This project contains a Visual Studio solution with a project to recognize train identification numbers. Results are not promising though. So this represents of how not to do it.
See https://en.wikipedia.org/wiki/UIC_identification_marking_for_tractive_stock
This repository contains the Tensorflow models to do train detection.
Further along there might be a model to detect isolation deficiencies.
python -m scripts.retrain `
--bottleneck_dir=tf_files/bottlenecks `
--how_many_training_steps=4000 `
--model_dir=tf_files/models/ `
--summaries_dir=tf_files/training_summaries/"$ARCHITECTURE" `
--output_graph=tf_files/train_graph.pb `
--output_labels=tf_files/train_labels.txt `
--architecture="$ARCHITECTURE" `
--image_dir=tf_files/trains