| Specific emitter identification is one of the key technologies for spectrum sensing and understanding,and is widely used in military and civil fields.Its main processes include preprocessing,feature extraction and classification.However,the large amount of specific emitter signal data and the high extracted feature dimension disaster increasingly serious.Learning high-dimensional signal features will spend a lot of computing time and storage space,which is not conducive to scientific research and practical application.Aiming at the above problems,this paper studies the feature selection problem for specific emitter identification.(1)Too large search space in feature selection often makes the feature selection problem unable to be solved effectively.The swarm intelligence optimization algorithm does not need domain knowledge and make any assumptions about the search space,the advantages can be used to solve the problems.A specific emitter identification method based on ant colony optimization feature selection is proposed.Twelve statistical feature parameters and normalized relative energy are selected,lifting wavelet packet decomposition and reconstruction method is used to extract features and build feature parameter system.A mathematical model of feature selection with the goal of maximum classification accuracy and minimum feature subset size is given.The improved ant colony algorithm is used as the search strategy of wrapped feature selection.Based on the construction diagram of the subset problem,the path transfer probability is used to search the path and the phenomenon update strategy formula based on equivalent path enhancement is used to solve the emitter signal feature selection model.Compared to the traditional search strategies in feature selection,the maximum classification accuracy obtained based on ant colony optimization in the original signal,signal to noise ratio of 10 d B,and signal to noise ratio of 5d B has been improved,and the minimum feature subset size obtained is also relatively small.(2)In order to improve the operational efficiency of specific emitter identification,a specific emitter identification of LightGBM based on ant colony parameters optimization is proposed using embedded feature selection.This method combines lifting wavelet packet transform,ant colony optimization,LightGBM algorithm and feature selection method.Ant colony optimization is used to optimize the six parameters of LightGBM.The value of the parameters will affect the feature selection and final classification results.The optimized LightGBM is used to obtain the importance value of each value and sorting it.The sequence backward search strategy is used to select features.Comparing the classification accuracy of each feature subset in the search process to obtain the best feature subset.Compared to traditional embedded feature selection methods,the maximum classification accuracy obtained by ACO-LightGBM in the original signal,signal to noise ratio of 10 d B,and signal to noise ratio of 5d B has been improved,and the minimum feature subset size obtained is also relatively small.(3)Aiming at the low classification accuracy of embedded feature selection and the low efficiency of wrapped feature selection,a hybrid feature selection method is proposed.Three embedded methods are used to calculate and sort the importance value of each feature in the first dimension reduction process,and the first several most importance features are selected to input into the second dimension reduction method.The wrapped feature selection method is used for the second dimension reduction.The sequential backward search strategy and the ant colony optimization are used as search strategies respectively.Combining the process of two dimensional reduction processes,six hybrid feature selection models are generated.Experiments are carried out to select the most suitable hybrid feature selection method for each feature set.Compared to the feature complete set and only one dimensionality reduction method,the maximum classification accuracy obtained by hybrid feature selection in the original signal,signal to noise ratio of 10 d B,and signal to noise ratio of 5d B has been improved,and the minimum feature subset size obtained is also relatively small.In view of the low accuracy and low efficiency of specific emitter identification,focusing on the feature selection methods to improve the accuracy of specific emitter identification and reduce feature dimension.Guided by the technology involved in specific emitter identification and feature selection,the research is carried out on feature extraction,wrapped feature selection,embedded feature selection and hybrid feature selection.Corresponding methods are proposed respectively,which has theoretical value and application value for solving the problems caused by high dimensional data in specific emitter identification. |
