Rail transit heterogeneous network seamless roaming method and system based on multi-level switching

Abstract

The invention provides a rail transit heterogeneous network seamless roaming method and system based on multi-level switching, and relates to the technical field of rail transit network communication, and the method comprises the steps: constructing a space-time transfer matrix to predict a motion track, carrying out the level division of a candidate access network, and calculating an optimal switching path; and calculating a resource reservation time length and reserving resources in advance in combination with the prediction trajectory and the network coverage range, and setting a sampling point to determine a switching trigger point to complete network switching. According to the invention, smooth transition between network hierarchies can be realized, the seamless roaming success rate is improved, and switching delay and service interruption are reduced.

Description

Technical Field

[0001] This invention relates to the field of rail transit network communication technology, and in particular to a seamless roaming method and system for heterogeneous rail transit networks based on multi-level switching. Background Technology

[0002] With the rapid development of urban rail transit networks, passengers' demand for wireless communication services within rail transit systems is increasing daily. Wireless communication networks in rail transit environments typically consist of various heterogeneous networks, with significant differences in coverage, bandwidth, and latency between different types. In the high-speed moving environment of rail transit, mobile terminals need to frequently switch between different heterogeneous networks to maintain communication continuity and service quality.

[0003] Seamless roaming technology in the heterogeneous network environment of rail transit aims to achieve smooth switching of mobile terminals between different networks, ensuring the continuity and stability of communication services. Traditional network switching methods are mainly based on the received signal strength index. When the signal strength is lower than a preset threshold, network switching is triggered. However, these methods lack the ability to accurately predict the movement trajectory of mobile terminals and cannot detect potential changes in network coverage in advance, resulting in delayed switching decisions and increasing the risk of switching failure and service interruption. Furthermore, the use of a single-level network selection strategy fails to effectively utilize the multi-level characteristics of heterogeneous networks, making it difficult to find the optimal network switching path in the complex and ever-changing rail transit environment. Additionally, the failure to effectively reserve target network resources before switching leads to resource contention and other problems during the switching process. Summary of the Invention

[0004] This invention provides a seamless roaming method and system for heterogeneous rail transit networks based on multi-level switching, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of this invention provides a seamless roaming method for heterogeneous rail transit networks based on multi-level switching, comprising:

[0006] The historical position and speed of the mobile terminal in different track sections are obtained and a spatiotemporal transition matrix is ​​constructed. The spatiotemporal transition matrix is ​​recursively evolved by the Markov state prediction algorithm, and the predicted motion trajectory is calculated by combining the path characteristics of the track line.

[0007] The system acquires real-time network status information and divides each candidate access network into levels. It defines the state transition cost between candidate access networks at different levels, and uses a dynamic programming algorithm to calculate the optimal path from the current network level to each candidate network level. This yields a multi-level switching strategy and determines the target network.

[0008] The predicted motion trajectory is spatially intersected with the coverage area of ​​each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. A multivariate Gaussian distribution is constructed based on the distribution parameters, and the statistical upper bound is extracted to obtain the resource reservation duration. The resource reservation trigger time is calculated based on the resource reservation duration and the expected access time, and a resource reservation instruction is sent to the target network.

[0009] Based on the resource reservation instruction, sampling points are set on the predicted motion trajectory. The trend of the received signal strength at each sampling point is extracted and combined with causal convolution to determine the handover tendency score. The sampling point with the highest handover tendency score is used as the handover trigger point. When the mobile terminal reaches the handover trigger point, a handover request is sent and the network handover is completed according to the multi-level handover strategy.

[0010] In one alternative implementation,

[0011] The historical position and speed of the mobile terminal in different track segments are obtained and a spatiotemporal transition matrix is ​​constructed. The spatiotemporal transition matrix is ​​recursively evolved using a Markov state prediction algorithm. Combined with the path characteristics of the track line, the predicted motion trajectory is calculated, including:

[0012] The historical location of the mobile terminal is discretized into a spatial grid. The historical velocity in each track segment is statistically clustered to obtain a velocity state set. A joint state space is constructed based on the spatial grid and the velocity state set. The transition frequency of the joint state between adjacent time steps is calculated and normalized to obtain the spatiotemporal transition matrix.

[0013] The current position and velocity of the mobile terminal are mapped to the joint state space to obtain an initial state vector. The spatiotemporal transition matrix is ​​recursively evolved in multiple steps using a Markov state prediction algorithm to obtain the state probability distribution for the next moment. The joint state with the highest probability in the state probability distribution is extracted and decoded into a position prediction value and a velocity prediction value. A preliminary trajectory is constructed based on the position prediction value and the velocity prediction value.

[0014] The curvature and slope information of the track are acquired and integrated into path features. The path features are interpolated and sampled along the initial trajectory to obtain local path constraints. Based on the local path constraints, the predicted position values ​​in the initial trajectory are geometrically corrected. The deviation energy functional between the initial trajectory and the local path constraints is minimized by a variational optimization algorithm, and the gradient flow of the energy functional is calculated. The coordinates of each predicted position point are iteratively updated along the gradient flow direction until the energy functional converges, thus obtaining the predicted motion trajectory.

[0015] In one alternative implementation,

[0016] Based on the local path constraints, the predicted position values ​​in the preliminary trajectory are geometrically corrected. The deviation energy functional between the preliminary trajectory and the local path constraints is minimized using a variational optimization algorithm, and the gradient flow of the energy functional is calculated, including:

[0017] The local path constraints are discretized into constraint vectors, the position prediction values ​​in the preliminary trajectory are discretized into trajectory vectors, the normal distance from each position point in the trajectory vector to the constraint surface defined by the corresponding constraint vector is calculated and a deviation field is constructed, the deviation field is mapped by a kernel function to obtain a high-dimensional deviation tensor, and the high-dimensional deviation tensor is subjected to tensor contraction operation and integrated to obtain the deviation energy functional.

[0018] The curvature tensor of the initial trajectory is constructed, and the trace of the curvature tensor is calculated as a trajectory curvature metric. A smoothness penalty term is constructed based on the trajectory curvature metric. The deviation energy functional and the smoothness penalty term are expanded into a series representation of basis functions in the function space and subjected to tensor product operation to obtain a coupled tensor. The coupled tensor is subjected to higher-order singular value decomposition to extract a multimodal core tensor and a factor matrix. The multimodal core tensor and the factor matrix are reconstructed into a total energy functional through multilinear transformation. The partial derivatives of the total energy functional with respect to the coordinate components of each position point in the trajectory vector are calculated, and the partial derivatives are assembled into an energy gradient vector. The gradient vector is subjected to gradient projection on the Riemannian manifold to obtain the tangent space gradient, which is used as the gradient flow of the energy functional.

[0019] In one alternative implementation,

[0020] Real-time network state information is acquired and candidate access networks are hierarchically divided. The state transition cost between candidate access networks at different levels is defined. A dynamic programming algorithm is used to calculate the optimal path from the current network level to each candidate network level, resulting in a multi-level switching strategy and determining the target networks, including:

[0021] The carrying capacity and latency indicators of each candidate access network are obtained as real-time network status information. Based on the real-time network status information, a status feature vector is constructed and cluster analysis is performed to obtain the network performance level. Based on the network performance level, the candidate access networks are hierarchically divided and a hierarchical topology is established.

[0022] Candidate access network pairs between adjacent layers in the hierarchical topology are extracted and handover cost indices are calculated. The handover cost indices are constructed into a multidimensional cost vector. The covariance matrix of the multidimensional cost vector is calculated and eigenvalue decomposition is performed to obtain the eigenvalue spectrum. Based on the eigenvalue spectrum, the dominant cost dimension is determined and the corresponding eigenvector is extracted to construct a projection matrix. Based on the projection matrix, the multidimensional cost vector is mapped to a low-dimensional cost space and the Euclidean distance of each candidate access network pair is calculated to obtain the state transition cost.

[0023] The current network level is taken as the initial state, the candidate access network level is taken as the target state, and a cost matrix is ​​constructed based on the state transition cost. The cost matrix is ​​recursively calculated using a dynamic programming algorithm to determine the cumulative cost from the initial state to each target state via intermediate levels and to determine the optimal path. The candidate access network identifiers in the optimal path are recorded and combined with the corresponding state transition sequence to obtain a multi-level handover strategy. The candidate access network corresponding to the next hop level is extracted from the multi-level handover strategy to obtain the target network.

[0024] In one alternative implementation,

[0025] The predicted motion trajectory is spatially intersected with the coverage area of ​​each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the statistical upper bound is extracted to obtain the resource reservation duration, including:

[0026] The predicted motion trajectory is represented as a parameterized curve, and an adaptive sampling is performed based on the rate of curvature change of the parameterized curve to obtain a trajectory point sequence. The coverage of each candidate access network is obtained, and a spatial boundary polyhedron is constructed using a convex hull algorithm. A ray projection intersection operation is performed on the trajectory point sequence and the spatial boundary polyhedron to obtain a set of crossing points. The entry point and the exit point are extracted from the set of crossing points. The trajectory parameter values ​​corresponding to the entry point are calculated, and the expected access time is obtained through parameter inversion. The arc length integral and trajectory velocity vector between the entry point and the exit point are calculated. The expected service duration is calculated based on the magnitude of the arc length integral and the trajectory velocity vector.

[0027] The expected service duration is mapped to the latent space through multi-layer nonlinear transformation to obtain the probabilistic encoding of latent variables. A reparameterization technique is applied to the probabilistic encoding to generate a latent variable sample set. The sample mean and sample covariance in the latent variable sample set are calculated to obtain the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the Mahalanobis distance metric defined by the multivariate Gaussian distribution is calculated. The chi-square quantile corresponding to the preset confidence level is solved in the space corresponding to the Mahalanobis distance metric. The maximum projection value of the Mahalanobis distance isosurface corresponding to the chi-square quantile in the duration dimension is determined to obtain the statistical upper bound of the resource reservation duration. The statistical upper bound is used as the resource reservation duration and recorded.

