The segmentation-by-classification method has become a popular way to detect sound events. It uses neural networks, trained to detect sound sources on short signals (one tenth of a second to a second), or long-term signals divided into a succession of short segments. In this paper, we focus on detecting train pass-bys from long term signals, where it is assumed that railway noise is higher than ambient noise. Using the neural network YAMNet, we show that a processing stage is necessary to improve the segmentation of such events and reduce the high false positive rate. The applied criteria are related to the nature of a train pass-by, lasting several seconds and with broad frequency band. Around 90% of events are detected with proper boundaries. However, we observe that YAMNet is not designed to distinguish railway vehicles from other vehicles. False positive rate remains high whenever other vehicles travel near the measurement location. Additional AI models are proposed to perform this specific distinction. Their efficiency depends on train type, as recent passenger train noise resembles roadway noise. Otherwise, false positive rate decreases below 10%, and the railway noise contribution estimate aligns with the reference level, within a 0.5 dB(A) margin on average.