Radar jamming recognition method based on binary depth separable convolution and electronic equipment

Abstract

The application provides a radar jamming recognition method based on a binary deep separable convolution and electronic equipment, based on a plurality of radar jamming signal models, echo signals corresponding to jamming signals and the jamming signals are acquired, contour maps of cross ambiguity functions CAF and range-doppler RD spectrum maps are respectively calculated; a double-flow adaptive real value preserving binary network ARR-BiNet network model including two same input branches is constructed, and the CAF contour maps and the RD spectrum map images are respectively received; each branch sequentially passes through a convolution layer, a deep separable convolution layer, a skip connection layer, a deep separable convolution layer, a convolution layer and a flattening layer for feature extraction; binary activation is performed after batch normalization in the convolution layer and the deep separable convolution layer; the CAF contour maps and the RD spectrum maps of the to-be-detected signals are input into the trained ARR-BiNet network model to perform a jamming signal classification task.

Description

Technical Field

[0001] This invention relates to the field of radar information processing technology, specifically to a radar interference identification method and electronic device based on binary depth-separable convolution. Background Technology

[0002] With the rapid development of modern electronic technology, the electromagnetic environment encountered by radar systems on the battlefield has become increasingly complex and variable. Accurately sensing interference is the most critical aspect of the closed-loop intelligent anti-jamming decision-making process of radar systems. At the same time, radar systems face operational limitations, such as the need for mobility, which restricts available computing resources. Therefore, researching effective methods to identify interference patterns affecting radar systems with limited computing power is crucial.

[0003] According to the jamming mechanism, jamming signals can be divided into suppression and deception. Active radar suppression jamming uses high-power jamming to completely suppress useful radar signals in both the time and frequency domains, preventing the radar from receiving and extracting target parameters. Active radar deception jamming refers to the jammer deliberately emitting or relaying signals that approach the target echo, using false targets or target information to confuse the enemy receiver, preventing the radar from functioning properly, and even causing it to receive false information created by the enemy.

[0004] Current research on active radar jamming identification primarily treats it as a pattern recognition problem utilizing the differences in characteristic parameters of various jamming signals. Classification algorithms, such as Support Vector Machines (SVM), tree models, and Logistic Regression, are then applied to identify these patterns. For example, multidimensional features are extracted and threshold estimation is performed to classify jamming signals; time-frequency domain features are used as inputs for classification via SVM and decision trees. However, these methods rely on manually extracting different features, which limits their transferability and introduces challenges such as complexity, inefficiency, and reduced accuracy.

[0005] In recent years, deep neural networks have become popular for extracting high-dimensional features from interference clutter to identify interference types and estimate parameters. For example, recognition networks with ensemble residuals and asymmetric convolutional blocks have been used; however, they are limited to the perception and recognition of suppressed interference. Attention networks specifically designed for deceiving interference patterns have been built based on the SqueezeNet model. However, these methods are generally effective only for a single interference pattern and lack scalability and generalization ability, thus limiting their effectiveness in complex and variable interference environments. Summary of the Invention

[0006] This invention proposes a radar interference identification method and electronic device based on binary depth separable convolution to solve the technical problems of low efficiency and low accuracy of existing technologies, and difficulty in dealing with non-single interference.

[0007] To address the aforementioned technical problems, this invention provides a radar interference identification method based on binary depthwise separable convolution, comprising the following steps:

[0008] Step S1: Based on several radar jamming signal models, obtain the echo signal and jamming signal corresponding to the jamming signal, calculate the mutual ambiguity function (CAF) contour map and range-Doppler (RD) spectrum map respectively, and use the CAF contour map as the spatial stream input and the RD spectrum map as the temporal stream input to construct a dual-stream dataset for training.

[0009] Step S2: Construct a dual-stream ARR-BiNet network model and train it using the dual-stream dataset;

[0010] The dual-stream ARR-BiNet network model includes two identical input branches, which receive CAF contour plots and RD spectral plots, respectively.