[0028] In one alternative implementation,

[0029] Calculating the resource reservation trigger time based on the resource reservation duration and the expected access time, and sending the resource reservation instruction to the target network includes:

[0030] The network-side reserved response delay of the target network is obtained. The initial trigger time is calculated based on the expected access time and the network-side reserved response delay. The channel propagation delay between the current terminal and the target network is obtained and a delay jitter distribution is constructed. The delay jitter distribution is subjected to tail probability estimation to obtain the delay buffer amount. The resource reserved trigger time is calculated based on the initial trigger time and the delay buffer amount.

[0031] At the resource reservation trigger time, a resource reservation request message is constructed, and the resource reservation duration and expected access time are encapsulated into the time parameter field of the resource reservation request message. The bandwidth requirement value and latency constraint value of the terminal are obtained and encapsulated into the quality parameter field of the resource reservation request message. The resource reservation request message is serialized and encoded to obtain a resource reservation instruction and sent to the target network.

[0032] In one alternative implementation,

[0033] Based on the resource reservation instruction, sampling points are set on the predicted motion trajectory. The trend of the received signal strength at each sampling point is extracted and combined with causal convolution to determine the handover tendency score. The sampling point with the highest handover tendency score is used as the handover trigger point. When the mobile terminal reaches the handover trigger point, a handover request is sent and network handover is completed according to the multi-level handover strategy, including:

[0034] Based on the resource reservation instruction, a set of sampling points is obtained by dividing the predicted motion trajectory into equal time intervals. The sampling points in the set are spatially located, and the received signal strength measurements of the target network and the current network at each sampling point are obtained. The received signal strength measurements are arranged in chronological order to construct a received signal strength time series sequence, and a first-order difference operation is performed to obtain a signal strength change gradient sequence. The temporal dependency relationship corresponding to the signal strength change gradient sequence is extracted by a causal convolutional neural network to obtain a causal feature tensor. The time dimension is weighted and aggregated to obtain the switching tendency score of each sampling point. The sampling point with the highest switching tendency score is used as the switching trigger point.

[0035] The mobile terminal is tracked in real time and its geographic coordinates are obtained at fixed intervals. The geodesic distance between the geographic coordinates and the handover trigger point is calculated. When the geodesic distance is less than a preset trigger threshold, the mobile terminal is determined to have reached the handover trigger point. The target network corresponding to the next hop level is extracted from the multi-level handover strategy and a handover request message is constructed. The handover request message is sent to the target network and a handover confirmation response is received. The resource allocation result is parsed from the handover confirmation response and the protocol stack reconfiguration from the physical layer to the network layer is executed to complete the network handover.

[0036] A second aspect of this invention provides a seamless roaming system for heterogeneous rail transit networks based on multi-level switching, comprising:

[0037] The trajectory prediction module is used to obtain the historical position and historical speed of the mobile terminal in different track segments and construct a spatiotemporal transition matrix. The spatiotemporal transition matrix is ​​recursively evolved through the Markov state prediction algorithm, and the predicted motion trajectory is calculated by combining the path characteristics of the track line.

[0038] The strategy planning module is used to acquire real-time network status information and divide each candidate access network into levels, define the state transition cost between different levels of candidate access networks, and calculate the optimal path from the current network level to each candidate network level by combining dynamic programming algorithm, so as to obtain multi-level switching strategy and determine the target network.

[0039] The resource reservation module is used to perform spatial intersection calculation between the predicted motion trajectory and the coverage of each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the statistical upper bound is extracted to obtain the resource reservation duration. The resource reservation trigger time is calculated according to the resource reservation duration and the expected access time, and a resource reservation instruction is sent to the target network.

[0040] The handover execution module is used to set sampling points on the predicted motion trajectory based on the resource reservation instruction, extract the trend of the received signal strength of each sampling point and determine the handover tendency score by combining causal convolution, take the sampling point with the highest handover tendency score as the handover trigger point, and send a handover request when the mobile terminal reaches the handover trigger point and complete the network handover according to the multi-level handover strategy.

[0041] A third aspect of the present invention provides an electronic device, comprising:

[0042] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0043] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0044] In this invention, by constructing a spatiotemporal transition matrix and combining it with a Markov state prediction algorithm, the motion trajectory of the mobile terminal can be accurately predicted, providing a reliable spatial location basis for subsequent network handover and improving prediction accuracy. A multi-level handover strategy is determined by combining hierarchical partitioning and dynamic programming, avoiding the limitations of single-network handover and making network handover more flexible and efficient, reducing the handover failure rate. By performing spatial intersection calculations between the predicted motion trajectory and the network coverage area, and extracting statistical upper bounds based on multivariate Gaussian distribution to determine resource reservation time, precise reservation of network resources is achieved, avoiding resource waste and ensuring service quality during handover. Setting sampling points on the predicted trajectory and combining causal convolution to determine the handover tendency score enables network handover to be triggered at the optimal time, effectively reducing handover latency and service interruption time, and achieving seamless roaming between heterogeneous networks in a rail transit environment. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the seamless roaming method for heterogeneous rail transit networks based on multi-level switching, according to an embodiment of the present invention.

[0046] Figure 2 This is a flowchart illustrating the network resource reservation and trajectory prediction process of the seamless roaming method for heterogeneous rail transit networks based on multi-level switching, as described in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0049] Figure 1 This is a flowchart illustrating the seamless roaming method for heterogeneous rail transit networks based on multi-level switching, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0050] The historical position and speed of the mobile terminal in different track sections are obtained and a spatiotemporal transition matrix is ​​constructed. The spatiotemporal transition matrix is ​​recursively evolved by the Markov state prediction algorithm, and the predicted motion trajectory is calculated by combining the path characteristics of the track line.

[0051] The system acquires real-time network status information and divides each candidate access network into levels. It defines the state transition cost between candidate access networks at different levels, and uses a dynamic programming algorithm to calculate the optimal path from the current network level to each candidate network level. This yields a multi-level switching strategy and determines the target network.

[0052] The predicted motion trajectory is spatially intersected with the coverage area of ​​each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. A multivariate Gaussian distribution is constructed based on the distribution parameters, and the statistical upper bound is extracted to obtain the resource reservation duration. The resource reservation trigger time is calculated based on the resource reservation duration and the expected access time, and a resource reservation instruction is sent to the target network.

[0053] Based on the resource reservation instruction, sampling points are set on the predicted motion trajectory. The trend of the received signal strength at each sampling point is extracted and combined with causal convolution to determine the handover tendency score. The sampling point with the highest handover tendency score is used as the handover trigger point. When the mobile terminal reaches the handover trigger point, a handover request is sent and the network handover is completed according to the multi-level handover strategy.

[0054] In one alternative implementation,

[0055] The historical position and speed of the mobile terminal in different track segments are obtained and a spatiotemporal transition matrix is ​​constructed. The spatiotemporal transition matrix is ​​recursively evolved using a Markov state prediction algorithm. Combined with the path characteristics of the track line, the predicted motion trajectory is calculated, including:

[0056] The historical location of the mobile terminal is discretized into a spatial grid. The historical velocity in each track segment is statistically clustered to obtain a velocity state set. A joint state space is constructed based on the spatial grid and the velocity state set. The transition frequency of the joint state between adjacent time steps is calculated and normalized to obtain the spatiotemporal transition matrix.

[0057] The current position and velocity of the mobile terminal are mapped to the joint state space to obtain an initial state vector. The spatiotemporal transition matrix is ​​recursively evolved in multiple steps using a Markov state prediction algorithm to obtain the state probability distribution for the next moment. The joint state with the highest probability in the state probability distribution is extracted and decoded into a position prediction value and a velocity prediction value. A preliminary trajectory is constructed based on the position prediction value and the velocity prediction value.

[0058] The curvature and slope information of the track are acquired and integrated into path features. The path features are interpolated and sampled along the initial trajectory to obtain local path constraints. Based on the local path constraints, the predicted position values ​​in the initial trajectory are geometrically corrected. The deviation energy functional between the initial trajectory and the local path constraints is minimized by a variational optimization algorithm, and the gradient flow of the energy functional is calculated. The coordinates of each predicted position point are iteratively updated along the gradient flow direction until the energy functional converges, thus obtaining the predicted motion trajectory.

[0059] The historical location data of the mobile terminal is discretized into a spatial grid, and the track line is divided into several grid units, each with a size of 20 meters × 20 meters. For example, for a certain track line, the historical trajectory data of the mobile terminal over the past 30 minutes can be obtained, including timestamps, latitude and longitude coordinates, and speed information. The aforementioned historical location points are mapped to a predefined spatial grid according to their geographic coordinates, and each location point is assigned a unique grid index. Within each track segment, the historical speed values ​​of the mobile terminal are collected, and the speed values ​​are divided into five typical states using the K-means clustering algorithm: stationary state (0 to 5 km / h), low speed state (5 to 20 km / h), medium speed state (20 to 40 km / h), high speed state (40 to 80 km / h), and ultra-high speed state (above 80 km / h). Combining the spatial grid index and the speed state index, a joint state space is constructed. For example, if a mobile terminal is located in grid (25, 37) and is in a medium speed state, its joint state can be represented as (25, 37, 3).

[0060] The frequency of joint state transitions between two adjacent time steps is statistically analyzed, with a time step size of 10 seconds. Historical data is used to observe the frequency with which a mobile terminal transitions from one state to another 10 seconds later. By normalizing each set of data, the probability of transitioning from one state to any other state is calculated, resulting in a complete spatiotemporal transition matrix.

