Beidou-based edge model optimization and collaboration method and terminal

Abstract

The invention provides a Beidou-based edge model optimization and collaboration method and a terminal. The Beidou-based edge model optimization and collaboration method comprises the following steps: acquiring current geographical environment characteristics of current equipment through a Beidou positioning system; determining an adaptive target calculation model based on the current geographical environment characteristics; the current device and at least one target device form a device network based on the real-time position in the current geographical environment feature, collaborative optimization of a target calculation model is carried out based on the device network, and a model library is updated according to a collaborative optimization result; and performing predictive model deployment based on the movement track prediction of the current equipment and the model library. According to the method, edge model dynamic adaptation driven by geographical environment perception, multi-device efficient collaborative learning based on position weighting and predictive model deployment oriented to mobile continuous service are realized, and the reasoning performance, the collaborative efficiency and the service continuity of an edge intelligent system in a complex real environment are improved.

Description

Technical Field

[0001] This invention relates to the fields of edge computing and artificial intelligence technology, and in particular to an edge model optimization and collaboration method and terminal based on BeiDou. Background Technology

[0002] With the convergence of edge computing and artificial intelligence technologies, edge AI systems primarily employ static configuration methods for model selection and deployment, lacking adaptability to geographical environments and application scenarios. Traditional edge AI inference systems typically use fixed AI models, failing to dynamically optimize based on the characteristics of different geographical environments, leading to compromises inference accuracy and efficiency in certain specific environments. The model collaboration mechanisms of related technologies are mainly based on computing resource allocation, neglecting the impact of geographical location factors on model performance, thus failing to achieve true intelligent collaboration. When mobile devices move between different geographical environments, the lack of predictive model deployment mechanisms easily leads to model switching delays and service interruptions. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a BeiDou-based edge model optimization and collaboration method and terminal, which improves the inference performance and adaptability of edge systems through intelligent selection of geographic perception models, geographic weighted collaboration and predictive model deployment.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: An edge model optimization and coordination method based on BeiDou includes: The current geographical environment characteristics of the device are obtained through the BeiDou positioning system; Based on the current geographical environment characteristics, determine the appropriate target calculation model; Based on the real-time location in the current geographic environment features, the current device and at least one target device form a device network, and the target computing model is collaboratively optimized based on the device network, and the model library is updated according to the optimization results; Based on the current device's movement trajectory prediction, predictive model deployment is performed using the model library.

[0005] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A terminal for edge model optimization and coordination based on BeiDou includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the various steps in the aforementioned edge model optimization and coordination method based on BeiDou.

[0006] The beneficial effects of this invention are as follows: By using the current geographical environment features acquired by the BeiDou system as the core driver, dynamic model adaptation is achieved, providing a precise perception foundation for subsequent optimization. The optimal computing model is intelligently selected for different environments, overcoming the limitations of static model configuration or reliance solely on hardware resources in traditional edge systems, thus improving the accuracy and efficiency of edge inference. Through collaborative optimization of the target computing model based on current geographical environment features, this collaborative mechanism, based on geographical and environmental similarity rather than random or full-scale network construction, ensures high consistency in data distribution and task objectives among participating devices. This allows devices with similar geographical environment features and tasks to efficiently share knowledge, enhancing the regional generalization ability and collaborative training efficiency of the target computing model. By combining a model pre-deployment mechanism with mobile trajectory prediction, seamless and continuous service switching is achieved when devices cross different geographical regions, eliminating delays and interruptions caused by model loading. Deployment complexity is reduced, effectively solving the problems of unstable model performance, low collaborative efficiency, and discontinuous mobile services caused by the lack of environmental perception in traditional edge systems. Attached Figure Description

[0007] Figure 1 A flowchart illustrating the steps of a BeiDou-based edge model optimization and coordination method provided in this embodiment of the invention; Figure 2 This is a schematic diagram illustrating the working principle of an edge model optimization and collaboration method based on BeiDou provided in an embodiment of the present invention. Figure 3 A schematic diagram of the structure of a terminal based on BeiDou edge model optimization and collaboration is provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a system for edge model optimization and coordination based on BeiDou, provided in an embodiment of the present invention. Label Explanation: 1. A system for edge model optimization and coordination based on BeiDou; 2. Processor; 3. Memory. Detailed Implementation

[0008] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0009] Please refer to Figure 1 An edge model optimization and collaboration method based on BeiDou includes steps 110 to 140.

[0010] Step 110: Obtain the current geographic environment features of the device through the BeiDou positioning system. For example, the device obtains the current geographic environment features through the integrated BeiDou positioning system module, constructing a multi-dimensional feature set that comprehensively describes the scene in which the device is located.

