A Beidou multi-domain characteristic jamming signal identification method, system, device and medium

Abstract

This invention belongs to the field of electromagnetic interference detection technology for the BeiDou Navigation Satellite System, and discloses a method, system, device, and medium for identifying BeiDou multi-domain characteristic interference signals. This invention constructs a two-stage step-by-step identification strategy combining time-frequency maps and fourth-order cumulant maps. In each stage, an interference signal identification model based on a lightweight multi-scale network is built. In the first stage, based on the time-frequency map and the interference signal identification model, highly distinguishable interference types are identified. For easily confused interference signals that cannot be identified in the first stage, the second stage, based on the fourth-order cumulant map and the interference signal identification model, improves the accuracy of identifying easily confused interference types, ultimately improving the overall identification accuracy for complex interference. Simultaneously, the introduction of the lightweight multi-scale network effectively solves the problems of insufficient key feature capture and high computational cost of multi-scale structures in existing neural network technologies.

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic interference detection technology of BeiDou satellite navigation system, specifically involving a method, system, device and medium for identifying BeiDou multi-domain characteristic interference signals. Background Technology

[0002] The BeiDou Navigation Satellite System (BDS) is a global navigation satellite system independently constructed and operated by my country. With its continuous expansion of global networking and applications, BDS has been widely used in many fields such as transportation, agriculture, communications, and aviation, becoming a key infrastructure supporting the digital economy and social development. However, due to its vulnerability, BeiDou signals are susceptible to interference signals such as suppression, spoofing, and multipath effects. Common anti-interference methods generally rely on prior knowledge of the types of interference. Therefore, interference identification is a prerequisite for subsequent anti-interference efforts and is crucial for the stable operation of BDS.

[0003] With the development of neural networks, outlier detection technology based on neural networks has been increasingly widely used in satellite signal processing. Signal interference identification tasks based on neural networks typically involve two key steps: first, extracting highly discriminative features from the interference signal; and second, using neural network algorithms to perform classification and identification.

[0004] However, existing interference signal recognition methods based on traditional neural networks are limited by a single receptive field, making it difficult to effectively capture changes in key features of interference signals at different scales, especially lacking the ability to detect low-power, small-scale interference features. Although introducing multi-scale structures can help improve feature representation capabilities, their network structures are often too complex, leading to increased computational complexity and memory consumption. In addition, relying solely on a single type of interference signal feature makes it difficult to effectively distinguish between multiple types of interference, especially for interference signals with high feature similarity, where the recognition accuracy of existing methods remains unsatisfactory.

[0005] In summary, a method for identifying interference in BeiDou multi-domain features based on lightweight multi-scale neural networks needs to be proposed. Summary of the Invention

[0006] The purpose of this invention is to propose a method for identifying interference signals with multi-domain features of BeiDou. This method constructs a two-stage step-by-step identification strategy that combines time-frequency diagrams and fourth-order cumulative quantity diagrams. At each stage, an interference signal identification model based on a lightweight multi-scale network is built. By introducing the multi-domain features of BeiDou signals, it is beneficial to improve the identification accuracy of easily confused interference types. At the same time, it helps to solve the problems of insufficient capture of key features and high computational consumption of multi-scale structures in existing neural networks.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for identifying multi-domain interference signals in the BeiDou Navigation Satellite System includes the following steps:

[0009] Step 1. Generate interference-free BeiDou signal samples and interference signal samples with different interference signals added respectively;

[0010] The added interference signals include 12 types, including continuous wave interference, pulse interference, narrowband FM interference, band-limited Gaussian noise interference, and chirp interference with 6 different scanning shapes, repeater spoofing interference, and multipath interference.

[0011] For both interference-free signal samples and all interference signal samples, extract the corresponding time-frequency plots and construct the first training dataset;

[0012] Meanwhile, for interference signal samples with narrowband FM interference, repeater spoofing interference and multipath interference, the corresponding fourth-order cumulant map is extracted and a second training dataset is constructed.

[0013] Step 2. Construct a two-stage step-by-step identification strategy that combines time-frequency plots and fourth-order cumulant plots. In each stage, an interference signal identification model based on a lightweight multi-scale network is built.

[0014] The first stage of the interference signal identification model takes the time-frequency diagram of the signal to be identified as input and outputs any one of the following: no interference, continuous wave interference, pulse interference, band-limited Gaussian noise interference, chirp interference with six different scanning shapes, and interference types that cannot be identified. For interference signals that cannot be identified in the first stage, they are further input into the second stage for identification.

[0015] The input to the second-stage interference signal identification model is the fourth-order cumulative graph of the interference signal to be identified, and the output is narrowband FM interference, repeater spoofing interference, or multipath interference.

[0016] Step 3. Train the interference signal recognition models for the first and second stages based on the first and second training datasets respectively, and use the trained interference signal recognition models for the two stages to collaboratively identify various interference signals.

[0017] Furthermore, based on the aforementioned method for identifying BeiDou multi-domain feature interference signals, this invention also proposes a corresponding BeiDou multi-domain feature interference signal identification system. Both are based on the same inventive concept and employ the following technical solutions:

[0018] A BeiDou multi-domain feature interference signal identification system includes the following modules:

[0019] The preprocessing module is used to generate interference-free BeiDou signal samples and interference signal samples with different interference signals added. There are 12 types of interference signals added, including continuous wave interference, pulse interference, narrowband frequency modulation interference, band-limited Gaussian noise interference, and chirp interference with 6 different scanning shapes, repeater spoofing interference, and multipath interference.

[0020] For both interference-free signal samples and all interference signal samples, extract the corresponding time-frequency plots and construct the first training dataset;

[0021] Meanwhile, for interference signal samples with narrowband FM interference, repeater spoofing interference and multipath interference, the corresponding fourth-order cumulant map is extracted and a second training dataset is constructed.

[0022] And an interference signal identification module, used to construct a two-stage step-by-step identification strategy that combines time-frequency graphs and fourth-order cumulative graphs, with an interference signal identification model based on a lightweight multi-scale network built in each stage.

[0023] The first stage of the interference signal identification model takes the time-frequency diagram of the signal to be identified as input and outputs any one of the following: no interference, continuous wave interference, pulse interference, band-limited Gaussian noise interference, chirp interference with six different scanning shapes, and interference types that cannot be identified. For interference signals that cannot be identified in the first stage, they are further input into the second stage for identification.

[0024] The input to the second-stage interference signal identification model is the fourth-order cumulative graph of the interference signal to be identified, and the output is narrowband FM interference, repeater spoofing interference, or multipath interference.

[0025] The interference signal recognition models for the first and second stages are trained based on the first and second training datasets, respectively, and the trained interference signal recognition models for the two stages are used to collaboratively identify various interference signals.

[0026] Furthermore, based on the aforementioned method for identifying BeiDou multi-domain feature interference signals, this invention also proposes a computer device comprising a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the aforementioned method for identifying BeiDou multi-domain feature interference signals.

[0027] Furthermore, based on the aforementioned method for identifying BeiDou multi-domain feature interference signals, this invention also proposes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the aforementioned method for identifying BeiDou multi-domain feature interference signals.

[0028] The present invention has the following advantages:

[0029] As described above, this invention proposes a method for identifying BeiDou multi-domain feature interference signals. This method utilizes the complementarity of the multi-domain features of BeiDou interference signals in the time-frequency domain and the high-order cumulant domain to construct a two-stage step-by-step identification strategy combining the time-frequency map and the fourth-order cumulant map. In each stage, an interference signal identification model based on a lightweight multi-scale MMDC-CSPDarknet network is built. In the first stage, based on the time-frequency map and using the lightweight interference signal identification model, highly distinguishable interference types are identified. For easily confused interference signals such as narrowband noise FM interference, spoofing interference, and multipath interference that cannot be identified in the first stage, the second stage further improves the identification accuracy of easily confused interference types based on the fourth-order cumulant map and using the lightweight interference signal identification model, thereby improving the overall identification accuracy of complex interference. At the same time, the introduction of the lightweight multi-scale MMDC-CSPDarknet network in each stage helps to solve the problems of insufficient key feature capture and high computational cost of multi-scale structures in existing neural network technologies. In this invention, a lightweight multi-scale MMDC-CSP module is designed in the lightweight multi-scale MMDC-CSPDarknet network. The MMDC-CSP module integrates the MSBlock, DC, and MLCA modules and replaces the CSP module in the CSPDarknet model. To solve the problem that forced feature segmentation may cause some feature information to be diluted or lost during propagation, multi-scale features are introduced to improve the recognition accuracy of BeiDou signals and their interference signals when features are not obvious, while also effectively taking into account the lightweight goal. Experimental results show that the overall recognition accuracy of BeiDou interference using the MMDC-CSPDarknet model proposed in this invention can reach 97.5%. Attached Figure Description

[0030] Figure 1 This is a flowchart of the BeiDou multi-domain feature interference signal identification method in an embodiment of the present invention;

[0031] Figure 2 This is a network structure diagram of the interference signal identification model in the first stage of this embodiment of the invention;

[0032] Figure 3 This is a network structure diagram of the MMDC_CSP module in an embodiment of the present invention;

[0033] Figure 4 This is a network structure diagram of the MMDC submodule in an embodiment of the present invention;

[0034] Figure 5 This is a network structure diagram of the CBM module in an embodiment of the present invention;

[0035] Figure 6 This is a network structure diagram of the SPP module in an embodiment of the present invention;

[0036] Figure 7 This is a network structure diagram of the interference signal identification model in the second stage of this invention.

[0037] Figure 8 This is a comparison of the recognition accuracy of the method of the present invention with that of other neural networks at different interference-to-information ratios. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0039] Example 1

[0040] This embodiment 1 describes a method for identifying BeiDou multi-domain interference signals. The method first generates a dataset of BeiDou signals (e.g., BeiDou B1C signals) and 12 types of interference signals with different added interferences. Then, it constructs a step-by-step identification strategy based on multiple features from time-frequency maps and fourth-order cumulant maps. Next, a lightweight, multi-scale MMDC-CSPDarknet neural network is designed and applied to each step of the identification strategy, utilizing neural network image classification technology to identify and classify multi-domain interference features. Finally, a trained step-by-step identification classifier is obtained, and the network's recognition performance is verified using a test set. This invention effectively solves the problem of insufficient discrimination ability of single time-frequency domain features in distinguishing easily confused interference types.

[0041] like Figure 1 As shown, the BeiDou multi-domain feature interference signal identification method in this embodiment includes the following steps:

[0042] Step 1. Generate interference-free BeiDou signal samples and interference signal samples with different interference signals added respectively.

[0043] The added interference signals include 12 types, such as continuous wave interference, pulse interference, narrowband FM interference, band-limited Gaussian noise interference, and chirp interference with 6 different scanning shapes, repeater spoofing interference, and multipath interference.

[0044] For both interference-free signal samples and all interference signal samples, extract the corresponding time-frequency plots and construct the first training dataset.

[0045] Furthermore, for interference signal samples with narrowband FM interference, repeater spoofing interference, and multipath interference, the corresponding fourth-order cumulant maps are extracted, and a second training dataset is constructed.

[0046] The generated first and second training datasets are divided into training samples, test samples, and validation samples in a 6:2:2 ratio, respectively, for subsequent feature map extraction to generate datasets.

[0047] Considering the typical interference forms faced by BDS, this embodiment selects suppression interference, repeater-type spoofing interference, and multipath interference as signal samples to be identified. Among them, suppression interference includes continuous wave interference, pulse interference, narrowband FM interference, band-limited Gaussian noise interference, and chirp interference with six different scanning shapes.

[0048] Based on the different scanning shapes and scanning frequencies of the chirp signal, chirp interference can be further subdivided into sinusoidal chirp interference, sawtooth chirp interference, hook-sawtooth chirp interference, triangular chirp interference, linear chirp interference, and frequency-hopping chirp interference.

[0049] The B1C signal of the BeiDou-3 system is one of the primary signals adopted by the International Civil Aviation Organization.

[0050] This embodiment takes the B1C signal as an example, superimposing the BeiDou B1C signal, additive white Gaussian noise and 12 kinds of interference signals in the time domain, and using MATLAB to generate the signal sample to be identified.

[0051] The signal parameters are set as follows: the power of the BeiDou B1C signal is fixed at -159dBW, the power spectral density of additive white Gaussian noise is -205dBW / Hz, and the receiver bandwidth is 15MHz. The initial phase and carrier frequency of the interference signal carrier are uniformly distributed within [0,2π] and [0,20]MHz, respectively. Other settings are as follows: for continuous wave interference, the number of sinusoidal signals is set to 1, 2, or 3. When the value is 1, it represents single-tone continuous interference, and the rest represent multi-tone continuous interference; for pulse interference, the repetition period is uniformly distributed within the range of [0.02,1]ms; for narrowband frequency modulation interference, the bandwidth is uniformly distributed within the range of [0.1,2]MHz; for band-limited Gaussian noise interference, the bandwidth is uniformly distributed within the range of [2,20]MHz; and for each type of chirp interference, the scanning period is uniformly distributed within the range of [6.25,50]s, and the scanning bandwidth is uniformly distributed within the range of [1,20]MHz. To cover different interference intensities from weak to strong, the jam-to-signal ratio (JSR) range for each type of interference signal was determined as follows: the power level of the suppression interference signal varied between -144 and -109 dBW, with a total of 36 power levels, i.e., JSR ∈ [15, 50] dB, generating 210 samples under a 1-dBW power step of suppression interference, for a total of 7560 samples for each type of suppression interference; the power level of the repeater spoofing interference varied between -140 and -130 dBW, with a total of 10 power levels, i.e., JSR ∈ [19, 29] dB, generating 756 samples under a 1-dBW power step; and 756 samples were generated for each power step of multipath interference.

[0052] Using the generated signal samples, the time-frequency domain and higher-order statistical feature images of the signal to be identified are extracted as a dataset. By utilizing the complementarity of the multi-domain features of the signal, a step-by-step identification strategy combining the time-frequency map and the fourth-order cumulant map is constructed.

[0053] First, in the first stage, the time-frequency graph is calculated and fed into the recognition neural network to directly identify highly distinguishable interference types. Then, for easily confused interference types that were difficult to distinguish in the first stage, the process proceeds to the second stage.

[0054] In the second stage, the fourth-order cumulogram is recalculated as input, and a neural network is used for secondary recognition.

[0055] First, using STFT as a time-frequency analysis tool, time-frequency transformation is performed on the interference-free signal samples and various BeiDou interference signal samples in step 1. The calculated time-frequency map dataset is then used to directly identify interference types with high distinguishability.

[0056] The specific steps of STFT are as follows:

[0057] Choose a sliding window function to translate along the entire time axis of the signal, and perform a Fourier transform on each windowed signal segment. Then, take the square of the magnitude of the transform result to obtain the energy spectral density distribution within that time period.

[0058] The process is defined as follows:

[0059] (1)

[0060] in, This indicates the observed signal received by the receiver. It is a window function of length m. It is the step size of the window function at each slide, and its value is equal to the size of the window minus the overlap length. ,Right now .

[0061] In this invention, the window function is mainly calculated using the Hamming window.

[0062] Arranging these energy distributions at different times in chronological order creates a time-frequency graph, where the horizontal axis represents time and the vertical axis represents frequency. Different color intensities indicate the signal power at different times and frequencies. This graph visually demonstrates the characteristics of signal power variation with time and frequency, and utilizes a window function to localize the spectrum near a specific moment.

[0063] Using time-frequency plots as distinguishing features, it is relatively easy to identify interference-free, continuous wave, impulse, band-limited Gaussian noise, and six different types of chirped interference with varying scanning shapes. However, the time-frequency plots of narrowband FM interference, spoofing interference, and multipath interference exhibit high similarity in their narrowband spectral characteristics, making them easily confused during identification. Therefore, for interference-free, continuous wave, impulse, band-limited Gaussian noise, and chirped interference, the time-frequency plot is directly output as the result; for narrowband FM interference, spoofing interference, and multipath interference, a secondary identification process is performed, using a fourth-order cumulant plot as the basis for the second step of identification.

[0064] Next, the fourth-order cumulant map dataset obtained from the signal samples in step 1 is used for secondary recognition. The fourth-order cumulant is one of the higher-order statistical features, which can more effectively capture the non-Gaussianity and nonlinear structure information of the signal, suppress environmental Gaussian background noise, and reduce the impact of noise on the representation of BeiDou signal and interference characteristics.

[0065] In practice, to eliminate the influence of mean shift on the fourth-order cumulant, the signal is first centered (zero mean), and then its diagonal slice form is selected for cumulant estimation. The specific expression for the fourth-order cumulant is as follows:

[0066] (2)

[0067] in, This represents the mathematical expectation. In practical discrete computing, the above formula can be further expressed as:

[0068] (3)

[0069] in, The time delay is represented by a sliding delay window, which is used to calculate the fourth-order cumulative characteristic of the signal under different delays. Indicates the length of the overlapping portion. This indicates the total number of signal sampling points.

[0070] The signal representing the received signal after zero-mean processing is calculated as follows:

[0071] (4)

[0072] based on Construct two offset signal sequences and .

[0073] Ultimately The x-axis represents the corresponding fourth-order cumulant. The fourth-order cumulant two-dimensional image of the signal is obtained by using the vertical axis.

[0074] Step 2. Construct a two-stage, step-by-step identification strategy combining time-frequency plots and fourth-order cumulant plots. At each stage, an interference signal identification model based on a lightweight multi-scale network is built, such as... Figure 2 and Figure 7 As shown.

[0075] The input to the interference signal identification model in the first stage is the time-frequency diagram of the signal to be identified;

[0076] The first stage of the interference signal identification model outputs one of the following: no interference, continuous wave interference, impulse interference, band-limited Gaussian noise interference, chirped interference with six different scanning shapes, and any one of the interference types that cannot be identified.

[0077] For interference signals that cannot be identified in the first stage, they are further input into the second stage for identification.

[0078] The input to the second-stage interference signal identification model is the fourth-order cumulative quantity map of the interference signal to be identified, and the output of the second-stage interference signal identification model is narrowband FM interference, repeater-type spoofing interference, or multipath interference.

[0079] The lightweight multi-scale network constructed in this embodiment adopts the MMDC-CSPDarknet neural network, which is an improvement obtained by replacing the CSP module with the MMDC-CSP module on the basis of the CSPDarknet network architecture.

[0080] The network structure mainly consists of data augmentation, convolution-batch normalization-mish (CBM) module, MMDC-CSP module, and spatial pyramid pooling (SPP) module.

[0081] The MMDC-CSP module replaces the Bottleneck structure in the CSP module with a Multi-scale Dual Conv (MMDC) sub-module. It introduces convolutional branches with different receptive fields in the same layer and solves the problem of insufficient utilization of multi-scale information through a parallel multi-branch structure with branch residuals. This optimizes signal processing, enhances feature representation capabilities, and improves the recognition accuracy of BeiDou signals with and without interference when features are not obvious. At the same time, a Hybrid Local Channel Attention (MLCA) mechanism is added at the end of the MMDC-CSP module. This mechanism uses the spatial position of the channels to compensate for the information loss that may be caused by the previous segmentation, reduces feature redundancy, highlights the target region, and significantly enhances feature representation capabilities and computational efficiency.

[0082] To address the issue that the mandatory feature segmentation of existing CSP modules may lead to the dilution or loss of some feature information during propagation, the MMDC-CSP module introduces the MMDC module to improve the recognition accuracy of BeiDou B1C and its interference signals when features are not obvious. At the same time, it introduces MLCA at the end to compensate for the information loss that may be caused by segmentation, while also taking into account the goal of lightweight design.

[0083] This invention designs a lightweight multi-scale MMDC submodule, introduces multi-scale features to improve the overall recognition accuracy, especially the detection and recognition capability of low-power interference signals, while also reducing the computational complexity and resource consumption of the model.

[0084] Specifically, the processing flow of the two-stage step-by-step identification strategy in this embodiment is as follows:

[0085] Step 2.1. First, input the time-frequency graph of the signal to be identified into the first-stage interference signal identification model; the first-stage interference signal identification model includes a data augmentation module, a multi-scale feature extraction module, and a classification head module, such as... Figure 2 As shown.

[0086] The overall processing flow of the interference signal identification model in the first stage is as follows:

[0087] First, the data augmentation module normalizes the (time-frequency map) feature image extracted in step 1, and then enhances the data using Mosaic methods such as rotation, scaling, and flipping to increase sample diversity and improve the robustness of the model.

[0088] The multi-scale feature extraction module then employs a five-layer pyramid structure (C1, C2, C3, C4, C5) with downsampling ratios of ×2, ×4, ×8, ×16, and ×32, respectively. As the network depth increases, the feature map size decreases progressively, thereby constructing a multi-scale feature representation covering different receptive fields. For targets with different feature sizes, feature extraction is performed on the signal feature map through feature map output layers of different scales.

[0089] The C1 layer consists of a CBM convolution, which converts the RGB image into a feature map for subsequent processing.

[0090] like Figure 5 As shown, in this embodiment, the CBM module is used for preliminary feature extraction. The CBM module contains a 2D convolution, a batch normalization layer, and a Mish activation function to achieve fast downsampling of the input image and basic feature extraction.

[0091] 2D convolution is a mathematical operation applied to two spatial dimensions: height and width. It involves sliding a kernel or filter across the input data, performing element-wise multiplication, and summing the results to generate a feature map. The 2D convolution in the CBM convolution module uses a 3×3 convolution kernel with a stride of 2 for downsampling, reducing the size of the feature map and increasing the number of channels.

