A high-speed rail train positioning method based on improved LSTM

Abstract

The application 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 application comprises the following steps: firstly, pre-processing the collected inertial and speed data based on the kinematic constraint of the rail train; secondly, constructing and training an IND-BN-LSTM neural network model, the model adopts a multi-channel parallel structure decoupled by latitude, longitude and height, and a batch normalization (BN) layer is introduced after the LSTM layer of each channel to stabilize the feature distribution; finally, applying the motion constraint again in the prediction output stage to correct the trajectory. The application combines structural innovation with physical constraints, significantly improves the stability and prediction accuracy of the model under high dynamics, and is mainly used for realizing continuous and high-precision autonomous positioning of the high-speed rail train in GNSS denial environments such as tunnels and mountainous areas, and providing reliable position information for train operation control.

Description

Technical Field

[0001] This invention belongs to the field of rail transit navigation and positioning technology, specifically relating to a high-speed rail train positioning method based on an improved LSTM. Background Technology

[0002] In train operation control systems, positioning technology plays a crucial role. The system needs to calculate safe protection distances based on the real-time position, speed, and direction of travel of each train to prevent rear-end collisions or head-on collisions. Simultaneously, the dispatch center needs to monitor the real-time positions of all trains to more effectively adjust operational plans, handle emergencies, and manage vehicle turnaround. Real-time, reliable, and accurate acquisition of train position information is a prerequisite for achieving safe, efficient, and green rail transit operations. Currently, rail trains primarily use ground transponders and speed-measuring radar for positioning. Ground transponders only provide position and speed information as the train passes, failing to achieve continuous train positioning. Furthermore, they are costly to install, test, and maintain, and suffer from static and dynamic positioning errors. Additionally, in remote and harsh environments, it is difficult to deploy trackside equipment. Therefore, relying solely on transponders for positioning cannot meet the high-precision positioning requirements of rail trains. With the development of Global Navigation Satellite Systems (GNSS), the application of satellite / inertial integrated navigation to rail train positioning has become a research hotspot. However, satellite signals are easily blocked or even interrupted in environments such as tunnels and urban canyons, causing inertial navigation errors to diverge rapidly over time, making it impossible to meet the requirements for long-term high-precision positioning.

[0003] In recent years, Recurrent Neural Networks (RNNs) have been widely used in the navigation field due to their powerful ability to extract useful information from time-series data. Their applications can be mainly categorized into four types: First, directly predicting position instead of using the Kalman filter algorithm; second, predicting GNSS increments to assist Kalman filter updates; third, predicting the state vector of the Kalman filter and then correcting the positioning results; and fourth, predicting the noise covariance matrix of the Kalman filter.

[0004] However, the aforementioned existing technologies mainly focus on highway vehicle navigation, with relatively little research on rail trains. Although some technologies attempt to apply recurrent neural networks to train localization, these characteristics pose the following challenges to traditional LSTM models when dealing with the unique high dynamics of rail trains (specifically manifested as rapid changes in speed and acceleration, nonlinear coupling between dimensions, and significant instability in feature distribution);

[0005] (1) High dynamics exacerbate internal covariate shift; the rapid changes in speed and acceleration during high-speed train operation cause significant fluctuations in the output features of LSTM over time. Due to the lack of a normalization mechanism in intermediate layers, the input feature distribution of the fully connected layer (FC) exhibits instability. This internal covariate shift forces the FC layer to continuously adapt to new data distributions during training, resulting in unstable weight update directions, making it difficult for the model to converge and significantly reducing prediction accuracy;

[0006] (2) Inter-dimensional interference caused by significant differences in the statistical characteristics of each target variable: Due to the influence of strong nonlinear coupling and complex motion states, the statistical characteristics of each target variable are significantly different, which leads to serious inter-dimensional interference in the traditional single-model joint training framework. Specifically, high-variance dimensions generate large error signals during training due to their drastic fluctuations, making gradient updates mainly affected by these dimensions. However, due to their complex nonlinearity and high uncertainty, these dimensions are difficult to fit stably, and the prediction accuracy remains low. As a result, gradient resources are occupied by high-variance dimensions, the effective learning of low-variance dimensions by the model is suppressed, and ultimately the improvement of overall performance is limited.

[0007] (3) Divergence of non-driving direction errors due to lack of prior physical knowledge: Most existing deep learning algorithms lack physical understanding of the kinematic constraints of trains. Especially in high-dynamic scenarios, sensor data contains more significant lateral and vertical interference components. Due to the lack of an inherent kinematic constraint mechanism, LSTM models struggle to remove this invalid information and are prone to misjudging it as valid motion features. During periods of GNSS signal loss, these incorrectly extracted features are mapped as false cumulative displacements at the output and are rapidly amplified as the train moves at high speed, causing the predicted trajectory to deviate rapidly from the true track and severely reducing the system's positioning accuracy.

