ConvLSTM-Attention-Based Wheel Slip and Zero Velocity Detection for Inertial Navigation

As a widely used navigation technique, inertial navigation system (INS), however, suffers from the issue of cumulative errors. Based on convolutional neural network (CNN), Long-Short Term Memory (LSTM) and Multi-Head Attention, we propose the ConvLSTM-Attention model to detect wheel slip and zero velocity, which introduces the focal loss as the training criterion to address data imbalance and put more focus on hard samples. Then we use the detection values and innovation filtering to improve the ES-EKF navigation algorithm, which is based on Zero-Velocity Update (ZUPT), Non-Holonomic Constraints (NHC) and Wheel Odometry (WO) Update. In detection experiments, the proposed model has the highest detection performance. The F1score of ConvLSTM-Attention is about <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$78.54 \%, 1.44 \%, 71.92 \%$</tex>, 78.83 % and 74.98 % higher than that of ConvLSTM-Attention with the cross entropy loss, ConvLSTM, CNN, k-nearest neighbors (KNN) and Random Forest, respectively. In navigation experiments, our method brings improvements to positioning accuracy at least <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$46.24 \%, 61.57 \%$</tex> and 6.17 % in the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$E, N$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$U$</tex> directions, respectively.