[0092] Batch normalization layers standardize features in each mini-batch of data, making the mean of each feature close to 0 and the variance close to 1 in each mini-batch, thus improving training stability and convergence speed. The Mish activation function is used after the convolutional layers and batch normalization layers to enhance the model's non-linear representation capabilities. The mathematical expression of the Mish activation function is:

[0093] (5)

[0094] The intermediate C1, C2, and C3 layers are the core layers for feature recognition, each consisting of a CBM convolution and an MMDC-CSP module. Since the forced feature segmentation of the original CSP module may lead to the dilution or loss of some feature information during propagation, this invention designs the MMDC-CSP module to address this issue. By utilizing a branching structure, it achieves efficient gradient and feature flow, enhancing the model's multi-scale feature extraction capability while also improving computational efficiency.

[0095] The final C5 layer consists of a CBM convolution, an MMDC-CSP module, and an SPP module. The C5 layer has the strongest semantic information and the largest receptive field, enabling it to recognize overall features.

[0096] like Figure 6 As shown, the SPP module first extracts basic semantics through CBM, then captures multi-scale contextual information through max pooling operations in three parallel branches. The pooling kernel sizes are 5×5, 9×9 and 13×13, and the stride is 1. After the outputs of each branch are concatenated, they are finally fused by convolution through CBM to form the final feature representation.

[0097] Finally, the classification head module is used to map the features extracted by the multi-scale feature extraction module to the classification prediction vector. Specifically, the features are compressed into a fixed-length feature vector by global average pooling, and then the final classification probability and interference signal category number are output through a linear fully connected layer to achieve the first-stage identification and classification of the interference feature image.

[0098] The first-stage identification and classification results were passed through Function processing, The processing procedure is as follows:

[0099] (6)

[0100] If the output interference signal category number The identified interference types are, in order: no interference, continuous wave interference, pulse interference, chirped interference with six different scanning shapes, and band-limited Gaussian noise interference.

[0101] If the output This indicates that the identified interference type is narrowband FM interference, repeater-type spoofing interference, or multipath interference; at this point, proceed to step 2.2, and further identify the interference signal in the second stage based on the fourth-order cumulant diagram.

[0102] Step 2.2. Input the fourth-order cumulative map of the interference signal to be identified into the interference signal identification model in the second stage; the interference signal identification model in the second stage also includes a data augmentation module, a multi-scale feature extraction module, and a classification head module.

[0103] like Figure 7 As shown, compared to the first-stage interference signal identification model, the second-stage interference signal identification model employs a three-layer downsampling factor structure of C1, C2, and C3, corresponding to feature map sizes of ×2, ×4, and ×8, respectively. Feature extraction of the signal is performed through feature map output layers of different scales.

[0104] The C1 layer consists of a CBM convolution; the intermediate C2 layer consists of a CBM convolution and an MMDC-CSP module; and the C3 layer consists of a CBM convolution, an MMDC-CSP module, and an SPP module.

[0105] Finally, the classification head module is used to map the features extracted by the multi-scale feature extraction module to the classification prediction vector, and output the interference signal category number, which corresponds to narrowband FM interference, repeater spoofing interference or multipath interference respectively.

[0106] To balance performance and efficiency, the first-stage interference signal identification model is based on time-frequency graph recognition and adopts a five-layer downsampling factor structure. In contrast, the second-stage interference signal identification model is based on fourth-order cumulant image recognition. Since there are fewer types of interference and the features are easy to distinguish, only a three-layer downsampling factor structure is used, thus saving computational resources.

[0107] like Figure 3 As shown, the processing flow of the MMDC-CSP module in this embodiment is as follows:

[0108] I. The input features are first compressed by 1×1 convolution; then split into two branches by a split operation: one branch is retained as a shortcut connection, and the other branch is fed into n multi-scale dual convolutional MMDC sub-modules.

[0109] like Figure 4As shown, the MMDC submodule uses a multi-scale block (MSBlock) as its framework, introducing multi-scale branches to enhance spatial feature diversity. To address the issue of insufficient utilization of multi-scale information, each scale branch is combined with a dual convolutional module (DC). Through parallel convolutional paths, the feature representation capability is collaboratively improved, information fusion efficiency is optimized, and recognition accuracy is enhanced when strong noise and weak interference signals are not obvious. The MSBlock module processes the data through a layer-by-layer residual addition mechanism across N parallel scales, realizing the hierarchical aggregation of multi-scale contextual information, enhancing the multi-scale feature representation capability, and reducing feature homogenization between branches, thereby improving the overall target detection accuracy and efficiency.

[0110] The processing flow of the MSBlock module is as follows: First, for a given input feature map... Divided into N sub-branches Each sub-branch undergoes feature transformation through a 1×1 convolution. To further address the issue of insufficient multi-scale information utilization, a DC convolution is introduced. By combining 3×3 and 1×1 convolution kernels to process the same input feature mapping channels, information processing and feature extraction are optimized. Simultaneously, group convolution techniques are used to efficiently arrange convolutional filters, reducing computational cost and the number of parameters. A further 1×1 convolution is then applied to enhance features and reduce the number of features. Finally, after processing all scales, a 1×1 convolution is used for fusion, integrating the features from each branch and compressing them to the target dimension.

[0111] The specific process of DC convolution is as follows: First, the M channels of the input features and the N filters required for the output are divided into G groups. Second, within each group, the corresponding input channels are further divided into two parts:

[0112] One part undergoes both 3×3 group convolutions and 1×1 point convolutions, and the outputs of the two are summed to balance local feature extraction and preservation of original information; the other part maintains cross-channel information flow only through 1×1 convolutions.