[0061] For the trajectory prediction stage, the current location coordinates (116.432, 39.915) and speed (25 km / h) of the mobile terminal are obtained and mapped to a state vector in the joint state space, assuming the mapping result is state (42, 18, 3). Based on the pre-constructed spatiotemporal transition matrix, a Markov state prediction algorithm is used to perform multi-step recursive evolution. The state distribution for the next 6 time steps (i.e., the next 60 seconds) is predicted. For each predicted time step, the joint state with the highest probability value in the state probability distribution is extracted, such as (43, 19, 3), and decoded into the predicted location value (116.436, 39.918) and the predicted speed value (26 km / h). All predicted location points are connected to construct a preliminary trajectory path.

[0062] The curvature and gradient information of the track line are acquired and integrated as path features. Curvature information describes the degree of curvature of the track, expressed as the radius of curvature per 100 meters. For example, the curvature of a straight section is 0, the curvature of a large curve is 0.005, and the curvature of a small curve is 0.02. Gradient information describes the degree of inclination of the track, expressed as a percentage. For example, the gradient of a flat section is 0%, the gradient of an uphill section is 2%, and the gradient of a downhill section is -3%. Interpolation sampling is performed along the preliminary predicted trajectory, with a sampling point set every 5 meters. The curvature and gradient values ​​corresponding to each sampling point are obtained to form a local path constraint set.

[0063] Geometric correction is performed on the initial trajectory based on local path constraints. A deviation energy functional is defined, which measures the degree of inconsistency between the predicted trajectory and the local path constraints. The deviation energy consists of the sum of squared distances between the trajectory point and the orbital centerline, and the sum of squared differences between the trajectory curvature and the orbital curvature. The gradient direction of the energy functional is calculated, representing the direction of the fastest energy decrease. The initial number of iterations is 0, the maximum number of iterations is 100, and the convergence threshold is 0.0001. In each iteration, the gradient of the current trajectory point relative to the deviation energy is calculated, and the trajectory point position is updated in the opposite direction of the gradient with an update step size of 0.05. The deviation energy is recalculated after each update, and iteration stops when the energy change is less than the convergence threshold or the maximum number of iterations is reached. Convergence is typically achieved after 20 to 30 iterations, yielding a predicted trajectory that satisfies the orbital geometric constraints.

[0064] In this embodiment, by constructing a joint state space from historical position and velocity features and performing Markov multi-step prediction based on the spatiotemporal transition matrix, the temporal evolution of the mobile terminal in the track scenario can be more accurately characterized, improving the stability and foresight of position and velocity prediction. By introducing track curvature and slope to form path constraints and performing geometric correction on the initial trajectory, the prediction process can better conform to the actual structural characteristics of the track line, improving the spatial rationality of trajectory points. By employing a variational optimization algorithm to optimize the trajectory for global energy convergence, the trajectory deviation can be reduced overall and the trajectory continuity enhanced, improving the consistency and reliability between the final predicted trajectory and the actual trajectory.

[0065] In one alternative implementation,

[0066] Based on the local path constraints, the predicted position values ​​in the preliminary trajectory are geometrically corrected. The deviation energy functional between the preliminary trajectory and the local path constraints is minimized using a variational optimization algorithm, and the gradient flow of the energy functional is calculated, including:

[0067] The local path constraints are discretized into constraint vectors, the position prediction values ​​in the preliminary trajectory are discretized into trajectory vectors, the normal distance from each position point in the trajectory vector to the constraint surface defined by the corresponding constraint vector is calculated and a deviation field is constructed, the deviation field is mapped by a kernel function to obtain a high-dimensional deviation tensor, and the high-dimensional deviation tensor is subjected to tensor contraction operation and integrated to obtain the deviation energy functional.

[0068] The curvature tensor of the initial trajectory is constructed, and the trace of the curvature tensor is calculated as a trajectory curvature metric. A smoothness penalty term is constructed based on the trajectory curvature metric. The deviation energy functional and the smoothness penalty term are expanded into a series representation of basis functions in the function space and subjected to tensor product operation to obtain a coupled tensor. The coupled tensor is subjected to higher-order singular value decomposition to extract a multimodal core tensor and a factor matrix. The multimodal core tensor and the factor matrix are reconstructed into a total energy functional through multilinear transformation. The partial derivatives of the total energy functional with respect to the coordinate components of each position point in the trajectory vector are calculated, and the partial derivatives are assembled into an energy gradient vector. The gradient vector is subjected to gradient projection on the Riemannian manifold to obtain the tangent space gradient, which is used as the gradient flow of the energy functional.

[0069] The local path constraints are discretized into constraint vectors. A sampling point is set every 2 meters along the track centerline. Each sampling point contains position coordinates, a normal vector, and curvature information. For example, at a track curve, 10 adjacent sampling points are collected to form a constraint vector set. Each constraint point contains three-dimensional spatial position coordinates (x, y, z), a normal vector (nx, ny, nz), and a curvature value k. The predicted position values ​​in the initial trajectory are discretized into trajectory vectors, maintaining the same sampling interval as the constraint vectors. For example, a mobile terminal predicts the trajectory and samples 30 trajectory points at 2-meter intervals over the next 60 seconds, forming the trajectory vector. For each position point in the trajectory vector, the normal distance to the constraint surface defined by the corresponding constraint vector is calculated. The constraint surface is determined by the constraint point and its corresponding normal vector. The distance from the trajectory point to the constraint surface is calculated using the point-to-plane distance formula. For example, the calculated normal distance from the trajectory point (115.872, 39.764, 42.5) to the constraint surface is 0.75 meters. The normal distance values ​​of all trajectory points are combined to construct a deviation field.

[0070] A kernel function mapping is performed on the deviation field, using a Gaussian radial basis function with a bandwidth parameter set to 0.8. This kernel function mapping transforms the deviation field from three-dimensional space to a high-dimensional feature space, resulting in a high-dimensional deviation tensor. This high-dimensional deviation tensor can be represented as a fourth-order tensor with dimensions 30×30×30×3, where 30 represents the number of trajectory points and 3 represents the spatial dimension. Tensor contraction is then performed on the high-dimensional deviation tensor, shrinking it to scalar values. Integrating the contraction result yields the deviation energy functional, representing the overall degree of deviation between the trajectory and the constraints. An example calculation result of 12.6 is provided.

[0071] The curvature tensor of the initial trajectory is constructed by calculating the second derivative information. The first and second derivatives of the position points in the trajectory vector are calculated using the central difference method, and assembled into a three-dimensional curvature tensor with dimensions of 30×3×3, where 30 represents the number of trajectory points and the 3×3 part represents the curvature information at each point. The trace of the curvature tensor, i.e., the sum of its diagonal elements, is calculated as a measure of trajectory curvature. For example, a curvature tensor trace of 0.024 at a certain point indicates a relatively high degree of trajectory curvature at that point. A smoothness penalty term is constructed based on the trajectory curvature measure, with a penalty coefficient set to 0.5, so that excessively curved trajectory segments will be subject to a higher penalty value.

[0072] The bias energy functional and the smoothness penalty term are represented as a series of basis functions in the function space. B-spline basis functions are chosen as the basis for the expansion, with an order of 3 and a node spacing of 2 meters. The bias energy functional expands into a linear combination of 60 basis functions, with a coefficient vector 2-norm of 8.3; the smoothness penalty term is also expanded into a linear combination of 60 basis functions, with a coefficient vector 2-norm of 3.7. The two terms are combined using a tensor product operation to obtain a coupled tensor, with a dimension of 60×60, representing the mutual influence relationships between the energy terms.

[0073] A higher-order singular value decomposition (SVD) was performed on the coupled tensor, with a truncation threshold of 0.01, to extract the multimodal core tensor and factor matrix. The core tensor, with a dimension of 15×15, captured 95% of the variability of the original coupled tensor; the factor matrices, with dimensions of 60×15 and 60×15 respectively, represented the mapping relationship from the original basis functions to the core tensor. A multilinear transformation was used to reconstruct the core tensor and factor matrix into a total energy functional, which comprehensively considers both the degree of fit of the trajectory to the constraints and the smoothness of the trajectory itself.

[0074] Taking the partial derivatives of the total energy functional with respect to the coordinate components of the trajectory vector at each position point yields the rate of energy change for each trajectory point in the x, y, and z directions. For example, the energy gradient at a trajectory point (116.423, 39.891, 45.2) is (0.42, -0.35, 0.12), indicating that moving the trajectory point along this direction will result in the fastest increase in total energy. The energy gradients of all trajectory points are assembled into an energy gradient vector with dimensions 30×3, representing the energy change trend of the entire trajectory.

[0075] The constraints formed by the trajectory typically constitute a curvilinear manifold. The energy gradient vector is then projected onto the Riemannian manifold. Using the Gram-Schmidt orthogonalization method, the energy gradient is projected onto the tangent space of the constraint surface, yielding the tangent space gradient. This tangent space gradient ensures that the trajectory optimization process always follows the constraint surface, preventing the optimization result from deviating from the trajectory constraints. The tangent space gradient is then used as the gradient flow of the energy functional to guide the iterative update direction of the trajectory points.

[0076] In this embodiment, by discretizing the local path constraints and the initial trajectory into vectors and calculating the normal distance to construct the deviation field, the degree of deviation of the trajectory from the real path structure can be quantified at a finer granularity, improving the accuracy and sensitivity of deviation measurement. By mapping the deviation field with a kernel function and generating a high-dimensional deviation tensor, the ability to express nonlinear deviation modes can be enhanced, improving the robustness of deviation analysis in complex path scenarios. By constructing a curvature tensor and introducing a smoothness penalty term, local oscillations and irregular bends in the initial trajectory can be effectively suppressed, improving the smoothness and structural rationality of the predicted trajectory. By unifying the deviation energy and smoothness term under the function space basis and performing tensor product operations, and then using high-order singular value decomposition to extract the core tensor structure, efficient decoupling and optimization between multi-modal constraints can be achieved, improving the stability and controllability of energy expression under multi-constraint fusion.

[0077] In one alternative implementation,

[0078] Real-time network state information is acquired and candidate access networks are hierarchically divided. The state transition cost between candidate access networks at different levels is defined. A dynamic programming algorithm is used to calculate the optimal path from the current network level to each candidate network level, resulting in a multi-level switching strategy and determining the target networks, including:

[0079] The carrying capacity and latency indicators of each candidate access network are obtained as real-time network status information. Based on the real-time network status information, a status feature vector is constructed and cluster analysis is performed to obtain the network performance level. Based on the network performance level, the candidate access networks are hierarchically divided and a hierarchical topology is established.

[0080] Candidate access network pairs between adjacent layers in the hierarchical topology are extracted and handover cost indices are calculated. The handover cost indices are constructed into a multidimensional cost vector. The covariance matrix of the multidimensional cost vector is calculated and eigenvalue decomposition is performed to obtain the eigenvalue spectrum. Based on the eigenvalue spectrum, the dominant cost dimension is determined and the corresponding eigenvector is extracted to construct a projection matrix. Based on the projection matrix, the multidimensional cost vector is mapped to a low-dimensional cost space and the Euclidean distance of each candidate access network pair is calculated to obtain the state transition cost.

[0081] The current network level is taken as the initial state, the candidate access network level is taken as the target state, and a cost matrix is ​​constructed based on the state transition cost. The cost matrix is ​​recursively calculated using a dynamic programming algorithm to determine the cumulative cost from the initial state to each target state via intermediate levels and to determine the optimal path. The candidate access network identifiers in the optimal path are recorded and combined with the corresponding state transition sequence to obtain a multi-level handover strategy. The candidate access network corresponding to the next hop level is extracted from the multi-level handover strategy to obtain the target network.

[0082] The network monitoring module acquires the carrying capacity and latency indicators of each candidate access network as real-time network status information. For heterogeneous networks in the rail transit environment, the monitorable candidate access networks include cellular networks, trackside private networks, station high-speed local area networks, and onboard wireless networks. Real-time status information of each network is collected, including five key indicators: downlink bandwidth, uplink bandwidth, network latency, packet loss rate, and signal strength. For example, at a certain moment, the status data of four candidate networks were collected as follows: cellular network: downlink bandwidth 15.6Mbps, uplink bandwidth 5.2Mbps, network latency 45ms, packet loss rate 0.8%, signal strength -75dBm; trackside private network: downlink bandwidth 48.3Mbps, uplink bandwidth 25.1Mbps, network latency 12ms, packet loss rate 0.2%, signal strength -62dBm; station high-speed LAN: downlink bandwidth 152.6Mbps, uplink bandwidth 78.5Mbps, network latency 5ms, packet loss rate 0.1%, signal strength -48dBm; vehicle-mounted wireless network: downlink bandwidth 25.8Mbps, uplink bandwidth 15.3Mbps, network latency 28ms, packet loss rate 0.5%, signal strength -68dBm.

[0083] Based on the aforementioned real-time network state information, a state feature vector is constructed, combining the five indicators of each candidate network into a five-dimensional vector. Each indicator is normalized to a value range of 0 to 1, making the normalized state feature vector easier for subsequent cluster analysis. The K-means clustering algorithm is used to cluster the normalized state feature vectors, with K set to 3, corresponding to high-performance, medium-performance, and low-performance networks, respectively. The cluster centers are for high-performance (0.85, 0.83, 0.90, 0.92, 0.88), medium-performance (0.62, 0.58, 0.65, 0.70, 0.60), and low-performance (0.35, 0.30, 0.40, 0.45, 0.38) levels. The performance level is determined by calculating the Euclidean distance between the state feature vector of each candidate network and the cluster center. In the aforementioned example, cellular networks are classified as low-performance networks, trackside private networks and vehicle-mounted wireless networks are classified as medium-performance networks, and high-speed local area networks at stations are classified as high-performance networks.

[0084] Candidate access networks are hierarchically divided based on network performance levels to establish a hierarchical topology. High-performance networks are located at the top layer (Layer 1), medium-performance networks at the middle layer (Layer 2), and low-performance networks at the bottom layer (Layer 3). In rail transit scenarios, station areas typically contain multiple high-performance networks, while sections mainly cover medium- and low-performance networks. In the hierarchical topology, networks at the same level are horizontally connected, and networks at different levels are vertically connected, forming a complete topology. For example, a station area may simultaneously cover a high-performance station high-speed LAN (Layer 1), a medium-performance trackside private network (Layer 2), and a low-performance cellular network (Layer 3), constituting a three-layer network architecture.

[0085] Candidate access network pairs between adjacent layers in the hierarchical topology are extracted, and handover cost metrics are calculated. In the aforementioned example, possible network pairs include station high-speed LAN and trackside private network, station high-speed LAN and vehicle-mounted wireless network, trackside private network and cellular network, vehicle-mounted wireless network and cellular network, etc. The handover cost metric consists of four dimensions: handover latency, service interruption time, authentication overhead, and handover failure rate. For example, the cost metrics for switching from station high-speed LAN to trackside private network are: handover latency 85ms, service interruption time 120ms, authentication overhead 0.6, and handover failure rate 0.5%; the cost metrics for switching from trackside private network to cellular network are: handover latency 150ms, service interruption time 220ms, authentication overhead 0.8, and handover failure rate 1.2%. These handover cost metrics are constructed into a four-dimensional cost vector, in the form of (85, 120, 0.6, 0.5) and (150, 220, 0.8, 1.2).

[0086] The covariance matrix of the multidimensional cost vectors is calculated, forming a cost matrix from all cost vectors, with each row representing the switching cost of a network pair. The covariance matrix of the cost matrix reflects the correlation between different cost dimensions. Eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvalue spectrum. In this embodiment, the eigenvalues ​​are 2.35, 1.28, 0.31, and 0.06, indicating that the first two eigenvalues ​​are dominant, explaining more than 90% of the variance. Based on the eigenvalue spectrum, the dominant cost dimension is determined, and the corresponding eigenvectors are extracted to construct a projection matrix. The projection matrix has a 4×2 dimension, mapping the original four-dimensional cost vectors to a two-dimensional cost space, reducing computational complexity.

[0087] The multidimensional cost vector is mapped to a low-dimensional cost space using a projection matrix, and the state transition cost is obtained by calculating the Euclidean distance between each candidate access network pair. For example, the state transition cost from the station's high-speed LAN to the trackside private network is 1.25, and the state transition cost from the trackside private network to the cellular network is 2.38. The state transition cost reflects the comprehensive cost of switching between different networks.

[0088] Using the current network layer as the initial state and the candidate access network layers as the target state, a cost matrix is ​​constructed based on the state transition costs. Assuming the mobile terminal is currently connected to a trackside private network (Layer 2), possible candidate access networks include the station high-speed local area network (Layer 1) and the cellular network (Layer 3). Cost matrices are constructed from Layer 2 to Layer 1 and Layer 3, with matrix elements representing the state transition costs of corresponding network pairs. The cost matrix is ​​recursively calculated using a dynamic programming algorithm to determine the cumulative cost from the initial state through intermediate layers to each target state, and the optimal path is determined. During train entry into the station, the optimal path to switch from the trackside private network to the station high-speed local area network is a direct switch, with a cumulative cost of 1.25. During train exit from the station, the optimal path to switch from the station high-speed local area network to the cellular network is to first switch to the trackside private network and then to the cellular network, with a cumulative cost of 3.63, lower than the cost of a direct switch (4.52).

[0089] Record the candidate access network identifiers in the optimal path and combine them with the corresponding state transition sequences to obtain a multi-level handover strategy. For example, the multi-level handover strategy when a train leaves the station is "station high-speed LAN → trackside private network → cellular network", and the corresponding state transition sequence is "layer 1 → layer 2 → layer 3". Extract the candidate access network corresponding to the next hop layer from the multi-level handover strategy to obtain the target network.

[0090] In this embodiment, by constructing features and performing cluster analysis on the carrying capacity and latency information of candidate access networks, the complex and ever-changing network states can be abstracted into clear performance levels, improving the accuracy of network state identification and the rationality of hierarchical organization. By constructing a hierarchical topology based on performance levels, a structured hierarchical relationship can be formed in network handover decisions, improving the manageability and search efficiency of the candidate network selection process. By calculating the handover costs of adjacent hierarchical network pairs and constructing a multi-dimensional cost vector, multiple key factors can be considered simultaneously in the handover cost evaluation, improving the comprehensiveness of cost characterization. By extracting the dominant cost dimension and constructing a projection matrix through covariance analysis and eigenvalue decomposition, dimensionality reduction can be achieved while maintaining the main cost features, improving the stability and discriminativeness of cost calculation.

[0091] In one alternative implementation,

[0092] The predicted motion trajectory is spatially intersected with the coverage area of ​​each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the statistical upper bound is extracted to obtain the resource reservation duration, including:

[0093] The predicted motion trajectory is represented as a parameterized curve, and an adaptive sampling is performed based on the rate of curvature change of the parameterized curve to obtain a trajectory point sequence. The coverage of each candidate access network is obtained, and a spatial boundary polyhedron is constructed using a convex hull algorithm. A ray projection intersection operation is performed on the trajectory point sequence and the spatial boundary polyhedron to obtain a set of crossing points. The entry point and the exit point are extracted from the set of crossing points. The trajectory parameter values ​​corresponding to the entry point are calculated, and the expected access time is obtained through parameter inversion. The arc length integral and trajectory velocity vector between the entry point and the exit point are calculated. The expected service duration is calculated based on the magnitude of the arc length integral and the trajectory velocity vector.

