quintenmm/challenge-classification

Bearing classification challenge

Mission

Our mission is to make an automated bearing testing system and to create a model in order to execute in a scheduled maintenance system. The goal is to find the more convenient algorithm to make some predictions about this system.

Content

• For the experiments device shown on the picture was constructed. Two bearings were installed on the shaft. The rotation speed changed from 0 to 1500 rpm, was held for 10 seconds, and decreased to 250 rpm.
• The shaft was rotated using an DC motor connected to the shaft through a coupling. A radial load of 3.5 kg is applied to the shaft using a balanced weight. The bearings were mounted on the shaft as shown in Figure 1.
• GY-61 ADXL3353 accelerometers were mounted on the bearing housing
• The sensor location is also shown in Figure.
• The recording was saved along the x, y, z axes.

Datasets:

• bearing_signals.csv. Contains signals recordings.
• bearing_classes.csv. Classes whole or defective for every bearing.

Credits

• Anne Jungers (@Annejungers)
• Hoang Minh (@Minh6019)
• Minh Hien (@minhhienvo368)
• Quinten (@QuintenMM)

Method

Below are provided the steps that were followed for this project. Each step and classifiers have their own document.

1. Data visualization: ploting data to detect missing values, outliers, data relations and usefulness of features

We detected the outliers of data in Hertz[Hz] column, the majority data in this column distributes from 0 to 25.5

3. Preprocessing: with the knowledge acquired with the preceding step, apply preprocessing of data including dealing with missing values, drop unuseful features and build new features

   • Option 1:

     • Feature selection: 5 new representative features (i.e. min, max, median, std, mean) derived from the orginal features (timestamp, a1_x, a2_x, a1_y, a2_y, a1_z, a2_z, hz, w). We have 45 features.
     • Target: status of bearings

   • Option 2:

     • Feature selection: 12 representative features (i.e. min, max, median, std, entropy, impulse factor, margin factor, frequency center, mean_square_frequency, root_mean_square_frequency, root_variance_frequency,crest_factor) derived from the orginal features (a1_x, a2_x, a1_y, a2_y, a1_z, a2_z, hz: range of (24-25.5))
     • Target: status of bearings {1: good, 0: bad}
4. Classifier: build classifiers based on the preprocessed data using a variety of techniques

Classification techniques with the relative scores

• Option 1: 45 features

  Classifier Test	Parameter	F1-score	CV_ROC_AUC_score
  KNN	k=5	0.95	0.94
  KNN with validation	k=5	0.95	0.94
  KNN with GridSearchCV	k=1	0.95	0.94
  Random Forest with GridSearchCV	k=100,200	0.95	1.0

  

  Conclusion: With the model Random Forest, GridSearchCV gives the highest score.

• Option 2: 84 features

  Classifier Test	Parameter	F1-Score	CV_ROC_AUC_score
  KNN	k=5	0.91	0.89
  KNN with validation	k=5	0.91	0.84
  KNN with GridSearchCV	k=4	0.91	0.87
  Random Forest with GridSearchCV	K=100,200	0.87	0.95

  

Conclusion: With the model Random Forest, GridSearchCV gives the highest score.

Folder structures

• Contains all of the jupyter's notebooks including classifiers, preprocessing and data visualization

  File	Description
  plot folder	Contains plots' images
  1.challenge-classification_01.ipynb	Python code written in "Jupyter Notebook"  Code used to get the data ready for Machine Learning.
  2.challenge-classification_02.ipynb	Python code written in "Jupyter Notebook" More in depth version.
  3.README.md	Information on the assignment

Installation instructions

1. Install Python and clone this repository
2. Install required Python modules with pip install -r requirements.txt to run the jupyter's notebooks just go with jupyter notebook

rainbow-song