Anomaly Detection

Introduction

For farmers across Australia, drought is a very difficult problem to deal with. Should their livestock not get access to adequate water, they will sustain organ damage within a day, and will die soon after. Proper water monitoring is essential to farmers’ profitability, and to keeping livestock healthy.


The Problem

Farmbot, a remote monitoring solutions and data company, wanted to be able to proactively detect anomalous behavior in their customer water tanks as well as faults in their sensor systems. Previously, Farmbot’s team had to manually scan through hundreds of time-series each day to screen for anomalies. This process, while necessary to ensure product quality, was extremely time consuming, non-scalable and costly to the business.


Our Approach

Through working closely with Farmbot to understand the intricacies of the problem, a machine learning system was developed and productionised to flag anomalous behavior of the water levels as well as the sensors. The performance of the system was evaluated, and it was proven to be sensitive to a wide range of anomalies. This was due to the use of several time-series unsupervised learning models that were combined together to generate an anomaly risk score, better capturing the true risk of anomalies in the data. The anomaly detection system was then deployed on Amazon SageMaker to deliver a robust and scalable serving layer and the model was operationalized through AWS Lambda and an API Gateway.

Results

By creating a scalable anomaly detection system by which water level sensors can be continuously monitored, Farmbot’s customers and product both greatly benefited. This will result in less time (and money) monitoring the hundreds of streams of data that Farmbot manages while at the same time ensuring more consistent product quality, leading to better management of Australian livestock farms.