A method for identifying frequency modulation fuze jamming signals based on a lightweight network

Abstract

This invention discloses a method for identifying frequency modulation (FM) fuze interference signals based on a lightweight network, belonging to the field of radio fuze technology. The method comprises the following steps: S1, simulating a triangular wave FM fuze system containing interference sources, acquiring target echo signals and interference signals, generating simulated signals, and constructing a dataset; S2, performing time-frequency transformation on the simulated signals to generate and process a time-frequency graph; S3, extracting high-order statistical features of the simulated signals and constructing a multi-dimensional statistical feature vector; S4, constructing a lightweight dual-branch network to extract depth features and statistical features from the time-frequency graph; S5, fusing the depth features and statistical features, and inputting the fused features into the network model for training. After training, the interference types of the test set signals are identified, and the model performance is verified. This method for identifying FM fuze interference signals based on a lightweight network can identify and detect various interference signals, exhibiting advantages such as high recognition rate, strong robustness, and lightweight model.

Description

Technical Field

[0001] This invention relates to the field of radio fuze technology, and in particular to a method for identifying frequency modulation fuze interference signals based on a lightweight network. Background Technology

[0002] Frequency modulation (FM) fuses are proximity fuses based on the principle of frequency modulation continuous wave ranging. They have the advantages of high ranging accuracy and strong anti-interference capability and have been widely used in conventional munitions and guided weapons. Symmetrical triangular wave modulation is one of the most mainstream modulation methods for FM fuses.

[0003] In complex electromagnetic environments, frequency-modulated (FM) fuses face various interference threats, which can be categorized into three main types based on their mechanisms and effects: First, suppression interference, including noise amplitude modulation (AM) and noise frequency modulation (FM) interference, which uses strong noise to cover the target echo signal; second, DRFM (Dual-Reference-Frequency-Frequency) relay interference, including intermittent sampling and direct relay and intermittent sampling and repeated relay interference, which generates coherent false targets by sampling and relaying the fuse signal; and third, deception interference, including range deception and dense false target interference, which misleads the fuse's operation by falsifying target information. Interference signal identification is the core prerequisite for FM fuse anti-interference technology and a crucial first step in its stable operation in complex electromagnetic environments. Therefore, efficient and accurate interference identification methods are of significant research value.

[0004] Existing interference identification methods mostly rely on manual extraction of time-frequency domain features, such as support vector machine-based identification schemes that extract features through the peak-to-peak ratio of the spectrum. However, these methods have significant drawbacks: first, in low interference-to-signal ratio scenarios, the representational power of manual features is greatly reduced, resulting in a significant decrease in identification accuracy; second, feature extraction is highly dependent on expert experience, lacking generalization ability and robustness, making it difficult to adapt to diverse and complex interference scenarios.

[0005] In recent years, deep learning technology has demonstrated tremendous potential in the field of signal recognition. To address the need for interference identification, algorithms must simultaneously meet both high accuracy and high real-time performance; therefore, lightweight models have become a core research direction. Existing lightweight deep learning models (such as MobileNet, GhostNet, and ShuffleNet) are all image-based classification algorithms that rely solely on time-frequency graph features for interference identification. However, these models have significant limitations: for interference types with highly similar time-frequency features, such as noise amplitude modulation and noise frequency modulation, accurate differentiation using only a single time-frequency graph feature is difficult, and the recognition accuracy and robustness still cannot meet the needs of practical applications. Summary of the Invention

[0006] The purpose of this invention is to provide a method for identifying frequency modulation fuze interference signals based on a lightweight network, thereby solving the aforementioned technical problems.

[0007] To achieve the above objectives, this invention provides a method for identifying frequency modulation fuze interference signals based on lightweight networks, comprising the following steps: S1. Simulate the triangular wave frequency modulation fuze system containing interference sources, collect the target echo signal and interference signal, generate simulation signals and construct a dataset; S2. Perform time-frequency transformation on the simulation signal, generate a time-frequency graph, and process it. S3. Extract higher-order statistical features from the simulation signal and construct a multidimensional statistical feature vector; S4. Construct a lightweight dual-branch network to extract the deep and statistical features of the time-frequency graph; S5. Fuse deep features and statistical features, and input the fused features into the network model for training. After training, identify the interference types of the test set signals and verify the model performance.

