Mobile terminal GIS data preloading system and method based on deep learning

Abstract

This invention discloses a mobile GIS data preloading system and method based on deep learning, belonging to the field of mobile geographic information system technology. The method includes: acquiring motion trajectory data and historical network connectivity pattern data; predicting connectivity window parameters using a connectivity window prediction model fused with a temporal convolutional network and a physical motion model; acquiring GIS data metadata and constructing a data dependency graph containing dependencies; inputting the connectivity window parameters and the data dependency graph into a decision model; filtering the smallest independent available unit under bandwidth constraints based on dependency integrity; and acquiring the corresponding data block and storing it in a local cache when the connectivity window start time arrives. This invention integrates a physical model with deep learning to accurately predict pulsed network windows, and constructs the smallest independent available unit based on data dependencies, solving the problem of preloaded data being unusable due to missing dependencies in extremely weak network environments, and achieving complete delivery of offline functional units within limited bandwidth.

Description

Technical Field

[0001] This invention relates to the field of mobile geographic information system technology, and in particular to a mobile GIS data preloading system and method based on deep learning. Background Technology

[0002] Mobile GIS applications play a vital role in scenarios such as user travel, emergency rescue, and field operations. However, mobile devices often face network instability, especially in high-speed travel (high-speed rail, subway), remote mountainous areas, and underground spaces, where network connectivity exhibits a pulse-like characteristic of "intermittent connectivity and long-term interruptions." Traditional map data preloading methods, typically based on smooth predictions of user trajectories and only considering simple pre-fetching of tiles around the current viewport, cannot address the following technical challenges:

[0003] First, existing methods assume that the network is continuously available or only experiences slight fluctuations, making it difficult to accurately predict the start time, duration, and bandwidth capacity of the connectivity window for pulsed networks, resulting in a misalignment between preloading tasks and network availability.

[0004] Secondly, there are hierarchical dependencies between GIS data (e.g., high-precision 3D models depend on terrain data, and terrain data depends on base maps). Traditional preloading strategies select data blocks in isolation, resulting in preloaded data being unusable offline due to the lack of prerequisites, leading to wasted bandwidth. Furthermore, under conditions of extremely weak network and severely limited bandwidth, existing solutions cannot intelligently select the smallest data unit capable of independently supporting full functionality offline, based on bandwidth budget and data dependencies, within a limited connectivity window. Summary of the Invention

[0005] This invention aims to provide a mobile GIS data preloading system and method based on deep learning to solve the intelligent preloading problem of simultaneously predicting connected window characteristics, evaluating the dependencies between data blocks, and calculating the "minimum independent usable data unit" under conditions of extreme pulsed networks and extremely limited single-window bandwidth, ensuring that the data transmitted by each limited connected window can form a complete and usable functional module in offline state.

[0006] To achieve the above objectives, this invention provides a deep learning-based method for preloading mobile GIS data, comprising the following steps:

[0007] (1) Obtain motion trajectory data and historical network connectivity pattern data of mobile terminal device, input the motion trajectory data and the historical network connectivity pattern data into a pre-trained connectivity window prediction model, and predict the connectivity window parameters within a preset time period through the connectivity window prediction model. The connectivity window parameters include the start time, duration and estimated bandwidth of the connectivity window.

[0008] (2) Obtain GIS data metadata of the current user’s location, and construct a data dependency graph based on the GIS data metadata. The data dependency graph includes multiple data block nodes and multiple directed edges, and the directed edges are used to represent the dependency relationships between the data block nodes.

[0009] (3) Input the connected window parameters and the data dependency graph into the pre-trained decision model. Under the constraint of the estimated bandwidth, the decision model selects at least one minimum independent available unit from the multiple data block nodes, with the integrity of the dependency relationship in the data dependency graph as the condition. The minimum independent available unit refers to a data set that can independently support the preset GIS function in offline state, consisting of at least one data block node and all its dependent preceding data block nodes.

