Kan-transformer radar signal classification method based on global context gating enhancement

Abstract

The application discloses a KAN-Transformer radar signal classification method based on global context gate enhancement. It aims to solve the technical problem of poor radar signal sorting effect in a complex electromagnetic environment. It includes: 1) standardizing and preprocessing the input PDW sequence; 2) generating a time series feature tensor by using an independent feature embedding strategy; 3) constructing a KAN-based sequence encoding unit by stacking an adaptive spline nonlinear coding layer, using a residual topology structure with a pre-normalization Pre-Norm at each layer, and extracting a time series feature sequence; 4) introducing a gating mechanism at the output end of the encoding unit to suppress interference; and 5) based on the feature vector calibrated by the gating mechanism, using a multilayer perceptron MLP classifier to complete the classification decision of the radar signal. The application can realize efficient and accurate identification of signals in small sample, low signal-to-noise ratio and high pulse loss rate scenarios, and meet the real-time needs of electronic warfare systems.

Description

Technical Field

[0001] This invention belongs to the field of radar reconnaissance and electronic countermeasures technology, and further relates to intelligent radar signal sorting and identification technology. Specifically, it is a radar signal classification method based on global context-gated enhancement of the Kolmogorov-Arnold Transformer (KAN-Transformer), which can be used for dense radar pulse stream sorting, individual radiation source identification and real-time signal processing of electronic warfare systems in complex electromagnetic environments. Background Technology

[0002] In the field of radar reconnaissance, radar signal sorting (RSS) is a core component. Efficient and reliable sorting technology is the prerequisite and guarantee for information processing and analysis in reconnaissance systems. It aims to separate pulses from different radiation sources from densely interwoven pulse streams, providing a foundation for radiation source identification and countermeasures. Radar signal sorting refers to the process of separating received radar pulse stream data according to their respective radiation sources; its essence is a "matching" problem. After receiving a signal, the primary task is to extract the pulse descriptor word (PDW) composed of instantaneous parameters (such as TOA, RF, PW, PA, DOA) and secondary derived parameters (such as PRI). From the perspective of signal characteristics, these parameters essentially belong to inter-pulse modulation features. Based on this, early classical sorting research mainly focused on inter-pulse feature mining. With the increasing complexity of the modern electromagnetic environment, radar signal overlap is becoming more frequent. Simultaneously, the continuous development of electronic counter-reconnaissance technology, the variability of radar parameters, and the increasing complexity of systems pose significant challenges to radar signal classification. Existing radar signal sorting and classification technologies are no longer sufficient to meet practical needs.

[0003] Radar signal sorting is a core component of electronic support systems, but it faces multiple challenges in the modern electromagnetic environment: the widespread use of multi-function radars (MFRs) leads to complex modulation of pulse parameters, such as PRI, PW, and RF, resulting in jitter, fluctuations, and slippage, causing severe parameter overlap; pulse loss and measurement noise further exacerbate the sorting difficulty, and traditional methods are prone to batch over-sorting or missed detections. Existing technical solutions can be divided into three categories: first, PRI-based methods, such as the cumulative difference histogram method, rely on the pulse arrival time difference, but are sensitive to PRI jitter and experience a sharp performance drop when pulses are lost; second, pulse parameter density-based clustering methods, such as DBSCAN and K-means, utilize multi-dimensional parameter clustering, but high-dimensional spatial distance metrics fail, easily confusing different radiation sources; third, deep learning-based methods, such as RNN and CNN, can automatically extract features, but RNN suffers from gradient vanishing and CNN has weak temporal modeling, and also has high computational complexity and insufficient generalization ability. These methods are all unable to effectively solve the problems of parameter overlap and pulse loss, limiting their application in real-world complex environments.

[0004] Early traditional methods based on pulse descriptor words (PDWs), such as template matching and PRI analysis, suffer from poor adaptability to non-cooperative targets and complex intra-pulse modulated signals due to their reliance on periodic assumptions and manual rules. They also suffer from computational redundancy, harmonic interference, and insufficient real-time performance. Subsequent methods for intra-pulse feature extraction, such as time-frequency analysis, fuzzy functions, and higher-order statistics, can uncover more details, but their adaptability to low signal-to-noise ratio and multi-signal overlapping scenarios remains limited by time-frequency resolution, cross-term interference, or computational complexity. Individual recognition techniques based on unintentional modulation features improve discriminative power but are limited by high computational cost and insufficient robustness. Traditional machine learning methods, while improving flexibility to some extent, rely on manual feature engineering and struggle to capture deep nonlinear patterns. Among the closest existing technologies, deep learning models, represented by Transformer, have been used for radar signal processing. For example, the MSCformer network based on multiple self-attention coupling mechanism proposed by Zhou et al. has shown high sorting accuracy and computational efficiency under conditions of small sample size, parameter overlap, and high pulse loss rate. However, such models still have obvious defects: on the one hand, they rely on fixed activation functions and linear weight adjustments, making it difficult to accurately fit radar signal parameters that are affected by environmental noise, multipath effects, and other complex nonlinear distortions; on the other hand, excessive stacking of network layers and widening of hidden layers to improve fitting ability leads to redundant parameters, which can easily cause overfitting when radar samples are limited (especially in small sample scenarios), and the inference computation is large, which cannot meet the real-time requirements of electronic warfare systems; at the same time, although pure Transformer models are good at capturing global dependencies, their feedforward network has insufficient function approximation ability for local fine features, making it difficult to take into account both contextual dependencies and local features.

