FACIAL RECOGNITION FOR TIES: DEVELOPMENT OF A DEEP NEURAL NETWORK TO FACILITATE TIE LIFECYCLE STUDY

摘要

Railroads have been collecting massive amounts of track maintenance data in recent years. Most of these datasets have been used for capital planning and immediate remediation actions. The next logical step to facilitate maintenance efforts and maximize efficiency is to develop predictive models; one of such models is a predictive tie life cycle. In order to create a tie life model and successfully analyze decay trends, one must be able to study and assess tie condition of the same asset over many years. This step has proven to be quite challenging; GPS is often not sufficient for positive identification, and regular maintenance and tie replacement efforts further complicate the matter. Placement of physical markers such as tags is not feasible on a large scale, and furthermore, losing markers, a common occurrence when exposed to the elements, would significantly decimate sample size. To solve this problem, Georgetown Rail Equipment Company created a database of Aurora® tie images and used it to develop an image recognition model powered by Embeddings and a Deep Neural Network to match ties from different collections throughout the years. The following paper describes the rationale behind the Artificial Intelligence approach, the creation of the model, the parameters used, results and the authors' future steps.