A neural network interference identification method based on Conv-KANformer

Abstract

The application belongs to the field of interference identification, and particularly relates to a neural network interference identification method based on Conv-KANformer, which comprises the following steps: S1: constructing an interference database, adopting short-time Fourier transform to extract an interference time-frequency graph; performing preprocessing operations such as clipping, scaling and label information construction on the time-frequency graph to ensure the accuracy and consistency of the data; S2: building a Conv-KANformer neural network framework, which is composed of a multi-scale convolution feature extraction module, a KANformer global feature extraction module and a classification module, wherein the multi-scale convolution feature extraction module contains four convolution kernels of different sizes to extract different receptive field features; the KANformer global feature extraction module introduces a Kolmogorov-Arnold Network (KAN) to improve the traditional transformer structure; and the classification module introduces the KAN to complete the classification of seven kinds of interference signals; S3: completing offline training of the model through a cross-entropy loss function to obtain an interference identification model; and S4: completing online interference identification by using the interference identification model. The neural network interference identification method based on Conv-Kanformer can identify seven kinds of interference types at a low signal-to-noise ratio.

Description

Technical Field

[0001] This invention belongs to the field of interference recognition and relates to a neural network interference recognition method based on Conv-KANformer. Background Technology

[0002] In today's information-explosive era, wireless communication technology stands at the forefront of rapid technological development. From 5G to 6G, and to the many more exciting mobile communication technologies to come, each leap forward breaks down technological boundaries, leading us into a new era of communication.

[0003] However, with the continuous development of wireless communication technology, we also face increasing challenges. The growing complexity of the electromagnetic environment severely tests the stability and reliability of wireless communication. In the civilian communications field, interference signals can lead to problems such as degraded communication quality, reduced communication efficiency, and threats to communication security, causing great inconvenience to the user experience. In the military communications field, interference signals can lead to serious consequences such as communication delays and interruptions, information leaks, and damage to communication systems, directly affecting the success or failure of operations and the safety of the military.

[0004] In today's information warfare and digital society, communication anti-jamming technology is particularly important to ensure ultra-reliable information transmission. Especially in the field of military communications, efficient and feasible anti-jamming technology is key to improving communication anti-jamming capabilities. The field of communication anti-jamming typically encompasses multiple aspects, including interference recognition, interference suppression, anti-jamming decision-making, and interference avoidance. Interference identification technology plays a crucial role in the anti-jamming field, and its importance is self-evident. First, interference identification is the prerequisite and foundation of anti-jamming technology. In communication systems, whether civilian or military, interference is a significant factor affecting communication quality and stability. Through interference identification technology, we can accurately detect the source, type, and key information of interference signals, providing a strong basis for subsequent anti-jamming measures. Second, interference identification is of great significance for improving the security of communication systems. In the field of military communications, the enemy may use various means to interfere with communications in order to disrupt our communication systems or steal important information. Through interference identification technology, we can promptly detect and respond to these intentionally created interferences, protecting the confidentiality and integrity of military communications. In the field of civilian communications, interference identification also plays an important role. By identifying and processing malicious interference signals, we can prevent malicious attacks and information theft, protecting user privacy and data security. Finally, with the continuous advancement and improvement of interference identification technology, the anti-interference capability of communication systems has been significantly enhanced. These technologies can not only more accurately identify interference signals, but also achieve rapid response and effective suppression of interference signals through intelligent algorithms and adaptive techniques.

[0005] Traditional interference identification methods do indeed have many shortcomings, such as reliance on signal processing and feature extraction techniques, the need for extensive expert knowledge, poor algorithm generalization ability, low identification accuracy, and difficulty in classification. However, with the rapid development of deep learning, these problems are being gradually solved. Deep learning, especially Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), can automatically learn and extract deep feature representations from raw data without relying on manually designed feature engineering. This capability gives deep learning a significant advantage in handling complex, high-dimensional, and nonlinear problems. Through training on large amounts of data, deep learning models can learn the inherent laws and patterns of the data, thus possessing powerful generalization capabilities. This means that the model can exhibit good performance on new and unseen data, adapting to various practical application scenarios. In the field of interference identification, deep learning, through its powerful learning capabilities, automatically extracts deep features of interference information, enabling rapid reasoning and judgment, and thus achieving real-time detection and identification of interference signals. This helps improve the anti-interference performance of communication systems, benefiting both civilian and military communications. Summary of the Invention

[0006] To address the above technical problems, this invention provides a neural network interference identification method based on Conv-KANformer. This invention primarily solves the problem of identifying typical interference signals, achieving interference signal identification even at low interference-to-noise ratios.

[0007] This invention first constructs an interference signal database, divided into a training set, a validation set, and a test set. The interference types include seven categories: single-tone interference, multi-tone interference, linear sweep interference, impulse interference, partial-band noise interference, sinusoidal frequency modulation interference, and noise-frequency modulation interference. A basic neural network is trained based on the training set, using the Conv-KANformer neural network. During testing, the test data is sensed online, and the trained interference recognition model is used to determine the interference type of the test data in real time.

[0008] The solution proposed in this invention includes the following steps:

[0009] S1: Construct an interference database and extract interference time-frequency maps using short-time Fourier transform; perform preprocessing operations such as cropping, scaling, and labeling on the time-frequency maps to ensure that the pixel size of the time-frequency maps is 3x128x128, ensuring data accuracy and consistency and improving training speed; interference types may include, but are not limited to, seven types of interference: single-tone interference, multi-tone interference, linear sweep interference, impulse interference, partial frequency band noise interference, sine wave frequency modulation interference, and noise frequency modulation interference;

[0010] S2: Conv-KANformer neural network framework is built. This network consists of a multi-scale convolutional feature extraction module, a KANformer global feature extraction module, and a classification module. The multi-scale convolutional feature extraction module contains four layers of convolutional kernels of different sizes to extract features of different receptive fields. The KANformer global feature extraction module introduces KAN to improve the traditional transformer structure. The classification module introduces KAN to complete the classification of 7 kinds of interference signals.

[0011] S3: The model is trained offline using the cross-entropy loss function to obtain the interference recognition model;

[0012] S4: Use the interference identification model to complete online interference identification.

[0013] According to the present invention, step S1 above includes the following steps:

[0014] An interference database was constructed from samples of seven types of interference signals (single-tone interference, multi-tone interference, linear sweep interference, pulse interference, partial-band noise interference, sinusoidal FM interference, and noise FM interference). Sample information consisted of input samples and classification labels. The sampling frequency was 50MHz, with 1024 sampling points. Under an interference-to-noise ratio (INR) range of -10dB to 20dB and a step size of 2dB, 100 training samples were generated for each type of interference signal at each INR, totaling 11200 samples. The dataset was partitioned at a 4:1 ratio. The training set consisted of 8960 samples, and the validation set consisted of 2240 samples.

[0015] According to the present invention, step S2 above includes the following steps:

[0016] A Conv-KANformer neural network framework was constructed, consisting of a multi-scale convolutional feature extraction module, a KANformer global feature extraction module, and a classification module. The multi-scale convolutional feature extraction module contains four layers of convolutional kernels of different sizes to extract features from different receptive fields. The KANformer global feature extraction module incorporates Kolmogorov-Arnold Networks (KAN) to improve the traditional transformer structure. The classification module uses KAN to classify seven types of interference signals. The KANformer consists of a classification module... The model consists of five parts: block flattening, adding special characters, positional encoding information, self-attention layer, and KAN layer. Block flattening transforms image features into sequence features. Adding special characters serves as the input for subsequent KAN. Positional encoding information effectively utilizes image spatial information, enhancing the model's ability to perceive contextual information and better understand the positional information between blocks. The encoding part employs a multi-head attention mechanism to effectively capture global information of the input sequence. The KAN layer ensures that the output dimension is consistent with the multi-scale convolution fusion dimension. Finally, the multi-scale convolution features are fused and concatenated with global information to construct high-order semantic information, which is then classified and recognized by a classification module.

[0017] According to the present invention, step S3 above includes the following steps:

[0018] Training sample information is obtained from S1, and a neural network model is obtained from S2. Offline training is completed using the cross-entropy loss function to obtain an interference recognition model.

[0019] According to the present invention, step S4 above includes the following steps:

[0020] The system senses the data to be tested online and uses the S3 interference identification model to complete online interference identification.

[0021] Beneficial effects

[0022] This invention achieves the identification of seven typical interference signals based on the Conv-KANformer neural network. The multi-scale convolutional layers introduce four convolutional blocks of different sizes, while simultaneously introducing 1x1 convolutional kernels to reduce computational load. The KAN network replaces the multi-layer perceptron (MLP) in the traditional transformer; KAN uses a non-linear kernel function instead of the linear function of MLP, and the activation function is learnable. Basic splines are used to fit the function curve, which can quickly approximate any function with higher accuracy. By fusing multi-scale convolutional features and global features based on the Conv-KANformer neural network, richer semantic information is obtained, improving the accuracy of interference identification at a lower noise-to-interference ratio. This invention provides a new technical approach for interference identification. Attached Figure Description