[0005] In summary, although existing technologies have made significant progress in the field of radar signal sorting, many challenges remain. Exploring more efficient and robust radar signal classification methods to meet the dual requirements of real-time performance and accuracy in radar signal sorting is a key issue that urgently needs to be addressed in this field. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a KAN-Transformer radar signal classification method based on global context gating enhancement. This method solves the technical problems of insufficient fitting ability of nonlinear features of radar signals in complex electromagnetic environments, model parameter redundancy and overfitting, and the difficulty in simultaneously considering global context dependence and local fine features. First, a KAN-Linear layer is constructed to replace the fully connected layer in the Transformer feedforward network (FFN), and a learnable B-spline activation function is used to replace the fixed activation function and linear weights, achieving accurate fitting of the nonlinear distortion features of radar signals. Then, a KAN-based global context gating mechanism is introduced, combined with a dual-stream feature extraction branch, to dynamically filter globally dependent features and local fine features, while reducing the number of network parameters. Through these improvements, this invention can achieve efficient approximation of nonlinear features of radar signals and dynamic filtering of key features in scenarios with small samples, low signal-to-noise ratio, and high pulse loss rate, thereby achieving higher recognition accuracy with fewer parameters and meeting the real-time requirements of electronic warfare systems.

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

[0008] (1) Standardize the input pulse descriptor word (PDW) sequence to make it have zero mean and unit variance, so as to ensure that the input data falls within the effective activation range of the B-spline grid;

[0009] (2) An independent feature embedding strategy is adopted. The preprocessed PDW sequence is first input into a high-dimensional linear projection layer to map the physical features of the pulse into a high-dimensional semantic feature vector. Then, a learnable position code is superimposed to inject the temporal position information of the sequence into the feature space and generate a temporal feature tensor.

[0010] (3) A sequence coding unit based on the Kolmogorov-Arnold network (KAN) is constructed by stacking L layers of adaptive spline nonlinear coding layers. The unit takes the temporal feature tensor as input data. Each layer adopts a pre-normalized Pre-Norm residual topology structure, which contains two cascaded sub-layers: the first sub-layer is a global context association modeling layer, and the second sub-layer is a hybrid learnable activation function processing module, namely a hybrid KAN linear transformation layer. The unit first uses the multi-head self-attention mechanism MHSA to capture the long-range correlation between impulses in the temporal feature tensor. Then, through the hybrid KAN linear transformation layer, the dynamic change curve is accurately fitted by the B-spline activation function, and the temporal feature sequence extracted by the unit is output.

[0011] (4) A gating mechanism is introduced at the output of the encoding unit to suppress the interference of measurement noise and abnormal features on the recognition results. This includes compressing the temporal feature sequence extracted by the encoding unit in the time dimension through global average pooling to generate a context vector rich in global statistical information; constructing a global context gating module based on KAN, generating adaptive gating coefficients through a combination network of "linear layer + GELU + KAN + Sigmoid", and weighting and enhancing the context vector to obtain the gating calibration feature vector.

[0012] (5) The feature vector of the gate calibration is processed by layer normalization to stabilize the feature distribution, and then input into the multilayer perceptron (MLP) classifier. The classifier outputs the probability distribution of each category to complete the classification decision of the radar signal.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] First, this invention employs a dual-stream parallel computing architecture based on B-spline basis functions. By replacing fixed activation logic with learnable spline functions, each neuron in the network possesses the ability to adaptively adjust the shape of its response curve. Experimental results demonstrate that when processing radar pulse data of the same scale, this invention can reproduce the time-frequency domain distortion patterns of the signal with extremely high fidelity, achieving a recognition accuracy of up to 97.95% on the test set, and the recognition rate for various radar targets approaches 100%, proving that this architecture represents a substantial technological breakthrough in refined feature extraction.

[0015] Secondly, this invention introduces a global context gating unit based on a nonlinear sensing mechanism. By setting an independent gating adjustment loop at the end of the encoding link, a high-order nonlinear network is used to extract global signal-to-noise ratio context information from the encoded features, and dynamic gating coefficients are generated to reweight the backbone features element by element. This mechanism can automatically sense and suppress abnormal feature dimensions caused by measurement noise, pulse loss, or environmental interference, while enhancing the weight of effective fingerprint features. Experimental data show that the model can converge to an extremely low loss value (0.0691) within a very short training period, and the final weighted F1 score reaches 0.9796, which greatly enhances the model's noise robustness and feature selection ability in complex electromagnetic environments.

[0016] Third, because this invention configures independent linear projection channels with non-shared weights for each physical parameter, it maps scalar features into high-dimensional independent feature vectors. This preserves the independent semantic space of each physical attribute to the greatest extent, enabling the subsequent self-attention mechanism to more sensitively capture the potential nonlinear coupling and modulation rules between different physical parameters. This effectively solves the semantic aliasing problem of heterogeneous physical parameters in the feature fusion process and significantly improves the model's ability to analyze and distinguish complex modulation modes of multifunctional radar.

[0017] Fourth, the recognition model constructed in this invention only requires 3 encoder layers to achieve the expected fitting effect. Experimental verification shows that the model has approximately 0.22M parameters and only 0.01M floating-point operations (FLOPs). This lightweight feature enables the invention to be directly deployed on resource-constrained computing platforms such as spaceborne, missile-borne, or UAV platforms, while maintaining high recognition accuracy while meeting the stringent real-time requirements of electronic warfare systems for signal sorting. Attached Figure Description

[0018] Figure 1 This is a diagram of the overall architecture of the KAN converter network with global context gating enhancement in this invention.

[0019] Figure 2 This is a schematic diagram of the encoder layer structure that integrates KAN in this invention;

[0020] Figure 3 This is a schematic diagram of the internal calculation principle of the hybrid KAN linear transform layer in this invention;

[0021] Figure 4 This is a schematic diagram of the global context gating mechanism in this invention;

[0022] Figure 5 The figure shows the simulation results of the model validation set accuracy of this invention changing with the number of training rounds under a 20dB signal-to-noise ratio environment.

[0023] Figure 6 A schematic diagram of the classification confusion matrix of the model of this invention at a signal-to-noise ratio of 20dB. Detailed Implementation

[0024] The present invention will now be further described with reference to the accompanying drawings.

[0025] Example 1: Refer to Appendix Figure 1-4 This invention proposes a KAN-Transformer radar signal classification method based on global context gating enhancement, comprising the following steps:

[0026] Step 1) Standardize the input pulse descriptor word (PDW) sequence to ensure it has zero mean and unit variance, thus guaranteeing that the input data falls within the effective activation range of the B-spline grid. This embodiment achieves this as follows:

[0027] (1.1) Calculate the mean of each feature in the training set. and standard deviation ;

[0028] (1.2) For the input sample Perform the following transformations:

[0029] ,

[0030] Ensure that the input data is distributed in a zero-mean, unit-variance distribution space to prevent gradient vanishing or exploding.

[0031] Step 2) An independent feature embedding strategy is adopted. The preprocessed PDW sequence is first input into a high-dimensional linear projection layer to map the physical features of the pulses into high-dimensional semantic feature vectors. Then, a learnable positional encoding is superimposed to inject the temporal positional information of the sequence into the feature space, generating a temporal feature tensor. In this embodiment, the specific implementation steps of the above independent feature embedding strategy include:

[0032] (2.1) First, set the dimension of the hidden layer of the model to be... For each input scalar physical parameter, a separate linear projection layer is configured to map it to a... A feature vector of dimension , i.e., a word; assuming the number of input features is . Then generate Each word is a token; the physical parameters in this step of the embodiment include radio frequency RF, pulse width PW, and pulse repetition interval PRI.

[0033] (2.2) Initialize a learnable category lexical CLS Token, whose dimension is... Used to aggregate global classification information;

[0034] (2.3) Connect the CLS Token with the aforementioned Each word token is concatenated along the sequence dimension to generate a shape of ( ). Input temporal feature tensor ;in This refers to the number of samples processed in a batch. It is the number of lexical tokens plus the number of category lexical tokens (CLS tokens); It is the dimension of the feature.

[0035] Step 3) Refer to Figure 2 The sequence coding unit is constructed by stacking L layers of adaptive spline nonlinear coding layers based on the Kolmogorov-Arnold network (KAN). This unit takes temporal feature tensors as input data. Each layer adopts a residual topology structure with pre-normalization (Pre-Norm), which contains two cascaded sub-layers: the first sub-layer is a global context association modeling layer, and the second sub-layer is a hybrid learnable activation function processing module, namely a hybrid KAN linear transformation layer. This unit first uses the multi-head self-attention mechanism (MHSA) to capture the long-range correlation between impulses in the temporal feature tensor. Then, after passing through the hybrid KAN linear transformation layer, the dynamic change curve is accurately fitted using the B-spline activation function, and the temporal feature sequence extracted by this unit is output.

