技术领域Technical Field
本发明属于雷达信号处理技术领域，具体涉及一种基于多损失特征自校准网络的雷达有源干扰识别方法。The invention belongs to the technical field of radar signal processing, and in particular relates to a radar active interference identification method based on a multi-loss feature self-calibration network.
背景技术Background Art
日益复杂的电磁环境严重威胁了雷达的探测效能与生存安全。由于现有的雷达抗干扰技术大多数针对特定类型的干扰，只有准确地识别出干扰样式，才能采取相应有效的反制措施。因此，研究雷达干扰识别算法能够为后续雷达抗干扰提供先验信息，具有重要的应用价值。The increasingly complex electromagnetic environment has seriously threatened the detection efficiency and survival safety of radars. Since most of the existing radar anti-interference technologies are targeted at specific types of interference, only by accurately identifying the interference pattern can corresponding effective countermeasures be taken. Therefore, studying radar interference recognition algorithms can provide prior information for subsequent radar anti-interference and has important application value.
传统雷达干扰识别算法需要人工分析和提取各类特征，导致通用性差、泛化能力弱。而且基于人工筛选特征的方法容易受到干噪比和其它干扰信号参数变化的影响。Traditional radar interference recognition algorithms require manual analysis and extraction of various features, resulting in poor versatility and generalization. Moreover, methods based on manual feature screening are easily affected by changes in the interference-to-noise ratio and other interference signal parameters.
深度卷积神经网络(Deep Convolutional Neural Network，DCNN)由于具有自适应特征提取与表征能力，在图像处理领域获得了广泛应用。目前，已有一些经典的基于深度学习的雷达有源干扰识别方法。深度融合卷积网络(Deep Fusion Convolution NeuralNetwork，DFCNN)结构由一维卷积神经网络、二维卷积神经网络和融合网络三个子网络组成，四个一维卷积神经网络分别提取原始雷达回波中干扰取实部、虚部、相位和幅度后的高维特征，二维卷积神经网络提取干扰的时频特征，两部分特征拼接后输入融合网络进行进一步的特征融合提取，另外还提出软标签平滑用于缓解过拟合。在此基础上，西安电子科技大学的Lv等人提出了一种基于迁移学习的加权集成CNN(Weighted Ensemble CNN withTransfer Learning，WECNN-TL)的雷达有源欺骗干扰识别算法；首先，该算法网络训练的基准数据集由仿真和实测干扰的时频谱混合构成，其次，为了充分挖掘干扰信号的潜在信息，对干扰时频谱提取实部、虚部、模和相位后，组合成15个子数据集；该网络利用引导聚类算法(Bootstrap aggregating，Bagging)的思想，设计了15个子分类器，分别挖掘各子数据集的结构特征并做出个体预测的结果，最终利用一个加权投票算法来获得集合模型的整体预测结果，达到进一步提升测试阶段雷达干扰识别精度的效果。干扰识别网络(JammingRecognition Network，JRNet)利用非对称卷积块(Asymmetric Convolution Block，ACB)在不增加额外计算量的前提下，能够增强存在旋转变形和细微差异等情况下的干扰识别的鲁棒性。迁移学习利用从源域上学到的知识帮助目标域的学习任务，AlexNet在数据集ImageNet上预训练后，分别调整输入和输出层为含干扰雷达回波的时频图和干扰识别的结果。识别卷积神经网络(Recognition Convolution Neural Network，RCNN)利用OS-CFAR从回波信号的时频图中测量干扰参数，然后提取出干扰，送到预先训练的CNN网络中进行分类。Deep Convolutional Neural Network (DCNN) has been widely used in the field of image processing due to its adaptive feature extraction and representation capabilities. At present, there are some classic radar active interference recognition methods based on deep learning. The Deep Fusion Convolution Neural Network (DFCNN) structure consists of three sub-networks: a one-dimensional convolutional neural network, a two-dimensional convolutional neural network, and a fusion network. The four one-dimensional convolutional neural networks extract the high-dimensional features of the real part, imaginary part, phase, and amplitude of the interference in the original radar echo. The two-dimensional convolutional neural network extracts the time-frequency features of the interference. The two parts of the features are spliced and input into the fusion network for further feature fusion extraction. In addition, soft label smoothing is proposed to alleviate overfitting. On this basis, Lv et al. from Xidian University proposed a radar active deception jamming recognition algorithm based on transfer learning weighted ensemble CNN (WECNN-TL); first, the benchmark data set for network training of the algorithm consists of a mixture of simulated and measured jamming time-frequency spectra; second, in order to fully explore the potential information of jamming signals, the real part, imaginary part, modulus and phase of the jamming time-frequency spectra are extracted and combined into 15 sub-datasets; the network uses the idea of bootstrap aggregating (Bagging) to design 15 sub-classifiers, respectively mining the structural features of each sub-dataset and making individual prediction results, and finally using a weighted voting algorithm to obtain the overall prediction results of the ensemble model, so as to further improve the accuracy of radar jamming recognition in the test phase. The Jamming Recognition Network (JRNet) uses asymmetric convolution blocks (ACB) to enhance the robustness of jamming recognition in the presence of rotational deformation and subtle differences without increasing the amount of additional calculations. Transfer learning uses the knowledge learned from the source domain to help the learning tasks in the target domain. After pre-training on the ImageNet dataset, AlexNet adjusts the input and output layers to the time-frequency diagram of the radar echo containing interference and the interference identification results respectively. The Recognition Convolution Neural Network (RCNN) uses OS-CFAR to measure the interference parameters from the time-frequency diagram of the echo signal, and then extracts the interference and sends it to the pre-trained CNN network for classification.
利用深度学习理论实现雷达有源干扰识别，克服了传统识别方法人工依赖性和鲁棒性差的缺点。但是，现有的基于深度学习的雷达有源干扰识别方法存在着特征参数对干扰样式敏感、识别的干扰类型有限、鲁棒性不足等问题，尤其在面对干扰间可辨识特征差异小和干扰样式繁多的识别任务时，对干扰特征提取与聚合能力降低。Using deep learning theory to realize radar active jamming recognition overcomes the shortcomings of traditional recognition methods, such as artificial dependence and poor robustness. However, existing radar active jamming recognition methods based on deep learning have problems such as feature parameters being sensitive to jamming patterns, limited jamming types being recognized, and insufficient robustness. In particular, when faced with recognition tasks with small differences in identifiable features between jamming and a variety of jamming patterns, the ability to extract and aggregate jamming features is reduced.
发明内容Summary of the invention
为了解决现有技术中存在的上述问题，本发明提供了一种基于多损失特征自校准网络的雷达有源干扰识别方法。本发明要解决的技术问题通过以下技术方案实现：In order to solve the above problems existing in the prior art, the present invention provides a radar active interference identification method based on a multi-loss feature self-calibration network. The technical problem to be solved by the present invention is achieved through the following technical solutions:
本发明实施例提供了一种基于多损失特征自校准网络的雷达有源干扰识别方法，包括步骤：The embodiment of the present invention provides a radar active interference identification method based on a multi-loss feature self-calibration network, comprising the steps of:
获取包括雷达有源干扰信号的待识别时频谱图数据；Acquire time-frequency spectrum data including radar active interference signals to be identified;
将所述待识别时频谱图数据输入训练好的多损失特征自校准网络进行干扰类型识别，得到识别分类结果；Inputting the to-be-identified spectrum data into a trained multi-loss feature self-calibration network to identify the interference type, and obtaining an identification and classification result;
其中，所述训练好的多损失特征自校准网络包括依次连接的特征提取模块、特征映射模块和重构模块，所述训练好的多损失特征自校准网络为利用混合损失函数对多损失特征自校准网络的参数进行训练更新得到，所述混合损失函数由训练集样本标签传播的交叉熵损失、所述特征映射模块的聚类损失和所述重构模块的均方误差损失确定，按照如下步骤进行干扰类型识别：The trained multi-loss feature self-calibration network includes a feature extraction module, a feature mapping module and a reconstruction module connected in sequence. The trained multi-loss feature self-calibration network is obtained by training and updating the parameters of the multi-loss feature self-calibration network using a mixed loss function. The mixed loss function is determined by the cross entropy loss of the training set sample label propagation, the clustering loss of the feature mapping module and the mean square error loss of the reconstruction module. Interference type identification is performed according to the following steps:
通过所述特征提取模块对所述待识别时频谱图数据进行干扰特征自适应提取，并且自适应缩小同类别干扰特征在特征空间的距离，拉大不同类别干扰特征在特征空间的距离，得到不同类型干扰特征向量；The feature extraction module adaptively extracts interference features from the to-be-identified time spectrum data, and adaptively reduces the distance between interference features of the same category in the feature space, and increases the distance between interference features of different categories in the feature space, so as to obtain interference feature vectors of different types;
通过所述特征映射模块对所述不同类型干扰特征向量降维映射后进行分类，得到所述识别分类结果。The feature mapping module performs dimensionality reduction mapping on the interference feature vectors of different types and classifies them to obtain the identification and classification results.
在本发明的一个实施例中，所述多损失特征自校准网络的训练方法包括：In one embodiment of the present invention, the training method of the multi-loss feature self-calibration network includes:
获取包括雷达有源干扰信号和无干扰雷达回波的原始时频谱图数据集，其中，所述原始时频谱图数据集包括训练集样本、训练集样本标签、验证集样本和验证集样本标签；Acquire an original time-spectrogram data set including radar active jamming signals and non-interference radar echoes, wherein the original time-spectrogram data set includes training set samples, training set sample labels, validation set samples, and validation set sample labels;
将所述训练集样本和所述训练集样本标签输入所述多损失特征自校准网络中进行训练；Inputting the training set samples and the training set sample labels into the multi-loss feature self-calibration network for training;
根据所述训练集样本标签传播的交叉熵损失、所述特征映射模块的聚类损失以及重构模块的均方误差损失确定所述混合损失函数；Determine the hybrid loss function according to the cross entropy loss of the training set sample label propagation, the clustering loss of the feature mapping module, and the mean square error loss of the reconstruction module;
利用所述混合损失函数对所述多损失特征自校准网络的参数进行更新；Using the hybrid loss function to update the parameters of the multi-loss feature self-calibration network;
利用所述验证集样本和所述验证集样本标签对每轮训练结束后的网络进行选择，将识别准确率最高的模型作为所述训练好的多损失特征自校准网络。The network after each round of training is selected using the validation set samples and the validation set sample labels, and the model with the highest recognition accuracy is used as the trained multi-loss feature self-calibration network.
