The method which we call dimensional function synthesis, can be used to augment existing machine learning techniques applied to sensor data, enabling many orders of magnitude performance improvement over the existing traditional machine learning techniques that ignore information about the units associated with data.

Dr Stanley-Marbell

The paper ‘Deriving Equations from Sensor Data Using Dimensional Function Synthesis’ presents the method dimensional function synthesis, developed with the objective of creating inference models that fit within the memory, computation, and energy constraints of low-power embedded systems, such as those powered by scavenged energy. It applies to data streams where the dimensions of the signals are known. Details are reported in the journal ACM Transactions on Embedded Computing Systems.

The Award was presented to Dr Vasileios Tsoutsouras, Research Associate in Embedded Multi-Sensor Computing Systems, on behalf of co-authors Sam Willis, Youchao Wang and Dr Phillip Stanley-Marbell.

The work is part of a larger effort within the Physical Computation Laboratory, based in the Department’s Electrical Engineering division, to exploit information about the physical world to make more efficient computing systems that interact with nature.

Large volumes of data can be generated by physical systems instrumented with sensors. The data is useful when it comes to understanding previous behaviours of the systems that generate them (e.g. monitoring properties of components in aircraft) as well as predicting future behaviours of those systems (e.g. predicting failures of components in machinery). Dimensional function synthesis generates the family of functions from which to learn a model, based on information about the physical dimensions of the signals in the system. It exploits information on the physics of signals that has until now been ignored.

Dr Stanley-Marbell, Head of the Physical Computation Laboratory, said: “The method which we call dimensional function synthesis, can be used to augment existing machine learning techniques applied to sensor data, enabling many orders of magnitude performance improvement over the existing traditional machine learning techniques that ignore information about the units associated with data.”

Reference: 
Wang, Y., Willis, S., Tsoutsouras, V., and Stanley-Marbell, P. ‘Deriving Equations from Sensor Data Using Dimensional Function Synthesis’. ACM Transactions on Embedded Computing Systems (TECS) (2019). DOI: 10.1145/3358218

View the open access version.