[0011] Step 120: Based on the current geographic environment characteristics, determine the appropriate target computing model. For example, based on the acquired current geographic environment characteristics, the system identifies the target computing model that is suitable for those characteristics.

[0012] Step 130: Based on the real-time location in the current geographic environment features, establish a device network between the current device and at least one target device, and perform collaborative optimization of the target computing model based on the device network. Update the model library according to the results of the collaborative optimization. For example, the system takes the real-time location in the current geographic environment features of the current device as the center, actively discovers and invites surrounding devices with similar geographic environment features to jointly form a device network, performs collaborative optimization of the target computing model based on the device network, and updates the model library according to the results of the collaborative optimization.

[0013] Step 140: Deploy a predictive model based on the current device's movement trajectory prediction and the model library. For example, the system deploys a predictive model in the model library based on the current device's movement trajectory prediction, thereby achieving seamless switching from the current model to the new model and ensuring the continuity and stability of the task.

[0014] As described above, the beneficial effects of this invention are as follows: By constructing a system driven by BeiDou high-precision geographic information, the current geographic environment features enhanced by the BeiDou system provide a precise perception foundation for all subsequent optimizations. Through dynamic model selection based on current geographic environment features, it overcomes the limitations of static model configuration or reliance solely on hardware resources in traditional edge systems, enabling intelligent matching of optimal computational models for different environments. This improves the accuracy and efficiency of edge inference. By constructing a device network based on real-time location and performing geographically weighted collaborative optimization, devices with similar geographical locations and task scenarios can be efficiently organized and share knowledge, improving the efficiency of model collaborative training and regional generalization ability. Simultaneously, the collaborative optimization results are fed back to the model library, driving its continuous evolution. Through proactive model deployment based on mobile trajectory prediction, the required models can be pre-loaded before the device arrives in a new environment, achieving seamless service switching during device movement and fundamentally eliminating service interruptions caused by model switching delays. This invention solves the technical problems of unstable model performance, low collaborative efficiency, and discontinuous mobile services caused by the lack of environmental perception in traditional edge systems.

[0015] Further, step 120 includes steps 121 and 122.

[0016] Step 121: Based on the current geographic environment features, query the feature database for the performance of the current model that matches the current geographic environment features. The feature database stores the correlation between geographic environment features and model performance. For example, after obtaining the current geographic environment features, the system performs matching and retrieval of the current geographic environment features in the feature database to find the model that performs best under the current geographic environment features and its specific performance record. This feature database not only stores data from macro to micro levels, but also records and correlates historical performance data of different computing models under various geographic environment features through continuous performance monitoring. Step 122: Determine the target computational model based on the current geographic environment characteristics and the current model performance. For example, the system inputs the current geographic environment characteristics and the current model performance set into a pre-trained geographic perception decision model. This decision model, such as a dynamic selection decision tree built based on reinforcement learning, comprehensively analyzes the uniqueness of the current geographic environment and the performance, resource requirements, and generalization ability of each candidate model, performs multi-objective trade-off evaluation, and calculates the comprehensive adaptability score of each candidate model for the current task. Finally, the system dynamically selects the candidate model with the highest comprehensive score and deploys it as the target computational model to perform the current edge inference task.

[0017] As described above, this invention constitutes a geographic perception model selection mechanism that combines data-driven and intelligent decision-making, overcoming the limitations of traditional solutions based on rules or simple historical best-fit matching. By querying the feature database to find the current model performance that matches the current geographic environment features of the current device, the system demonstrates its ability to perform preliminary screening using massive amounts of historical experience. Furthermore, through intelligent decision-making based on the current geographic environment features and the current model performance, the system further achieves adaptive judgment capabilities for complex, dynamic, and even unseen geographic environments. This not only ensures high accuracy in model selection but also enhances the system's robustness, thus laying a core foundation for the performance and efficiency of the entire edge intelligent system.

[0018] Furthermore, steps 123 to 126 are included before step 120.

[0019] Step 123: Record the historical performance data of different computing models under different multi-scale geographic environmental characteristics. For example, during long-term operation, the system continuously collects and structures and stores the actual historical performance data of each computing model under different geographic environmental characteristics.

[0020] Step 124: Based on historical performance data, train an evaluation model to assess the adaptability of the computational model to multi-scale geographic environmental features. For example, the system uses machine learning techniques, with the historical performance database as the training set, to learn the complex nonlinear mapping relationship between geographic environmental features and model performance indicators. The evaluation model adopts a gradient boosting decision tree model. The input of this model includes: multi-scale geographic environmental feature vectors and the identification information of candidate computational models. The multi-scale geographic environmental feature vectors include micro-features (terrain slope, building density), meso-features (road type, traffic flow), and macro-features (climate conditions, time period information). The output is the adaptability score of the computational model in the corresponding geographic environment, with a score range of 0 to 1. The training process includes: dividing historical performance data into training and validation sets according to a preset ratio; using geographic environmental feature vectors and model identification information as input features, and normalized model performance metrics as labels for training, wherein the model performance metrics include a weighted combination of inference accuracy, inference latency, and resource utilization; iteratively training a decision tree through multiple rounds to enable the evaluation model to learn the mapping relationship between geographic environmental features and model performance; evaluating model performance on the validation set, and completing training when the prediction error meets a preset threshold. The trained evaluation model can predict the performance of each candidate computing model in the input geographic environmental features, thereby providing a quantitative basis for model selection.