[0008] In summary, existing technologies suffer from severely limited positioning accuracy due to their inability to effectively address internal covariate shifts and interdimensional interference, as well as the lack of effective kinematic constraint mechanisms.

[0009] Therefore, there is an urgent need for an intelligent positioning method that can adapt to the high dynamic characteristics of trains, overcome dimensional interference, and incorporate kinematic constraints. Summary of the Invention

[0010] To address the aforementioned technical problems in existing technologies, this invention proposes a high-speed rail train positioning method based on an improved LSTM, which effectively solves the problems of internal covariate offset, inter-dimensional interference, and non-driving direction error accumulation in high-dynamic scenarios, thereby achieving continuous and high-precision autonomous positioning of trains even in environments where satellite navigation signals are frequently unavailable.

[0011] To achieve the above objectives, the present invention adopts the following technical solution;

[0012] A high-speed rail train positioning method based on improved LSTM, applied to environments where satellite navigation signals fail, the method includes the following steps: S1, data acquisition and input preprocessing based on motion constraints;

[0013] S2. Construct the IND-BN-LSTM neural network model;

[0014] Construct a network model containing three independent and parallel sub-channels: latitude, longitude, and altitude. Each sub-channel independently receives the same input features, and the weight parameters are not shared. Each sub-channel includes, in sequence: a Long Short-Term Memory (LSTM) layer, a Batch Normalization (BN) layer, and a fully connected (FC) layer.

[0015] S3, Model Training;

[0016] Prediction under S4 and GNSS satellite failure and output post-processing based on kinematic constraints.

[0017] Preferably, step S1 specifically includes the following steps;

[0018] S101, Data Acquisition;

[0019] Collect the triaxial specific force output by the inertial measurement unit during train operation. Triaxial angular velocity And the high-precision latitude output by the integrated navigation system when satellite signals are valid. ,longitude and height And obtain the velocity in the navigation coordinate system. and coordinate transformation matrix ;

[0020] S102, Velocity coordinate transformation;

[0021] Using coordinate transformation matrix The velocity vector in the navigation coordinate system Transform to carrier coordinate system ;

[0022] S103, Input end motion constraint processing; based on the kinematic characteristics of the train mainly traveling along the forward direction in the carrier coordinate system, the carrier coordinate system... lateral velocity component and vertical velocity components Suppression processing is performed to obtain the corrected velocity in the carrier coordinate system. Then Transform back to the navigation coordinate system to obtain the velocity after kinematic constraint preprocessing. As one of the input features of the neural network; the specific formula is as follows;

[0023] (1);

[0024] (2);

[0025] (3);

[0026] (4);

[0027] in, It is the forward velocity in the carrier coordinate system. Indicates the eastward, northward, and upward velocities in the navigation coordinate system; This is the transformation matrix from the navigation coordinate system to the vehicle coordinate system; The velocity is in the carrier coordinate system; The corrected velocity in the carrier coordinate system; These are constraint coefficients; The velocity in the corrected navigation coordinate system. , , These represent the eastward, northward, and upward velocities in the corrected navigation coordinate system.

[0028] Preferably, in step S2, the Long Short-Term Memory (LSTM) network layer is used to extract temporal features from the input feature sequence, and its cell state... With hidden state Updates follow the following gating mechanism: Forget Gate, (5); Input gate, (6); Candidate state, (7); Output gate, (8); Cell state, (9);

[0029] Hidden state (10); Among them, , These are the output vectors of the forget gate, input gate, candidate state, and output gate, respectively. For the Sigmoid function, This indicates element-wise multiplication, and tanh is the hyperbolic tangent function. Enter the current time. This is the hidden state from the previous moment. This represents the current state of the cell. The current hidden state; , These are the weight matrices for the forget gate, input gate, candidate state, and output gate, respectively. , , , These are the bias vectors for the forget gate, input gate, candidate state, and output gate, respectively.

[0030] Batch normalization (BN) layers, connected after LSTM layers, are used to normalize the features output by the LSTM layers. The standardization and affine transformation process includes: calculating the input features of this layer in the current training batch. mean With variance ;

[0031] (11);

[0032] in, This represents the number of samples in the current training batch.

