Signal Subtle Feature Dimension Reduction Optimization And GPU Parallel Optimization

The identification of communication satellite signal source is of great significance for modern electronic warfare.Signal subtle feature analysis is a very important part of signal recognition.Due to the increasingly complex electromagnetic environment and the rapid development of radio technology,the identification of signal source is more High requirements.The signal subtle features commonly used have higher feature dimensions and computational complexity.There are two problems: First,the development of signal acquisition technology,the calculation of the subtle features of the signal is getting larger and larger,and the operation time becomes longer;The feature dimension is getting higher and higher,which will lead to curse of dimensionality.In thesis,we use two kinds of signal subtle features: bispectrum and fractal dimension of envelope.The two features are manifolded and reduced as the input of the classifier.The spectral transformation algorithm performs GPU(Graphics Processing Unit)parallel optimization.First,the principle of the subtle features of the two signals is analyzed.Then,for the dimensional disaster of high-dimensional feature data faced in signal recognition,the traditional dimension reduction method in machine learning is introduced,and the principle,applicable environment,advantages and disadvantages of the dimension reduction algorithm are analyzed.The manifold dimension reduction algorithm is applied to the reduction dimension reduction of signal subtle features.The idea and calculation method of typical manifold dimension reduction algorithm are analyzed.The performance of different dimensionality reduction algorithms is analyzed by using the signal characteristics of actual data to reduce the dimensionality of data points in manifolds.Finally,the signal recognition technology based on t-SNE(tDistribution Stochastic Neighbour Embedding)algorithm and SVM(Support Vector Machine)classifier is studied.The steps of the single feature recognition process are optimized,including changing the bispectrum subtle feature extraction algorithm of the signal to the GPU-based parallel bispectral extraction algorithm,the influence of the feature dimension on the classification and the reason of the communication satellite signal feature layering.On this basis,the t-SNE algorithm is used to reduce the dimension of different features,and is classified by SVM classifier.Depending on the electromagnetic environment,using three sets of signal data are used to five experiments.The results show that the GPU-based parallel bispectral feature extraction algorithm can greatly improve the speed of feature extraction.The subtle features of the dimensionality reduction using the t-SNE algorithm will affect the recognition accuracy,but basically preserve the topological information of the samples in the highdimensional features,and reduce the computational amount when dealing with the lowdimensional features after the dimensionality reduction.The real-time nature of classification recognition is significantly improved.