[0023] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0024] Figure 1 This is a schematic diagram of interference recognition based on the Conv-KANformer neural network of the present invention;

[0025] Figure 2 This is a schematic diagram of the Conv-KANformer neural network framework of the present invention;

[0026] Figure 3 This is a schematic diagram of the multi-scale convolutional layer of the present invention; Detailed Implementation

[0027] To make the steps of the present invention more detailed and clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0028] The interference recognition algorithm based on Conv-KANformer neural network proposed in this invention is as follows: Figure 1 As shown in step S1, an interference database is constructed. The interference signals may include, but are not limited to, seven types: single-tone interference, multi-tone interference, linear frequency sweep interference, pulse interference, partial frequency band noise interference, sinusoidal frequency modulation interference, and noise frequency modulation interference, as shown in S1. The specific mathematical models for the seven typical interference signals are as follows:

[0029] (1) Single-tone interference

[0030] The expression for a continuous wave (CW) signal is:

[0031]

[0032] Where A is the amplitude of the single-tone interference signal, f J For single-tone interference signal frequency, This is the initial phase of the single-tone interference.

[0033] (2) Multi-tone interference

[0034] The expression for a multi-tone jamming (MTJ) signal is:

[0035]

[0036] Where N is the number of pitches; f i θ is the center frequency of the i-th tone; i Let be the initial phase of the i-th tone.

[0037] (3) Linear frequency sweep interference

[0038] The expression for Linear Frequency Modulation (LFM) is:

[0039]

[0040] Where A(t) is the amplitude of the linear sweep interference, T is the sweep period, f0 is the center frequency of the sweep, and K is the linear sweep frequency. Sweep interference is a time- and frequency-discontinuous blocking interference that can be used to interfere with conventional communication signals as well as frequency-hopping communication signals.

[0041] (4) Pulse interference

[0042] The expression for Periodic Pulse Noise Jamming (PPNJ) is:

[0043]

[0044] Where N(t) represents the Gaussian white noise signal.

[0045] (5) Partial frequency band noise interference

[0046] Partial-band noise jamming (PBNJ) manifests as Gaussian white noise within a specific frequency band, and its expression is as follows:

[0047]

[0048] Among them, U n (t) represents a function with a mean of zero and a variance of . baseband noise, f J The center frequency of the signal. The phases are uniformly distributed and independent within [0, 2π].

[0049] (6) Sinusoidal frequency modulation interference

[0050] The expression for sinusoidal frequency modulation (SFM) interference is:

[0051]

[0052] Where S(t) is the cosine modulation signal; K FM This represents the frequency modulation coefficient.

[0053] (7) Noise frequency modulation interference

[0054] The expression for Noise Frequency Modulation (NFM) is:

[0055]

[0056] Where A is the amplitude of the noise FM signal, f0 is the carrier frequency of the noise FM signal, and k fm Let ξ(t) be the frequency modulation index, and let ξ(t) be the mean and variance of zero. It is a narrowband Gaussian white noise of a certain value. It is a Wiener process, belonging to a The Gaussian distribution. Frequency modulation index k. fm and variance Together, they determine the effective bandwidth of noise frequency modulation.

[0057] In step S1, an interference database is constructed. A portion of the samples from a large number of interference signals is randomly selected as the training set for the neural network. Seven typical interference signals are generated using MATLAB, with a sampling rate of 50MHz, 1024 sampling points, and an interference-to-noise ratio (IRR) ranging from -10dB to 20dB with a step size of 2dB. 100 samples are generated for each type of interference signal at each IRR, totaling 11200 samples. The dataset is divided into a 4:1 partition. The training set contains 8960 samples, and the validation set contains 2240 samples. The true classification label for each interference sample is also labeled. The interference time-frequency graph, after preprocessing, has a size of 3*128*128.

[0058] In step S2, a Conv-KANformer neural network is constructed, which includes three parts: a multi-scale convolutional feature extraction module, a KANformer global feature extraction module, and a classification module. Figure 2 This is a Conv-KANformer neural network interference recognition framework.

[0059] The S2.1 multi-scale convolutional feature extractor uses a cascaded approach to construct different convolutional kernels, capturing features from different receptive fields, such as... Figure 3 As shown, the time-frequency image with dimensions 3*100*100 first becomes 192*10*10 after passing through the first feature extraction layer; then it passes through four multi-scale convolutional layers, with the dimensions changing from top to bottom as follows: 96*10*10, 64*10*10, 96*10*10, and 64*10*10; then the output dimensions of the four layers are fused and concatenated along the channel feature vectors to form a 320*10*10 output vector; finally, after passing through a 1x1 convolutional kernel, adaptive global average pooling, and flattening, the dimension becomes 512.