[0033] (12);

[0034] Standardize the input features;

[0035] (13);

[0036] in, It is a tiny constant. These are the standardized features;

[0037] Perform an affine transformation on the standardized features;

[0038] (14);

[0039] in, and For learnable scaling factors and offsets, This is the final output of the batch normalization layer;

[0040] A fully connected (FC) layer is used to map the features output by the BN layer into a single-dimensional position increment prediction value.

[0041] The three sub-channels are trained independently using a dataset containing the input features and corresponding high-precision position increment labels.

[0042] Preferably, in step S3, during model training, the input features are selected as the triaxial specific force output from the accelerometer in the inertial sensor. The three-axis angular velocity output by the gyroscope and the speed after preprocessing in step S1 As input features;

[0043] Output characteristics: The position increment y output by the inertial satellite integrated navigation system when the satellite signal is valid is used as the output characteristic;

[0044] (15);

[0045] in, For latitude increments, For longitude increments, For high increments; among them, (16);

[0046] in For integrated navigation systems at time Output high-precision position; For integrated navigation systems High-precision position output at all times.

[0047] Preferably, step S4 specifically includes the following steps;

[0048] Step 4.1: Predict the location increment;

[0049] When satellite signals fail, the input features acquired in real time and preprocessed by S1 are fed into the trained IND-BN-LSTM model to predict the latitude increments. Longitude increment and height increment Step 4.2, Post-processing: Convert the predicted geodetic coordinate system position increment into navigation coordinate system displacement. Displace the navigation coordinate system Transform to carrier coordinate system Based on train kinematic constraints, the lateral displacement component in the carrier coordinate system is suppressed. and vertical displacement components Finally, the corrected carrier coordinate system is displaced. First, reverse the displacement back to the navigation coordinate system. , and update the position in the geodetic coordinate system accordingly to obtain the positioning result of the train at the current moment .

[0050] Preferably, in step 4.2, it specifically includes the following steps;

[0051] Step 4.2.1, Convert the position increment to displacement;

[0052] Convert the predicted position increment in the geodetic coordinate system to the displacement in the navigation coordinate system ;

[0053] (17);

[0054] Where, is the radius of curvature of the meridian, is the radius of curvature of the prime vertical; , , are the position increments in the east, north, and up directions in the navigation coordinate system respectively, is the ellipsoidal height, that is, the geometric height relative to the standard ellipsoid surface;

[0055] Step 4.2.2, Project the displacement onto the vehicle coordinate system;

[0056] Use the coordinate transformation matrix , and project the displacement in the navigation coordinate system onto the displacement in the vehicle coordinate system;

[0057] (18);

[0058] Where, , , are the position increments in the front, right, and up directions in the vehicle coordinate system respectively;

[0059] Step 4.2.3, Process the motion constraints at the output end;

[0060] According to the kinematic constraints of the train, suppress the lateral displacement in and the vertical displacement component in the displacement of the vehicle coordinate system to obtain the corrected displacement in the vehicle coordinate system;

[0061] (19);

[0062] Where, is the constraint coefficient; is the corrected displacement in the vehicle coordinate system; END]], , These represent the corrected displacements in the front, right, and top directions in the carrier coordinate system, respectively.

[0063] Step 4.2.4: Correct the inverse displacement transformation and update the coordinates;

[0064] Corrected displacement in the carrier coordinate system Corrected displacement after inverse transformation back to navigation coordinate system And based on this, the position in the geodetic coordinate system is cumulatively updated;

[0065] (20);

[0066] in, This is the corrected displacement in the carrier coordinate system; This is the corrected displacement in the navigation coordinate system; , , These are the corrected displacements in the east-west, north-south, and sky directions of the navigation coordinate system, respectively.

[0067] (twenty one);

[0068] in, , , They are respectively The latitude, longitude, and altitude at any given time; , , They are respectively Latitude, longitude, and altitude at time -1;

[0069] Step 4.2.5: Output the positioning results;

[0070] Output This serves as the train's real-time location at the current moment.

[0071] Preferably, the constraint coefficient The value range is between 0 and 1.

[0072] The beneficial technical effects of this invention are as follows: 1. Effectively suppressing internal covariate shift and improving model stability; introducing a BN layer between the LSTM layer and the FC layer to enhance adaptability to highly dynamic operating data. This design effectively suppresses internal covariate shift by normalizing the feature distribution of the LSTM output, thereby improving training speed and model stability, and giving it better accuracy and generalization ability compared to traditional LSTM models.

