Edge analytics

Sensor networks require computation on the edge and communication both among edge devices and with a central hub to achieve various mission-critical goals. In optimizing the efficiency of communication between these devices and maximizing the capabilities of the network as a whole, several considerations deserve particular emphasis:

• The separation of the central computing hub from the data-collecting edge devices, with constrained communications limiting what and how much information can be sent to the central hub
• Limited computational power on the edge
• How the geometry of a collection of edge devices can improve individual device efficiency

Our Solutions

Our topological signal-processing toolkit operates at the edge to extract parsimonious but information-rich summaries of massive amounts of mission sensor data, allowing the transmission of key findings from edge to a central hub. These summaries allow powerful hub machine learning to take place while respecting the limited communication allowances imposed by satellite data transmission.

We also apply shape analytics to transform positional broadcasts from individual sensors into a holistic picture of sensor field capabilities (e.g. coverage of certain critical observation areas). These sensor field measurements lead to recommendations to specific sensors such as changing a sampling rate, which permits massive battery preservation while ensuring that at least a subset of the sensors are able to detect nearby objects. These energy saving principles are key in the development of edge computing resources for any Internet of Things application in a resource-constrained setting.

Examples

In our study of IoT in the ocean, we considered the standard problems of edge analytics with the additional challenge of being unable to control the motion of the devices.

In our recent work with algorithms on the edge, GDA has developed a graph spectral representation of time series data that is parsimonious, unsupervised, performant in data-constrained scenarios, and computationally efficient. This representation, which we call Laplacian Events Signal Segmentation (LESS), can be computed on time series of arbitrary dimension and time series originating from sensors of arbitrary type, and can also be combined with our fusion algorithms.