[0021] Step 125: Extract current multi-scale geographic environment features from the current geographic environment features. For example, when it is necessary to select a model for the current device, the system first processes the acquired raw current geographic environment features. Using a multi-scale analytical framework, the system extracts standardized micro, meso, and macro feature vectors from the raw information. This step ensures that the data format input to the evaluation model is consistent with the training data, and is a key preprocessing step to ensure the accuracy of the evaluation.

[0022] Step 126: Input the current multi-scale geographic environment features into the evaluation model to obtain the target computational model. For example, the system inputs the extracted standardized feature vectors into the trained evaluation model. Based on the mapping relationships learned internally, the evaluation model quickly calculates a list of adaptability scores for all candidate computational models for the current environment. The system automatically determines and outputs the final target computational model from this adaptability score list according to a preset decision-making strategy.

[0023] As described above, this invention achieves intelligent edge model selection by constructing an adaptive decision-making system for geographic perception models that combines offline training and online evaluation. The system trains an intelligent evaluation kernel offline based on historical data, accurately modeling the adaptation relationship between the model and the geographic environment. In the online phase, it performs standardized analysis of real-time environmental characteristics and uses the evaluation kernel to quickly make decisions, dynamically determining the optimal model. This architecture design resolves the contradiction between traditional systems' reliance on static rules and the difficulty in balancing experience utilization and real-time response. The standardized and quantifiable optimization benchmark it provides lays a unified and efficient technical foundation for subsequent collaborative optimization, cache updates, and predictive deployment, thereby systematically improving the decision-making accuracy and response agility of edge intelligence in complex geographic scenarios.

[0024] Further, step 130 performs collaborative optimization of the target computation model based on the device network, and updates the model library according to the results of the collaborative optimization, including steps 131 to 133.

[0025] Step 131: Calculate the collaborative weight of each device based on the geographical location relationships among all devices constituting the device network. For example, the system calculates the geographical distance between any two devices in the network based on the precise location of the devices obtained from the BeiDou positioning system. Simultaneously, it calculates the environmental similarity between devices by comparing environmental feature vectors of their environments, such as terrain type, building density, and climate data. Subsequently, the system uses a dynamic weighting function to comprehensively calculate a unique collaborative weight for each device.

[0026] Step 132: Based on the collaborative weights, perform federated learning in the device network to collaboratively optimize the target computation model. For example, the optimization process of federated learning includes: First, the central server distributes the current global model parameters to each device in the device network; second, each device, based on locally collected geographic environment data and task data, uses the gradient descent algorithm to train the received model parameters locally, performs a preset number of iterative updates, and calculates the local model parameter update amount; then, each device uploads the local model parameter update amount to the central server; on the server side, the system does not simply average the update amount of all devices, but rather performs weighted aggregation of the received model parameter update amounts according to the collaborative weights of each device calculated in step 131. Specifically, the global model parameter update amount is equal to the sum of the products of the local model parameter update amounts of each device and their corresponding collaborative weights; finally, the central server applies the weighted aggregated global model parameter update amount to the global model to obtain the optimized target computation model. Through multiple iterations of the above process, until the performance indicators of the global model on the validation dataset converge or reach the preset number of training rounds, the collaborative optimization is completed. In this process, devices that are geographically close and have similar environments have higher collaborative weights, and their local training results have a greater impact on the global model, thus making the optimized model more suitable for the current geographical environment.

[0027] Step 133: Update the model library based on the results of collaborative optimization and the current geographical characteristics of the devices. For example, the new model version generated by collaborative optimization, along with a summary of the geographical distribution characteristics of the device network on which it depends, is submitted to the model library as a complete knowledge unit, and the model library is updated.

[0028] As described above, this invention implements a geographically-aware collaborative learning and knowledge accumulation mechanism. Through refined, multi-dimensional geographic relationship calculations, it introduces prior spatial and environmental knowledge into federated learning, improving the efficiency of collaborative training and the relevance of regional models. The system performs geographically weighted model aggregation based on this prior knowledge, ensuring consistency between priority directions and geographic tasks. Finally, the results of collaborative optimization are sent to the model library in a structured, versioned form, thus forming a complete closed loop from distributed training to geographically weighted aggregation.

