Application of lightweight neural network models for nuclear pulse parameter prediction

BackgroundIn nuclear radiation measurement, pulse distortion is inevitable due to the interference of the measurement system itself and the measurement environment. If the parameters of such pulses cannot be accurately estimated, the resolution performance of the energy spectrum will be reduced.PurposeThis study aims to accurately estimate the height of distorted pulses using neural network model.MethodsFirstly, six lightweight neural network models, i.e., LeNet5, LSTM, GRU, UNet, CNN-GRU, CNN-LSTM, were applied to parameter prediction of distorted nuclear pulses, including pulse amplitude parameters and distortion time parameters. Then, based on the distorted pulses generated by predefined mathematical models, the dataset required for model training was obtained through digital triangulation shaping. Finally, parameter prediction performances of those neural network models on test set with additional white noise, Gaussian noise and flicker noise were compared with each other, as well as with the traditional digital forming method.ResultsWhen evaluating the parameter prediction performance of six neural network models, the UNet model achieves the lowest relative error on the test set, with a relative error of approximately 0.57% for amplitude parameters and 3.51% for time parameters. In the signal-to-noise ratio experiment, noise is superimposed on the test set to obtain noise test sets with different signal-to-noise ratios.ConclusionsThe results of this study show that the proposed models can achieve accurate estimation of the parameters of distorted pulses.