accuracy.

[0070] Compared to existing technologies, this invention explores the importance of multi-domain characterization and feature fusion in communication interference source signal detection. It can be used in any scenario where communication interference source signal detection is designed, especially in complex electromagnetic environments. This invention offers higher detection performance and greater versatility.

[0071] The intelligent detection method for communication interference signals proposed in the present invention addresses the problems in the existing technology of relying on a single signal characterization method, limited detection accuracy, and poor adaptability to complex interference forms. It introduces three feature extraction mechanisms for the first time: adaptive variational mode decomposition, Gram angular field, and cyclic spectrum. It extracts time domain, image domain, and frequency domain representations respectively, constructs a multi-branch fusion signal feature system, and fundamentally improves the information utilization and discrimination of the original interference signal.

[0072] Traditional interference detection methods often rely on rule-based discrimination or single deep models, which can easily lead to misjudgments when encountering diverse interference types or complex channel conditions. This paper designs a three-branch deep detection network architecture, constructing dedicated feature extraction paths for different modal features. This architecture integrates a time series convolutional model with an attention mechanism and a lightweight image convolutional model, significantly improving the targetedness and expressiveness of feature extraction, and effectively enhancing the model's robustness and adaptability.

[0073] In terms of detection statistic construction, the present invention constructs a probability detection criterion based on the output of a neural network, introduces a selection vector and a statistical function, and dynamically estimates the detection threshold in combination with noise-only samples. For the first time, it realizes an adjustable control mechanism for the false alarm probability of the detection threshold, breaking through the bottleneck that the fixed threshold mechanism is difficult to adapt to complex channel backgrounds, and improving the generalization ability of the detection system to different interference intensities and forms.

[0074] The present invention realizes high-precision, high-robustness, and high-versatility intelligent detection of communication interference signals through modal decomposition and fusion modeling at the algorithm layer, branch optimization and attention coupling at the network structure layer, and threshold adaptive adjustment at the statistical decision layer. Compared with the existing technology, it has achieved significant improvements in adaptability to complex environments, detection accuracy and computational efficiency.

[0075] Second, the intelligent detection technology for communication interference signals proposed in the present invention achieves high-precision and high-robustness detection of interference signals in complex electromagnetic environments by constructing an innovative multi-modal intelligent characterization method and a fusion feature detection network, significantly improving the accuracy and generalization of interference signal detection. Once this technology is industrialized, it can be widely used in satellite Internet, Internet of Things, smart transportation, military communications and other fields, significantly improving the anti-interference capability of wireless communication systems, ensuring safe and reliable information transmission, and thus bringing about significant improvements in the quality of communication services. In addition, the technical solution of the present invention has good scalability and adaptability, and can be combined with existing communication equipment to quickly implement deployment and upgrades, which is expected to create considerable economic benefits and market competitive advantages for related equipment manufacturers and communication operators.

[0076] Existing communication interference signal detection methods usually perform signal detection based on a single domain or a single type of feature representation, and are unable to fully utilize the complementary characteristics of multi-domain information within the signal. Especially in complex environments with low interference-to-noise ratios, the detection accuracy and generalization are significantly limited. This invention organically combines three feature extraction methods, namely adaptive variational mode decomposition, Gram angular field, and cyclic spectrum, for the first time, extracting and fusing the multimodal features of the interference signal from the time domain, image domain, and frequency domain respectively, breaking through the technical bottleneck of single feature representation and insufficient fusion. This invention successfully realizes the collaborative learning and fusion of features from different domains, effectively improving the interference signal detection performance, filling the technical gap in the field of multi-domain feature fusion intelligent detection at home and abroad, and has outstanding technical innovation. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a flow chart of the communication interference signal intelligent detection method, system, medium and equipment provided by an embodiment of the present invention.

[0078] Figure 2 It is a schematic diagram of a communication interference source signal detection network in the communication interference source signal detection method provided by an embodiment of the present invention.

[0079] Figure 3 It is a schematic diagram of simulation experiment results of the detection performance of the communication interference signal intelligent detection method provided by an embodiment of the present invention.

[0080] Figure 4 It is a schematic diagram of simulation experiment results comparing the detection performance of the communication interference signal intelligent detection method provided by an embodiment of the present invention with that of existing detection methods.

