decomposition for koopman analysis of
play

Decomposition for Koopman Analysis of Time-Variant Systems Naoya - PowerPoint PPT Presentation

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; Mezi


  1. 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

  2. Koopman Operator [Koopman 31; Meziฤ‡ 05] ๐’˜ ๐ฎ+๐Ÿ โ„’ โ„ณ ๐‘• ๐’ˆ ๐’˜ ๐’ˆ ๐’ง ๐’˜ ๐‘ข ๐‘• ๐’˜ ๐’˜ ๐‘ข+1 = ๐’ˆ ๐’˜ ๐‘ข ๐‘• ๐’ˆ ๐’˜ = ๐’ง๐‘• ๐’˜ ๐’˜ โˆˆ โ„ณ, ๐’ˆ: โ„ณ โ†’ โ„ณ ๐‘• โˆˆ โ„’: โ„ณ โ†’ โ„‚, ๐’ง: โ„’ โ†’ โ„’ ๐’ˆ may be nonlinear ๐’ง is a linear operator usually, dim โ„ณ < โˆž in general, dim โ„’ = โˆž 2

  3. Modal Decomposition Based on Koopman Operator [ Meziฤ‡ 05] For simplicity, suppose ๐’ง has only discrete spectra (eigenvalues) ๐’ง๐œ’ ๐‘˜ ๐’˜ = ๐œ‡ ๐‘˜ ๐œ’ ๐‘˜ ๐’˜ ( ๐œ‡ ๐‘˜ โˆˆ โ„‚ , ๐œ’: โ„ณ โ†’ โ„‚ , and ๐‘˜ โˆˆ โ„• ) Assume that observable ๐‘• is in the span of the eigenfunctions: ๐‘• ๐’˜ = เท ๐‘ฅ ๐‘˜ ๐œ’ ๐‘˜ ๐’˜ ๐‘˜ With these assumptions, because ๐œ’ ๐‘˜ ๐’ˆ ๐’˜ = ๐’ง๐œ’ ๐‘˜ ๐’˜ = ๐œ‡๐œ’ ๐‘˜ ๐’˜ , ๐‘ข ๐‘ฅ ๐‘• ๐’ˆ ๐‘ข ๐’˜ = เท ๐œ‡ ๐‘˜ ๐‘˜ ๐œ’ ๐‘˜ ๐’˜ ๐‘˜ frequency / mode decay rate 3

  4. Dynamic Mode Decomposition [Rowley+ 09; Schmid 10] Dynamic mode decomposition (DMD) can compute the Koopman-based modal decomposition under some conditions. Let ๐’‰: โ„ณ โ†’ โ„‚ ๐‘› or โ„ ๐‘› (vector-valued observable). Suppose we have time-series data from time ๐‘ข 0 to ๐‘ข ๐‘œ ( ๐‘ข ๐‘— = ๐‘ข 0 + ๐‘—ฮ”๐‘ข ). ๐’‰ ๐’˜ ๐‘ข 0 , ๐’‰ ๐’˜ ๐‘ข 1 , โ€ฆ , ๐’‰ ๐’˜ ๐‘ข ๐‘— , โ€ฆ , ๐’‰ ๐’˜ ๐‘ข ๐‘œโˆ’1 , ๐’‰ ๐’˜ ๐‘ข ๐‘œ ๐’€ = ๐’‰ ๐’˜ ๐‘ข 0 โ‹ฏ ๐’‰ ๐’˜ ๐‘ข ๐‘œโˆ’1 and ๐’ = ๐’‰ ๐’˜ ๐‘ข 1 โ‹ฏ ๐’‰ ๐’˜ ๐‘ข ๐‘œ DMD computes eigenvalues ๐œ‡ & eigenvectors ๐’™ of ๐‘ฉ = ๐’๐’€ + . ๐‘— ๐’™ ๐‘˜ ๐’œ ๐‘˜ โˆ— ๐’‰ ๐’˜ ๐‘ข 0 ๐’‰ ๐’˜ ๐‘ข ๐‘— = เท ๐œ‡ ๐‘˜ Under some conditions, these yield ๐‘˜ 4

  5. Dynamic Mode Decomposition (contโ€™d) [Rowley+ 09; Schmid 10] frequency / data decay rate coherent mode ๐‘— ๐’™ ๐‘˜ ๐’œ ๐‘˜ โˆ— ๐’‰ ๐’˜ ๐‘ข 0 ๐’‰ ๐’˜ ๐‘ข ๐‘— = เท ๐œ‡ ๐‘˜ ๐‘˜ ร— + = ร— 5

  6. 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, โ—ฆ This assumption may not hold. (e.g., switching systems) โ—ฆ 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: โ—ฆ Manual separation as preprocessing โ—ฆ Multi-resolution DMD [Kutz+ 16] 6

