Introduction
Traditional state estimation often fails in complex environments because it "lumps" various noises into a single, simplified disturbance term. This paper introduces Composite Disturbance Filtering (CDF), a framework that explicitly separates and addresses multi-source, heterogeneous, and isomeric disturbances to achieve significantly higher accuracy and robustness in systems like UAVs and high-precision navigation.
Problem Statement: Beyond "White Noise"
In the modern era of complex systems, disturbances—such as those affecting a drone in a windy corridor—are rarely simple white noise. They are characterized by three complexities:
- Multi-source: Originating from distinct physical sources (e.g., aerodynamic drag vs. sensor glitches).
- Heterogeneous: Possessing diverse mathematical properties (e.g., Gaussian, norm-bounded, or unknown dynamic signals).
- Isomeric: Entering the system through different channels (e.g., additive noise, multiplicative scaling errors, or recessive parameter uncertainties).
Standard tools like the Kalman Filter (KF) or $H_\infty$ filters typically treat these as one "lumped" disturbance. This oversimplification leads to sub-optimal performance because the filter cannot distinguish between a predictable physical force and random sensor jitter. The CDF framework bridges this gap by identifying and treating each disturbance type according to its unique characteristics.
The Composite Disturbance Filtering Framework
The core of the CDF framework is the "X-DO plus Y-Filter" architecture. Instead of a monolithic algorithm, CDF combines two specialized modules to handle different noise profiles simultaneously.
- Disturbance Observer (DO): This module focuses on "rejecting" Unknown Dynamic Signals (UDS)—disturbances with a recognizable physical or mathematical structure, such as constant biases or periodic oscillations.
- Stochastic/Robust Filter: This module focuses on "attenuating" the remaining noise, such as random Gaussian jitter or norm-bounded uncertainties.
Beyond the dual-module architecture, the framework incorporates Deeply Coupled Modeling to account for disturbance interactions, Refined Quantification using tools like mixed $H_2/H_\infty$ norms, and Separability Analysis to determine if disturbances can be theoretically distinguished based on available sensor data.
Experimental Validation and Results
The CDF framework was validated across three high-stakes application scenarios:
- Inertial Navigation (INS): In initial alignment simulations, misalignment angle errors were reduced by 85% to 92% for leveling and 54% for azimuth compared to the standard Unscented Kalman Filter (UKF).
- UAV Localization: Using Ultra-Wideband (UWB) sensors in an indoor environment with non-Gaussian noise, the "DO + Variational Bayes" version of the filter successfully separated wind drag from sensor noise, achieving the lowest localization error.
- Polarization Navigation: In vehicle tests using bio-inspired sensors for GPS-denied heading determination, the CDF approach reduced the standard deviation of heading errors by over 70% compared to robust Particle Filters.
Conclusion and Future Directions
The fundamental insight of CDF is that "separation is power." By explicitly modeling structured disturbances rather than treating them as generic noise, systems can utilize that information to refine state estimates. While powerful, the framework requires prior physical knowledge of disturbances and involves higher computational complexity.
Future research directions include "Risk-Aware" filtering to defend against malicious data injection and "Green" estimation schemes designed to optimize sensor usage and reduce power consumption without sacrificing precision.
Summary updated