[0011] Each branch sequentially passes through a convolutional layer, a depthwise separable convolutional layer, at least one skip connection layer, a depthwise separable convolutional layer, a convolutional layer, and a flattening layer for feature extraction;

[0012] Both the convolutional layers and the depthwise separable convolutional layers undergo binary activation after batch normalization.

[0013] Step S3: Input the CAF contour map and RD spectrum of the signal to be detected into the trained ARR-BiNet network model to perform the interference signal classification task.

[0014] Preferably, in step S2, when training the dual-stream ARR-BiNet network model, an adaptive scaling factor based on the maximum entropy principle is introduced during the forward propagation process to scale the binarized weights.

[0015] Preferably, the adaptive scaling factor The expression is:

[0016] ;

[0017] In the formula, and These represent the number of neurons in the previous layer and the next layer, respectively.

[0018] Preferably, in step S2, during the training of the dual-stream ARR-BiNet network model, a learning rate compensation mechanism is used to compensate the adaptive scaling factor to dynamically adjust the learning rate during the parameter update process.

[0019] Preferably, the learning rate compensation mechanism The expression is:

[0020] ;

[0021] ;

[0022] In the formula, The initial learning rate is E; E represents the training period. Indicates the final learning rate; This represents the learning rate compensation coefficient.

[0023] Preferably, in step S2, when training the dual-stream ARR-BiNet network model, an adaptive pruning constraint is applied to the network weights during backpropagation. The pruning range is dynamically set based on the adaptive scaling factor, and real-valued weights exceeding the pruning range are pruned.

[0024] Preferably, the expression for the adaptive clipping constraint is:

[0025] ;

[0026] in Indicates the clipping weight; express Layer weights; β = α; This represents the clipping function, which applies to the range -β to β. Cut it.

[0027] Preferably, the skip connection layer includes a 1×1 binary convolutional layer, a 3x3 binarized depthwise separable convolutional layer, and a binarized activation layer;

[0028] The original input tensor is processed by a 1×1 binary convolutional layer and a 3×3 binarized depthwise separable convolutional layer. The outputs processed by the 1×1 binary convolutional layer and the outputs processed by the 3×3 binarized depthwise separable convolutional layer are added element-wise. Then, the outputs processed by the binarized activation layer and the 1×1 binary convolutional layer are added element-wise to the original input tensor.

[0029] Preferably, the skip connection layer is configured as two.

[0030] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program, wherein the computer program is stored in the memory and configured to be executed by the processor to implement the above-described method.

[0031] The beneficial effects of this invention include at least the following: By constructing a dual-stream data structure with mutually fuzzy function (CAF) contour maps and range-Doppler (RD) spectra as inputs, this invention fully utilizes the coupling characteristics of interference signals in the spatial and time-frequency domains. The CAF contour maps can characterize the distribution of interference targets in the time-delay-Doppler dimension, while the RD spectra reflect the dynamic evolution of interference with distance and velocity. Through this dual-stream input approach, the model can capture the deep semantic features of interference signals from a complementary perspective, significantly improving the ability to identify complex interference patterns.

[0032] In terms of network structure, the ARR-BiNet model proposed in this invention integrates a bi-branch symmetric structure and a skip connection mechanism to ensure the preservation and fusion of information between spatial and temporal flows. Simultaneously, batch normalization and binary activation mechanisms are introduced into the convolutional and depthwise separable convolutional modules, enabling the model to significantly reduce computational complexity and parameter redundancy while maintaining strong expressive power.

[0033] To address the severe information loss problem in traditional binary neural networks, this invention further introduces an adaptive real-value preservation mechanism. By introducing a scaling factor, the feature distribution after binarization is made closer to the original real-valued network, effectively alleviating the expression degradation problem caused by binary approximation. Combined with a skip connection structure, it enhances feature cross-layer transfer and residual information utilization, improving overall recognition stability.