[0036] In this embodiment, the first sub-layer first normalizes the received temporal feature tensor, and then uses a multi-head self-attention module (MHSA) to compute attention weight matrices in parallel across multiple feature subspaces to model the global dependencies between impulses. The output, after random deactivation and Dropout regularization, is added to the original input via residual connections to generate context-enhanced features. (Refer to...) Figure 3 The second sub-layer described above employs a parallel dual-branch architecture consisting of a global residual approximation branch and a local nonlinear fitting branch. Context enhancement features are processed collaboratively in the two parallel branches. In the global residual approximation branch, the input features are first pre-activated by the SiLU or GELU function, and then input to the basic linear projection module. The learned linear weight matrix is ​​used to map the input features to the target output dimension, generating basic output features that contain the overall trend of the signal. In the local nonlinear fitting branch, based on KAN theory, a high-order nonlinear fitting is performed on the local fingerprint features of the input signal using a learnable B-spline basis function. Finally, the basic output features of the global residual approximation branch and the nonlinear fitting features of the local nonlinear fitting branch are added element-wise, and the final output of the module is obtained through dual-stream feature fusion.

[0037] Furthermore, the above-mentioned high-order nonlinear fitting of the local fingerprint features of the input signal using learnable B-spline basis functions specifically includes the following steps:

[0038] (3b1) Flatten the three-dimensional sequence tensor into a two-dimensional matrix, discretize the input domain using a preset grid, and generate the spline basis response matrix using the Cox-deBoor recursive algorithm;

[0039] (3b2) Introduce learnable gain coefficients, which are then dimensionally aligned and weighted and modulated with the original spline control weights to generate calibrated control weights;

[0040] (3b3) Input the spline basis response matrix and the calibrated control weights into the sparse linear projection module to generate nonlinear fitting features.

[0041] In this embodiment, to achieve high-precision nonlinear mapping of radar pulse parameters, a vectorized computation unit based on the Cox-de Boor recursive algorithm is preferably used. The data processing flow of this unit mainly includes two stages:

[0042] Phase 1: Grid localization and zero-order activation.

[0043] The system receives normalized input feature components and a preset discretized feature node vector. Through grid positioning logic, it calculates the relative positions of the input features within the discrete grid in parallel, generating an interval indicator tensor, i.e., the zeroth-order response. Mathematically, this tensor is represented by a set of non-overlapping rectangular pulses, used to indicate the local interval index where the input signal is currently located.

[0044] Second stage: Recursive progression.

[0045] The system generates high-order continuous basis functions iteratively based on the set spline order. Each iteration involves the following two steps: a. Mixed weight calculation: The system calculates the mixed weights of the basis functions based on the distance relationship between the input features and the node vectors; b. Recursive aggregation: The recursive calculation unit receives the lower-order left-neighborhood response and the lower-order right-neighborhood response from the previous order and performs neighborhood-weighted aggregation on them using the aforementioned mixed weights.

[0046] Through multiple iterations, the originally discrete rectangular pulses gradually evolve into smooth curves with local characteristics, ultimately outputting a high-order nonlinear basis function response, providing a high-dimensional feature space for subsequent adaptive weight calibration.

[0047] Step 4) Refer to Figure 4A gating mechanism is introduced at the output of the encoding unit to suppress the interference of measurement noise and abnormal features on the recognition results. This includes compressing the temporal feature sequence extracted by the encoding unit in the time dimension through global average pooling to generate a context vector rich in global statistical information; constructing a global context gating module based on KAN, generating adaptive gating coefficients through a combination network of "linear layer + GELU + KAN + Sigmoid", and weighting the context vector to suppress interference such as noise and abnormal features, thereby obtaining a gating calibration feature vector. The gating mechanism introduced in this step of this embodiment specifically includes: taking the first token in the output sequence of the encoding unit, which corresponds to the CLS token or the feature mean, and projecting it through the linear layer to input it into an independent KAN linear layer; using KAN to learn the nonlinear distribution law of signal quality from the features, i.e., the global context; outputting the gating coefficient vector G with a value range of [0,1] through the Sigmoid activation function; and performing element-wise multiplication of the gating coefficient G with the original encoded feature F.

[0048]

[0049] This step can automatically enhance key fingerprint features and suppress the weight of noise dimensions based on global signal-to-noise ratio information.

[0050] Step 5) The gating calibration feature vector is processed by layer normalization to stabilize the feature distribution, and then input into a multilayer perceptron (MLP) classifier. This classifier outputs the probability distribution of each category to complete the classification decision of the radar signal. In this embodiment, the MLP classifier is specifically a classifier composed of "linear layer → GELU activation → Dropout → linear layer", which is used to map high-dimensional features to the target radar category space. If there is a class imbalance in the classification output stage, i.e., some radar target samples have very few samples, causing the model to favor the category with more samples, then in the training stage, the number of samples of each category in the training set is first counted, the reciprocal of the category weight is calculated, and then a weighted random sampler is used to build a data loader so that the model focuses on the small sample category during training. At the same time, a weighted cross-entropy loss function is used to increase the loss weight of the small sample category, which is used to solve the recognition bias problem caused by the long tail distribution of the dataset.

[0051] Example 2: Refer to Appendix Figure 1 The overall implementation steps of the radar signal classification method proposed in this embodiment are the same as those in Embodiment 1. The implementation process of the model of this invention will now be further described from the perspective of data processing:

[0052] The global context-gated enhanced KAN transformer network model constructed in this invention adopts an end-to-end sequence processing architecture. The network input is a normalized preprocessed PDW pulse sequence matrix, and the data flow sequentially goes through four stages: feature embedding, KAN encoder stacking, global feature aggregation, and gating classification.

[0053] (i) Semantic feature embedding stage: The standardized preprocessed PDW pulse sequence is first input into the high-dimensional linear projection layer to map the physical features into semantic feature vectors. Then, learnable position codes are superimposed to inject the temporal position information of the sequence into the feature space, providing a temporal benchmark for the model.

[0054] (ii) KAN encoder stacking stage: Feature sequences enter the stack The layer encoder employs a pre-normalized residual structure in each layer. First, it uses a multi-head self-attention mechanism (MHSA) to capture long-range correlations between pulses. Then, the data stream passes through a hybrid KAN linear transform layer, and the dynamic change curve is accurately fitted using a B-spline activation function.

[0055] (III) Global Aggregation and Gating Stage: The temporal feature sequence extracted by the encoder is first compressed by global average pooling in the time dimension to generate a context vector rich in global statistical information. This vector is then fed into the global context gating module, where an adaptive gating coefficient is generated through a combination network of "linear layer + GELU + KAN + Sigmoid" to weight and enhance the feature vector.

[0056] (iv) Classification and decision stage: The feature vector after gating calibration is input into the MLP classifier through the final layer normalization process, mapped to the preset radar signal category space, and the final probability distribution is output.

[0057] like Figure 1 As shown, the signal processing flow of this invention begins with the data input and preprocessing stage. The system receives radar pulse descriptor (PDW) sequences intercepted by a reconnaissance receiver. Considering the significant differences in the dimensions of physical parameters such as carrier frequency, pulse width, and pulse repetition interval, a standardization module first maps each dimension of the parameters to a unified numerical range to adapt to the input requirements of the subsequent nonlinear network. Subsequently, the data enters an independent feature lexicalization module. This module abandons the traditional feature concatenation method and instead uses multiple parallel linear projection units to independently map each scalar physical parameter into a high-dimensional feature embedding vector, which is then concatenated with a learnable category lexical, thereby constructing an input temporal feature sequence that retains the independent semantics of each physical quantity.

[0058] Next, the aforementioned feature sequences enter the core feature extraction part of the system, namely the encoder, which consists of several cascaded stacked sequence coding units based on a Kolmogorov-Arnold network. In this stage, the data stream undergoes global correlation modeling via a multi-subspace self-attention mechanism within each coding unit, and a hybrid learnable activation function processing module is used to perform high-order nonlinear fitting and transformation on the minute fingerprint features in the radar signal, thereby outputting a coded vector rich in temporal dependency information and nonlinear fingerprint features.

[0059] At the encoder output, this invention innovatively incorporates a global context gating module. This module receives the encoded feature vector as input and uses a nonlinear sensing network to learn the signal quality distribution from the global context, generating a dynamic gating coefficient with a value between 0 and 1. This coefficient is then weighted and modulated element-wise with the feature vector of the backbone network, thereby achieving adaptive enhancement of key fingerprint features and automatic suppression of measurement noise or anomalous feature dimensions. Finally, the gated and enhanced feature vector is processed by layer normalization and input to the classification decision output module. The classifier maps the vector to a preset radar radiation source category space, outputting the final individual identification probability distribution result.

[0060] Example 3: The overall implementation steps of the radar signal classification method proposed in this example are the same as in Example 1. Please refer to the appendix for details. Figure 2-3 A more detailed description of the specific architecture of the sequence coding unit based on the Kolmogorov-Arnold network (KAN) is provided below:

[0061] The sequence coding unit based on the Kolmogorov-Arnold network (KAN) described above is the core feature extraction component constructed in this invention. For example... Figure 2 As shown, this unit consists of multiple stacked adaptive spline nonlinear coding layers (FlexibleEncoder Layer), each layer employing a pre-normalized residual topology, and mainly contains two cascaded processing sub-layers:

[0062] The first sublayer is a global context association modeling layer used to capture the correlation between different pulse parameters, aiming to capture the long-range dependencies between different pulses in a radar pulse sequence. The input sequence first passes through a feature distribution normalization module to stabilize the data distribution, then enters a multi-subspace self-attention module to calculate the correlation weights between pulses in different feature subspaces. Finally, random feature sparsification, i.e., Dropout, is used to prevent overfitting, and information flow shortcut aggregation is performed with the original input.