在本发明的一个实施例中，所述特征提取模块包括第一特征自校准卷积块、第二特征自校准卷积块、第三特征自校准卷积块和第四特征自校准卷积块，其中，In one embodiment of the present invention, the feature extraction module includes a first feature self-calibration convolution block, a second feature self-calibration convolution block, a third feature self-calibration convolution block and a fourth feature self-calibration convolution block, wherein:
所述第一特征自校准卷积块、第二特征自校准卷积块、第三特征自校准卷积块和第四特征自校准卷积块依次连接，用于对所述待识别时频谱图数据依次进行特征自校准卷积，得到所述不同类型干扰特征向量。The first feature self-calibration convolution block, the second feature self-calibration convolution block, the third feature self-calibration convolution block and the fourth feature self-calibration convolution block are connected in sequence to perform feature self-calibration convolution on the time spectrum data to be identified in sequence to obtain the different types of interference feature vectors.
在本发明的一个实施例中，所述第一特征自校准卷积块、所述第二特征自校准卷积块、所述第三特征自校准卷积块和所述第四特征自校准卷积块的结构相同，均包括第一卷积块、第二卷积块、第三卷积块、通道特征自校准模块、空间特征自校准模块、升维模块、第一相加模块和第一最大池化层，其中，In one embodiment of the present invention, the first feature self-calibration convolution block, the second feature self-calibration convolution block, the third feature self-calibration convolution block and the fourth feature self-calibration convolution block have the same structure, and all include a first convolution block, a second convolution block, a third convolution block, a channel feature self-calibration module, a spatial feature self-calibration module, a dimensionality increase module, a first addition module and a first maximum pooling layer, wherein:
所述第一卷积块、所述第二卷积块、所述第三卷积块依次连接，用于对特征自校准卷积块的输入特征图依次进行卷积处理，得到第三卷积块的输出特征图；The first convolution block, the second convolution block, and the third convolution block are connected in sequence, and are used to sequentially perform convolution processing on the input feature map of the feature self-calibration convolution block to obtain the output feature map of the third convolution block;
所述通道特征自校准模块用于对所述第三卷积块的输出特征图进行通道特征自校准，得到通道特征自校准特征图；The channel feature self-calibration module is used to perform channel feature self-calibration on the output feature map of the third convolution block to obtain a channel feature self-calibration feature map;
所述空间特征自校准模块用于对所述通道特征自校准特征图进行空间特征自校准，得到空间特征自校准特征图；The spatial feature self-calibration module is used to perform spatial feature self-calibration on the channel feature self-calibration feature map to obtain a spatial feature self-calibration feature map;
所述升维模块用于对特征自校准卷积块的输入特征图进行升维操作，得到升维特征图；The dimension-increasing module is used to perform a dimension-increasing operation on the input feature map of the feature self-calibration convolution block to obtain a dimension-increasing feature map;
所述第一相加模块用于将所述升维特征图加在所述空间特征自校准特征图上，得到相加特征图；The first adding module is used to add the dimension-increased feature map to the spatial feature self-calibration feature map to obtain an added feature map;
所述第一最大池化层用于对所述相加特征图进行下采样，得到特征自校准卷积块的输出特征图：The first maximum pooling layer is used to downsample the added feature map to obtain an output feature map of the feature self-calibration convolution block:
； ;
其中，表示升维操作，表示提取特征，表示第个特征自校准卷积块的输入特征图，表示第个特征自校准卷积块的权重参数，表示最大池化。in, represents the dimension-raising operation, represents the extracted features, Indicates The input feature map of the feature self-calibration convolution block, Indicates The weight parameters of the feature self-calibration convolution block, Represents maximum pooling.
在本发明的一个实施例中，所述通道特征自校准模块包括自适应最大池化层、自适应平均池化层、多层感知机、第二相加模块、通道权重归一化模块和第一相乘模块，其中，In one embodiment of the present invention, the channel feature self-calibration module includes an adaptive maximum pooling layer, an adaptive average pooling layer, a multi-layer perceptron, a second addition module, a channel weight normalization module and a first multiplication module, wherein:
所述自适应最大池化层用于对第三卷积块的输出特征图中特征图高度和特征图宽度同时进行自适应最大池化，归纳出每一通道维上的最大值响应；The adaptive maximum pooling layer is used to simultaneously perform adaptive maximum pooling on the feature map height and feature map width in the output feature map of the third convolution block, and summarize the maximum value response on each channel dimension;
所述自适应平均池化层用于对第三卷积块的输出特征图中特征图高度和特征图宽度同时进行自适应平均池化，归纳出每一通道维上的平均响应；The adaptive average pooling layer is used to simultaneously perform adaptive average pooling on the feature map height and feature map width in the output feature map of the third convolutional block, and summarize the average response on each channel dimension;
所述多层感知机用于对每一通道维上的最大值响应依次进行特征缩放、特征还原、提取通道维信息得到最大值响应输出，并对所述每一通道维上的平均响应依次进行特征缩放、特征还原、提取通道维信息得到平均响应输出；The multilayer perceptron is used to perform feature scaling, feature restoration, and channel dimension information extraction on the maximum value response on each channel dimension in sequence to obtain a maximum value response output, and to perform feature scaling, feature restoration, and channel dimension information extraction on the average response on each channel dimension in sequence to obtain an average response output;
所述第二相加模块用于将所述最大值响应输出和所述平均响应输出进行相加融合，得到相加特征图；The second adding module is used to add and fuse the maximum value response output and the average response output to obtain an addition feature map;
所述通道权重归一化模块用于利用激活函数将所述相加特征图中通道的权重归一化，得到通道归一化权重；The channel weight normalization module is used to normalize the weights of the channels in the added feature map using an activation function to obtain channel normalization weights;
所述第一相乘模块用于将所述通道归一化权重与所述第三卷积块的输出特征图相乘得到通道特征自校准特征图：The first multiplication module is used to multiply the channel normalization weight and the output feature map of the third convolution block to obtain a channel feature self-calibration feature map:
； ;
其中，表示通道特征自校准模板，表示激活函数，表示多层感知机，表示自适应平均池化，表示自适应最大池化，表示第三卷积块的输出特征图，，表示向量空间，表示特征图通道数，表示特征图高度，表示特征图宽度。in, represents the channel feature self-calibration template, represents the activation function, represents a multi-layer perceptron, represents adaptive average pooling, represents adaptive maximum pooling, represents the output feature map of the third convolutional block, , represents a vector space, Representation feature map Number of channels, Representation feature map high, Representation feature map width.
在本发明的一个实施例中，所述空间特征自校准模块包括第二最大池化层、平均池化层、拼接模块、卷积模块、空间权重归一化模块和第二相乘模块，其中，In one embodiment of the present invention, the spatial feature self-calibration module includes a second maximum pooling layer, an average pooling layer, a splicing module, a convolution module, a spatial weight normalization module and a second multiplication module, wherein:
所述第二最大池化层用于对所述通道特征自校准特征图进行通道维最大池化，得到压缩到空间维的最大池化特征图；The second maximum pooling layer is used to perform channel-dimensional maximum pooling on the channel feature self-calibration feature map to obtain a maximum pooling feature map compressed to a spatial dimension;
所述平均池化层用于对所述通道特征自校准特征图进行通道维平均池化，得到压缩到空间维的平均池化特征图；The average pooling layer is used to perform channel-dimensional average pooling on the channel feature self-calibration feature map to obtain an average pooling feature map compressed to the spatial dimension;
所述拼接模块用于采用拼接方法将所述最大池化特征图和所述平均池化特征图进行融合，得到拼接特征图；The splicing module is used to fuse the maximum pooling feature map and the average pooling feature map using a splicing method to obtain a splicing feature map;
所述卷积模块用于对所述拼接特征图进行卷积映射，得到映射特征图；The convolution module is used to perform convolution mapping on the spliced feature map to obtain a mapping feature map;
所述空间权重归一化模块用于利用激活函数将所述映射特征图中空间的权重归一化，得到空间归一化权重；The spatial weight normalization module is used to normalize the spatial weights in the mapping feature map using an activation function to obtain spatial normalized weights;
所述第二相乘模块用于将所述空间归一化权重与所述通道特征自校准特征图相乘得到空间特征自校准特征图：The second multiplication module is used to multiply the spatial normalization weight and the channel feature self-calibration feature map to obtain a spatial feature self-calibration feature map:
； ;
其中，表示空间特征自校准模板，表示卷积权重，表示激活函数，表示平均池化，表示最大池化，表示通道特征自校准特征图，，表示向量空间，表示特征图通道数，表示特征图高度，表示特征图宽度。in, represents the spatial feature self-calibration template, represents the convolution weight, represents the activation function, represents average pooling, represents the maximum pooling, represents the channel characteristic self-calibration feature map, , represents a vector space, Representation feature map Number of channels, Representation feature map high, Representation feature map width.
在本发明的一个实施例中，所述通道特征自校准模块和空间特征自校准模块的整体公式表示为：In one embodiment of the present invention, the overall formula of the channel feature self-calibration module and the spatial feature self-calibration module is expressed as:
； ;
其中，表示空间特征自校准模块的输出特征图，表示第三卷积块的输出特征图，，表示向量空间，表示特征图通道数，表示特征图高度，表示特征图宽度，表示通道特征自校准模板，表示空间特征自校准模板，表示逐元素相乘。in, represents the output feature map of the spatial feature self-calibration module, represents the output feature map of the third convolutional block, , represents a vector space, Representation feature map Number of channels, Representation feature map high, Representation feature map width, represents the channel feature self-calibration template, represents the spatial feature self-calibration template, Represents element-wise multiplication.
在本发明的一个实施例中，所述特征映射模块包括第一全连接层和第二全连接层，其中，In one embodiment of the present invention, the feature mapping module includes a first fully connected layer and a second fully connected layer, wherein:
所述第一全连接层用于将所述不同类型干扰特征向量的展平特征投影到特征空间，得到嵌入向量；The first fully connected layer is used to project the flattened features of the different types of interference feature vectors into the feature space to obtain an embedding vector;
所述第二全连接层用于对所述嵌入向量进行分类，得到所述识别分类结果。The second fully connected layer is used to classify the embedded vector to obtain the recognition classification result.
在本发明的一个实施例中，所述重构模块包括依次连接的第三全连接层、第四非线性层、第四全连接层、第五非线性层、第五全连接层和激活函数层。In one embodiment of the present invention, the reconstruction module includes a third fully connected layer, a fourth nonlinear layer, a fourth fully connected layer, a fifth nonlinear layer, a fifth fully connected layer and an activation function layer connected in sequence.
