Applying state of the art machine learning techniques to a suitable problem. Improve the effectiveness of maintenance workers, make their lives easier, and preserve the wonderful countryside train service.
What it does
Identifies the location of malfunctioning equipment by detecting it in open-air photos. Uses this location to provide a map and directions to the human maintenance worker who will repair it.
How we built it
Train a convolutional neural network on some provided training data to recognize the devices of interest Use coordinates in meta-data of each image to place devices on a map Also use these coordinates to map the train's movement over land, thus providing a function from distance along the track to coordinates in space Compose this function with the given function from device identifiers to distance along the track, thus allowing us to determine gps coordinates given a device id
Challenges we ran into
High resolution images which are computationallly expensive to process Discrepancies between different sources of provided data, eg pdf sketches of tracks and digitized distances along tracks Inaccuracies in gps measurements
Accomplishments that we're proud of
A convolutional neural networks that achieves 99% classification accuracy on unseen and class-balanced data A clear and effective user interface Using multiple disconnected data sources to build a valuable function from device identifiers to map coordinates and routes
What we learned
A direct experience of using academic ideas from ai and machine learning to solve real-world problems That the difficult part of a project is often not finding enough sources of information, but rather of linking them together in the right way How much can be accomplished in a weekend!
What's next for Trained to Train
Sleep :) then maybe more practice at model building and development so as to continue to create satisfying and effect tools