[0034] In summary, this invention not only significantly outperforms traditional single-modal recognition methods in terms of interference recognition accuracy, but also possesses high computational efficiency and model deployment flexibility, demonstrating good engineering practical value and promising prospects for widespread application. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the dual-stream ARR-BiNet network model structure according to an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the binary neural network convolution process according to an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of a traditional 1-bit skip connection layer;

[0038] Figure 4 This is a schematic diagram of the logic flow of the RRS connection structure in Embodiment 3 of the present invention;

[0039] Figure 5 This is a schematic diagram of the network structure of the RRS connection structure in Embodiment 3 of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0041] Example 1

[0042] like Figure 1 As shown, this embodiment of the invention provides a radar interference identification method based on binary depthwise separable convolution, including the following steps:

[0043] Step S1: Based on several radar jamming signal models, obtain the echo signal and jamming signal corresponding to the jamming signal, calculate the mutual ambiguity function (CAF) contour map and range-Doppler (RD) spectrum map respectively, and use the CAF contour map as the spatial stream input and the RD spectrum map as the temporal stream input to construct a dual-stream dataset for training.

[0044] In this embodiment of the invention, to more comprehensively analyze the characteristics of the interference signals, the cross-ambiguity function (CAF) contour lines and range-Doppler (RD) spectral images of the above nine types of interference signals were extracted and used as input to the two-stream structure. The CAF domain and RD domain allow the model to learn the characteristics of the interference signals from spatial and temporal perspectives, thereby capturing supplementary information.

[0045] CAF (Concurrent Array) is used to describe the correlation between two signals under time delay and Doppler shift. It is a key tool in radar signal processing, used to analyze the resolution capability of target detection systems. The general form of CAF is expressed as follows: Represented as:

[0046] (1)

[0047] In the formula, where and These represent the echo signal and the interference signal, respectively. Indicates a time delay; It is the Doppler frequency shift parameter; express Complex conjugates with time delay.

[0048] This equation calculates the correlation between two signals under a given time delay and Doppler shift through integration. and When both are zero, the peak of the function is usually the maximum value, indicating that the two signals are perfectly correlated. With... and As the signal intensity increases, the peak value typically decreases, reflecting a reduction in the correlation between signals. This characteristic makes the Coherent Accuracy Filtering (CAF) an effective tool for analyzing the signal similarity and resolution capabilities of radar systems. In this study, the CAF was calculated using coherently accumulated interference signals and transmitted radar signals. By coherently accumulating the signals, the accumulated CAF helps reduce the impact of random noise, thereby making target features more prominent and improving detection accuracy.

[0049] The RD method is a signal processing technique used in radar systems to extract range and Doppler information of a target. This involves performing a two-dimensional Fourier transform on the received radar echo: one along the fast time axis (range dimension) to calculate the target's range, and another along the slow time axis (Doppler dimension) to determine its relative velocity. This transform is crucial for identifying the position and motion of targets in a radar scene.

[0050] CAF contour maps and RD spectrograms of nine types of interfering signals were used as input to the two-stream architecture. CAF contour maps were considered part of the spatial stream. This is because contour maps can reveal the combination of correlations and spatial hierarchies of interfering signals macroscopically, similar to how they reflect terrain features. By reducing the dimensionality of the CAF using contour lines, more representative feature parameters that enhance recognition performance can be extracted. Conversely, RD spectrograms were considered part of the temporal stream because they represent the temporal dynamics of the signal by showing information about target distance and velocity. In this two-stream architecture, the integration of spatial and temporal streams allows the model to learn the characteristics of interfering signals from both spatial and temporal perspectives. This dual approach enhances the model's generalization ability and improves its overall performance in recognizing and classifying interfering patterns. By capturing complementary information through these different dimensions, the architecture provides a comprehensive understanding of the signal, resulting in more robust and accurate results.

[0051] Table 1 lists the simulation parameter settings.

