Abstract: A better knowledge about the interrelationship between convolutional neural network （CNN） performance and its sample selection, training procedure, structure, etc., will be beneficial to employ this technique efficiently. We use CNN to detect tele-seismic P-S and local seismic Pg-Sg phases recorded by the Beijing National Earth Observatory. From the results by different parameter associations, it shows that the moderate layer depth, proper regularization, and data wash can significantly enhance the CNN performance while residual blocks giving only marginal improvement. Furthermore, we employ class model visualization and smooth GradCAM++ techniques to interpret the optimal CNN model. The results show that our model has learned the fundamental features of the seismic phases, with decision-sensitive distribution agreeing well with a priori knowledge. Also we use CNN model to scan the continuous seismic waveform, which exhibits its potentiality in seismic phase detection. Lastly, topics on sample selection, model framework, sample labelling, and ensemble learning are discussed for further work.