Fast Subspace Identification for Large Input-Output Data

Abstract

In this work, a fast subspace identification method for estimating LTI state-space models corresponding to large input-output data is proposed. The algorithm achieves lesser RAM usage, reduced data movement between slow (RAM) and fast memory (processor cache), and introduces a novel method to estimate input (B) and feedforward (D) matrices. By design, the proposed algorithm is specially well-suited to identify multi-scale systems with both fast and slow dynamics. Identification of these systems require high-frequency data recordings over prolonged periods, leading to large input-output data sizes. For such large data sizes, the proposed algorithm outperforms the MATLAB-MOESP method in terms of memory cost, flop-count, and computation time. The effectiveness of the proposed algorithm is established by theoretical computations and various case studies.