Inspiration
The project was based around the ConocoPhillips challenge to predict equipment failures using the given dataset.
What it does
The model takes in a wide variety of data from 107 sensors to predict the equipment failure.
How I built it
I built the model using Google Colab and the libraries listed above. I first converted the data from the csv file into an array and performed some initial refining of the data, including converting the "na" values for the sensors into a negative number since none of the sensors are able to read negative values. The data was then randomly shuffled and passed into the model. The model consists of a Dense layer that takes the input tensor of size 170 and converts it to a tensor of size 16. This is then followed by a dropout layer with a dropout rate of 0.15, which drops the values of about 15% of the input values, preventing the model from overfitting. This is then passed into another Dense layer which converts the tensor of size 16 to a tensor of size 32. This is finally passed into a Dense layer of size 1 with a sigmoid activation. This returns a value between 0 and 1 that represents the probability that the target value is 1. The model was optimized using a RMSprop optimizer. The training data was split into training and validation data, with training data taking about 80% of the data and validation taking the rest. The model was then tested on the test data, and the predictions were uploaded to Kaggle. The model achieved a Mean F1 score of 0.99125.
Challenges I ran into
Some of the challenges I ran into was trying to find a good model architecture. I first started with several Dense layers, and realized that the model was a little too complex and could be simplified. I then tested several other model architectures such as a Recurrent Neural Network and a simple Dense layer model. I finally tested a simple Dense layer model with dropout and it increased the performance somewhat, and it was the model that I finally settled on.
Accomplishments that I'm proud of
I'm proud of the wide variety of models that I tested and was able to get to work.
What I learned
I learned how certain ML techniques can vastly effect how the model learns and performs.
What's next for ML solution to ConocoPhillips Challenge
The model could definitely be further improved to achieve a better Mean F1 score and gain better accuracy. This could be done by trying even more model architectures and playing around with tensor sizes and dropout rates.
Log in or sign up for Devpost to join the conversation.