[0029] Furthermore, it also includes step 134.

[0030] Step 134: Based on the geographical environment characteristics of each device in the device network and the model call frequency of all computing models in the device, obtain the caching and distribution strategy for each computing model. For example, the system formulates differentiated caching and distribution strategies for each computing model, specifically including: First, statistically analyzing the call frequency of each computing model in the device network within a preset time window, and combining the geographical environment characteristics of each device to analyze the spatiotemporal distribution characteristics of model calls; then, determining the caching strategy and distribution strategy based on the analysis results. The caching strategy includes: for high-frequency models with a call frequency higher than a first threshold, a network-wide caching strategy is adopted, caching the model to all devices or edge nodes in the device network; for medium-frequency models with a call frequency between the first and second thresholds, a regional caching strategy is adopted, caching the model to device groups within geographical areas with higher call frequencies; for low-frequency models with a call frequency lower than the second threshold, an on-demand caching strategy is adopted, loading only temporarily when a device requests it. The distribution strategy includes: for device clusters with similar geographical characteristics, multicast distribution is used to synchronously distribute the collaboratively optimized model to all devices within the cluster; for geographically dispersed devices, unicast distribution is used to selectively distribute models based on the network status and storage space of each device; for newly added devices, incremental distribution is used to prioritize the distribution of highly adaptable models corresponding to their current geographical characteristics. Through these strategies, the system achieves intelligent scheduling and efficient utilization of model resources.

[0031] As described above, by integrating geographical environmental features, model invocation behavior, and network resource status, an intelligent caching and distribution system with spatiotemporal awareness and dynamic adaptability has been constructed. This system not only achieves accurate device clustering from "physical proximity" to "task similarity," but also significantly reduces redundant transmission and storage overhead through hierarchical caching and targeted distribution mechanisms, improving the timeliness of model supply and service continuity. This results in a systematic optimization of resource efficiency and inference performance in complex and dynamic edge environments.

[0032] Furthermore, step 133 includes steps 1341 and 1342.

[0033] Step 1341: Obtain device clusters similar to the current geographical environment features of the current device. For example, the system maintains a dynamic device topology and feature map. When the model library needs to be updated, the system uses the current geographical environment features of the current device as the query benchmark and performs a nearest neighbor search in the map. The search comprehensively considers two types of key similarities: spatial proximity, i.e., whether the geographical distance between devices is within a set threshold; and environmental homogeneity. The system automatically classifies devices that simultaneously meet the dual conditions of spatial proximity and environmental similarity into a logically high-similarity device cluster. The definition of this cluster is dynamic and is adjusted in real time as devices move or the environment changes.

[0034] Step 1342: Share the target computing model within the device cluster. For example, after confirming the target cluster, the system initiates an efficient model distribution protocol. It proactively and preferentially pushes the new version of the target computing model generated by collaborative optimization, along with necessary model metadata, to all devices within the cluster or edge nodes serving the cluster. This sharing is not a simple broadcast, but rather employs a differentiated distribution strategy based on the computing power, storage space, and network status of the devices within the cluster. Simultaneously, the system establishes a model version synchronization mechanism within the cluster to ensure that all members hold consistent model versions and supports conflict detection and resolution.

[0035] As described above, this invention intelligently clusters devices by integrating multi-dimensional similarity measures of spatial proximity and environmental homogeneity. This accurately identifies highly similar device groups that can benefit most from optimizing the same model, achieving a precise mapping from physical proximity to task similarity. The system then efficiently and purposefully pushes the collaboratively optimized model to the target cluster. This process not only reduces network and storage overhead caused by network-wide broadcasting but also fundamentally improves the overall resource utilization, service consistency, and response efficiency and execution accuracy of common regional tasks through precise model-environment matching and version unification within the cluster.

[0036] Furthermore, step 1342 includes step 1343.

[0037] Step 1343: Obtain shared computing models with call frequencies exceeding a threshold in the device cluster, and cache these shared computing models in device groups within the device cluster. Device groups include multiple geographically proximate devices. For example, the system continuously monitors real-time access metrics for all shared models within the device cluster, including but not limited to the number of calls per unit time, the average number of concurrent response requests, and the coverage rate of the model requested by different devices. When the overall call frequency of a shared computing model within the cluster consistently exceeds a dynamically adjusted intelligent threshold, the system determines that the model is a high-hot model. Subsequently, the system no longer relies solely on the on-demand retrieval sharing mode but proactively implements a hot caching strategy: it first analyzes the geographical distribution of devices within the cluster, identifying one or more subsets of devices with the most concentrated geographical locations and the best network interconnection quality, forming device groups. Then, the system pre-caches the full or lightweight version of the high-hot model to the core devices within this group or the edge servers serving them.