  7. Core Idea: Introducing โ€œOn -off Switchingโ€ to Dynamic Modes ร— + = ร— off ร— + = off off ร— โ†’ Implement this idea via probabilistic formulation. 7

  8. Preliminary: Probabilistic DMD [Takeishi+ 17] Dataset: ๐’€ = ๐’‰ ๐’˜ ๐‘ข 0 โ‹ฏ ๐’‰ ๐’˜ ๐‘ข ๐‘œโˆ’1 and ๐’ = ๐’‰ ๐’˜ ๐‘ข 1 โ‹ฏ ๐’‰ ๐’˜ ๐‘ข ๐‘œ = ๐’š 1 โ‹ฏ ๐’š ๐‘œ = ๐’› 1 โ‹ฏ ๐’› ๐‘œ ๐Œ ๐‘— Likelihood (observation model): ๐’š ๐‘— ฯƒ ๐‘˜ ๐’™ ๐‘˜ ๐œ’ ๐‘—,๐‘˜ , ๐œ 2 ๐‘ฑ ๐‘ž ๐’š ๐‘— โˆฃ ๐Œ ๐‘— = ๐’Ÿ๐’ช ๐’› ๐‘— ฯƒ ๐‘˜ ๐œ‡ ๐‘˜ ๐’™ ๐‘˜ ๐œ’ ๐‘—,๐‘˜ , ๐œ 2 ๐‘ฑ ๐‘ž ๐’› ๐‘— โˆฃ ๐Œ ๐‘— = ๐’Ÿ๐’ช ๐’š ๐‘— ๐’› ๐‘— Prior: ๐’™, ๐œ‡, ๐œ 2 ๐‘ž ๐œ’ ๐‘—,๐‘˜ = ๐’Ÿ๐’ช ๐œ’ ๐‘—,๐‘˜ 0, 1 for ๐‘— = 1, โ€ฆ , ๐‘œ โ†’ MLE in ๐œ 2 โ†’ 0 coincides with TLS-DMD 8

  9. Proposed Model: Factorially-Switching DMD Likelihood (observation model): ๐’š ๐‘— ฯƒ ๐‘˜ ๐’™ ๐‘˜ ๐œ“ ๐‘—,๐‘˜ , ๐œ 2 ๐‘ฑ ๐‘ž ๐’š ๐‘— โˆฃ ๐ ๐‘— = ๐’Ÿ๐’ช ๐’› ๐‘— ฯƒ ๐‘˜ ๐œ‡ ๐‘˜ ๐’™ ๐‘˜ ๐œ” ๐‘—,๐‘˜ , ๐œ 2 ๐‘ฑ ๐‘ž ๐’› ๐‘— โˆฃ ๐Ž ๐‘— = ๐’Ÿ๐’ช ๐’œ ๐,๐‘— ๐’œ ๐Ž,๐‘— ๐Œ ๐‘— Priors: ๐ ๐‘— ๐Ž ๐‘— 1โˆ’๐‘จ ๐œ“,๐‘—,๐‘˜ ๐œ€ ๐œ’ ๐‘—,๐‘˜ โˆ’ ๐œ“ ๐‘—,๐‘˜ ๐‘จ ๐œ“,๐‘—,๐‘˜ ๐‘ž ๐œ“ ๐‘—,๐‘˜ = ๐œ€ ๐œ“ ๐‘—,๐‘˜ ๐’š ๐‘— ๐’› ๐‘— 1โˆ’๐‘จ ๐œ”,๐‘—,๐‘˜ ๐œ€ ๐œ’ ๐‘—,๐‘˜ โˆ’ ๐œ” ๐‘—,๐‘˜ ๐‘จ ๐œ”,๐‘—,๐‘˜ ๐‘ž ๐œ” ๐‘—,๐‘˜ = ๐œ€ ๐œ” ๐‘—,๐‘˜ ๐’™, ๐œ‡, ๐œ 2 ๐‘ž ๐œ’ ๐‘—,๐‘˜ = ๐’Ÿ๐’ช ๐œ’ ๐‘—,๐‘˜ 0, 1 โ†’ ๐‘จ ๐‘—,๐‘˜ controls on-off of ๐‘˜ -th mode at time ๐‘— : ๐‘จ ๐‘—,๐‘˜ = 1 (on) / ๐‘จ ๐‘—,๐‘˜ = 0 (off) 9

