Abstract:At present, electromagnetic environment is becoming more and more complex. Using machine learning methods to achieve radio direction finding has become a research hotspot. The existing research work attempts to use convolution neural network to complete broadband radio direction finding based on end-to-end method. It can solve the problem of wide-band phase ambiguity. However, the dimension of the feature after convolution increases greatly, which results in the phenomenon of feature sparse and further, affects the performance of the fully connected feedforward neural network to a certain extent. To solve this problem, we divided radio direction finding into feature learning task and direction prediction task, and used convolution neural network as feature extractor to obtain secondary extracted features,which were the inputs of the direction prediction task. Further, we used the principal component analysis algorithm to reduce the dimension of the features. In addition, we explored the performance of several classification models as the final classifier, including decision tree, random forest, radial basis function neural network and K-nearest neighbor. The experimental results showed that using principal component analysis algorithm to reduce the dimension of features could improve the efficiency of training and testing, and the accuracy of K-nearest neighbor classifier was significantly higher than that of the original convolution neural network. If both accuracy and efficiency were considered, random forest classifier was the best.