[0038] As described above, the introduction of a caching decision layer based on usage frequency represents a significant enhancement to the geographic awareness sharing mechanism. Through real-time monitoring and threshold determination, it dynamically identifies knowledge hotspots within the cluster and, based on the principle of geographic concentration, proactively allocates these hotspots to the most suitable physical locations. This strategy transforms cache resources from static allocation to dynamic allocation following business hotspots. It not only further reduces network latency and back-to-origin bandwidth pressure for high-frequency model calls but also significantly improves the service capacity and response speed of the entire device cluster for sudden, high-concurrency regional tasks by pre-positioning computing power in demand-intensive areas. This is a key optimization step in maximizing the performance of the model sharing mechanism.

[0039] Further, step 140 includes steps 141 to 143.

[0040] Step 141: Analyze the device's historical movement trajectory and predict its future movement trajectory based on this trajectory. The future movement trajectory includes the target location and its time information. For example, the system continuously collects and stores time-stamped location sequences reported by the device through the BeiDou positioning system, forming high-precision historical movement trajectory data. Based on this data, the system uses time series analysis or machine learning prediction models to comprehensively analyze the device's movement patterns, path preferences, speed patterns, and historical destinations. The prediction model not only outputs one or more possible future target locations but also simultaneously outputs the estimated time window for the device to arrive at each target location.

[0041] Step 142: Based on the predicted geographic environmental characteristics of the target location, match computational models adapted to the predicted geographic environmental characteristics from the model library. For example, for each predicted target location, the system combines GIS and real-time environmental data sources to proactively deduce the predicted geographic environmental characteristics of that location at the expected arrival time of the equipment. Subsequently, the system uses these predicted characteristics as query conditions to perform intelligent retrieval and matching in the model library. The system identifies those adapted computational models from the model library whose historical performance data indicates optimal performance in similar predicted environments, forming a list of pre-deployed models for this future journey.

[0042] Step 143: Preload the matched computational model to the edge node corresponding to the current device or target location. For example, the system formulates a phased and target-specific preloading plan based on the predicted arrival time window, the model size, and the network conditions between the current device and the target area edge nodes. For nearby trips or small models, the system may choose to preload them directly into the local cache of the mobile device. For longer trips, large models, or computationally intensive models, the system is more likely to pre-deploy the model on the edge server node with the lowest network latency to the target location.

[0043] As described above, this invention constructs a forward-looking service assurance mechanism that transforms mobility from a challenge into a plannable known quantity, the future environment into a pre-prepared computing requirement, and resource deployment from reactive to proactive. This interconnected mechanism achieves end-to-end automation from movement trajectory prediction to computing resource pre-provisioning. Its core value lies in advancing the service interruption time window to the background preparation stage, which is imperceptible to the user, thereby achieving absolute continuity of services across geographical regions at the user experience level.

[0044] Furthermore, it also includes steps 150 and 151.

[0045] Step 150: Collect performance data of the target computing model on the current device. For example, after the target computing model is deployed on the current device and executes inference tasks, the system's built-in performance monitoring module will comprehensively and in real-time collect multi-dimensional data about its runtime. This includes core performance metrics, resource consumption metrics, and contextual metrics. Simultaneously, the system will also collect implicit or explicit user feedback data through preset interfaces.

[0046] Step 151: Based on operational performance data, optimize the adaptation mechanism and collaborative optimization strategy between the target computing model and the current geographic environment features. For example, for optimizing the model adaptation mechanism, the system uses the collected operational performance data as new samples to incrementally update the evaluation model trained in step 124. Specifically, the current geographic environment features, the selected model identifier, and the actual operational performance indicators constitute new training samples. The evaluation model is fine-tuned through an online learning algorithm to update the model parameters, making its evaluation of the current environmental features more accurate. When the accumulated number of new samples reaches a preset threshold, the evaluation model is retrained to more accurately predict the model's performance under new emerging environmental modes. To optimize the collaborative optimization strategy, the system evaluates the effectiveness of the current collaborative weight allocation strategy based on operational performance data. Specifically, this includes: first, statistically analyzing the actual performance of the optimized model on each device and calculating the performance improvement rate; then, analyzing devices with poor performance and extracting their geographical environment characteristics, collaborative weights, and geographical relationships with other devices; next, identifying unreasonable weight allocation patterns, such as assigning excessively low weights to devices with similar environments but large geographical distances, or assigning excessively high weights to devices with similar geographical locations but large environmental differences; finally, adjusting the weight calculation function in step 131 based on the identification results, including adjusting the geographical distance decay coefficient, environmental similarity weight coefficient, or optimizing the similarity threshold for device clustering, so that the weight allocation is more reasonable in the next collaborative learning iteration, resulting in a model with stronger generalization ability. Furthermore, the system dynamically adjusts the training parameters of federated learning based on operational performance data, including the number of local training rounds, learning rate, and aggregation frequency, to improve the efficiency and effectiveness of collaborative optimization.

