Research Details

Introduction

>Research Topics

Methods

Compressed sensing

  • This is a technique for reconstructing high-dimensional data from a limited number of sampled data points and has been primarily studied in the field of medical imaging, such as MRI, with the goal of achieving super-resolution in the spatial domain [A]. In this study, we apply compressed sensing in the time domain, enabling the measurement of vibrations in the hundreds of hertz range using data sampled at only a few hertz.
  • As shown in Figure A, compressed sensing consists of two main components: random sampling and L1-norm minimization. The use of random sampling is essential to avoid aliasing. For example, as illustrated on the left side of the figure (x-plot), if a 10 Hz wave is sampled uniformly at 4 Hz, it becomes indistinguishable from a 2 Hz wave, since both pass through the same sampled points—making it impossible to distinguish between them.
  • Compressed sensing avoids this problem by employing random sampling. This makes it theoretically possible to measure frequency components that exceed the sampling rate. However, because the number of samples becomes small, the problem becomes underdetermined—meaning there are fewer observations than unknown variables, as shown on the right side of Figure A—leading to an indeterminate solution.
  • To overcome this, we utilize the sparsity of the solution. Sparsity refers to the property where most of the variables to be estimated are zero, with only a small number ss having nonzero values. When the number of measurements mm is much larger than the number of nonzero variables ss (i.e., m≫s), it has been mathematically proven that an exact solution can be obtained through L1-norm minimization, which minimizes the sum of the absolute values of the variables. In this study, we apply this principle to reconstruct the signal.
  • Figure B visually illustrates why a sparse solution can be identified through L1-norm minimization. When using the commonly applied Euclidean distance, or L2-norm minimization, the problem becomes one of minimizing the distance between variables. As shown on the right side of Figure B, the solution space takes on a rounded shape, making it less likely to yield a sparse solution.
  • In contrast, L1-norm minimization—defined by the sum of the absolute values of the variables—creates a solution space with sharp corners. This geometric property increases the likelihood that the solution lies on one of the coordinate axes, meaning that many of the variables will be zero. As a result, L1-norm minimization naturally promotes sparsity and enables the identification of sparse solutions.
  • Under construction

  • A

    Figure A: Overview of compressed sensing


    B

    Figure B: Solving underdetermined problems using L1 norm minimization

    [A] *D. L. Donoho and J. Tanner, in IEEE Transactions on Information Theory, vol. 56, no. 4, pp. 2002-2016, April 2010

    DIC(Digital Image Correlation)

  • Under construction

  • PIV(Particle Image Velocimetry)

  • Under construction

  • POD(Proper Orthogonal Decomposition)

  • Under construction

  • DMD(Dynamic Mode Decomposition)

  • Under construction

  • Applications

    Research findings have been implemented in the following areas: