Algorithm Research Of γ-ray Spectrum Analysis For UAV Radioactivity Monitoring Application Based On Artificial Intelligence

Radiation monitoring system is a prerequisite guarantee for dealing with the problems of sudden nuclear accident,radioactive pollution and loss of radioactive sources.UAV radioactivity monitoring system has high mobility and flexibility,and has been widely used in radiation monitoring.Due to the limitation of load and time of flight,unmanned aerial vehicles have put forward higher requirements for the fast measurement of gamma spectrum analysis.The traditional γ-spectrum analysis method has high requirements for energy spectrum data,and has a poor recognition effect on low-count and low-resolution γ-spectrum.It does not perform well in UAV radioactivity monitoring systems.Aiming at the remaining shortcomings of rapid measurement of energy spectrum nuclide identification of drones,this topic proposes a multi-nuclides identification method based on deep learning,and the algorithm is implemented in a small drone monitoring system Rotor designed by this research group.Application verification.The research contents and conclusions of this article are as follows:(1)Based on MCNP simulation,a nuclide data set for deep learning training is established.Using the characteristic peaks of different radioactive sources to fit the relationship between energy and FWHM,the γ energy spectra of different particle number responses are simulated according to MCNP,and a data enhancement method is proposed.The simulated γ energy spectra are randomly combined,adding different backgrounds and energy spectrum drift,etc.Way to increase the energy spectrum sample capacity of the nuclide dataset.The results show that the simulated energy spectrum is close to the measured energy spectrum,ensuring the generalization ability and reliability of subsequent algorithms.(2)A hybrid model nuclide recognition algorithm based on CNN and MLP is proposed.The process is to identify possible radionuclides in the gamma spectrum by CNN,and further analyze the overlapping peaks of the output results in combination with MLP to eliminate misidentified nuclide.The CNN structure is an optimization improvement based on Alex Net,including the input of two-dimensional transformation of the gamma energy spectrum,and the acceleration of network convergence at the BN layer.By analyzing and selecting the overlapping peak ROI in the nuclide library,the overlapping peaks were resolved by MLP.Experimental results show that CNN has high accuracy and generalization ability,and can respond well to multi-source and drift energy spectrum.The proposed hybrid model algorithm also effectively improves the recognition accuracy.(3)An energy spectrum thinning algorithm based on SNIP is proposed.The background subtraction effect of SNIP is modified by the detector energy scale.The iterative calculation of background subtraction and convolution thins the energy spectrum vector.The experimental results show that the algorithm can effectively retain the energy spectrum characteristics,and enhance the counting ratio of characteristic peaks in the energy spectrum to suppress noise and background interference.Compare the effect of the original energy spectrum and the energy spectrum processed by CSNIP on the recognition results of the neural network.The latter energy spectrum loss function can effectively converge,and it can improve the accuracy rate by 4.5%under the same data set.The research content of this article has a very good effect on the application of Rotor drones,provides a guarantee for the rapid measurement of drones and complex environmental monitoring,and helps to improve the automatic nuclide identification and activity measurement of existing radiation monitoring instruments Play a greater role in emergency monitoring.