Dr Liang Wang1, Dr Yi-Fan Zhang2, Mr Nigel Bajema3, Dr Scott Mill3
1CSIRO, Douglas, Australia, 2CSIRO, St Lucia, Australia, 3eResearch Center, James Cook University, Douglas, Australia
There has been growing interest and application of Low Power Wide Area Networks (LPWAN) in the field of the Internet of Things (IoT). Current GPS technology provides accurate location information, however, it is limited to functions due to the high power consumption demands of the technology. LPWAN offers a potential alternative to GPS for location finding, where power consumption can be minimized by using Time Difference of Arrival (TDOA). This technique falls into the category of miltilateration, and three TDoAs and the locations of four signal receivers are sufficient for the localization each time. Thus far the accuracy of these solutions is not at an optimal level since the obstructions on the ground can introduce signification error and/or cause data missing in the localization, such as the effects of multipathing and blocking.
We proposed a machine learning-based method to reduce localization errors and impute the missing locations. Firstly, we implement the one-class support vector machine to mark the TDoAs with significant error as missing TDoAs and K Nearest Neighbors to impute the missing TDoAs if there is at least one TDoA received. Then The Radial based function neural network reduces the error of each TDoA, followed by an analytic algorithm to calculate the position by the denoised TDoAs. Lastly, the positions where none TDoA received are imputed by a Sequence to Sequence model with the attention mechanism. The outcome of the model is the denoised full path of moving objects.
Liang Wang received a bachelor of science in physics from Sichuan University, Chengdu, Sichuan, China, in 2009, and the Ph.D. degree in Astrophysics from the University of Chines Academy of Science, Beijing, China, in 2015. He is currently a Post-Doctoral Fellow with the Department of Agriculture and Food, CSIRO, Townsville, QLD, Australia. His current research interests include machine/deep learning and time series analysis.