Performance Analysis and Comparison of Two Deep RNNs in MEMS Gyroscope Raw Signals Processing

摘要

Global Navigation Satellite System (GNSS) has been a feasible andflexible apparatus for providing Positioning, Navigation and Timing (PNT)information globally. In GNSS, the navigation satellites broadcast thesignals to the earth, and the user receives the signals for PNT informationdetermination. However, a standalone GNSS is not sufficient to construct aseamless navigation system, especially in some signal challengingenvironments. Thus, GNSS is always integrated with an Inertial NavigationSystem (INS) for providing reliable PNT information, since the INS is ableto provide moderate PNT information during short time.Micro-Electro-Mechanical IMU (MEMS IMU) is popular in the navigationcommunity, due to its low cost, smaller volume and less power consumption.However, the MEMS IMU experiences complicated noise, whichcontributes the dramatically errors divergence in navigation solutions. Forsolving the problem and improving the navigation solution accuracy, thispaper introduced Artificial Intelligence methods for addressing this issue.Specifically, deep recurrent neural networks (DRNN) gained excellentperformance in processing time series. Inspired by this, this paper firstlyemployed a deep Gated Recurrent Unit – Recurrent Neural Networks(GRU-RNN) to model the noise with the aim to improve the accuracy of thenavigation solutions. Two different MEMS gyroscopes (MSI3200,STIM300) from two different companies were employed in the experimentsfor testing the evaluating the proposed GRU-RNN, and a Long Short TermMemory Recurrent Neural Networks (LSTM-RNN) was also employed forcomparing with the GRU-RNN. Following conclusions were drawnaccording to the results: 1) the employed GRU-RNN and LSTM-RNN wereboth effective for MSI3200 and STIM300 gyroscope raw signals processing.The results showed that the standard deviation (STD) of the noise decreasedby 27.1%, 36.1% and 51.1% in MSI3200, and 37.5%, 60.0% and 57.1% inSTIM300 dataset after processed by the GRU-RNN. The correspondingthree-axis attitude errors decreased by 11.4%, 21.0% and 25.7% inMSI3200 dataset, and 60.0%, 36.8%, and 34.7% in STIM300. 2) Furtherly,GRU-RNN and LSTM-RNN obtained similar performance in bothMSI3200 and STIM300 gyroscopes de-nosing. However, they bothobtained better performance in STIM300 gyroscopes de-noising.