Electromagnetic interference sources (EMIS) must be identified in order to locate them promptly. Because representative features of EMIS broadband signals are difficult to extract, we propose a new identification method based on convolutional neural network (CNN) to extract EMIS deep features from spectrum signals and increase recognition accuracy. To achieve noise reduction, we added a noise reduction layer (NRL) to the network, which uses background noise data as the weight to determine its correlation with the input data. Furthermore, a new loss function based on intra-class and inter-class relative distances is presented, which is paired with the Softmax loss function to make the network converge fast and consistently. Experiments on three data sets are used to validate the created method's overall performance. Simulated results demonstrate that the suggested method can effectively extract the deep features of the EMIS signal, enhance signal classification speed and accuracy, and achieve 100% accuracy on our data set.

必须识别电磁干扰源 (EMIS)，以便及时找到它们。由于 EMIS 宽带信号的代表性特征难以提取，我们提出了一种基于卷积神经网络 (CNN) 的新识别方法，从频谱信号中提取 EMIS 深度特征，提高识别精度。为了实现降噪，我们在网络中添加了降噪层（NRL），它使用背景噪声数据作为权重来确定其与输入数据的相关性。此外，提出了一种基于类内和类间相对距离的新损失函数，与 Softmax 损失函数配对，使网络快速一致地收敛。三个数据集的实验用于验证所创建方法的整体性能。