[0047] As described above, by transforming each service into an experiment and accurately recording the experimental data, and using this data to feed back into and calibrate the core decision-making modules of the system, a complete closed loop is formed, from "perception, decision-making, and execution" to "evaluation, feedback, and optimization." This closed loop enables the system to continuously adapt to long-term changes in the geographical environment, the evolution of model technology, and the migration of user needs, thereby maintaining and improving its overall performance, robustness, and user satisfaction in complex edge scenarios over the long term. This represents a qualitative leap from one-time optimization to lifelong learning and adaptation.

[0048] Please refer to Figure 2 The following are application embodiments of the present invention, specifically applying the above solution to edge intelligence real-time optimization scenarios for autonomous vehicles in real road environments, particularly in the fields of advanced driver assistance systems (ADAS) and autonomous driving that require high environmental adaptability, low-latency decision-making, and continuous cross-regional services. Taking the operation of an autonomous driving company's urban fleet as an example, its vehicles are equipped with an edge intelligence system based on the present invention, and the specific implementation includes the following steps: S1. Obtain the precise geographical location and environmental feature information of the device through the BeiDou positioning system. The system acquires the current geographical environment features through the BeiDou high-precision positioning module and environmental sensors integrated into the vehicle. When the vehicle is driving on urban roads, the system acquires centimeter-level location information in real time through the BeiDou system, and extracts geographical environment feature vectors by combining high-precision maps and real-time vehicle network data. By combining GIS data to obtain environmental information, the system analyzes the multi-scale geographical environment features of the current environment, including: micro-features, meso-features, and macro-features. The system receives the inference task of "execute real-time obstacle perception and path planning" issued by the cloud. This is equivalent to step 110 above.

[0049] S2. Establish a geographic perception AI model selection mechanism. Based on the current geographic environment characteristics, construct a geographic environment feature database and establish a correlation mapping between location and model performance. In specific implementation, the system dynamically determines the suitable target computing model, and inputs the above feature combination into the locally maintained feature database for query matching. Database records show that in the "evening rush hour light rain in a commercial area" environment, by designing a geographic feature-based model evaluation algorithm, it was found that the lightweight model "Model_A" had the lowest false alarm rate under raindrop interference, while the high-precision model "Model_B" had the lowest false negative rate under complex pedestrian flow. By creating a dynamic model selection decision tree, the geographic perception evaluation model comprehensively weighs real-time computing resources and task safety level, and calculates that "Model_B" has the highest comprehensive adaptation score. Therefore, the system dynamically loads and activates "Model_B" as the "target computing model" for currently performing obstacle perception. This is equivalent to steps 120 and 121 to 122 above.

[0050] S3. The system automatically selects the most suitable AI inference model based on different geographical environmental characteristics. It analyzes the geographical environmental characteristics of the current location and matches them with historical model data. In actual collaborative optimization scenarios, the system forms a temporary device network with neighboring vehicles based on real-time location and performs model collaborative optimization. Vehicles enter a parking area consisting of multiple autonomous vehicles equipped with similar systems. Each vehicle automatically networks via C-V2X communication, forming a "parking lot low-speed patrol device network." Within the network, the system assigns nearly equal high collaborative weights to each vehicle based on their extremely close proximity and identical environmental characteristics. Subsequently, each vehicle uses locally collected unique obstacle data to perform federated learning training on a shared copy of the target computation model. After several rounds of local training and weighted aggregation, the network collaboratively optimizes a new model version that significantly improves the recognition accuracy of "small static obstacles indoors." After optimization, the new model and its applicable "indoor parking lot" environmental feature summary are uploaded to the regional edge server and updated to the global model library. This is equivalent to steps 130 and 131 to 133 above.

[0051] S4. Implement AI model collaborative learning and intelligent caching strategies among multiple devices. The system establishes a communication and collaboration network between devices and implements a federated learning algorithm with geographical weights. Based on vehicle trajectory prediction, the system deploys predictive models. A vehicle plans to leave the parking lot and head to a center 10 kilometers away. The system analyzes historical travel data to predict its future trajectory: passing through an urban expressway (target location A, expected arrival in 5 minutes), and finally arriving at the underground parking lot of a center (target location B, expected arrival in 25 minutes). The system then proactively queries the model library: for the "urban expressway light rain" environment, the optimal model is "Model_C," which focuses on long-distance vehicle recognition; for the "underground parking lot of a center" environment, the optimal model is the recently co-optimized "Indoor Obstacle Recognition Enhanced Model_B." Therefore, while the vehicle is still in the parking lot, the system preloads "Model_C" into the onboard computing unit via the 5G network for later use; simultaneously, it sends a command to the edge server of the convention center to preheat "Model_B" into the computing cache. This is equivalent to steps 140 and 141 to 143 above.

