A MEMS IMU De-Noising Method Using Long Short Term Memory Recurrent Neural Networks (LSTM-RNN)

Changhui Jiang, Shuai Chen, Yuwei Chen, Boya Zhang, Ziyi Feng, Hui Zhou, Yuming BoSensors 2018 · DOI

TL;DR

This paper introduces a novel approach to de-noise MEMS Inertial Measurement Units (IMUs) by leveraging Long Short-Term Memory Recurrent Neural Networks (LSTM-RNNs). By treating IMU data as time-series, their LSTM-RNN method significantly reduces gyroscope noise and attitude errors, outperforming traditional ARMA models and offering a more robust solution for navigation systems, especially in challenging GPS environments.

What Problem Does This Paper Address?

Accurate positioning and navigation are crucial for many applications, from autonomous vehicles to handheld devices. While Global Navigation Satellite Systems (GNSS) like GPS provide reliable positioning, they struggle in environments with signal blockage (e.g., urban canyons, dense forests). Inertial Navigation Systems (INS), which use IMUs (gyroscopes and accelerometers), can bridge these GNSS outages. However, MEMS IMUs, despite their low cost and small size, suffer from significant non-linear and random errors that cause INS solutions to drift over time. Traditional methods to mitigate these errors, like Allan Variance or Auto-Regressive Moving Average (ARMA) models, have limitations, particularly in capturing complex, time-varying noise patterns. This paper aims to address this by developing a more effective de-noising technique for MEMS IMU data to improve INS accuracy, especially during GNSS signal outages.

Methods

The core of this paper is the application of Long Short-Term Memory Recurrent Neural Networks (LSTM-RNNs) for de-noising MEMS IMU gyroscope data. LSTM-RNNs are a type of Recurrent Neural Network (RNN) specifically designed to handle sequential data and capture long-term dependencies, making them well-suited for time-series analysis.

The authors compare their LSTM-RNN approach against a traditional statistical method:

  • Auto-Regressive Moving Average (ARMA) Model: This is a standard statistical technique for time-series analysis. The paper describes the process of determining the ARMA model order using auto-correlation (ACF) and partial auto-correlation (PACF) functions, followed by parameter estimation.
  • LSTM-RNN Method: The paper details the architecture of an LSTM unit, explaining its "forget gate," "input gate," and "output gate" mechanisms, which allow it to selectively remember or discard information over time. They explore both single-layer and multi-layer LSTM-RNN configurations. The LSTM-RNN treats the gyroscope data as a time series, where a sequence of past data points (input vector) is used to predict the next data point (output vector). The "step" parameter defines the length of this input vector.

The authors also investigate the impact of input vector length and compare single-layer versus multi-layer LSTM-RNNs in terms of training time, computation load, and de-noising performance.

Datasets

The experiments were conducted using raw gyroscope data collected from a MEMS IMU (MSI3200) manufactured by MT Microsystems Company. A 2-minute dataset was recorded at a sampling rate of 400 Hz, resulting in 48,000 data points. This dataset was divided into training and testing sets. No new benchmark datasets were introduced; the focus was on evaluating the proposed method on existing IMU data.

Key Insights and Ideas

  • LSTM-RNNs Excel at Capturing Complex IMU Noise: The paper strongly suggests that the sophisticated gating mechanisms within LSTM-RNNs allow them to better model and remove the complex, non-linear, and time-varying noise inherent in MEMS IMUs compared to traditional statistical methods like ARMA.
  • Time-Series Nature of IMU Data is Key: By treating IMU outputs as time series, LSTM-RNNs can leverage their ability to learn temporal dependencies, which is crucial for understanding and mitigating drift.
  • Trade-offs in LSTM-RNN Design: The study highlights that while multi-layer LSTMs can achieve lower training loss with fewer epochs, they don't necessarily offer superior de-noising performance (in terms of STD) compared to a well-trained single-layer LSTM. There's a clear trade-off between model complexity, training time, and prediction accuracy.
  • ARMA's Limitations in Dynamic Environments: The paper implies that ARMA's fixed parameters make it less adaptable to the changing noise characteristics that might occur in dynamic motion scenarios, whereas LSTM-RNNs, with their adaptive learning, are more promising.

Key Results

  • Significant Noise Reduction: The LSTM-RNN method achieved substantial reductions in the standard deviation (STD) of the gyroscope data:
  • Compared to raw signals: 60.3% (X-axis), 37% (Y-axis), and 44.6% (Z-axis).
  • Compared to ARMA: 42.4% (X-axis), 21.4% (Y-axis), and 21.4% (Z-axis).
  • Attitude Error Reduction: The de-noising also led to significant improvements in attitude errors:
  • Compared to raw signals: 15.8% (pitch), 18.3% (roll), and 51.3% (yaw).
  • Compared to ARMA: 47.6% (pitch), 42.3% (roll), and 52.0% (yaw).
  • ARMA Performance: The ARMA method also showed improvement over raw data, reducing STD by 31.1%, 20.0%, and 25.0% for the three axes.
  • LSTM-RNN vs. ARMA: The LSTM-RNN consistently outperformed the ARMA model in both STD reduction and attitude error reduction.
  • Input Vector Length: An input vector length of 20 was found to be the optimal trade-off between training time and prediction accuracy for the LSTM-RNN.
  • Single vs. Multi-layer LSTM: While multi-layer LSTMs showed lower training loss with fewer epochs, single-layer LSTMs trained for more epochs achieved comparable or slightly better STD values.

Limitations and Future Work

  • Computational Capacity: The authors acknowledge that the computational capacity of their computer limited the depth of the LSTM-RNN network, potentially affecting long-term prediction ability.
  • Limited RNN Variants: Only LSTM-RNN was explored. Future work could investigate other RNN variants and more complex LSTM architectures.
  • Static Data Only: The current study used static data. Evaluating the method with dynamic trajectory data is crucial, as noise characteristics can differ significantly in dynamic environments.
  • Further Investigation: The authors suggest exploring deeper LSTM-RNN networks trained on larger datasets and comparing a wider range of RNN structures for MEMS IMU error modeling and de-noising.

Summary updated

BibTeX
@article{jiang2018mems,
  author = {Changhui Jiang and Shuai Chen and Yuwei Chen and Boya Zhang and Ziyi Feng and Hui Zhou and Yuming Bo},
  title = {{A MEMS IMU De-Noising Method Using Long Short Term Memory Recurrent Neural Networks (LSTM-RNN)}},
  journal = {Sensors},
  year = {2018},
  doi = {10.3390/s18103470},
}
Figure