在本发明的一个实施例中，所述混合损失函数为：In one embodiment of the present invention, the hybrid loss function is:
； ;
其中，和表示可调权重超参数，表示交叉熵损失函数，，表示批次大小，表示训练样本的真实标签，表示预测结果的概率分布，表示聚类损失函数，，表示特征向量所属类别的特征中心，表示均方误差损失函数，，表示重构像素点，表示原始图像像素点。in, and represents an adjustable weight hyperparameter, represents the cross entropy loss function, , Indicates the batch size, Represents training samples The real label, represents the probability distribution of the prediction results, represents the clustering loss function, , Represents the feature vector The characteristic center of the category to which it belongs, represents the mean square error loss function, , represents the reconstructed pixel point, Represents the original image pixel.
与现有技术相比，本发明的有益效果：Compared with the prior art, the present invention has the following beneficial effects:
本发明的识别方法中，多损失特征自校准网络利用特征自校准机制对提取的干扰特征进行细化，通过特征提取模块对待识别时频谱图数据进行干扰特征自适应提取，并且自适应缩小同类别干扰特征在特征空间的距离，拉大不同类别干扰特征在特征空间的距离，一方面能够提取长距离依赖的像素关系，另一方面能够针对性地提取关键特征，有利于特征差异不明显的干扰识别任务，克服了传统卷积模块只关注局部信息，而往往忽略全局信息的弊端，聚合能力较高；同时，本发明采用混合损失函数，结合聚类损失的约束使得网络最终归纳组合的高级特征更具有可辨识性，利用聚合类内特征和重构输入任务的辅助信息，达到提高模型在识别任务上的泛化性的效果，可以识别较多类型的干扰，具有鲁棒性。因此，该方法对复杂混合干扰细微特征的捕捉与表征更为精准，在干扰样式繁多的识别任务中表现出更高的识别精度和更稳健的性能。In the recognition method of the present invention, the multi-loss feature self-calibration network uses the feature self-calibration mechanism to refine the extracted interference features, and adaptively extracts interference features from the spectrum data to be recognized through the feature extraction module, and adaptively reduces the distance of the same category interference features in the feature space, and enlarges the distance of the different category interference features in the feature space. On the one hand, it can extract the pixel relationship of long-distance dependence, and on the other hand, it can extract key features in a targeted manner, which is conducive to the interference recognition task with unclear feature differences, overcomes the disadvantage that the traditional convolution module only focuses on local information and often ignores the global information, and has a higher aggregation ability; at the same time, the present invention adopts a mixed loss function, combined with the constraints of clustering loss, so that the high-level features finally summarized and combined by the network are more recognizable, and the auxiliary information of the aggregated intra-class features and the reconstruction input task is used to achieve the effect of improving the generalization of the model in the recognition task, and can identify more types of interference, with robustness. Therefore, the method is more accurate in capturing and characterizing the subtle features of complex mixed interference, and shows higher recognition accuracy and more robust performance in recognition tasks with a variety of interference styles.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
图1为本发明实施例提供的基于多损失特征自校准网络的雷达有源干扰识别方法的流程示意图；FIG1 is a schematic flow chart of a radar active interference identification method based on a multi-loss feature self-calibration network provided by an embodiment of the present invention;
图2为本发明实施例提供的多损失特征自校准网络的训练方法的流程示意图；FIG2 is a schematic diagram of a flow chart of a method for training a multi-loss feature self-calibration network according to an embodiment of the present invention;
图3为本发明实施例提供的多损失特征自校准网络的结构示意图；FIG3 is a schematic diagram of the structure of a multi-loss feature self-calibration network provided by an embodiment of the present invention;
图4为本发明实施例构建的19类雷达有源干扰和1类无干扰雷达回波的时频图像示意图；FIG4 is a schematic diagram of time-frequency images of 19 types of radar active jammers and 1 type of non-interference radar echo constructed according to an embodiment of the present invention;
图5a-图5h为本发明实施例提供的不同方法下t-SNE聚类结果可视化示意图；FIG5a-FIG5h are schematic diagrams of visualization of t-SNE clustering results under different methods provided in embodiments of the present invention;
图6为本发明所提方法和其它对比方法的特征可视化结果示意图。FIG6 is a schematic diagram of feature visualization results of the method proposed in the present invention and other comparative methods.
具体实施方式DETAILED DESCRIPTION
下面结合具体实施例对本发明做进一步详细的描述，但本发明的实施方式不限于此。The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.
实施例一Embodiment 1
请参见图1，图1为本发明实施例提供的基于多损失特征自校准网络的雷达有源干扰识别方法的流程示意图，该方法包括步骤：Please refer to FIG. 1 , which is a flow chart of a radar active interference identification method based on a multi-loss feature self-calibration network provided by an embodiment of the present invention. The method comprises the following steps:
S1、获取包括雷达有源干扰信号的待识别时频谱图数据。S1. Obtaining time-frequency spectrum data to be identified including radar active interference signals.
S2、将待识别时频谱图数据输入训练好的多损失特征自校准网络进行干扰类型识别，得到识别分类结果。S2. Input the frequency spectrum data to be identified into the trained multi-loss feature self-calibration network to identify the interference type and obtain the identification and classification results.
其中，训练好的多损失特征自校准网络包括依次连接的特征提取模块、特征映射模块和重构模块，训练好的多损失特征自校准网络为利用混合损失函数对多损失特征自校准网络的参数进行训练更新得到，混合损失函数由训练集样本标签传播的交叉熵损失、特征映射模块的聚类损失和重构模块的均方误差损失确定。Among them, the trained multi-loss feature self-calibration network includes a feature extraction module, a feature mapping module and a reconstruction module connected in sequence. The trained multi-loss feature self-calibration network is obtained by training and updating the parameters of the multi-loss feature self-calibration network using a mixed loss function. The mixed loss function is determined by the cross entropy loss of the training set sample label propagation, the clustering loss of the feature mapping module and the mean square error loss of the reconstruction module.
按照如下步骤进行干扰类型识别：Follow the steps below to identify the interference type:
S21、通过特征提取模块对待识别时频谱图数据进行干扰特征自适应提取，并且自适应缩小同类别干扰特征在特征空间的距离，拉大不同类别干扰特征在特征空间的距离，得到不同类型干扰特征向量。S21. Adaptively extract interference features from the frequency spectrum data to be identified through a feature extraction module, and adaptively reduce the distance between interference features of the same category in the feature space, and increase the distance between interference features of different categories in the feature space, so as to obtain interference feature vectors of different types.
S22、通过特征映射模块对不同类型干扰特征向量降维映射后进行分类，得到识别分类结果。S22. Using a feature mapping module, different types of interference feature vectors are mapped and classified after dimension reduction to obtain identification and classification results.
需要强调的是，重构模块用于多损失特征自校准网络的训练过程，训练好的多损失特征自校准网络在对待识别时频谱图数据进行干扰类型识别时，重构模块不参与识别过程。It should be emphasized that the reconstruction module is used in the training process of the multi-loss feature self-calibration network. When the trained multi-loss feature self-calibration network identifies the interference type of the spectrum data to be identified, the reconstruction module does not participate in the identification process.
本实施例中，多损失特征自校准网络利用特征自校准机制对提取的干扰特征进行细化，通过特征提取模块对待识别时频谱图数据进行干扰特征自适应提取，并且自适应缩小同类别干扰特征在特征空间的距离，拉大不同类别干扰特征在特征空间的距离，一方面能够提取长距离依赖的像素关系，另一方面能够针对性地提取关键特征，有利于特征差异不明显的干扰识别任务，克服了传统卷积模块只关注局部信息，而往往忽略全局信息的弊端，聚合能力较高；同时，本实施例采用混合损失函数，结合聚类损失的约束使得网络最终归纳组合的高级特征更具有可辨识性，利用聚合类内特征和重构输入任务的辅助信息，达到提高模型在识别任务上的泛化性的效果，可以识别较多类型的干扰，具有鲁棒性。因此，该方法对复杂混合干扰细微特征的捕捉与表征更为精准，在干扰样式繁多的识别任务中表现出更高的识别精度和更稳健的性能。In this embodiment, the multi-loss feature self-calibration network uses the feature self-calibration mechanism to refine the extracted interference features, and adaptively extracts interference features from the spectrum data to be identified through the feature extraction module, and adaptively reduces the distance between interference features of the same category in the feature space, and increases the distance between interference features of different categories in the feature space. On the one hand, it can extract long-distance dependent pixel relationships, and on the other hand, it can extract key features in a targeted manner, which is beneficial to interference identification tasks with unclear feature differences, overcomes the disadvantage that traditional convolution modules only focus on local information and often ignore global information, and has high aggregation ability; at the same time, this embodiment adopts a mixed loss function, combined with the constraints of clustering loss, so that the high-level features finally summarized and combined by the network are more recognizable, and the auxiliary information of the aggregated intra-class features and the reconstructed input tasks is used to achieve the effect of improving the generalization of the model in the identification task, and can identify more types of interference, with robustness. Therefore, this method is more accurate in capturing and characterizing the subtle features of complex mixed interference, and shows higher recognition accuracy and more robust performance in recognition tasks with a variety of interference styles.
为得到训练好的多损失特征自校准网络，本实施例的思路为：首先，建立雷达有源干扰信号模型，并通过短时傅里叶变换得到干扰时频图像，以此作为干扰识别的基准数据集，记为原始时频谱图数据集。然后，构建多损失特征自校准网络，其主要组成模块包括：特征提取模块、特征映射模块和重构模块。之后，将原始时频谱图数据集的数据输入到多损失特征自校准网络中，通过网络将干扰回波时频谱图投影在高维抽象特征可分空间，精细化表征不同类型干扰的可辨识特征，并且利用跳跃连接结构简化网络的学习目标和优化难度；同时，利用特征提取模块中的特征自校准卷积块精细化模型提升其对不同类型干扰本征特征的表示能力，在特征空间内实现类内干扰特征向量的聚合，以及类间特征向量的分离。最后，在分类层中完成对训练集真实标签的预测，并计算对应的聚类损失、交叉熵损失和均方误差损失函数，得到增强的多元混合损失函数来更新模型的参数，直至网络收敛。In order to obtain a trained multi-loss feature self-calibration network, the idea of this embodiment is: first, establish a radar active interference signal model, and obtain the interference time-frequency image through short-time Fourier transform, which is used as the benchmark data set for interference identification, recorded as the original time-frequency spectrum data set. Then, construct a multi-loss feature self-calibration network, whose main components include: feature extraction module, feature mapping module and reconstruction module. After that, the data of the original time-frequency spectrum data set is input into the multi-loss feature self-calibration network, and the interference echo time-frequency spectrum is projected into a high-dimensional abstract feature separable space through the network, and the identifiable features of different types of interference are refined, and the jump connection structure is used to simplify the learning objectives and optimization difficulty of the network; at the same time, the feature self-calibration convolution block in the feature extraction module is used to refine the model to improve its representation ability of the intrinsic features of different types of interference, and realize the aggregation of intra-class interference feature vectors and the separation of inter-class feature vectors in the feature space. Finally, the true labels of the training set are predicted in the classification layer, and the corresponding clustering loss, cross entropy loss, and mean square error loss functions are calculated to obtain an enhanced multivariate mixed loss function to update the parameters of the model until the network converges.
