为了提高寄生式时栅传感器的测量精度， 分析了它的工作原理和动态误差组成,得到其主要误差分量为常值误差、周期误差和随机误差等。针对寄生式时栅误差特点，建立了寄生式时栅动态误差高精度预测模型，并与其他建模方法进行了比较。选用插入标准值的贝叶斯预测模型，以实际测量的传感器第一个对极动态误差数据进行建模，在后续对极特定位置插入部分实际误差测量数据，建立误差预测模型，预测了传感器后83个对极的动态误差。 另选用三次样条插值和BP神经网络建模方法对寄生式时栅整圈动态误差建模，并与建立的误差模型进行了对比。验证实验表明，三次样条插值建模时间最短(0.62 s)，但其建模精度不高(16.050 0″ ); 贝叶斯动态模型建模时间(0.86 s)略长于三次样条插值，但建模精度最高(0.415 3″); BP神经网络建模时间最长(32 min)，但建模精度最低(19.680 2″)。同时贝叶斯插入标准值建模方法所需数据点(69395个)远少于三次样条和BP神经网络建模数据点(235526个)，节省了大量的标定时间和建模数据量，因此可用于寄生式时栅传感器的动态测量误差高精度建模修正。

To improve the measurement accuracy of a parasitic time grating sensor, the working principle and dynamic error composition of the sensor were analyzed deeply and the main error components including constant error, periodic error and random error were obtained. According to the error characteristics of parasitic time grating, a high precise prediction model for dynamic error of the parasitic time grating was established and the modeling method was compared with other modeling methods. The Bayesian prediction model interpolated with standard values was chosen to build the error prediction model based on the actually measured dynamic error data of first pole in the sensor. Then, a part of actual measurement error data were inserted in the specific location of subsequent pole to establish the error prediction model to predict the dynamic error of 83 poles of the sensor. The modeling method of cubic spline interpolation and BP neural network were used to build the whole circle dynamic error model of parasitic time grating sensor and compared with the above Bayesian model. The modeling verification experiment results show that the modeling time of cubic spline interpolation method is the shortest (0.62 s), but the modeling accuracy is not high(16.050 0″). The modeling time of Bayesian prediction model is slightly longer than that of the cubic spline interpolation(0.86s), but the modeling accuracy is the highest one(0.415 3″). The modeling time of BP neural network method is the longest one (32 min), and the modeling accuracy is the lowest one (19.680 2″). Moreover, the modeling data points of Bayesian prediction model interpolated with standard value(69395) is far less than that of cubic spline interpolation and BP neural network(235526). Therefore, Bayesian prediction model interpolated with standard values saves a lot of calibration time and modeling data points, and can be used for high precision modeling and dynamic measurement error correction of parasitic time grating sensors.