Abstract: Vehicular Non-holonomic constraint (NHC) is a commonly used enhancement technique in combined vehicle-mounted GNSS/SINS navigation in order to enhance the positioning effect. The constraint effect of traditional NHC methods is enhanced by applying machine learning methods to establish a complex mapping relationship between IMU outputs and NHC pseudo-observations, and to adjust the size of NHC pseudo-observations directly in the observation domain. Existing machine learning methods do not consider the influence of vehicle motion state, resulting in poor NHC prediction accuracy and reliability. Recent studies have shown that machine learning can predict the forward speed of vehicles, i.e., virtual odometry(ODO). However, current research mainly discusses predicting virtual NHC and virtual odometry separately, without fully exploring the coupling relationship between the two and the ability of 3D speed to fully constrain the vehicle. Therefore, the study in this paper addresses this issue in depth and proposes an LSTM neural network based on vehicle speed classification for predicting the 3D speed of a vehicle and self-adaptively adjusting its variance domain using 3D speed constraints on the new interest. In order to verify the effectiveness of this paper's method, an experimental test of vehicle-mounted GNSS/SINS combined navigation is carried out. According to the experimental results, the average accuracy of this paper's method in forward velocity prediction is about 0.4 m/s, and the average accuracy in lateral and vertical velocity prediction is about 2 cm/s. In addition, in the case of simulated GNSS signals being out-of-lock for 460 seconds, compared with the inertial derivation results, this paper's method improves the horizontal localization accuracy under the three-dimensional velocity constraint by 99.40%.