TL;DR
This paper introduces a novel adaptive denoising method for gyroscope signals, addressing the challenge of random drift errors that degrade accuracy in inertial navigation systems. By combining Empirical Mode Decomposition (EMD) with a new classification scheme for intrinsic mode functions (IMFs) based on fractal Gaussian noise and Hausdorff distance, the method effectively separates noise from useful signal components for improved reconstruction.
What Problem Does This Paper Address?
Gyroscope signals are crucial for inertial navigation systems (SINS), but they are plagued by random drift errors that accumulate over time, significantly impacting the accuracy of position, velocity, and attitude estimations. While traditional methods exist to model or directly denoise these signals, they often struggle with the complex, time-varying, and non-stationary nature of gyroscope data, which includes not only white Gaussian noise but also colored noise. Existing EMD-based denoising methods can be suboptimal, either by discarding potentially useful information in mixed IMFs or by failing to adequately process low-order noise IMFs. This paper aims to develop a more robust and adaptive denoising technique that can effectively handle these challenges.
Methods
The core of the proposed method lies in a refined approach to Empirical Mode Decomposition (EMD) and signal reconstruction.
- Empirical Mode Decomposition (EMD): The paper starts by applying EMD to decompose the noisy gyroscope signal into a set of intrinsic mode functions (IMFs) and a residual. EMD is a data-driven technique that adaptively breaks down complex signals into simpler oscillatory components ordered by frequency.
- Fractal Gaussian Noise (FGN) for Noise Characterization: Recognizing that gyroscope noise is often colored, the authors introduce Fractal Gaussian Noise (FGN) to model this characteristic. They estimate the Hurst parameter (H) of the FGN to quantify the noise properties within the IMFs.
- IMF Classification: This is the key innovation. The IMFs are adaptively classified into three categories:
- Pure Noise IMFs: Identified by comparing their actual energy to theoretical energy values derived from FGN.
- Mixed IMFs (Noise and Information): Determined using an improved Hausdorff Distance (HD) method. HD measures the similarity between the probability density functions (PDFs) of the input signal and each IMF, helping to distinguish components that contain both noise and signal.
- Information IMFs: These are the IMFs deemed to contain the primary signal information.
- Interval Thresholding: A modified interval thresholding method is applied to the "mixed IMFs" to denoise them while preserving essential signal components. This thresholding is designed to avoid discontinuities and bias introduced by simpler methods.
- Optimal Signal Reconstruction: The denoised signal is reconstructed by combining the thresholded mixed IMFs, the pure information IMFs, and the residual. Pure noise IMFs are discarded.
Datasets
The authors evaluated their method using real static and dynamic test data collected from a fiber optic gyroscope (FOG).
- Static Test Data: A six-hour output signal from a FOG with a zero offset of 0.3 °/h was used. 100,000 data points from the middle section of this signal were analyzed.
- Dynamic Test Data: Five sets of dynamic rotation data were collected at different rotational speeds (10°/s, 20°/s, 30°/s, 40°/s, and 50°/s). For each set, 100,000 stable rotation data points were selected for analysis.
The paper does not introduce new benchmark datasets but rather uses real-world FOG data to compare its method against existing ones.
Key Insights and Ideas
- Colored Noise is Crucial: The paper highlights that gyroscope noise isn't just white Gaussian noise; colored noise is a significant factor. Modeling this with FGN and estimating the Hurst parameter is key to accurately characterizing the noise.
- Adaptive IMF Classification is Essential: Simply discarding low-order IMFs or applying a blanket filter is insufficient. The proposed method's strength lies in its ability to intelligently classify IMFs into noise, mixed, and information categories, allowing for targeted processing.
- Hausdorff Distance for Similarity: Using Hausdorff Distance on PDFs is an effective way to measure the similarity between the original signal and its decomposed components, aiding in the identification of mixed IMFs.
- Improved Thresholding for Mixed Components: The interval thresholding method, adapted for FGN characteristics, offers a more nuanced approach to denoising mixed IMFs, balancing noise reduction with signal preservation.
Key Results
The proposed method demonstrated superior performance compared to traditional EMD-based denoising techniques (EMD-l2-norm and EMD-SA) in both static and dynamic tests.
- Static Test: The proposed method significantly reduced noise, as visualized in time-domain plots and amplitude spectra. Allan variance analysis also showed the best noise reduction.
- Quantitative Metrics (Static Test): The paper reports improvements in bias drift, angle random walking (ARW), quantization noise, and RMSE. Specifically, compared to the input signal (Scheme 1), the proposed method (Scheme 4) improved these metrics by 87.7%, 63.4%, 68.4%, and 85.6%, respectively.
- Dynamic Test: The method effectively reduced noise, with performance improvements observed across various rotational speeds. Table 2 shows that the proposed method achieved the minimum RMSE and standard deviation (STD) values for dynamic signals, improving RMSE by up to 4.5 times compared to the original signal.
Limitations and Future Work
- Limitations: The authors acknowledge that the FOG used in their experiments was not highly accurate, which might limit the direct generalizability of the results to high-precision gyroscopes. The reported ARW and bias drift values, while improved, are still not ideal compared to state-of-the-art models.
- Future Work: The authors plan to explore improved EMD algorithms and different thresholding techniques for mixed IMFs in their future studies to further enhance denoising performance. They also intend to investigate more advanced EMD algorithms and adaptive IMF selection strategies.
Summary updated