[0063] The second sublayer is a hybrid learnable activation function processing module (KAN layer), designed to perform fine nonlinear transformations on radar parameters to extract fingerprint features. To overcome the limitation of traditional neural networks where fixed activation functions cannot simultaneously preserve linear features and fit high-order nonlinearities, this sublayer innovatively introduces a vectorized nonlinear mapping component to replace the traditional feedforward neural network. This component first transforms the three-dimensional temporal tensor into a two-dimensional matrix form suitable for spline computation by rearranging the temporal-feature dimensions. Then, it uses an adaptive weight-calibrated two-stream spline transformation unit, i.e., the improved KAN, to perform high-order function fitting. After the transformation, the temporal dimension of the data is restored through a temporal structure restoration module. This process not only retains the Transformer's ability to process sequences but also endows the model with the ability to approximate complex radar fingerprints using B-spline functions.

[0064] like Figure 3 As shown, the second sub-layer adopts a parallel dual-branch architecture. The input feature tensors are fed in parallel into the global residual approximation branch (Base Branch) and the local nonlinear fitting branch (Spline Branch) for collaborative processing, wherein:

[0065] a) Global residual approximation branch:

[0066] This branch aims to capture the overall linear trend of the input signal and acts as a shortcut, similar to that in ResNet, providing a stable gradient propagation path and convergence benchmark for model training. Processing flow: The input features are first pre-activated using a non-linear activation function (SiLU is preferred in this embodiment), and then input to the Base Linear Projection module. A learnable linear weight matrix maps the features to the target output dimension, generating a base output that contains the overall trend of the signal.

[0067] b) Local nonlinear fitting branch:

[0068] This branch utilizes Kolmogorov-Arnold Network (KAN) theory to perform high-order nonlinear fitting of the local fingerprint features of the input signal using learnable B-spline basis functions. Internally, this branch is further subdivided into two cooperating sub-modules: signal space transformation (data flow) and adaptive weight calibration (weight flow).

[0069] b1) Signal Space Transformation Submodule (Data Flow): First, the input 3D sequence tensor is flattened into a 2D matrix to accommodate subsequent grid computation requirements; this step is called Flattening. Then, the input domain is discretized using a predefined grid, and the Cox-de Boor recursive algorithm is used to calculate the response values ​​of the input data in each grid interval and at different orders; this step is called Basis Expansion. This process maps low-dimensional input features to a high-dimensional B-spline basis function space, generating a spline basis response matrix.

[0070] b2) Adaptive Weight Calibration Submodule (Weight Flow): To achieve adaptive gain adjustment for each input feature dimension and address the common dimension broadcasting error problem in parallel matrix operations, this embodiment introduces a learnable gain coefficient (Spline Scaler). Dimension Alignment and Modulation: This coefficient is first expanded to the same dimension as the weight tensor through an unsqueeze operation, and then broadcast-parameter-weighted modulation is performed with the original spline control weights. This step generates calibrated control weights with amplitude correction.

[0071] b3) Sparse linear projection and feature fusion:

[0072] After the above two sub-modules have completed their processing, the generated spline basis response matrix and the calibrated control weights are input into the Sparse Linear Projection module. Through matrix multiplication, the features of the high-dimensional basis function space are mapped back to the target feature space, generating nonlinear fitting features (Spline Output).

[0073] Finally, the system performs an element-wise summation of the basic output features generated by the global residual approximation branch and the nonlinear fitting features generated by the local nonlinear fitting branch, obtaining the final output of the module through dual-stream feature fusion. This design preserves the linear principal components of the signal while accurately characterizing the nonlinear details of the radar fingerprint, significantly improving the model's feature representation capability.

[0074] The effects of the present invention will be further explained below with reference to simulation experiments.

[0075] 1. Simulation conditions:

[0076] The simulation experiment hardware and software environment of this invention is shown in the table below:

[0077] Table 1. Experimental Software and Hardware Environment Configuration Table

[0078]

[0079] 2. Simulation content:

[0080] A radar signal dataset containing various inter-pulse modulation types, complex channel environments, and non-ideal reception conditions was constructed. Ten typical radar inter-pulse modulation signals were selected as sorting objects, covering various types such as fixed PRI, PRI staggered, PRI jitter, PRI slip, frequency agility, and PRI coincident modulation. 10,000 samples were generated for each signal to test the network's ability to learn and generalize inter-pulse modulation patterns under complex conditions such as pulse loss and noise interference. The specific parameter settings are shown in the table below.

[0081] Table 2 Radar Signal Inter-Pulse Modulation Parameter Settings

[0082]

[0083] The simulated signal carrier frequencies are all located in the X-band, and the sampling frequencies are set according to the bandpass sampling theorem. The following channel non-ideal factors are introduced into the signal generation process to improve data fidelity: phase offset is added (…). ), frequency offset ( )MHz, introducing a Rician distribution ( The multipath effect (dB) causes the intensity to decrease to... dB, delay per path is and overlay AWGN+ The dataset is stably mixed with a signal-to-noise ratio of 20 dB, and a dropout rate of 20% to 30% is independently generated for each pulse to simulate random missed detections in a real-world environment. The dataset is divided into training, testing, and validation sets in a ratio of 0.6:0.2:0.2.

