n-γ discrimination in low energy region using Artificial Neural Network based on improved traditional methods

Abstract The n- γ signals in low energy region are difficult to be discriminated because of the ambiguous energy loss, partial overlapping of energy spectra and the presence of noise. In this paper, an n- γ discrimination method combining the improved traditional methods with Artificial Neural Network (ANN) in low energy region is proposed. Firstly, this paper improves the Charge Comparison Method (CCM) and Discrete Wavelet Transform (DWT). With respect to the original method, the discrimination parameter of Improved CCM adds different charge time distributions, and the discrimination parameter of Improved DWT adds the second smallest scale. Then, the n- γ data co-screened by Improved CCM and Improved DWT are used as the dataset of ANN. These low-energy data co-screened by Improved CCM and Improved DWT are more discriminant, which is conducive to improving the reliability and generalization ability of the ANN model. Finally, elastic net regularization technique is added to the ANN model, different ANNs are compared using F1 score ( F 1 ) and Discrimination Error Ratio (DER). In this paper, the pulse waveforms generated by the EJ301 liquid scintillator detector in a real environment are processed. The experimental results indicate that, compared to traditional methods, the FoM values of Improved CCM and Improved DWT increased by 7.7% and 76% respectively in 0–25 keV, 2.8% and 28.7% respectively in 25–50 keV. The comparison results of F 1 and DER demonstrate that the ANN using the n- γ dataset determined by Improved CCM and Improved DWT has higher F 1 and lower DER. Therefore, the ANN proposed in this paper, based on improved traditional methods and elastic net regularization, exhibits higher reliability and generalization ability in low energy region. Additionally, increasing the batch size of model is beneficial for higher efficiency.