Thief Hatch Neural Network

• 

  Example Data

• 

  Neural Network Training

• 

  Prediction Comparison

• 

  Confusion Matrix

(This project is for the general track)

Inspiration

We actually got the idea for this project from school. An energy company (who will remain anonymous) posed a challenge to help them solve a problem that they were having. In oil fields, they have oil tanks which have something called a 'thief hatch', which helps relieve pressure and other safety issues. Sometimes they get left open, and gas leaks into the environment. They have pressure sensors in some of the tanks, and the challenge is to make a model that can read the data, and predict whether or not the thief hatch is open.

What it does

Our program builds a training set of data that we manually verified, then uses that to train a neural network. The neural network model can then take any dataframe and time, then predict whether or not the thief hatch is open, with a certain confidence. As of right now it works with about 96% accuracy.

How we built it

The program is all in python, and we used the pandas library to manipulate the data. The neural network itself was made with tensorflow and keras.

Challenges we ran into

Neither of us had ever touched a neural network before this hackathon, and our knowledge on the subject was close to non-existent.

Accomplishments that we're proud of

We are proud that we were able to get a working model within 24 hours. When we started, we weren't even sure we were going to have a working model by the end. Persevering and not giving up definitely played a big part in our success.

What we learned

We learned a ton about neural networks, how they work, and how to make one. Like we said, our knowledge on the subject was extremely thin, and we feel like we have come a very long way during this competition.

What's next for Thief Hatch Neural Network

There is definitely a lot of room for improvement in the neural network, there are a lot of parameters that we would like to fiddle around with. There is also more data that we got as a part of the challenge, so we would like to try to work that into the model to make it even more accurate.

Note: The code is available to view, but the data is not available for privacy reasons.

Built With

• keras
• pandas
• python
• tensorflow