Research In Measurement Of Mems Gyroscope Random Drift Compensation

Compared to other kinds of gyroscopes, MEMS (Micro Electromechanical System) gyroscopes have many advantages such as low cost, low power, small volume, light weight and can be mass-produced. Thus they have been applied into civil market including vehicle navigations system, airbag tripping devices, camera anti-shake platforms, attitude measurement system of robots, electronic toys, virtual somatosensory games etc. In military field, It can be predicted that the weapon systems and UAVs will be developed more digitalized, Intellectualized, miniaturized and higher motorized, so MEMS gyroscopes have great potential developing values.But the measurement precision of MEMS gyroscopes is much lower than others, which has become the bottleneck of micro navigation system, so the low precision is the key problem to be solved. There are two ways to solve this problem, the first one is to analyze the internal structure of hardware targetly, and the second one aims at the software and the processing algorithm. The drift error is the primary content of MEMS precision improvement research. We starts from the second way to study the static output noise and random drift error compensation methods of MEMS gyroscopes.Firstly, several performance indicators of MEMS gyroscopes are introduced and then the datum collection of hardware system is set up. The Allan variance is employed to analyze the output noise..3 principle is applied to filter the singular points, and regression algorithm is applied to removing the drift trend, and then the output is converted to a smooth, zero-mean random drift data. According to the autocorrelation coefficients ACF, partial correlation coefficient PCF, AIC selected criteria, time-series model is built and parameters are calculated. Then making use of Kalman filter based on AR model is expounded. The compensating results of MPU-6050 MEMS gyroscope show that the performance can be increased up to 50% effectively.Furthermore, AR & SVM model is introduced supplementally dealing with the weak points of Kalman filter method. AR model above is proposed to illustrate the linear data and SVM model is applied to train and predict. Allan variance and the other performance index are calculated by finishing all the steps including normalization, phase space reconstruction, data training and verification. The results demonstrate that this model obtained better denoising capability and the performance can be increased over 80%.