Boltryk, Peter J., Hill, Martyn and White, Paul R.
Improving the resolution of peak estimation on a sparsely sampled surface with high variance using Gaussian processes and radial basis functions
Measurement Science and Technology, 16, (4), . (doi:10.1088/0957-0233/16/4/007).
Full text not available from this repository.
A correlation velocity log (CVL) estimates the velocity of a marine vehicle using a sonar array. The resolution of the velocity estimate provided by the device is dependent upon the ability of the device to estimate the position of a peak on a surface of calculated data points. Interpolation techniques are therefore employed to improve the resolution of the peak estimate.
The task of peak estimation is challenging because the surface is inherently asymmetric, exhibits a significant variance within a short distance from the peak location and is sparsely sampled. Previous work has concentrated on fitting a quadratic model to a selection of the data points using either a least-squares (LS) approach or an iterative maximum likelihood estimation (MLE) algorithm.
Both LS and MLE methods have proved to be reliable in both numerical simulations and when applied to data from sea trials of a newly developed CVL system, particularly when peak locations fall within the central region of the measurement area. However, the numerical simulations suggest a significant reduction in the ability of both LS and MLE to reliably estimate peak positions located near to the edge of the measurement area.
In the present study radial basis functions (RBF) and Gaussian processes (GP) are used to estimate the location of the peak position using networks that have been trained offline using example datasets. Both RBF and GP techniques are shown to achieve impressive performance throughout the measurement area, including the edges of the measurement area where LS and MLE tend to fail.
Actions (login required)