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.
Introduction
Many real-world systems—such as aircraft, chemical reactors, or automotive engines—exhibit non-linear behavior that changes depending on external conditions like speed, temperature, or altitude. Linear Parameter-Varying (LPV) models offer a powerful framework for describing these systems by treating them as linear models whose parameters vary according to an external scheduling signal.
However, identifying these models from noisy input/output data remains a significant challenge. Traditional subspace methods often require strict "local observability" conditions and suffer from exponential computational complexity as the time horizon increases. Moreover, existing research typically focuses on either purely deterministic or purely stochastic systems. This paper introduces a robust algorithm to identify stochastic LPV models by decomposing them into independent deterministic and stochastic components, ensuring computational efficiency and statistical consistency.
Methodology
The proposed approach utilizes a "divide and conquer" strategy rooted in realization theory. The methodology consists of several key steps:
- Decomposition: The authors prove that a stationary LPV system can be formally split into a deterministic component (driven by known inputs) and a stochastic component (driven by unknown white noise).
- Correlation Analysis: For the deterministic part, correlation analysis is used to estimate "sub-Markov parameters," which serve as the LPV equivalent of impulse responses.
- Covariance Realization: The stochastic part is handled via a covariance-based approach to estimate noise propagation.
- Ho-Kalman Adaptation: The classic Ho-Kalman algorithm is adapted for LPV systems. By employing an "n-selection" technique to choose specific rows and columns of the data matrix, the algorithm maintains low computational overhead.
- Innovation Form: The final model is represented in an innovation form, facilitating the design of optimal filters and controllers.
Experimental Setup
The algorithm's performance was evaluated through a synthetic numerical case study involving a 3rd-order LPV system with one input, one output, and two scheduling signals. A dataset of 100,000 samples was generated for training, with an additional 100,000 for validation. To test robustness, the output was corrupted with white Gaussian noise, resulting in a Signal-to-Noise Ratio (SNR) of 4.7 dB. Both exogenous inputs and scheduling signals were generated using uniform distributions.
Results and Discussion
The identification algorithm demonstrated high precision and remarkable speed. A key insight is that the deterministic and stochastic components can be identified independently without compromising statistical consistency, significantly simplifying the mathematical requirements. Unlike many subspace methods, this algorithm does not require local observability at every fixed point, relying instead on global properties.
As shown in the parameter comparison above, the estimated sub-Markov parameters were highly accurate. For instance, a true value of 0.80 was estimated as 0.7957, and 0.40 as 0.3914.
The overall model performance was strong, as evidenced by the metrics in the table above, achieving a Best Fit Rate (BFR) of 93.56% and a Variance Accounted For (VAF) of 99.58%. Furthermore, the entire identification process for the large dataset took only 1.55 seconds on a standard i5 laptop, highlighting its computational efficiency.
Conclusion
This work provides a consistent and efficient framework for LPV system identification. Future research will focus on relaxing the assumption that scheduling signals are i.i.d. white noise, as many engineering applications involve slowly changing, highly correlated scheduling signals. Additionally, extending the theory to handle highly transient or unstable systems remains an open challenge.
Summary updated