深度学习下MEMS陀螺温度误差补偿方法

Temperature Error Compensation Method of MEMS Gyroscope Under Deep Learning

针对MEMS陀螺仪因材质特性,制造工艺等差异导致输出数据受温度影响的问题。本文在传统温度误差补偿的基础上,提出将深度学习与神经网络相结合,通过LSTM神经网络进行温度误差补偿,从而减小温度变化引起的陀螺的温度漂移。分析了MEMS陀螺仪的温度特性,并在RNN神经网络模型的基础上,建立多层LSTM神经网络模型,利用基于ADAM的优化算法和时间反向传播BPTT算法对LSTM网络进行训练。将训练好的网络模型植入到STM32硬件中,进而实现对MEMS陀螺仪输出的实时温度补偿。实验表明,LSTM模型与RBF温度补偿模型相比,陀螺仪补偿后的零偏稳定性、零偏不稳定性和角度随机游走等性能指标,以及MAE、MSE、RMSE三个模型评价指标提高了90%以上。

For MEMS gyroscopes, the output data is affected by temperature due to differences in material characteristics and manufacturing processes. On the basis of traditional temperature error compensation, this paper proposes to combine deep learning with neural network, and perform temperature error compensation through LSTM neural network, thereby reducing the temperature drift of the gyro caused by temperature changes. The temperature characteristics of the MEMS gyroscope are analyzed, and on the basis of the RNN neural network model, a multilayer LSTM neural network model is established, and the LSTM network is trained using the ADAM-based optimization algorithm and the time back propagation BPTT algorithm. The trained network model is implanted into the STM32 hardware to realize real-time temperature compensation for the output of the MEMS gyroscope. Experiments show that, compared with the RBF temperature compensation model, the LSTM model has improved performance indicators such as bias stability, bias instability, and angle random walk after gyroscope compensation by more than 90%.