roldan87/challenge-classification

Challenge Bearing Classification

Description

This was an assignment we received during our training at BeCode.
The main goal was to get used to Machine Learning, specifically using classification algorithms.
For this we used a database from Kaggle, on testing bearings.
Our job was to predict if a bearing was defective or not, with an accuracy as high as possible.

Installation

Python version

• Python 3.9

Databases

Github has a 100MB file limit, so the files can't be found in this repository.
Below however, you can find the links to both files.

• Target
• Features

Packages used

• pandas
• numpy
• matplotlib.pyplot
• seaborn
• sklearn

Usage

File	Description
main.py	File containing Python code.    Used for cleaning and feature engineering the data
plots.py	File containing Python code.   Used for making some explanatory plots for this README.
utils/model.py	File containing Python code, using ML - Random Forest.   Fitting our data to the model and use to it make predictions.
utils/manipulate_dataset.py	File containing Python code. Functions made for ease of use in a team enviroment.
utils/plotting.py	File containing Python code. Used for getting to know the data. Made plots to find correlations between features.
csv_output	Folder containing some of the csv-files we used for our coding. Not all of our outputted files are in here,   since Github has a file limit of 100MB.
visuals	Folder containing plots we deemed interesting and helped us gain insights on the data.

Feature engineering

Column name of feature	Change made	Reason
timestamp	Only keeping rows above 0,25	We found some outliers where the "rpm" and "hz" values spiked in the first parts of the test.  With the use of plotting, we discovered a cut off point.

Column name of feature	Change made	Reason
timestamp	Only keeping rows equal to or below 1,5	We found that the biggest differences between it being a bad or good bearing, could be found in the first parts of the test. With the use of plotting, we discovered a cut off point.

Column names of feature	Changes made	Reason
a1_x  a1_y  a1_z  a2_x  a2_y  a2_z	For every "experiment_id", took the mean of every column mentioned. For every row, changed the value in every column mentioned to its mean.	The model had an easier time of fitting and was still able to make accurate predictions with these changes.

Visuals

Machine used to gather the data on bearings

Plot showing the min-max-difference of every axis, on every bearing.

Plot that gave us the idea to look into the first seconds.

Plot that showed possible clusters

Ready for future exploration

Contributors

Name	Github
Patrick Brunswyck	https://github.com/brunswyck
Jose Roldan	https://github.com/Roldan87
Matthew Samyn	https://github.com/matthew-samyn
Maarten Van den Bulcke	https://github.com/MaartenVdBulcke

Timeline

29/07/2021 - 03/08/2021