[0052] S5. Based on movement trajectory prediction, a predictive model is deployed. The system collects model performance data and provides feedback to optimize the core mechanism. During the expressway driving phase, system monitoring revealed that "Model_C" had a 15ms higher delay in recognizing a suddenly entering motorcycle than expected. This performance data, along with a specific environmental snapshot, was recorded and encrypted before being uploaded. After receiving feedback data from a large number of vehicles, the cloud analysis center determined that the current geographic feature model evaluation algorithm underestimated the weight of the impact of wet road surface splash interference. Therefore, an incremental training of the evaluation model was automatically initiated, and the "environmental similarity" parameter in the weight calculation of the device network was fine-tuned, so that the system would be more inclined to select model variants with stronger anti-interference capabilities in similar environments in the future. This is equivalent to steps 150 and 151 above.

[0053] Through the above application examples, this solution achieves a complete edge intelligence lifecycle management system in real and complex urban scenarios, encompassing environmental perception, dynamic adaptation, collaborative learning, predictive deployment, and closed-loop optimization. Vehicles can intelligently switch computing models based on their environment, collaborate with surrounding vehicles for optimization, and make predictive deployments. This solution effectively addresses core challenges faced by autonomous driving deployment, such as high environmental complexity, numerous long-tail scenarios, difficulties in model updates, and discontinuous cross-regional services. It demonstrates significant practical value and technological advancement in improving driving safety, system intelligence, and operational continuity.

[0054] Please refer to Figure 3. A terminal 1 for edge model optimization and collaboration based on BeiDou includes a memory 3, a processor 2, and a computer program stored on the memory 3 and running on the processor 2. When the processor 2 executes the computer program, it implements the various steps of an edge model optimization and collaboration method based on BeiDou.

[0055] Please refer to Figure 4 The processor 11 can be a CPU, which can realize BeiDou positioning processing, environmental feature analysis, model selection algorithm, collaborative learning control, and trajectory prediction analysis. The memory 22 can be RAM, used for geographic feature database, AI model storage, trajectory data caching, performance mapping storage, and configuration parameter caching; External interfaces include the BeiDou satellite positioning receiving interface, the geographic information system interface, the AI ​​model management interface, and the device collaborative communication interface.

[0056] The beneficial effects of a BeiDou-based edge model optimization and collaboration terminal are the same as those of the aforementioned BeiDou-based edge model optimization and collaboration method, and will not be repeated here.

[0057] In summary, this invention constructs an edge model optimization and collaboration method and terminal based on BeiDou. By deeply integrating the current geographic environment features acquired by the BeiDou positioning system into the entire lifecycle of edge intelligence, an intelligent model management system with geographic environment perception as the core driving engine is constructed, realizing full-process adaptive and intelligent operation from dynamic model adaptation, geographic weighted collaborative optimization, predictive model deployment to closed-loop feedback learning.

[0058] The system can automatically trigger the optimal model selection strategy based on the real-time geographical environment characteristics of the devices, and dynamically construct a device network through device location relationships to achieve intelligent organization and resource allocation of the collaborative training process. For task requirements in different geographical regions and mobile scenarios, the system can execute differentiated model deployment strategies and synchronously optimize the model library and caching strategies in real time.

[0059] Meanwhile, based on a pre-defined multi-scale geographic feature system and model performance evaluation criteria, the system defines a structured decision-making and weight allocation mechanism in model selection and collaborative optimization. This mechanism serves as the core intelligent engine for edge inference and collaborative learning, enabling dynamic optimization that is highly adaptable to complex geographic scenarios. Regarding data flow, the system performs standardized collection and multi-dimensional correlation analysis on the massive amounts of performance data, environmental features, and user feedback generated during device operation. This data is then used for continuous iterative training and evaluation of models and optimization strategies, ensuring the system possesses continuous evolution capabilities.

[0060] Based on the collaborative characteristics of federated learning and the dynamic decision-making mechanism of reinforcement learning, the system performs cross-device and cross-regional correlation analysis and knowledge accumulation on model operation records, collaborative optimization effects, and feedback data at each stage, achieving global perception and performance evaluation of edge intelligent service status. This method differs from traditional static configuration edge AI deployment approaches that ignore geographical connections. Through geographical perception-driven, data feedback-driven, and collaborative knowledge-driven approaches, it constructs an efficient, accurate, and scalable edge model optimization and collaborative workflow, significantly improving the model's environmental adaptability, collaborative training efficiency, and service continuity.

