Method for Fault Diagnosis of Temperature-Related MEMS Inertial Sensors by Combining Hilbert-Huang Transform and Deep Learning

Tong Gao, Wei Sheng, Mingliang Zhou, Bin Fang, Futing Luo, Jiajun Li

TL;DR

This paper introduces a novel fault diagnosis method for MEMS inertial sensors in UAVs, tackling the challenge of temperature-related faults. By combining a BLSTM-enhanced Hilbert-Huang Transform (HHT) for feature extraction with a multi-scale CNN for classification, the method achieves high accuracy and improved computational efficiency compared to existing approaches.

What Problem Does This Paper Address?

Unmanned Aerial Vehicles (UAVs) rely heavily on MEMS inertial sensors for navigation and control. However, these sensors are susceptible to performance degradation due to temperature fluctuations, leading to potential faults that can compromise UAV reliability. Traditional fault diagnosis methods often struggle with the non-stationary and non-linear nature of these temperature-induced faults. Existing deep learning approaches also face limitations in effectively converting 1D sensor data into 2D representations suitable for CNNs, and traditional HHT methods can be computationally expensive. This paper aims to address these challenges by proposing a more efficient and accurate fault diagnosis system for MEMS inertial sensors.

Methods

The proposed method integrates several key techniques:

  • Hilbert-Huang Transform (HHT) with BLSTM-based Empirical Mode Decomposition (EMD): The core innovation lies in enhancing the EMD process, a crucial part of HHT. Instead of traditional iterative EMD, the authors propose a BLSTM-based regression model to directly estimate the trend signal. This approach significantly reduces the number of iterations required for IMF (Intrinsic Mode Function) computation, leading to improved computational efficiency. The BLSTM is trained on piecewise polynomial fitting parameters derived from the residual signal.
  • Frequency Shifting and Noise Assistance: To further improve the EMD process and mitigate mode mixing, frequency shifting and noise assistance techniques are employed, building upon existing research.
  • Multi-Scale Convolutional Neural Network (MS-CNN): After HHT transforms the 1D inertial data into 2D time-frequency spectra (Hilbert spectra), an MS-CNN is used for fault classification. This CNN architecture is designed to capture features at multiple scales, enhancing its ability to learn complex fault patterns.

Datasets

The experiments were conducted using a MEMS inertial sensor dataset collected from UAV micro guidance and navigation controllers (UAV-MGNCs). This dataset was acquired under variable temperature conditions (-40°C to 68°C) at a sample rate of 100 Hz. The dataset includes four types of temperature-related faults:

  • Normal condition (Fault 0)
  • Fitting tendency error (Fault 1)
  • Bulge in a range of temperature (Fault 2)
  • Output hopping (Fault 3)

The dataset was split into 80% for training and 20% for testing.

Key Insights and Ideas

  • BLSTM for EMD Trend Estimation: The most significant insight is the successful application of BLSTM for directly estimating the trend signal in EMD. This bypasses the iterative nature of traditional EMD, offering a substantial computational speedup.
  • Data-Driven Feature Engineering for EMD: The paper highlights the effectiveness of using piecewise polynomial fitting parameters as features to train the BLSTM for trend estimation. This data-driven approach makes the EMD process more robust and efficient.
  • HHT's Advantage for Non-Stationary Data: The study reinforces that HHT is well-suited for analyzing non-stationary signals like those from MEMS inertial sensors affected by temperature. Combining it with deep learning further amplifies its diagnostic power.
  • Multi-Scale CNN for Robust Classification: The use of an MS-CNN demonstrates its capability in effectively classifying faults from the complex time-frequency representations generated by HHT.

Key Results

  • Improved EMD Efficiency: The proposed BLSTM-based EMD method achieved the lowest time consumption, with a mean of 7.6475 seconds per EMD operation, significantly outperforming traditional EMD and other advanced EMD methods.
  • High Fault Classification Accuracy: The proposed method achieved a mean fault classification accuracy of 97.3007%, outperforming several state-of-the-art data-driven FD methods, including ANN, SVM, and other CNN-based approaches.
  • Confusion Matrix Performance: The confusion matrix analysis showed that the proposed method achieved the best overall accuracy (97.2434%), with strong performance across all fault types, particularly for Fault 2.

Limitations and Future Work

  • Computational Cost of HHT: While the BLSTM-based EMD improves efficiency, the overall HHT process still involves frequency shifting and EEMD, which can add to the computational burden.
  • Feature Coding Strategy: The authors acknowledge that the preferable feature coding strategy for BLSTM-based EMD was not fully explored. They suggest that in the future, an improved feature coding method that integrates the state of EEMD and frequency shifting could be studied to further enhance the efficiency of the BLSTM-based regression model.
  • Dataset Annotation: The proposed dataset generation method is automatic, avoiding complex data annotation. However, future work could explore more sophisticated annotation strategies if needed.

Summary updated

Abstract

In this paper, we propose a novel method for fault diagnosis in micro-electromechanical system (MEMS) inertial sensors using a bidirectional long short-term memory (BLSTM)-based Hilbert–Huang transform (HHT) and a convolutional neural network (CNN). First, the method for fault diagnosis of inertial sensors is formulated into an HHT-based deep learning problem. Second, we present a new BLSTM-based empirical mode decomposition (EMD) method for converting one-dimensional inertial data into two-dimensional Hilbert spectra. Finally, a CNN is used to perform fault classification tasks that use time–frequency HHT spectrums as input. According to our experimental results, significantly improved performance can be achieved, on average, for the proposed BLSTM-based EMD algorithm in terms of EMD computational efficiency compared with state-of-the-art algorithms. In addition, the proposed fault diagnosis method achieves high accuracy in fault classification.
BibTeX
@inproceedings{gao2020method,
  author = {Tong Gao and Wei Sheng and Mingliang Zhou and Bin Fang and Futing Luo and Jiajun Li},
  title = {{Method for Fault Diagnosis of Temperature-Related MEMS Inertial Sensors by Combining Hilbert-Huang Transform and Deep Learning}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=H9N3nreN8N},
}
Figure