[0073] 2. Eliminate interdimensional interference and optimize overall positioning accuracy; design a multi-channel parallel structure, dividing the model into three independent channels: latitude, longitude, and altitude. Each channel adopts the above-mentioned BN-LSTM structure, enabling each channel to focus on predicting a single location increment and learn the mapping relationship more accurately.

[0074] 3. Incorporating physical priors to prevent trajectory divergence: Motion constraint processing was implemented at both the input and output ends of the model. By eliminating noise components in non-travel directions (i.e., lateral and vertical directions) at the model input end, the model can focus on learning the true motion law and position increment mapping relationship of the train in the forward dimension. At the same time, at the model output end, the lateral and vertical position increments in the carrier coordinate system were suppressed, thereby physically cutting off the accumulation path of inertial error in non-travel directions and effectively improving positioning accuracy. Attached Figure Description

[0075] Figure 1 This is an overall flowchart of the high-speed rail train positioning method based on improved LSTM provided in the embodiments of the present invention.

[0076] Figure 2 This is a schematic diagram of the structure of the IND-BN-LSTM model in an embodiment of the present invention.

[0077] Figure 3 This is a comparative diagram of the traditional LSTM structure and the BN-LSTM structure used in this invention.

[0078] Figure 4 This is a diagram of the LSTM neuron structure.

[0079] Figure 5 This is a northward position prediction map of the embodiment of the present invention and the comparative method (LSTM, GRU) during GNSS failure.

[0080] Figure 6 This is an eastward position prediction map of the embodiment of the present invention and the comparative method (LSTM, GRU) during GNSS failure.

[0081] Figure 7 This is a height-direction position prediction map of the embodiment of the present invention and the comparative method (LSTM, GRU) during GNSS failure.

[0082] Figure 8 This is a northbound positioning error diagram of the embodiment of the present invention and the comparative method (LSTM, GRU) during GNSS failure.

[0083] Figure 9 This is a map showing the eastward positioning error of the embodiment of the present invention and the comparative method (LSTM, GRU) during GNSS failure.

[0084] Figure 10 This is a height orientation positioning error diagram of the embodiment of the present invention and the comparative method (LSTM, GRU) during GNSS failure. Detailed Implementation

[0085] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0086] This invention provides a high-speed rail train positioning method based on an improved LSTM, the overall process of which is as follows: Figure 1 As shown. The core lies in constructing the IND-BN-LSTM neural network model (structure as shown). Figure 2 As shown in the figure, train kinematic constraints were incorporated.

[0087] S1. Data acquisition and motion-constraint-based input preprocessing; specifically including the following steps;

[0088] S101, Data Acquisition;

[0089] Raw data on train operation is collected via onboard satellite receiver, inertial measurement unit, and computing unit; three-axis specific force... Triaxial angular velocity And the high-precision latitude output by the integrated navigation system when satellite signals are valid. ,longitude and height And obtain the velocity in the navigation coordinate system output by the integrated navigation system. and coordinate transformation matrix ;

[0090] S102, Velocity coordinate transformation;

[0091] Using coordinate transformation matrix The velocity vector in the navigation coordinate system Transform to the carrier coordinate system to obtain ;

[0092] S103, Input end motion constraint processing; Considering that the train is subject to the physical constraints of the track and mainly moves along the direction of the vehicle's forward movement, with lateral and vertical movements restricted, therefore, the lateral velocity component in the carrier coordinate system is processed. and vertical velocity components Apply suppression (multiplied by a coefficient) (e.g., 0.1) to obtain the corrected velocity in the carrier coordinate system. Then Transform back to the navigation coordinate system to obtain the velocity after kinematic constraint preprocessing. As one of the input features of the neural network, this process uses prior physical knowledge to weaken interference from non-primary motion directions in the input data. The specific formula is as follows;

[0093] (1);

[0094] (2);

[0095] (3);

[0096] (4);

[0097] in, , , These represent the eastward, northward, and upward velocities in the navigation coordinate system, respectively. This is the transformation matrix from the navigation coordinate system to the vehicle coordinate system; The velocity is in the carrier coordinate system; The corrected velocity in the carrier coordinate system; This is a constraint coefficient, with a value ranging from 0 to 1; The velocity in the corrected navigation coordinate system; , , These represent the eastward, northward, and upward velocities in the corrected navigation coordinate system.

[0098] S2. Construct the IND-BN-LSTM neural network model;

[0099] like Figure 2 As shown, the model consists of three independent sub-channels (latitude, longitude, and altitude) operating in parallel. Each sub-channel employs a cascaded "LSTM-BN-FC" structure, i.e., a BN-LSTM structure (compared to traditional LSTM, for example...). Figure 3 ).

