TL;DR
This paper introduces and rigorously compares a class of "Generalized Method of Wavelet Functional Moments" (GMWFM) estimators for calibrating inertial sensors. They demonstrate that the simplest GMWFM estimator, the Generalized Method of Wavelet Moments (GMWM), offers superior performance in terms of bias and dispersion compared to existing methods, making it the recommended choice for practical applications.
What Problem Does This Paper Address?
Inertial Measurement Units (IMUs) are crucial for navigation and motion tracking, but their accuracy is degraded by stochastic errors. Traditionally, the "Allan variance slope method" (AVSM) has been used to calibrate these errors. However, AVSM has statistical inconsistencies and requires subjective manual intervention, making it prone to errors. Recent research has shifted towards "automatic" methods that match statistical moments of the sensor's error signal (like Allan Variance or Wavelet Variance) to theoretical models. This paper addresses the lack of a clear, theoretically grounded comparison between these automatic moment-matching techniques, aiming to identify the optimal approach for practitioners.
Methods
The paper formalizes a class of estimators called "Generalized Method of Wavelet Functional Moments" (GMWFM). These estimators work by minimizing the difference between observed and model-implied moments of the sensor's error signal, using a general function $f(\cdot)$ and a weighting matrix $\Omega$. This framework encompasses existing methods like the Generalized Method of Wavelet Moments (GMWM) (where $f(x)=x$) and the Autonomous Regression Method for Allan Variance (ARMAV) (where $f(x) = \log_{10}(x)$).
The authors provide theoretical analysis, establishing conditions for consistency and asymptotic normality of these GMWFM estimators. They then focus on the choice of the function $f(\cdot)$ and the weighting matrix $\Omega$. A key finding is that the choice $f(x)=x$ (leading to the GMWM estimator) simplifies the estimation process, often allowing for closed-form solutions and exhibiting desirable asymptotic properties, including reduced bias. They also derive conditions for asymptotic efficiency, showing that an optimally chosen weighting matrix can lead to an asymptotically efficient estimator.
Datasets
The paper uses real calibration data from the X-axis accelerometer and X-axis gyroscope of a STIM-300 IMU for simulation purposes. The parameter estimates obtained from this real data, using the GMWM approach, were treated as the "true" parameter values for the Monte Carlo simulations. No new benchmarks or datasets were introduced.
Key Insights and Ideas
- GMWFM Framework Unifies Methods: The paper elegantly unifies various moment-matching techniques under a single GMWFM framework, allowing for a systematic comparison.
- Simplicity is Key (GMWM): The simplest choice for the functional form, $f(x)=x$ (corresponding to the GMWM estimator), leads to significant practical advantages, including closed-form solutions and reduced bias.
- Bias Matters: The choice of $f(\cdot)$ directly impacts the bias of the objective function, and $f(x)=x$ is shown to minimize this bias, leading to more accurate parameter estimates.
- Asymptotic Efficiency is Achievable: While many GMWFM estimators are asymptotically normal, the paper shows how to construct asymptotically efficient estimators by carefully selecting the weighting matrix.
Key Results
- Theoretical Guarantees: The paper provides theoretical proofs for the consistency and asymptotic normality of GMWFM estimators under specific assumptions.
- GMWM Superiority: Simulation results consistently show that the GMWM estimator (using $f(x)=x$) outperforms ARMAV and AVSM in terms of Root Mean Square Error (RMSE), exhibiting lower bias and dispersion across various parameters and signal lengths.
- Parameter-Specific Performance: While GMWM is generally best, the performance differences between estimators can vary depending on the specific stochastic error parameter being estimated. For instance, AVSM tends to have larger RMSE, particularly for quantization noise.
- Sample Size Impact: Increasing the sample size generally improves the estimation performance (reduces RMSE) for all estimators, as expected.
Limitations and Future Work
- Assumptions: The theoretical results rely on several assumptions (e.g., compactness of the parameter space, continuity of functions, differentiability) which, while common in statistical estimation, might not hold in all practical scenarios.
- True Parameter Values: The simulations used estimated parameters from real data as "true" values. While this is a practical approach, it means the simulations are based on an approximation of the ground truth.
- Future Work: The paper suggests that further investigation into the "exact form of the asymptotic variance" for specific models and exploring approximate methods for complex models could be valuable. The paper also implicitly points towards the need for robust methods that can handle situations where the assumptions might be violated.
Summary updated