Low-cost MEMS gyroscope performance improvement under unknown disturbances through deep learning-based array

Abstract

The performance of low-cost Micro Electro Mechanical systems (MEMS) gyroscope severely deteriorates under unknown disturbances, such as temperature, vibration and shock. In this paper, gyroscope array based on deep learning is proposed to improve the accuracy of single gyroscope during unknown disturbances. First, the gyroscope raw data are subjected to modeling and estimation within long short-term memory (LSTM) neural network. By training the model within the LSTM network, the intricate patterns and dynamic variations present in the gyroscope data could be captured, resulting in the generation of high-quality state estimates. Furthermore, the dropout technique is introduced to finely adjust the LSTM model, with the aim of reducing overfitting risks and effectively diminishing the computational complexity of the network. Lastly, state estimates generated by the LSTM are ingeniously combined with the raw gyroscope observations, furnishing the Kalman filter (KF) with input. The simulation and experiment results demonstrate that compared with the gyroscope array based on KF and H∞, the proposed method exhibits better performance and robustness.