[0094] The expected service duration is mapped to the latent space through multi-layer nonlinear transformation to obtain the probabilistic encoding of latent variables. A reparameterization technique is applied to the probabilistic encoding to generate a latent variable sample set. The sample mean and sample covariance in the latent variable sample set are calculated to obtain the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the Mahalanobis distance metric defined by the multivariate Gaussian distribution is calculated. The chi-square quantile corresponding to the preset confidence level is solved in the space corresponding to the Mahalanobis distance metric. The maximum projection value of the Mahalanobis distance isosurface corresponding to the chi-square quantile in the duration dimension is determined to obtain the statistical upper bound of the resource reservation duration. The statistical upper bound is used as the resource reservation duration and recorded.

[0095] Cubic spline interpolation is used to represent the predicted motion trajectory as a parametric curve. For rail transit lines, the set of key points on the predicted trajectory is taken as control points to construct the parametric curve, with parameters ranging from 0 to 1. A parameter of 0 corresponds to the starting point, a parameter of 1 corresponds to the ending point, and intermediate values ​​correspond to interpolation points on the curve. The rate of curvature change of the parametric curve is calculated, and dense sampling is performed in areas with large curvature changes, while sparse sampling is performed in areas with small curvature changes, achieving adaptive sampling. The sampling threshold is set to a rate of curvature change of 0.01 per meter. When the rate of curvature change is greater than this threshold, the sampling interval is 5 meters; when the rate of curvature change is less than this threshold, the sampling interval is 20 meters, resulting in a trajectory point sequence. Each trajectory point contains its position coordinates and corresponding parameter values. For example, in a certain rail transit section, the generated trajectory point sequence contains 87 sampling points, with a sampling interval of approximately 20 meters in straight sections and a sampling interval reduced to 5 meters in curved sections.

[0096] The coverage area of ​​each candidate access network is obtained by collecting the location coordinates and coverage radius of the network access points. For example, there are 5 access points along the rail transit line with location coordinates of (120.32, 30.45, 5.2), (120.36, 30.47, 5.8), (120.41, 30.49, 6.5), (120.46, 30.52, 7.0), and (120.51, 30.54, 7.2), corresponding to coverage radii of 200 meters, 250 meters, 180 meters, 300 meters, and 220 meters, respectively. Based on the locations and coverage radii of the access points, a spatial boundary polyhedron is constructed using a convex hull algorithm. A sphere is constructed centered on the location corresponding to each access point, and the surface of the sphere is discretized into 32 vertices. The vertex set generated by all access points is input into the 3D convex hull algorithm to calculate the boundary polyhedron. In the example above, the generated spatial boundary polyhedron contains 28 faces and 56 vertices.

[0097] Ray projection intersection operations are performed on the trajectory point sequence and the spatial boundary polyhedron to determine the intersection points between the trajectory and the network coverage area. For each adjacent pair of points in the trajectory point sequence, a ray is constructed, and the intersection of the ray with each face of the spatial boundary polyhedron is determined to identify valid intersection points. The intersection point calculation uses a method of simultaneously solving the parametric line equation and the plane equation, and the calculation results include the intersection point coordinates and the corresponding parameter values. All intersection points are grouped into a set of crossing points and sorted in ascending order of parameter values. In the above case, there are 8 intersection points between the trajectory and the network coverage area, with parameter values ​​of 0.23, 0.31, 0.45, 0.52, 0.67, 0.74, 0.86, and 0.93.

[0098] Entry and exit points are extracted from the set of crossing points using an alternating labeling method. When a trajectory enters the network coverage area from outside, the intersection point is marked as an entry point; when a trajectory leaves the network coverage area from inside, the intersection point is marked as an exit point. For the aforementioned eight intersection points, the parameter values ​​for entry points are 0.23, 0.45, 0.67, and 0.86, and the parameter values ​​for exit points are 0.31, 0.52, 0.74, and 0.93, forming four pairs of entry-exit points. The trajectory parameter values ​​corresponding to the entry points are calculated, and the expected access time is obtained through parameter inversion. Assuming the current time is 8:00:00, and there is a linear mapping relationship between trajectory parameters and time (parameter 0 corresponds to 8:00:00, parameter 1 corresponds to 8:15:00), the expected access times corresponding to the four entry points are 8:03:27, 8:06:45, 8:10:03, and 8:12:54, respectively.

[0099] Calculate the arc length integral between the entry and exit points, i.e., the length of the curve segment. For each pair of entry-exit points, the parameter interval is evenly divided into 100 sub-intervals. The arc length infinitesimal is calculated for each sub-interval, and then summed to obtain the arc length integral. For example, the arc length integral corresponding to the first pair of entry-exit points is 327.5 meters, the second pair is 295.8 meters, the third pair is 312.4 meters, and the fourth pair is 284.6 meters. Calculate the velocity vector of the trajectory at the entry point using the parametric derivative of the curve. For example, the velocity vector at the parameter value of 0.23 is (42.3, 15.7, 0.5) meters per second, with a velocity magnitude of 45.2 meters per second. Calculate the expected service duration based on the arc length integral and the magnitude of the trajectory velocity vector, i.e., the time required for the trajectory to traverse the network coverage area. For example, the expected service duration for the first pair of entry-exit points is 7.25 seconds, and for the other three pairs are 6.42 seconds, 6.86 seconds, and 6.15 seconds, respectively.

[0100] The expected service duration is mapped to the latent space through multi-layer nonlinear transformation to obtain the probabilistic encoding of the latent variables. A variational autoencoder architecture is adopted, with the encoder containing three fully connected neural networks, with 64, 32, and 16 hidden layer nodes respectively, and ReLU activation function. The input is the expected service duration, and the output is a mean vector and a log-variance vector, both with a dimension of 8, which constitute the probabilistic encoding of the latent variables. For example, for a service duration of 7.25 seconds, the generated mean vector is (-0.25, 0.37, 0.12, -0.44, 0.56, -0.18, 0.29, -0.31), and the log-variance vector is (-1.28, -0.95, -1.32, -1.05, -0.89, -1.15, -1.02, -1.21).

[0101] A reparameterization technique is applied to the probabilistic encoding to generate a latent variable sample set. A noise vector with dimension 8 is sampled from a standard normal distribution and transformed to calculate latent variable samples. This process is repeated 100 times to obtain a latent variable sample set containing 100 samples. The sample mean and sample covariance in the latent variable sample set are calculated to obtain the distribution parameters of the latent space. The sample mean vector is (-0.22, 0.35, 0.10, -0.42, 0.53, -0.16, 0.27, -0.28), and the sample covariance matrix is ​​an 8×8 symmetric matrix with diagonal elements of (0.28, 0.39, 0.26, 0.35, 0.41, 0.32, 0.37, 0.30).

[0102] A multivariate Gaussian distribution is constructed based on the distribution parameters, and the Mahalanobis distance metric defined by the multivariate Gaussian distribution is calculated. The chi-square quantile corresponding to a pre-set confidence level is then calculated in the space corresponding to the Mahalanobis distance metric. With a confidence level of 95%, the corresponding chi-square quantile is 15.51, with 8 degrees of freedom. The maximum projection value of the Mahalanobis distance isosurface corresponding to the chi-square quantile onto the duration dimension is determined, yielding the statistical upper bound of the resource reservation duration. By sampling the latent space and mapping it back to the duration space, the calculated statistical upper bound is 1.35 times the original expected service duration. For example, for an expected service duration of 7.25 seconds, the resource reservation duration is 9.79 seconds; for the other three pairs of entry-departure points, the resource reservation durations are 8.67 seconds, 9.26 seconds, and 8.30 seconds, respectively. This statistical upper bound is used as the resource reservation duration and recorded for subsequent network resource scheduling and handover decisions.

[0103] In this embodiment, by representing the predicted motion trajectory as a parameterized curve and adaptively sampling based on the rate of curvature change, the trajectory point sequence can be obtained in a more refined and structure-sensitive manner, improving the adaptability and accuracy of trajectory sampling to the complexity of motion paths. By constructing a spatial boundary polyhedron of the candidate access network coverage area and intersecting it with the trajectory point sequence by ray projection, the entry and exit points of the trajectory crossing different network coverage areas can be accurately identified, improving the accuracy of access time and service interval identification. By performing parameter inversion on the entry point and combining arc length integral and trajectory velocity vector to calculate the expected service duration, the expected dwell time of the terminal in the target network can be obtained in a geometrically accurate manner, improving the reliability of service duration prediction. By mapping the service duration to the latent space through multi-layer nonlinear transformation and combining reparameterization techniques to generate a latent variable sample set, the potential distribution structure of the duration can be captured while maintaining the statistical characteristics of the duration, improving the expressive power of uncertainty modeling.

[0104] Figure 2 This is a flowchart illustrating the network resource reservation and trajectory prediction process of the seamless roaming method for heterogeneous rail transit networks based on multi-level switching, as described in an embodiment of the present invention.

[0105] In one alternative implementation,

[0106] Calculating the resource reservation trigger time based on the resource reservation duration and the expected access time, and sending the resource reservation instruction to the target network includes:

[0107] The network-side reserved response delay of the target network is obtained. The initial trigger time is calculated based on the expected access time and the network-side reserved response delay. The channel propagation delay between the current terminal and the target network is obtained and a delay jitter distribution is constructed. The delay jitter distribution is subjected to tail probability estimation to obtain the delay buffer amount. The resource reserved trigger time is calculated based on the initial trigger time and the delay buffer amount.

[0108] At the resource reservation trigger time, a resource reservation request message is constructed, and the resource reservation duration and expected access time are encapsulated into the time parameter field of the resource reservation request message. The bandwidth requirement value and latency constraint value of the terminal are obtained and encapsulated into the quality parameter field of the resource reservation request message. The resource reservation request message is serialized and encoded to obtain a resource reservation instruction and sent to the target network.

