Predicting blown geyser elements with AWS SageMaker

Blown elements are the most frequent homeowner’s insurance claim.

Burst geysers and consequential water damage are a pain point for homeowners and insurance providers alike, resulting in millions of rands in claims for providers, and increased premiums for policy holders. Sensor Networks’ IoT platform has enabled insurers to significantly reduce this risk through bespoke sensors that detect and prevent this threat. Our last 18 months of geyser usage and claims data have revealed hidden costs amounting to millions of rands that are significantly more frequent than burst geysers: blown elements.

Electronic geyser components – most notably heating elements – are the most frequently claimed-for items for insurance providers, and inconvenience homeowners the most. Most geyser manufacturers only offer a year’s warranty for these parts, averaging just under a two-year lifespan. Dry heating is the number-one cause for elements blowing up or being damaged, so we designed a machine-learning algorithm to predict when an anomaly was detected for a possible overheating scenario on Amazon Web Services (AWS) SageMaker.

Sensor Networks’ predict-and-prevent AWS SageMaker framework.

The solution involves tunneling the sensor data into an AWS IoT channel, through a pipeline where various transformations are stored, and then transferred to a containerised instance of the SageMaker Jupyter Notebook app. SageMaker’s Random Cut Forest (RCF) model is an unsupervised algorithm for detecting anomalous data points within a data set that diverge from the expected pattern and manifest as unexpected spikes in time series data. After training RCF on the sensor data, the model produces an anomaly score for each data point that is thresholded against the absolute deviation from each device’s mean score to detect anomalies. Based on these thresholds, an output with the anomalies is sent into an S3 bucket, triggering a lambda event to send the anomalous device information into Sensor Networks’ IoT platform, Sensor Desk – the device-management system for analysis – before remotely preventing the geyser from overheating.

Based on these anomaly scores, we are able to detect anomalous behaviour through our sensors detecting dry heating. The figure below displays one such dry-heating scenario, and the model predicting the event five hours before the element overheated.

Sensor Networks anomaly detection through AWS SageMaker at work

Dry heating is the most frequent of all geyser faults, particularly in the Gauteng region of South Africa, where water interruptions are frequent and often unannounced.

Check out this video Amazon Web Services has made about Sensor Networks’s anomaly detection with SageMaker:

Featured image by Kavinda Herath/Stuff (