TL;DR
This paper introduces a novel robust Hammerstein adaptive filter (HAF) that excels in identifying nonlinear systems, even under challenging non-Gaussian noise conditions. By combining a new "kernel mean p-power error" (KMPE) cost function with efficient "random Fourier features" for nonlinear modeling, the proposed HAF-RFF-KMPE demonstrates superior accuracy and robustness compared to existing methods.
What Problem Does This Paper Address?
Hammerstein adaptive filters (HAFs) are powerful tools for modeling systems with both static nonlinearities and linear dynamics. However, their performance can significantly degrade when dealing with real-world signals that often contain impulsive disturbances or other non-Gaussian noise. Traditional cost functions like Mean Squared Error (MSE) are particularly sensitive to these outliers. While existing robust methods offer improvements, they often face challenges in balancing robustness, convergence speed, and steady-state accuracy. This paper aims to address this by developing a HAF that is inherently more resilient to non-Gaussian noise and outliers, while maintaining efficient modeling capabilities.
Methods
The core of this work lies in two key innovations:
- Kernel Mean p-Power Error (KMPE) Criterion: The authors extend the "p-power loss" concept into kernel space. This KMPE criterion is designed to be more robust to outliers than traditional error metrics. By mapping data into a reproducing kernel Hilbert space, KMPE can capture higher-order statistical information and create a more resilient cost surface. The paper shows that KMPE is a generalization of existing cost functions like mean-p-power error, squared error, and sign error, suggesting its potential for broader applicability.
- Random Fourier Features (RFF) for Nonlinear Modeling: To efficiently model the nonlinear component of the Hammerstein system, the paper employs Random Fourier Features (RFF). RFF offers a flexible and computationally efficient way to approximate nonlinear mappings, balancing accuracy with reduced complexity compared to other kernel-based methods.
These two components are integrated into a Hammerstein adaptive filter framework, leading to the proposed HAF-RFF-KMPE algorithm. The algorithm's weights are updated using a steepest descent approach based on the KMPE cost function.
Datasets
The paper does not introduce new datasets. Instead, it synthesizes data for training and evaluation using specific equations to simulate a Hammerstein system with controlled nonlinearities and noise. The noise model used is a mixture of Gaussian noise and impulsive outliers, designed to mimic challenging real-world scenarios.
Key Insights and Ideas
- KMPE as a Unified Robustness Framework: The KMPE criterion is shown to be a powerful generalization of several existing robust cost functions. This suggests that KMPE can offer a unified approach to achieving robustness in adaptive filtering.
- Synergy of KMPE and RFF: The combination of KMPE for robust cost function design and RFF for efficient nonlinear modeling proves to be highly effective. This synergy allows the HAF to achieve better performance in challenging environments.
- Theoretical Guarantees for Robustness: The paper provides theoretical analyses of the steady-state excess mean square error (EMSE), offering insights into the algorithm's convergence behavior and robustness under specific assumptions.
- Parameter Sensitivity: The study highlights the importance of carefully selecting parameters like L (RFF dimension), τ (RFF parameter), p (KMPE power parameter), and σ (Gaussian kernel parameter) for optimal performance.
Key Results
- Superior Robustness: Simulations demonstrate that the HAF-RFF-KMPE consistently achieves lower steady-state Mean Squared Error (MSE) compared to several benchmark algorithms (HAF-Polynomial-MSE, HAF-Polynomial-MCC, HAF-Spline-LMP, HAF-RFF-KRSL) across various nonlinear sub-block settings and different non-Gaussian noise environments.
- Faster Convergence: The proposed method often exhibits faster convergence speed in the initial stages of learning.
- Theoretical Validation: The simulated steady-state EMSE values closely match the theoretical predictions when the prior error is small, validating the theoretical analysis.
- Effectiveness Across Noise Types: The KMPE criterion proved more effective than MSE, MCC, and KRSL in handling non-Gaussian noise, leading to lower MSE. Similarly, RFF outperformed polynomial and spline mapping functions in capturing nonlinear structures.
Limitations and Future Work
- Parameter Tuning: While the paper investigates parameter sensitivity, optimal parameter selection can still be a challenge and requires careful tuning for specific applications.
- Computational Complexity: Although RFF offers efficiency, increasing the RFF dimension (L) can lead to higher computational complexity. The paper suggests selecting L to balance accuracy and efficiency.
- Theoretical Assumptions: The theoretical analysis relies on certain assumptions (e.g., small prior error at steady state), and deviations from these assumptions might lead to discrepancies between theoretical and actual performance, as noted in Remark 2.
The authors suggest exploring future applications of the HAF-RFF-KMPE in more challenging scenarios, such as acoustic echo cancellation.
Summary updated