Challenge:

A European city has many, large, high-voltage, electrical power transformers.
Access inside the enclosures is severely limited to people with very special training. Traditional power measurement equipment is very expensive, so it can’t be left permanently installed, and its installation and removal are involved processes. These devices are distributed around the city, close to the neighborhoods they service, so they power company needs to “roll trucks” when they need to send someone out to examine a device.

My client develops a much less expensive hardware/software IoT solution that is permanently installed within the building’s enclosure, but is not directly connected to the transformer. It remotely measures multiple parameters of a transformer’s operation.

The challenge is to permantently monitor the transformer to determine whether it is operating well or needs service, given only the available IoT sensors.

Solution:

We did an initial Exploratory Data Analysis (EDA) to examine and get a baseline understanding of the datasets.

The operators proposed a rule which seemed to detect over-heating before a transformer shuts down and plunges its neighborhood into darkness. Our analysis showed the rule worked in some cases, but in the majority of cases, results were either not statistically significant or were opposite to expectations. The rule was abandoned.

The second approach is to build and train a machine learning model to detect anomalous transformer behavior and is currently being evaluated.

Results:

The operators were pleased to discover their originally proposed rule did not work.

Current results of the anomaly detection analysis are encouraging and on-going.