Realization and identification algorithm for stochastic LPV state-space models with exogenous inputs

Manas Mejari, Mihály Petreczky

This paper, titled "Realization and identification algorithm for stochastic LPV state-space models with exogenous inputs," presents a robust mathematical framework for learning the dynamics of complex, time-varying systems from noisy data.

TL;DR

The authors propose a new algorithm to identify stochastic Linear Parameter-Varying (LPV) models by decomposing them into independent deterministic and stochastic components. This approach is computationally efficient, avoids the "curse of dimensionality" common in subspace methods, and is mathematically proven to be statistically consistent.

What Problem Does This Paper Address?

Many real-world systems—like aircraft, chemical reactors, or automotive engines—exhibit non-linear behavior that changes depending on external conditions (e.g., speed, temperature, or altitude). Linear Parameter-Varying (LPV) models are a powerful way to describe these systems by treating them as linear models whose parameters "drift" based on an external scheduling signal.

The challenge lies in system identification: how do we take a stream of noisy input/output data and reconstruct the underlying state-space equations? Existing "subspace" methods often require the system to be "locally observable" (a strict condition) and suffer from exponential growth in complexity as the time horizon increases. Furthermore, most previous research handled either purely deterministic systems (no noise) or purely stochastic systems (no known inputs), but rarely both in a unified, consistent way.

Methods

The researchers introduce a "divide and conquer" strategy based on realization theory.

  1. Decomposition: They prove that a stationary LPV system can be formally split into two parts: a deterministic component (driven by the known input $u$) and a stochastic component (driven by unknown white noise $v$).
  2. Correlation Analysis: To identify the deterministic part, they use correlation analysis to estimate "sub-Markov parameters"—essentially the LPV version of impulse responses.
  3. Covariance Realization: For the stochastic part, they use a covariance-based approach to estimate how noise propagates through the system.
  4. Ho-Kalman Adaptation: They adapt the classic Ho-Kalman algorithm (a staple of linear system theory) to work with LPV systems. By using a technique called n-selection (carefully choosing specific rows and columns of a data matrix), they keep the computational cost low.
  5. Innovation Form: The final model is represented in an "innovation form," which is particularly useful for designing optimal filters (like Kalman filters) and controllers.

Datasets

The paper evaluates the algorithm using a synthetic numerical case study.

  • System: A 3rd-order LPV system with one input, one output, and two scheduling signals.
  • Data Generation: 100,000 samples were generated for training and another 100,000 for validation.
  • Noise: The output was corrupted with white Gaussian noise, resulting in a Signal-to-Noise Ratio (SNR) of 4.7 dB (quite noisy).
  • Inputs: Both the exogenous input and the scheduling signals were generated using uniform distributions.

Key Insights and Ideas

  • Independence of Components: The most significant insight is that the deterministic and stochastic parts can be identified independently without losing statistical consistency. This simplifies the math significantly.
  • Avoiding Local Observability: Unlike many subspace methods, this algorithm does not require the system to be observable at every fixed point of the scheduling signal. It only requires "global" properties to hold.
  • Efficiency: By focusing on sub-Markov parameters rather than large data projections, the algorithm avoids the massive memory and CPU requirements that plague other LPV identification methods.

Key Results

The algorithm demonstrated high precision and speed in the numerical simulation:

  • Best Fit Rate (BFR): 93.56% on noise-free validation data.
  • Variance Accounted For (VAF): 99.58%, indicating the model captured nearly all the system's dynamics.
  • Parameter Accuracy: The estimated sub-Markov parameters were remarkably close to the ground truth (e.g., an actual value of 0.80 was estimated as 0.7957).
  • Speed: The entire identification process for 100,000 data points took only 1.55 seconds on a standard i5 laptop.

Limitations and Future Work

  • Scheduling Signal Constraints: The current proof of consistency assumes the scheduling signal is "white noise" (i.i.d.). In many engineering applications, scheduling signals (like temperature) change slowly and are highly correlated over time. Extending the proof to these cases is a key next step.
  • Prior Knowledge: While the algorithm avoids the curse of dimensionality, it may require some "n-selection" tuning or prior knowledge about the system's structure to choose the best rows/columns for the Hankel matrix.
  • Stationarity: The theory relies on the system being in a "stationary" regime. Handling highly transient or unstable LPV systems remains a challenge.

Summary updated

Abstract

In this paper, we present a realization and an identification algorithm for stochastic Linear Parameter-Varying State-Space Affine (LPV-SSA) representations. The proposed realization algorithm combines the deterministic LPV input output to LPV state-space realization scheme based on correlation analysis with a stochastic covariance realization algorithm. Based on this realization algorithm, a computationally efficient and statistically consistent identification algorithm is proposed to estimate the LPV model matrices, which are computed from the empirical covariance matrices of outputs, inputs and scheduling signal observations. The effectiveness of the proposed algorithm is shown via a numerical case study.
BibTeX
@inproceedings{mejari2019realization,
  author = {Manas Mejari and Mihály Petreczky},
  title = {{Realization and identification algorithm for stochastic LPV state-space models with exogenous inputs}},
  booktitle = {IFAC-PapersOnLine},
  year = {2019},
  doi = {10.1016/j.ifacol.2019.12.340},
}

Loading related content...

Figure