Condition Based Monitoring Burning Motors

Feature Engineering | Thresholding Setting | Condition Quantification

Preamble

This one day project was an emergency response to the spate of motor burns on the BPLRT. The objective was to find an engineering-based scoring metric that could be used to filter out the motors that were at higher risk of burning so that they could be removed from the network and serviced. Since there was no time for validation, the model's effectiveness was based on backtest results.

Aside from the rapid identification and direct application of engineering insights based on research which correlated motor resistance, temperature and life to carbon brush-rotor interfaces (Shin, WG., Song, YS. & Seo, YK. Correlation analysis of brush temperature in brush-type DC motor for predicting motor life. J Mech Sci Technol 26, 2151–2154 (2012). https://doi.org/10.1007/s12206-012-0534-0), this project relied on direct feedback on visualizations of the data over time from the maintenance team to determine their preferred mode of quantifying a motor's risk of burning. The ideation process based on visualized data allowed for different ideas to be tested and the backtesting of the methods from the ideas eventually quantitatively determined the best idea.

Data Processing

EDA

Profile Burnt Motors

Detect Burning Motors (Method 1)

Detect Burning Motors (Method 2)

  1. Score scales positively with recency
  2. Score multiplier kicks in for multiple wear rates above the threshold
  3. Score scales with how far above the threshold the wear rate is (hinge loss)
  4. Variance of length of carbon brushes at each point of measurement (separate state as it detects specific failure mode)

Detect Burning Motors (Method 2 + Factor Replacement Count)

Plot Profiles Ranked by Exceedance

Plot Profiles Ranked by Score

Decay Vector Generator

Length and Wear Variance Analysis

For the detection of imbalanced wear of carbon brushes