[0100] (1) The LSTM layer is used as a feature extraction layer to capture long dependencies in time series.

[0101] LSTM neuron structure such as Figure 4 As shown. LSTM adds an input gate (Input Gate, Forget Gate Output Gate ) and cell state (Cell State) LSTM is used to handle long dependencies in sequences. The core of LSTM is the memory cell, which is essentially a cell state that can transmit information. In theory, it allows information to remain unchanged during transmission and controls the forgetting and updating of cell state information through gates based on the Sigmoid activation function.

[0102] The Gate of Oblivion decides to start from the previous moment Which data is discarded? It discards the hidden state from the previous moment. and current input As input, the sigmoid function outputs a vector with the same dimensions as the cell state. Each element in the vector has a value between 0 and 1, representing the state of the cell. The degree of retention of the corresponding information (1 for complete retention, 0 for complete rejection). It is the Sigmoid activation function, whose output is between (0, 1).

[0103] (5).

[0104] The input gate determines which new information will be stored in the current cell state. This process involves two steps: first, the sigmoid layer generates the input gate vector; second, the tanh layer creates a candidate state. This vector contains any new information that may be added. tanh is the hyperbolic tangent activation function, whose output is between (-1, 1).

[0105] (6);

[0106] (7).

[0107] The output gate controls the transition from the current cell state. What information is output to the hidden state? The hidden state is a filtered version based on the cell state; it serves as both the output of the current time step and one of the inputs for the next time step.

[0108] (8);

[0109] The core workflow of LSTM is as follows: First, the current cell state. It is calculated through a "forget-update" mechanism that combines the output of the forget gate and the candidate states controlled by the input gate; then the updated cell state is calculated. pass The function undergoes a nonlinear transformation; finally, the output gate... The output is multiplied element-wise with this result to obtain the hidden state at the current time step. ; This represents the element-wise product (Hadamard product).

[0110] (9);

[0111] (10);

[0112] in, , These are the output vectors of the forget gate, input gate, candidate state, and output gate, respectively. For the Sigmoid function, This indicates element-wise multiplication, and tanh is the hyperbolic tangent function. Enter the current time. This is the hidden state from the previous moment. This represents the current state of the cell. The current hidden state; , These are the weight matrices for the forget gate, input gate, candidate state, and output gate, respectively. , , , These are the bias vectors for the forget gate, input gate, candidate state, and output gate, respectively.

[0113] (2) Batch normalization layer.

[0114] Normalization or standardization in the data preprocessing stage mainly targets the global distribution of the data, eliminating scale differences between features, but it is difficult to eliminate the drastic fluctuations caused by high dynamics. Although LSTM can capture long-term dependencies, under the high dynamic operation of a train, its output features will change significantly due to the drastic fluctuations in the input. This instability will be directly passed to subsequent fully connected (FC) layers. As training progresses, this fluctuation will exacerbate the internal covariate shift phenomenon—that is, the network layer input distribution changes continuously with parameter updates, leading to slow training convergence, unstable gradients, and affecting the final positioning accuracy.

[0115] The operation of the Batch Normalization (BN) layer can be divided into two parts: normalization and affine transformation. Assume a mini-batch of size i. ,in This is the activation vector of a certain layer in the network. First, the Batch Normalization (BN) layer calculates the mean of any feature dimension of this feature vector within the current mini-batch. and variance ;

[0116] (11);

[0117] (12);

[0118] in, This represents the number of samples in the current training batch.

[0119] Using the calculated mean and variance, for each activation value Standardize to obtain ;

[0120] (13);

[0121] in It is a tiny constant used to ensure that the denominator is not zero, thus maintaining the stability of numerical calculations; These are the standardized features.

[0122] Forcing activation values ​​to have a mean of 0 and a variance of 1 might affect the network's expressive power. Therefore, Batch Normalization (BN) layers introduce learnable scaling factors. and It is used to perform affine transformations on the standardized activation values.

[0123] (14);

[0124] in, and For learnable scaling factors and offsets, This is the final output of the batch normalization layer.

[0125] During model training, the weights of the LSTM layers are updated after each batch. This causes the way the LSTM extracts features to change continuously, and the statistical distribution of its output hidden state vector fluctuates accordingly, with more dramatic fluctuations under high dynamic conditions. Since the fully connected (FC) layers need to learn stable mapping relationships based on these features, the instability of the input distribution may increase the learning difficulty, leading to unstable training and reduced final model performance.