[0084] 3. Simulation results:

[0085] Figure 5 The training curves show the accuracy of the model on the validation set under a 20dB signal-to-noise ratio. In the first 50 epochs, the model's accuracy shows a rapid upward trend, quickly exceeding 80%, mainly due to the design of the basic branch in the hybrid KAN linear transform layer, which provides a stable gradient flow for the model. It eventually converges to around 98.61%, and the curve is smooth and oscillating, verifying the robustness of the model's pre-normalized residual structure and spline function interpolation in handling discontinuous time-series data.

[0086] Table 3 summarizes the test performance under different signal-to-noise ratios (SNRs). When the SNR is greater than or equal to 5 dB, the sorting accuracy remains stable at over 94%; even in extreme noise environments with an SNR of 0 dB, it can still maintain a sorting accuracy of 85.12%. This result reflects its extremely strong anti-interference capability, even in non-ideal signal environments where the signal is severely contaminated by noise and the pulse sequence is broken, it can still uncover potential modulation patterns from a macroscopic time dimension.

[0087] Table 3. Test set sorting accuracy under different signal-to-noise ratios

[0088]

[0089] Figure 6 The model's classification confusion matrix at a 20dB signal-to-noise ratio is shown. Experimental results demonstrate that the model exhibits extremely high recognition accuracy for all 10 types of inter-pulse modulated signals. The classification accuracy for categories 0, 2, 5, and 7 (jitter-related PRI) and categories 3 and 8 (slip-modulated PRI) is close to 100%. A small number of misclassifications exist between categories 1 and 4. Analysis suggests that the high pulse miss rate of 20%–30% may be due to the physical degradation of uneven pulse sequences into fixed-period sequences, which is the main reason for the slight confusion in these signals. However, even under these extreme conditions, the model maintains extremely high overall accuracy, demonstrating that the global context gating mechanism effectively fills in the semantic gaps caused by local misses by extracting macroscopic information from the sequence.

[0090] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0091] The above simulation analysis proves the correctness and effectiveness of the method proposed in this invention.

[0092] This invention can be directly deployed in spaceborne, airborne, and shipborne electronic reconnaissance (ELINT) and electronic support measures (ESM) systems, and is suitable for dense signal sorting and situational awareness in complex electromagnetic environments. This technology can serve as a core module for radar signal preprocessing, performing high-precision de-interleaving and classification of aliased pulse streams, significantly reducing the false alarm rate and missed alarm rate in subsequent signal processing. Simultaneously, the lightweight architecture based on KAN significantly reduces computational latency and possesses powerful few-sample learning capabilities, enabling it to quickly adapt to new targets through online learning or few-sample fine-tuning, accurately outputting target attributes and categories, significantly improving the system's adaptability and survivability. Furthermore, the technical architecture is highly versatile and easily extended to IoT physical layer security and RF fingerprint authentication fields, demonstrating broad application prospects.

[0093] The parts of this invention not described in detail are common knowledge to those skilled in the art.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention. These include: using Chebyshev polynomials, Legendre polynomials, or other orthogonal polynomials to replace the B-splines in this embodiment to implement the KAN layer; using adaptive grid technology, i.e., dynamically adjusting the density of the grid according to the gradient distribution during training, instead of using the fixed grid_size in the present invention; the KANLinear layer can not only replace the traditional feedforward neural network, but also replace the Q, K, V projection matrices in the attention mechanism, or only be used in the final classifier part; or not using tokenization, but concatenating all PDW parameters into a long vector and directly inputting it into the fully connected KAN network, without going through the Transformer attention mechanism, i.e., a pure KAN grid scheme. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A KAN-Transformer radar signal classification method based on global context gating enhancement, characterized in that, Includes the following steps: (1) Standardize the input pulse descriptor word (PDW) sequence to make it have zero mean and unit variance, so as to ensure that the input data falls within the effective activation range of the B-spline grid; (2) An independent feature embedding strategy is adopted, and the preprocessed PDW sequence is first input into a high-dimensional linear projection layer to map the physical features of the pulse into a high-dimensional semantic feature vector; Then, learnable positional encoding is superimposed to inject the temporal positional information of the sequence into the feature space, generating a temporal feature tensor; (3) A sequence coding unit based on the Kolmogorov-Arnold network (KAN) is constructed by stacking L layers of adaptive spline nonlinear coding layers. The unit takes the temporal feature tensor as input data. Each layer adopts a pre-normalized Pre-Norm residual topology structure, which contains two cascaded sub-layers: the first sub-layer is a global context association modeling layer, and the second sub-layer is a hybrid learnable activation function processing module, namely a hybrid KAN linear transformation layer. The unit first uses the multi-head self-attention mechanism MHSA to capture the long-range correlation between impulses in the temporal feature tensor. Then, through the hybrid KAN linear transformation layer, the dynamic change curve is accurately fitted by the B-spline activation function, and the temporal feature sequence extracted by the unit is output. (4) A gating mechanism is introduced at the output of the encoding unit to suppress the interference of measurement noise and abnormal features on the recognition results. This includes compressing the temporal feature sequence extracted by the encoding unit in the time dimension through global average pooling to generate a context vector rich in global statistical information; constructing a global context gating module based on KAN, generating adaptive gating coefficients through a combination network of "linear layer + GELU + KAN + Sigmoid", and weighting and enhancing the context vector to obtain the gating calibration feature vector. (5) The feature vector of the gate calibration is processed by layer normalization to stabilize the feature distribution, and then input into the multilayer perceptron (MLP) classifier. The classifier outputs the probability distribution of each category to complete the classification decision of the radar signal.

