TrackBuster

Keeping train tracks under control via live status monitoring, anomaly detection and stress prediction. We leverage robust statistical correlations in dataset provided to ensure guaranteed control.

Comment

Inspiration

Of utmost importance is a clean, fast, easy-to-use interface that shows reliable and robust predictions regarding future anomalies. We drew inspiration from nuclear-reactor dashboards: a powerful example of how to show many and important information clearly. Last but not least we got inspired by Ghostbusters..

What it does

  • live status monitoring of train tracks
  • anomaly detection (via multi poly-fit regression)
  • stress prediction and predictive maintenance of track anomalies
  • ensuring guaranteed control by combining and leveraging emergent statistical correlations
  • comparison and optimization of in-place works
  • visualizing the main results in an intuitive and user-friendly way

How we built it

  • Python
  • React
  • scikit-learn
  • R
  • GitHub
  • Asana

Challenges we ran into

  • UI building (not enough experience in this team)
  • accuracy of model results (ideally more data should be available)
  • coordination between team members (hybrid +3h timezone delta)
  • lack of coffee
  • lack of sleep

Accomplishments that we're proud of

  • timeframe of 120 days (best found yet) allows high-level precision and confidence of time and place of prediction
  • submission (nearly) on time

What we learned

  • data visualization is the first step to gather knowledge about a dataset
  • we need more coffee
  • we need to sleep more

What's next for TrackBuster

  • model validation on bigger dataset
  • UI-refinement
  • backend-scaling (+ moving to a SaaS-architecture)
Share this project:

Updates

posted an update —

We firstly filtered out the data and reduced it by aggregating all of the values for a day-by-day basis for different areas and tracks using mean, positionnoleap, A2_RSSI, A1 and A2 ratio of signal sent vs signal received and current velocity. We then first draw a polyfit over time vs A2_RSSI value by converting time to ordinal format and then creating a 1D regression between time and A2_RSS1. For this we used the input from user of the positionnoleap, date (from which date and date-120 days range is taken), areanumber and track. We isolate the track by track and area by area regression and then calculate a polyfit regression between positionnoleap values for the given area and track with A2_RSS1 for the same over the period of time that has been provided by the user. We average out A1 signal loss ratio and A2 signal loss ratio and create another 1D polyfit wit respect to A2_RSSI. Upon finding the slope and intercept from the 3 regression polyfit models. We use the slope and intercept from the time vs RSSI value to first calculate time for achieving weak, fair, good and strong A2_RSSI signals. We then identify for the given user input of positionnoleap, what would be the value of y by using slope_2 and intercept_2. We then multiply this value with the 3rd regression model slope and subtract from the time intercept to accurate the time prediction further. We did not have time to successfully use CurrentVelocity, disruption code and event code to further improve the accuracy for the system. Moreover, in some cases we had less datapoints for some tracks and area numbers for which we could not aggregate many values and create a better regression model but our model seems to work good for about 79% of areanumber, tracks and date values. We also don't have metric to cross validate it yet. In case of any questions please let us know!

Log in or sign up for Devpost to join the conversation.