Predicting Accelerometer Baseline Correction and Nondivergent Deformation Velocity Based on Convolutional Neural Network (CNN) During GNSS Downgrade

Abstract

Accelerometer and global navigation satellite system (GNSS) can be effectively combined to establish a robust multisensor deformation monitoring system. However, GNSS signals may get downgraded in challenging environments, and then destroy the Kalman filter data fusion model. As a result, the accelerometer becomes the only reliable sensor for deformation monitoring, but relying on only accelerometer data may lead to rapid error accumulation due to its potential baseline shift error. To mitigate this challenge, especially in the slow-moving deformation scenarios, we propose a baseline correction prediction algorithm named CNN-based baseline correction (CNN-BC), based on convolutional neural networks. This algorithm utilizes high-frequency acceleration and baseline correction as input and output features, respectively. The baseline correction of the training dataset is derived from the accelerometer and GNSS coupled algorithm. By incorporating the reliable prediction from the network, we can correct the original accelerometer data and reduce error accumulation. To further address the divergence in deformation velocity, we develop a convolutional neural network (CNN)-dVel, which uses high-frequency acceleration and velocity difference as input and output features, respectively. We validated the proposed algorithms through two slow deformation experiments utilizing both high-precision and low-cost accelerometers. The results demonstrate that the CNN-BC can predict reliable baseline correction, with an average root mean square (rms) of $\boldsymbol {\textbf {0}.\textbf {37}~\textbf {cm}/\textbf {s}^{\textbf {2}}}$ , and the CNN-dVel achieves nondivergent deformation velocity prediction, with an average rms of 0.42 cm/s. Furthermore, optimizing the training dataset with an acceleration standard deviation (STD) basis enhances prediction accuracy.