请参见图2，图2为本发明实施例提供的多损失特征自校准网络的训练方法的流程示意图。该多损失特征自校准网络的训练方法包括步骤：Please refer to Figure 2, which is a flow chart of a method for training a multi-loss feature self-calibration network provided by an embodiment of the present invention. The method for training a multi-loss feature self-calibration network comprises the following steps:
S201、获取包括雷达有源干扰信号和无干扰雷达回波的原始时频谱图数据集，其中，原始时频谱图数据集包括训练集样本、训练集样本标签、验证样本和验证集样本标签。S201. Acquire an original time-spectrogram dataset including radar active jamming signals and non-interference radar echoes, wherein the original time-spectrogram dataset includes training set samples, training set sample labels, verification samples, and verification set sample labels.
具体的，根据每种干扰的调制原理和不同的干噪比(Jamming-to-Noise Ratio，JNR)，利用多类型雷达有源干扰产生机理和短时傅里叶变换建立包含19类雷达有源干扰信号和1类无干扰雷达回波的原始时频谱图数据集，并对原始时频谱图数据集按照不同干扰类别打上标签，之后，将原始时频谱图数据集和标签文件按照6：2：2的比例划分训练集、验证集和测试集，训练集样本对应有训练集样本标签，验证集样本对应有验证集样本标签，测试集样本对应有测试集样本标签。Specifically, according to the modulation principle of each interference and different Jamming-to-Noise Ratio (JNR), the generation mechanism of multi-type radar active interference and short-time Fourier transform are used to establish an original time-spectrogram dataset containing 19 types of radar active interference signals and 1 type of non-interference radar echo, and the original time-spectrogram dataset is labeled according to different interference categories. After that, the original time-spectrogram dataset and the label file are divided into training set, validation set and test set in a ratio of 6:2:2. The training set samples have corresponding training set sample labels, the validation set samples have corresponding validation set sample labels, and the test set samples have corresponding test set sample labels.
本实施例中，随机生成干噪比为0dB、5dB、10dB、15dB、20dB、25dB和30dB这7种不同干噪比的信号，每种信号下有19类雷达有源干扰时频图和1类无干扰雷达回波时频图作为样本集合。具体的，样本数据大小为，批次大小为64。In this embodiment, seven signals with different interference-to-noise ratios of 0dB, 5dB, 10dB, 15dB, 20dB, 25dB and 30dB are randomly generated. Under each signal, there are 19 types of radar active interference time-frequency diagrams and 1 type of non-interference radar echo time-frequency diagram as sample sets. Specifically, the sample data size is , the batch size is 64.
S202、将训练集样本和训练集样本标签输入多损失特征自校准网络中进行训练。S202: Input the training set samples and training set sample labels into the multi-loss feature self-calibration network for training.
首先，构建多损失特征自校准网络。First, a multi-loss feature self-calibration network is constructed.
请参见图3，图3为本发明实施例提供的多损失特征自校准网络的结构示意图。该多损失特征自校准网络包括依次连接的特征提取模块、特征映射模块和重构模块。Please refer to Fig. 3, which is a schematic diagram of the structure of a multi-loss feature self-calibration network provided by an embodiment of the present invention. The multi-loss feature self-calibration network includes a feature extraction module, a feature mapping module and a reconstruction module connected in sequence.
在一个具体实施例中，特征提取模块包括第一特征自校准卷积块、第二特征自校准卷积块、第三特征自校准卷积块和第四特征自校准卷积块。其中，第一特征自校准卷积块、第二特征自校准卷积块、第三特征自校准卷积块和第四特征自校准卷积块依次连接，用于对待识别时频谱图数据依次进行特征自校准卷积，得到不同类型干扰特征向量。In a specific embodiment, the feature extraction module includes a first feature self-calibration convolution block, a second feature self-calibration convolution block, a third feature self-calibration convolution block, and a fourth feature self-calibration convolution block. The first feature self-calibration convolution block, the second feature self-calibration convolution block, the third feature self-calibration convolution block, and the fourth feature self-calibration convolution block are connected in sequence to perform feature self-calibration convolution on the spectrum data to be identified in sequence to obtain different types of interference feature vectors.
在一个具体实施例中，第一特征自校准卷积块、第二特征自校准卷积块、第三特征自校准卷积块和第四特征自校准卷积块的结构相同，均包括第一卷积块、第二卷积块、第三卷积块、通道特征自校准模块、空间特征自校准模块、升维模块、第一相加模块和第一最大池化层。In a specific embodiment, the first feature self-calibration convolution block, the second feature self-calibration convolution block, the third feature self-calibration convolution block and the fourth feature self-calibration convolution block have the same structure, and all include a first convolution block, a second convolution block, a third convolution block, a channel feature self-calibration module, a spatial feature self-calibration module, a dimensionality increase module, a first addition module and a first maximum pooling layer.
其中，第一卷积块、第二卷积块、第三卷积块依次连接，用于对特征自校准卷积块的输入特征图依次进行卷积处理，得到第三卷积块的输出特征图。通道特征自校准模块用于对第三卷积块的输出特征图进行通道特征自校准，得到通道特征自校准特征图。空间特征自校准模块用于对通道特征自校准特征图进行空间特征自校准，得到空间特征自校准特征图。升维模块用于对特征自校准卷积块的输入特征图进行升维操作，得到升维特征图。第一相加模块用于将升维特征图加在空间特征自校准特征图上，得到相加特征图。第一最大池化层用于对相加特征图进行下采样，得到特征自校准卷积块的输出特征图。可以理解的是，第一卷积块、第二卷积块、第三卷积块、通道特征自校准模块、空间特征自校准模块依次串联。Among them, the first convolution block, the second convolution block, and the third convolution block are connected in sequence, and are used to perform convolution processing on the input feature map of the feature self-calibration convolution block in sequence to obtain the output feature map of the third convolution block. The channel feature self-calibration module is used to perform channel feature self-calibration on the output feature map of the third convolution block to obtain the channel feature self-calibration feature map. The spatial feature self-calibration module is used to perform spatial feature self-calibration on the channel feature self-calibration feature map to obtain the spatial feature self-calibration feature map. The dimension increase module is used to perform dimension increase operation on the input feature map of the feature self-calibration convolution block to obtain the dimension increase feature map. The first addition module is used to add the dimension increase feature map to the spatial feature self-calibration feature map to obtain the added feature map. The first maximum pooling layer is used to downsample the added feature map to obtain the output feature map of the feature self-calibration convolution block. It can be understood that the first convolution block, the second convolution block, the third convolution block, the channel feature self-calibration module, and the spatial feature self-calibration module are connected in series in sequence.
进一步的，第一卷积块包括依次连接的第一卷积层、第一批规范化层和第一非线性层。第二卷积块包括依次连接的第二卷积层、第二批规范化层和第二非线性层；第三卷积块包括依次连接的第三卷积层和第三非线性层。其中，第一非线性层、第二非线性层、第三非线性层可以均采用LeakyReLU非线性层。Furthermore, the first convolution block includes a first convolution layer, a first batch of normalization layers, and a first nonlinear layer connected in sequence. The second convolution block includes a second convolution layer, a second batch of normalization layers, and a second nonlinear layer connected in sequence; the third convolution block includes a third convolution layer and a third nonlinear layer connected in sequence. Among them, the first nonlinear layer, the second nonlinear layer, and the third nonlinear layer can all use LeakyReLU nonlinear layers.
具体的，第一特征自校准卷积块中，第一卷积层、第二卷积层、第三卷积层均有64个卷积核，卷积核大小均为，步长为1，填充为1。第二特征自校准卷积块中，第一卷积层、第二卷积层、第三卷积层均有128个卷积核，卷积核大小均为，步长为1，填充为1。第三特征自校准卷积块中，第一卷积层、第二卷积层、第三卷积层均有256个卷积核，卷积核大小均为，步长为1，填充为1。第四特征自校准卷积块中，第一卷积层、第二卷积层、第三卷积层均有512个卷积核，卷积核大小均为，步长为1，填充为1。每个特征自校准卷积块中，第三卷积块的输出接入串联的通道特征自校准模块和空间特征自校准模块。Specifically, in the first feature self-calibration convolution block, the first convolution layer, the second convolution layer, and the third convolution layer all have 64 convolution kernels, and the convolution kernel sizes are , the step size is 1, and the padding is 1. In the second feature self-calibration convolution block, the first convolution layer, the second convolution layer, and the third convolution layer all have 128 convolution kernels, and the convolution kernel size is , the step size is 1, and the padding is 1. In the third feature self-calibration convolution block, the first convolution layer, the second convolution layer, and the third convolution layer all have 256 convolution kernels, and the convolution kernel size is , the step size is 1, and the padding is 1. In the fourth feature self-calibration convolution block, the first convolution layer, the second convolution layer, and the third convolution layer all have 512 convolution kernels, and the convolution kernel size is , the step size is 1, and the padding is 1. In each feature self-calibration convolution block, the output of the third convolution block is connected to the series-connected channel feature self-calibration module and spatial feature self-calibration module.