[0126] To mitigate the drastic fluctuations in LSTM output features caused by highly dynamic train operation and reduce their impact on subsequent network layers, a BN-LSTM structure was constructed as the foundational component of the overall model. This involves introducing a BN layer between the LSTM and FC layers to enhance adaptability to highly dynamic operating data. This design effectively suppresses internal covariate shifts by normalizing the feature distribution of the LSTM output, thereby improving training speed and model stability, resulting in better accuracy and generalization ability compared to traditional LSTM models.

[0127] S3, Model Training;

[0128] In areas with good satellite signal, a large amount of input features are collected ( ) and the corresponding real location incremental label (Obtained by differential calculation of high-precision integrated navigation results, formula (16)). Using this dataset, the backpropagation algorithm is employed to minimize the error (such as mean square error) between the predicted increment and the actual increment, and the three sub-channels are trained independently until the model converges.

[0129] Input characteristics; select the triaxial specific force output from the accelerometer. The three-axis angular velocity output by the gyroscope and the speed after preprocessing in step S1 As input features.

[0130] Output characteristics; the position increment y output by the integrated navigation system when satellite signals are valid is used as the output characteristic;

[0131] (15);

[0132] in, For latitude increments, For longitude increments, For high increments; among them, (16);

[0133] in, For integrated navigation systems at time Output high-precision position; For integrated navigation systems High-precision position output at all times.

[0134] S4. Prediction under GNSS failure and post-processing of output based on kinematic constraints. The prediction process is initiated when the train enters a GNSS failure area such as a tunnel.

[0135] Step 4.1: Predict the location increment;

[0136] When satellite signals fail, the input features acquired in real time and preprocessed by S1 are fed into the trained IND-BN-LSTM model to predict the latitude increments. Longitude increment and height increment . Step 4.2, Output Post-processing; this is the second key constraint to prevent error accumulation and divergence. It specifically includes the following steps;

[0137] Step 4.2.1: Convert position increment to displacement;

[0138] Convert the predicted position increment in the geodetic coordinate system to the displacement in the navigation coordinate system ;

[0139] (17);

[0140] where is the radius of curvature of the meridian, is the radius of curvature of the prime vertical; , , are the position increments in the east, north, and up directions in the navigation coordinate system respectively; is the ellipsoidal height, i.e., the geometric height relative to the standard ellipsoid surface.

[0141] Step 4.2.2, Project the displacement to the vehicle coordinate system;

[0142] Use the coordinate transformation matrix to project the displacement in the navigation coordinate system to the displacement in the vehicle coordinate system ;

[0143] (18);

[0144] where , , are the position increments in the forward, right, and up directions in the vehicle coordinate system respectively.

[0145] Step 4.2.3, Process the motion constraints at the output end;

[0146] According to the kinematic constraints of the train, suppress the lateral displacement in the displacement in the vehicle coordinate system and the vertical displacement component to obtain the corrected displacement in the vehicle coordinate system;

[0147] (19);

[0148] where is the constraint coefficient; is the corrected displacement in the vehicle coordinate system;<00005​​​​​​​​​​​​​ And based on this, the position in the geodetic coordinate system is cumulatively updated;

[0151] (20);

[0152] (twenty one);

[0153] in, This is the corrected displacement in the carrier coordinate system; This is the corrected displacement in the navigation coordinate system; These are the corrected displacements in the east-west, north-south, and sky directions of the navigation coordinate system, respectively. This represents the current latitude, longitude, and altitude. , , for Latitude, longitude, and altitude at time -1.

[0154] Step 4.2.5: Output the positioning results to complete one positioning cycle;

[0155] Output This serves as the train's real-time location at the current moment.

[0156] Key points of this invention: 1) To address the issues of drastic fluctuations in data distribution and internal covariate shifts caused by high-speed train operation, this invention designs a BN-LSTM structure. This structure enhances its adaptability to highly dynamic operating data by introducing a BN layer between the LSTM layer and the FC layer. This design effectively suppresses internal covariate shifts by normalizing the feature distribution of the LSTM output, thereby improving training speed and model stability, resulting in better accuracy and generalization ability compared to traditional LSTM models.

[0157] 2) To address the problem of inter-dimensional interference caused by significant differences in the statistical characteristics of each target variable, a multi-channel parallel structure with high latitude and longitude decoupling was designed to eliminate inter-dimensional interference and improve the overall accuracy of the three-dimensional positioning results.