[0081] Figure 5 It is a schematic diagram of simulation results comparing the contributions of different branch feature combinations to the detection performance of the communication interference signal intelligent detection method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0083] The invention discloses an intelligent detection method, system, device and medium for communication interference signals. The method comprises: processing communication interference source signals using adaptive variational mode decomposition, Gram angular field and cyclic spectrum to obtain intelligent representations of the communication interference source signals in time domain, graph domain and frequency domain respectively; constructing and training a communication interference source signal detection network comprising three branches, extracting and fusing deep features of each intelligent representation respectively; constructing a detection statistic using the fused features to obtain detection thresholds under different false alarm probabilities; inputting the processed communication interference source signal to be detected into the trained detection network, and obtaining a detection result by comparing it with the detection threshold under a preset false alarm probability; performing intelligent representation of the communication interference source signal in time domain, graph domain and frequency domain to effectively extract multivariate information of the original signal; the constructed communication interference source signal detection network selects a suitable branch network according to the modal characteristics of each representation, thereby improving the feature extraction effect and achieving complementarity and fusion between different deep features. The invention has better performance and wider versatility. The system, device and medium are used to implement an intelligent detection method for communication interference signals.

[0084] The communication interference signal intelligent detection method, system, medium and equipment provided by the present invention can also be implemented by ordinary technicians in the industry using other steps. Figure 1 The communication interference signal intelligent detection method, medium and device provided by the present invention are only a specific embodiment.

[0085] like Figure 1 As shown, the communication interference signal intelligent detection method provided by the embodiment of the present invention has the following specific steps:

[0086] Step S1: Use adaptive variational mode decomposition, Gram angle field, and cyclic spectrum to process the communication interference source signal to obtain intelligent representations of the communication interference source in the time domain, image domain, and frequency domain, respectively. These three representations are used as inputs to the communication interference source signal detection network. The specific process is as follows:

[0087] Step S1.1: Process the original communication interference source signal X using the adaptive variational mode decomposition method to obtain its time domain representation vector m, which serves as the time domain intelligent representation of the communication interference source signal;

[0088] The adaptive variational mode decomposition method uses the global search advantage of the PSO algorithm to obtain the variational mode decomposition mode number K and penalty factor α that best suits the signal through continuous iteration of the fitness function. The fitness function is the envelope entropy E of the signal. p , the mathematical expression is:

[0089]

[0090] Where a(j) represents the signal envelope, p j Represents the probability distribution sequence of a(j).

[0091] In order to unify the size of the time domain representation vector, a maximum decomposition number M is set based on the modal number K. If K < M, the missing vectors are filled with 0.

[0092] Step S1.2: Process the original communication interference source signal s(t) using the Gram angular field algorithm to obtain an image n of the processed signal as an image domain intelligent representation of the communication interference source signal;

[0093] The Gram angular field algorithm is an algorithm that encodes a time series into an image, which can maintain the complete time series characteristics while mapping. If the number of sampling points of the communication interference source signal X is a, X is first normalized to [-1, 1] or [0, 1] to reduce the influence of the inner product on the maximum value. The mathematical expression is:

[0094]

[0095] in, Represents the signal value normalized to [-1,1], Indicates the signal value normalized to [0,1]. The normalized signal is obtained at this time. Then, Mapped into polar coordinates, Encoded as angle Time t is encoded as radius r, and its mathematical expression is:

[0096]

[0097] Among them, t i is the timestamp.

[0098] Based on the different trigonometric functions calculated, the Gram angle field can be divided into the Gram angle difference field and the Gram angle sum field. The Gram angle difference field uses the sine difference between angles, while the Gram angle sum field uses the cosine sum of angles. For an interference signal with a number of sampling points, a Gram angle field image with a maximum size of a×a can be converted. This method allows the communication interference source signal to be converted into a two-dimensional image while preserving its time dependence, facilitating feature extraction by deep learning networks.

[0099] Step S1.3: Process the original communication interference source signal s(t) using a cyclic spectrum algorithm to obtain an image q of the processed signal as a graph-domain intelligent representation of the communication interference source signal;

[0100] The cyclic spectrum algorithm can simultaneously reveal the frequency domain distribution and implicit modulation characteristics of a signal, significantly enhancing the noise immunity of the resulting representation compared to traditional frequency domain analysis methods. The cyclic spectrum can effectively distinguish communication interference signals from noise. This is because Gaussian white noise is completely uncorrelated in the time domain and lacks any periodicity or cyclic characteristics. Therefore, its cyclic spectrum only appears as a horizontal line when the cyclic frequency α = 0. When α ≠ 0, it theoretically has no components.

