TL;DR
This paper introduces a deep learning approach using a Simple Recurrent Unit Recurrent Neural Network (SRU-RNN) to denoise MEMS gyroscope data, a critical step for improving navigation system accuracy. The study demonstrates that SRU-RNN effectively reduces noise and improves attitude estimation, offering a promising alternative to traditional methods.
What Problem Does This Paper Address?
Microelectromechanical System (MEMS) Inertial Measurement Units (IMUs) are widely used in navigation systems due to their small size and low cost. However, MEMS IMUs are susceptible to various noises and errors that degrade the accuracy of navigation solutions. While Global Navigation Satellite Systems (GNSS) can provide accurate positioning, they are vulnerable to signal blockages in urban canyons or dense environments. Integrating IMUs with GNSS (GNSS/INS) offers a robust solution, but the accuracy of the IMU component, particularly gyroscopes, needs improvement, especially during GNSS outages. Traditional noise modeling and suppression methods have limitations, prompting research into more advanced techniques like deep learning. This paper specifically addresses the challenge of MEMS gyroscope noise by evaluating the performance of a deep SRU-RNN for de-noising.
Methods
The core of this paper is the application of a deep Simple Recurrent Unit Recurrent Neural Network (SRU-RNN) for MEMS gyroscope de-noising. The SRU is a recently developed variant of RNN designed for faster training due to its simplified structure and parallelizable computations. The SRU-RNN architecture consists of multiple SRU units stacked to form a deep network, allowing it to learn complex temporal dependencies in the time-series data from the gyroscope. The authors investigated the influence of training data length on the SRU-RNN's performance, exploring a trade-off between computation load and prediction accuracy. They also analyzed the compensation of key MEMS gyroscope noise parameters: quantization noise, angle random walk, and bias instability.
Datasets
The experiments were conducted using raw data collected from a MEMS IMU, specifically the MSI3200 from MT Microsystems Company. The data was collected statically on a table for approximately 10 minutes at a sampling frequency of 400 Hz. No new benchmark datasets were introduced; the study utilized real-world data from a commercially available MEMS IMU.
Key Insights and Ideas
- SRU-RNN's Efficacy in De-noising: The study strongly suggests that deep SRU-RNNs are effective in denoising MEMS gyroscope signals, significantly reducing noise compared to raw data.
- Training Data Length Matters: There's a clear trade-off between training data length and de-noising performance. Insufficient data hinders the SRU-RNN's ability to learn and converge, while excessively long datasets increase computation without proportional gains.
- Parameter Compensation Varies: The degree of improvement in noise parameters (quantization noise, angle random walk, bias instability) varied across the three axes of the gyroscope. This suggests that the specific characteristics of each axis, potentially due to manufacturing variations, influence the effectiveness of the de-noising.
- Attitude Accuracy Improvement: The de-noising directly translates to improved attitude estimation accuracy, with significant reductions in pitch, roll, and yaw angle errors.
Key Results
- Sufficient Training Data Lengths: For the MSI3200 IMU, training data lengths of 500 samples for the X-axis gyroscope and 3000 samples for the Y- and Z-axis gyroscopes were found to be sufficient for reliable and stable prediction performance within 100 training epochs.
- Noise Parameter Improvements:
- X-axis: Quantization noise (0.6%), angle random walk (6.8%), bias instability (12.5%).
- Y-axis: Quantization noise (60.5%), angle random walk (17.3%), bias instability (34.1%).
- Z-axis: Quantization noise (11.3%), angle random walk (22.7%), bias instability (35.7%).
- Attitude Error Reduction: The de-noising resulted in significant reductions in attitude errors: pitch (19.2%), roll (82.1%), and yaw (69.4%).
Limitations and Future Work
- Fixed Parameters: The SRU-RNN was trained with fixed parameters (learning rate and batch size). The authors suggest that parameter optimization could further improve performance.
- Single-Layer SRU-RNN: The study used a single-layer SRU-RNN. They hypothesize that a multi-layer SRU-RNN might offer even better performance.
- Static Dataset: Experiments were limited to a static dataset due to lab equipment constraints. Future work should include dynamic and field testing to evaluate performance in more realistic scenarios.
- Future Work Suggestions:
- Further analysis of signal characteristics to understand the differences in training data length requirements across axes.
- Comparison of SRU-RNN with other RNN variants like LSTM-RNN.
- Dynamic testing and the use of field testing datasets.
- Investigating the embedding of deep learning modules directly into MEMS IMUs.
- Feasibility study of integrating SRU-RNN into a GNSS/MEMS IMU system for performance evaluation during GPS outages.
Summary updated