Study Of γ Spectrum Characteristic Recognition Technology Based On Similarity

The latest changes of nuclear security situation at home and abroad have put forward new demands for not only the enhancement on the management and supervision of nuclear components and nuclear materials but also the international non-proliferation mechanism, however, there still exists a big gap between demand and development of the corresponding verification technology for nuclear arms control and disarmament; what’s more, with the rapid development of China’s nuclear industry, it’s become increasingly urgent to improve the processing technique and the control ability for the vast nuclear materials and nuclear wastes consequently produced. All these above require the timely research and development of the relevant non-destructive analysis techniques and recognition technologies. For the above mentioned nuclear objects, y-ray radiation is one of the most important outgoing particles, and the judgment and recognition can be realized for measured objects under the nondestructive conditions through high-resolution γ-ray spectra analysis.This paper centers on the object characteristic Recognition technology based on Similarity of high resolution γ spectra and combines the pattern recognition technology and image registration technology which have witnessed rapid development in the past ten years. It researches on the classified recognition of γ spectra objects, aiming to enhance nuclear safeguard technology, to provide technical reserve for nuclear arms control and nuclear disarmament and to offer support for further research work. Researches in the paper are made on the basis of VC++2010 and related programs have been developed in the process in combination with ROOT. Major works and research results accomplished are as follows:1) Analyzing the high resolution γ spectra with traditional γ spectra analysis methods. Main analysis works include spectral smoothing and denoising, spectral peak searching, background continuum deduction, peak area fitting, etc. Peak-searching algorithm refers to symmetrical zero area method and deconvolution method, background continuum deduction is conducted by SNIP and peak area fitting is realized with the method of AWMI. For the first time, the research combines wavelet denoising and polynomial least squares smoothing to realize smoothing and denoising for the spectra and the smoothing and denoising degree is enhanced.2) Putting forward the concept of γ spectra registration specific to their correction with a reference to image registration technology and initiating the research on γ spectra registration. The research on character-based γ spectra registration covers γ spectra preprocessing, character extraction, character matching, interpolation resampling, similarity measurement, etc. The paper focuses on researching the effects of γ spectra interpolation of four interpolation reconstruction algorithms, including Linear Interpolation, Lagrange Interpolation, Hermite Interpolation and Cubic Spline Interpolation, and results show that the effects of Hermite Interpolation is better than that of others. The similarity measurement is researched with algorithms such as χ2/(N-1), Goodness of Fit, Euclidean Distance, Cosine Similarity, Pearson Product-moment Correlation Coefficient, etc. After interpolation resampling, the similarity measurement is researched following the multilevel registration after wavelet transform.3) Researching the classified recognition of γ spectra by Principal Component Analysis, Linear Discriminant Analysis and Support Vector Machine in pattern recognition. For the first time in China, Support Vector Machine research and linear discriminant analysis are applied in γ spectra analysis.From the perspective of classification results, Linear Discriminant Analysis is superior to Principal Component Analysis while the physical significance of projection space of the latter is relatively obvious. The classification training of Support Vector Machine for the high resolution γ spectra in plutonium sample demonstrates that cross validation can reach an accuracy rate of above 99% in classified-learning.