  10. Proposed Model: Factorially-Switching DMD (contโ€™d) Likelihood (observation model): GP ๐œˆ, ฮฃ t ๐œน ๐,๐‘— ๐œน ๐Ž,๐‘— ๐’š ๐‘— ฯƒ ๐‘˜ ๐’™ ๐‘˜ ๐œ“ ๐‘—,๐‘˜ , ๐œ 2 ๐‘ฑ ๐‘ž ๐’š ๐‘— โˆฃ ๐ ๐‘— = ๐’Ÿ๐’ช ๐’› ๐‘— ฯƒ ๐‘˜ ๐œ‡ ๐‘˜ ๐’™ ๐‘˜ ๐œ” ๐‘—,๐‘˜ , ๐œ 2 ๐‘ฑ ๐‘ž ๐’› ๐‘— โˆฃ ๐Ž ๐‘— = ๐’Ÿ๐’ช ๐’œ ๐,๐‘— ๐’œ ๐Ž,๐‘— ๐Œ ๐‘— Priors: ๐ ๐‘— ๐Ž ๐‘— 1โˆ’๐‘จ ๐œ“,๐‘—,๐‘˜ ๐œ€ ๐œ’ ๐‘—,๐‘˜ โˆ’ ๐œ“ ๐‘—,๐‘˜ ๐‘จ ๐œ“,๐‘—,๐‘˜ ๐‘ž ๐œ“ ๐‘—,๐‘˜ = ๐œ€ ๐œ“ ๐‘—,๐‘˜ ๐’š ๐‘— ๐’› ๐‘— 1โˆ’๐‘จ ๐œ”,๐‘—,๐‘˜ ๐œ€ ๐œ’ ๐‘—,๐‘˜ โˆ’ ๐œ” ๐‘—,๐‘˜ ๐‘จ ๐œ”,๐‘—,๐‘˜ ๐‘ž ๐œ” ๐‘—,๐‘˜ = ๐œ€ ๐œ” ๐‘—,๐‘˜ ๐’™, ๐œ‡, ๐œ 2 ๐‘ž ๐œ’ ๐‘—,๐‘˜ = ๐’Ÿ๐’ช ๐œ’ ๐‘—,๐‘˜ 0, 1 ๐‘ž ๐‘จ ฮค ๐œ“ ๐œ”,๐‘—,๐‘˜ โˆฃ ๐›ฟ ฮค ๐œ“ ๐œ”,๐‘—,๐‘˜ = Bernoulli ๐›ธ ๐›ฟ ฮค ( ๐›ธ : cdfof normal) ๐œ“ ๐œ”,๐‘—,๐‘˜ โ†’ GPmakes ๐‘จ smooth along time ๐‘ž ๐œน ฮค ๐œ“ ๐œ”,โˆถ,๐‘˜ = GaussianProcess ๐œˆ ฮค ๐œ“ ๐œ”,๐‘˜ ๐Ÿ, ๐šป ๐‘ข 10

  11. Parameter Estimation by Approx. EM Input Number of modes, kernel function, GP ๐œˆ, ฮฃ t & data matrices ๐’€, ๐’ ๐œน ๐,๐‘— ๐œน ๐Ž,๐‘— 1. Initialize quantities using DMD. ๐’œ ๐,๐‘— ๐’œ ๐Ž,๐‘— ๐Œ ๐‘— 2. E-step Approximate posterior of ๐œ’, ๐œ“, ๐œ”, z, ๐›ฟ ๐ ๐‘— ๐Ž ๐‘— using expectation propagation [Minka 01] . 3. M-step ๐’š ๐‘— ๐’› ๐‘— Maximize ๐”ฝ โ„’ ๐’™, ๐œ‡, ๐œ 2 , ๐œˆ ( ๐”ฝ is wrt. distribution from E-step). ๐’™, ๐œ‡, ๐œ 2 4. Repeat 2. and 3. until convergence. Posterior statistics of ๐œ’, z & estimated values of ๐’™, ๐œ‡, ๐œ 2 , ๐œˆ Output 11

  12. Toy Example (Local Traveling Wave) DATA RESULTS DMD FSDMD (proposed) = + bumpy because of sudden changes 12

  13. Transient Fluid Flow DATA RESULTS time on/off states of each dynamic mode 13

  14. Summary Objective โ—ฆ Compute dynamic mode decomposition on time-varying systems & transient phenomena Method โ—ฆ Idea: Introducing on-off switching GP ๐œˆ, ฮฃ ๐‘ข of each dynamic mode at each timestep ๐œน ๐,๐‘— ๐œน ๐Ž,๐‘— โ—ฆ Implemented it via probabilistic modeling/inference ๐’œ ๐,๐‘— ๐’œ ๐Ž,๐‘— ๐Œ ๐‘— ๐ ๐‘— ๐Ž ๐‘— Future Work โ—ฆ Developing faster & more stable inference ๐’š ๐‘— ๐’› ๐‘— โ—ฆ Considering interaction between dynamic modes ๐’™, ๐œ‡, ๐œ 2 14

Recommend


More recommend