Best Linear Approximation of Nonlinear Continuous-Time Systems Subject to Process Noise and Operating in Feedback

Rik Pintelon, Maarten Schoukens, John Lataire

TL;DR

This paper extends the "Best Linear Approximation" (BLA) framework—a tool used to model nonlinear systems as linear ones—to handle systems that generate their own internal noise (process noise) while operating in feedback loops. The authors prove that existing nonparametric estimation methods remain valid in these complex scenarios, allowing engineers to accurately quantify nonlinear distortions even when they are "masked" by system noise.

What Problem Does This Paper Address?

In the real world, no system is perfectly linear. Engineers often use the Best Linear Approximation (BLA) to find the "closest" linear model to a nonlinear system for a specific operating point. However, classical BLA theory has two major blind spots:

  1. Measurement vs. Process Noise: Most theories assume noise only happens at the sensors (measurement noise). In reality, systems often generate their own internal "process noise" (e.g., turbulence in a pipe or thermal fluctuations in a circuit), which is often the dominant noise source.
  2. Feedback Loops: Many systems operate in closed loops (feedback). In these cases, the noise and the nonlinearities interact and circulate through the system, making it much harder to separate the "true" system behavior from the noise.

This paper bridges the gap by providing a rigorous theoretical foundation for BLA in continuous-time nonlinear systems that are both noisy "from the inside" and operating in feedback.

Methods

The authors use a combination of theoretical modeling and nonparametric estimation:

  • Volterra Series Framework: They use Volterra series (a mathematical expansion for nonlinear systems) to define the class of systems where BLA is applicable.
  • Random Phase Multisines: The system is excited using "multisines"—signals composed of many harmonically related sines with random phases. These signals are "Gaussian-like," which is a requirement for the BLA to be well-defined.
  • Nonparametric Estimation: They employ "robust" and "fast" methods to estimate the BLA and its variance. These methods involve measuring multiple periods of the system's response to different realizations of the multisine signal.
  • Multiple Experiment Procedure: A novel contribution is a procedure using different excitation power levels to distinguish between Type I nonlinearities (direct input-output distortions) and Type II nonlinearities (complex interactions between the input signal and the process noise).

Datasets

As this is a control theory and instrumentation paper, it does not use standard machine learning datasets. Instead, the authors evaluate their theory using:

  • Simulated NFIR System: A nonlinear finite impulse response (NFIR) system operating in a feedback loop, used to verify the mathematical proofs.
  • Physical Electronic Circuit: A custom-built high-gain bandpass filter circuit featuring operational amplifiers and voltage-dependent resistors (LED-LDR pairs) to introduce controlled nonlinearities and process noise.

Key Insights and Ideas

  • The BLA is Robust: The most important "aha moment" is that the BLA of a system subject to process noise is identical to the BLA of a "noise-free" version of that same system (where the noise is replaced by its expected value). This means engineers can trust their linear approximations even in noisy environments.
  • Noise as a Mask: Process noise can "hide" nonlinearities. If the noise is high enough, the system might appear linear because the nonlinear distortions are smaller than the noise variance.
  • Type I vs. Type II: The paper distinguishes between distortions that are independent of noise (Type I) and those caused by the input signal "modulating" the noise (Type II). This is crucial for troubleshooting whether a problem is in the system's core dynamics or its noise handling.
  • Power Matters: To truly see the nonlinearity, you must change the power of the reference signal. If the BLA changes when you turn up the volume, you have a Type I nonlinearity. If the BLA stays the same but the "noise" variance changes unexpectedly, you likely have a Type II interaction.

Key Results

  • Theoretical Proof: The authors successfully proved that the BLA remains the "best" linear fit in the mean-square sense, even with feedback and process noise.
  • Simulation Accuracy: In simulations, the estimated BLA and its predicted variance matched the theoretical values perfectly, even when the process noise was high.
  • Experimental Validation: In the physical circuit experiment, the authors showed that increasing process noise caused a visible shift in the system's resonance frequency—a clear sign of nonlinear interaction that their framework was able to quantify and explain.

Limitations and Future Work

  • Gaussian Assumption: The theory relies on the process noise being a stationary Gaussian process. While common, this doesn't cover systems with highly "spiky" or non-stationary noise.
  • System Constraints: The Volterra framework excludes systems with "hard" chaotic behavior, subharmonics, or bifurcations. It is best suited for "fading memory" nonlinearities (where the impact of past inputs eventually dies out).
  • Future Directions: The authors suggest that this framework could be extended to help with nonlinear model selection, using the BLA's dependence on the excitation spectrum to choose the right structure for a more complex nonlinear model.

Summary updated

Abstract

In many engineering applications, the level of nonlinear distortions in frequency response function (FRF) measurements is quantified using specially designed periodic excitation signals called random phase multisines and periodic noise. The technique is based on the concept of the best linear approximation (BLA), and it allows one to check the validity of the linear framework with a simple experiment. Although the classical BLA theory can handle measurement noise only, in most applications, the noise generated by the system-called process noise-is the dominant noise source. Therefore, there is a need to extend the existing BLA theory to the process noise case. In this article, we study in detail the impact of the process noise on the BLA of nonlinear continuous-time systems operating in a closed loop. It is shown that the existing nonparametric estimation methods for detecting and quantifying the level of nonlinear distortions in FRF measurements are still applicable in the presence of process noise. All results are also valid for discrete-time systems and systems operating in an open loop.
BibTeX
@article{pintelon2020best,
  author = {Rik Pintelon and Maarten Schoukens and John Lataire},
  title = {{Best Linear Approximation of Nonlinear Continuous-Time Systems Subject to Process Noise and Operating in Feedback}},
  journal = {IEEE Trans. Instrumentation \& Measurement},
  year = {2020},
  url = {https://openreview.net/forum?id=GIrNRGznrz},
}

Loading related content...

Figure