进一步的，通道特征自校准模块包括自适应最大池化层、自适应平均池化层、多层感知机、第二相加模块、通道权重归一化模块和第一相乘模块。其中，自适应最大池化层用于对第三卷积块的输出特征图中特征图高度和特征图宽度同时进行自适应最大池化，归纳出每一通道维上的最大值响应。自适应平均池化层用于对第三卷积块的输出特征图中特征图高度和特征图宽度同时进行自适应平均池化，归纳出每一通道维上的平均响应。多层感知机用于对每一通道维上的最大值响应依次进行特征缩放、特征还原、提取通道维信息得到最大值响应输出，并对每一通道维上的平均响应依次进行特征缩放、特征还原、提取通道维信息得到平均响应输出。第二相加模块用于将最大值响应输出和平均响应输出进行相加融合，得到相加特征图。通道权重归一化模块用于利用激活函数将相加特征图中通道的权重归一化，得到通道归一化权重。第一相乘模块用于将通道归一化权重与第三卷积块的输出特征图相乘得到通道特征自校准特征图。Furthermore, the channel feature self-calibration module includes an adaptive maximum pooling layer, an adaptive average pooling layer, a multi-layer perceptron, a second addition module, a channel weight normalization module and a first multiplication module. Among them, the adaptive maximum pooling layer is used to simultaneously perform adaptive maximum pooling on the feature map height and feature map width in the output feature map of the third convolution block, and summarize the maximum value response on each channel dimension. The adaptive average pooling layer is used to simultaneously perform adaptive average pooling on the feature map height and feature map width in the output feature map of the third convolution block, and summarize the average response on each channel dimension. The multi-layer perceptron is used to perform feature scaling, feature restoration, and extract channel dimension information on the maximum value response on each channel dimension in turn to obtain the maximum value response output, and perform feature scaling, feature restoration, and extract channel dimension information on the average response on each channel dimension in turn to obtain the average response output. The second addition module is used to add and fuse the maximum value response output and the average response output to obtain the added feature map. The channel weight normalization module is used to normalize the weights of the channels in the added feature map using an activation function to obtain channel normalization weights. The first multiplication module is used to multiply the channel normalization weight and the output feature map of the third convolution block to obtain a channel feature self-calibration feature map.
具体的，假设输入通道特征自校准模块的特征图为，表示向量空间，表示特征图通道数，为特征图高度，为特征图宽度。首先，自适应最大池化层对中的第二维(特征图高度)和第三维(特征图宽度)进行自适应最大池化(Adaptive Max Pooling)，归纳出每一通道维上的最大值响应；同时，自适应平均池化层对中的第二维(特征图高度)和第三维(特征图宽度)并行进行自适应平均池化(Adaptive Average Pooling)后，归纳出每一通道维上的平均响应。接着，和送入参数共享的多层感知机(Multilayer Perceptron，MLP)，经特征缩放并特征还原后进一步提取通道维的有用信息，得到感知机的并行输出：最大值响应输出和平均响应输出。之后，第二相加模块对感知机的并行输出进行求和操作以将最大值响应输出和平均响应输出相加融合，相加特征图再经过通道权重归一化模块以利用Sigmoid激活函数将通道的权重归一化到0-1之间，得到通道归一化权重；最后第一相乘模块将通道归一化权重和输入特征图即第三卷积块的输出特征图相乘得到通道特征自校准特征图。通道特征自校准特征图的公式可以表示为：Specifically, assuming that the feature map of the input channel feature self-calibration module is , represents a vector space, Representation feature map Number of channels, The feature map high, The feature map Width. First, the adaptive maximum pooling layer The second dimension in (feature map height) and the third dimension (feature map Width) to perform adaptive max pooling (Adaptive Max Pooling) to summarize the maximum value response on each channel dimension ; At the same time, the adaptive average pooling layer The second dimension in (feature map height) and the third dimension (feature map After performing adaptive average pooling in parallel (width), the average response on each channel dimension is summarized. .then, and It is sent to the parameter-sharing multilayer perceptron (MLP), and after feature scaling and feature restoration, useful information of the channel dimension is further extracted to obtain the parallel output of the perceptron: the maximum response output and the average response output. After that, the second addition module performs a summation operation on the parallel output of the perceptron to add and fuse the maximum response output and the average response output. The added feature map is then passed through the channel weight normalization module to use the Sigmoid activation function to normalize the channel weight to between 0 and 1 to obtain the channel normalized weight; finally, the first multiplication module multiplies the channel normalization weight and the input feature map, that is, the output feature map of the third convolution block, to obtain the channel feature self-calibration feature map. The formula of the channel feature self-calibration feature map can be expressed as:
； ;
其中，表示通道特征自校准模板，表示激活函数，表示多层感知机，表示自适应平均池化，表示自适应最大池化，表示第三卷积块的输出特征图，，表示向量空间，表示特征图通道数，表示特征图高度，表示特征图宽度。in, represents the channel feature self-calibration template, represents the activation function, represents a multi-layer perceptron, represents adaptive average pooling, represents adaptive maximum pooling, represents the output feature map of the third convolutional block, , represents a vector space, Representation feature map Number of channels, Representation feature map high, Representation feature map width.
在一个具体实施例中，多层感知机包括第四卷积层、ReLU非线性激活层和第五卷积层。第四卷积层、第五卷积层的卷积核大小均为；第四卷积层中卷积核个数为，为缩放率，可以根据输入特征图通道数进行调整，本实施例中；第五卷积层中卷积核个数为。设置第一特征自校准卷积块、第二特征自校准卷积块、第三特征自校准卷积块中缩放率为16，第四特征自校准卷积块缩放率为32。In a specific embodiment, the multi-layer perceptron includes a fourth convolutional layer, a ReLU nonlinear activation layer, and a fifth convolutional layer. The convolution kernel sizes of the fourth convolutional layer and the fifth convolutional layer are both ; The number of convolution kernels in the fourth convolution layer is , is the scaling rate, which can be adjusted according to the number of input feature map channels. ; The number of convolution kernels in the fifth convolution layer is . Set the scaling ratio in the first feature self-calibration convolution block, the second feature self-calibration convolution block, and the third feature self-calibration convolution block The fourth feature self-calibration convolution block scaling rate is 16. is 32.
进一步的，空间特征自校准模块包括第二最大池化层、平均池化层、拼接模块、卷积模块、空间权重归一化模块和第二相乘模块。其中，第二最大池化层用于对通道特征自校准特征图进行通道维最大池化，得到压缩到空间维的最大池化特征图。平均池化层用于对通道特征自校准特征图进行通道维平均池化，得到压缩到空间维的平均池化特征图。拼接模块用于采用拼接方法将最大池化特征图和平均池化特征图进行融合，得到拼接特征图。卷积模块用于对拼接特征图进行卷积映射，得到映射特征图。空间权重归一化模块用于利用激活函数将映射特征图中空间的权重归一化，得到空间归一化权重。第二相乘模块用于将空间归一化权重与通道特征自校准特征图相乘得到空间特征自校准特征图。Furthermore, the spatial feature self-calibration module includes a second maximum pooling layer, an average pooling layer, a splicing module, a convolution module, a spatial weight normalization module and a second multiplication module. Among them, the second maximum pooling layer is used to perform channel dimension maximum pooling on the channel feature self-calibration feature map to obtain a maximum pooling feature map compressed to the spatial dimension. The average pooling layer is used to perform channel dimension average pooling on the channel feature self-calibration feature map to obtain an average pooling feature map compressed to the spatial dimension. The splicing module is used to fuse the maximum pooling feature map and the average pooling feature map using a splicing method to obtain a splicing feature map. The convolution module is used to perform convolution mapping on the splicing feature map to obtain a mapping feature map. The spatial weight normalization module is used to normalize the spatial weights in the mapping feature map using an activation function to obtain a spatial normalization weight. The second multiplication module is used to multiply the spatial normalization weight with the channel feature self-calibration feature map to obtain a spatial feature self-calibration feature map.
具体的，经过通道特征自校准模块处理的特征图输入空间特征自校准模块中，其中，表示向量空间，表示特征图通道数，表示特征图高度，表示特征图宽度。特征图并行经通道维的第二最大池化和通道维的平均池化得到压缩到空间维的最大池化特征图和压缩到空间维的平均池化特征图。拼接模块采用拼接方法将得到的两个特征图和信息融合，得到拼接特征图。卷积模块对拼接特征图进行卷积映射，进一步得出映射特征图；其中，卷积核个数为1，卷积核大小为，填充为取整，取值可人为根据输入特征图尺寸进行调整；例如，设置第一、第二和第三特征自校准卷积块中空间特征自校准模块的卷积核大小为7，设置第四特征自校准卷积块中空间特征自校准模块的卷积核大小为5。空间权重归一化模块利用Sigmoid激活函数将映射特征图的空间的权重归一化到0-1之间，得到空间归一化权重。最后，第二相乘模块将空间归一化权重和通道特征自校准特征图相乘得到空间特征校准后的特征图输出，即空间特征自校准特征图。空间特征自校准特征图的公式可以表示为：Specifically, the feature map processed by the channel feature self-calibration module Input spatial feature self-calibration module, where represents a vector space, Representation feature map Number of channels, Representation feature map high, Representation feature map Width. Feature map The second maximum pooling of the channel dimension and the average pooling of the channel dimension are performed in parallel to obtain the maximum pooling feature map compressed to the spatial dimension. And the average pooling feature map compressed to the spatial dimension The splicing module uses the splicing method to obtain the two feature maps and The information is fused to obtain the spliced feature map. The convolution module performs convolution mapping on the spliced feature map to further obtain the mapping feature map; where the number of convolution kernels is 1 and the size of the convolution kernel is , filled with Rounding, The value can be adjusted manually according to the input feature map size; for example, setting the convolution kernel size of the spatial feature self-calibration module in the first, second, and third feature self-calibration convolution blocks 7, setting the convolution kernel size of the spatial feature self-calibration module in the fourth feature self-calibration convolution block is 5. The spatial weight normalization module uses the Sigmoid activation function to normalize the spatial weight of the mapping feature map to between 0 and 1 to obtain the spatial normalization weight. Finally, the second multiplication module multiplies the spatial normalization weight and the channel feature self-calibration feature map to obtain the feature map output after spatial feature calibration, that is, the spatial feature self-calibration feature map. The formula of the spatial feature self-calibration feature map can be expressed as:
； ;
其中，表示空间特征自校准模板，表示卷积权重，表示激活函数，表示平均池化，表示最大池化，表示通道特征自校准特征图，，表示向量空间，表示特征图通道数，表示特征图高度，表示特征图宽度。in, represents the spatial feature self-calibration template, represents the convolution weight, represents the activation function, represents average pooling, represents the maximum pooling, represents the channel characteristic self-calibration feature map, , represents a vector space, Representation feature map Number of channels, Representation feature map high, Representation feature map width.
