Data-Driven State-Space Models of Acoustical Systems (vor Ort)
* Presenting author
Abstract:
Recent advances in the area of numerical linear algebra have brought forth randomized rank-revealing matrix factorizations. These computationally efficient methods can be used to apply classical system identification methods, e.g. the Eigensystem Realization Algorithm, to high-dimensional input data which can arise from impulse response or transfer function measurements of acoustical transmission systems. The size of these measurements is not uncommonly in the order of 10e+6 samples which have made many system identification methods unfeasible in the past because they rely on the singular value decomposition.The apparent benefits of state-space models such as model order reduction, parametric interpolation or robust control can now be reaped in scenarios where high-fidelity analytical models are hard to obtain. This sheds new light onto many modelling challenges. Our studies have shown that reduced data-driven state-space models can already be more efficient when it comes to forced-response computations compared to established (spectral) convolution methods.