Abstract: As an important pillar of modern science, the safety of nuclear technology is highly dependent on accurate neutron detection, and the n/γ discrimination technique in the composite radiation field is directly related to the reliability of detection. Among the n/γ discrimination techniques, the pulse shape discrimination (PSD) method is an important technique in nuclear radiation composite measurements, which works on the basis of the different decay times of fluorescence excitation of different charged particles in scintillators. In order to solve the problems of high hardware cost and hardware impact that existed in the early implementation of the PSD method through electronics hardware, the use of digital technology to achieve n/γ discrimination has gradually become the mainstream. The traditional discrimination methods are time-domain PSD method and frequency-domain PSD method. Among them, the time-domain PSD method has the problem of being affected by noise, and the frequency-domain PSD method has the problem of high hardware requirements. Currently, neural network-based discrimination algorithms were used to screen signals by extracting the variability of their internal information features, so they can circumvent the shortcomings of the time-domain and frequency-domain PSD methods while taking into account the high discrimination effect, which can greatly improve the upper limit of the discrimination performance, even though it requires complex algorithmic modelling and training processes, and has a large operating overhead. Therefore, in this study, in order to make full use of the dual features of time and frequency domains of impulse signals to improve the n/γ discrimination performance, an n/γ discrimination algorithm based on convolutional neural network, attention mechanism, and long and short-term memory neural network (CNN+Attention+LSTM) was proposed. Based on the CLYC detector experimental platform to obtain n/γ hybrid pulse data, the experimental data were screened using the charge comparison method and frequency gradient method. In this study, an intelligent n/γ discrimination was implemented using the module splicing method. Firstly, a sliding convolution filter was applied to the input to process the sequence data via CNN, and the data frequency domain features were extracted for discrimination using convolution kernel and pooling operation to reduce the dimensionality and retain the key modes. Next, feature weights were learned through network loss in the attention mechanism, allowing the model to focus more on important information. Finally, the sequential data were processed through the long-term dependency between the LSTM loop time step and the learning time step, using a gating mechanism to selectively memorize the long-term patterns and extract the data time-domain features for discrimination. This approach enables better feature discrimination by allowing effective feature maps to receive greater weights while ignoring unimportant features. The results show that the CNN-Attention-LSTM discrimination method can effectively improve the effectiveness of n/γ discrimination and provide a new idea for utilizing deeper impulse features.