[0101] To match the structure of the neural network, the present invention normalizes the cyclic spectrum to the range of [0, 1] on the Z axis, and then extracts a two-dimensional image of the cyclic spectrum in the XY plane, thereby converting the three-dimensional cyclic spectrum into a two-dimensional spectrum graph.

[0102] Step S2: Construct and train a communication interference source signal detection network containing three branches to extract the deep features of each intelligent representation and fuse the deep features. The specific process is as follows:

[0103] Step S2.1: For the temporal representation vector obtained in step S1, a temporal convolutional network with a spatial-channel coupled attention mechanism is constructed as branch network 1 to extract features;

[0104] The spatial-channel coupled attention mechanism module in the branch network 1 is composed of a shared multi-semantic spatial attention mechanism and a progressive channel self-attention mechanism. The feature map is first decomposed into multiple groups of independent sub-features according to height and width, and then each group of sub-features is subjected to a multi-receptive field shared depth one-dimensional convolution of different scales to capture the multi-semantic spatial structure. Then, the groups of features are spliced and each group of features is independently normalized to retain the semantic independence between the sub-features and avoid feature dilution. Next, the output enhanced spatial features are input into the progressive channel self-attention mechanism module, which successively performs progressive compression, channel self-attention mechanism calculation, global average pooling and channel weighted output on the spatial features, and finally obtains the feature map enhanced by the attention mechanism. The synergy of SCSA is reflected in the multi-semantic information of the shared multi-semantic spatial attention mechanism helping the progressive channel self-attention mechanism to more accurately allocate channel weights, and the global channel interaction of the progressive channel self-attention mechanism compensates for the local deviation of the shared multi-semantic spatial attention mechanism.

[0105] The temporal convolutional network in the branch network 1 uses convolution instead of recursive operation to process time series data. While inheriting the advantages of recurrent neural networks, it avoids the problems of gradient vanishing and low computational efficiency. Compared with traditional networks such as LSTM and GRU, it has the characteristics of being able to efficiently process long-term dependencies, flexible channel number setting, and low model complexity, and is very suitable for extracting multi-channel time series features.

[0106] Step S2.2: Based on the graph domain intelligent representation and frequency domain intelligent representation obtained in step S1, a lightweight ConvNeXt network is constructed as branch network 2 and branch network 3 respectively to extract features;

[0107] The network structures of branch network 2 and branch network 3 are exactly the same, and ConvNeXt is used as the basic network. This network draws on the basic structure of the ResNet network and the design concept of Swin Transformer (Swin-T). While surpassing the Transformer network in multiple tasks such as image classification, it has faster reasoning speed and smaller number of parameters. The network is further improved in two aspects. First, for the stacking ratio of 3:3:9:3 between the blocks of the original ConvNeXt-Tiny layers, the lightweight ConvNeXt is modified to 1:1:3:1 while maintaining the relative ratio, significantly reducing the number of parameters while reducing the degree of network fragmentation; secondly, the Drop Path of the third layer of convolution in the ConvNeXt block is removed to further reduce the computational complexity of the network;

[0108] Step S2.3: Combine the branch networks 1 to 3 obtained in step S2.2 into a communication interference source signal detection network, and use the representation obtained in step S1 to train the network to achieve deep feature fusion.

[0109] During the training of the communication interference source signal detection network, binary cross entropy is used as the loss function. Binary cross entropy can measure the difference between the probability predicted by the model and the true label. The mathematical expression is:

[0110]

[0111] Branch networks 1 to 3 are trained in parallel, each outputting corresponding deep features, which are then fused through splicing. Training ends when the number of training rounds meets the required number of iterations or the loss function falls below a threshold, resulting in a detection model with optimal parameters.

[0112] Step S3: Utilize the fusion features to construct the detection statistics, calculate and obtain the detection threshold under different false alarm probabilities according to the different false alarm probabilities. The specific process is as follows:

[0113] Step S3.1: Introduce the selection vector e i , the mathematical expression is:

[0114]

[0115] After the test set is input into the network, the output is They represent the probability of detecting the absence of a signal and the probability of detecting the presence of a signal. and All probability values are within (0,1), so we can refer to the construction of the test statistic. The mathematical expression of the test statistic is:

[0116]

[0117] Step S3.2: Extract the sample data with only noise in the training set and input it into the network. The output is The mathematical expression of the false alarm probability at this time is:

[0118]

[0119] The K T noise The values are sorted in descending order, and the A set of sets represented by

[0120]

[0121] The detection threshold is expressed as:

[0122]

[0123] in, Indicates rounding down to the nearest integer.

[0124] Step S4: The communication interference source signal to be detected is processed using adaptive variational mode decomposition, Gram angle field and cyclic spectrum, and then input into the trained detection network. The detection result is obtained by comparing it with the detection threshold under the preset false alarm probability. The specific process is as follows:

[0125] After the test set is input into the network, the obtained detection statistic T and the detection threshold η are used to make a detection decision according to the following formula:

[0126]

[0127] When T>η, it is assumed that H1 is true, and it is determined that an interference signal exists; when T<η, it is assumed that H0 is true, and it is determined that no interference signal exists.

[0128] The communication interference signal intelligent detection method proposed in the present invention can be widely applied in the field of communications, and is particularly suitable for interference detection scenarios in complex electromagnetic environments, including but not limited to:

[0129] (1) Military communications

[0130] It is suitable for military communication interference detection and suppression systems, and plays an important role in military applications such as electronic countermeasures, military reconnaissance, and radar detection, improving anti-interference capabilities and ensuring the security and reliability of military communications.

[0131] (2) Satellite Internet System

[0132] The signal monitoring and interference detection equipment is suitable for low-orbit satellite Internet, which improves the real-time response capability of satellite communication systems to malicious interference, spectrum occupancy and other issues, and ensures the stable operation of satellite communication links.

[0133] (3) Internet of Things (IoT) communication equipment

[0134] In the fields of smart cities, smart homes, and smart industrial manufacturing, it can be applied to signal detection modules of various IoT communication node devices to improve the signal quality monitoring and anti-interference capabilities of IoT devices.

[0135] (4) Intelligent transportation system

[0136] Applicable to vehicle-to-everything (V2X) equipment in smart transportation scenarios, especially communication signal detection equipment for on-board communication terminals or roadside units (RSUs), to improve the reliability and safety of traffic communication systems in complex road environments.

[0137] (5) Communication network operators and regulatory authorities

[0138] Provide communication network operators with advanced interference detection and positioning systems to help operators and regulatory authorities monitor and manage the wireless communication environment and ensure the operation quality and information security of public communication networks.

[0139] (6) Emergency communication equipment

[0140] The technical solution of the present invention can be integrated into emergency communication equipment deployed during major disasters, emergencies or special events to quickly discover and identify interference sources and ensure smooth emergency communication channels.

[0141] In summary, the technical solution of the present invention has a wide range of applications, and related products include but are not limited to communication interference detection terminal equipment, spectrum monitoring systems, intelligent interference positioning instruments, military electronic countermeasures equipment and communication network intelligent monitoring and management platforms, etc., which have broad market prospects and significant social and economic benefits.

[0142] In order to evaluate the technical effect obtained by the embodiment of the present invention, simulation verification is carried out. In order to simulate the distribution of different interference sources in a complex electromagnetic environment, Matlab software is used to simulate seven types of interference, including single tone jamming (STJ), multi tone jamming (MTJ), noise frequency modulation jamming (NFMJ), linear frequency modulation jamming (LFMJ), comb spectrum jamming (COMB), partial band noise jamming (PBNJ) and BPSK interference. During the simulation process, the sampling rate f s The frequency spectrum is 20 MHz, the number of sampling points is 1000, and the channel model is an additive white Gaussian noise channel. The simulated jamming-to-noise ratio (JNR) range is [-16:2:16] dB. 200 samples of each interference signal are generated at each JNR. All generated interference signals are considered positive samples, and an equal number of white Gaussian noise samples are generated as negative samples, so that the ratio of positive and negative samples is 1:1. Then, all signals are processed according to the proposed representation method, and three corresponding representations are obtained for each signal. The size of the Gram angle field map obtained by the image domain representation and the CS map obtained by the frequency domain representation are both 224×224×3. The maximum decomposition number M set for the time domain representation is 8, resulting in a size of each representation matrix of 9×1000. After alignment, the training set, validation set, and test set are divided into training set, validation set, and test set in a ratio of 9:1:10.