[0109] Obtain the network-side reserved response latency of the target network. Periodically probe the control plane signaling processing capability of the target network using network measurement tools, and calculate the average time interval from receiving a resource reservation request to completing resource allocation. In the heterogeneous network environment of rail transit, the network-side reserved response latency varies significantly depending on the access technology. For example, the reserved response latency of microcell networks is approximately 15 milliseconds, while that of macrocell networks is approximately 35 milliseconds, and that of satellite communication networks can reach 200 milliseconds. Calculate the initial trigger time based on the expected access time and the network-side reserved response latency, and subtract the network-side reserved response latency from the expected access time. Taking a rail transit train passing through a station as an example, with an expected access time of 8:03:27, a target network being a microcell network, and a reserved response latency of 15 milliseconds, the initial trigger time is 8:03:26.985.

[0110] The channel propagation delay between the current terminal and the target network is obtained by measuring the round-trip time using measurement messages. In high-speed mobile scenarios, the channel propagation delay fluctuates with changes in the terminal's location. Therefore, 100 measurement messages are continuously sent, and the round-trip time is recorded. Messages that exceed the time limit are discarded, and the mean and standard deviation of the delay for the effective samples are calculated. In the rail transit scenario, the measured mean channel propagation delay is 8.3 milliseconds, and the standard deviation is 2.1 milliseconds. The delay jitter distribution usually approximates a Gaussian distribution, but a long-tail effect may occur in real network environments. Therefore, a generalized extreme value distribution is used for fitting to accurately characterize the delay fluctuation characteristics. For the fitted generalized extreme value distribution, the tail probability is estimated using extreme value theory, and the delay threshold corresponding to exceeding 99.9% confidence is calculated to obtain the delay buffer. In the aforementioned example, the delay buffer corresponding to 99.9% confidence is 16.8 milliseconds. The resource reservation trigger time is calculated based on the initial trigger time and the delay buffer, by subtracting the delay buffer from the initial trigger time. For the example above, the resource reservation trigger time is 8:03:26.9682.

[0111] A resource reservation request message is constructed at the resource reservation trigger moment. This message uses a lightweight binary format, with a 12-byte header including: message type (1 byte, value 0x01 indicating a resource reservation request); message version (1 byte, current version 0x03); session identifier (4 bytes, uniquely identifying the current resource reservation session); terminal identifier (4 bytes, a unique device identifier); and message length (2 bytes, indicating the message body length). The resource reservation duration and expected access time are encapsulated in the time parameter field of the resource reservation request message. The time parameter field contains 16 bytes, where the expected access time is represented by an 8-byte UNIX timestamp with millisecond precision; the resource reservation duration is represented by a 4-byte unsigned integer in milliseconds; and 4 bytes are reserved for future expansion. For the scenario of a rail transit train passing through a station, the expected access time is 8:03:27, the resource reservation time is 9790 milliseconds, and the encapsulated time parameter field is {0x01, 0x98, 0x76, 0x25, 0xFE, 0x39, 0x48, 0x00, 0x00, 0x00, 0x26, 0x3E, 0x00, 0x00, 0x00, 0x00}.

[0112] Obtain the terminal's bandwidth requirement and latency constraint. The bandwidth requirement is dynamically calculated based on the type and number of applications currently running on the terminal. For example, video streaming applications require 5Mbps, VoIP calls require 64Kbps, and web browsing requires 1Mbps. The latency constraint is determined based on the application's interactivity requirements. For example, VoIP calls require end-to-end latency of no more than 150 milliseconds, and online games require end-to-end latency of no more than 50 milliseconds. Encapsulate the bandwidth requirement and latency constraint into the quality parameter field of the resource reservation request message. The quality parameter field contains 12 bytes, where the bandwidth requirement is represented by 4 bytes in Kbps; the latency constraint is represented by 2 bytes in milliseconds; the upper limit of packet loss rate is represented by 2 bytes with a precision of 0.01%; the service priority is represented by 1 byte, ranging from 0 to 255, with higher values ​​indicating higher priority; and 3 bytes are reserved for future expansion. In a high-speed rail transit scenario, assuming the terminal is simultaneously running video conferencing and data synchronization applications, the total bandwidth requirement is 6144Kbps, the latency constraint is 100 milliseconds, the maximum packet loss rate is 0.1%, the service priority is 192, and the encapsulated quality parameter fields are {0x00, 0x00, 0x18, 0x00, 0x00, 0x64, 0x00, 0x0A, 0xC0, 0x00, 0x00, 0x00}.

[0113] Serializing and encoding resource reservation request messages to obtain resource reservation instructions is a necessary step for network communication. The serialization process uses TLV encoding to ensure message compatibility and efficient transmission in heterogeneous network environments. First, the total length of the message header and body is calculated, and a buffer of corresponding size is created. Then, each field is written to the buffer according to a predefined byte order. Finally, the CRC32 checksum of the message is calculated and appended to the end of the message to form a complete resource reservation instruction. The complete resource reservation instruction is 44 bytes long, including a 12-byte header, 16 bytes of time parameters, 12 bytes of quality parameters, and a 4-byte checksum. The resource reservation instruction is sent to the target network through the control channel, using reliable transmission protocols to ensure transmission reliability, such as TCP at the transport layer and automatic retransmission requests at the link layer. By setting a message priority flag, the resource reservation instruction is given priority processing during network congestion, further improving the timeliness of resource reservation.

[0114] After receiving the resource reservation instruction, the target network will parse the resource reservation request message, perform resource planning and allocation based on time and quality parameters, and send a resource reservation response message to the terminal.

[0115] In this embodiment, by calculating the initial trigger time based on the expected access time and the reserved response delay of the target network, the trigger window for resource reservation actions can be planned in advance, improving the foresight of resource reservation scheduling. By obtaining the channel propagation delay between the terminal and the target network and constructing the delay jitter distribution, and then estimating the tail probability to obtain the delay buffer, the uncertainty caused by random link jitter can be effectively absorbed in the trigger time calculation, improving the stability and risk protection capability of the trigger time setting. After calculating the resource reservation trigger time based on the initial trigger time and the delay buffer, the reliable triggering of resource reservation actions can still be maintained even when the network state fluctuates, improving the timeliness of the reservation process. By constructing a resource reservation request message at the resource reservation trigger time and encapsulating parameters such as resource reservation duration, expected access time, bandwidth requirements, and delay constraints into complete time and quality parameter fields, a clear and consistent description of resource requirements can be provided to the target network, improving the accuracy of the resource allocation process.

[0116] In one alternative implementation,

[0117] Based on the resource reservation instruction, sampling points are set on the predicted motion trajectory. The trend of the received signal strength at each sampling point is extracted and combined with causal convolution to determine the handover tendency score. The sampling point with the highest handover tendency score is used as the handover trigger point. When the mobile terminal reaches the handover trigger point, a handover request is sent and network handover is completed according to the multi-level handover strategy, including:

[0118] Based on the resource reservation instruction, a set of sampling points is obtained by dividing the predicted motion trajectory into equal time intervals. The sampling points in the set are spatially located, and the received signal strength measurements of the target network and the current network at each sampling point are obtained. The received signal strength measurements are arranged in chronological order to construct a received signal strength time series sequence, and a first-order difference operation is performed to obtain a signal strength change gradient sequence. The temporal dependency relationship corresponding to the signal strength change gradient sequence is extracted by a causal convolutional neural network to obtain a causal feature tensor. The time dimension is weighted and aggregated to obtain the switching tendency score of each sampling point. The sampling point with the highest switching tendency score is used as the switching trigger point.

[0119] The mobile terminal is tracked in real time and its geographic coordinates are obtained at fixed intervals. The geodesic distance between the geographic coordinates and the handover trigger point is calculated. When the geodesic distance is less than a preset trigger threshold, the mobile terminal is determined to have reached the handover trigger point. The target network corresponding to the next hop level is extracted from the multi-level handover strategy and a handover request message is constructed. The handover request message is sent to the target network and a handover confirmation response is received. The resource allocation result is parsed from the handover confirmation response and the protocol stack reconfiguration from the physical layer to the network layer is executed to complete the network handover.

[0120] Based on resource reservation instructions, a set of sampling points is obtained by dividing the predicted motion trajectory into equal time intervals. Sampling points are generated sequentially on the predicted motion trajectory at 500 millisecond intervals. For example, for the predicted trajectory of a rail transit train from station A to station B, with a total duration of 120 seconds, 240 sampling points are obtained through equal time interval division, denoted as P1 to P240. Each sampling point in the set is spatially located and converted into geographic coordinates. Simultaneously, the received signal strength measurements of the target network and the current network are obtained. Spatial positioning uses trajectory point interpolation. Taking P30 as an example, its geographic coordinates are (120.3645, 30.4823, 6.75). The received signal strength of the current microcell network is -75 dBm, and the received signal strength of the target macrocell network is -92 dBm. The received signal strength measurements are arranged chronologically to construct a received signal strength time series, resulting in time series of length 240 for both the current network and the target network.

[0121] First-order differencing is performed on the time series to obtain the gradient sequence of signal intensity changes, and the signal intensity change between adjacent sampling points is calculated. For the time series of the current network, a gradient sequence of length 239 is obtained after first-order differencing; for the time series of the target network, the gradient sequence is obtained after first-order differencing. Taking P30 and P31 as an example, the signal intensity of the current network changes from -75 dBm to -76 dBm, with a difference of -1 dBm; the signal intensity of the target network changes from -92 dBm to -90 dBm, with a difference of 2 dBm. The temporal dependencies corresponding to the gradient sequence of signal intensity changes are extracted using a causal convolutional neural network to obtain a causal feature tensor. The causal convolutional neural network adopts a multi-layer causal convolutional structure with a kernel size of 3. The input is a concatenation of gradient sequences, and the output is a causal feature tensor with a feature dimension of 64. The network contains four causal convolutional layers, with 16, 32, 48, and 64 kernels in each layer, and the activation function is a rectified linear unit. The dilation rates are 1, 2, 4, and 8, respectively, with an effective receptive field of 31 time steps, capable of capturing signal change patterns within approximately 15 seconds. Residual connections are added to the network to prevent gradient vanishing. After processing by a causal convolutional neural network, a causal feature tensor with dimensions of 239×64 is obtained.