[0158] 3) To address the issue that traditional LSTM models are prone to interference and thus prone to trajectory divergence due to a lack of prior physical knowledge, a dual mechanism based on train kinematic constraints for input preprocessing and output correction was designed. By eliminating lateral and vertical noise components of the velocity vector at the input end and suppressing lateral and vertical increments of the displacement vector at the output end, the propagation of accumulated errors in non-travel directions is effectively blocked, thereby improving the navigation reliability of the system in scenarios with no satellite signal for extended periods.

[0159] Effectiveness Verification: The effectiveness of this invention was verified by testing based on actual train operation data. The experiment used a Spatial integrated navigation module, with the satellite antenna fixed above the central axis of the train carriage. The inertial navigation module was fixed inside the carriage, aligned vertically with the satellite antenna. The experimental data collection section was the Lanzhou-Zhangye Railway from Wuwei Station to Lanzhou Station, with a maximum speed of 234 km / h. The total data collection time was approximately 92 minutes, with a sampling frequency of 100 Hz. Data from 14:28:32 to 14:40:31 on April 15, 2025, a total duration of 720 seconds, was selected from the collected data. At this sampling frequency, the experimental dataset included 72,000 samples. The training period was from 14:28:32 to 14:38:31, a total of 600 seconds, with 60,000 samples; the prediction period was from 14:38:32 to 14:40:31, a total of 120 seconds, with 12,000 samples. The solution results of the Spatial integrated navigation chip were used as experimental reference values.

[0160] 1) Before model training, kinematic constraints are first applied to the velocity to reduce the random fluctuations in the input data, thus providing a smoother input for subsequent model training and improving the stability of the training process and the reliability of the prediction results. Experimental results show that the variance of the northward velocity decreased from 22.5951 to 21.2284, the variance of the eastward velocity decreased from 125.5129 to 106.1216, and the variance of the celestial velocity decreased from 0.1311 to 0.0864.

[0161] 2) To verify the function of each module, ablation experiments were conducted under the same experimental conditions. The hidden layer dimension of the model was 200, the maximum number of iterations was 200, and the learning rate was 2e-3. The experiments included a traditional LSTM model; a BN-LSTM model with a BN layer; an IND-LSTM model decoupled from LSTM; an LSTM+kinematic constraint model with kinematic constraints; and the model proposed in this invention. Root Mean Square Error (RMSE) was used as the evaluation metric, which can effectively reflect the overall deviation level of the predicted trajectory from the true trajectory. The experimental results are listed in Table 1.

[0162] Table 1: Ablation Experiment Results;

[0163] .

[0164] As shown in Table 1, the 3D RMSE of the LSTM model is as high as 818.0170 meters; the 3D RMSE of the BN-LSTM model is reduced to 665.6110 meters, an improvement of 18.6% compared to the traditional LSTM model; the 3D RMSE of the IND-LSTM model is reduced to 519.7383 meters, an improvement of 36.5% compared to the traditional LSTM model; the 3D RMSE of the LSTM+kinematic constraint model is reduced to 759.9961 meters, an improvement of 7.1% compared to the traditional LSTM model; and the 3D RMSE of the model proposed in this invention is reduced to a low of 291.2651 meters, an improvement of 64.4% compared to the traditional LSTM model. In summary, the three innovations of this invention all have varying degrees of effect on improving positioning accuracy, and the combined effect is the best.

[0165] 3) To further evaluate the performance ceiling of the models under real-world application conditions, each model in this section adopts a suitable network structure and training parameters, and is compared with the LSTM and GRU models. The hidden layer dimension of the LSTM and GRU models is 500, and the learning rate is 1e-3. The parameters of the IND-BN-LSTM model are set as follows: the hidden layer dimension of the latitude model LSTM is 500, and the learning rate is 1e-3; the hidden layer dimension of the longitude model LSTM is 500, and the learning rate is 5e-3; the hidden layer dimension of the altitude model LSTM is 400, and the learning rate is 1e-3. The maximum number of iterations in this part of the experiment is 200. The experimental results are listed in Table 2. The results for the north, east, and altitude directions in the location prediction are as follows: Figure 5 , Figure 6 , Figure 7 As shown, the positioning errors in the north, east, and altitude directions are respectively as follows: Figure 8 , Figure 9 , Figure 10 As shown.

[0166] Table 2: Results of Model Comparison Experiment;

[0167] .

[0168] As shown in Table 2, although the GRU model improves the 3D accuracy by 65.2% compared to the traditional LSTM model, initially improving the prediction performance, its RMSE remains as high as 181.6008 meters in the dynamically fluctuating eastward dimension. This indicates that the GRU model struggles to cope with the drastic fluctuations in data and the interference between dimensions. In contrast, the model proposed in this invention significantly outperforms other comparative models in all dimensions, with its overall 3D RMSE further reduced to 66.1147 meters, representing an 87.4% improvement in overall positioning accuracy compared to the traditional LSTM model. This confirms that the model of this invention has stronger error suppression capabilities and higher positioning stability in the high-dynamic scenario of rail trains.

