Research On Intelligent Identification Algorithm Of Radar Radiation Source Based On Deep Neural Network

Radar radiation source identification is a key link in the process of electronic reconnaissance system,which provides the necessary intelligence support for decision-making by identifying the signal type of the target radiation source.However,the non-cooperative nature of the radar radiation source to be identified makes the source identification task face many difficulties and challenges,which can be summarized as the low signal-to-noise ratio of the acquired radar radiation source data,small training sample set,limited data types covered by the training sample library and insufficient arithmetic power of miniaturized hardware.These difficulties limit the application of radar radiation source identification technology in practical engineering.To address the above difficulties,this paper conducts an in-depth study on radar radiation source identification technology,which includes the following four aspects:(1)For the problem of low signal-to-noise ratio of acquired radar radiation source data leading to difficult model feature extraction and degradation of recognition performance,this paper designs a new radar radiation source recognition algorithm based on Transformer and feature fusion.The algorithm firstly adopts a deep residual systolic network for soft threshold denoising of the radiation source signal to improve the signal-to-noise ratio of the input signal;then adopts a Transformer encoder for temporal modeling of the radiation source features to capture the bi-directional correlation of the internal features of the radiation source samples;finally,the fused features are processed by the phase-amplitude time-domain sequence to improve the problem of single input features.The experimental results show that the proposed algorithm can achieve better recognition results under the condition of low signal-to-noise ratio signal.(2)In order to improve the problem that the model is easily overfitted due to the insufficient number of training samples,a new semi-supervised learning-based radiation source classification algorithm is proposed in this paper.The algorithm firstly adopts a semi-supervised algorithm framework and uses simulated RF fingerprint data to improve the problem of plummeting recognition performance due to insufficient number of labeled samples by mixing pseudo-labeled and labeled training forms;secondly,the problem of lack of attention to multi-scale and global features of radiation sources is improved by combining the selected convolutional kernel module and the feature encoding module in the designed recognition model.Experimental results demonstrate that the proposed algorithm not only improves the adverse effects of insufficient training samples,but also shows excellent recognition performance with a larger number of samples.(3)To address the problem that the traditional recognition model for new radiation source task data cannot be effectively scaled and suffers from catastrophic forgetting,a new incremental learning algorithm for radar radiation sources based on the conditional diffusion model is proposed in this paper.The algorithm first generates the historical data sample set by the conditional diffusion model;then mixes the generated historical data sample set and the new category data set to obtain the current overall data set to train the classification model;finally,the conditional diffusion model is trained to have the ability to generate the current overall data set.The advantages of this algorithm are: firstly,the migration process of past task knowledge does not need to store and revisit the real historical data of radar radiation sources,which is more suitable for long time and small batch update learning;secondly,compared with the online learning model based on regularization method,the overall parameter drift of this method is smaller,and it has better performance under the practical engineering conditions of multiple sample acquisition and multiple batch learning;finally,this algorithm has better performance in Finally,the algorithm has better data confidentiality in practical engineering.(4)To address the problem of insufficient computing power of miniaturized hardware and the difficulty of real time edge computing in practical engineering,this paper builds a software and hardware platform for radar radiation source identification based on Jetson Nano development board and Tensor RT inference engine,and achieves model acceleration at the edge based on this platform.The experimental results verify that the proposed radar radiation source identification method has high real-time performance and noise robustness on embedded devices with low computing power,which lays the foundation for the future application of the algorithm in practical engineering.