[0113] Finally, the outputs of all paths are concatenated along the channel dimension to form a complete N-dimensional output feature.

[0114] II. The two features from step I are concatenated in the Concat layer to fuse cross-layer information and alleviate gradient vanishing.

[0115] III. Feature remapping and dimension restoration are completed through 1×1 convolution.

[0116] IV. Finally, MLCA combines information from local and global features, as well as channel and spatial features, to compensate for information loss that may be caused by segmentation, thereby enhancing sensitivity to interference features while maintaining computational efficiency.

[0117] The specific processing flow of the MLCA module is as follows:

[0118] First, the input feature map (C, W, H) undergoes local average pooling to calculate the average value within a local region, preserving spatial details while reducing computational cost. Then, a two-branch structure transforms the input features into a one-dimensional vector for processing. The first branch captures global information; after one-dimensional convolution and rearrangement, it combines with the locally pooled features through addition, fusing global contextual information. The second branch preserves local spatial information, performs one-dimensional convolution and rearrangement, and combines with the original input features through multiplication for feature selection, strengthening the focus on locally useful signal features. Both branches then undergo unpooling to restore the spatial dimension, reconstructing the feature map. Finally, information fusion is performed to generate hybrid attention weights. This attention mechanism achieves efficient cross-channel and spatial domain modeling with low computational overhead, improving recognition performance in complex interference environments.

[0119] Step 3. Train the interference signal recognition models for the first and second stages based on the first and second training datasets respectively, and use the trained interference signal recognition models for the two stages to collaboratively identify various interference signals.

[0120] Using the two image training datasets extracted in step 1, and combining them with the constructed multi-domain feature-based stepwise recognition strategy, the MMDC-CSPDarknet network designed in step 3 is used as the classifier in the strategy for training.

[0121] During training, validation samples are used to evaluate the network and fine-tune its hyperparameters, ultimately training the optimal classifier.

[0122] During the optimization process, the random horizontal flip of the Mosaic data augmentation was set to 0.5, the translation to ±10%, the scaling to ±50%, and the occlusion ratio to 40%. The optimization function SGD with linear decay and an initial learning rate of 0.01 was used, and the loss function was the cross-entropy loss function.

[0123] The formula for calculating the cross-entropy loss function is:

[0124]

[0125] in, It is an indicator, if the sample belong If it is 1, then it is 1; otherwise, it is 0. The model predicts the sample. Category The probability of.

[0126] In addition, to verify the effectiveness of the method of the present invention, the test samples generated in step 1 are input into the trained step-by-step identification classifier obtained in step 3 to verify and predict the performance of the proposed method for BeiDou interference identification.

[0127] Use the test samples generated in step 1 to verify the step-by-step recognition strategy and the network's recognition performance.

[0128] The performance of the lightweight multi-scale MMDC-CSPDarknet network used in this invention in terms of recognition accuracy, computational complexity, and running efficiency is compared with CSPDarknet and traditional neural networks (CNN, VGG16, Resnet-18, Mobilenet-V2, Swin Transformer, Vision Transformer) as shown in Table 1.

[0129] Table 1. Comparison of performance results of different neural networks

[0130]

[0131] In terms of recognition accuracy, MMDC-CSPDarknet achieves the best accuracy and macro-average F1 score of 97.5%, which is 2.19% higher than the second-best Swin Transformer network. Regarding computational cost, it has the lowest FLOPS at only 2.9G. Although its parameter count is slightly higher than CNN and Mobilenet-V2, both of which have poor recognition accuracy. In terms of efficiency, MMDC-CSPDarknet has the lowest inference time of only 1.7ms, but its FPS performance is not as good as networks like VGG16. It is difficult for a model to simultaneously achieve optimal recognition accuracy, computational cost, and efficiency. This invention prioritizes high recognition accuracy and low resource consumption, and the proposed MMDC-CSPDarknet offers the best overall performance.

[0132] Figure 8 The results show a comparison of the recognition accuracy of this invention with other neural networks at different interference-to-signal ratios. Figure 8 The horizontal axis represents the interference signal power for different JSRs, and the vertical axis represents the accuracy of each network at different interference power levels. Through comparison, this invention achieves the highest accuracy across almost all JSRs, and also performs well with low JSRs. When JSR = 16dB, the accuracy reaches 85.45%, which is 9.85% higher than the second-best CSPDarknet and 29.69% higher than Swin Transformer. This indicates that the multi-scale feature dynamic enhancement introduced by the MMDC submodule can improve the sensitivity to weak interference features.

[0133] Example 2

[0134] This embodiment 2 describes a BeiDou multi-domain feature interference signal identification system, which is based on the same inventive concept as the BeiDou multi-domain feature interference signal identification method in embodiment 1 above.

[0135] The system for identifying BeiDou multi-domain feature interference signals in this embodiment includes the following modules:

[0136] The preprocessing module is used to generate interference-free BeiDou signal samples and interference signal samples with different interference signals added. There are 12 types of interference signals added, including continuous wave interference, pulse interference, narrowband frequency modulation interference, band-limited Gaussian noise interference, and chirp interference with 6 different scanning shapes, repeater spoofing interference, and multipath interference.

[0137] For both interference-free signal samples and all interference signal samples, extract the corresponding time-frequency plots and construct the first training dataset;

[0138] Meanwhile, for interference signal samples with narrowband FM interference, repeater spoofing interference and multipath interference, the corresponding fourth-order cumulant map is extracted and a second training dataset is constructed.

