Investigation of pattern recognition algorithms to determine depth and volume of water inside the sump of a pumping station

Abstract: Pattern recognition is now considered a fundamental building block in many fields.  The ability to interact with a computer or vice versa is no longer limited by how fast the computer is, but rather what an application developer can think of. Today many modern, real-time applications, such as high performance and high quality graphics can be combined with a Xbox Kinect to do object tracking and Google Glass to provide a heads up display. These applications can also be combined with other sensors and actuators to produce monitoring systems that can give facilities' operators "telepresence" throughout a facility. To be able to computationally interpret movement or patterns in an image it is imperative to investigate the application of this technology. The research conducted at Xylem has focused on a very specific problem: How can pattern recognition be utilized to dynamically determine the volume and depth of water in a sump at a pumping station. The equipment currently used to determine the water level depends upon being either below or alongside the water's surface, this puts the equipment under great stress due to the nature of a pumping station. Xylem is one of the leading global water technology companies, hence sewage-pumps are one of its main products. The main equipment utilized in this thesis project consists of a camera attached to the interior at the top of the sump in a pumping station connected to a computer. The software developed includes a simple graphical user interface (GUI). This GUI was implemented in C# and is designed to continuously collect data from a camera for subsequent analysis. Our algorithm utilizes anti-correlation between many images taken during a short interval to determine the actual water level. The known dimensions of the sump are then used to calculate the volume of water. Most of the depth values produced by our software were correct and we were able to correctly estimate the water level with an error of less than 4 cm, this corresponds to a volume error of 62 liters for a 140 cm wide sump. Our algorithm was able to monitor the depth over time inside a pumping station in a simulated live environment.  This accuracy is obtained with a time window of 1 second. The results of this system are important because it shows that it is possible to use a camera to measure water depth. This provides pump owners and operators with valuable information regarding the current state of the pump, both in terms of current water depth and an image of possible anomalies such as the presence of foreign objects in the sump.