FloodBud Week 9

Hello potential CDS intern! We are the FloodBud team, with Maddie Carlini (Colby ‘21), Kutub Gandhi (Rice ‘20), and Jade Wu (UNC Chapel Hill ‘20) and we worked as part of the Smart Sea Level Sensors (SLS) team in the City of Savannah and Chatham County. The complex geography of Chatham County causes intricate differences in flooding patterns across the region, affecting local residents in different ways. So far, a single tide gauge at Ft. Pulaski has been used to monitor the entire Georgia coastline. However, this one gauge does not provide enough granularity to capture trends in flooding across the entire region. The SLS project is working to quantify the intricacies of this tidal system. 

Over the past year, the SLS  team has deployed over 30 sensors across Chatham County, with a goal of over 100 installed by 2020. Each sensor records water level, air temperature, and barometric pressure measurements every five minutes. As the sensor network expands, there is a growing need to monitor and maintain the sensors. Thus far, researchers must manually check the datastream for each sensor to assess its health. This process can become inefficient and time consuming. We aim to devise tools that will (1) streamline the monitoring of the sensor network for researchers and (2) make the sensor data more accessible to the public. 

In order to assist with sensor maintenance, we built an anomaly detection model to flag sensors that are outputting anomalous data. To give a primer on what our data looks like, refer to the plot below. The top graph displays normal, health sensor data, with normal fluctuations in the high and low tide. The highlighted chunks in the bottom plot show examples of anomalous data that we would want our model to detect. 

The issues with the sensors fall into two main categories: issues with the sensor itself and environmental signals. Issues with the sensor involves fuzzy, incorrect, or missing data points. On the other hand, environmental signals are more complex and are caused by weather events. Our model aims to flag both kinds of anomalies. 

The general idea of our model is to take Ft. Pulaski water level readings and apply vertical and phase shifts based of off each sensors’ output to create a unique fitted function for each sensor. We then use the past three days as our testing period. We calculate the least squared difference between the testing period data and the fitted prediction function for each sensor. The testing comparison are made over one hour, one day, and three day time windows. We then flag sensors if the test errors for any of the time windows are above a set threshold. 

We came up with two main deliverables to answer our research questions. To address sensor maintenance, we created an email alert service that sends our research mentor the sensors that are currently flagged for being anomalous. The email includes the names of the seniors sorted into groups for type of anomaly. In addition, plots of the sensors for the flagged time period are attached so that a human can check for what kind of event occurred. 

Our second deliverable is a public facing website to host the data visualization and exploration tools that we created. 

The website connects to our anomaly detection model by displaying the current day’s flagged sensors. The user can also see what kind of anomaly each sensor was flagged for. 

Thank you for tuning in! Hopefully by the time you read this, our website will be up and running 🙂