trainAI

User Interface for alerting technicians
Taking the Technician to the faulty device on the tracks
Logical Breakdown of the App
Team Work!

Inspiration

Use AI to identify the gps position of field devices and guide technicians to the exact position of faulty devices

What we don't know

Having a set of pictures from the track, we don't know 1) Object - It could of 3 types

Main Signal
Distance Signal
Antenna

2) GPS Location Detection of Objects

We have two GPS locations

Mapping of object IDs with their stored GPS locations
GPS Location embedded in the image

What we do

Pinpointing faulty equipment on train tracks using AI to map the location of equipment from video taken from train cockpit to have accurate GPS location

How we built it

Frontend: Android for the navigation to the faulty devices Backend : Keras and tensorflow to identify the devices from the images Python to match the output to the documentation data And extract a data base for the frontend

Challenges we ran into

-Identifying the devices as they don't have a unique ID -having split tracks with different devices without knowing the exact path of the train

Accomplishments that we're proud of

An android application with a built-in Ticketing System that takes the technician to the faulty issue using Google Maps
The proof of concept of building a final database of object ID and their GPS coordinates
Solving the issue of parallel tracks and detecting signals within same GPS location
Implementing State of the Art Object Detection Model on GPU using Faster RCNN and getting an accuracy of 81%.

What we learned

Limitations of CPU on training for Machine Learning Model
Training on GPU using of tensorflow-gpu
Limitations of the architecture which subsamples the original image and changes the input shape. The inconsistency in the input shape of the model will not allow for TPU processing
CPU takes 30/50 and even 90 hours per epoch depending on the system
GPU allows us to do faster operations using 4 hours per epoch
Data Size Reduction to half for faster training

Unexpected Issues

Dataset: Uploading the huge dataset on Google Drive as AWS was not operational

What's next for trainAI

Identification of foreign objects and debris on the train tracks and tagging its location Adding multiple tracks and detecting objects using transfer learning