The Research And Application Of Deep Learning Model In The Recognition Of Piled-up Nuclear Pulses

Digital shaping method has become an important shaping method for nuclear pulse signals.This method uses the relevant principles of digital signal processing to identify nuclear signal parameters,which can improve and optimize the performance of nuclear instruments greatly.However,in actual detection process,regardless of the shaping method used,the pile-up phenomenon between adjacent nuclear pulses is inevitable.Therefore,the identification of pile-up pulses after digital shaping is crucial.Although many scholars and research teams at home and abroad have conducted in-depth research on nuclear pulse shaping,acquisition,and parameter identification,the identification and separation of pile-up pulses after digital shaping is still a problem because of the performance limitations of traditional methods.With the emergence and continuous development of deep learning,some internal neuron structures are more complex,the connection methods between neurons are more diverse,and the deep learning models have deeper layers.Several data can be trained to enable them to perceive and understand external things,such as the human brain.At present,related studies that introduce deep learning technology into nuclear pulse parameter identification and spectrum correction are still in their preliminary exploration stage.This paper initially discretizes the continuous pulse signal and subsequently uses the classic digital shaping method to shape the discrete exponential pulse to make it exhibit the characteristics of time series.Then,an efficient and stable nuclear pulse parameter identification method is studied using the deep learning model's powerful learning ability for complex non-linear time series.Finally,a set of classification and grading deep learning spectrum correction techniques is investigated on the basis of the research of piled-up pulse recognition.Experimental results show:?1?Whether it is Gaussian shaping or trapezoidal?triangular?shaping,under the conditions that the data set is rich enough and the deep learning model is sufficiently trained,the deep learning-based piled-up pulse recognition technology can identify and separate piled-up pulses effectively.?2?The use of a compound deep learning model can improve the calculation efficiency and save time in identifying piled-up pulses.?3?For ¹³⁷Csenergy spectrum,the Na I?Tl?and the N-G275 detectors are used to measure 200 s.When the shaping method uses SK Gaussian shaping and the pulse width is 202 Ts,the probability of piled-up pulses in this spectrum is approximately3.39%.The use of the LSTM model can identify piled-up pulses accurately under the condition that the shortest interval between adjacent pulses is 50 Ts.The deep learning model's ability to resolve piled-up pulses is 74.39%.When the ¹³⁷Csenergy spectrum adopts a trapezoidal shape and the shaping time is 340 Ts,the probability of pulse piled-up of this spectrum is approximately 5.65%.The CNN-LSTM model can identify piled-up pulses accurately under the condition that the shortest interval between adjacent pulses is 20 Ts.The deep learning model's ability to decompose piled-up pulses is 93.96%.?4?The FAST-FDD detector is used as the experimental research platform for high count rate X-ray spectroscopy.For the standard source of ²³⁸Pu,when the detection time is 120 s,and the shaping time is 70 Ts,the probability of pulse piled-up is 16.789%.The CNN-LSTM model can identify piled-up pulses accurately under the condition that the shortest interval between adjacent pulses is 30 Ts.The deep learning model's ability to decompose piled-up pulses is 54.89%.For the ⁵⁵Fe standard source,when the detection time is 500 s and the triangle shaping time is 60 Ts,the probability of pulse piled-up is 2.240%.The CNN-GRU model can identify piled-up pulses accurately under the condition that the shortest interval between adjacent pulses is 40 Ts.The deep learning model's ability to identify piled-up pulses is 33.07%.?5?For the correction of the X-ray spectrum of the ⁵⁵Fe standard source,the FAST-SDD detector is used as the experimental research platform.The deep learning spectrum correction model based on classification and grading can achieve 80.78%piled-up pulse recognition efficiency for the measured adjacent pulse sequence with an average interval time of 145 Ts.Thus,the count rate of the X-ray spectrum of the entire⁵⁵Fe standard source is increased by approximately 4.7361%.In summary,the proposed method can overcome the difficulties in the parameter identification process of noise-containing piled-up pulses effectively.Moreover,the identified parameters have high accuracy and can correct the spectral count rate effectively under the background of high count rate.It is a set of excellent pile-up pulse recognition and spectrum correction methods.