Factorially Switching Dynamic Mode Decomposition for Koopman Analysis of Time-Variant Systems Naoya Takeishi (RIKEN) Takehisa Yairi (The University of Tokyo) Yoshinobu Kawahara (Osaka University / RIKEN) 1
Koopman Operator [Koopman 31; Meziฤ 05] ๐ ๐ฎ+๐ โ โณ ๐ ๐ ๐ ๐ ๐ง ๐ ๐ข ๐ ๐ ๐ ๐ข+1 = ๐ ๐ ๐ข ๐ ๐ ๐ = ๐ง๐ ๐ ๐ โ โณ, ๐: โณ โ โณ ๐ โ โ: โณ โ โ, ๐ง: โ โ โ ๐ may be nonlinear ๐ง is a linear operator usually, dim โณ < โ in general, dim โ = โ 2
Modal Decomposition Based on Koopman Operator [ Meziฤ 05] For simplicity, suppose ๐ง has only discrete spectra (eigenvalues) ๐ง๐ ๐ ๐ = ๐ ๐ ๐ ๐ ๐ ( ๐ ๐ โ โ , ๐: โณ โ โ , and ๐ โ โ ) Assume that observable ๐ is in the span of the eigenfunctions: ๐ ๐ = เท ๐ฅ ๐ ๐ ๐ ๐ ๐ With these assumptions, because ๐ ๐ ๐ ๐ = ๐ง๐ ๐ ๐ = ๐๐ ๐ ๐ , ๐ข ๐ฅ ๐ ๐ ๐ข ๐ = เท ๐ ๐ ๐ ๐ ๐ ๐ ๐ frequency / mode decay rate 3
Dynamic Mode Decomposition [Rowley+ 09; Schmid 10] Dynamic mode decomposition (DMD) can compute the Koopman-based modal decomposition under some conditions. Let ๐: โณ โ โ ๐ or โ ๐ (vector-valued observable). Suppose we have time-series data from time ๐ข 0 to ๐ข ๐ ( ๐ข ๐ = ๐ข 0 + ๐ฮ๐ข ). ๐ ๐ ๐ข 0 , ๐ ๐ ๐ข 1 , โฆ , ๐ ๐ ๐ข ๐ , โฆ , ๐ ๐ ๐ข ๐โ1 , ๐ ๐ ๐ข ๐ ๐ = ๐ ๐ ๐ข 0 โฏ ๐ ๐ ๐ข ๐โ1 and ๐ = ๐ ๐ ๐ข 1 โฏ ๐ ๐ ๐ข ๐ DMD computes eigenvalues ๐ & eigenvectors ๐ of ๐ฉ = ๐๐ + . ๐ ๐ ๐ ๐ ๐ โ ๐ ๐ ๐ข 0 ๐ ๐ ๐ข ๐ = เท ๐ ๐ Under some conditions, these yield ๐ 4
Dynamic Mode Decomposition (contโd) [Rowley+ 09; Schmid 10] frequency / data decay rate coherent mode ๐ ๐ ๐ ๐ ๐ โ ๐ ๐ ๐ข 0 ๐ ๐ ๐ข ๐ = เท ๐ ๐ ๐ ร + = ร 5
Limitation of Standard DMDs Within the dataset at hand, system is assumed to be time-invariant, and only a single set of dynamic modes is computed for the dataset. In practice, however, โฆ This assumption may not hold. (e.g., switching systems) โฆ Even if ๐ is time-invariant, within finite-data regime, dynamic modes adequate for different periods of data may vary with time. (e.g., transient phenomena) Existing approaches: โฆ Manual separation as preprocessing โฆ Multi-resolution DMD [Kutz+ 16] 6
Core Idea: Introducing โOn -off Switchingโ to Dynamic Modes ร + = ร off ร + = off off ร โ Implement this idea via probabilistic formulation. 7
