Introduction The Baker Hughes Challenge presents a common issue in modern-day algorithmic modeling: creating accurate data representations given limited computational power. With a mission of improving energy in an efficient and safe manner, having accurate models that better predict the lifespan of products is critical for maintenance and repair procedures. The following documentation outlines the process of preparing model training for an electric motor.

Background The dataset provided gives 500,000 points that each have a corresponding frequency, power, and vibration level. The variable-frequency drive determines a Hertz value which corresponds to the input rate of electricity. The power is a measurement of watts of the engine’s output and the vibration level is the value to be forecasted by the model. Both of these units are normalized to one another so that they can be graphed effectively. The Baker Hughes Challenge requires downsampling, reducing the dataset to 2,500 data points that model the data in three ways:

1. Density-Based Selection - Identify 200 different clusters and find the best match within each cluster to be selected among the 2,500 data points. The size of the clusters should vary depending on the density of the frequency-power pair. Areas with high data density should be represented with smaller cluster sizes while areas with lower data density should be represented with large clusters. In a broader context, the overall distribution of the data should match that of the original dataset only but with 0.5% of the size.

2. Uniform Selection - The uniform distribution selection model should produce a similar output as the density-based selection in terms of the range of values covered, but should have a more even distribution. The cluster size across different densities should remain relatively constant to provide a different data distribution.

3. Hybrid Selection - The hybrid selection provides an in-between solution set of 2,500 data points. Different parameters of the hybrid selection algorithm should outline a distribution that is affected by data density, but to a lesser extent than the density-based selection. The purpose of creating three different datasets is to diversify the output of the prediction models. While in testing, the model is trained on multiple datasets and monitored for accuracy. Because of the unique performance of specific products, in real production environments, certain models may be a more accurate fit in different stages of the product’s life. By providing multiple datasets, the likelihood of creating an accurate prediction model is substantially increased.

More information can be found in the complete documentation linked here: https://drive.google.com/file/d/1JMLWMsQDiDozFRWFXhn752RjBwWNU_9C/view?usp=sharing

The Google Drive folder containing the CSV files, data visualizations, and project presentation can be found here: https://drive.google.com/drive/folders/1I0dAvUnOGW33BLSLVn-N_0BPp0K12vEa?usp=drive_link