Feature data set creation method for loose particle localization based on multi-channel characteristic encoding

The existing research on loose particle localization using machine learning methods takes one segment of loose particle signal as the classification object, and uses the single-channel feature data set creation method to create localization data sets, ignoring the correlation and complementarity between four-channel loose particle signals, thus the trained classifier achieved a limited classification accuracy of 84.08%. In this paper, for the first time, the authors took four-channel loose particle signals as the classification object, and seriously considered the correlation and complementarity between them, thus proposed a feature data set creation method for loose particle localization based on multi-channel characteristic encoding. Specifically, for the four-channel loose particle signals, the three-threshold pulse extraction algorithm was used to extract effective pulses, the barrel-principle-based pulse matching algorithm was newly proposed to match effective pulse groups. The values of the eleven applicable time–frequency-domain features were respectively calculated on effective pulse groups, the mean of the four values corresponding to the same time–frequency-domain feature was calculated, and eleven new values were obtained. Four-bit encoding was newly introduced to clearly quantify the correlation and complementarity between four-channel loose particle signals, thus four values were obtained. On this basis, fused feature vectors were constructed, the encoding localization data set was created. Test results on one type of spaceborne electronic equipment show that, compared with the localization data sets created by existing methods, multiple classifiers trained on the encoding localization data set achieves the highest classification accuracy. The feasibility and superiority of the proposed method are fully demonstrated. The in-depth validation results represented by random forest show that, it achieves the most significant and stable classification effect on the encoding localization data set, with the least time loss. The stability and efficiency of the proposed method are fully demonstrated. Currently, the proposed method achieves the highest classification accuracy in loose particle localization research, at 94.03%, an increase of 9.95% compared to previous.