Preliminary: Probabilistic DMD [Takeishi+ 17] Dataset: ๐ = ๐ ๐ ๐ข 0 โฏ ๐ ๐ ๐ข ๐โ1 and ๐ = ๐ ๐ ๐ข 1 โฏ ๐ ๐ ๐ข ๐ = ๐ 1 โฏ ๐ ๐ = ๐ 1 โฏ ๐ ๐ ๐ ๐ Likelihood (observation model): ๐ ๐ ฯ ๐ ๐ ๐ ๐ ๐,๐ , ๐ 2 ๐ฑ ๐ ๐ ๐ โฃ ๐ ๐ = ๐๐ช ๐ ๐ ฯ ๐ ๐ ๐ ๐ ๐ ๐ ๐,๐ , ๐ 2 ๐ฑ ๐ ๐ ๐ โฃ ๐ ๐ = ๐๐ช ๐ ๐ ๐ ๐ Prior: ๐, ๐, ๐ 2 ๐ ๐ ๐,๐ = ๐๐ช ๐ ๐,๐ 0, 1 for ๐ = 1, โฆ , ๐ โ MLE in ๐ 2 โ 0 coincides with TLS-DMD 8
Proposed Model: Factorially-Switching DMD Likelihood (observation model): ๐ ๐ ฯ ๐ ๐ ๐ ๐ ๐,๐ , ๐ 2 ๐ฑ ๐ ๐ ๐ โฃ ๐ ๐ = ๐๐ช ๐ ๐ ฯ ๐ ๐ ๐ ๐ ๐ ๐ ๐,๐ , ๐ 2 ๐ฑ ๐ ๐ ๐ โฃ ๐ ๐ = ๐๐ช ๐ ๐,๐ ๐ ๐,๐ ๐ ๐ Priors: ๐ ๐ ๐ ๐ 1โ๐จ ๐,๐,๐ ๐ ๐ ๐,๐ โ ๐ ๐,๐ ๐จ ๐,๐,๐ ๐ ๐ ๐,๐ = ๐ ๐ ๐,๐ ๐ ๐ ๐ ๐ 1โ๐จ ๐,๐,๐ ๐ ๐ ๐,๐ โ ๐ ๐,๐ ๐จ ๐,๐,๐ ๐ ๐ ๐,๐ = ๐ ๐ ๐,๐ ๐, ๐, ๐ 2 ๐ ๐ ๐,๐ = ๐๐ช ๐ ๐,๐ 0, 1 โ ๐จ ๐,๐ controls on-off of ๐ -th mode at time ๐ : ๐จ ๐,๐ = 1 (on) / ๐จ ๐,๐ = 0 (off) 9
Proposed Model: Factorially-Switching DMD (contโd) Likelihood (observation model): GP ๐, ฮฃ t ๐น ๐,๐ ๐น ๐,๐ ๐ ๐ ฯ ๐ ๐ ๐ ๐ ๐,๐ , ๐ 2 ๐ฑ ๐ ๐ ๐ โฃ ๐ ๐ = ๐๐ช ๐ ๐ ฯ ๐ ๐ ๐ ๐ ๐ ๐ ๐,๐ , ๐ 2 ๐ฑ ๐ ๐ ๐ โฃ ๐ ๐ = ๐๐ช ๐ ๐,๐ ๐ ๐,๐ ๐ ๐ Priors: ๐ ๐ ๐ ๐ 1โ๐จ ๐,๐,๐ ๐ ๐ ๐,๐ โ ๐ ๐,๐ ๐จ ๐,๐,๐ ๐ ๐ ๐,๐ = ๐ ๐ ๐,๐ ๐ ๐ ๐ ๐ 1โ๐จ ๐,๐,๐ ๐ ๐ ๐,๐ โ ๐ ๐,๐ ๐จ ๐,๐,๐ ๐ ๐ ๐,๐ = ๐ ๐ ๐,๐ ๐, ๐, ๐ 2 ๐ ๐ ๐,๐ = ๐๐ช ๐ ๐,๐ 0, 1 ๐ ๐จ ฮค ๐ ๐,๐,๐ โฃ ๐ฟ ฮค ๐ ๐,๐,๐ = Bernoulli ๐ธ ๐ฟ ฮค ( ๐ธ : cdfof normal) ๐ ๐,๐,๐ โ GPmakes ๐จ smooth along time ๐ ๐น ฮค ๐ ๐,โถ,๐ = GaussianProcess ๐ ฮค ๐ ๐,๐ ๐, ๐ป ๐ข 10
Parameter Estimation by Approx. EM Input Number of modes, kernel function, GP ๐, ฮฃ t & data matrices ๐, ๐ ๐น ๐,๐ ๐น ๐,๐ 1. Initialize quantities using DMD. ๐ ๐,๐ ๐ ๐,๐ ๐ ๐ 2. E-step Approximate posterior of ๐, ๐, ๐, z, ๐ฟ ๐ ๐ ๐ ๐ using expectation propagation [Minka 01] . 3. M-step ๐ ๐ ๐ ๐ Maximize ๐ฝ โ ๐, ๐, ๐ 2 , ๐ ( ๐ฝ is wrt. distribution from E-step). ๐, ๐, ๐ 2 4. Repeat 2. and 3. until convergence. Posterior statistics of ๐, z & estimated values of ๐, ๐, ๐ 2 , ๐ Output 11
Toy Example (Local Traveling Wave) DATA RESULTS DMD FSDMD (proposed) = + bumpy because of sudden changes 12
Transient Fluid Flow DATA RESULTS time on/off states of each dynamic mode 13
Summary Objective โฆ Compute dynamic mode decomposition on time-varying systems & transient phenomena Method โฆ Idea: Introducing on-off switching GP ๐, ฮฃ ๐ข of each dynamic mode at each timestep ๐น ๐,๐ ๐น ๐,๐ โฆ Implemented it via probabilistic modeling/inference ๐ ๐,๐ ๐ ๐,๐ ๐ ๐ ๐ ๐ ๐ ๐ Future Work โฆ Developing faster & more stable inference ๐ ๐ ๐ ๐ โฆ Considering interaction between dynamic modes ๐, ๐, ๐ 2 14
Recommend
More recommend