[0122] A time-dimensional weighted aggregation of the causal feature tensor is performed to obtain the switching tendency score for each sampling point. The weighted aggregation employs an attention mechanism to learn the importance weights of different positions in the time dimension. The attention weights are calculated through a fully connected layer, with the causal feature vector as input and scalar weight values ​​as output. The weighted feature vectors are summed and normalized to between 0 and 1 using a sigmoid activation function to obtain the switching tendency score. In rail transit scenarios, higher switching tendency scores typically appear near signal strength intersections. In the aforementioned example, sampling point P120 has the highest switching tendency score, reaching 0.935. At this point, the current network signal strength is -86 dBm, and the target network signal strength is -83 dBm, showing an intersection trend, with signal strength gradients of -0.8 dBm and 1.2 dBm, respectively. Sampling point P120 with the highest switching tendency score is taken as the switching trigger point, with its geographical coordinates (120.4125, 30.5103, 7.15).

[0123] The mobile terminal's location is tracked in real time, and its geographic coordinates are acquired at fixed intervals. Location tracking employs fusion positioning technology, combining satellite positioning, inertial navigation, and trajectory matching, updating the terminal's location every 50 milliseconds. For example, at a certain moment, the mobile terminal's real-time geographic coordinates are (120.4080, 30.5072, 7.05). The geodesic distance between these geographic coordinates and the handover trigger point is calculated using Vincent's formula to calculate the spherical distance, taking into account altitude differences, to obtain a comprehensive distance value. Using the aforementioned coordinates as an example, the calculated geodesic distance is 53.7 meters. A preset trigger threshold of 50 meters is set. When the geodesic distance is less than this threshold, the mobile terminal is determined to have reached the handover trigger point. In the next location update, the terminal coordinates are updated to (120.4082, 30.5075, 7.06), and the calculated geodesic distance is 48.3 meters, which is less than the preset trigger threshold of 50 meters, indicating that the mobile terminal has reached the handover trigger point.

[0124] The target network corresponding to the next hop level is extracted from the multi-level handover strategy. The multi-level handover strategy is stored in a strategy database and organized in the form of a decision tree. In the heterogeneous network environment of rail transit, a typical multi-level handover strategy includes multiple levels such as microcell networks, macrocell networks, and satellite communication networks. Based on the current network level of microcell network and the environmental conditions of high-speed movement and signal crossing, the next hop level is found to be macrocell network from the decision tree. A handover request message is constructed. The message format is a binary structure, containing a message header and a message body. The message header contains 8 bytes, including a 1-byte message type with a value of 0x02 indicating a handover request, a 1-byte message version, a 4-byte session identifier, and a 2-byte message length. The message body contains 28 bytes, including an 8-byte terminal identifier, a 4-byte current network identifier, a 4-byte target network identifier, and 12 bytes of terminal location coordinates, with each dimension consisting of 4 bytes.

[0125] A handover request message is sent to the target network and a handover confirmation response is received. A reliable transmission protocol, such as Transmission Control Protocol (TCP) or Flow Control Protocol (Flow Control Protocol), is used to ensure successful message delivery. In the example above, the handover request message is sent to the macrocell network's access controller within 5 milliseconds and a handover confirmation response is received after 25 milliseconds. The resource allocation result is parsed from the handover confirmation response. The confirmation response uses a binary format similar to the request message, containing an 8-byte header and a 20-byte body. The body includes a 2-byte status code (0x0000 indicating success), an 8-byte allocated radio resource identifier, 4 bytes of IP address configuration information, 4 bytes of quality of service parameters, and 2 bytes of security parameters. The parsed resource allocation result includes: 12 radio resource blocks, a center frequency of 2.6 GHz, a bandwidth of 10 MHz, a modulation scheme of 64 quadrature amplitude modulation (64), an IP address of 192.168.2.45, a guaranteed uplink bandwidth of 2 Mbps, and a guaranteed downlink bandwidth of 5 Mbps.

[0126] The network handover is completed by reconfiguring the protocol stack from the physical layer to the network layer. The reconfiguration process includes several steps: Physical layer reconfiguration includes adjusting the radio frequency parameters, setting the center frequency to 2.6 GHz, and synchronizing the new reference signal with a synchronization time of 3 milliseconds; Media Access Control layer reconfiguration includes updating the radio resource allocation information, setting the number of resource blocks to 12, and adjusting the Hybrid Automatic Repeat Request parameter redundancy version number to 4; Radio Link Control layer reconfiguration includes resetting the state machine and updating the window size, adjusting the transmission window to 32; Packet Data Convergence Protocol layer reconfiguration includes updating the security key and resetting the sequence number from 0; Network layer reconfiguration includes updating the IP address to 192.168.2.45 and adjusting the Quality of Service (QoS) parameter priority to high.

[0127] In this embodiment, by sampling at equal time intervals based on resource reservation instructions on the predicted motion trajectory and obtaining the received signal strength measurements of the target network and the current network at each sampling point, continuous observation of the signal evolution trend can be formed in the time dimension, improving the precision of signal change monitoring. By constructing a signal strength time series sequence and performing first-order difference to obtain a change gradient sequence, the sensitivity to dynamic processes such as signal attenuation and enhancement can be enhanced, improving the responsiveness of handover decisions to signal fluctuations. By extracting the temporal dependency of the signal strength change gradient sequence through a causal convolutional neural network and obtaining a causal feature tensor, the directional influence of signal evolution can be captured, improving the accuracy of handover tendency assessment. The handover tendency score of the sampling points is obtained by weighted aggregation based on the time dimension, and the position with the highest tendency is selected as the handover trigger point. This avoids the problems of premature or delayed handover caused by traditional handover methods based on instantaneous signal strength or fixed thresholds, improving the rationality and stability of trigger point selection.

[0128] A second aspect of this invention provides a seamless roaming system for heterogeneous rail transit networks based on multi-level switching, comprising:

[0129] The trajectory prediction module is used to obtain the historical position and historical speed of the mobile terminal in different track segments and construct a spatiotemporal transition matrix. The spatiotemporal transition matrix is ​​recursively evolved through the Markov state prediction algorithm, and the predicted motion trajectory is calculated by combining the path characteristics of the track line.

[0130] The strategy planning module is used to acquire real-time network status information and divide each candidate access network into levels, define the state transition cost between different levels of candidate access networks, and calculate the optimal path from the current network level to each candidate network level by combining dynamic programming algorithm, so as to obtain multi-level switching strategy and determine the target network.

[0131] The resource reservation module is used to perform spatial intersection calculation between the predicted motion trajectory and the coverage of each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the statistical upper bound is extracted to obtain the resource reservation duration. The resource reservation trigger time is calculated according to the resource reservation duration and the expected access time, and a resource reservation instruction is sent to the target network.

[0132] The handover execution module is used to set sampling points on the predicted motion trajectory based on the resource reservation instruction, extract the trend of the received signal strength of each sampling point and determine the handover tendency score by combining causal convolution, take the sampling point with the highest handover tendency score as the handover trigger point, and send a handover request when the mobile terminal reaches the handover trigger point and complete the network handover according to the multi-level handover strategy.

[0133] A third aspect of the present invention provides an electronic device, comprising:

[0134] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0135] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0136] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A seamless roaming method for heterogeneous rail transit networks based on multi-level switching, characterized in that, include: The historical position and speed of the mobile terminal in different track sections are obtained and a spatiotemporal transition matrix is ​​constructed. The spatiotemporal transition matrix is ​​recursively evolved by the Markov state prediction algorithm, and the predicted motion trajectory is calculated by combining the path characteristics of the track line. The system acquires real-time network status information and divides each candidate access network into levels. It defines the state transition cost between candidate access networks at different levels, and uses a dynamic programming algorithm to calculate the optimal path from the current network level to each candidate network level. This yields a multi-level switching strategy and determines the target network. The predicted motion trajectory is spatially intersected with the coverage area of ​​each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. A multivariate Gaussian distribution is constructed based on the distribution parameters, and the statistical upper bound is extracted to obtain the resource reservation duration. The resource reservation trigger time is calculated based on the resource reservation duration and the expected access time, and a resource reservation instruction is sent to the target network. Based on the resource reservation instruction, sampling points are set on the predicted motion trajectory. The trend of the received signal strength at each sampling point is extracted and combined with causal convolution to determine the handover tendency score. The sampling point with the highest handover tendency score is used as the handover trigger point. When the mobile terminal reaches the handover trigger point, a handover request is sent and the network handover is completed according to the multi-level handover strategy.