[0008] Preferably, in S1: Expression for the transmitted signal of a triangular wave frequency modulation fuze As shown below: ; In the formula, t For time variables, n This is the signal modulation period number. The amplitude of the transmitted signal. For the carrier frequency of the transmitted signal, For modulation bandwidth, For the modulation period, and These are the initial phases of the rising and falling edges of the triangular wave, respectively; After the fuze's emitted signal reaches the target, it is reflected by the target, generating a target echo signal. The expression for the target echo signal is... S r ( t As shown below: ; In the formula, The target echo signal amplitude, To delay time, The distance between the fuse and the target. The speed of light; After being subjected to human interference and environmental noise in the battlefield environment, the signal received by the fuse receiver As shown below: ; In the formula, This is an interference signal. It is additive white Gaussian noise.

[0009] Preferably, the interference signals in S1 include noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, range deception interference, and dense false target interference. These six types of interference signals, together with the target echo signal, constitute the dataset.

[0010] Preferably, based on the interference-to-signal ratio, a consistent number of simulated signals are generated for each type of signal in the dataset, and then all the simulated signals are divided into training set, validation set and test set in a 4:1:1 ratio.

[0011] Preferably, in S2, the time-frequency transformation adopts short-time Fourier transform, the window function is selected as Hamming window, the window length is 256 points, the number of overlapping points is 200 points, the time-frequency image is extracted after transformation and logarithmically compressed, and the size of the time-frequency image is adjusted to 224×224 pixels.

[0012] Preferably, the higher-order statistical features in S3 are 11-dimensional, including mean amplitude, variance, root mean square, peak-to-peak value, skewness, kurtosis, mean instantaneous frequency, variance instantaneous frequency, spectral entropy, spectral centroid, and spectral spread. After standardizing the 11-dimensional features, they are spliced ​​together in a fixed order to construct a multi-dimensional statistical feature vector.

[0013] Preferably, in S4, the first branch of the lightweight dual-branch network uses MobileNetV3-small to extract the deep feature vector of the time-frequency map, and the second branch uses a multilayer perceptron to extract the statistical feature vector. The output of the lightweight dual-branch network uses the Softmax function to convert the network output into a probability distribution.

[0014] Preferably, the structure of the multilayer perceptron includes a first linear layer and a second linear layer. The first linear layer maps 11-dimensional statistical features to 128-dimensional features. After normalization, ReLU activation, and Dropout processing, the second linear layer maps to 64-dimensional features. After batch normalization and ReLU activation, it outputs 64-dimensional statistical features.

[0015] Preferably, the AdamW optimizer is used for network training in S5, with an initial learning rate of [missing value]. α The weight decays to β At the same time, a cosine annealing learning rate scheduling strategy is adopted.

[0016] Preferably, S5 employs an early stopping strategy for network training, where the validation set accuracy is continuously... S Training stops when there is no improvement after a certain number of epochs, and the model with the highest accuracy on the validation set is saved as the optimal model.

[0017] Therefore, the present invention employs the above-mentioned method for identifying frequency modulation fuze interference signals based on lightweight networks, which has the following beneficial effects: 1. To address the problem that traditional methods have difficulty distinguishing between amplitude modulation (AM) noise interference and frequency modulation (FM) noise interference, this paper introduces higher-order statistical features including skewness and kurtosis to effectively distinguish between the two types of interference.

[0018] 2. The model employs a dual-branch network structure that integrates high-order statistical features and deep features. Even under a low interference-to-information ratio of -5dB, the average recognition accuracy is still greater than 95%, demonstrating excellent model robustness.

[0019] 3. It adopts a lightweight network with MobileNetV3-small as the backbone, with only 2.5M model parameters, low computational load, and can be deployed on embedded platforms such as FPGA to meet the real-time processing requirements of frequency modulation fuses.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for identifying frequency modulation fuze interference signals based on a lightweight network according to the present invention. Figure 2 This is a diagram of the lightweight dual-branch feature fusion architecture of the present invention; Figure 3 The recognition accuracy curves of the present invention under different interference-to-signal ratios; Figure 4 The above is the overall normalized confusion matrix diagram of the present invention, wherein the target echo signal is denoted as 1, the noise amplitude modulation interference is denoted as 201, the noise frequency modulation interference is denoted as 202, the intermittent sampling direct forwarding interference is denoted as 203, the intermittent sampling repeated forwarding interference is denoted as 204, the range spoofing interference is denoted as 205, and the dense false target interference is denoted as 206. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0023] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] like Figures 1-4 As shown, this invention provides a method for identifying interference signals in frequency-modulated (FM) fuses based on lightweight networks, meeting the industrial application requirements of real-time performance and lightweight design for FM fuses. The method includes the following steps: S1, simulating a triangular wave FM fuse system containing interference sources, acquiring target echo signals and interference signals, generating simulated signals, and constructing a dataset; S2, performing time-frequency transformation on the acquired signals, generating a time-frequency graph, and preprocessing it; S3, extracting high-order statistical features of the simulated signals, and constructing a multi-dimensional statistical feature vector; S4, constructing a lightweight dual-branch network, extracting deep features and statistical features from the time-frequency graph; S5, fusing the deep features and statistical features, and inputting the fused features into the network model for training. After training, identifying the interference type of the test set signals and verifying the model performance.

