CN-121829571-B - High-speed rail train positioning method based on improved LSTM

Abstract

The invention discloses a high-speed rail train positioning method based on an improved LSTM, and belongs to the technical field of rail transit navigation and positioning. The method comprises the steps of firstly preprocessing collected inertia and speed data based on rail train kinematics constraint, secondly constructing and training an IND-BN-LSTM neural network model which adopts a multichannel parallel structure with latitude, longitude and high decoupling, introducing a Batch Normalization (BN) layer after the LSTM layer of each channel to stabilize characteristic distribution, and finally applying motion constraint again in a prediction output stage to correct the track. The invention combines structural innovation with physical constraint, remarkably improves the stability and prediction precision of the model under high dynamic state, is mainly used for realizing continuous and high-precision autonomous positioning of a high-speed rail train under GNSS refusing environments such as tunnels, mountain areas and the like, and provides reliable position information for train operation control.

Inventors

• HAN ZHIFENG
• Cong Xiangcheng
• XU YING
• MENG KUN
• SUN HUI
• XING LEI
• LIU DAN
• WEI ZHAOKUN

Assignees

• 山东科技大学

Dates

Publication Date

20260512

Application Date

20260316

Claims (3)

1. 1. A method for locating a high speed rail train based on an improved LSTM, applied to a satellite navigation signal failure environment, the method comprising the steps of: S1, data acquisition and motion constraint-based input preprocessing; S2, constructing an IND-BN-LSTM neural network model; Constructing a network model comprising three mutually independent and parallel sub-channels, wherein the three sub-channels are respectively a latitude channel, a longitude channel and a height channel, and the sub-channels independently receive the same input characteristics and have mutually not shared weight parameters; s3, training a model; s4, predicting under GNSS satellite failure and performing output post-processing based on kinematic constraint; The step S1 specifically comprises the following steps: s101, data acquisition, namely acquiring triaxial specific force output by an inertia measurement unit in the running process of a train Angular velocity of three axes And high-precision latitude output by integrated navigation system when satellite signal is valid Longitude and longitude And height of And obtain the speed under the navigation coordinate system And a coordinate conversion matrix ; S102, converting speed coordinates by using a coordinate conversion matrix Velocity vector in navigation coordinate system Conversion to a carrier coordinate system ; S103, input end motion constraint processing, namely, according to the kinematic characteristics of the train running mainly along the advancing direction under the carrier coordinate system, carrying out motion constraint processing on the carrier coordinate system Transverse velocity component in (a) And a vertical velocity component Performing inhibition processing to obtain corrected carrier coordinate system speed And then will Converting back to the navigation coordinate system to obtain the speed after the preprocessing of the kinematic constraint As one of the input features of the neural network; In step S3, in the model training, the input characteristics are that the triaxial specific force output by the accelerometer in the inertial sensor is selected Triaxial angular velocity of gyroscope output Speed after pretreatment in step S1 As an input feature; the output characteristic is that the position increment y output by the inertial satellite integrated navigation system when the satellite signal is effective is taken as the output characteristic; the step S4 specifically includes the following steps: Step 4.1, predicting position increment; When satellite signals fail, input features which are acquired in real time and subjected to S1 pretreatment are input into a trained IND-BN-LSTM model, and latitude increment is respectively predicted and obtained Longitude increment And height increment ; Step 4.2, outputting post-treatment; Converting the predicted geodetic position increment into a navigation coordinate system displacement Displacing the navigation coordinate system Conversion to a carrier coordinate system And suppresses the transverse displacement component under the carrier coordinate system according to the kinematic constraint of the train And a vertical displacement component And finally, shifting the corrected carrier coordinate system First reverse conversion back to navigation coordinate system displacement And updating the position under the geodetic coordinate system according to the position to obtain the positioning result of the current moment of the train 。
2. 2. The improved LSTM based high-speed rail train positioning method as recited in claim 1, wherein in step S2, the long-short-term memory network layer is an LSTM layer for extracting timing characteristics from the input characteristic sequence; bulk normalization layer, i.e., BN layer, features for output to LSTM layer after connection to LSTM layer The normalization and affine transformation are carried out by calculating the input characteristics of the layer in the current training batch Mean of (2) And variance of ; Normalizing the input features; Affine transformation is carried out on the standardized characteristics; the fully connected FC layer is used for mapping the characteristics output by the BN layer into position increment predicted values with single dimension; and independently training the three sub-channels by using a data set containing the input features and corresponding high-precision position increment labels.
3. 3. The method for locating a high-speed rail train based on the improved LSTM as claimed in claim 1, wherein in step 4.2, the method specifically comprises the following steps: step 4.2.1, position increment rotary displacement; Converting the predicted geodetic position increment into a navigation coordinate system displacement ; Step 4.2.2, displacement projection is carried out on the carrier coordinate system; using a coordinate transformation matrix Displacing the navigation coordinate system Projection to the displacement of the carrier coordinate system ; Step 4.2.3, output end motion constraint processing; Inhibiting the displacement of the carrier coordinate system according to the kinematic constraint of the train The amount of lateral displacement in (a) And a vertical displacement component Obtaining the correction displacement under the carrier coordinate system ; Step 4.2.4, correcting the displacement inverse transformation and updating the coordinates; correction displacement under carrier coordinate system Correction displacement under reverse conversion back navigation coordinate system And accordingly, accumulating and updating the position under the geodetic coordinate system; and 4.2.5, outputting a positioning result.