[0052] Table 1

[0053] Step S2: Construct a dual-stream ARR-BiNet network model and train it using a dual-stream dataset.

[0054] Specifically, the dual-stream ARR-BiNet model proposed in this embodiment of the invention includes two independent input branches, which respectively receive CAF contour maps and RD spectral maps, such as... Figure 1As shown, Bi-DS Conv is a depthwise separable convolution after binarization, Bi-Conv is a standard convolution after binarization, MaxPooling represents max pooling, BatchNorm represents batch normalization, Bi-Act represents an activation layer after binarization, Flatten represents a flattened layer, and Dense represents a fully connected layer.

[0055] Each branch processes a different dataset to capture different features or perspectives, ensuring the extraction of complementary information. The network uses Bi-DS Conv as the main connection and skip connection layers. These components retain essential features while maintaining a lightweight architecture. The two-stream ARR-BiNet network model includes two identical input branches that receive CAF contour maps and RD spectral maps, respectively. Each branch sequentially passes through a convolutional layer, a depthwise separable convolutional layer, at least one skip connection layer, a depthwise separable convolutional layer, a convolutional layer, and a flattening layer for feature extraction; Both convolutional and depthwise separable convolutional layers undergo binary activation after batch normalization.

[0056] Binarization is applied throughout the network, including convolutional layers and activation functions. Furthermore, except for the last layer, all layers use binary activation after a batch normalization (BN) layer, allowing the loss function to make decisions. This converts weights and activations into binary values, significantly reducing memory usage and accelerating inference while maintaining model performance.

[0057] Bi-DS Conv, a binarized depthwise separable convolution, plays a crucial role in improving computational efficiency without compromising performance. By decomposing convolution into depthwise convolution (DW-Conv) and pointwise convolution (PW-Conv), Bi-DS Conv significantly reduces the computational complexity within each branch while maintaining accuracy.

[0058] Compared to standard convolutions, DS-Conv significantly reduces parameter counting and computational cost. Traditional binary networks apply binary values ​​to standard convolutions, while we use Bi-DS Conv. By representing weights as 1-bit values, Bi-DS Conv reduces memory usage by 32 times. The XNOR operation replaces floating-point multiplication, enabling more efficient bit-by-bit computation and further reducing computational complexity and inference time.

[0059] Step S3: Input the CAF contour map and RD spectrum of the signal to be detected into the trained ARR-BiNet network model to perform the interference signal classification task.

[0060] Example 2

[0061] Based on Example 1, this embodiment proposes a training method for the dual-stream ARR-BiNet network model in Example 1.

[0062] Binary Neural Networks (BNNs) suffer from information loss during training. Forward sign function, backward gradient approximation, and parameter updates all contribute to this loss, impacting network accuracy. The training method proposed in this embodiment preserves crucial information during training, thereby improving model accuracy. The basic principles of network binarization are first briefly described. Then, the three adaptation strategies proposed in this embodiment are explained to mitigate information loss during the forward propagation, backward propagation, and parameter update stages of the training process.

[0063] In full-precision FP-CNN, given input, activation... and weight Output It is calculated through convolution operations, as shown in the following formula.

[0064] (2)

[0065] In the formula, It is a nonlinear function. This represents the convolution operation.

[0066] To obtain the binarized function Using symbolic functions Indicate:

[0067] (3)

[0068] Where x represents the floating-point parameter, including the floating-point weight. and activation , Indicates the inclusion of binary weights and activation The binary number. In forward propagation, vector multiplication can be reconstructed:

[0069] (4)

[0070] in This represents element-wise vector multiplication using bitwise XOR bit counting operations. Figure 2 This explains the convolution process in BNNs.

[0071] Similar to training FP networks, BNNs can use gradient-based backpropagation to update parameters. However, quantization functions, such as sign functions, require different approaches. The gradient is non-differentiable, causing gradient vanishing in some functions, such as the sign function, which is almost zero everywhere. This renders traditional gradient-based backpropagation ineffective for updating binary weights. To address this issue, "pass-through estimators" are typically used to train binary models, with gradient propagation approximated using the Hardtanh function.