2. The method according to claim 1, characterized in that: The normalization preprocessing of the input pulse descriptor word (PDW) sequence described in step (1) is performed as follows: (1.1) Calculate the mean of each feature in the training set. and standard deviation ; (1.2) For the input sample Perform the following transformations: ， Ensure that the input data is distributed in a zero-mean, unit-variance distribution space to prevent gradient vanishing or exploding.

3. The method according to claim 1, characterized in that: The independent feature embedding strategy described in step (2) includes the following specific implementation steps: (2.1) First, set the dimension of the hidden layer of the model to be... For each input scalar physical parameter, a separate linear projection layer is configured to map it to a... A feature vector of dimension , i.e., a word; assuming the number of input features is . Then generate Each word token; (2.2) Initialize a learnable category lexical CLS Token, whose dimension is... Used to aggregate global classification information; (2.3) Connect the CLS Token with the aforementioned Each word token is concatenated along the sequence dimension to generate a shape of ( ). Input temporal feature tensor ;in This refers to the number of samples processed in a batch. It is the number of lexical tokens plus the number of category lexical tokens (CLS tokens); It is the dimension of the feature.

4. The method according to claim 3, characterized in that: The physical parameters mentioned in step (2.1) include radio frequency RF, pulse width PW, and pulse repetition interval PRI.

5. The method according to claim 1, characterized in that: In step (3), the first sub-layer first normalizes the received temporal feature tensor, and then calculates the attention weight matrix in parallel in multiple feature subspaces through the multi-head self-attention module MHSA to model the global dependency between pulses; after the output is regularized by random deactivation Dropout, it is added to the original input through residual connection to generate context-enhanced features.

6. The method according to claim 5, characterized in that: In step (3), the second sub-layer adopts a parallel dual-branch architecture consisting of a global residual approximation branch and a local nonlinear fitting branch, and the context enhancement features are processed collaboratively in the two parallel branches. In the global residual approximation branch, the input features are first pre-activated by the GELU function and then input to the basic linear projection module. The learned linear weight matrix is ​​used to map the features to the target output dimension, generating basic output features that contain the overall trend of the signal. In the local nonlinear fitting branch, based on KAN theory, a high-order nonlinear fitting is performed on the local fingerprint features of the input signal using a learnable B-spline basis function. Finally, the basic output features of the global residual approximation branch and the nonlinear fitting features of the local nonlinear fitting branch are added element-wise, and the final output of the module is obtained through dual-stream feature fusion.

7. The method according to claim 6, characterized in that: The step of performing high-order nonlinear fitting of the local fingerprint features of the input signal using learnable B-spline basis functions includes: (3b1) Flatten the three-dimensional sequence tensor into a two-dimensional matrix, discretize the input domain using a preset grid, and generate the spline basis response matrix using the Cox-deBoor recursive algorithm; (3b2) Introduce learnable gain coefficients, which are then dimensionally aligned and weighted and modulated with the original spline control weights to generate calibrated control weights; (3b3) Input the spline basis response matrix and the calibrated control weights into the sparse linear projection module to generate nonlinear fitting features.

8. The method according to claim 1, characterized in that: The gating mechanism described in step (4) includes the following steps: taking the first token in the output sequence of the encoding unit, which corresponds to the CLS token or the mean of features, and projecting it into an independent KAN linear layer; using KAN to learn the nonlinear distribution law of signal quality from the features, i.e., the global context; outputting the gating coefficient vector G with a value range of [0,1] through the Sigmoid activation function; and performing element-wise multiplication of the gating coefficient G with the original encoded features F to enhance the key fingerprint features and suppress the weight of the noise dimension.

9. The method according to claim 1, characterized in that: The multilayer perceptron (MLP) classifier mentioned in step (5) is a classifier composed of "linear layer → GELU activation → Dropout → linear layer", which is used to map high-dimensional features to the target radar category space.

10. The method according to claim 1, characterized in that: If there is a class imbalance in the classification output stage, i.e., some radar target samples have very few samples, causing the model to be biased towards identifying the class with more samples, then in the training stage, the number of samples of each class in the training set is counted first, the reciprocal of the class weight is calculated, and then a weighted random sampler is used to build a data loader so that the model pays attention to the small sample class during training; at the same time, a weighted cross-entropy loss function is used to increase the loss weight of the small sample class. This is used to address the identification bias caused by the long-tailed distribution of the dataset.

Images