2. The method according to claim 1, characterized in that, The historical position and speed of the mobile terminal in different track segments are obtained and a spatiotemporal transition matrix is ​​constructed. The spatiotemporal transition matrix is ​​recursively evolved using a Markov state prediction algorithm. Combined with the path characteristics of the track line, the predicted motion trajectory is calculated, including: The historical location of the mobile terminal is discretized into a spatial grid. The historical velocity in each track segment is statistically clustered to obtain a velocity state set. A joint state space is constructed based on the spatial grid and the velocity state set. The transition frequency of the joint state between adjacent time steps is calculated and normalized to obtain the spatiotemporal transition matrix. The current position and velocity of the mobile terminal are mapped to the joint state space to obtain an initial state vector. The spatiotemporal transition matrix is ​​recursively evolved in multiple steps using a Markov state prediction algorithm to obtain the state probability distribution for the next moment. The joint state with the highest probability in the state probability distribution is extracted and decoded into a position prediction value and a velocity prediction value. A preliminary trajectory is constructed based on the position prediction value and the velocity prediction value. The curvature and slope information of the track are acquired and integrated into path features. The path features are interpolated and sampled along the initial trajectory to obtain local path constraints. Based on the local path constraints, the predicted position values ​​in the initial trajectory are geometrically corrected. The deviation energy functional between the initial trajectory and the local path constraints is minimized by a variational optimization algorithm, and the gradient flow of the energy functional is calculated. The coordinates of each predicted position point are iteratively updated along the gradient flow direction until the energy functional converges, thus obtaining the predicted motion trajectory.

3. The method according to claim 2, characterized in that, Based on the local path constraints, the predicted position values ​​in the preliminary trajectory are geometrically corrected. The deviation energy functional between the preliminary trajectory and the local path constraints is minimized using a variational optimization algorithm, and the gradient flow of the energy functional is calculated, including: The local path constraints are discretized into constraint vectors, the position prediction values ​​in the preliminary trajectory are discretized into trajectory vectors, the normal distance from each position point in the trajectory vector to the constraint surface defined by the corresponding constraint vector is calculated and a deviation field is constructed, the deviation field is mapped by a kernel function to obtain a high-dimensional deviation tensor, and the high-dimensional deviation tensor is subjected to tensor contraction operation and integrated to obtain the deviation energy functional. The curvature tensor of the initial trajectory is constructed, and the trace of the curvature tensor is calculated as a trajectory curvature metric. A smoothness penalty term is constructed based on the trajectory curvature metric. The deviation energy functional and the smoothness penalty term are expanded into a series representation of basis functions in the function space and subjected to tensor product operation to obtain a coupled tensor. The coupled tensor is subjected to higher-order singular value decomposition to extract a multimodal core tensor and a factor matrix. The multimodal core tensor and the factor matrix are reconstructed into a total energy functional through multilinear transformation. The partial derivatives of the total energy functional with respect to the coordinate components of each position point in the trajectory vector are calculated, and the partial derivatives are assembled into an energy gradient vector. The gradient vector is subjected to gradient projection on the Riemannian manifold to obtain the tangent space gradient, which is used as the gradient flow of the energy functional.

4. The method according to claim 1, characterized in that, Real-time network state information is acquired and candidate access networks are hierarchically divided. The state transition cost between candidate access networks at different levels is defined. A dynamic programming algorithm is used to calculate the optimal path from the current network level to each candidate network level, resulting in a multi-level switching strategy and determining the target networks, including: The carrying capacity and latency indicators of each candidate access network are obtained as real-time network status information. Based on the real-time network status information, a status feature vector is constructed and cluster analysis is performed to obtain the network performance level. Based on the network performance level, the candidate access networks are hierarchically divided and a hierarchical topology is established. Candidate access network pairs between adjacent layers in the hierarchical topology are extracted and handover cost indices are calculated. The handover cost indices are constructed into a multidimensional cost vector. The covariance matrix of the multidimensional cost vector is calculated and eigenvalue decomposition is performed to obtain the eigenvalue spectrum. Based on the eigenvalue spectrum, the dominant cost dimension is determined and the corresponding eigenvector is extracted to construct a projection matrix. Based on the projection matrix, the multidimensional cost vector is mapped to a low-dimensional cost space and the Euclidean distance of each candidate access network pair is calculated to obtain the state transition cost. The current network level is taken as the initial state, the candidate access network level is taken as the target state, and a cost matrix is ​​constructed based on the state transition cost. The cost matrix is ​​recursively calculated using a dynamic programming algorithm to determine the cumulative cost from the initial state to each target state via intermediate levels and to determine the optimal path. The candidate access network identifiers in the optimal path are recorded and combined with the corresponding state transition sequence to obtain a multi-level handover strategy. The candidate access network corresponding to the next hop level is extracted from the multi-level handover strategy to obtain the target network.

5. The method according to claim 1, characterized in that, The predicted motion trajectory is spatially intersected with the coverage area of ​​each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the statistical upper bound is extracted to obtain the resource reservation duration, including: The predicted motion trajectory is represented as a parameterized curve, and an adaptive sampling is performed based on the rate of curvature change of the parameterized curve to obtain a trajectory point sequence. The coverage of each candidate access network is obtained, and a spatial boundary polyhedron is constructed using a convex hull algorithm. A ray projection intersection operation is performed on the trajectory point sequence and the spatial boundary polyhedron to obtain a set of crossing points. The entry point and the exit point are extracted from the set of crossing points. The trajectory parameter values ​​corresponding to the entry point are calculated, and the expected access time is obtained through parameter inversion. The arc length integral and trajectory velocity vector between the entry point and the exit point are calculated. The expected service duration is calculated based on the magnitude of the arc length integral and the trajectory velocity vector. The expected service duration is mapped to the latent space through multi-layer nonlinear transformation to obtain the probabilistic encoding of latent variables. A reparameterization technique is applied to the probabilistic encoding to generate a latent variable sample set. The sample mean and sample covariance in the latent variable sample set are calculated to obtain the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the Mahalanobis distance metric defined by the multivariate Gaussian distribution is calculated. The chi-square quantile corresponding to the preset confidence level is solved in the space corresponding to the Mahalanobis distance metric. The maximum projection value of the Mahalanobis distance isosurface corresponding to the chi-square quantile in the duration dimension is determined to obtain the statistical upper bound of the resource reservation duration. The statistical upper bound is used as the resource reservation duration and recorded.

6. The method according to claim 1, characterized in that, Calculating the resource reservation trigger time based on the resource reservation duration and the expected access time, and sending the resource reservation instruction to the target network includes: The network-side reserved response delay of the target network is obtained. The initial trigger time is calculated based on the expected access time and the network-side reserved response delay. The channel propagation delay between the current terminal and the target network is obtained and a delay jitter distribution is constructed. The delay jitter distribution is subjected to tail probability estimation to obtain the delay buffer amount. The resource reserved trigger time is calculated based on the initial trigger time and the delay buffer amount. At the resource reservation trigger time, a resource reservation request message is constructed, and the resource reservation duration and expected access time are encapsulated into the time parameter field of the resource reservation request message. The bandwidth requirement value and latency constraint value of the terminal are obtained and encapsulated into the quality parameter field of the resource reservation request message. The resource reservation request message is serialized and encoded to obtain a resource reservation instruction and sent to the target network.

7. The method according to claim 1, characterized in that, Based on the resource reservation instruction, sampling points are set on the predicted motion trajectory. The trend of the received signal strength at each sampling point is extracted and combined with causal convolution to determine the handover tendency score. The sampling point with the highest handover tendency score is used as the handover trigger point. When the mobile terminal reaches the handover trigger point, a handover request is sent and network handover is completed according to the multi-level handover strategy, including: Based on the resource reservation instruction, a set of sampling points is obtained by dividing the predicted motion trajectory into equal time intervals. The sampling points in the set are spatially located, and the received signal strength measurements of the target network and the current network at each sampling point are obtained. The received signal strength measurements are arranged in chronological order to construct a received signal strength time series sequence, and a first-order difference operation is performed to obtain a signal strength change gradient sequence. The temporal dependency relationship corresponding to the signal strength change gradient sequence is extracted by a causal convolutional neural network to obtain a causal feature tensor. The time dimension is weighted and aggregated to obtain the switching tendency score of each sampling point. The sampling point with the highest switching tendency score is used as the switching trigger point. The mobile terminal is tracked in real time and its geographic coordinates are obtained at fixed intervals. The geodesic distance between the geographic coordinates and the handover trigger point is calculated. When the geodesic distance is less than a preset trigger threshold, the mobile terminal is determined to have reached the handover trigger point. The target network corresponding to the next hop level is extracted from the multi-level handover strategy and a handover request message is constructed. The handover request message is sent to the target network and a handover confirmation response is received. The resource allocation result is parsed from the handover confirmation response and the protocol stack reconfiguration from the physical layer to the network layer is executed to complete the network handover.

8. A seamless roaming system for heterogeneous rail transit networks based on multi-level switching, used to implement the method of any one of claims 1-7, characterized in that, include: The trajectory prediction module is used to obtain the historical position and historical speed of the mobile terminal in different track segments and construct a spatiotemporal transition matrix. The spatiotemporal transition matrix is ​​recursively evolved through the Markov state prediction algorithm, and the predicted motion trajectory is calculated by combining the path characteristics of the track line. The strategy planning module is used to acquire real-time network status information and divide each candidate access network into levels, define the state transition cost between different levels of candidate access networks, and calculate the optimal path from the current network level to each candidate network level by combining dynamic programming algorithm, so as to obtain multi-level switching strategy and determine the target network. The resource reservation module is used to perform spatial intersection calculation between the predicted motion trajectory and the coverage of each candidate access network to obtain the expected access time and expected service duration. The expected service duration is encoded and mapped to the distribution parameters of the latent space. Based on the distribution parameters, a multivariate Gaussian distribution is constructed and the statistical upper bound is extracted to obtain the resource reservation duration. The resource reservation trigger time is calculated according to the resource reservation duration and the expected access time, and a resource reservation instruction is sent to the target network. The handover execution module is used to set sampling points on the predicted motion trajectory based on the resource reservation instruction, extract the trend of the received signal strength of each sampling point and determine the handover tendency score by combining causal convolution, take the sampling point with the highest handover tendency score as the handover trigger point, and send a handover request when the mobile terminal reaches the handover trigger point and complete the network handover according to the multi-level handover strategy.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images