[0026] Specifically, in S1: Expression for the transmitted signal of a triangular wave frequency modulation fuze As shown below: ; In the formula, t For time variables, n This is the signal modulation period number. The amplitude of the transmitted signal. For the carrier frequency of the transmitted signal, For modulation bandwidth, For the modulation period, and These represent the initial phases of the rising and falling edges of the triangular wave, respectively; after the fuze's transmitted signal reaches the target, it is reflected by the target, generating a target echo signal, the expression for which is given. S r ( t As shown below: ; In the formula, The target echo signal amplitude, To delay time, The distance between the fuse and the target. The speed of light; after passing through human interference and environmental noise in the battlefield environment, the signal received by the fuse receiver As shown below: ; In the formula, This is an interference signal. It is additive white Gaussian noise.

[0027] In addition, the interference signals in S1 include: noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, range spoofing interference, and dense false target interference. These six types of interference signals, together with the target echo signal, constitute the dataset. The signals in the dataset are divided into simulated signals every 3dB from -5dB to 10dB based on their interference-to-signal ratio (ISR), with 480 simulated signals generated for each type. Therefore, the total number of simulated signals is 20160. Based on the ISR of the simulated signals, all generated simulated signals are divided into a 4:1:1 dataset. The training set includes 13440 simulated signals, the validation set includes 3360 simulated signals, and the test set includes 3360 simulated signals. This hierarchical partitioning method ensures a balanced distribution of samples under different ISR ratios, effectively improving the model's generalization ability and recognition stability in low ISR scenarios.

[0028] In S2, the acquired signal is processed by time-frequency transformation, which uses short-time Fourier transform with a Hamming window as the window function. The window length is 256 points and the number of overlapping points is 200. After transformation, the time-frequency image is extracted and logarithmically compressed. The size of the time-frequency image is then adjusted to 224×224 pixels. This processing can preserve the time-frequency details of the signal while adapting to the input format requirements of the MobileNetV3-small network, ensuring the consistency and effectiveness of deep feature extraction.

[0029] The higher-order statistical features in S3 are 11-dimensional, including mean amplitude, variance, root mean square, peak-to-peak value, skewness, kurtosis, mean instantaneous frequency, variance instantaneous frequency, spectral entropy, spectral centroid, and spectral spread. After standardizing the 11-dimensional features, they are spliced ​​together in a fixed order to construct a multi-dimensional statistical feature vector. Among them, the skewness and kurtosis features can accurately distinguish between noise amplitude modulation interference and noise frequency modulation interference, which are difficult to identify by traditional methods, thus making up for the identification defects of single time-frequency plot features from a statistical dimension.

[0030] The lightweight dual-branch network structure in S4 employs a parallel processing mode, significantly improving signal processing efficiency. The first branch of the lightweight dual-branch network uses MobileNetV3-small to extract deep feature vectors from the time-frequency map. This lightweight backbone network has only 2.5M parameters, resulting in low computational cost and excellent feature extraction capabilities. The second branch uses a multilayer perceptron to extract statistical feature vectors. The output of the lightweight dual-branch network uses the Softmax function to convert the network output into a probability distribution, providing intuitive and reliable classification results. The multilayer perceptron structure includes a first linear layer and a second linear layer. The first linear layer maps 11-dimensional statistical features to 128 dimensions. After normalization, ReLU activation, and Dropout processing, the second linear layer maps to 64 dimensions. After batch normalization and ReLU activation, it outputs 64-dimensional statistical features, forming complementary features with dimensionally adapted deep features.