Description

High-speed rail train positioning method based on improved LSTM Technical Field The invention belongs to the technical field of rail transit navigation and positioning, and particularly relates to a high-speed rail train positioning method based on an improved LSTM. Background In a train operation control system, a positioning technology plays an important role, and the system needs to calculate a safe protection distance according to the real-time position, speed and operation direction of each train so as to prevent rear-end collision or collision of the trains. Meanwhile, the dispatching center needs to master all train positions in real time, so that the operation plan is adjusted, the emergency is processed, and the vehicle turnover is performed more effectively. The train position information is obtained reliably and accurately in real time, and is a precondition for realizing safe, efficient and green operation of rail transit. Currently, rail trains are mainly positioned using ground transponders and speed radars. The ground transponder only provides position and speed information when the train passes, continuous positioning of the train cannot be realized, and the ground transponder has high installation, test and maintenance costs and static positioning deviation and dynamic positioning deviation. In addition, in remote, harsh environments, it is difficult to deploy trackside equipment. Therefore, the requirement of high-precision positioning of the rail train cannot be met by merely positioning by means of a transponder and the like, and with the development of a global satellite navigation system (Global Navigation SATELLITE SYSTEM, GNSS), the satellite/inertial integrated navigation is gradually applied to rail train positioning and becomes a research hot spot. However, satellite signals are easily shielded or even interrupted in environments such as tunnels, urban canyons and the like, so that inertial navigation errors are rapidly dispersed along with time, and the long-time high-precision positioning requirement cannot be met. In recent years, a recurrent neural network (Recurrent Neural Network, RNN) has been widely used in the navigation field due to its strong ability to mine effective information from time-series data, and its applications are mainly classified into the following four types. The first class is to predict the position directly without using Kalman filtering algorithm, the second class is to predict GNSS increment to assist Kalman filtering update, the third class is to predict Kalman filtering state vector and then correct the positioning result, and the fourth class is to predict Kalman filtering noise covariance matrix. However, the above prior art is mainly focused on the field of road car navigation, and relatively little research is conducted on rail trains. While some techniques have attempted to apply recurrent neural networks to train positioning, these characteristics present the following challenges to conventional LSTM models when dealing with the high dynamics specific to rail trains (embodied by rapid changes in speed and acceleration, nonlinear coupling between dimensions, and significant instability of feature distribution); (1) The high dynamics exacerbates the internal covariate offset (Internal Covariate Shift) and the rapid changes in speed and acceleration during high speed operation of the train cause significant fluctuations in LSTM output characteristics over time. The input feature profile of the fully connected layer (Fully Connected Layer, FC) exhibits instability due to the lack of a normalization mechanism for the intermediate layer. The internal covariate offset forces the FC layer to continuously adapt to new data distribution in the training process, so that the weight updating direction is unstable, the model is difficult to converge, and the prediction precision is obviously reduced; (2) The statistical characteristics of all target variables have significant differences, which cause serious inter-dimensional mutual interference under the traditional single-model joint training framework. In particular, the high variance dimensions, due to their strongly fluctuating nature, produce a large error signal during training, making gradient updates largely affected by these dimensions. However, these dimensions are difficult to fit stably, and the prediction accuracy is still low, due to their complex nonlinearity and high uncertainty. As a result, gradient resources are occupied by high variance dimensions, efficient learning of the model on low variance dimensions is suppressed, and overall performance improvement is ultimately limited. (3) The prior deep learning algorithm mostly lacks physical cognition on the kinematic constraint characteristics of the train. Particularly in high dynamic scenarios, more significant lateral and vertical disturbance components are contained in the sensor data. Due to the lack of an inherent kinematic constra