[0139] And an interference signal identification module, used to construct a two-stage step-by-step identification strategy that combines time-frequency graphs and fourth-order cumulative graphs, with an interference signal identification model based on a lightweight multi-scale network built in each stage.

[0140] The first stage of the interference signal identification model takes the time-frequency diagram of the signal to be identified as input and outputs any one of the following: no interference, continuous wave interference, pulse interference, band-limited Gaussian noise interference, chirp interference with six different scanning shapes, and interference types that cannot be identified. For interference signals that cannot be identified in the first stage, they are further input into the second stage for identification.

[0141] The input to the second-stage interference signal identification model is the fourth-order cumulative graph of the interference signal to be identified, and the output is narrowband FM interference, repeater spoofing interference, or multipath interference.

[0142] The interference signal recognition models for the first and second stages are trained based on the first and second training datasets, respectively, and the trained interference signal recognition models for the two stages are used to collaboratively identify various interference signals.

[0143] It should be noted that any content not mentioned in the above-described functional modules of the BeiDou multi-domain feature interference signal identification system described in this embodiment can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be elaborated here.

[0144] Example 3

[0145] This embodiment 3 describes a computer device, which includes a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the BeiDou multi-domain feature interference signal identification method in embodiment 1 above.

[0146] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0147] Example 4

[0148] This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the BeiDou multi-domain feature interference signal identification method in embodiment 1 above.

[0149] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0150] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for identifying multi-domain characteristic interference signals in BeiDou navigation, characterized in that, Includes the following steps: Step 1. Generate interference-free BeiDou signal samples and interference signal samples with different interference signals added respectively; The added interference signals include 12 types, including continuous wave interference, pulse interference, narrowband FM interference, band-limited Gaussian noise interference, and chirp interference with 6 different scanning shapes, repeater spoofing interference, and multipath interference. For both interference-free signal samples and all interference signal samples, extract the corresponding time-frequency plots and construct the first training dataset; Meanwhile, for interference signal samples with narrowband FM interference, repeater spoofing interference and multipath interference, the corresponding fourth-order cumulant map is extracted and a second training dataset is constructed. Step 2. Construct a two-stage step-by-step identification strategy that combines time-frequency plots and fourth-order cumulant plots. In each stage, an interference signal identification model based on a lightweight multi-scale network is built. The first-stage interference signal identification model takes the time-frequency diagram of the signal to be identified as input and outputs any one of the following: no interference, continuous wave interference, pulse interference, band-limited Gaussian noise interference, chirp interference with six different scanning shapes, and interference types that cannot be identified. For interference signals that cannot be identified in the first stage, they are further input into the second stage for identification. The input to the second-stage interference signal identification model is the fourth-order cumulative graph of the interference signal to be identified, and the output is narrowband FM interference, repeater spoofing interference, or multipath interference. The lightweight multi-scale network adopts the MMDC-CSPDarknet neural network, which is an improvement on the CSPDarknet network architecture by replacing the CSP module with the MMDC-CSP module. The MMDC-CSP module design replaces the Bottleneck structure in the CSP module with a multi-scale dual-convolutional MMDC submodule to introduce convolutional branches with different receptive fields in the same layer, and improves feature representation capability through the parallel multi-branch structure of branch residuals. Meanwhile, a hybrid local channel attention mechanism (MLCA) is added to the end of the MMDC-CSP module. MLCA uses channel spatial location to compensate for the information loss that may be caused by the previous segmentation, so as to reduce feature redundancy and highlight the target region. The processing flow of the MMDC-CSP module is as follows: I. The input features are first compressed by 1×1 convolution; then split into two branches by a split operation: one branch is retained as a shortcut connection, and the other branch is fed into n multi-scale dual convolutional MMDC sub-modules; The MMDC submodule uses the multi-scale block MSBlock as its framework and introduces multi-scale branches to enhance the diversity of spatial features. To address the insufficient utilization of multi-scale information, each scale branch is combined with the dual convolution module Dual Conv, which collaboratively enhances feature representation capabilities through parallel convolution paths, thereby improving recognition accuracy when strong noise and weak interference signal features are not obvious. II. The two features are concatenated in the Concat layer to fuse cross-layer information and alleviate gradient vanishing. III. Feature remapping and dimensionality restoration are completed through 1×1 convolution; IV. Finally, the MLCA module combines information from local and global features as well as channel and spatial features; Step 3. Train the interference signal recognition models for the first and second stages based on the first and second training datasets respectively, and use the trained interference signal recognition models for the two stages to collaboratively identify various interference signals.

2. The method for identifying BeiDou multi-domain feature interference signals according to claim 1, characterized in that, In step 1, based on the different scanning shapes and scanning frequencies of the chirp signals, the chirp signals are classified into sinusoidal chirp interference, sawtooth chirp interference, hook-shaped sawtooth chirp interference, triangular chirp interference, linear chirp interference, and frequency hopping chirp interference.

3. The method for identifying BeiDou multi-domain feature interference signals according to claim 1, characterized in that, In step 1, STFT is used as a time-frequency analysis tool to perform time-frequency transformation on the interference-free signal samples and all interference signal samples in step 1, and the time-frequency diagram is calculated. The specific process is as follows: A sliding window function is selected and shifted across the entire time axis of the signal. A Fourier transform is performed on each windowed signal segment. The square of the magnitude of the transform result is then taken to obtain the energy distribution in the time-frequency domain. ; ； in, This indicates the observed signal received by the receiver. It is a length of Window function, It is the step size of the window function at each slide, and its value is equal to the size of the window minus the overlap length. ,Right now ; The energy distribution at different times is arranged in chronological order to form a time-frequency diagram, where the horizontal axis represents time and the vertical axis represents frequency. Different shades of color indicate the power of the signal at different times and frequencies.

4. The method for identifying BeiDou multi-domain feature interference signals according to claim 1, characterized in that, In step 1, the process of obtaining the fourth-order cumulative quantity map is as follows: First, the easily confused signals entering the second stage, namely narrowband FM interference, repeater-type spoofing interference, and multipath interference samples, are centered (i.e., zero-mean processed). Then, their diagonal slices are selected for cumulative estimation, and the formula expression is as follows: ； in, It is a fourth-order cumulant. The time delay is represented by a sliding delay window, which is used to calculate the fourth-order cumulative characteristic of the signal under different delays. Indicates the length of the overlapping portion. Indicates the total number of signal sampling points; The signal representing the received signal after zero-mean processing is calculated as follows: ； based on Construct two offset signal sequences and ; Ultimately The x-axis represents the corresponding fourth-order cumulants. The fourth-order cumulant plot of the signal is obtained by using the vertical axis.

5. The method for identifying BeiDou multi-domain feature interference signals according to claim 1, characterized in that, In step 2, the processing flow of the two-stage step-by-step identification strategy is as follows: Step 2.

1. First, input the time-frequency graph of the signal to be identified into the first-stage interference signal identification model; the first-stage interference signal identification model includes a data augmentation module, a multi-scale feature extraction module, and a classification head module; First, the data is augmented using a data augmentation module to enhance sample diversity; Then, the multi-scale feature extraction module adopts a five-layer downsampling factor structure of C1, C2, C3, C4, and C5, and extracts features from the signal's feature map through feature map output layers of different scales; The C1 layer consists of a CBM convolution; The intermediate C1, C2, and C3 layers each consist of a CBM convolution and an MMDC-CSP module; the C5 layer consists of a CBM convolution, an SPP module, and an MMDC-CSP module. Finally, the classification head module is used to map the features extracted by the multi-scale feature extraction module to the classification prediction vector and output the interference signal category number to achieve the first-stage identification and classification of the interference feature image. The first-stage identification and classification results were passed through Function processing, The processing procedure is as follows: ； If the output interference signal category number The identified interference types are, in order: no interference, continuous wave interference, pulse interference, chirped interference with six different scanning shapes, and band-limited Gaussian noise interference. If the output This indicates that the identified interference type is narrowband FM interference, repeater-type spoofing interference, or multipath interference; at this point, proceed to step 2.2, and further identify the interference signal in the second stage based on the fourth-order cumulant diagram; Step 2.

2. Input the fourth-order cumulative map of the interference signal to be identified into the interference signal identification model in the second stage; the interference signal identification model in the second stage also includes a data augmentation module, a multi-scale feature extraction module, and a classification head module; In the second stage of the interference signal identification model, the multi-scale feature extraction module adopts a three-layer downsampling multiple structure of C1, C2 and C3, and extracts features from the feature map of the signal through feature map output layers of different scales; The C1 layer consists of a CBM convolution; the intermediate C2 layer consists of a CBM convolution and an MMDC-CSP module; the C3 layer consists of a CBM convolution, an SPP module, and an MMDC-CSP module. Finally, the classification head module is used to map the features extracted by the multi-scale feature extraction module to the classification prediction vector, and output the interference signal category number, which corresponds to narrowband FM interference, repeater spoofing interference or multipath interference respectively.

6. A BeiDou multi-domain feature interference signal identification system, used to implement the BeiDou multi-domain feature interference signal identification method as described in any one of claims 1 to 5, characterized in that, The BeiDou multi-domain feature interference signal identification system includes the following modules: The preprocessing module is used to generate interference-free BeiDou signal samples and interference signal samples with different interference signals added. There are 12 types of interference signals added, including continuous wave interference, pulse interference, narrowband frequency modulation interference, band-limited Gaussian noise interference, and chirp interference with 6 different scanning shapes, repeater spoofing interference, and multipath interference. For both interference-free signal samples and all interference signal samples, extract the corresponding time-frequency plots and construct the first training dataset; Meanwhile, for interference signal samples with narrowband FM interference, repeater spoofing interference and multipath interference, the corresponding fourth-order cumulant map is extracted and a second training dataset is constructed. And an interference signal identification module, used to construct a two-stage step-by-step identification strategy that combines time-frequency graphs and fourth-order cumulative graphs, with an interference signal identification model based on a lightweight multi-scale network built in each stage. The first-stage interference signal identification model takes the time-frequency diagram of the signal to be identified as input and outputs any one of the following: no interference, continuous wave interference, pulse interference, band-limited Gaussian noise interference, chirp interference with six different scanning shapes, and interference types that cannot be identified. For interference signals that cannot be identified in the first stage, they are further input into the second stage for identification. The input to the second-stage interference signal identification model is the fourth-order cumulative graph of the interference signal to be identified, and the output is narrowband FM interference, repeater spoofing interference, or multipath interference. The interference signal recognition models for the first and second stages are trained based on the first and second training datasets, respectively, and the trained interference signal recognition models for the two stages are used to collaboratively identify various interference signals.

7. A computer device, comprising a memory and one or more processors; characterized in that, The memory stores executable code, which, when executed by the processor, is used to implement the steps of the BeiDou multi-domain feature interference signal identification method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a program stored thereon; characterized in that, When executed by the processor, the program is used to implement the steps of the BeiDou multi-domain feature interference signal identification method according to any one of claims 1 to 5.

Images