[0031] In S5, the network training uses the AdamW optimizer with an initial learning rate of [missing information]. α Set to 0.001, weight decay is... β The learning rate is set to 1e-4, and a cosine annealing learning rate scheduling strategy is employed. This optimizes model convergence speed and training stability. The network training uses an early stopping strategy: training stops when the validation set accuracy does not improve for 15 consecutive epochs, and the model with the highest validation set accuracy is saved as the optimal model, avoiding overtraining that could lead to a decline in generalization ability. (See attached...) Figure 3 As shown in the accuracy curves for different interference-to-signal ratios, the overall accuracy of this method reaches 98.21%. Even under extremely low interference-to-signal ratio conditions of around -5dB, it still maintains a high accuracy of 96.43%, demonstrating robustness far superior to traditional interference identification methods. (Combined with the attached...) Figure 4 The normalized confusion matrix shown can further verify that this method can achieve 100% accurate identification of noise frequency modulation interference, intermittent sampling direct forwarding interference, range spoofing interference, and dense false target interference, with only a very small amount of category confusion. It completely solves the industry pain point of difficulty in distinguishing similar interference. At the same time, the lightweight characteristics of the model can meet the deployment requirements of embedded platforms such as FPGA, and are fully compatible with the industrial application requirements of real-time anti-interference of frequency modulation fuses.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying frequency modulation fuze interference signals based on lightweight networks, characterized in that, Includes the following steps: S1. Simulate the triangular wave frequency modulation fuze system containing interference sources, collect the target echo signal and interference signal, generate simulation signals and construct a dataset; S2. Perform time-frequency transformation on the simulation signal, generate a time-frequency graph, and process it. S3. Extract higher-order statistical features of the simulation signal and construct a multidimensional statistical feature vector; S4. Construct a lightweight dual-branch network to extract the deep and statistical features of the time-frequency graph; S5. Fuse deep features and statistical features, and input the fused features into the network model for training. After training, identify the interference types of the test set signals and verify the model performance.

2. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 1, characterized in that, In S1: Expression for the transmitted signal of a triangular wave frequency modulation fuze As shown below: ； In the formula, t For time variables, n This is the signal modulation period number. The amplitude of the transmitted signal. For the carrier frequency of the transmitted signal, For modulation bandwidth, For the modulation period, and These are the initial phases of the rising and falling edges of the triangular wave, respectively; After the fuze's emitted signal reaches the target, it is reflected by the target, generating a target echo signal. The expression for the target echo signal is... S r ( t As shown below: ； In the formula, The target echo signal amplitude, To delay time, The distance between the fuse and the target. The speed of light; After being subjected to human interference and environmental noise in the battlefield environment, the signal received by the fuse receiver As shown below: ； In the formula, This is an interference signal. It is additive white Gaussian noise.

3. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 2, characterized in that: The interference signals in S1 include noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, range deception interference, and dense false target interference. These six types of interference signals, together with the target echo signal, constitute the dataset.

4. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 3, characterized in that: Based on the interference-to-signal ratio, a consistent number of simulated signals are generated for each type of signal in the dataset. Then, all the simulated signals are divided into training set, validation set, and test set in a 4:1:1 ratio.

5. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 4, characterized in that: In S2, the time-frequency transformation adopts short-time Fourier transform, the window function is Hamming window, the window length is 256 points, the number of overlapping points is 200 points, the time-frequency image is extracted after transformation and logarithmically compressed, and the size of the time-frequency image is adjusted to 224×224 pixels.

6. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 5, characterized in that: The higher-order statistical features in S3 are 11-dimensional, including mean amplitude, variance, root mean square, peak-to-peak value, skewness, kurtosis, mean instantaneous frequency, variance instantaneous frequency, spectral entropy, spectral centroid, and spectral spread. After standardizing the 11-dimensional features, they are concatenated in a fixed order to construct a multi-dimensional statistical feature vector.

7. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 6, characterized in that: The first branch of the lightweight dual-branch network in S4 uses MobileNetV3-small to extract the deep feature vector of the time-frequency map, and the second branch uses a multilayer perceptron to extract the statistical feature vector. The output of the lightweight dual-branch network uses the Softmax function to convert the network output into a probability distribution.

8. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 7, characterized in that: The structure of a multilayer perceptron includes a first linear layer and a second linear layer. The first linear layer maps 11-dimensional statistical features to 128-dimensional features. After normalization, ReLU activation, and Dropout processing, the second linear layer maps to 64-dimensional features. After batch normalization and ReLU activation, it outputs 64-dimensional statistical features.

9. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 8, characterized in that: In S5, the network training uses the AdamW optimizer, with an initial learning rate of [value missing]. α The weight decays to β At the same time, a cosine annealing learning rate scheduling strategy is adopted.

10. The method for identifying frequency modulation fuze interference signals based on a lightweight network according to claim 9, characterized in that: S5 employs an early stopping strategy for network training, where the validation set accuracy is continuously improved. S Training stops when there is no improvement after a certain number of epochs, and the model with the highest accuracy on the validation set is saved as the optimal model.

Images