基于神经网络和Kalman滤波的陀螺阵列融合算法

Fusion algorithm of gyroscope array based on neural network and Kalman filter

针对微机械陀螺测量精度和可靠性低的问题,提出了基于神经网络和Kalman滤波的陀螺阵列融合算法。将神经网络与Kalman滤波相结合,利用LSTM-RNN计算每个陀螺的置信度;再将每个陀螺的置信度、测量值及Kalman滤波估计的角速度输入给BP神经网络进行数据融合,使得BP网络在训练时拥有更多关于陀螺的特征信息,从而提高角速度的融合精度。由于率先得到了每个陀螺的置信度,使得BP网络可以更容易辨识出故障陀螺,从而减小对故障陀螺测量数据的利用率。经实际系统验证,在陀螺存在故障的情况下,所提算法的陀螺阵列MAE和RMSE相比Kalman滤波分别降低了80.25%、81.39%,相比只有测量值输入的LSTM-RNN融合算法分别降低了60.33%、63.41%,具有较强的容错性和鲁棒性。

Aiming at the low precision and reliability of micromechanical gyroscopes,a fusion algorithm of gyroscope array based on neural network and Kalman filter is proposed.By combining the neural network with Kalman filter,LSTM-RNN is used to calculate the confidence degree of each gyroscope.The confidence degree,measured value and angular velocity estimated by Kalman filter of each gyroscope are input to BP neural network for data fusion,so that BP network has more characteristic information about gyroscopes during training,so as to improve the angular velocity fusion accuracy.Since the confidence degree of each gyroscope is obtained first,BP network can identify the fault gyroscope more easily,thus reducing the utilization rate of the fault gyroscope measurement data.The actual system verification shows that in the case of gyroscope fault,the MAE and RMSE of gyroscope array of the proposed algorithm are reduced by 80.25%and 81.39%respectively compared with Kalman filter,and reduced by 60.33%and 63.41%respectively compared with LSTM-RNN fusion algorithm with only measurement input,which has strong fault tolerance and robustness.