[0061] Furthermore, during model deployment and service, the system has established a pre-deployment mechanism based on trajectory prediction and proactive caching. When it is predicted that a device will enter a new geographical area, the system can automatically trigger model matching and preloading processes based on the characteristics of the target environment, completing seamless preparation of computing resources before the device arrives. This ensures a smooth transition and zero-interruption awareness in the service process, thereby improving service efficiency while guaranteeing the absolute reliability of critical mobile applications.

[0062] This terminal is suitable for edge intelligence applications that are sensitive to geographical environments and require high reliability, low latency, and continuous service, such as autonomous driving, drone inspection, and intelligent transportation. By constructing an intelligent, adaptive, and sustainably evolving geographic-aware edge computing system, it effectively reduces the blindness of model deployment, improves resource utilization efficiency and collaborative learning effects, and provides a quantifiable, optimizable, and scalable engineering solution foundation for the large-scale and reliable deployment of edge intelligence in the complex real world, thus possessing significant industrial application value.

[0063] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention's specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A BeiDou-based edge model optimization and collaborative method, characterized in that, include: The current geographical environment characteristics of the device are obtained through the BeiDou positioning system; Based on the current geographical environment characteristics, determine the appropriate target calculation model; Based on the real-time location in the current geographic environment features, the current device and at least one target device form a device network, and the target computing model is collaboratively optimized based on the device network. The model library is then updated based on the results of the collaborative optimization. Based on the current device's movement trajectory prediction, predictive model deployment is performed using the model library.

2. The edge model optimization and coordination method based on BeiDou as described in claim 1, characterized in that, The step of determining the appropriate target calculation model based on the current geographical environment characteristics includes: Based on the current geographic environment features, query the feature database for the current model performance that matches the current geographic environment features; the feature database stores the correlation between geographic environment features and model performance; Based on the current geographical environment characteristics and the current model performance, the target calculation model is determined.

3. The edge model optimization and coordination method based on BeiDou as described in claim 1, characterized in that, Before determining the appropriate target calculation model based on the current geographical environment characteristics, the process also includes: Record historical performance data of different computational models under different multi-scale geographical environmental characteristics; Based on the historical operational performance data, an evaluation model is trained to assess the adaptation of the computational model to the multi-scale geographical environmental features; Extract the current multi-scale geographic environment features from the current geographic environment features; The current multi-scale geographical environment features are input into the evaluation model to obtain the target calculation model.

4. The edge model optimization and coordination method based on BeiDou as described in claim 1, characterized in that, The step of collaboratively optimizing the target computation model based on the device network and updating the model library according to the results of the collaborative optimization includes: Calculate the collaborative weight of each device based on the geographical location relationships among all devices constituting the device network; Based on the collaborative weights, federated learning is performed in the device network to collaboratively optimize the target computation model; The model library is updated based on the results of the collaborative optimization and the current geographical environment characteristics of the current device.

5. The edge model optimization and coordination method based on BeiDou as described in claim 4, characterized in that, Also includes: Based on the geographical environment characteristics of each device in the device network and the model call frequency of all computing models in the device, a caching and distribution strategy for each computing model is obtained.

6. The edge model optimization and coordination method based on BeiDou as described in claim 5, characterized in that, The caching and distribution strategy for each computing model, based on the geographical environment characteristics of each device in the device network and the model call frequency of all computing models in the device, includes: Obtain a cluster of devices with similar current geographical environment characteristics to the current device; The target computing model is shared within the device cluster.

7. The edge model optimization and coordination method based on BeiDou as described in claim 6, characterized in that, Sharing the target computing model in the device cluster includes: Obtain the shared computing model in the device cluster whose call frequency is higher than a threshold, and cache the shared computing model in the device group in the device cluster. The device group includes multiple devices with similar geographical locations.

8. The edge model optimization and coordination method based on BeiDou as described in claim 1, characterized in that, The prediction of the movement trajectory based on the current device, and the deployment of predictive models in the model library, include: Analyze the historical movement trajectory of the device, and predict the future movement trajectory of the device based on the historical movement trajectory; the future movement trajectory includes the target location and the time information of the target location. Based on the predicted geographical environmental characteristics of the target location, a computational model that is compatible with the predicted geographical environmental characteristics is matched from the model library; The matching computational model is preloaded into the edge node corresponding to the current device or the target location.

9. The edge model optimization and coordination method based on BeiDou as described in claim 1, characterized in that, Also includes: Collect the performance data of the target computing model on the current device; Based on the operational performance data, the adaptation mechanism between the target calculation model and the current geographical environment features, as well as the collaborative optimization strategy, are optimized.

10. A terminal based on BeiDou edge model optimization and collaboration, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement each step of the BeiDou-based edge model optimization and coordination method as described in any one of claims 1 to 9.

Images