[0010] (4) When the start time of the connection window arrives, according to the estimated bandwidth constraint, the data block corresponding to the smallest independent available unit is obtained, and the obtained data block is stored in the local cache of the mobile terminal.

[0011] Furthermore, the connected window prediction model is a fusion model of temporal convolutional network and physical motion model. In step (1), the connected window parameters are predicted through the connected window prediction model, including:

[0012] The temporal convolutional network extracts temporal features from the motion trajectory data and the historical network connectivity pattern data, and outputs the initial window prediction result; the physical motion model calculates the path loss based on the position and velocity information in the motion trajectory data, combined with the signal propagation attenuation model. ,in For path loss, For reference distance Path loss at the location, This is the path loss index. This refers to the real-time distance between the mobile device and the base station. The shadow fading random variable is used in conjunction with the Doppler frequency shift model. Calculate the Doppler frequency shift, where This is the Doppler frequency shift. For the movement speed of mobile devices, The angle between the direction of motion and the direction of signal incidence. For the signal wavelength, The carrier frequency is used as the reference. The initial window prediction result is fused with the path loss and the Doppler frequency shift to obtain the corrected connectivity window parameters.

[0013] Furthermore, the data dependency graph is a dynamic data dependency graph, which also includes a dynamic weight matrix. This dynamic weight matrix characterizes the decay factor of the dependency strength of each directed edge over time. The decay factor is expressed by the formula... Calculation, where For a moment Dependence strength For the initial dependency strength, The attenuation coefficient is... This is the moment when dependencies are generated.

[0014] Further, step (2) involves constructing a data dependency graph based on the GIS data metadata, including:

[0015] Parse the GIS data metadata to extract the unique identifier, data type, dependency identifier, and timeliness label for each data block; construct an initial directed graph using the unique identifier of each data block as a node and the dependency identifier as the pointing relationship of directed edges; perform cyclic dependency detection on the initial directed graph, and when cyclic dependency is detected, decoupling is performed by introducing virtual nodes to generate an acyclic data dependency graph; set a dynamic weight for each data block according to the timeliness label, and reduce the filtering priority of the data block in the smallest independent available unit when the validity period of the data block is less than the duration of the connected window.

[0016] Furthermore, the decision model is a deep reinforcement learning model, and step (3) involves selecting at least one minimum independent usable unit through the decision model, including:

[0017] The connected window parameters and the data dependency graph are encoded into a state vector. ; through the policy network based on the state vector Output action probability distribution The actions The decision to select a data block node from candidate data block nodes to join the set of minimum independent available units; through a reward function. The learning of the policy network is guided, in which For indicator functions, when The value is 1 when dependency integrity is satisfied, and the volume is 1. For the present Total volume, Latency To obtain Estimated latency for all data blocks in the middle. For the estimated bandwidth, The duration is... , , These are the preset weighting coefficients.

[0018] Furthermore, the minimum independent available unit satisfies the condition and ,in For the set of the smallest independent available units, For data blocks volume, For data blocks The set of all preceding dependent data blocks.

[0019] Further, obtaining the data block corresponding to the smallest independent available unit in step (4) includes:

[0020] The system obtains the current user's geographic location information, current viewport range, and current user's behavioral heatmap value. The behavioral heatmap value is obtained using a formula... Calculation, where For users in the first The frequency of accesses to the current region within a historical time period The weighting coefficients are related to the time distance of the current period; based on the geographical location information, the current viewport range, and the behavior heatmap value, the data blocks in the smallest independent available unit are weighted and sorted to generate a preloading queue. The weighting coefficients for the weighted sorting are determined by the formula... Calculation, where For data blocks The distance between the corresponding spatial area and the user's current location. For data blocks The predicted access probability, For data blocks Data freshness, , , The preset weighting coefficients are used; within the duration of the connected window, the corresponding data blocks are acquired in the order of the preloaded queue.

