Predictive Equipment Failures

Inspiration

When costs go up, cash goes down. There is a need to bring the failure costs down and to do that we must learn from our data about failures.

What it does

Given various data from various sensors, our model tries to predict if failure event will occur on surface or down-hole, so that it can be addressed well before it occurs.

How we built it

We came up with our model after trying out many plausible classification techniques. Our models were developed using the scikit-learn, pandas and numpy libraries. We tried following techiques - Regression, Regression with regularization, Random Forest, Support vector machines and finally Adaboost with random forest. We got best results from Adaboost with random forest.

Challenges we ran into

We had to look up and try various models. Not all models gave good results. Those which gave promising results without proper hyper parameters, the models were over-fitting most of the time. It took number of attempts to get the model training right and to get good results.

Accomplishments that we're proud of

We are quite happy with our model performance given we could dedicate only limited time because of other commitments. What we achieved in that limited time frame at the expense of our not so great computing resource ( our laptops!) is quite overwhelming.

What we learned

We learnt a great deal on implementation perspective. It was possible to try few other complex techniques to increase model score further, something to look forward to in coming challenges in future.