进一步的，通道特征自校准模块和空间特征自校准模块以串联的形式置于三层卷积之后，通道特征自校准模块和空间特征自校准模块的整体公式表示为：Furthermore, the channel feature self-calibration module and the spatial feature self-calibration module are placed in series after the three-layer convolution. The overall formula of the channel feature self-calibration module and the spatial feature self-calibration module is expressed as:
； ;
其中，表示空间特征自校准模块的输出特征图，表示第三卷积块的输出特征图，，表示向量空间，为特征图通道数，为特征图高度，为特征图宽度，表示通道特征自校准模板，表示空间特征自校准模板，表示逐元素相乘。in, represents the output feature map of the spatial feature self-calibration module, represents the output feature map of the third convolutional block, , represents a vector space, The feature map Number of channels, The feature map high, The feature map width, represents the channel feature self-calibration template, represents the spatial feature self-calibration template, Represents element-wise multiplication.
进一步的，由升维模块和第一相加模块形成跳跃连接。跳跃连接使特征自校准卷积块的输入通过跨层数据通路，跳过特征自校准卷积块的计算，通过升维模块的升维操作后直接利用第一相加模块加在通道特征自校准和空间特征自校准后的输出特征图上。具体的，升维模块由的第六卷积层和第三批规范化层完成，卷积核个数和特征自校准模块输出的通道数一致。Furthermore, a skip connection is formed by the dimension-raising module and the first addition module. The skip connection allows the input of the feature self-calibration convolution block to pass through the cross-layer data path, skipping the calculation of the feature self-calibration convolution block, and directly using the first addition module to add to the output feature map after channel feature self-calibration and spatial feature self-calibration after the dimension-raising operation of the dimension-raising module. Specifically, the dimension-raising module consists of The sixth convolutional layer and the third batch of normalization layers are completed, and the number of convolution kernels is consistent with the number of channels output by the feature self-calibration module.
进一步的，每个特征自校准卷积块中，跳跃连接之后添加池化内核为的第一最大池化层进行下采样，池化层步长为2。Furthermore, in each feature self-calibration convolution block, a pooling kernel is added after the jump connection as The first maximum pooling layer is downsampled, and the pooling layer stride is 2.
具体的，特征自校准卷积块的输出特征图的公式可以表示为：Specifically, the formula for the output feature map of the feature self-calibration convolution block can be expressed as:
； ;
其中，表示升维操作，表示提取特征，表示第个特征自校准卷积块的输入特征图，表示第个特征自校准卷积块的权重参数，表示最大池化。in, represents the dimension-raising operation, represents the extracted features, Indicates The input feature map of the feature self-calibration convolution block, Indicates The weight parameters of the feature self-calibration convolution block, Represents maximum pooling.
在一个具体实施例中，特征映射模块包括第一全连接层和第二全连接层。其中，第一全连接层用于将不同类型干扰特征向量的展平特征投影到特征空间，得到嵌入向量。第二全连接层用于对嵌入向量进行分类，得到识别分类结果。In a specific embodiment, the feature mapping module includes a first fully connected layer and a second fully connected layer. The first fully connected layer is used to project the flattened features of different types of interference feature vectors into the feature space to obtain an embedded vector. The second fully connected layer is used to classify the embedded vector to obtain a recognition classification result.
具体的，特征映射模块接收来自特征提取模块的维度为的特征图，将其展平后，通过节点数为256的第一全连接层得到嵌入向量，最后通过节点数等于类别数20的第二全连接层得到识别分类结果。Specifically, the dimension of the feature mapping module received from the feature extraction module is The feature graph is flattened and passed through the first fully connected layer with 256 nodes to obtain the embedding vector. Finally, the recognition and classification results are obtained through the second fully connected layer with 20 nodes.
在一个具体实施例中，重构模块包括依次连接的第三全连接层、第四非线性层、第四全连接层、第五非线性层、第五全连接层和激活函数层。In a specific embodiment, the reconstruction module includes a third fully connected layer, a fourth nonlinear layer, a fourth fully connected layer, a fifth nonlinear layer, a fifth fully connected layer and an activation function layer connected in sequence.
具体的，重构网络由三层全连接层和两层ReLU非线性层交替组成，第三全连接层有1024个节点，第四非线性层有4096个节点，第五全连接层有16384个节点；最后经Sigmoid函数激活输出，重组成与原始时频谱图相同的尺寸。Specifically, the reconstruction network consists of three layers of fully connected layers and two layers of ReLU nonlinear layers. The third fully connected layer has 1024 nodes, the fourth nonlinear layer has 4096 nodes, and the fifth fully connected layer has 16384 nodes. Finally, the output is activated by the Sigmoid function and reconstructed into the same as the original time-frequency spectrum. size.
然后，将训练集样本和训练集样本标签输入构建好的多损失特征自校准网络中以对网络进行训练。Then, the training set samples and training set sample labels are input into the constructed multi-loss feature self-calibration network to train the network.
S203、根据训练集样本标签传播的交叉熵损失、特征映射模块的聚类损失、以及重构模块的均方误差损失确定混合损失函数。S203, determining a mixed loss function according to the cross entropy loss of the training set sample label propagation, the clustering loss of the feature mapping module, and the mean square error loss of the reconstruction module.
具体的，对聚类损失、交叉熵损失和均方误差损失这三项损失函数求和，得到混合损失函数。Specifically, the three loss functions of clustering loss, cross entropy loss and mean square error loss are summed to obtain a hybrid loss function.
首先，计算特征映射模块输出的嵌入向量的聚类损失。对于每次迭代训练过程中特征映射模块提取的特征向量，计算它们的聚类损失：First, the clustering loss of the embedding vector output by the feature mapping module is calculated. For each feature vector extracted by the feature mapping module during the training process, , calculate their clustering loss:
； ;
式中，表示批次大小，表示特征向量所属类别的特征中心，它会随着同类别特征向量的变化而更新位置。从聚类损失公式可以看出聚类损失希望同类样本到特征中心的距离越小越好，达到约束类内紧凑的效果。In the formula, Indicates the batch size, Represents the feature vector The feature center of the category to which it belongs will update its position as the feature vector of the same category changes. From the clustering loss formula, it can be seen that the clustering loss hopes that the distance between the same type of samples and the feature center is as small as possible, so as to achieve the compact effect within the constraint class.
然后，计算每个训练集样本标签传播对应的交叉熵损失函数：Then, the cross entropy loss function corresponding to the label propagation of each training set sample is calculated:
； ;
式中，表示训练样本的真实标签，表示预测结果的概率分布。In the formula, Represents training samples The real label, Represents the probability distribution of the prediction results.
最后，重构模块用归纳的预测结果重新构建出该类别代表的输入干扰时频谱图，均方误差损失函数由计算输出图像像素与原始图像像素点的欧式距离构建，它的计算公式为：Finally, the reconstruction module uses the summarized prediction results to reconstruct the input interference time-frequency spectrum of the category. The mean square error loss function is constructed by calculating the Euclidean distance between the output image pixel and the original image pixel. Its calculation formula is:
； ;
式中，为重构像素点，为原始图像像素点。In the formula, To reconstruct pixels, is the original image pixel.
综合上述三个损失函数，得到一个增强的混合损失函数:Combining the above three loss functions, we get an enhanced hybrid loss function:
； ;
式中和为可调权重超参数，在本例中设置为0.5，设置为5e-4。In the formula and is an adjustable weight hyperparameter, in this case Set to 0.5, Set to 5e-4 .
增强的混合损失函数的设计基于多任务学习的思想，在常用的交叉熵损失基础上加入聚类损失和均方误差损失。聚类损失能够在高维特征空间实现同类型干扰特征向量的聚合，增强网络所提取特征的鲁棒性，而均方误差损失能够减少网络传播中特征的丢失。增强的混合损失函数利用不同优化过程之间的信息交流和辅助训练，提高标签传播在嵌入空间中的分类性能。The design of the enhanced hybrid loss function is based on the idea of multi-task learning. Clustering loss and mean square error loss are added to the commonly used cross entropy loss. Clustering loss can achieve the aggregation of the same type of interference feature vectors in the high-dimensional feature space and enhance the robustness of the features extracted by the network, while mean square error loss can reduce the loss of features in network propagation. The enhanced hybrid loss function uses information exchange and auxiliary training between different optimization processes to improve the classification performance of label propagation in the embedding space.
S204、利用混合损失函数对多损失特征自校准网络的参数进行更新。S204, using a mixed loss function to update the parameters of the multi-loss feature self-calibration network.
具体的，根据增强的混合损失函数对网络的参数进行更新，重复上述步骤直至完成训练次数，达到网络收敛。Specifically, the parameters of the network are updated according to the enhanced hybrid loss function, and the above steps are repeated until the training times are completed and the network converges.
S205、利用验证集样本和验证集样本标签对每轮训练结束后的网络进行选择，将识别准确率最高的模型作为训练好的多损失特征自校准网络。S205, using the validation set samples and validation set sample labels to select the network after each round of training, and taking the model with the highest recognition accuracy as the trained multi-loss feature self-calibration network.
具体的，利用验证集样本和验证集样本标签对每轮训练结束后固定超参数的网络进行选择，将识别准确率最高的模型作为训练好的多损失特征自校准网络。Specifically, the validation set samples and validation set sample labels are used to select the network with fixed hyperparameters after each round of training, and the model with the highest recognition accuracy is used as the trained multi-loss feature self-calibration network.
进一步，在得到训练好的多损失特征自校准网络后，可以将测试集样本输入网络进行识别测试，并计算识别准确率。Furthermore, after obtaining the trained multi-loss feature self-calibration network, the test set samples can be input into the network for recognition testing, and the recognition accuracy can be calculated.