[0021] Another objective of this invention is to provide a mobile GIS data preloading system based on deep learning, the system comprising:

[0022] The connectivity window prediction module is used to acquire motion trajectory data and historical network connectivity pattern data of mobile devices, input the motion trajectory data and the historical network connectivity pattern data into a pre-trained connectivity window prediction model, and predict the connectivity window parameters within a preset time period through the connectivity window prediction model. The connectivity window parameters include the start time, duration and estimated bandwidth of the connectivity window.

[0023] The dependency relationship construction module is used to obtain GIS data metadata of the current user's location and construct a data dependency relationship graph based on the GIS data metadata. The data dependency relationship graph includes multiple data block nodes and multiple directed edges, and the directed edges are used to represent the dependency relationship between the data block nodes.

[0024] The decision module is used to input the connectivity window parameters and the data dependency graph into a pre-trained decision model. Under the constraint of the estimated bandwidth and with the integrity of the dependency relationship in the data dependency graph as a condition, the decision model selects at least one minimum independent available unit from the multiple data block nodes. The minimum independent available unit refers to a data set that consists of at least one data block node and all its dependent preceding data block nodes, which can independently support the preset GIS functions in offline mode.

[0025] The preload execution module is used to obtain the data block corresponding to the smallest independent available unit according to the estimated bandwidth constraint when the start time of the connection window arrives, and store the obtained data block to the local cache of the mobile device.

[0026] Furthermore, the connected window prediction model is a fusion model of temporal convolutional networks and physical motion models, and the connected window prediction module includes:

[0027] The temporal feature extraction unit is used to extract the temporal features of the motion trajectory data and the historical network connectivity pattern data through the temporal convolutional network, and output the initial window prediction result; the physical attenuation calculation unit is used to calculate the path loss and Doppler frequency shift based on the position and velocity information in the motion trajectory data, combined with the signal propagation attenuation model and the Doppler frequency shift model, using the physical motion model; the fusion correction unit is used to fuse the initial window prediction result with the path loss and the Doppler frequency shift to obtain the corrected connectivity window parameters.

[0028] Furthermore, the decision-making model is a deep reinforcement learning model, and the decision-making module includes:

[0029] A state encoding unit is used to encode the connectivity window parameters and the data dependency graph into a state vector; a policy output unit is used to output an action probability distribution through a policy network based on the state vector, the action probability distribution being used to select data block nodes from candidate data block nodes to add to the minimum independent available unit set; a reward calculation unit is used to guide the learning of the policy network through a reward function, the reward function being constructed based on dependency integrity, the ratio of the total data block volume to the estimated bandwidth, and the ratio of the data block acquisition latency to the duration.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. By integrating temporal convolutional networks with physical motion models, signal propagation attenuation, and Doppler frequency shift, high-precision prediction of the start time, duration, and estimated bandwidth of the connectivity window of pulsed networks is achieved, solving the problem that traditional methods cannot handle intermittent networks.

[0032] 2. Construct a data dependency graph and introduce mathematical constraints on the smallest independent available unit to ensure that the data preloaded in each connected window can independently support the complete GIS function in offline mode, thus fundamentally eliminating bandwidth waste;

[0033] 3. A deep reinforcement learning decision model is adopted to unify the modeling of multi-dimensional constraints such as bandwidth budget, window duration, data dependence, and data timeliness. The reward function guides the policy network to learn the optimal data selection strategy, thereby maximizing utility under extremely narrow bandwidth.

[0034] 4. By combining multiple factors such as user geographic location, viewport range, behavioral heatmap value, and data freshness, a smart preloading queue is generated, further enhancing the practical value of data within a limited window period. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0036] Figure 1 A flowchart illustrating the overall process of a deep learning-based mobile GIS data preloading method provided in this embodiment of the invention.

[0037] Figure 2 This is a schematic diagram of the internal structure of the connected window prediction model in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram illustrating the construction and dynamic weight update of the data dependency graph in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the training and inference process of the deep reinforcement learning decision model in an embodiment of the present invention;

[0040] Figure 5 The module structure diagram of the mobile GIS data preloading system based on deep learning provided in the embodiments of the present invention is shown. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] This embodiment provides a mobile GIS data preloading method based on deep learning, such as... Figure 1 As shown, it includes the following steps:

[0043] Step S10: Predict connectivity window parameters: The mobile device (such as a smartphone) collects motion trajectory data and historical network connectivity pattern data in real time. The motion trajectory data includes the position coordinate sequence obtained by the GPS module and the speed and direction information obtained by the accelerometer and gyroscope; the historical network connectivity pattern data includes time series such as network signal strength (RSSI), signal-to-noise ratio (SNR), and measured bandwidth values ​​recorded over a period of time.

[0044] These data are then fed into a pre-trained connected window prediction model. This model is a fusion of a temporal convolutional network (TCN) and a physical motion model, such as... Figure 2 As shown. The TCN part receives the time series of motion trajectories and network patterns, extracts temporal features through causal convolution and dilated convolution, and outputs the initial window prediction results (start time, duration, and estimated bandwidth). The physical motion model calculates path loss and Doppler shift based on the position and velocity in the current motion trajectory, combined with the signal propagation attenuation model and the Doppler shift model. The path loss calculation formula is:

[0045]

[0046] In the formula, This represents path loss (dB). For reference distance Path loss at the location, The path loss index is 3-4 for urban environments and 4-5 for tunnels. This refers to the real-time distance between the mobile device and the base station. Let be a random variable representing shadow fading (following a zero-mean Gaussian distribution). The Doppler shift calculation formula is:

[0047]

[0048] In the formula, The value is the Doppler frequency shift (Hz). The speed of movement of the mobile device (m / s). The angle between the direction of motion and the direction of signal incidence. For the signal wavelength, The carrier frequency is used. The path loss and Doppler shift output by the physical model are used to correct the initial prediction of the TCN, for example, by using Kalman filtering fusion to obtain the final connectivity window parameters.

[0049] Step S20: Construct a data dependency graph: such as Figure 3 As shown, the system acquires GIS data metadata for the current user's location, including unique identifiers for data blocks, data types (base map, vector, raster, 3D model, etc.), dependency identifiers (listing other data blocks necessary for the normal operation of the current data block), and expiration labels (data validity period). Based on this information, a data dependency graph is constructed: each data block is used as a node, and directed edges are defined using dependency identifiers to form an initial directed graph. To prevent circular dependencies from causing optimization difficulties, circular dependency detection is performed on the initial directed graph. If a cycle exists (e.g., A depends on B, B depends on A), a virtual node is introduced to break the cycle, generating an acyclic directed graph.

[0050] Furthermore, to reflect the decay characteristics of dependencies over time, a dynamic weight matrix is ​​assigned to each directed edge, with the weight factors decaying exponentially:

[0051]

[0052] In the formula, For a moment Dependence strength For the initial dependency strength, This is the attenuation coefficient (preset based on data type). This indicates the time at which dependencies are generated. Additionally, based on the timeliness label, if the validity period of a data block is less than the duration of the connected window, its selection priority is reduced.

[0053] Step S30: Filtering the smallest independent available unit: Input the connected window parameters predicted in step S10 and the data dependency graph constructed in step S20 into the pre-trained decision model. In this embodiment, the decision model adopts a deep reinforcement learning (DRL) framework, the core of which is a policy network used to select the optimal action given a state. Figure 4 As shown, the state coding unit will connect the window parameters (estimated bandwidth) Duration The data dependency graph (node ​​features, dependency edges, dynamic weights) is encoded into a state vector. The policy network is based on... Output action probability distribution The actions This indicates selecting a node from the candidate data block nodes to add to the current set. After the system executes the action, the environment returns a reward. The reward function is designed as follows:

[0054]

[0055] In the formula, For indicator functions, when Satisfying dependency integrity (i.e.) If the value is 1, then the value is 1; otherwise, the value is 0. This represents the total volume of the current set; Latency To estimate download delay; The preset weights are used. By maximizing the cumulative reward, the policy network learns to select the set of nodes that constitute the smallest independent and usable unit under the dual constraints of bandwidth and time, and this set satisfies:

[0056]

[0057] That is, set It is the smallest set among all sets that satisfy dependency integrity, thus ensuring that at least one complete functional module can be loaded each time a connected window is connected.

[0058] Step S40: Perform preloading: When the connection window start time arrives, the system generates a preloading queue based on the minimum independent available unit set output by the decision model and executes the download. To improve bandwidth utilization efficiency, a weighted sorting is performed based on the user's current behavior: obtaining the user's current geographic location information (GPS), current viewport range (the map area displayed on the screen), and the user's behavior heatmap value. The formula for calculating the behavioral thermal value is:

[0059]

[0060] In the formula, For users in the first The frequency of accesses to the current region within a historical time period Time-weighted (higher weight for recent data). Then, for each data block in the smallest independent available unit... Calculate the weights:

[0061]

[0062] In the formula, This represents the distance between the spatial region corresponding to the data block and the user's current location. This represents the probability predicted by the model that a user will access this data block. For the freshness of data blocks (such as update time). These are preset coefficients. (Press...) A preload queue is generated in descending order, and data blocks are sequentially acquired and stored in the local cache within the duration of the connected window.

[0063] Application example:

[0064] Taking a user traveling by high-speed rail from city A to city B as an example, the high-speed rail is about to enter a tunnel complex approximately 10 kilometers long, where historical data shows a complete network outage. The system executes step S10, predicting a 5-minute connectivity window with a bandwidth of 200KB within the next 15 minutes based on GPS trajectory (high-speed movement) and historical network patterns (no signal in the tunnel). In step S20, the system acquires GIS metadata for the tunnel area and constructs a dependency map: a 3D model of the cross-river bridge (150KB) depends on terrain data (80KB), and the terrain data depends on a base map (50KB). In step S30, the decision model calculates that the total demand of 280KB exceeds the 200KB bandwidth requirement, so the 3D model is abandoned, and the "base map + terrain data" is selected as the smallest independent available unit, which allows the user to view the terrain offline. In step S40, 5 seconds before entering the tunnel, the system starts preloading, prioritizing the loading of terrain data near the window based on spatial distance and user behavior, and successfully downloads two data blocks within the window period. After entering the tunnel, users can still browse the topographic map of the area offline without waiting for the network to be restored.

[0065] This embodiment provides a mobile GIS data preloading system based on deep learning, such as... Figure 5 As shown, it includes:

[0066] Connectivity window prediction module (100): used to acquire motion trajectory data and historical network connectivity pattern data of mobile devices, input the motion trajectory data and historical network connectivity pattern data into a pre-trained connectivity window prediction model, and predict the connectivity window parameters within a preset time period through the connectivity window prediction model. The connectivity window parameters include the start time, duration and estimated bandwidth of the connectivity window.

[0067] The connected window prediction module (100) specifically includes: a temporal feature extraction unit (101), which is used to extract the temporal features of motion trajectory data and historical network connectivity pattern data through a temporal convolutional network and output the initial window prediction result; a physical attenuation calculation unit (102), which is used to calculate the path loss and Doppler frequency shift based on the position and velocity information in the motion trajectory data through a physical motion model, combined with the signal propagation attenuation model and the Doppler frequency shift model; and a fusion correction unit (103), which is used to fuse the initial window prediction result with the path loss and the Doppler frequency shift to obtain the corrected connected window parameters.

[0068] Dependency building module (200): Used to obtain GIS data metadata of the current user’s region and build a data dependency graph based on the GIS data metadata. The data dependency graph includes multiple data block nodes and multiple directed edges. The directed edges are used to represent the dependency relationships between data block nodes.

[0069] The dependency construction module (200) specifically includes: a metadata parsing unit (201), used to parse GIS data metadata and extract the unique identifier, data type, dependency identifier and timeliness label of each data block; a directed graph construction unit (202), used to construct an initial directed graph with the unique identifier of each data block as the node and the dependency identifier as the pointing relationship of the directed edge; a circular dependency decoupling unit (203), used to perform circular dependency detection on the initial directed graph, and when a circular dependency is detected, decoupling is performed by introducing virtual nodes to generate an acyclic data dependency graph; and a weight initialization unit (204), used to set the dynamic weight of each data block according to the timeliness label, and when the validity period of a data block is less than the duration of the connected window, the filtering priority of the data block in the smallest independent available unit is reduced.

[0070] Decision module (300): It is used to input the connectivity window parameters and data dependency graph into the pre-trained decision model. Under the constraint of estimated bandwidth, the decision model selects at least one minimum independent available unit from multiple data block nodes based on the integrity of the dependency relationship in the data dependency graph. The minimum independent available unit refers to a data set that consists of at least one data block node and all its dependent preceding data block nodes, which can independently support the preset GIS functions in offline state.

[0071] The decision module (300) specifically includes: a state encoding unit (301) for encoding the connected window parameters and the data dependency graph into a state vector; a policy output unit (302) for outputting an action probability distribution based on the state vector through the policy network, the action probability distribution being used to select data block nodes from candidate data block nodes to join the minimum independent available unit set; and a reward calculation unit (303) for guiding the learning of the policy network through a reward function, the reward function being constructed based on dependency integrity, the ratio of the total data block volume to the estimated bandwidth, and the ratio of the data block acquisition delay to the duration.

[0072] Preload execution module (400): When the start time of the connection window arrives, according to the estimated bandwidth constraint, it obtains the data block corresponding to the smallest independent available unit and stores the obtained data block in the local cache of the mobile terminal.

[0073] The preloading execution module (400) specifically includes: a behavior heat calculation unit (401), used to obtain the current user's geographic location information, current viewport range, and current user's behavior heat value; a weighted sorting unit (402), used to perform weighted sorting on the data blocks in the smallest independent available unit according to the geographic location information, current viewport range, and behavior heat value, and generate a preloading queue; and a data acquisition unit (403), used to send data requests to the GIS data server in sequence according to the order of the preloading queue within the duration of the connected window, and obtain the corresponding data blocks.

[0074] The modules work together to achieve efficient preloading under extreme pulsed network conditions.

[0075] This embodiment briefly illustrates the training process of the connected window prediction model and the decision model.

[0076] Connectivity window prediction model training: A large amount of motion trajectory data and network connectivity pattern data from mobile devices in various scenarios (high-speed rail, subway, mountainous areas, cities) are collected. The start time, duration, and bandwidth of the connectivity window are manually labeled. The motion trajectory data and historical network connectivity pattern data are used as input, and the labels are used as output. A temporal convolutional network (TCN) is trained using supervised learning. The physical motion model does not require training; its parameters are preset based on the physical environment (e.g., path loss index). (Set according to the scenario type). After training, the TCN output is fused with the physical model output during actual prediction.

[0077] Training the deep reinforcement learning decision model: A simulated environment is constructed that can return the next state and reward based on the current state (connectivity window parameters, data dependency graph) and the agent's chosen action. Algorithms such as Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) are used to train the policy network and value network. During training, the agent continuously explores different data selection strategies, learning to minimize the overall volume while maintaining dependency integrity, and avoiding exceeding bandwidth and time budgets through reward signals. After training, the policy network can be directly used for online decision-making.

[0078] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A deep learning-based mobile GIS data preloading method, characterized in that, Includes the following steps: (1) Obtain motion trajectory data and historical network connectivity pattern data of mobile terminal device, input the motion trajectory data and the historical network connectivity pattern data into a pre-trained connectivity window prediction model, and predict the connectivity window parameters within a preset time period through the connectivity window prediction model. The connectivity window parameters include the start time, duration and estimated bandwidth of the connectivity window. (2) Obtain GIS data metadata of the current user’s location, and construct a data dependency graph based on the GIS data metadata. The data dependency graph includes multiple data block nodes and multiple directed edges, and the directed edges are used to represent the dependency relationships between the data block nodes. (3) Input the connected window parameters and the data dependency graph into the pre-trained decision model. Under the constraint of the estimated bandwidth, the decision model selects at least one minimum independent available unit from the multiple data block nodes, with the integrity of the dependency relationship in the data dependency graph as the condition. The minimum independent available unit refers to a data set that can independently support the preset GIS function in offline state, consisting of at least one data block node and all its dependent preceding data block nodes. (4) When the start time of the connection window arrives, according to the estimated bandwidth constraint, the data block corresponding to the smallest independent available unit is obtained, and the obtained data block is stored in the local cache of the mobile terminal. 2.The deep learning-based mobile GIS data preloading method of claim 1, wherein, The connected window prediction model is a fusion of a temporal convolutional network and a physical motion model. In step (1), the connected window parameters are predicted using the connected window prediction model, including: The temporal convolutional network extracts temporal features from the motion trajectory data and the historical network connectivity pattern data, and outputs the initial window prediction result; the physical motion model calculates the path loss based on the position and velocity information in the motion trajectory data, combined with the signal propagation attenuation model. ,in For path loss, For reference distance Path loss at the location, This is the path loss index. This refers to the real-time distance between the mobile device and the base station. The shadow fading random variable is used in conjunction with the Doppler frequency shift model. Calculate the Doppler frequency shift, where This is the Doppler frequency shift. For the movement speed of mobile devices, The angle between the direction of motion and the direction of signal incidence. For the signal wavelength, The carrier frequency is used as the reference. The initial window prediction result is fused with the path loss and the Doppler frequency shift to obtain the corrected connectivity window parameters. 3.The deep learning-based mobile GIS data preloading method of claim 1, wherein, The data dependency graph is a dynamic data dependency graph, which also includes a dynamic weight matrix. This dynamic weight matrix characterizes the decay factor of the dependency strength of each directed edge over time. The decay factor is expressed by the formula... Calculation, where For a moment Dependence strength For the initial dependency strength, The attenuation coefficient is... This is the moment when dependencies are generated. 4.The deep learning-based mobile GIS data preloading method of claim 1, wherein, Step (2) involves constructing a data dependency graph based on the GIS data metadata, including: Parse the GIS data metadata to extract the unique identifier, data type, dependency identifier, and timeliness label for each data block; construct an initial directed graph using the unique identifier of each data block as a node and the dependency identifier as the pointing relationship of directed edges; perform cyclic dependency detection on the initial directed graph, and when cyclic dependency is detected, decoupling is performed by introducing virtual nodes to generate an acyclic data dependency graph; set a dynamic weight for each data block according to the timeliness label, and reduce the filtering priority of the data block in the smallest independent available unit when the validity period of the data block is less than the duration of the connected window. 5.The deep learning-based mobile GIS data preloading method of claim 1, wherein, The decision model is a deep reinforcement learning model. In step (3), at least one minimum independent and usable unit is selected through the decision model, including: The connected window parameters and the data dependency graph are encoded into a state vector. ; through the policy network based on the state vector Output action probability distribution The actions The decision to select a data block node from candidate data block nodes to join the set of minimum independent available units; through a reward function. The learning of the policy network is guided, in which For indicator functions, when The value is 1 when dependency integrity is satisfied, and the volume is 1. For the present Total volume, Latency To obtain Estimated latency for all data blocks in the middle. For the estimated bandwidth, The duration is... , , These are the preset weighting coefficients. 6.The deep learning-based mobile GIS data preloading method of claim 1, wherein, The minimum independent available unit satisfies the condition. and ,in For the set of the smallest independent available units, For data blocks volume, For data blocks The set of all preceding dependent data blocks. 7.The deep learning-based mobile GIS data preloading method of claim 1, wherein, Step (4) involves obtaining the data block corresponding to the smallest independent available unit, including: The system obtains the current user's geographic location information, current viewport range, and current user's behavioral heatmap value. The behavioral heatmap value is obtained using a formula... Calculation, where For users in the first The frequency of accesses to the current region within a historical time period The weighting coefficients are related to the time distance of the current period; based on the geographical location information, the current viewport range, and the behavior heatmap value, the data blocks in the smallest independent available unit are weighted and sorted to generate a preloading queue. The weighting coefficients for the weighted sorting are determined by the formula... Calculation, where For data blocks The distance between the corresponding spatial area and the user's current location. For data blocks Predicted access probability, For data blocks Data freshness, , , The preset weighting coefficients are used; within the duration of the connected window, the corresponding data blocks are acquired in the order of the preloaded queue.

