This project focused on optimizing a condition monitoring system to track the health of electric motors and their driven loads. The objective was to develop a data selection algorithm that reduces a large dataset (500,000 data points) to a manageable subset of 2,500 representative data points, enabling efficient training of a predictive model. This model uses frequency, power, and vibration levels to monitor motor health under real-world resource constraints.
- Implementation of grid-based and density-based sampling methods.
- Flexible interpolation using a user-controlled parameter, alpha, for blending grid and density sampling.
- Visualizations, including scatter plots and histograms, to compare sampling methods.
- Normalized visualizations for fair comparison across datasets.
Condition monitoring is critical in industries such as manufacturing, oil and gas, and energy, where unexpected motor failures can result in costly downtime and safety concerns. By applying machine learning to historical data, we aim to detect early warning signs of motor health deterioration and enhance productivity and safety.
Designed for uniform data distribution across the operational range:
- Data is divided into bins (max 50 bins), with up to 50 data points per bin.
- Ensures balanced training datasets by accounting for underrepresented regions.
- Challenges: Determining the optimal number of bins and points per bin while minimizing computational time.
- Example: Visualized in histograms, this method compensates for low-value clusters, producing a balanced distribution.
Prioritizes high-density regions where the motor frequently operates:
- Assigns higher selection probabilities to clustered data points.
- Improves model accuracy by focusing on critical operational areas.
- Still includes low-density regions for comprehensive coverage and removes outliers to avoid skewed results.
- Example: Histogram comparisons highlight its similarity to the original dataset while focusing on dense areas.
Allows users to blend between grid-based and density-based sampling using the alpha parameter (range: 0 to 1):
- Linear Interpolation: Connects data points with straight lines, but struggles with non-linear patterns.
- Radial Basis Function (RBF) Interpolation: Provides smooth, non-linear approximations, better capturing complex data trends.
- Users can customize their approach based on the desired distribution balance.
The visualization module generates scatter plots and histograms for comparison:
-
Scatter Plots:
- Grid-Based Data (blue): Uniform spread.
- Density-Based Data (green): Clusters in high-density areas.
- Interpolated Data (red): A blend of grid and density distributions.
- Original Data (purple): Reference for comparison.
-
Histograms:
- Show frequency density of the
X1_tvariable across datasets. - Highlight differences in sampling methods and their effects on data distribution.
- Show frequency density of the
Normalization ensures fair comparisons by bounding data ranges based on frequency density.
This project demonstrates how targeted data selection algorithms improve predictive accuracy in condition monitoring systems, even under tight computational constraints. We extend our gratitude to Baker Hughes for providing the data and inspiration for this challenge.
The challenge aimed to optimize a condition monitoring system for electric motors, selecting 2,500 representative data points from 500,000 while maintaining predictive accuracy. Using grid-based and density-based sampling, we enabled users to balance uniform and cluster-focused sampling. Visual comparisons (scatter plots, histograms) illustrate how each method retains operational insights while ensuring efficient training. Our approach significantly enhances real-world applicability for machine learning in motor health monitoring.
- Python for data processing and visualization
- Matplotlib for scatter plots and histograms
- Scikit-learn for density clustering
- NumPy for data manipulation
- Radial Basis Function Interpolation for smooth data blending