[0072] (5)

[0073] In the formula, This represents the clipping function.

[0074] 1) Adaptive scaling weights in forward propagation

[0075] In BNNs, the parameters are restricted to two values, which can lead to information loss. To address this issue, we introduce adaptive scaling to approximate the statistical properties of the original weights and reduce information loss during forward propagation.

[0076] For random variables with Bernoulli distribution probability mass function Defined as:

[0077] (6)

[0078] Where p is the probability of taking the value +1. Each element in can be considered a sample of b. entropy It is given by the following formula.

[0079] (7)

[0080] Solving for the derivative of the above equation yields the following equation:

[0081] (8)

[0082] From the analysis of the second derivative of the above equation, we can conclude that when... When the entropy reaches its maximum value, maximizing the information entropy maintains the diversity of quantization parameters, reduces information loss, and improves the representational power of the BNN. This indicates that the values ​​of the binary function should be uniformly distributed.

[0083] To minimize information loss during binarization, this embodiment introduces a continuous scaling factor. The scaled binary variance Approximate full-precision variance This method statistically makes the binarized weights... Compared with the original weights The distribution of the weights is kept consistent in order to preserve the statistical properties of the original network weights as much as possible.

[0084] Construct the following constraint equations :

[0085] (9)

[0086] Considering that variance is a scalar value, expanding the above equation yields:

[0087] (10)

[0088] In the formula, Represents the binary variance, since ,therefore The above formula is written as In order to minimize , need to obtain The derivative of is then set to zero.

[0089] (11)

[0090] Solving for the given information, we get:

[0091] (12)

[0092] Considering It is a positive value. The optimal solution is:

[0093] (13)

[0094] Therefore, the optimal scaling factor yes The standard deviation of makes it possible to... Scaling and Similarly, it achieves the effect of simulating continuous weights.

[0095] In traditional neural networks, weight initialization relies on continuous distributions, such as uniform or normal distributions, to ensure relatively stable activation values ​​and gradient distributions between layers, thereby promoting effective learning. However, binarization introduces significant nonlinearity and discontinuity, causing the loss of most of the information generated by the initialization method. Therefore, the initial purpose of weight initialization cannot be fully preserved. To mitigate this limitation, the Xavier uniform distribution initialization method is used. Initialization process, making approximate .

[0096] based on The principle of entropy maximization It should follow a uniform distribution , and Both represent random variables, and their variances are given by the following formula:

[0097] (14)

[0098] Under Xavier unified initialization Initialize to:

[0099] (15)

[0100] in and These represent the number of neurons in the previous and next layers, respectively. Substituting these values ​​into the above formula, the variance... It becomes:

[0101] (16)

[0102] The resulting scaling factor for:

[0103] (17)

[0104] This method considers the number of input and output units in each layer and dynamically calculates... The value of . This ensures The scaling is neither arbitrary nor fixed. Instead, it is optimized to meet the specific needs of each layer, promoting faster convergence and improving model stability during training.

[0105] 2) Adaptive learning rate for parameter updates

[0106] Traditional binary networks use a fixed scaling factor, neglecting the implicit parameter scaling effect. This embodiment rigorously analyzes this phenomenon through derivative propagation. The binarized weights can be reformulated as:

[0107] (18)

[0108] In a strictly mathematical sense, when When, the sign function is not differentiable, when hour However, in actual backpropagation, the STE approximation of the pass-through estimator in equation (5) is used:

[0109] therefore,

[0110] (19)

[0111] in, For indicator functions, it means The value is 1 if it is true, and 0 otherwise.

[0112] The gradient of the floating-point weights is then:

[0113] (20)

[0114] Update volume compared to regular networks In comparison, the order of magnitude is:

[0115] (twenty one)

[0116] It is evident that parameter updates suffer from implicit scaling issues, due to... This results in the gradient magnitude being implicitly reduced. If the original learning rate η is used directly, the parameter update step size will decrease as α increases, leading to slow convergence. Therefore, a compensation mechanism can be used to ensure that the parameter update magnitudes of the binary network and the full-precision network are consistent, satisfying the convergence condition of gradient descent. This is achieved by introducing a learning rate compensation coefficient. This causes the parameter update amount of the binary network, i.e., the floating-point weights, to be... Redefined as:

[0117] (twenty two)

[0118] in, The initial learning rate decays exponentially to the final learning rate over the training period E. .

[0119] Therefore, the weights of each layer were optimized, and the learning rate was scaled to match its initialization, resulting in smoother training and potentially better performance.

[0120] 3) Adaptive pruning constraints in backpropagation

[0121] Weight initialization provides a fixed starting point for each network layer; however, the weights are continuously updated throughout the training process. Although W does not directly participate in gradient calculation during backpropagation, it is based on the formula shown in Equation (20). The gradient is updated.

[0122] (twenty three)

[0123] in It is the updated weight; L is the updated learning rate; L is the loss function.

[0124] Traditional clipping methods limit real-valued weights to Within a certain range, this limits the gradient flow of weights beyond this range and can lead to the loss of important feature information, especially in sparse BNNs. Therefore, setting an appropriate weight range is crucial. A range that is too large will waste the representational power of the quantized tensor, while a range that is too small will cause floating-point truncation, thus degrading performance.

[0125] To address this issue, this embodiment employs a dynamically constrained adaptive pruning constraint to ensure stable gradient updates. This constraint prunes the weights within the following range:

[0126] (twenty four)

[0127] in This represents the clipping weight, where β matches α (i.e., β=α), ensuring that the weight range is consistent with its proportional value.

[0128] Pruning constraints are applied after each weight update to preserve valuable information during training while preventing weights from exceeding their expected range. This reduces noise and error and improves stability. Combined with weight initialization and pruning, this ensures consistent and efficient training. This adaptive adjustment modifies the weight range to suit different network layers, thereby enhancing the model's ability to capture complex patterns.

[0129] Example 3

[0130] In traditional binary networks, the output of each convolutional layer passes through a Batch Normalization (BN) layer to produce real values. These values ​​are usually immediately binary, leading to information loss.

[0131] To alleviate this problem, methods such as... are commonly used. Figure 3 The 1-bit skip connection layer shown bypasses the binarization step to preserve real-valued information. This embodiment, based on the traditional skip connection layer technique, adds a 1×1 convolutional block to the bypass path and replaces it with a 3×3 Bi-DS-Conv. Figure 3 The 3×3 binary standard convolution in the model further enhances this structure, as shown in the figure. Figure 4 As shown, this is simply referred to as RRS connection, which improves efficiency and representation capability. The 1×1Bi-Conv, 3×3Bi-Conv, and 3×3Bi-DS Conv in the figure each contain a binary convolutional layer and a BN layer.

[0132] In this embodiment, the RRS connection structure in the binary network adopts a combination of Bi-DS Conv and Bi-Conv. For example... Figure 5As shown, the input tensor first undergoes a 1×1 binary convolution to reduce dimensionality and introduce non-linearity. Simultaneously, the real-valued features from the BN layer output are directly forwarded to the append nodes of the next block. Then, the tensor undergoes a 3×3 binary DW-Conv, followed by BN, and the final output is element-wise added to the real-valued features from the previous BN layer before activation.

[0133] Compared to traditional 1-bit skip connection layers, this structure significantly improves the model's representational power, with negligible computational cost for element-wise addition and no additional memory overhead. Compared to the traditional ResNet design, the proposed one-layer-per-block design achieves greater efficiency by breaking the assumption that the traditional shortcut must bypass at least two layers.

[0134] The process continues with an additional 1×1 Bi-Conv and bypass connections. By adding shortcuts that are fused with the main path, this configuration enhances the information flow and reduces the sparsity typically observed in binary networks. This setup forms a complete RRS connection, which is applied to two identical blocks for deeper feature extraction.

[0135] It should be noted that in this embodiment, at least one skip connection layer is used, and two layers are optimal.

[0136] Example 4

[0137] In order to verify the effectiveness of the proposed method, this embodiment uses the interference signals shown in Table 1 to conduct experiments to create CAF contour maps and RD spectrum datasets.

[0138] For the training set, the interference noise ratio (IRR) is varied from -5dB to 5dB in increments of 2.5dB. In each... Below, 1000 training samples across 9 categories were generated by randomly varying the initial distance over distances of 3-8 km, with 50 samples for each signal type. For the test set, the JNR was varied from -4 dB to 4 dB in 2 dB increments. In each... With initial distances randomly varying between 3 and 8 km, we generated 50 test samples for each signal type, forming 45 test sets containing 200 samples each, where the JNR was fixed. Furthermore, CAF contour maps and RD spectra of 0-5 real target echoes were added to both training and test samples, with randomized signal-to-noise ratios and locations.

[0139] Ablation experiments were conducted using the sample distribution and corresponding training parameters described above, and the results are shown in Table 2.

[0140] Table 2

[0141] To evaluate the impact of dual-stream input, each branch was removed independently, creating a single-stream network with a single CAF contour plot and a single RD spectrogram. For comparison with the FP network, all binarization processes were removed, and the model was converted to FP.

[0142] Furthermore, to verify the effectiveness of the three adaptive strategies, the network structure was kept constant while disabling each strategy. 1) Adaptive scaling weights were removed in forward propagation; 2) Adaptive clipping constraints were disabled in backpropagation with a static clipping range of [-1, 1]; 3) 1e -2 1e -3 and 1e -4 A fixed learning rate was used instead of an adaptive learning rate. Finally, to independently assess the impact of the network structure, all skip connections, including RRS connections, were removed, while only the main information paths were retained.

[0143] Table 2 shows that the parameters and MAC of the single-stream network are only 50% of those of the method proposed in this invention. However, the single-stream network performs poorly on all five evaluation metrics. Specifically, compared with the method in this embodiment, the accuracy, precision, F1 score, recall, and kappa of the single-stream CAF contour network are reduced by 0.36%, 0.34%, 0.35%, 0.36%, and 0.40%, respectively. The results of the single-stream RD network are even lower than those of the single-stream CAF contour network, highlighting the crucial role of dual-stream input in fusing multi-domain features to improve network performance.

[0144] The FP network achieves similar performance to the method in this embodiment of the invention, but at the cost of significantly higher parameters. Although the MACs are the same, the bitwise operations of the proposed method are faster. This demonstrates that the binarization technique in this embodiment of the invention can match the performance of FP while significantly reducing memory usage.

[0145] When adaptive weight scaling is removed during forward propagation, the model's performance degrades. The same phenomenon was observed when the weights were scaled down to the range [-1, 1], highlighting the improvement effect of the adaptive constraints in both forward and backward propagation in this embodiment. Furthermore, using a fixed learning rate instead of an adaptive rate leads to poor performance, even at 1×10⁻⁶. -3 At the learning rate, compared with the method proposed in this embodiment, the accuracy, F1 score and recall also decreased by 0.36%, the precision decreased by 0.35%, and the Kappa decreased by 0.40%.

[0146] Similarly, when this embodiment retains only the main path and removes all hop connections, the results are not satisfactory, highlighting the necessity of the proposed hop connection structure for achieving optimal network performance.

[0147] As can be seen from the above analysis, the dual-stream ARR-BiNet structure for multi-pulse interference identification in Example 1, by utilizing dual-stream input, allows ARR-BiNet to simultaneously capture spatial and temporal interference features, thereby achieving complementary information extraction. This improves training stability and recognition accuracy, balancing high recognition accuracy with reduced computational load, making it an ideal choice for deployment on resource-constrained platforms. Further improvements in Examples 2 and 3 further enhance the model's performance.

[0148] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0149] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A radar interference identification method based on binary depthwise separable convolution, characterized in that: Includes the following steps: Step S1: Based on several radar jamming signal models, obtain the echo signal and jamming signal corresponding to the jamming signal, calculate the mutual ambiguity function (CAF) contour map and range-Doppler (RD) spectrum map respectively, and use the CAF contour map as the spatial stream input and the RD spectrum map as the temporal stream input to construct a dual-stream dataset for training. Step S2: Construct a dual-stream ARR-BiNet network model and train it using the dual-stream dataset; The dual-stream ARR-BiNet network model includes two identical input branches, which receive CAF contour plots and RD spectral plots, respectively. Each branch sequentially passes through a convolutional layer, a depthwise separable convolutional layer, at least one skip connection layer, a depthwise separable convolutional layer, a convolutional layer, and a flattening layer for feature extraction; Both the convolutional layers and the depthwise separable convolutional layers undergo binary activation after batch normalization. Step S3: Input the CAF contour map and RD spectrum of the signal to be detected into the trained ARR-BiNet network model to perform the interference signal classification task.

2. The radar interference identification method based on binary depthwise separable convolution according to claim 1, characterized in that: In step S2, during the training of the dual-stream ARR-BiNet network model, an adaptive scaling factor based on the maximum entropy principle is introduced to scale the binarized weights during the forward propagation process.

3. The radar interference identification method based on binary depthwise separable convolution according to claim 2, characterized in that: The adaptive scaling factor The expression is: ； In the formula, and These represent the number of neurons in the previous layer and the next layer, respectively.

4. The radar interference identification method based on binary depthwise separable convolution according to claim 3, characterized in that: In step S2, during the training of the dual-stream ARR-BiNet network model, a learning rate compensation mechanism is used to compensate the adaptive scaling factor to dynamically adjust the learning rate during the parameter update process.

5. The radar interference identification method based on binary depthwise separable convolution according to claim 4, characterized in that: The learning rate compensation mechanism The expression is: ； ； In the formula, The initial learning rate is E; E represents the training period. Indicates the final learning rate; This represents the learning rate compensation coefficient.

6. The radar interference identification method based on binary depthwise separable convolution according to claim 3, characterized in that: In step S2, during the training of the dual-stream ARR-BiNet network model, an adaptive pruning constraint is applied to the network weights during backpropagation. The pruning range is dynamically set based on the adaptive scaling factor, and real-valued weights exceeding the pruning range are pruned.

7. The radar interference identification method based on binary depthwise separable convolution according to claim 6, characterized in that: The expression for the adaptive clipping constraint is: ； in Indicates the clipping weight; express Layer weights; β=α; This represents the clipping function, which applies to the range -β to β. Cut it.

8. The radar interference identification method based on binary depthwise separable convolution according to claim 1, characterized in that: The skip connection layer includes a 1×1 binary convolutional layer, a 3x3 binarized depthwise separable convolutional layer, and a binarized activation layer; The original input tensor is passed through a 1×1 binary convolutional layer and a 3×3 binarized depthwise separable convolutional layer; The outputs after processing by a 1×1 binary convolutional layer and the outputs after processing by a 3×3 binarized depthwise separable convolutional layer are added element-wise. Then, the outputs after processing by a binarized activation layer and a 1×1 binary convolutional layer are added element-wise to the original input tensor.

9. A radar interference identification method based on binary depthwise separable convolution according to claim 8, characterized in that: The skip connection layer is configured as two.

10. An electronic device, comprising: A memory, a processor, and a computer program, characterized in that: the computer program is stored in the memory and configured to be executed by the processor to implement the method of any one of claims 1 to 9.

Images