[0143] In order to effectively evaluate the performance of the intelligent detection method for communication interference signals, the detection probability P commonly used in detection problems is used. d As an evaluation index of detection performance. In particular, in practical applications, it is usually necessary to f Detection is performed under the condition of poor detection, so the present invention introduces the receiver operating characteristic (ROC) curve as another important indicator for evaluating detection performance. The ROC curve is a tool for evaluating the performance of a binary classification model. Its horizontal axis is the false positive rate (FPR), which indicates the probability that a negative class sample is predicted as a positive class. The vertical axis is the true positive rate (TPR), which indicates the probability that a positive class sample is correctly predicted. The area under the ROC curve is the area under the curve (AUC). The larger the AUC, the better the detection performance.

[0144] Figure 3 The detection performance of the method proposed in the present invention is specifically demonstrated in FIG. Figure 3 It can be seen that the detection probability of the seven interference signals increases with the increase of JNR. Specifically, the detection probability of STJ and MTJ has reached 1 when the interference-to-noise ratio is about -10dB, and when the JNR continues to rise to -8dB, the detection probability of the seven interference signals is already higher than 0.8, and when the interference-to-noise ratio is -2dB, the detection probability of the seven interference signals is close to 1. It can be seen that the proposed method can effectively detect the seven interference signals, and achieve good detection performance under low JNR, and when the interference-to-noise ratio is in the range of [-14dB, -6dB], the detection probability of the seven interference signals increases significantly, which also shows that the generalization and robustness of the proposed method are good. This shows that the clustering results of the present invention have the characteristics of tight intra-class and high inter-class distinction, which effectively improves the clustering effect when there is overlap in parameters.

[0145] Figure 4 The comparison of the detection performance of the proposed method with the existing methods is specifically shown in the paper. The selected comparison methods are Energy Detection (ED), Continuous Mean Excision (CME), Support Vector Data Description (SVDD) and DetectNet. Figure 4As can be seen from the figure, the AUC values of the five methods are 0.9602, 0.8756, 0.8013, 0.6776, and 0.5905, respectively. The proposed method achieves the highest AUC value, which again demonstrates that the proposed method performs well and maintains a certain degree of sensitivity for different false alarm probabilities. In particular, when the false alarm probability is 0.1, the detection probability of the proposed method is 0.93, which is more than 20% higher than that of the other methods. When the false alarm probability continues to decrease, the detection performance of the proposed method also outperforms the comparison methods.

[0146] Figure 5 The contribution of different branch feature combinations to the detection performance of the proposed method is specifically shown in FIG. CS represents the cyclic spectrum and GADF represents the Gram angle difference field. Figure 5It can be seen that when a single-branch feature input is used, the AUC values corresponding to the CS graph, GADF graph, and time domain characterization vector are 0.8821, 0.7777, and 0.6863, respectively. The AUC value corresponding to the CS graph is the highest among the three features. This is because this experiment uses the NFMJ signal under JNR = -8dB. The impact of low JNR on the signal frequency domain is often smaller than the impact on the time domain, and the cyclic spectrum is a frequency domain analysis method with strong noise resistance, so it can achieve better results. The GADF graph and the time domain characterization vector are both obtained from the time domain of the original interference signal, and the AUC value corresponding to the GADF graph is higher than the AUC value corresponding to the time domain characterization vector. This is because although the present invention proposes a time domain characterization method based on adaptive VMD to address the problem that the interference signal characteristics are not obvious under low JNR, and the communication interference signal is characterized by a time domain vector to achieve noise suppression and signal enhancement, the key features of the signal under low JNR may have been affected. At this time, the use of VMD decomposition will bring certain improvements, but it fails to completely overcome the limitations of time domain processing. The GADF map maps the communication interference signal into a two-dimensional image. During this mapping process, the noise's interference with the overall image texture is relatively weakened, thereby converting the communication interference signal into a more robust input. When using a dual-branch feature input, the combination of the GADF map and the CS map achieves the best AUC value of 0.9457. Comparing this combination with the results of Experiment 2, it is found that using only the GADF map and the CS map performs better than adding the original signal to these two branches. This is because at low JNRs, the signal is submerged in noise, making it difficult for the network to learn useful features of the original signal, and may even have a negative impact on the results. The combination of a time-domain representation vector and a CS map performs slightly worse, but still achieves a high AUC of 0.9218. The AUC of the GADF map and the time-domain representation vector is 0.7904, 0.1553 and 0.1314 lower than the two previous combinations, respectively. This confirms the previous analysis of single-branch input: for signals with extremely low JNR, after multi-branch feature fusion, the CS feature dominates the decision-making process, while the combination of the GADF map and the time-domain representation vector fails to utilize the signal's frequency domain information. The best performance is achieved when a three-branch feature input is used, as proposed in this chapter, with an AUC of 0.9602. This is a further improvement over the two-branch feature input, especially when the false alarm probability is low. This is because each branch network in the proposed TBFFNet learns rich information from the time, image, and frequency domains, and the fusion of features from different branches complements this information across multiple domains.

[0147] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0148] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A communication interference signal detection method, characterized in that: The following steps are involved: S1: The communication interference source signal is processed using adaptive variational mode decomposition, Gram's angle field, and cyclic spectrum methods to obtain the signal's time domain, image domain, and frequency domain representations, which serve as the input of the detection network. S2: Build and train a three-branch communication interference source signal detection network consisting of a temporal convolutional network and two image convolutional networks, extracting and fusing deep features of each representation; S3: Construct detection statistics based on the fused features and calculate the detection threshold based on the preset false alarm probability; S4: Input the signal to be detected into the trained detection network, compare the detection statistic with the threshold, and output the detection result of whether interference exists or not.

2. The communication interference signal detection method according to claim 1, wherein: The adaptive variational modal decomposition determines the optimal modal number and penalty factor through a particle swarm algorithm, and extracts time domain features using envelope entropy as a fitness function.

3. The communication interference signal detection method according to claim 1, wherein: The image domain representation is achieved through the Gram angle field, mapping the normalized time series into a two-dimensional image to retain the temporal characteristics; the frequency domain representation is generated by generating a two-dimensional spectrogram through cyclic spectrum extraction and normalization.

4. The communication interference signal detection method according to claim 1, wherein: The temporal convolutional network is a network that includes a spatial-channel coupled attention mechanism and is used to extract time domain features. The two image convolutional networks use a lightweight ConvNeXt structure to extract image domain and frequency domain features respectively.

5. The communication interference signal detection method according to claim 1, wherein: The detection statistic construction process includes performing a logarithmic ratio operation on the probability of existence of the signal output by the network, and constructing a threshold sequence based on training samples containing only noise to control the false alarm probability.

6. A communication interference signal detection system implementing the communication interference signal detection method according to any one of claims 1 to 5, characterized in that: The communication interference signal detection system comprises: Intelligent representation module, used to process the input signal into time domain, image domain and frequency domain representations respectively; The three-branch detection network module extracts the deep features of various representations and fuses them; Detection statistics construction module, used to construct detection values based on fusion features and set thresholds; The judgment module compares the detection value with the threshold and outputs the result of whether there is an interference signal.

7. A communication interference signal detection device implementing the communication interference signal detection method according to any one of claims 1 to 5, characterized in that: include: A memory, configured to store a communication interference signal detection program; The processor is used to call the program and execute a processing flow including representation generation, deep feature extraction, detection statistics construction and signal judgment.

8. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the program is used to implement the method according to any one of claims 1 to 5.

9. A time series feature extraction method based on a spatial-channel coupled attention mechanism for implementing the communication interference signal detection method according to any one of claims 1 to 5, characterized in that: First, the input feature map is decomposed into multiple semantic space features, then the semantic information is extracted through shared one-dimensional convolution, and finally the enhanced features are output through progressive compression and channel weighting mechanism.

10. A lightweight ConvNeXt image feature extraction method for implementing the communication interference signal detection method according to any one of claims 1 to 5, characterized in that: The adjusted network stacking ratio and convolution path configuration are used to extract features from the input image to adapt to the processing requirements of communication interference pattern domain or frequency domain representation.

Images