ConocoPhilips Challenge

SOM

Inspiration

Sooner or later, hardware is going to stop functioning. Failures are unexpected, costly, and affect everyone negatively. One way to minimize the repercussions of a hardware failure is to predict under what circumstances it is going to fail. An issue that arises in predicting failure is the fact that it might be very difficult to find a correlation between the values for which it might fail, especially in this challenge, which has readings from 107 sensors, several of which have 10 readings as a histogram. Seeing patterns with such a huge dataset is not humanely possible.

What it does

This is where Artificial Neural Networks (ANNs) step in. With sophisticated algorithms developed by mathematicians and scientists over several years, we can use ANNs to find patterns among large amount of data and predict things that aren’t possible to otherwise. Our team developed an ANN model to predict if, given the sensor values, the fault is going to be down-hole or surface.

How I built it

There are several models of neural network that are available free of use for public domain. We decided to use Keras with TensorFlow for backend because that was what our team had experience with, and it is popular for a reason—it is straightforward to implement in Python. We tuned some hyperparameters—we tuned the learning rate to a reasonably low value because the prediction rate was very high. We left the other hyperparameters to their usual values and saved the trained model. We also used SOM (miniSOM) to group the parameters together.