Acoustic Emission Source Localization using Approximate Discrete Wavelet Transform

Abstract

Approximate computing improves the hardware efficiency of a system by exploiting the disparity between the level of accuracy required by the application and that provided by the computing hardware. Therefore, its use has been limited to the trade-off between quality of the result and hardware cost in error-resilient applications. In this paper, we show that in addition to such a trade-off, it is possible to increase the system’s output quality, thanks to regularization that approximate processing introduces. Unlike the conventional noise injection techniques, approximate processing offers a strong correlation between the input signals and the output noise, which can be beneficial as a regulizer. We show using simulation results that the provided regularization by properly selected approximate adders in a source localization application not only improves the hardware efficiency, but also increase the regression accuracy in comparison with the exact implementation. Remarkably, these improvements are additional to a substantial decrease in the memory size as well as number of multiply-accumulate units of our proposed model in comparison to a state-of-the-art model in the literature.