[0060] S2.2 KANformer consists of five parts: block flattening, adding special characters, positional encoding information, self-attention layer, and KAN layer. First, the image with a time-frequency frequency of 3*100*100 size is divided into blocks of 10*10 size. A convolution operation is performed with a kernel size of 10, a stride of 10, and an output dimension of 300. The input image of 3*100*100 dimension is transformed into 300*10*10, and then the dimensions are swapped to obtain a sequence of size 100*300. Next, special characters and positional encoding information are added to generate a special character of size 1*300, which is then concatenated to the original sequence, resulting in a final output sequence of 101*300. Then, positional information of size 101*300 is generated for each of the 101 blocks. The positional encoding vector and the sequence are directly added together as the input vector for KANformer encoding. In the KANformer encoder, self-attention is used to obtain the entire sequence information. The number of multi-heads is 12. KAN is used to replace the MLP in the traditional transformer. The final output dimension is 512.

[0061] The S2.3 classification module uses KAN, with an input dimension of 1024 and an output of the number of interference types. Multi-scale convolution obtains features with different receptive fields, while the KANformer quickly captures global feature information and effectively concatenates them into a higher-dimensional feature vector. Through the KAN classifier, interference identification is quickly achieved.

[0062] Step S3, train the model. Training samples: S = {x} i ,y i}, where xi is the time-frequency diagram of the input interference data, and y i True classification labels for interference signals.

[0063] Training method: The number of output neurons is equal to the number of interference types. Cross-entropy is used as the model loss function, and its expression is:

[0064]

[0065] Where C is the number of categories, y i For the label, p i To predict probabilities.

[0066] Step S4 involves online sensing of the data to be tested and using the S3 interference identification model to complete online interference identification, which can accurately identify 7 types of interference at a low interference-to-noise ratio.

Claims

1. A neural network interference recognition method based on Conv-KANformer, characterized in that, Includes the following steps: S1: Construct an interference database and extract interference time-frequency maps using short-time Fourier transform; perform preprocessing operations such as cropping, scaling, and constructing label information on the time-frequency maps to ensure that the pixel size of the time-frequency maps is 3x128x128, ensuring the accuracy and consistency of the data and improving the training speed; S2: Conv-KANformer neural network framework is built. This network consists of a multi-scale convolutional feature extraction module, a KANformer global feature extraction module, and a classification module. The multi-scale convolutional feature extraction module contains four layers of convolutional kernels of different sizes to extract features from different receptive fields. The KANformer global feature extraction module incorporates Kolmogorov-Arnold Networks (KANformer). KAN improves upon the traditional transformer structure; the classification module introduces KAN to classify seven types of interference signals; the KANformer consists of five parts: block flattening, adding special characters, positional encoding information, self-attention layer, and KAN layer; block flattening transforms image features into sequence features, adding special characters as subsequent KAN input, positional encoding information effectively utilizes image spatial information, enhances the model's ability to perceive contextual information, and better understands the positional information between blocks; the encoding part adopts a multi-head attention mechanism to effectively capture global information of the input sequence, and the KAN layer ensures that the output dimension is consistent with the multi-scale convolution fusion dimension; finally, the multi-scale convolution features and global information are fused and concatenated to construct high-order semantic information, which is then classified and recognized by the classification module. S3: The model is trained offline using the cross-entropy loss function to obtain the interference recognition model; S4: Use the interference identification model to complete online interference identification.

2. The method according to claim 1, characterized in that, Step S1 includes: First, seven types of interference signal datasets were collected, and corresponding real signal label information was constructed. Short-time Fourier transform was selected to preprocess the original data, preserving time and frequency domain information. The time-frequency graph was cropped and scaled to ensure that the pixel size of the time-frequency graph was 3x128x128, ensuring the accuracy and consistency of the data and improving the training speed. After the dataset was generated and preprocessed, the real label of each interference sample was labeled for use in the offline training stage of step S3.

3. The method according to claim 1, characterized in that, Step S3 includes: Training samples: ,in Input interference time-frequency diagram, The current input is the perturbation of the true classification label information; training method in The parameters of the trained neural network, The loss function is defined using the AdamW optimizer, which enables fast model convergence; where... The loss function used is the cross-entropy loss function.

4. The method according to claim 1, characterized in that, Step S4 includes: Obtain online interference data and use the S3 interference identification model to complete online interference identification.

Images