Probabilistic Estimation of Instantaneous Frequencies of Chirp Signals

Zheng Zhao, Simo Särkkä, Jens Sjölund, Thomas B. Schön

TL;DR

This paper introduces a novel probabilistic framework for estimating the instantaneous frequency (IF) of chirp signals, even when their underlying forms are unknown. By modeling chirp signals and their IFs as non-linearly cascaded Gaussian Processes (GPs) within a continuous-time stochastic differential equation (SDE) framework, the authors develop an efficient filtering and smoothing approach that outperforms existing methods on synthetic and real-world data.

What Problem Does This Paper Address?

Estimating the instantaneous frequency (IF) of chirp signals is crucial in many fields, including radar, communications, and seismic analysis. Traditional methods often rely on assumptions about the signal's parametric form (e.g., linear IF) or require extensive manual tuning. This paper tackles the challenge of estimating IFs when these assumptions don't hold and explicit parametric forms are unknown. The goal is to develop a flexible, continuous-time, probabilistic approach that can handle complex chirp signals and quantify the uncertainty in the IF estimates.

Methods

The core of the proposed method lies in representing the chirp signal and its IF using non-linearly cascaded Gaussian Processes (GPs) formulated as continuous-time stochastic differential equations (SDEs). Specifically, the chirp signal's state is modeled by an SDE whose parameters are driven by another GP representing the IF. This hierarchical GP-SDE model allows for flexible modeling of complex IFs without requiring prior knowledge of their functional form.

To estimate the posterior distribution of the chirp signal and its IF, the authors employ computationally efficient stochastic filters and smoothers, specifically Gaussian filters and smoothers (GFSs). This approach allows for continuous-time inference and handles unevenly sampled measurements. The model parameters are estimated using maximum likelihood estimation combined with automatic differentiation. The paper also provides theoretical guarantees by deriving a posterior Cramér-Rao lower bound (CRLB) and a mean-square upper bound for the estimation error.

Datasets

The paper evaluates its method on:

  • Synthetic Data: Generated using the proposed chirp signal model with varying parameters and amplitude functions (constant, exponentially decaying, and random process).
  • Gravitational Wave (GW) Data: A real-world binary black hole collision event (GW150914) is used to estimate the GW signal's frequency.
  • Bat Calls: Two real-world European bat calls (Myotis myotis and Eptesicus nilssonii) are analyzed to estimate their harmonic frequencies.

The synthetic data serves as a benchmark to compare against various state-of-the-art methods, while the real-world datasets demonstrate the method's practical applicability "out-of-the-box."

Key Insights and Ideas

  • GP-SDE for Flexible IF Modeling: The use of cascaded GPs within an SDE framework provides a powerful and flexible way to model complex, non-parametric IFs, overcoming limitations of traditional parametric approaches.
  • Continuous-Time Inference: The continuous-time formulation allows for interpolation and extrapolation, making it robust to unevenly sampled data, a common issue in real-world applications.
  • Uncertainty Quantification: The probabilistic nature of the approach inherently provides uncertainty estimates for the IF, which is crucial for reliable signal analysis.
  • "Out-of-the-Box" Applicability: The combination of maximum likelihood estimation for parameter tuning and the GP-SDE model's flexibility allows the method to work effectively on real-world data without extensive manual configuration.

Key Results

  • Superior Performance on Synthetic Data: The proposed method consistently outperforms several state-of-the-art IF estimation techniques (including Hilbert transform, spectrogram, polynomial MLE, ANF, FastNLS, FHC, and existing GP-SDE models) in terms of Root Mean Squared Error (RMSE) across various scenarios, especially when dealing with random amplitude variations.
  • Accurate Real-World Signal Analysis: The method successfully estimates the IF of a gravitational wave signal and harmonic frequencies in bat calls, demonstrating its practical utility.
  • Theoretical Guarantees: The paper provides a theoretical upper bound on the estimation error and computes a posterior Cramér-Rao lower bound, offering insights into the achievable performance.
  • Computational Efficiency: The use of Gaussian filters and smoothers leads to a computational complexity that is linear in the number of measurements, making it suitable for long signals.

Limitations and Future Work

  • Computational Cost for Multiple Harmonics: While the method can handle multiple harmonics, the computational complexity increases significantly with the number of harmonics and the state dimension of the IF GP, making it challenging for very high-dimensional cases. The authors suggest using alternative filters like cubature Kalman filters (CKFS) for such scenarios.
  • Tightness of Error Bounds: The authors acknowledge that the derived error bounds might not be the tightest possible and suggest that improving the bound on the chirp residual could lead to tighter overall error bounds.
  • Future Work: The paper suggests exploring tighter posterior Cramér-Rao bounds and investigating more efficient methods for handling high-dimensional state spaces when dealing with multiple harmonic chirps.

Summary updated

Abstract

We present a continuous-time probabilistic approach for estimating the chirp signal and its instantaneous frequency function when the true forms of these functions are not accessible. Our model represents these functions by non-linearly cascaded Gaussian processes represented as non-linear stochastic differential equations. The posterior distribution of the functions is then estimated with stochastic filters and smoothers. We compute a (posterior) Cramér–Rao lower bound for the Gaussian process model, and derive a theoretical upper bound for the estimation error in the mean squared sense. The experiments show that the proposed method outperforms a number of state-of-the-art methods on a synthetic data. We also show that the method works out-of-the-box for two real-world datasets.
BibTeX
@article{zhao2023probabilistic,
  author = {Zheng Zhao and Simo Särkkä and Jens Sjölund and Thomas B. Schön},
  title = {{Probabilistic Estimation of Instantaneous Frequencies of Chirp Signals}},
  journal = {IEEE Trans. Signal Process.},
  year = {2023},
  doi = {10.1109/tsp.2023.3245720},
}
Figure