本实施例中，输入的原始时频谱图数据先后经过特征提取模块和特征映射模块，通过特征向量嵌入函数将干扰时频谱图从原空间映射到特征空间，其中表示学习参数，表示向量空间，表示映射前的向量空间维度，表示映射后的向量空间维度。经过训练后，多损失特征自校准网络能够自适应关注辨识性更强的干扰特征，并且能够自适应缩小同类别样本的特征向量在高维特征空间中的距离，拉大不同类别样本在特征空间的距离，以此来降低不同类型干扰特征向量降维映射后的分类难度；由于跳跃连接结构解决了因模型加深带来的梯度消失问题，且更容易优化，所以特征提取模块采用跳跃连接结构；进一步，在多层卷积操作后加入特征自校准模块，使网络能够进行自适应细化特征，增强不同类型干扰特征的可分性。最后，由于池化过程会带来信息损失，因此分类器后连接输入重构模块，直接建立潜在特征映射空间与输入的关系，辅助网络的优化。In this embodiment, the input raw time-frequency spectrum data is successively passed through the feature extraction module and the feature mapping module, and then the feature vector is embedded in the function. The interference spectrum is mapped from the original space to the feature space, where represents the learning parameters, represents a vector space, represents the dimension of the vector space before mapping, Represents the dimension of the vector space after mapping. After training, the multi-loss feature self-calibration network can adaptively focus on more recognizable interference features, and can adaptively reduce the distance between feature vectors of samples of the same category in the high-dimensional feature space, and increase the distance between samples of different categories in the feature space, so as to reduce the difficulty of classification after dimensionality reduction mapping of different types of interference feature vectors; since the jump connection structure solves the gradient vanishing problem caused by model deepening and is easier to optimize, the feature extraction module adopts the jump connection structure; further, the feature self-calibration module is added after the multi-layer convolution operation, so that the network can adaptively refine the features and enhance the separability of different types of interference features. Finally, since the pooling process will cause information loss, the input reconstruction module is connected after the classifier to directly establish the relationship between the potential feature mapping space and the input to assist in network optimization.
综上，本实施例提出的基于多损失特征自校准网络的雷达有源干扰识别方法，利用深度非线性神经网络深入挖掘不同类型干扰之间的时频特征，并将其映射到高维特征空间中拟合出可分离平面，实现高精度干扰智能识别。一方面设计了特征自校准卷积块，克服了传统卷积模块只关注局部信息，而往往忽略全局信息的弊端。通过引入空间维度和通道维度的特征校准模模块，自适应校准网络对干扰的特征提取，提高网络对不同类型干扰的提取精度与表征效果进行自适应特征细化，而跳跃连接结构能够消除网络深度增加带来的梯度消失问题；另一方面引入多重约束的混合损失函数聚合干扰在高维空间的类内特征，扩增干扰类间特征，提升网络的雷达有源干扰识别性能，为雷达干扰识别提供了新的方法。因此，该方法可用于雷达抗干扰处理中，为雷达抗干扰策略选择提供重要的先验信息，提升SAR在复杂电磁环境下的抗干扰能力和信息获取能力。In summary, the radar active interference identification method based on the multi-loss feature self-calibration network proposed in this embodiment uses a deep nonlinear neural network to deeply explore the time-frequency characteristics between different types of interference, and maps it to a high-dimensional feature space to fit a separable plane, thereby realizing high-precision interference intelligent identification. On the one hand, a feature self-calibration convolution block is designed to overcome the disadvantage that the traditional convolution module only focuses on local information and often ignores global information. By introducing feature calibration modules in spatial dimensions and channel dimensions, the adaptive calibration network extracts the features of interference, improves the extraction accuracy and characterization effect of the network for different types of interference, and performs adaptive feature refinement, and the jump connection structure can eliminate the gradient vanishing problem caused by the increase in network depth; on the other hand, the introduction of a multi-constrained mixed loss function aggregates the intra-class features of interference in high-dimensional space, amplifies the inter-class features of interference, and improves the radar active interference identification performance of the network, providing a new method for radar interference identification. Therefore, this method can be used in radar anti-interference processing, provides important prior information for the selection of radar anti-interference strategies, and improves the anti-interference ability and information acquisition ability of SAR in complex electromagnetic environments.
进一步，本实施例通过仿真实验对基于多损失特征自校准网络的雷达有源干扰识别方法的效果进行说明。Furthermore, this embodiment illustrates the effect of the radar active interference identification method based on the multi-loss feature self-calibration network through simulation experiments.
一、数据集1. Dataset
实验所用数据为十九类雷达干扰仿真时频图和一类无干扰雷达回波信号时频图，如图4所示，图4为本发明实施例构建的19类雷达有源干扰和1类无干扰雷达回波的时频图像示意图。干扰类型包括单个干扰和复合干扰两大类。单个干扰类型有：噪声调幅干扰、噪声调频干扰、噪声乘积干扰、噪声卷积干扰、多点频干扰、正弦波调制扫频干扰、锯齿波调制扫频干扰、方波调制扫频干扰、密集假目标干扰、间歇采样转发干扰、间歇采样重复转发干扰、示样脉冲干扰、多假目标干扰以及梳状谱调制干扰。复合干扰包括：调频+密集假目标干扰、调频+间歇采样转发干扰、调频+间歇采样重复转发干扰、噪声卷积+密集假目标干扰、噪声卷积+间歇采样转发干扰。根据每种干扰信号的调制原理和不同的干噪比(Jamming-to-Noise Ratio，JNR)，干噪比选择0dB、5dB、10dB、15dB、20dB、25dB和30dB共7种，通过短时傅里叶变换生成干扰信号的时频图像。同时为了保证样本中类别的平衡，每一种干扰随机产生1000个样本，每个样本的尺寸为。样本均按照训练集：验证集：测试集=6：2：2的比例进行划分。The data used in the experiment are 19 types of radar interference simulation time-frequency diagrams and a type of non-interference radar echo signal time-frequency diagram, as shown in Figure 4, which is a schematic diagram of the time-frequency images of 19 types of radar active interference and 1 type of non-interference radar echo constructed by an embodiment of the present invention. The interference types include two categories: single interference and composite interference. The single interference types are: noise amplitude modulation interference, noise frequency modulation interference, noise product interference, noise convolution interference, multi-point frequency interference, sine wave modulation sweep frequency interference, sawtooth wave modulation sweep frequency interference, square wave modulation sweep frequency interference, dense false target interference, intermittent sampling forwarding interference, intermittent sampling repeated forwarding interference, sample pulse interference, multiple false target interference and comb spectrum modulation interference. Composite interference includes: frequency modulation + dense false target interference, frequency modulation + intermittent sampling forwarding interference, frequency modulation + intermittent sampling repeated forwarding interference, noise convolution + dense false target interference, noise convolution + intermittent sampling forwarding interference. According to the modulation principle of each interference signal and different jamming-to-noise ratios (JNR), 7 types of jamming-to-noise ratios are selected: 0dB, 5dB, 10dB, 15dB, 20dB, 25dB and 30dB. The time-frequency image of the interference signal is generated by short-time Fourier transform. At the same time, in order to ensure the balance of categories in the sample, 1000 samples are randomly generated for each interference, and the size of each sample is The samples are divided into training set: validation set: test set = 6:2:2 ratio.
二、实现细节Implementation Details
1)按上述要求选取实验数据，划分训练集、验证集和测试集；1) Select experimental data according to the above requirements and divide them into training set, validation set and test set;
2)本发明的仿真实验的硬件平台为：CPU为AMD Ryzen 9 5900HX RadeonGraphics，十六核，主频为3.30GHz，内存大小为32GB；显存大小为16GB。本发明使用AdamW优化器以及StepLR的学习率更新方法，设置初始学习率为0.0005，衰减步长为10，衰减率为0.5。将训练样本输入到网络中进行训练，对于所有模型均训练30个epoch，根据每轮训练结束后固定超参数的网络在验证集上的效果，保存在验证集上识别准确率最高的模型。2) The hardware platform of the simulation experiment of the present invention is: the CPU is AMD Ryzen 9 5900HX RadeonGraphics, sixteen cores, the main frequency is 3.30GHz, the memory size is 32GB; the video memory size is 16GB. The present invention uses the AdamW optimizer and the learning rate update method of StepLR, setting the initial learning rate to 0.0005, the decay step size to 10, and the decay rate to 0.5. The training samples are input into the network for training. For all models, 30 epochs are trained. According to the effect of the network with fixed hyperparameters on the verification set after each round of training, the model with the highest recognition accuracy on the verification set is saved.
3)将测试样本输入到最优模型中进行测试并得到识别率，与其它智能干扰识别方法的识别准确率进行比较，结果如表1和表2所示：3) Input the test samples into the optimal model for testing and obtain the recognition rate, which is compared with the recognition accuracy of other intelligent interference recognition methods. The results are shown in Tables 1 and 2:
表1 本发明所提方法和其它识别网络识别准确率在JNR=30dB、25dB、20dB和15dB数据集的对比结果Table 1 Comparison results of the recognition accuracy of the proposed method and other recognition networks in the JNR=30dB, 25dB, 20dB and 15dB data sets
表2 本发明所提方法和其它识别网络识别准确率在JNR=10dB、5dB和0dB数据集的对比结果Table 2 Comparison results of the recognition accuracy of the proposed method and other recognition networks in the JNR=10dB, 5dB and 0dB data sets
上述结果表明，本发明提出的基于多损失特征自校准网络的雷达有源干扰识别方法，在0dB、5dB、10dB、15dB、20dB、25dB和30dB干噪比全样本训练的结果，均表现出更高的识别准确率，并且JNR的改变没有造成本发明所提方法识别精度的大范围波动，显示出本方法的鲁棒性。The above results show that the radar active interference identification method based on the multi-loss feature self-calibration network proposed in the present invention shows a higher recognition accuracy rate in the results of full sample training at 0dB, 5dB, 10dB, 15dB, 20dB, 25dB and 30dB interference-to-noise ratios, and the change of JNR does not cause a large range of fluctuations in the recognition accuracy of the method proposed in the present invention, which shows the robustness of the method.
请参见图5a-图5h，图5a-图5h为本发明实施例提供的不同方法下t-SNE(t-Distributed Stochastic Neighbor Embedding)聚类结果可视化示意图，其中，图5a为ResNet方法，图5b为JRNet方法，图5c为2DCNN方法，图5d为RCNN方法，图5e为AlexNet方法，图5f为1DCNN方法，图5g为DFCNN方法，图5h为本实施例方法。从图5a-图5h中可以发现，本发明所提方法在特征空间内形成了20个明显紧凑的聚类簇，且簇与簇之间相互远离。这说明相较于其它干扰识别方法，本方法在高维特征空间实现了类内更聚集，类间更分离的特征映射效果，显示了本方法的鲁棒性。Please refer to Figures 5a-5h, which are visualization diagrams of t-SNE (t-Distributed Stochastic Neighbor Embedding) clustering results under different methods provided in embodiments of the present invention, wherein Figure 5a is the ResNet method, Figure 5b is the JRNet method, Figure 5c is the 2DCNN method, Figure 5d is the RCNN method, Figure 5e is the AlexNet method, Figure 5f is the 1DCNN method, Figure 5g is the DFCNN method, and Figure 5h is the method of this embodiment. It can be found from Figures 5a-5h that the method proposed in the present invention forms 20 obviously compact clusters in the feature space, and the clusters are far away from each other. This shows that compared with other interference identification methods, this method achieves a feature mapping effect of more clustered within the class and more separated between classes in the high-dimensional feature space, showing the robustness of this method.