[0169] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A high-speed rail train positioning method based on an improved LSTM, characterized in that, Applied to environments where satellite navigation signals are unavailable, the method includes the following steps: S1. Data acquisition and motion-constraint-based input preprocessing; S2. Construct the IND-BN-LSTM neural network model; Construct a network model containing three independent and parallel sub-channels: latitude, longitude, and altitude. Each sub-channel independently receives the same input features, and the weight parameters are not shared. Each sub-channel includes, in sequence: a Long Short-Term Memory (LSTM) layer, a Batch Normalization (BN) layer, and a fully connected (FC) layer. S3, Model Training; Prediction under S4 and GNSS satellite failure and output post-processing based on kinematic constraints; Step S1 specifically includes the following steps: S101, Data Acquisition; Acquiring the three-axis specific force output by the inertial measurement unit during train operation. Triaxial angular velocity And the high-precision latitude output by the integrated navigation system when satellite signals are valid. ,longitude and height And obtain the velocity in the navigation coordinate system. and coordinate transformation matrix ; S102, Velocity coordinate transformation; using the coordinate transformation matrix The velocity vector in the navigation coordinate system Transform to carrier coordinate system ; S103, Input end motion constraint processing; based on the kinematic characteristics of the train mainly traveling along the forward direction in the carrier coordinate system, the carrier coordinate system... lateral velocity component and vertical velocity components Suppression processing is performed to obtain the corrected velocity in the carrier coordinate system. Then Transform back to the navigation coordinate system to obtain the velocity after kinematic constraint preprocessing. As one of the input features of a neural network; In step S3, during model training, the input feature is selected as the triaxial specific force output from the accelerometer in the inertial sensor. The three-axis angular velocity output by the gyroscope and the speed after preprocessing in step S1 As input features; Output characteristics: The position increment y output by the inertial satellite integrated navigation system when the satellite signal is valid is used as the output characteristic; Step S4 specifically includes the following steps: Step 4.1: Predict the location increment; When satellite signals fail, the input features acquired in real time and preprocessed by S1 are fed into the trained IND-BN-LSTM model to predict the latitude increments. Longitude increment and height increment ; Step 4.2, Post-processing of output; The predicted geodetic coordinate system position increment is converted into a navigation coordinate system displacement. Displace the navigation coordinate system Transform to carrier coordinate system Based on train kinematic constraints, the lateral displacement component in the carrier coordinate system is suppressed. and vertical displacement components Finally, the corrected carrier coordinate system is displaced. First, reverse the displacement back to the navigation coordinate system. Based on this, the position in the geodetic coordinate system is updated to obtain the train's current location. .

2. The high-speed rail train positioning method based on improved LSTM according to claim 1, characterized in that, In step S2, the Long Short-Term Memory (LSTM) network layer is used to extract temporal features from the input feature sequence. Batch normalization (BN) layer: Connected after the LSTM layer, it is used to normalize the features output by the LSTM layer. Perform standardization and affine transformation; The specific process includes: calculating the input features of this layer in the current training batch. mean With variance ; Standardize the input features; Perform an affine transformation on the standardized features; Fully connected (FC) layer: used to map the features output by the BN layer to a single-dimensional position increment prediction value; The three sub-channels are trained independently using a dataset containing the input features and corresponding high-precision position increment labels.

3. The high-speed rail train positioning method based on improved LSTM according to claim 1, characterized in that, Step 4.2 specifically includes the following steps: Step 4.2.1: Convert position increment to displacement; The predicted geodetic coordinate system position increment is converted into a navigation coordinate system displacement. ; Step 4.2.2: Project the displacement onto the carrier coordinate system; Using coordinate transformation matrix Displace the navigation coordinate system Displacement projected onto the carrier coordinate system ; Step 4.2.3: Output end motion constraint processing; Based on train kinematic constraints, suppress the displacement of the carrier coordinate system. Lateral displacement and vertical displacement components The corrected displacement in the carrier coordinate system is obtained. ; Step 4.2.4: Correct the inverse displacement transformation and update the coordinates; Corrected displacement in the carrier coordinate system Corrected displacement after inverse transformation back to navigation coordinate system And based on this, the position in the geodetic coordinate system is cumulatively updated; Step 4.2.5: Output the positioning results.

Images