8. A mobile GIS data preloading system based on deep learning, characterized in that, The system, which is applied to the deep learning-based mobile GIS data preloading method according to any one of claims 1-7, comprises: The connectivity window prediction module (100) is used to acquire motion trajectory data and historical network connectivity pattern data of the mobile device, input the motion trajectory data and the historical network connectivity pattern data into a pre-trained connectivity window prediction model, and predict the connectivity window parameters within a preset time period through the connectivity window prediction model. The connectivity window parameters include the start time, duration and estimated bandwidth of the connectivity window. The dependency construction module (200) is used to obtain GIS data metadata of the current user's location and construct a data dependency graph based on the GIS data metadata. The data dependency graph includes multiple data block nodes and multiple directed edges, and the directed edges are used to represent the dependency relationships between the data block nodes. The decision module (300) is used to input the connectivity window parameters and the data dependency graph into a pre-trained decision model. Under the constraint of the estimated bandwidth, the decision model selects at least one minimum independent available unit from the multiple data block nodes, based on the integrity of the dependency relationship in the data dependency graph. The minimum independent available unit refers to a data set that consists of at least one data block node and all its dependent preceding data block nodes, and can independently support the preset GIS functions in offline mode. The preload execution module (400) is used to obtain the data block corresponding to the smallest independent available unit according to the estimated bandwidth constraint when the start time of the connection window arrives, and store the obtained data block to the local cache of the mobile terminal. 9.The deep learning based mobile GIS data preloading system of claim 8, wherein, The connected window prediction model is a fusion of a temporal convolutional network and a physical motion model, and the connected window prediction module (100) is... include: The temporal feature extraction unit (101) is used to extract the temporal features of the motion trajectory data and the historical network connectivity pattern data through the temporal convolutional network, and output the initial window prediction result; the physical attenuation calculation unit (102) is used to calculate the path loss and Doppler frequency shift based on the position and velocity information in the motion trajectory data, combined with the signal propagation attenuation model and the Doppler frequency shift model, through the physical motion model; the fusion correction unit (103) is used to fuse the initial window prediction result with the path loss and the Doppler frequency shift to obtain the corrected connectivity window parameters. 10.The deep learning based mobile GIS data preloading system of claim 8, wherein, The decision model is a deep reinforcement learning model, and the decision module (300) includes: A state encoding unit (301) is used to encode the connected window parameters and the data dependency graph into a state vector; a policy output unit (302) is used to output an action probability distribution through a policy network based on the state vector, the action probability distribution being used to select data block nodes from candidate data block nodes to join the minimum independent available unit set; a reward calculation unit (303) is used to guide the learning of the policy network through a reward function, the reward function being constructed based on dependency integrity, the ratio of the total data block volume to the estimated bandwidth, and the ratio of the data block acquisition delay to the duration.

Images