TY - JOUR
T1 - Fast and High-Resolution Acoustic Beamforming
T2 - A Convolution Accelerated Deconvolution Implementation
AU - Chu, Ning
AU - Zhao, Han
AU - Yu, Liang
AU - Huang, Qian
AU - Ning, Yue
N1 - Publisher Copyright:
© 1963-2012 IEEE.
PY - 2021
Y1 - 2021
N2 - It is valuable to improve the resolution and speed of acoustic beamforming in real-time measurement, especially fault location, noise mapping, and so on. However, multiplication of the power transfer matrix takes much time in the iteration of the deconvolution algorithm. This article proposes a novel method based on the convolution approximation and GPU platform to obtain acoustic imaging with a high resolution quickly. First, the power propagation matrix is approximated by a symmetric Toeplitz block Toeplitz (STBT) matrix, which can transform the product of matrix and vector into the convolution of two smaller patterns. Three different convolution kernels are derived for the convolution operation, including 2-D variant convolution kernel, 2-D invariant convolution kernel, and 1-D separable invariant convolution kernel. Then, the regularization deconvolution algorithm with the convolution kernel is derived. Besides, the relation between approximation error, time cost, and setting parameters is discussed. The results show that the separable invariant convolution kernel can obtain the fastest imaging and relatively accurate localization, while the time consumption of the variant kernel is the highest. Simulations and experiments are operated on the CPU and GPU to make a comparison, which validates that our proposed deconvolution-GPU implementation can significantly improve the computation speed for the matrix on a large scale.
AB - It is valuable to improve the resolution and speed of acoustic beamforming in real-time measurement, especially fault location, noise mapping, and so on. However, multiplication of the power transfer matrix takes much time in the iteration of the deconvolution algorithm. This article proposes a novel method based on the convolution approximation and GPU platform to obtain acoustic imaging with a high resolution quickly. First, the power propagation matrix is approximated by a symmetric Toeplitz block Toeplitz (STBT) matrix, which can transform the product of matrix and vector into the convolution of two smaller patterns. Three different convolution kernels are derived for the convolution operation, including 2-D variant convolution kernel, 2-D invariant convolution kernel, and 1-D separable invariant convolution kernel. Then, the regularization deconvolution algorithm with the convolution kernel is derived. Besides, the relation between approximation error, time cost, and setting parameters is discussed. The results show that the separable invariant convolution kernel can obtain the fastest imaging and relatively accurate localization, while the time consumption of the variant kernel is the highest. Simulations and experiments are operated on the CPU and GPU to make a comparison, which validates that our proposed deconvolution-GPU implementation can significantly improve the computation speed for the matrix on a large scale.
KW - Acoustic imaging
KW - convolution
KW - deconvolution
KW - GPU acceleration
KW - real-time measurement
KW - symmetric Toeplitz block Toeplitz (STBT) matrix
UR - http://www.scopus.com/inward/record.url?scp=85097927075&partnerID=8YFLogxK
U2 - 10.1109/TIM.2020.3043869
DO - 10.1109/TIM.2020.3043869
M3 - 文章
AN - SCOPUS:85097927075
SN - 0018-9456
VL - 70
JO - IEEE Transactions on Instrumentation and Measurement
JF - IEEE Transactions on Instrumentation and Measurement
M1 - 9290103
ER -