Terminal trajectory generation method, apparatus, medium, and product

Abstract

This application provides a terminal trajectory generation method, apparatus, medium, and product, relating to the field of mobile communication technology. The method includes: determining mapping data between terminal signaling and location; using a spatiotemporal encoder to acquire time and location features from the mapping data; and inputting the time and location features into a preset attention mechanism model to determine the actual location of the terminal. The solution of this application reduces dependence on a single base station signal by deeply associating terminal signaling with location mapping data, thus overcoming the dependence on base station positioning accuracy; it utilizes a spatiotemporal encoder and attention mechanism model to achieve multimodal mobility modeling; and based on signaling and attention mechanisms, it strengthens the modeling of semantic and base station interaction relationships, improving the accuracy of trajectory production or trajectory reconstruction.

Description

Technical Field

[0001] This application relates to the field of mobile communication technology, specifically to a terminal trajectory generation method, apparatus, medium, and product. Background Technology

[0002] With the development of mobile communication technology, terminal trajectory restoration technology based on base station signaling data plays an important role in urban traffic planning, travel behavior analysis and other fields. Existing technologies mainly improve trajectory accuracy through signaling data clustering, road network matching and rule correction, but still face the following bottlenecks: (1) Inherent limitations of base station positioning accuracy: Traditional methods such as density clustering and stopping point removal in existing technologies can suppress the ping-pong effect, but rely on the geometric division of the base station coverage area. Due to the uneven distribution of base station density, rule correction is difficult to overcome physical positioning errors, resulting in trajectory breakage or drift in complex urban areas or sparse base station areas. (2) Lack of processing of multimodal mobility modes: One related technology extracts a single travel path through hierarchical clustering and path similarity calculation, but lacks the ability to model mixed transportation modes, such as walking-subway-shared bicycle combination travel. The spatiotemporal characteristics of different modes of transportation are significantly different, and traditional clustering algorithms cannot adaptively distinguish the dynamic switching of mobility modes, resulting in the accumulation of path fusion errors. (3) Insufficient utilization of semantic information and network topology: Although another related technology introduces road network matching, it does not deeply explore the topological relationships between base stations and the surrounding semantic environment. For example, the terminal migration patterns between base stations in dense commercial areas are significantly different from those in industrial areas, and the other related technology does not incorporate such information into the trajectory reasoning model.

[0003] Therefore, existing technologies suffer from limitations in trajectory reconstruction due to insufficient positioning accuracy relying on base stations, lack of multi-mode mobility modeling, and neglect of semantic interaction with base stations. Summary of the Invention

[0004] At least one embodiment of this application provides a terminal trajectory generation method, apparatus, medium, and product to address the problems in the prior art, such as insufficient reliance on base station positioning accuracy, lack of multi-mode mobility modeling, and neglect of semantic interaction with base stations, which result in limited trajectory restoration performance.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows:

[0006] In a first aspect, embodiments of this application provide a terminal trajectory generation method, including:

[0007] Determining the mapping data between terminal signaling and location includes: acquiring terminal communication signaling data and actual trajectory data; performing time sorting and outlier removal operations on the communication signaling data to generate a signaling sequence; based on the actual location in the signaling sequence and the actual trajectory data, determining the mapping data with a mapping relationship between signaling and location according to time alignment, and filtering out abnormal data;

[0008] Using a spatiotemporal encoder, the temporal and positional features in the mapped data are obtained;

[0009] The time features and location features are input into a preset attention mechanism model to determine the actual location of the terminal.

[0010] Optionally, a spatiotemporal encoder is used to obtain the temporal features in the mapped data, including:

[0011] By using the time branch in the spatiotemporal encoder, the timestamp information in the mapping data is normalized to determine the entry time parameter, exit time parameter, and duration parameter.

[0012] For the entry time parameter, the exit time parameter, and the duration parameter, corresponding time feature vectors are generated through trainable linear layers, respectively.

[0013] The periodicity of the time component is captured by using the cosine function to determine the periodic characteristics;

[0014] The corresponding time feature vector and the periodic feature are fused to obtain the time feature in the mapping data.

[0015] Optionally, a spatiotemporal encoder is used to obtain the positional features in the mapping data, including:

[0016] Using the multilayer perceptron in the spatiotemporal encoder, the base station locations and relative positional relationships between different base stations in the mapping data are obtained;

[0017] The location of the base station and the relative positional relationship are comprehensively encoded and represented to obtain the positional features in the mapping data.

[0018] Optionally, before inputting the time features and the location features into a preset attention mechanism model to determine the actual location of the terminal, the method further includes:

[0019] Obtain the base station adjacency relationship and determine the global spatiotemporal representation based on the base station adjacency relationship;

[0020] For the movement trajectory of a single terminal, a local transfer subgraph is constructed, and personalized movement pattern features are extracted based on the local transfer subgraph using a hierarchical attention mechanism;

[0021] Extract the road network topology features and point-of-interest distribution features within the preset range of the base station;

[0022] Based on the road network topology features and the interest point distribution features, a geographic semantic representation is generated through a spatial attention network.

[0023] The global spatiotemporal representation, the personalized mobility pattern features, and the geographic semantic representation are gated and fused to output comprehensive base station information containing multi-dimensional interaction relationships.

[0024] The initial attention mechanism model is trained using the integrated base station information to generate the attention mechanism model.

[0025] Optionally, the base station adjacency relationship is obtained, and a global spatiotemporal representation is determined based on the base station adjacency relationship, including:

[0026] Obtain the spatial location of all base stations within the target area, as well as the terminal transfer mode between different base stations;

[0027] The spatial location of the base station is used as a node, and the terminal transfer mode is used as an edge to determine the adjacency relationship of the base station;

[0028] Based on the aforementioned base station adjacency relationships, a global transfer graph of the base stations is constructed;

[0029] Based on the global transition graph, the global spatiotemporal representation is determined.

[0030] Optionally, the initial attention mechanism model is trained using the integrated base station information to generate the attention mechanism model, including:

[0031] Obtain sample time features and sample location features generated from samples;

[0032] Based on the integrated base station information, the sample time features, and the sample location features, a multi-head self-attention mechanism is used to generate multi-time-step interactive representations.

[0033] A multi-task loss function is constructed; the multi-task loss function includes a geographic coordinate reconstruction error term, a positive sample representation similarity term, and a negative sample comparison term, and the model parameters are optimized through self-supervised learning;

[0034] The attention mechanism model is constructed based on the interaction representation and the multi-task loss function.

[0035] Secondly, embodiments of this application provide a terminal trajectory generation device, comprising:

[0036] A first determining module is used to determine the mapping data between the signaling and location of the terminal; wherein, the first determining module includes: a first acquiring unit, used to acquire the terminal's communication signaling data and actual trajectory data; a first processing unit, used to perform time sorting and outlier removal operations on the communication signaling data to generate a signaling sequence; and a second processing unit, used to determine the mapping data with a mapping relationship between the signaling and location based on the actual location in the signaling sequence and the actual trajectory data, according to time alignment, and to filter out outlier data;

[0037] The first processing module is used to obtain the time features and location features in the mapping data using a spatiotemporal encoder;

[0038] The second processing module is used to input the time features and the location features into a preset attention mechanism model to determine the actual location of the terminal.

[0039] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0040] Fourthly, embodiments of this application provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0041] Compared with existing technologies, the terminal trajectory generation method, apparatus, medium, and product provided in this application combine the signaling and location of the determined terminal to generate mapping data between the two. A spatiotemporal encoder extracts the temporal and spatial features from the mapping data, reducing reliance on base station signals. Compared to traditional base station air interface data dependence, this method utilizes road network topology constraints to optimize the trajectory, reducing error accumulation caused by uneven base station coverage and improving positioning accuracy in complex environments. The spatiotemporal encoder acquires the temporal and location features from the mapping data, and time-segmented dynamic model selection is fused with the spatiotemporal features to solve the problem that a single model cannot adapt to complex movement states in traditional methods. Furthermore, an attention mechanism is used to input the temporal and location features into a preset attention mechanism model to determine the actual location of the terminal, improving the semantic relevance of trajectory reconstruction. Attached Figure Description

[0042] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0043] Figure 1 A flowchart illustrating the terminal trajectory generation method provided in this application embodiment;

[0044] Figure 2 A schematic diagram illustrating the model training process provided in this application embodiment;

[0045] Figure 3 This is a schematic diagram of the terminal trajectory generation device provided in the embodiments of this application. Detailed Implementation

[0046] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, without limiting the number of objects; for example, the first object can be one or more. Furthermore, "or" in this application indicates at least one of the connected objects. For example, "A or B" covers three scenarios: Scenario 1: including A but not B; Scenario 2: including B but not A; Scenario 3: including both A and B. The character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0047] The term "instruction" in this application can be either a direct instruction (or explicit instruction) or an indirect instruction (or implicit instruction). A direct instruction can be understood as one in which the sender explicitly informs the receiver of specific information, the operation to be performed, or the requested result, etc.; an indirect instruction can be understood as one in which the receiver determines the corresponding information based on the instruction sent by the sender, or makes a judgment and determines the operation to be performed or the requested result, etc., based on the judgment result.

[0048] As described in the background section, existing technologies for improving trajectory longitude often rely solely on rule-based methods to process signaling data, failing to overcome the limitations of base station-based positioning accuracy, which itself has low precision. Furthermore, existing technologies do not adequately consider complex terminal mobility patterns, lacking universal applicability and often only applicable to specific travel modes, rather than providing unified processing for data from multiple travel modes. Moreover, existing technologies merely utilize signaling data itself and supplementary information (such as road network information) to improve data quality, often neglecting semantic information surrounding the base station that contributes to trajectory accuracy improvement, as well as relevant information such as the interaction relationships between base stations. To address at least one of these issues, this application provides a terminal trajectory generation method that can reduce or avoid these problems.

[0049] Please refer to Figure 1This application provides a terminal trajectory generation method, comprising:

[0050] Step 11: Determine the mapping data between the terminal's signaling and location.

[0051] In this embodiment, the core of this step is to establish a mapping relationship between the terminal's signaling interactions with the base station (such as access events and handover records) and the base station's geographical location information. Optionally, data such as the terminal's communication event timestamps, base station identifiers (such as cell IDs), and unique terminal identifiers are obtained to collect the terminal's signaling. Combined with base station ledgers or geographical location data, the base station identifiers are mapped to specific geographical coordinates (latitude and longitude) and road network information. The signaling data is then associated with the road network information through the terminal identifiers to generate a raw dataset containing time, base station location, and the terminal's movement path.

[0052] For example, suppose a terminal accesses base station A at time t1 (location coordinates (x1, y1)) and then switches to base station B at time t2 (location coordinates (x2, y2)). By associating the event time in the signaling data with the base station location, an initial trajectory point sequence can be generated: [(t1, x1, y1), (t2, x2, y2), ...].

[0053] Optionally, determine the mapping data between the terminal's signaling and location, including:

[0054] Acquire the terminal's communication signaling data and real trajectory data;

[0055] The communication signaling data is sorted by time and anomalies are removed to generate a signaling sequence;

[0056] Based on the actual location in the signaling sequence and the real trajectory data, the mapping data that has a mapping relationship between the signaling and the location is determined according to the time alignment method, and abnormal data is filtered out.

[0057] In this embodiment, the communication signaling data includes event records of terminal-base station interactions (such as access, handover, disconnection, etc.), including but not limited to timestamps, base station identifiers (such as cell ID, LAC), and terminal identifiers (such as IMSI). Real trajectory data can be obtained through high-precision positioning technologies (such as GPS, BeiDou) or trajectory points reported by the user, including timestamps and geographic coordinates (latitude and longitude). The purpose of acquiring the terminal's communication signaling data and real trajectory data is to establish the correlation between signaling events and physical locations, providing foundational data for subsequent feature extraction and anomaly filtering.

[0058] For example, suppose a terminal interacts with a base station at the following times: t1=10:00:00, accessing base station A (LAC=12345, Cell=67890, coordinates (116.4, 39.9)); t2=10:05:00, switching to base station B (LAC=23456, Cell=89012, coordinates (116.5, 39.8)). The actual trajectory data recorded by GPS is: t1=10:00:00, location (116.41, 39.92); t2=10:05:00, location (116.52, 39.78). The two datasets are merged by timestamp to form the initial dataset.

[0059] Furthermore, the communication signaling data is sorted by time and anomaly removal operations are performed. Time sorting, such as sorting signaling times by timestamp, ensures temporal continuity. Anomaly removal operations include, but are not limited to: removing records with duplicate timestamps or excessively small intervals; filtering points where base station coordinates do not match the road network; and removing outliers through clustering or rate thresholds.

[0060] In this application, signaling times are matched with actual trajectory points according to a time window, and candidate association pairs are generated using time alignment. The determination of the mapping relationship includes direct matching and difference matching. Direct matching is used when the signaling time and the trajectory point time coincide; difference matching is used when, if there is no precise match, the signaling position is estimated through the difference between the trajectory points, and then association pairs where the signaling and trajectory points are too far apart or have contradictory directions are eliminated.

[0061] Screening out abnormal mapping data can be done using multi-dimensional verification, including: checking whether the mapping location is within the road network and applying road network constraints; calculating the position and direction changes between adjacent mapping points and eliminating abrupt changes (such as 90° sharp turns); and combining base station signal strength (RSRP) or MR (Measurement Report) data to filter out low-quality signaling.

[0062] The steps in this application combine multi-dimensional information such as signaling, trajectory, and road network to achieve multi-source data fusion and improve mapping accuracy. Through time sorting, cluster analysis, and road network constraints, dynamic anomaly filtering is implemented to effectively suppress signal drift. Interpolation matching and motion continuity verification reduce time window errors, achieving high-precision alignment suitable for high-speed movement scenarios. The refined data cleaning and spatiotemporal alignment mechanisms significantly improve the reliability of signaling and location mapping, providing high-quality input for subsequent trajectory generation.

[0063] In one specific implementation, step 11 of this application can be understood as a data preprocessing step. Communication signaling data and actual travel trajectory data of the same batch of terminals are collected, which can be obtained through mobile operators and related applications (APPs). A mapping pair from actual signaling trajectory to actual travel location is formed, and sample pairs are filtered and abnormal data is removed. Communication signaling data of terminals on mobile devices is collected, abnormal data is removed, and the remaining data is arranged in chronological order to form the original signaling sequence data. Actual location data of the same terminals within a time period are collected and arranged by time. Based on the time period corresponding to the terminal, a mapping from the communication signaling sequence to the terminal's actual trajectory sequence is constructed. Furthermore, based on the time period format, sample pairs are divided into a structured format, and abnormal sample pairs are removed. Thus, a mapping from the communication signaling sequence to the terminal's actual trajectory sequence is constructed.

[0064] Step 12: Use a spatiotemporal encoder to obtain the time and location features in the mapped data.

[0065] In this application, the spatiotemporal encoder converts the temporal and spatial information in the mapped data into high-dimensional feature vectors to capture spatiotemporal dependencies: it uses sine / cosine functions or learnable location encoding (such as location embedding in Transformer) to map timestamps into periodic features; it can optionally use BeiDou grid codes, geohashing, or Euclidean distance encoding to convert base station coordinates into spatial semantic features.

[0066] Furthermore, convolutional layers or Transformers can be used to interact with temporal and location encodings to generate joint spatiotemporal features.

[0067] Step 13: Input the time features and the location features into a preset attention mechanism model to determine the actual location of the terminal.

[0068] It should be noted that the attention mechanism model dynamically weights key spatiotemporal features, suppresses noise, and enhances effective information, ultimately outputting a high-precision trajectory. The attention layer in the attention mechanism model includes local attention and global attention. Local attention is used to calculate the association weights (such as cosine similarity) between the current trajectory point and its neighbors. Global attention is used to incorporate prior knowledge such as road direction and traffic flow to adjust the weight distribution. Here, assuming the model detects that the terminal frequently switches base stations near the highway (e.g., from base station C to D to E), the attention mechanism will enhance the weights of road direction features and suppress abnormal signals from remote base stations, ultimately ensuring that the error between the output trajectory point (x, y) and the actual road location is within a preset range.

[0069] In this embodiment, a pre-trained attention mechanism model is used, and by inputting temporal and positional features, the actual location of the terminal can be obtained from the model output. This application improves the accuracy and robustness of trajectory generation by integrating communication signaling, spatiotemporal coding, and the attention mechanism model, making it suitable for positioning needs in dynamic and complex environments.

[0070] Optionally, obtaining the temporal features in the mapping data in step 12 above includes:

[0071] By using the time branch in the spatiotemporal encoder, the timestamp information in the mapping data is normalized to determine the entry time parameter, exit time parameter, and duration parameter.

[0072] For the entry time parameter, the exit time parameter, and the duration parameter, corresponding time feature vectors are generated through trainable linear layers, respectively.

[0073] The periodicity of the time component is captured by using the cosine function to determine the periodic characteristics;

[0074] The corresponding time feature vector and the periodic feature are fused to obtain the time feature in the mapping data.

[0075] In this embodiment, the time branch in the spatiotemporal encoder is used to convert the original timestamps in the mapping data into interpretable time-series parameters and eliminate dimensional differences. After normalizing the timestamp information in the mapping data, the entry time parameter, exit time parameter, and duration parameter can be determined. The entry time parameter represents the time when the terminal enters the base station coverage area; the exit time parameter represents the time when the terminal leaves the base station coverage area; and the duration parameter represents the length of time the terminal stays in the base station coverage area.

[0076] By using trainable linear layers, such as fully connected networks, the three types of time parameters—entry time, exit time, and duration—are mapped into high-dimensional vectors, capturing the latent semantic information of these time parameters. The weights of the linear layers are optimized through gradient descent to adapt to the time representation needs of different scenarios.

[0077] Cosine functions are introduced to periodically encode time components, such as converting timestamps into periodic features. Periodic features can reflect repetitive patterns in terminal behavior. The time feature vector is then concatenated or weighted and fused with the periodic features to generate a comprehensive time feature. For example, an attention mechanism can be used to dynamically allocate the weights of the two types of features. This fusion process can suppress noise and improve the model's generalization ability to complex scenarios.

[0078] This step achieves multi-dimensional parameter extraction (entry / exit / duration), cosine function encoding of periodic components, and a trainable linear layer with fusion mechanism to enhance nonlinear representation. Through parameter refinement, periodic modeling, and dynamic fusion, the representation capability of temporal features is systematically improved, providing reliable input for downstream tasks (such as trajectory prediction and anomaly detection). This method improves the accuracy and robustness of spatiotemporal modeling through refined temporal feature extraction.

[0079] Optionally, obtaining the location features in the mapping data in step 12 above includes:

[0080] Using the multilayer perceptron in the spatiotemporal encoder, the base station locations and relative positional relationships between different base stations in the mapping data are obtained;

[0081] The location of the base station and the relative positional relationship are comprehensively encoded and represented to obtain the positional features in the mapping data.

[0082] In this embodiment, base station location features are extracted using a multilayer perceptron (MLP) in the spatiotemporal encoder. The specific process is as follows: The MLP is used to perform a nonlinear mapping of the base station's latitude and longitude coordinates, converting the absolute position in physical space into a high-dimensional vector, enhancing the model's semantic understanding of geographic coordinates. Activation functions (such as ReLU) in the MLP are used to capture base station density distribution features, such as the difference between densely populated base stations in urban areas and sparsely populated base stations in suburban areas. Parameters such as distance and orientation angle between base stations are calculated to construct a spatial topology network between base stations, representing the interaction relationships of base station coverage areas. The relative position parameters are normalized and weighted using the MLP to eliminate dimensional differences and highlight the influence of key neighboring base stations. The absolute position vector and the relative position weight matrix are concatenated or attention-weighted to generate joint spatial features. For example, a gating mechanism is used to dynamically adjust the contribution ratio of the two types of features. The fusion process introduces residual connections or batch normalization to avoid the gradient vanishing problem and improve the model's fault tolerance to base station signal drift. The fusion process in this application introduces residual connections or batch normalization to avoid the gradient vanishing problem and improve the model's fault tolerance to base station signal drift.

[0083] Specifically, the latitude and longitude positions in the signaling sequence and the spatial relationships between preceding and subsequent base stations are represented using the following encoding method:

[0084] , formula (1);

[0085] , formula (2);

[0086] in, These represent the longitude, latitude, and relative position of the base station to the preceding and following base stations, respectively. The feature representation of latitude and longitude information is used to obtain the contextual features before and after the base station and the location of the base station. MLP is a multilayer perceptron model. PosEncoder represents the comprehensive encoding representation of latitude and longitude information and relative position, which can be processed by MLP or by a custom neural network.

[0087] The timestamp information in the signaling sequence is represented using... As a representation of the comprehensive information of timestamps, the specific representation method is as follows:

[0088] , formula (3);

[0089] , formula (4);

[0090] , formula (5);

[0091] , formula (6);

[0092] in, These represent the duration of the terminal's time within the base station's detection range, its entry time, and its exit time, respectively. The method encodes the time characteristics of the terminal's time within the base station's detection range, capturing relevant features of the base station's coverage area and periodic characteristics of the terminal's behavior. Furthermore, by introducing time difference information and processing specific timestamps, this method can handle variable time intervals, making it more general. All are trainable parameters; cos refers to a trigonometric function, which can be used to learn periodic features in the representation, while also incorporating duration features. It is spliced ​​with periodic features.

[0093] Optionally, prior to step 13 of this application, the method further includes a step of generating the attention mechanism model, which includes:

[0094] Obtain the base station adjacency relationship and determine the global spatiotemporal representation based on the base station adjacency relationship;

[0095] For the movement trajectory of a single terminal, a local transfer subgraph is constructed, and personalized movement pattern features are extracted based on the local transfer subgraph using a hierarchical attention mechanism;

[0096] Extract the road network topology features and point-of-interest distribution features within the preset range of the base station;

[0097] Based on the road network topology features and the interest point distribution features, a geographic semantic representation is generated through a spatial attention network.

[0098] The global spatiotemporal representation, the personalized mobility pattern features, and the geographic semantic representation are gated and fused to output comprehensive base station information containing multi-dimensional interaction relationships.

[0099] The initial attention mechanism model is trained using the integrated base station information to generate the attention mechanism model.

[0100] In this embodiment, base station adjacency relationships are obtained. These relationships can be used to analyze the relationship between the spatial distribution of base station clusters and terminal transfer patterns (including terminal transfer frequencies), such as confirming the relationship between base station locations and terminal transfers. Determining a global spatiotemporal representation based on base station adjacency relationships solves the problem of excessive reliance on local base stations in traditional methods, supporting dynamic scheduling of cross-regional communication resources. The construction of local transfer subgraphs allows for modeling base station handover sequences as directed graphs based on the movement trajectory of individual terminals, capturing individual movement preferences.

[0101] The algorithm calculates the weights of the handover probability between base stations, focuses on high-frequency handover paths, and calculates spatial-level attention. It then dynamically adjusts the path weights based on dwell time periods, distinguishes between commuting and random movement patterns, and calculates temporal-level attention. This approach avoids individual trajectory ambiguity caused by overlapping base station signals, thereby improving trajectory prediction accuracy.

[0102] Based on parameters such as road nodes and road segment connectivity within the preset range of base stations, a graph structure is constructed to represent the regional traffic capacity, such as the connectivity differences between main roads and branch roads, thus realizing road network topology extraction. Density information of points of interest (POIs) such as commercial areas and transportation hubs is integrated to map regional functional attributes; for example, high POI density in commercial districts corresponds to high terminal aggregation, enabling POI distribution analysis. A spatial attention network design is achieved by dynamically allocating weights between road topology and POI features through a multi-head attention mechanism.

[0103] Input global spatiotemporal representation, personalized features, and geographic semantic representation, and dynamically adjust the contribution ratio of each feature through a trainable gating unit, such as the Sigmoid function, to train the initial attention mechanism model and generate the attention mechanism model.

[0104] For example, in areas with sparse base station coverage, the weights of geographic semantic representations are enhanced to compensate for signal loss; the model training strategy employs a multi-task loss function to jointly optimize objectives such as base station load prediction and terminal trajectory reconstruction. An adaptive learning rate adjustment strategy is introduced to balance the gradient update speed of different feature dimensions.

[0105] The method in this application achieves refined representation of base station information through joint modeling of global and local features, such as the distribution of base station clusters and individual trajectories, multimodal geographic semantic fusion, and dynamic gating mechanisms, providing high-precision model support for scenarios such as communication network optimization and user behavior analysis.

[0106] Furthermore, the adjacency relationships of base stations are obtained, and a global spatiotemporal representation is determined based on the adjacency relationships, including:

[0107] Obtain the spatial location of all base stations within the target area, as well as the terminal transfer mode between different base stations;

[0108] The spatial location of the base station is used as a node, and the terminal transfer mode is used as an edge to determine the adjacency relationship of the base station;

[0109] Based on the aforementioned base station adjacency relationships, a global transfer graph of the base stations is constructed;

[0110] Based on the global transition graph, the global spatiotemporal representation is determined.

[0111] In this application, base station adjacency relationships are constructed to capture base station transition patterns and spatial correlations. Based on the spatial locations of all base stations in the study area, as well as the overall transfer patterns and frequencies of terminals between different base stations, base station adjacency relationships are constructed, and a global base station transition graph is generated. Where V is the set of base station nodes, and E is the weighted edge defined by the transfer mode and frequency, capturing the global spatiotemporal features in the base station, thereby further enriching the base station's representation. For each node... Record its set of adjacent nodes as ; The original features of the nodes are given below. Subsequently, for each node in the model and its adjacent edges, the representation of each node is as follows:

[0112] , formula (7);

[0113] , formula (8);

[0114] , formula (9);

[0115] , formula (10);

[0116] , formula (11);

[0117] , formula (12);

[0118] Formula (7) is used to represent the initial representation of the node; A trainable weight matrix is ​​used to represent the global node feature transformation, mapping the original node features fi to a high-dimensional space; This represents the original characteristics of base station node i (such as coverage area, traffic volume, geolocation code, etc.). Used to represent the set of neighboring nodes of base station node i (adjacency relationship defined based on transfer frequency). This represents the attention coefficient vector, used for weighted adjacent edge features. ; The edge feature (output of formula (8)) represents the association strength from base station i to j. In formula (8) The trainable weight matrix and bias term representing the global edge features; in formula (8) This represents the original edge features (such as transition frequency, time interval, signal strength difference, etc.). Formula (9) is used for attention coefficient normalization; Used to represent the normalized attention weights, reflecting the adjacent edges The contribution ratio to node i. Formula (10) represents the dynamic edge feature update, where aglobal represents the global attention parameter vector, used to dynamically adjust the edge weights; and These represent the concatenated feature vectors of nodes i and j, respectively (the symbol ∥ indicates the concatenation operation). Used to represent the activation function with leakage, enhancing nonlinear expressive power. Equation (11) represents the updated attention coefficient. Indicates updating edge features The attention weights are recalculated. Equation (12) is used for node representation updates. σ(⋅) represents the activation function (such as Sigmoid or ReLU) to enhance nonlinearity. This represents the updated global node representation, a high-dimensional feature after fusing neighborhood information.

[0119] Through the above parameter design, the model achieves global-local feature complementarity (group patterns and individual preferences) and multimodal information fusion (communication data plus geographic semantics), providing a high-resolution representation basis for base station optimization and trajectory prediction.

[0120] The updated node features here are These are the original, trainable parameters, and , All are attention coefficients. Original edge features This is the updated edge feature. In this way, the node feature, based on the original Graph Attention Network (GAT) model, adds further consideration and fusion of edge representation, so that edge features are also considered in the global representation.

[0121] Based on the changes in base station location during the terminal's own travel patterns, a local transfer map is constructed, and the local transfer features are obtained using the GAT model. Specifically, for each terminal u, the constructed local transfer map... ,in For the terminal Base stations in the trajectory, This refers to the transition relationship in its movement trajectory. This represents a local representation, so for each terminal u, the local representation... The processing method used here is similar to that above, and the formula is:

[0122] , formula (13);

[0123] , formula (14);

[0124] , formula (15);

[0125] , formula (16);

[0126] , formula (17);

[0127] , formula (18).

[0128] Among them, the local transfer diagram ( In the relevant formulas, formula (13) represents the initial representation of the local nodes. A trainable weight matrix representing the feature transformation of local nodes (shared structure with the global matrix but with independent parameters). This represents the personalized characteristics of base station i in the trajectory of terminal u (such as dwell time, trajectory order, etc.). Formula (14) is used for local edge feature generation. This represents local edge features (such as the transfer frequency and timestamp interval of terminal u at base stations i to j). In formulas (15) to (18), the meaning of the parameters is the same as that of the global formulas (9) to (12), but the scope is limited to the local transfer graph of terminal u. alocal represents the local attention parameter vector, which captures individual mobility preferences. This represents the local node representation of terminal u, reflecting its personalized mobility pattern.

[0129] here That is all A set of.

[0130] Here it is used As a comprehensive representation of a base station, using This indicates additional features such as road network information, POI information, etc. .in, The global spatiotemporal representation (output of formula (12)) is represented, and the overall transfer pattern of the base station group is encoded. This represents a localized personalized representation (output of formula (18)) that captures the individual behavior patterns of the terminal. Additional environmental features (such as road network topology and POI distribution) are encoded into vectors via a neural network. Examples of road network topology features include road connectivity and traffic capacity; POI distribution features include the density of points of interest (POIs) such as commercial areas and transportation hubs. Feature fusion methods can utilize concatenation with fully connected layers, such as those described in this application. Gated weighting can be used, such as introducing adaptive weights to allocate the contribution ratio of each feature.

[0131] here This indicates that a comprehensive characterization is obtained by fusing multiple features, including interactions between base stations and interactions between base stations and their context. This represents the global characterization of the base station. By vectorizing environmental features, such as road network structure and POIs, a comprehensive representation can be obtained through neural networks.

[0132] For the road network structure, all roads are represented by vectors, while the types of POIs are represented using one-hot encoding. For each base station, various relevant information such as the number of roads within a certain range of the base station, the distance to the roads, the number and types of POIs around the base station are considered and statistically analyzed. An attention-based network is designed to fully consider the importance and relevance of geographic semantic information, thereby obtaining... .

[0133] Optionally, the initial attention mechanism model is trained using the integrated base station information to generate the attention mechanism model, including:

[0134] Obtain sample time features and sample location features generated from samples;

[0135] Based on the integrated base station information, the sample time features, and the sample location features, a multi-head self-attention mechanism is used to generate multi-time-step interactive representations.

[0136] A multi-task loss function is constructed, which includes a geographic coordinate reconstruction error term, a positive sample representation similarity term, and a negative sample comparison term. The model parameters are optimized through self-supervised learning.

[0137] The attention mechanism model is constructed based on the interaction representation and the multi-task loss function.

[0138] In this embodiment, temporal features (such as dwell time and transfer interval) and spatial features (such as base station latitude and longitude and POI distribution) of samples are extracted from terminal signaling data. These are combined with comprehensive information such as base station coverage and traffic load to construct an input feature matrix, converting the original signaling data into structured features and capturing the spatiotemporal regularity of user movement behavior. A multi-head self-attention mechanism is employed to dynamically weight and fuse the spatiotemporal features of the base station. Each attention head independently learns different dimensions of spatiotemporal association patterns (such as time dependence, spatial proximity, and semantic similarity), and the multi-head outputs are finally concatenated to obtain a comprehensive interactive representation. The input features are linearly mapped into query, key, and value vectors; through attention weight calculation, the focus is adaptively applied to key time steps and base station nodes.

[0139] The construction of a multi-task loss function can include designing a geographic coordinate reconstruction error term, a positive sample representation similarity term, and a negative sample comparison term. Then, based on these terms, a total loss function is constructed. The geographic coordinate reconstruction error term includes minimizing the root mean square error (RMSE) between the predicted and actual locations, ensuring that the latitude and longitude output by the model are consistent with the actual trajectory. The positive sample representation similarity term includes maximizing the representation similarity (e.g., cosine similarity) of the actual trajectories of the same user, improving the model's ability to capture user behavior patterns. The negative sample comparison term includes introducing adversarial negative examples (e.g., randomly perturbed trajectories or other user trajectories) to distinguish between positive and negative samples by comparing the loss, enhancing the model's robustness.

[0140] Based on interactive representations and multi-task loss functions, the constructed attention mechanism model can jointly optimize multi-task losses using backpropagation and update model weights through gradient descent. A self-supervised mechanism reduces reliance on manually labeled data and automatically learns representations using the spatiotemporal correlation of base stations. The final output of the constructed attention mechanism model is an attention weight matrix that integrates the global context of the base station (e.g., network topology) and local behavioral patterns (e.g., user mobility preferences), supporting downstream tasks such as trajectory generation and base station coverage correction.

[0141] This application employs a multi-head mechanism and differential attention to suppress noise interference from irrelevant base station nodes and increase the weight of key spatiotemporal features. By combining base station attributes (coverage range), environmental semantics (road network, POI), and user behavior (transfer frequency), the physical interpretability of the representation is enhanced. This application also achieves joint optimization of geographic coordinate regression and representation comparison tasks, avoiding information loss during staged training.

[0142] Specifically, the steps for generating an attention mechanism model include:

[0143] Using a TransformerEncoder based on a self-attention mechanism, multimodal representations are synthesized into spatiotemporal representations. ,in, Indicates the time characteristics of the sample. Indicates the location features of the sample. This is a comprehensive base station representation determined after processing the sample time and location features generated from the samples. Meanwhile, It contains information from multiple time steps. By using an attention mechanism to assign different weights to different time steps, the information from each time step is fully considered in the generation of each position in the actual trajectory.

[0144] By using actual location information as labels, combining a self-supervised learning module with a mean squared error loss function, a custom loss function is designed to train the model, thereby achieving the task of reconstructing precise locations from signaling trajectory sequences. In designing the custom loss function, coefficients control the weight of each of the above components. The specific formula is as follows:

[0145] , formula (19);

[0146] Wherein, formula (19) is the total loss function; is a hyperparameter used to adjust the weights of the contrastive loss term, balancing the importance of the position regression task and the representation learning task. The training objective of equation (19) is: to minimize... Used to improve the accuracy of latitude and longitude predictions. Maximize - This is used to enhance the model's ability to distinguish between real and abnormal trajectories. For calculations based on latitude and longitude, this is expressed as the root mean square error loss; and The calculations are performed using self-supervised learning, and are represented as positive sample contrast loss and negative sample contrast loss, respectively; the specific method is as follows: It is the loss function based on the existing movement trajectory of the terminal, and This is the loss function representing the data obtained by randomly sampling nodes of the same length, and it is calculated as follows:

[0147] , formula (20).

[0148] , formula (21).

[0149] , formula (22).

[0150] Where N represents the total number of samples, that is, the number of trajectory location points; This represents the actual latitude and longitude of the i-th location point; Represents the latitude and longitude of the i-th location predicted by the model. Formula (20) is used to measure the average error of the two-dimensional Euclidean distance between the predicted location and the actual location (the unit is consistent with the input data, usually in degrees or projected plane coordinate units). M represents the number of positive or negative samples; For positive samples, it represents the representation of the i-th true trajectory, which is obtained by averaging the representation vectors of all base station nodes in the trajectory; Negative samples represent the characteristics of the i-th abnormal trajectory and are typically constructed in the following ways: randomly sampling the trajectories of other users; perturbing the trajectory of the same user (e.g., randomly replacing some nodes); or generating adversarial negative examples (e.g., using a generative model). Used to represent the similarity score that maximizes the true trajectory. This is the representation of each actual trajectory, obtained by averaging the representations of the corresponding nodes in the trajectory from the above node representations. Similarly, we can obtain... Definition of . () represent the actual longitude and the predicted longitude, respectively. These represent the actual latitude and the predicted latitude, respectively, while λ is used to adjust the weights.

[0151] By adjusting the parameters using the aforementioned loss function, the model is trained, and the best-performing model is saved. Then, the terminal's signaling data is input into the trained model for processing, resulting in the generation of the terminal's actual trajectory location. Through this training process, the model can analyze the difference between the base station's location and the terminal's actual location, thereby generating the actual terminal's travel trajectory and applying it to more downstream tasks.

[0152] It should be noted that, referring to Figure 2As shown, the training process of the above model includes: starting by dividing the dataset; selecting model hyperparameters; evaluating model training; determining if the model performance is optimal; if not, continuing with the step of selecting model hyperparameters; if optimal, saving the model parameters; saving node representations; and ending the training process.

[0153] In one specific embodiment, the terminal trajectory generation method based on communication signaling data of this application comprises the following steps: Data preprocessing: collecting signaling data and the user's actual travel data. Constructing a mapping from the actual signaling trajectory to the actual travel location, and filtering out abnormal data. Feature extraction: processing the spatiotemporal location information in the collected signaling data through a spatiotemporal encoder to obtain a spatiotemporal representation. Base station interaction representation construction: capturing the base station's transition modes and spatial correlations, as well as its interaction relationships with the surrounding environment. Combining auxiliary information and graph neural networks to obtain a comprehensive representation. Using an attention-based model, learning the global and local spatiotemporal relationships of the signaling sequence, capturing the differences between it and the actual location, and generating the user's actual location.

[0154] In summary, the solution presented in this application significantly improves the accuracy and generalization ability of terminal trajectory generation through deep modeling of communication signaling data and fusion of multi-source information, and demonstrates clear advantages in processing efficiency, multimodal adaptability, and practical application scenarios. Specific technical effects are as follows:

[0155] (1) Significantly improved positioning accuracy: The global-local transition graph modeling of this application combines the adjacency relationship of base stations (handover frequency, spatial correlation) with the trajectory patterns of individual terminals to capture group patterns and individual preferences; and the designed multimodal attention mechanism dynamically fuses base station attributes, road network connectivity and POI semantic features to improve environmental perception capabilities. Compared with traditional rule-based and clustering methods, the positioning error is reduced, mainly due to the nonlinear modeling capability of deep learning models for complex spatiotemporal patterns. With the support of multi-source information fusion (POI, road network), the positioning accuracy is improved compared with methods that only rely on trajectory information, and the signal drift interference is reduced by enhancing the interactive perception capability between base stations and the environment.

[0156] (2) It supports unified modeling of various travel modes such as walking, driving, and public transportation, without the need to design special rules for a single travel mode, thus improving processing efficiency; through the spatiotemporal encoder and self-supervised learning module, it adaptively extracts the implicit features of different travel modes (such as speed changes and dwell patterns).

[0157] (3) Innovations were made in multi-source data fusion and modeling: environmental features such as road density and POI type distribution were encoded into high-dimensional vectors and weighted by attention mechanism to enhance the context awareness of base stations; the correlation matrix between base stations and roads and base stations and POIs was constructed to capture the correlation between user behavior and geographic semantics (such as high probability of staying in commercial areas).

[0158] (4) By using the multi-task loss function (geographic coordinate reconstruction plus positive and negative sample comparison) to reduce the dependence on labeled data, the generalization ability of the model in sparse signaling scenarios is improved, and self-supervised learning optimization is achieved; by introducing perturbation trajectories and cross-user negative examples, the robustness of the model to noisy data (such as abnormal base station handover) is enhanced.

[0159] (5) Downstream application scalability and diversified scenario support: Traffic management: Generate high-precision vehicle trajectories to optimize traffic light control and congestion prediction; Urban safety: Identify traffic accident areas through abnormal trajectory detection (such as long-term stays and frequent turns); Business analysis: Combine POI data to analyze user consumption behavior and regional business vitality.

[0160] The proposed solution can serve fields such as smart cities, transportation planning, and public safety, filling the technological gap in high-precision trajectory data generation; it also improves the utilization rate of signaling data, creating value-added data services for operators (such as user behavior analysis and regional traffic prediction). This proposal achieves reliable mapping from low-precision signaling data to high-resolution trajectories through multi-source fusion modeling, self-supervised optimization, and dynamic graph attention mechanisms, providing core technical support for smart cities and mobile behavior analysis.

[0161] The various methods of the embodiments of this application have been described above. Apparatus for implementing the above methods will now be provided.

[0162] Please refer to Figure 3 This application also provides a terminal trajectory generation device, comprising:

[0163] The first determining module 31 is used to determine the mapping data between the signaling and location of the terminal; wherein, the first determining module 31 includes: a first acquiring unit, used to acquire the terminal's communication signaling data and actual trajectory data; a first processing unit, used to perform time sorting and anomaly removal operations on the communication signaling data to generate a signaling sequence; and a second processing unit, used to determine the mapping data with a mapping relationship between the signaling and location based on the actual location in the signaling sequence and the actual trajectory data, according to the time alignment method, and to filter out abnormal data;

[0164] The first processing module 32 is used to obtain the time features and location features in the mapping data using a spatiotemporal encoder;

[0165] The second processing module 33 is used to input the time features and the location features into a preset attention mechanism model to determine the actual location of the terminal.

[0166] Optionally, the first processing module 32 described above includes:

[0167] The third processing unit is used to normalize the timestamp information in the mapping data by using the time branch in the space-time encoder, and to determine the entry time parameter, exit time parameter and duration parameter.

[0168] The fourth processing unit is used to generate corresponding time feature vectors for the entry time parameter, the exit time parameter, and the duration parameter respectively through a trainable linear layer;

[0169] The first determining unit is used to capture the periodicity of the time component using the cosine function and determine the periodic characteristics.

[0170] The second acquisition unit is used to perform feature fusion on the corresponding time feature vector and the periodic feature to obtain the time feature in the mapping data.

[0171] Optionally, the first processing module 32 described above further includes:

[0172] The fifth processing unit is used to obtain the base station locations and relative positional relationships between different base stations in the mapping data by utilizing the multilayer perceptron in the spatiotemporal encoder.

[0173] The third acquisition unit is used to perform comprehensive encoding and characterization of the base station location and the relative positional relationship to acquire the positional features in the mapping data.

[0174] Optionally, the terminal trajectory generation device of this application further includes:

[0175] A construction module is used to obtain the adjacency relationship of base stations and determine the global spatiotemporal representation based on the adjacency relationship of base stations;

[0176] The third processing module is used to construct a local transfer sub-graph for the movement trajectory of a single terminal, and extract personalized movement pattern features based on the local transfer sub-graph using a hierarchical attention mechanism.

[0177] The fourth processing module is used to extract the road network topology features and point-of-interest distribution features within a preset range of the base station;

[0178] The fifth processing module is used to generate a geographic semantic representation through a spatial attention network based on the road network topology features and the interest point distribution features.

[0179] The sixth processing module is used to perform gated fusion of the global spatiotemporal representation, the personalized mobility pattern features and the geographic semantic representation, and output comprehensive base station information containing multi-dimensional interaction relationships;

[0180] The seventh processing module is used to train the initial attention mechanism model using the integrated base station information to generate the attention mechanism model.

[0181] Optionally, the above-mentioned building modules include:

[0182] The fourth acquisition unit is used to acquire the spatial location of all base stations within the target area, as well as the terminal transfer mode of the terminal between different base stations;

[0183] The second determining unit is used to determine the adjacency relationship of the base station by taking the spatial location of the base station as a node and the terminal transfer mode as an edge.

[0184] The first construction unit is used to construct a global transfer graph of the base stations based on the adjacency relationship of the base stations;

[0185] The third determining unit is used to determine the global spatiotemporal representation based on the global transition graph.

[0186] Optionally, the seventh processing module mentioned above includes:

[0187] The fifth acquisition unit is used to acquire the sample time features and sample location features generated based on the samples;

[0188] The sixth processing unit is used to generate multi-time-step interactive representations based on the integrated base station information, the sample time features, and the sample location features using a multi-head self-attention mechanism.

[0189] The second building unit is used to construct a multi-task loss function; the multi-task loss function includes a geographic coordinate reconstruction error term, a positive sample representation similarity term, and a negative sample comparison term, and optimizes the model parameters through self-supervised learning.

[0190] The third building module is used to build the attention mechanism model based on the interaction representation and the multi-task loss function.

[0191] It should be noted that the device in this embodiment corresponds to the method described above, and the implementation methods in each of the above embodiments are applicable to the embodiment of this device, achieving the same technical effect. The device provided in this application embodiment can implement all the method steps implemented in the above method embodiments and achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0192] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described terminal trajectory generation method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0193] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described terminal trajectory generation method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0194] It should be noted that the collection, gathering, updating, analysis, processing, use, transmission, and storage of user personal information involved in this disclosed technical solution all comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken to prevent unauthorized access to user personal information data and to safeguard user personal information security and network security.

[0195] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0196] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0197] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for generating a terminal trajectory, characterized in that, include: Determining the mapping data between terminal signaling and location includes: acquiring terminal communication signaling data and actual trajectory data; performing time sorting and outlier removal operations on the communication signaling data to generate a signaling sequence; based on the actual location in the signaling sequence and the actual trajectory data, determining the mapping data with a mapping relationship between signaling and location according to time alignment, and filtering out abnormal data; The method of obtaining temporal and positional features from the mapped data using a spatiotemporal encoder includes: normalizing the timestamp information in the mapped data using the time branch of the spatiotemporal encoder to determine entry time parameters, exit time parameters, and duration parameters; generating corresponding time feature vectors for the entry time parameter, exit time parameter, and duration parameter respectively through trainable linear layers; capturing the periodicity of the time components using a cosine function to determine periodic features; and fusing the corresponding time feature vectors and the periodic features to obtain the temporal features from the mapped data. The method of obtaining positional features from the mapped data includes: using a multilayer perceptron in the spatiotemporal encoder to obtain the base station locations and relative positional relationships between different base stations in the mapped data; and comprehensively encoding and representing the base station locations and the relative positional relationships to obtain the positional features from the mapped data. The time features and location features are input into a preset attention mechanism model to determine the actual location of the terminal.

2. The method according to claim 1, characterized in that, Before inputting the time features and location features into a preset attention mechanism model to determine the actual location of the terminal, the method further includes: Obtain the base station adjacency relationship and determine the global spatiotemporal representation based on the base station adjacency relationship; For the movement trajectory of a single terminal, a local transfer subgraph is constructed, and personalized movement pattern features are extracted based on the local transfer subgraph using a hierarchical attention mechanism; Extract the road network topology features and point-of-interest distribution features within the preset range of the base station; Based on the road network topology features and the interest point distribution features, a geographic semantic representation is generated through a spatial attention network. The global spatiotemporal representation, the personalized mobility pattern features, and the geographic semantic representation are gated and fused to output comprehensive base station information containing multi-dimensional interaction relationships. The initial attention mechanism model is trained using the integrated base station information to generate the attention mechanism model.

3. The method according to claim 2, characterized in that, Obtaining base station adjacency relationships and determining a global spatiotemporal representation based on these relationships includes: Obtain the spatial location of all base stations within the target area, as well as the terminal transfer mode between different base stations; The spatial location of the base station is used as a node, and the terminal transfer mode is used as an edge to determine the adjacency relationship of the base station; Based on the aforementioned base station adjacency relationships, a global transfer graph of the base stations is constructed; Based on the global transition graph, the global spatiotemporal representation is determined.

4. The method according to claim 2, characterized in that, Using the integrated base station information, the initial attention mechanism model is trained to generate the attention mechanism model, including: Obtain the sample time features and sample location features generated based on the samples; Based on the integrated base station information, the sample time features, and the sample location features, a multi-head self-attention mechanism is used to generate multi-time-step interactive representations. A multi-task loss function is constructed; the multi-task loss function includes a geographic coordinate reconstruction error term, a positive sample representation similarity term, and a negative sample comparison term, and the model parameters are optimized through self-supervised learning; The attention mechanism model is constructed based on the interaction representation and the multi-task loss function.

5. A terminal trajectory generation device, characterized in that, include: A first determining module is used to determine the mapping data between the signaling and location of the terminal; wherein, the first determining module includes: a first acquiring unit, used to acquire the terminal's communication signaling data and actual trajectory data; a first processing unit, used to perform time sorting and outlier removal operations on the communication signaling data to generate a signaling sequence; and a second processing unit, used to determine the mapping data with a mapping relationship between the signaling and location based on the actual location in the signaling sequence and the actual trajectory data, according to time alignment, and to filter out outlier data; The first processing module is used to acquire time features and location features in the mapping data using a spatiotemporal encoder. The first processing module includes: a third processing unit used to normalize the timestamp information in the mapping data using a time branch in the spatiotemporal encoder to determine entry time parameters, exit time parameters, and duration parameters; a fourth processing unit used to generate corresponding time feature vectors for the entry time parameter, the exit time parameter, and the duration parameter respectively through a trainable linear layer; a first determining unit used to capture the periodicity of the time components using a cosine function to determine periodic features; and a second acquiring unit used to fuse the corresponding time feature vectors and the periodic features to acquire time features in the mapping data. The first processing module further includes: a fifth processing unit used to acquire the base station locations and relative positional relationships between different base stations in the mapping data using a multilayer perceptron in the spatiotemporal encoder; and a third acquiring unit used to comprehensively encode and represent the base station locations and the relative positional relationships to acquire location features in the mapping data. The second processing module is used to input the time features and the location features into a preset attention mechanism model to determine the actual location of the terminal.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 4.

7. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 4.

Images