请参见图6，图6为本发明所提方法和其它对比方法的特征可视化结果示意图。由于WECNN、ECNN、1DCNN和DFCNN方法所处理数据的特殊性，无法产生特征可视化的热力图。由图中结果可知，本发明所提方法分类的依据更集中于时频图的干扰区域。特征可视化的结果证明了本发明所提方法对于干扰特征的表征精度明显优于其它干扰识别方法。Please refer to Figure 6, which is a schematic diagram of the feature visualization results of the method proposed in the present invention and other comparative methods. Due to the particularity of the data processed by the WECNN, ECNN, 1DCNN and DFCNN methods, it is impossible to generate a heat map of feature visualization. From the results in the figure, it can be seen that the classification basis of the method proposed in the present invention is more concentrated on the interference area of the time-frequency diagram. The results of feature visualization prove that the characterization accuracy of the interference features of the method proposed in the present invention is significantly better than other interference identification methods.
因此，本实施例的基于多损失特征自校准网络的雷达有源干扰识别方法结合干扰时频图像，在传统卷积网络的基础上配合特征自校准模块和输入重构模块，前者能够自适应捕捉特征，实现特征的精细化表征，从而增强网络的特征提取能力；后者建立特征空间和输入之间的联系，进而提高标签传播在空间中的分类性能。通过交叉熵损失、聚类损失和均方误差损失三项构成的增强的混合损失函数对模型进行优化，使模型提取到的特征向量在高维特征空间类内更聚集，类间更分散。同时，在仿真数据集上的测试结果表明，本发明所提方法在不同干噪比上的表现优于其它智能干扰识别方法。Therefore, the radar active interference identification method based on the multi-loss feature self-calibration network of this embodiment is combined with the interference time-frequency image, and cooperates with the feature self-calibration module and the input reconstruction module on the basis of the traditional convolutional network. The former can adaptively capture features and realize the refined representation of features, thereby enhancing the feature extraction ability of the network; the latter establishes the connection between the feature space and the input, thereby improving the classification performance of label propagation in space. The model is optimized by an enhanced hybrid loss function composed of cross entropy loss, clustering loss and mean square error loss, so that the feature vectors extracted by the model are more clustered within the high-dimensional feature space class and more dispersed between classes. At the same time, the test results on the simulation data set show that the method proposed in the present invention performs better than other intelligent interference identification methods at different interference-to-noise ratios.
以上内容是结合具体的优选实施方式对本发明所作的进一步详细说明，不能认定本发明的具体实施只局限于这些说明。对于本发明所属技术领域的普通技术人员来说，在不脱离本发明构思的前提下，还可以做出若干简单推演或替换，都应当视为属于本发明的保护范围。The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310741199.2ACN116482618B (en) | 2023-06-21 | 2023-06-21 | Radar active interference identification method based on multi-loss characteristic self-calibration network |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310741199.2ACN116482618B (en) | 2023-06-21 | 2023-06-21 | Radar active interference identification method based on multi-loss characteristic self-calibration network |
| Publication Number | Publication Date |
|---|---|
| CN116482618A CN116482618A (en) | 2023-07-25 |
| CN116482618Btrue CN116482618B (en) | 2023-09-19 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310741199.2AActiveCN116482618B (en) | 2023-06-21 | 2023-06-21 | Radar active interference identification method based on multi-loss characteristic self-calibration network |
| Country | Link |
|---|---|
| CN (1) | CN116482618B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116894178B (en)* | 2023-08-02 | 2025-09-12 | 电子科技大学 | Method, device and medium for predicting transpiration of medicinal crops |
| CN116975603B (en)* | 2023-09-15 | 2025-04-25 | 西安电子科技大学 | SAR interference recognition method based on feature space distance measurement |
| CN117233706B (en)* | 2023-11-16 | 2024-02-06 | 西安电子科技大学 | Radar active interference identification method based on multilayer channel attention mechanism |
| CN118519097B (en)* | 2024-05-08 | 2024-11-22 | 南京航空航天大学 | Radar interference pattern recognition method and system based on operator vectorization representation |
| CN118628376B (en)* | 2024-06-13 | 2025-09-09 | 南京航空航天大学 | Image reconstruction method and device under multi-interference-source scene |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE2451710C1 (en)* | 1972-12-08 | 1992-05-14 | Siemens Ag | Arrangement for disturbing a monopulse tracking radar device by re-emission in cross polarization |
| US5239309A (en)* | 1991-06-27 | 1993-08-24 | Hughes Aircraft Company | Ultra wideband radar employing synthesized short pulses |
| RU2193782C2 (en)* | 2000-09-19 | 2002-11-27 | Федеральное государственное унитарное предприятие "Научно-исследовательский институт измерительных приборов" | Procedure evaluating characteristics of radar exposed to active jamming |
| CN112731309A (en)* | 2021-01-06 | 2021-04-30 | 哈尔滨工程大学 | Active interference identification method based on bilinear efficient neural network |
| CN114201987A (en)* | 2021-11-09 | 2022-03-18 | 北京理工大学 | An Active Interference Identification Method Based on Adaptive Identification Network |
| CN114488140A (en)* | 2022-01-24 | 2022-05-13 | 电子科技大学 | A small-sample radar one-dimensional image target recognition method based on deep transfer learning |
| CN114895263A (en)* | 2022-05-26 | 2022-08-12 | 西安电子科技大学 | Radar active interference signal identification method based on deep migration learning |
| CN115494466A (en)* | 2022-09-22 | 2022-12-20 | 东南大学 | A self-calibration method for distributed radar |
| CN116047427A (en)* | 2023-03-29 | 2023-05-02 | 西安电子科技大学 | A small-sample radar active jamming identification method |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20220349986A1 (en)* | 2021-04-30 | 2022-11-03 | Nxp B.V. | Radar communication with interference suppression |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE2451710C1 (en)* | 1972-12-08 | 1992-05-14 | Siemens Ag | Arrangement for disturbing a monopulse tracking radar device by re-emission in cross polarization |
| US5239309A (en)* | 1991-06-27 | 1993-08-24 | Hughes Aircraft Company | Ultra wideband radar employing synthesized short pulses |
| RU2193782C2 (en)* | 2000-09-19 | 2002-11-27 | Федеральное государственное унитарное предприятие "Научно-исследовательский институт измерительных приборов" | Procedure evaluating characteristics of radar exposed to active jamming |
| CN112731309A (en)* | 2021-01-06 | 2021-04-30 | 哈尔滨工程大学 | Active interference identification method based on bilinear efficient neural network |
| CN114201987A (en)* | 2021-11-09 | 2022-03-18 | 北京理工大学 | An Active Interference Identification Method Based on Adaptive Identification Network |
| CN114488140A (en)* | 2022-01-24 | 2022-05-13 | 电子科技大学 | A small-sample radar one-dimensional image target recognition method based on deep transfer learning |
| CN114895263A (en)* | 2022-05-26 | 2022-08-12 | 西安电子科技大学 | Radar active interference signal identification method based on deep migration learning |
| CN115494466A (en)* | 2022-09-22 | 2022-12-20 | 东南大学 | A self-calibration method for distributed radar |
| CN116047427A (en)* | 2023-03-29 | 2023-05-02 | 西安电子科技大学 | A small-sample radar active jamming identification method |
| Title |
|---|
| Deceptive jamming template synthesis for SAR based on generative adversarial nets;Weiwei Fan;Signal Processing;全文* |
| 基于无载波超宽带雷达的小样本人体动作识别;蒋留兵;周小龙;车俐;;电子学报(03);全文* |
| 基于贝叶斯深度学习的一维雷达有源干扰信号识别方法;马博俊;信号处理;第39卷(第2期);全文* |
| Publication number | Publication date |
|---|---|
| CN116482618A (en) | 2023-07-25 |
| Publication | Publication Date | Title |
|---|---|---|
| CN116482618B (en) | Radar active interference identification method based on multi-loss characteristic self-calibration network | |
| Wang et al. | TS-I3D based hand gesture recognition method with radar sensor | |
| CN104732240B (en) | A kind of Hyperspectral imaging band selection method using neural network sensitivity analysis | |
| Liu et al. | Crop disease recognition based on modified light-weight CNN with attention mechanism | |
| CN104063719B (en) | Pedestrian detection method and device based on depth convolutional network | |
| CN111368896A (en) | A classification method of hyperspectral remote sensing images based on dense residual 3D convolutional neural network | |
| CN115222994A (en) | Hyperspectral image classification method based on hybrid spectrum network and multi-head self-attention mechanism | |
| CN103366184B (en) | Polarization SAR data classification method based on hybrid classifer and system | |
| CN114866172B (en) | Interference identification method and device based on inverse residual deep neural network | |
| Hou et al. | Jamming recognition of carrier-free UWB cognitive radar based on MANet | |
| CN115439679A (en) | Hyperspectral image classification method combining multi-attention and Transformer | |
| WO2021179198A1 (en) | Image feature visualization method, image feature visualization apparatus, and electronic device | |
| Qiao et al. | A dual frequency transformer network for hyperspectral image classification | |
| CN118135392A (en) | Remote sensing image detection method based on dual-temporal interactive enhanced CNN-Transformer | |
| CN119206192A (en) | A multi-core SAR target detection method based on adaptive convolution | |
| CN118898010A (en) | Radiation source identification method based on the fusion of deep learning and DS evidence theory | |
| Pandhiani et al. | Time series forecasting by using hybrid models for monthly streamflow data | |
| Amalia et al. | The Application of Modified K-Nearest Neighbor Algorithm for Classification of Groundwater Quality Based on Image Processing and pH, TDS, and Temperature Sensors | |
| CN116030300B (en) | A progressive domain adaptive recognition method for zero-shot SAR target recognition | |
| CN117910346A (en) | A robust unsupervised domain-adaptive target localization method enabled by generalization theory | |
| CN115329804B (en) | A method for burst recognition in solar radio spectrum images based on attention mechanism | |
| Shang et al. | Knowledge Distillation Based on Adaptive Learning and Channel Amplification Features for PolSAR Image Classification | |
| Peng et al. | Toward masked face recognition: An effective facial feature extraction and refinement model in multiple scenes | |
| CN116631449A (en) | Speech emotion recognition method based on multi-scale space-time attention mechanism | |
| CN115329821A (en) | Ship noise identification method based on pairing coding network and comparison learning |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |