A rural scenic spot-oriented resilient edge intelligent tour guide system and method

Abstract

The present application relates to the field of information and communication technology, and discloses a kind of rural scenic spot-oriented resilience edge guide system and method.The basic principle of the method is that: the dynamic organization of tourist terminal is delay tolerant network node to realize data opportunity relay transmission under the environment without public network, while the lightweight artificial intelligence model is deployed on the local edge server to understand the semantics and knowledge reasoning of the request, and the model calculation accuracy and network communication power consumption are jointly regulated based on Lyapunov optimization theory. Therefore, the present application realizes the core technical effect that intelligent guide service can still run stably and continuously and provide depth associated narrative under the harsh conditions of rural scenic area network and resource double restriction.

Description

Technical Field

[0001] This invention relates to the field of information and communication technology, specifically to a resilient edge intelligent tour guide system and method for rural scenic areas. Background Technology

[0002] With the rapid development of rural tourism and the increasing demand from tourists for in-depth cultural experiences, the intelligent and personalized nature of scenic area tour services has become a clear trend. However, unlike urban scenic areas, most rural scenic areas are located in remote areas and generally face core constraints such as numerous mobile communication network coverage blind spots, unstable signal quality, and weak infrastructure such as electricity. This objective condition makes it difficult to directly implement traditional smart tourism solutions that heavily rely on continuous cloud connections and sufficient computing resources, thus posing a significant challenge to the intelligent construction of rural scenic areas.

[0003] Currently, attempts at guided tour technologies for rural environments mainly fall into two categories: one is online guided tour applications that rely on public mobile networks, whose service is completely interrupted in network blind spots; the other is simple offline audio guides or text and image displays, whose content is fixed, lacks interactivity, cannot respond to personalized queries, and has a very low level of intelligence. Furthermore, some existing solutions employing the concept of edge computing, while offloading some processing tasks, still largely pre-define stable backhaul networks and fail to fully consider resource constraints such as limited battery power of terminal devices and limited computing power of edge servers. These solutions lack systematic design for the combined and demanding conditions of random network connection interruptions and severely limited energy and computing resources.

[0004] Therefore, the fundamental problem that existing technologies have failed to effectively solve lies in how to build a navigation system that can adapt to environmental changes, intelligently allocate global resources, ensure long-term stable availability, and provide deep intelligent interaction in typical rural scenarios where there is no stable public network connection and terminal and edge node resources are highly limited. This leads to problems such as service interruptions, superficial experiences, and short system battery life in smart navigation services in rural scenic areas, hindering their high-quality development. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a stable, reliable, and intelligent interactive smart tour guide solution for rural scenic areas.

[0006] According to a first aspect of the present invention, a resilient edge-intelligent tour guide system for rural scenic areas is proposed, comprising: Multiple visitor terminals are used to initiate guided tour requests and receive feedback; Multiple network nodes are fixedly deployed in the scenic area or acted by tourist terminals. The network nodes form a dynamic communication network through self-organization, which is used to transmit data in an environment without a public network by opportunistic relay. An edge server, deployed locally within the scenic area, communicates with at least one network node; the edge server includes: A local knowledge base stores structured scenic area guide information; An artificial intelligence model library is used to perform semantic understanding of tour requests and to retrieve and reason about tour knowledge in order to generate tour answers; The resource scheduling optimizer is used to dynamically adjust the computational accuracy of the artificial intelligence model and the communication power consumption of network nodes based on the real-time resource status of the visitor terminal and the edge server, so as to maintain the long-term stable operation of the system under resource constraints.

[0007] According to some embodiments, in the system of the first aspect of the present invention, the artificial intelligence model library includes a speech recognition model, a natural language understanding model, and a speech synthesis model; all models are pruned and quantized, and are stored in an edge server for independently completing end-to-end processing from voice request to voice answer in an environment without a public network.

[0008] According to some embodiments, in the system of the first aspect of the present invention, the artificial intelligence model library further includes a graph neural network model; the data in the local knowledge base is organized into a spatiotemporal knowledge graph; the graph neural network model is used to perform multi-hop reasoning on the spatiotemporal knowledge graph to generate an associative guided narrative.

[0009] According to some embodiments, in the system of the first aspect of the present invention, the dynamic communication network is a delay-tolerant network; the network nodes are configured as DTN agents, and when two network nodes enter the communication range, they evaluate and decide whether to exchange and forward data packets based on a preset utility function.

[0010] According to some embodiments, in the system of the first aspect of the present invention, the calculation of the utility function follows the following formula: ; in: Indicates the remaining power factor. This represents the node's current remaining battery power. Fully charged; To predict the probability factor of intersection; This is a data packet urgency factor. It is the maximum lifespan of the data packet. This indicates the time the device has been in existence. This is a factor representing the historical delivery success rate. , , , These are weighting coefficients, all of which are positive numbers, used to adjust the relative importance of the four factors.

[0011] According to some embodiments, in the system of the first aspect of the present invention, the resource scheduling optimizer is specifically used for: Based on historical system operation data collected by edge servers, a virtual queue representing system task backlog and energy consumption deviation is constructed. Based on Lyapunov optimization theory, with the goal of minimizing the weighted sum of virtual queues, a macro-resource optimization problem is solved in each policy update cycle to generate a resource regulation policy containing at least one condition-action pair. The resource allocation strategy is encapsulated into a strategy data package and distributed to the visitor terminal via a dynamic communication network using an opportunistic relay method. The visitor terminal receives and stores policy data packets and continuously monitors its own status locally. When the monitored status meets the preset conditions in the condition-action pair, it autonomously executes the corresponding control action.

[0012] According to some embodiments, in the system of the first aspect of the present invention, the control actions in the resource control strategy include at least one of the following: adjusting the wireless transmission power of the tourist terminal; adjusting the confidence threshold of the local speech recognition module or the output sampling rate of the speech synthesis module in the tourist terminal; adjusting the local data caching strategy or data packet priority marking rule of the tourist terminal; wherein, when solving the macro-resource optimization problem, the resource scheduling optimizer controls the system's preference between long-term average energy consumption and service quality by adjusting the trade-off parameter V in the Lyapunov optimization framework.

[0013] According to a second aspect of the present invention, a resilient edge-based intelligent navigation method for rural scenic areas is proposed, implemented in rural scenic areas without stable public network connectivity. The method includes: S1. Initiate a guided tour request through the visitor terminal; S2. Guided tour requests are transmitted to locally deployed edge servers via an opportunistic relay method through a dynamic communication network composed of multiple self-organizing network nodes. S3. In the edge server, a lightweight artificial intelligence model is used to understand the request and generate guided answers by combining the local knowledge base; S4. In the edge server, the resource scheduling optimizer generates resource regulation strategies based on the system's historical operating data and distributes the strategies to the visitor terminals through a dynamic communication network. The visitor terminals receive and store the strategies and continuously monitor their own status locally. When the monitored status meets the preset conditions in the strategy, they autonomously execute the corresponding regulation actions. S5. Return the guide answers to the visitor terminal via a dynamic communication network.

[0014] According to some embodiments, in the method of the second aspect of the present invention, step S4 specifically includes: S41. The edge server collects historical system operation data, including the average battery decay rate of each visitor terminal, task backlog trend, and network encounter frequency; S42. Based on historical operating data, construct a virtual queue representing system task backlog and energy consumption deviation; S43. Based on the Lyapunov optimization framework, with the goal of minimizing the weighted sum of virtual queues, a macro-resource optimization problem is solved in each policy update cycle to generate a resource regulation policy containing at least one condition-action pair. S44. Encapsulate the resource allocation strategy into a strategy data packet and distribute it to the visitor terminal via a dynamic communication network in an opportunistic relay manner. S45. The visitor terminal receives and stores policy data packets and continuously monitors its own status locally; S46. When the visitor terminal detects that its own status meets the preset conditions in the condition-action pair, it autonomously executes the corresponding control action.

[0015] According to some embodiments, in the method of the second aspect of the present invention, the artificial intelligence model library includes a speech recognition model, a natural language understanding model, a graph neural network model, and a speech synthesis model; the generation of guide answers in step S3 includes: S31. Use a speech recognition model to convert the speech data in the tour guide request into text; S32. Use a natural language understanding model to parse the text in order to identify the request intent and extract at least one entity and relationship; S33. Based on the extracted entities and relationships, locate the relevant entity nodes in the spatiotemporal knowledge graph constructed in the local knowledge base; S34. Using a graph neural network model, starting from the located entity node, iterative message passing and information aggregation are performed along the edges of the spatiotemporal knowledge graph to infer the implicit knowledge nodes and associated paths related to the request intent. S35. Integrate the node information on the associated path and use a speech synthesis model to generate the corresponding navigation answer speech data.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. In response to the problem that the lack of a stable public network and limited resources in rural scenic areas render intelligent tour guide services unusable, this invention constructs a resilient system architecture that integrates terminals as nodes, edge intelligence, and joint resource regulation. This architecture achieves the core effect of ensuring the basic availability and long-term stable operation of intelligent tour guide services without relying on fixed communication and power infrastructure.

[0017] 2. To address the issues of limited computing power and fragmented traditional tour guide content in offline environments that cannot deeply respond to tourist inquiries, this invention achieves end-to-end intelligent processing and generates deeply related, narrative tour guide content by deploying a lightweight voice AI model on an edge server and combining it with graph neural networks to perform multi-hop reasoning on spatiotemporal knowledge graphs.

[0018] 3. To address the problem of data transmission failure in areas with no communication infrastructure, this invention defines the dynamic network as a delay-tolerant network (DTN) and designs an intelligent utility function based on power consumption and mobility prediction to guide the opportunistic relay of data packets. This enables the construction of an adaptive and energy-efficient data "carry-forward" network using tourist mobility, ensuring information delivery.

[0019] 4. To address the issue of system service interruptions caused by random fluctuations in resources such as network and power, this invention introduces a resource scheduler based on Lyapunov optimization theory. This scheduler generates resource regulation strategies containing condition-action pairs based on historical system data and distributes these strategies to user terminals via a dynamic communication network. Each terminal continuously monitors its own status locally and autonomously executes the corresponding regulation action when the conditions are met. Thus, this invention achieves macro-level strategy-level optimization of system resources and distributed autonomous execution by terminals under multiple random constraints, thereby ensuring stable long-term average service performance. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an embodiment 1000 of a resilient edge-guided intelligent tour system for rural scenic areas according to the present invention; Figure 2 for Figure 1 A schematic diagram of the structure of Example 1032A of the artificial intelligence model library of Example 1000; Figure 3 for Figure 1 A schematic diagram of the structure of Example 1032B of the artificial intelligence model library of Example 1000; Figure 4 This is a flowchart illustrating an embodiment 2000 of the resilient edge intelligent tour guide method for rural scenic areas according to the present invention. Figure 5 for Figure 4 A flowchart illustrating step S3 in Example 1000; Figure 6 for Figure 4 A flowchart illustrating step S4 in Example 1000. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Figure 1 This is a schematic diagram of an embodiment 1000 of the resilient edge-based intelligent tour guide system for rural scenic areas according to the present invention. Figure 1 As shown, embodiment 1000 includes multiple tourist terminals 101, multiple network nodes 102, and an edge server 103.

[0023] In some specific embodiments, in this invention: multiple tourist terminals 101 are used to initiate guided tour requests and receive feedback; multiple network nodes 102 are fixedly deployed in the scenic area or are acted by the tourist terminals 101; the network nodes 102 form a dynamic communication network through self-organization, which is used to transmit data in an opportunistic relay mode in an environment without a public network.

[0024] Optionally, the visitor terminal 101 serves as the direct interface between the system and the user, playing a dual role: firstly, as a client for requesting and receiving services, and secondly, as a mobile relay node in the dynamic communication network. The visitor terminal 101 is both the starting and ending point of the intelligent tour guide service, and also a key infrastructure component ensuring the service's connectivity in offline environments.

[0025] Specifically, the visitor terminal 101 is typically a visitor's smartphone or a dedicated smart rental device provided by the scenic area. For example, the scenic area can provide a ruggedized tablet or handheld tour guide as the visitor terminal 101. Optionally, in some specific embodiments, the core principle of the visitor terminal 101 is to run a customized application on the device, which integrates two main functional modules: a smart tour guide interaction front-end and a DTN communication proxy. Hardware-wise, it needs to ensure the presence of a microphone, speaker, GPS / Bluetooth chip, and sufficient storage space to support voice interaction and temporary data caching.

[0026] In some specific embodiments, the specific operational process by which the visitor terminal 101 implements its functional applications includes: (1) Initiating a request: The guest speaks through the application's voice button, and the application calls the local speech recognition engine to perform real-time transcription on the front end, converting the speech stream into a text request. Subsequently, this text request, along with the guest's anonymous ID, current GPS location, and other metadata, is encapsulated into a data packet with the target address of the edge server.

[0027] (2) Network Access and Transmission: The DTN agent within the application is awakened and attempts to find an available network connection. In the absence of a public network, the agent automatically turns on Bluetooth and begins scanning for other network nodes 102 in the vicinity. Once a neighboring node is found, a decision is made based on a utility function evaluation to determine whether to transmit immediately. After the data packet is sent, a copy is generated locally and stored in a storage queue until a successful reception confirmation is received from the server, or until the time-to-live (TTL) expires. Optionally, the neighboring node can be another guest mobile phone or a fixed beacon.

[0028] (3) Receiving and Presenting Feedback: When the navigation answer data packet is finally transmitted back to the terminal through the multi-hop network, the DTN agent delivers it to the application. The application parses the data packet; if the content is text, it calls the local lightweight speech synthesis engine to read it aloud; if it contains images or short videos, they are displayed synchronously. The entire interaction process does not require internet access on the terminal side.

[0029] In traditional technology, the visitor terminal is used only as a client and loses all functionality when the network is disconnected. According to... Figure 1 The illustrated implementation transforms terminals from consumers into participants, making them mobile and renewable network infrastructure. Even in areas without cellular signals and Wi-Fi hotspots, terminals can connect to each other, forming a personnel-based network. This significantly reduces reliance on and construction costs of fixed communication infrastructure, leveraging the inherent mobility of tourists to extend network coverage, enabling services to move with people and the network to go wherever they go.

[0030] In some specific embodiments, network nodes 102 are entities that constitute the system's resilient communication network, and are divided into two categories: fixed nodes pre-deployed at key locations within the scenic area, and mobile nodes acted by tourist terminals 101. These two types of nodes together form a hybrid, self-organizing, latency-tolerant network, whose core function is to transmit data in time and space to address the challenges of discontinuous network connectivity.

[0031] In some specific embodiments, the fixed nodes in network node 102 are typically low-power, dustproof, and waterproof embedded devices, such as single-board computers integrating Bluetooth 5.0 or LoRa communication modules, solar charging panels, and batteries. Optionally, the fixed nodes are fixed at locations such as viewing platforms, intersections, and rest stops, acting as network hubs or relay stations. Optionally, the common logical core of the fixed nodes and mobile nodes in network node 102 is running the same DTN agent software.

[0032] Optionally, the dynamic communication network is a latency-tolerant network. A latency-tolerant network is an architecture designed specifically for extreme network environments. Its core idea is to abandon the instant connectivity required by the traditional Internet and adopt a store-and-carry-forward working mode. Unlike traditional networks, the latency-tolerant network used in this invention does not assume a real-time end-to-end path. Nodes in the network can store received data packets locally and then physically carry the data while moving. When encountering other nodes, they determine whether to forward the data to them. Through this relay transmission method in time and space, the data eventually reaches its destination, but will experience uncertain delays. In this solution, the lack of stable cellular network coverage in rural scenic areas is a typical application scenario for DTN. Each tourist terminal and fixed device is configured as a DTN agent, enabling data to utilize the mobility of people to reach areas with network connectivity or where the target is located.

[0033] Network node 102 is configured as a DTN agent. When two network nodes 102 enter the communication range, they evaluate whether to exchange and forward data packets based on a preset utility function. In some specific embodiments, the utility function is calculated based at least on the remaining battery power of the network node 102 initiating the exchange, its movement trajectory, and the predicted intersection probability of the data packet destination.

[0034] In some specific embodiments, the process of node discovery and connection between network nodes 102 includes: each node's DTN agent continuously or periodically broadcasting a beacon to announce its presence. When two nodes enter communication range, they establish a temporary point-to-point connection and exchange their respective node IDs, remaining battery power, and stored packet digest lists.

[0035] The intelligent routing decision-making process among network nodes 102 in this invention mainly involves utility function calculation. In some specific embodiments, the formula for the utility function is designed as follows: ; in: This represents the remaining power factor. This represents the node's current remaining battery power. Fully charged.

[0036] This is the predicted intersection probability factor. It represents the estimated probability that the node will encounter the destination node of the data packet within a certain period of time, and is between 0 and 1.

[0037] This is the urgency factor for data packets. It is the maximum lifespan of the data packet. This is the data packet's lifetime. The higher the value, the more urgent the data packet.

[0038] This is the historical delivery success rate factor. It records the historical ratio of similar data packets successfully delivered by this node, ranging from 0 to 1.

[0039] , , , : Weighting coefficients, all of which are positive numbers, are used to adjust the relative importance of the four factors and need to be adjusted according to the actual scenario.

[0040] Optionally, the adjustment principles for the four weighting coefficients in the utility function are shown in the table below: Table 2. Adjustment principles for weighting coefficients in the utility function: .

[0041] Specifically, when node A carrying several data packets enters the communication range with node B, the following calculation and decision-making process is triggered: (a) Information exchange: First, A and B exchange basic information, including their respective node IDs, remaining power, and their respective data packet digest lists, which include destination address, time to live, etc., but not all data.

[0042] (ii) Packet-by-packet evaluation and decision-making: For each data packet that needs to be forwarded carried by A, A's DTN agent will calculate the forwarding utility value U(A->B) of the data packet for node B respectively.

[0043] Calculate the predicted intersection probability This is the most innovative computation. The system includes a mobility model based on historical encounter records. For a fixed node B, if B is a fixed gateway deployed at the observation deck, and the destination of the data packets is the edge server, then B's... The value for that destination is 1, and the two will inevitably meet.

[0044] Intelligent routing decisions between nodes rely on an accurate estimate of the probability of future encounters between mobile nodes and data packet destinations. For mobile nodes, this involves predicting the intersection probability. The following calculations were performed by integrating information from historical encounter probabilities, real-time movement trajectory factors, and scenic area movement pattern matching probabilities: 1. Historical encounter probability Each node maintains a record of its historical encounters with other nodes, estimating the base probability based on the encounter frequency. The more frequently two nodes encounter each other over a period of time, the greater the likelihood of them meeting again in the future. The historical encounter probability can be directly derived from the proportion of encounters to the total observation time.

[0045] 2. Real-time movement trajectory factor T: When two nodes meet, the node initiating the evaluation estimates the tendency of the other node to approach its destination based on the other node's current position, movement direction, and speed. Specifically, it calculates the distance d from the node to the destination and the angle between the movement direction and the destination direction. ,definition Where f(d) decreases with distance, As the included angle decreases; weight , It can be preset or dynamically adjusted. In practical applications, it can be simplified to a comprehensive evaluation based on distance and direction, corresponding to three levels: high, medium, and low.

[0046] 3. Scenic Area Movement Pattern Matching Probability M: The system pre-learns typical tourist movement patterns in rural scenic areas, constructs a database of popular routes and transfer probabilities between attractions. Based on the node's current location, movement direction, and historical trajectory, it matches the most likely tour route and obtains the probability of the node's future destinations. .

[0047] Final prediction of intersection probability The weighted sum of the three The adjustment principles for the weight coefficients w1, w2, and w3 are as follows: In the initial stage of system operation, historical data is sparse, so w2 and w3 are increased to rely on real-time information and the scenic area's prior model; after the system is running stably, historical data is abundant, so w1 and w3 are increased to trust historical statistics and stable models; for urgent data with high real-time requirements, w2 is temporarily increased to respond promptly to the latest developments.

[0048] (III) Utility Comparison and Action: Node A compares its calculated U(A->B) for a given data packet with its utility U(A->A) for continuing to carry that data packet. The calculation of U(A->A) is similar, but... It is a prediction based on A's own trajectory.

[0049] Forwarding rule: If U(A->B)>U(A->A) + Threshold, that is, the expected utility of B's ​​delivery is significantly higher than that of A carrying it itself, and B has enough storage space, then A decides to forward the data packet to B.

[0050] Non-forwarding rule: If the utility difference is small or negative, A retains the data packet and continues to wait for a better relay opportunity.

[0051] (iv) Dynamic updates After each successful delivery or encounter, the historical success rate of the relevant nodes. It will be updated. The mobility model will also be continuously learned and fine-tuned based on actual trajectories, enabling... The predictions are becoming increasingly accurate.

[0052] According to such Figure 1 The implementation shown in this invention employs a utility function decision process to upgrade Bluetooth data exchange to a value assessment based on energy, prediction, data urgency, and historical reputation. This ensures that limited routing opportunities and node energy are allocated to the most promising and needed data packets and the most suitable next-hop carrier, thereby significantly improving the overall data delivery success rate and network energy efficiency of the DTN network in sparsely populated, regularly moving scenarios such as rural scenic areas. Compared to the simple strategies in traditional ad hoc networks, this approach offers significant advantages in intelligence and adaptability.

[0053] Edge server 103 is deployed locally in the scenic area and communicates with at least one network node 102.

[0054] Optionally, the edge server 103 includes: a local knowledge base 1031 storing structured scenic area guide knowledge; an artificial intelligence model library 1032 used to perform semantic understanding of guide requests and retrieve and reason about guide knowledge to generate guide answers; and a resource scheduling optimizer 1033 used to dynamically adjust the computational accuracy of the artificial intelligence model and the communication power consumption of network nodes according to the real-time resource status of tourist terminals and edge servers, so as to maintain the long-term stable operation of the system under resource constraints.

[0055] In some specific embodiments, the core principle of the local knowledge base 1031 is to transform unstructured text and image knowledge into a structured, relation-rich spatiotemporal knowledge graph. The graph is stored in the form of "entity-relationship-attribute" triples, such as [ancient bridge]-[built in]-[Qing Dynasty], [ancient bridge]-[located in]-[stream east of the village], and time (such as seasonal characteristics) and space (GPS coordinates) labels are attached to entities and relationships.

[0056] Optionally, the specific operations involved in the local knowledge base 1031 include: (1) Initialization construction: Before the system is deployed, the administrator imports the pre-organized scenic area data into the knowledge base construction tool through a wired network. The tool automatically or semi-automatically extracts entities and relationships to form an initial knowledge graph.

[0057] (2) Request response: When a query request is received from the artificial intelligence model library, the local knowledge base 1031 query engine will: locate the entity; expand the multi-hop query along the relationship edge to find the node directly connected to it and indirectly associate it with more distant nodes; combine the context of the visitor's current location, current season and other contexts attached to the request, and prioritize returning the most relevant knowledge fragments in time and space.

[0058] (3) Dynamic updates: After the system is running, the administrator can update new knowledge packages incrementally when connecting to the network at a fixed node, thereby realizing the evolution of the knowledge base.

[0059] Optionally, the artificial intelligence model library 1032 includes a speech recognition model, a natural language understanding model, and a speech synthesis model, all of which have undergone pruning and quantization processing and are stored in an edge server for independent end-to-end processing from voice request to voice answer in an environment without a public network.

[0060] Optionally, the resource scheduling optimizer 1033 is specifically used to: construct a virtual queue representing system task backlog and energy consumption deviation based on historical system operation data collected by the edge server; solve a macro-level resource optimization problem in each policy update cycle based on Lyapunov optimization theory, with the goal of minimizing the weighted sum of the virtual queues, and generate a resource regulation policy containing at least one condition-action pair; encapsulate the resource regulation policy into a policy data packet and distribute it to the visitor terminal via a dynamic communication network in an opportunistic relay manner; the visitor terminal receives and stores the policy data packet and continuously monitors its own status locally, and autonomously executes the corresponding regulation action when the monitored status meets the preset conditions in the condition-action pair.

[0061] Optionally, the control actions in the resource control strategy generated by the resource scheduling optimizer 1033 include at least one of the following: adjusting the wireless transmission power of the tourist terminal; adjusting the confidence threshold of the local speech recognition module or the output sampling rate of the speech synthesis module in the tourist terminal; adjusting the local data caching strategy or data packet priority marking rule of the tourist terminal; wherein, when solving the macro-level resource optimization problem, the resource scheduling optimizer controls the system's preference between long-term average energy consumption and service quality by adjusting the trade-off parameter V in the Lyapunov optimization framework.

[0062] Lyapunov optimization is a classic mathematical framework for dealing with performance optimization problems under long-run average constraints in stochastic dynamic systems. Its core mechanism can be summarized by the following key elements: Construction of virtual queues: Virtual queues are a fundamental construct of Lyapunov optimization, used to transform long-term average constraints into queue stability problems; Based on the general Lyapunov optimization theory, this invention makes the following four innovative special treatments for the special application scenario of rural scenic spots without public network connection: (1) Macro-strategy generation replaces real-time control General Lyapunov optimization makes real-time decisions in each tiny time slot, requiring the decisions to take effect on the system immediately. However, in DTN networks, edge servers cannot connect to terminals in real time. This invention adjusts the time scale from real-time to hourly / day-level policy update cycles, and changes the optimization object from real-time actions to macro-level policy rules. In specific implementation, the resource scheduling optimizer solves a macro-level resource optimization problem based on the system's historical operating data from the past week, generating a control policy containing multiple condition-action pairs.

[0063] (2) Spatiotemporal Dimension Expansion of Virtual Queues General virtual queues only reflect the cumulative deviation in the time dimension. This invention extends the definition of virtual queues to include spatial dimension information as well. Optionally, the system constructs independent virtual queues for different areas of the scenic area. For example, the energy consumption virtual queue Znorth(t) in the North area reflects the average energy consumption deviation of terminals in the North area; the task backlog queue Qsouth(t) in the South area reflects the request processing delay of tourists in the South area.

[0064] When solving macro-optimization problems, the optimizer comprehensively evaluates the queue status of each region and generates region-specific control strategies. For example, it focuses on energy consumption control for the northern region and on service quality assurance for the southern region. This expansion of the spatiotemporal dimensions enables the strategy to adapt to the different resource needs of different areas within the rural scenic area.

[0065] (3) Introduction of terminal autonomous execution framework The general Lyapunov optimization implicitly assumes that the central controller executes actions. This invention innovatively introduces a terminal autonomous execution framework: Policy distribution phase: The edge server encapsulates the generated policy packets into DTN data packets and distributes them to each terminal in a best-effort manner using opportunistic networking.

[0066] Local storage phase: After receiving the policy packet, the terminal stores it in the local policy rule base and continuously monitors its own state variables.

[0067] Autonomous execution phase: When the monitored state meets the preset conditions in the strategy, the terminal immediately and autonomously executes the corresponding control action without waiting for server instructions.

[0068] This framework decouples centralized decision-making from distributed execution, ensuring global optimization while overcoming the physical limitations of DTN networks that cannot be controlled in real time.

[0069] (4) Adjustable action design that is deeply integrated with the business logic of the scenic area General resource scheduling typically only adjusts basic parameters such as transmission power. In some specific embodiments, this invention deeply integrates the adjustment actions with the scenic area navigation business logic, designing four types of adjustable actions, all of which are executed autonomously by the terminal, as shown in the table below: Table 1. Resource Scheduling Optimizer Control Actions: .

[0070] Among them, the adjustment of the trade-off parameter V is of particular significance: scenic area managers can adjust the value of V according to the operation strategy. The larger V is, the more the generated strategy tends to be energy-saving; the smaller V is, the more it tends to ensure service quality.

[0071] According to such Figure 1 The technical effects achieved by the resource scheduling optimizer in the present invention, as shown in the implementation method, include: theoretically ensuring the long-term average performance stability of the system under conditions of no public network connection, intermittent network connectivity, and limited terminal power in rural scenic areas; achieving near-optimal global resource allocation; and adaptive adjustment capability to environmental changes such as seasonal fluctuations in passenger flow. At the same time, the design of one-time policy distribution and local execution at the terminal achieves zero control overhead, fundamentally solving the fundamental problem that traditional centralized control cannot close the control loop in DTN networks.

[0072] Figure 2 for Figure 1 A schematic diagram of the structure of Example 1032A of the artificial intelligence model library of Example 1000. (See attached diagram.) Figure 2 The artificial intelligence model library 1032A shown in Example 1032A includes a speech recognition model A1, a natural language understanding model A2, and a speech synthesis model A3.

[0073] Optionally, models A1, A2, and A3 in the artificial intelligence model library 1032A are all pruned and quantized, and then stored in the edge server 103 for independent end-to-end processing from voice request to voice answer in an environment without a public network.

[0074] In some specific embodiments, the speech recognition model A1 converts continuous audio signals into corresponding text sequences. Deep learning models (such as Conformer and RNN-T) are typically used, which learn the mapping relationship from audio features to phonemes and then to text, handling long-term dependencies in speech through attention mechanisms. Optionally, the specific implementation of the speech recognition model A1 includes: using an open-source framework model pre-trained on a large amount of Chinese data as a base; and fine-tuning the base with audio data containing proper nouns from rural scenes and background noise to improve scene adaptability.

[0075] In some specific embodiments, the natural language understanding model A2 understands the intent of the text and extracts key information. Its principle is based on the Transformer architecture, using a self-attention mechanism to understand the semantic relationships between words. Optionally, the specific implementation process of the natural language understanding model A2 includes: 1. Pre-training fine-tuning: Lightweight Chinese pre-trained models (such as MiniLM and ALBERT) are used as the basis to fine-tune the intent classification and named entity recognition tasks on dialogue data in the tourism guide domain.

[0076] 2. Knowledge Enhancement: During training and inference, it works in conjunction with the entity dictionary in the local knowledge base to improve the accuracy of professional entity recognition.

[0077] In some specific embodiments, the speech synthesis model A3 converts text into natural, fluent speech. The principle is as follows: first, acoustic features are generated using a sequence-to-sequence model, and then the features are restored to waveform audio using a vocoder. Optionally, the speech synthesis model A3 employs an end-to-end model such as VITS to simplify the process and improve synthesis efficiency.

[0078] In some specific embodiments, such as Figure 2 As shown, the artificial intelligence model library 1032A proposed in this invention has the following design for the specific application scenario of this solution: At the deployment location: Current mainstream solutions employ centralized cloud deployment, where the original, complete large model runs on a cloud server. The terminal only collects voice data, uploads the audio stream over the network, and the cloud processes the data before returning the results. This solution, however, uses edge-side hardening deployment: a deeply compressed, lightweight model is directly burned into the storage chip or dedicated AI accelerator card of the edge server.

[0079] In terms of model form: Traditional approaches employ large floating-point precision models using 32-bit or 16-bit floating-point (FP32 / FP16) weights. While these models are large and highly accurate, they incur significant computational and memory overhead. This approach uses integer quantization to create smaller models: quantization techniques convert weights and activation values ​​to 8-bit integers (INT8) or even lower precision. Then, redundant neurons or connections that contribute minimally to the output are removed through pruning. This reduces the model size to tens of MB.

[0080] Optionally, the special handling of pruning quantification in this scheme is mainly reflected in the following aspects: (1) Unified compression across the entire chain: Considering the cascading error propagation between multiple models, the quantization accuracy of each model needs to be kept balanced; the models share the same quantization parameters, which facilitates unified hardware acceleration.

[0081] (2) Adjustable accuracy design: Confidence threshold, output sampling rate, etc. are designed as dynamically adjustable parameters; based on Lyapunov optimization theory, the resource scheduling optimizer generates control instructions in each strategy update cycle to achieve a flexible trade-off between accuracy and energy consumption.

[0082] (3) Solidified storage implementation: After pruning and quantization, the model is solidified and stored in a dedicated storage area of ​​the edge server. Solidified storage ensures that the model exists as an inherent capability of the system, with anti-tampering and fast startup characteristics; it is deeply optimized with the instruction set and memory architecture of specific hardware to achieve the ultimate energy efficiency ratio; it does not require the support of the operating system file system and is directly loaded into the local memory of the NPU at startup.

[0083] In some specific embodiments, the pruning quantization processing in this solution achieves end-to-end intelligent processing from voice input to deep knowledge reasoning on resource-constrained edge servers through the following mechanism: 1. Storage Space Dimension: Edge server storage space is limited. The total size of the four original FP32 precision models is approximately: ASR (200MB) + NLU (150MB) + TTS (300MB) + GNN (250MB) = 900MB. After INT8 quantization and pruning, the total size is reduced to approximately 200-250MB. This allows all models to be permanently stored on the edge server, eliminating the need for dynamic loading from the cloud and enabling offline intelligence.

[0084] 2. Memory Usage: Memory usage during model inference includes not only the parameters themselves but also intermediate activation values. INT8 quantization reduces memory bandwidth requirements, and combined with structured pruning, the model can reside simultaneously in the memory of the edge server. This eliminates the need for frequent model swapping throughout the entire process, ensuring that end-to-end response latency remains within acceptable limits.

[0085] 3. Computational power dimension: The AI ​​computing power of edge servers is far lower than that of cloud GPUs. INT8 integer operations are 2-4 times faster than FP32. After end-to-end compression, the total computational load can be reduced from hundreds of GFLOPs to tens of GFLOPs.

[0086] 4. Power Consumption: Edge servers in rural scenic areas may rely on unstable power sources such as solar power. The power consumption of INT8 operations is only 1 / 3 to 1 / 2 that of FP32. More importantly, the adjustable precision design allows the resource scheduler to dynamically reduce model precision when power is scarce, achieving a flexible trade-off between precision and power consumption, and ensuring long-term stable operation of the system.

[0087] In terms of operating mode: Existing solutions employ online service calls, where each request is a network API call, and the model is dynamically loaded and executed in a shared computing pool in the cloud. This solution, however, achieves local, embedded execution: the model becomes part of the edge server firmware. When processing requests, the model is directly loaded from embedded storage into the local memory of the NPU or GPU for execution, without complex system calls or external dependencies.

[0088] Traditional solutions exhibit shortcomings in rural scenarios characterized by unstable networks, expensive bandwidth, and latency sensitivity: service continuity cannot be guaranteed, response latency is high and unpredictable, and private data must be uploaded to the cloud. According to... Figure 2 The present invention solves the problem through the above-described technical path in the embodiments shown. Offline availability: The fixed storage model ensures that the core speech recognition, understanding, and synthesis capabilities remain stable and available in the event of any network outage.

[0089] Resource constraints: The pruned and quantized model reduces computational FLOPs and memory usage by an order of magnitude. This allows it to run smoothly on edge AI chips equipped with only low to medium computing power, greatly reducing hardware costs and power consumption, making large-scale deployment in rural scenarios economically and technically feasible.

[0090] Response latency and determinism: Eliminates network round-trip latency and cloud queue waiting time. All computations are performed locally, and end-to-end response time is strictly limited to local hardware processing time, providing a smooth interactive experience.

[0091] Figure 3 for Figure 1 A schematic diagram of the structure of Example 1032B of the artificial intelligence model library of Example 1000. (See attached diagram.) Figure 3 The artificial intelligence model library shown in Example 1032B includes: with Figure 2 The speech recognition model A1, natural language understanding model A2, and speech synthesis model A3 are the same as the speech recognition model B1, natural language understanding model B2, and speech synthesis model B3, and Example 1032B also includes a graph neural network model B4.

[0092] Optionally, the data in the local knowledge base 1031 is organized into a spatiotemporal knowledge graph. A spatiotemporal knowledge graph is an enhanced data structure built upon traditional knowledge graphs, by adding time and space dimension labels to entities and relationships. Its construction process includes: (1) Data collection and extraction: Information extraction technology is used to automatically extract triples from unstructured texts such as scenic area chronicles, folk tales, and ecological survey reports. At the same time, spatial coordinates GPS are bound to each entity, and time tags are bound to relationships involving specific seasons, festivals, or historical periods.

[0093] (2) Graph Pattern Design: Define the core categories of the ontology and their relationships. An entity can belong to multiple categories at the same time, forming multi-dimensional labels. For example, categories include buildings, plants, people, and events; relationships include location, built on, related legends, and best viewing season.

[0094] (3) Spatiotemporal attribute binding: Spatialization: Accurately mapping entities onto electronic maps. For example, not only is the existence of "Wang Family Ancestral Hall" recorded, but also its polygonal geofence coordinates are recorded, and relationships with spatial entities such as "east end of the village" and "starting point of the stone road" are established.

[0095] Temporalization: Add a temporal context to relationships. For example, a relationship (stream, visible animals, egret) can be appended with the attribute {"time condition": "summer morning"}; a historical event (ancient bridge, under renovation, Qing Dynasty) will have a built-in time point.

[0096] In some specific embodiments, the graph neural network model B4 is used for multi-hop reasoning on spatiotemporal knowledge graphs to generate associative guided narratives. During its learning process, it employs a combination of knowledge graph embedding and graph neural networks (such as CompGCN and R-GAT). The model learns through multiple rounds of iterative learning, with message passing as its core operation. Each entity node aggregates feature information from its neighbors through relational edges and updates its own vector representation. After learning, the model can map entities and relations into a continuous vector space, where semantically similar or spatiotemporally related entities have vector representations that are also close to each other in the space. The model is trained to accurately predict missing links in the graph.

[0097] Optionally, in a specific scenario embodiment of the multi-hop reasoning and narrative generation process of the graph neural network model B4, when a tourist asks, "What is the significance of the stone lions at the entrance of the Wang Family Ancestral Hall?", the workflow of the graph neural network model B4 is as follows: (1) Query parsing and embedding: The natural language understanding model extracts the core entity "stone lion at the entrance of Wang Family Ancestral Hall" and the intent "inquire about cultural connotations". This entity is transformed into its vector representation learned by the GNN.

[0098] (2) Multi-hop reasoning path search: In the vector space, the model starts from the "Stone Lion" node and performs multi-step intelligent roaming along the relation edges: First hop: Through the "belongs to" relation, it is associated with the concept of "door decoration art". Second hop: Through the "symbol" relation, it is associated with cultural connotation nodes such as "brave and powerful". Third hop: Through the "found in" relation, it is associated with the inductive knowledge node of "common characteristics of local Qing Dynasty architecture".

[0099] Spatiotemporal filtering: Throughout the process, the reasoning path is implicitly constrained by the current tourist's geographical location (at the entrance of the ancestral hall) and time (such as the current time being the Spring Festival), prioritizing the activation of the knowledge path most relevant to the current time and place.

[0100] (3) Relational Narrative Synthesis: The model integrates the node information on the activated knowledge path (stone lion -> door decoration art -> bravery and strength -> commonalities of Qing Dynasty architecture). Instead of simply listing facts, it organizes them into a coherent and in-depth explanatory text according to the logic of "specific object -> cultural category -> deeper meaning -> broader cultural context".

[0101] According to such Figure 3 The embodiments shown in this invention achieve the following technical effects: They construct a rural knowledge brain with spatiotemporal awareness capabilities. This distinguishes the system from most existing tour guide technologies that only provide information broadcasting, achieving a high level of cultural interpretation and knowledge association. This solves the common problems in rural smart tourism, such as superficial experiences and insufficient transmission of cultural connotations, thereby enhancing service value and competitiveness.

[0102] Figure 4 This is a flowchart illustrating an embodiment 2000 of the resilient edge-based intelligent tour guide method for rural scenic areas according to the present invention. Figure 4 As shown, Embodiment 2000 includes steps S1-S5. Optionally, the method of Embodiment 2000 is implemented in rural scenic areas without a stable public network connection.

[0103] In step S1, a guided tour request is initiated through the tourist terminal. Specifically, the tourist asks a question through the voice interface of the smart terminal. The terminal's built-in preprocessing module first reduces and enhances the recording to cope with environmental noise such as wind and flowing water in the countryside. Subsequently, instead of performing full online speech recognition, the compressed audio data and necessary metadata are encapsulated into a standard data packet. The preset destination of this data packet is an edge server. Optionally, the necessary metadata includes: a globally unique request ID, a terminal device ID, a request timestamp, and the terminal's real-time GPS coordinates.

[0104] In step S2, the tour request is transmitted to the locally deployed edge server via an opportunistic relay method through a dynamic communication network composed of multiple self-organized network nodes.

[0105] Optionally, the specific implementation process of step S2 includes: (1) Neighbor discovery: The DTN agent module of the node carrying the data packet periodically broadcasts its own identification information, including node ID and power status, via Bluetooth Low Energy.

[0106] (2) Utility-driven intelligent node contact: When another node (which may be another tourist's mobile phone, or a fixed repeater deployed on a tree stump or road sign) enters the range and exchanges identification information, the two parties will enter a distributed decision-making process. The core of the decision is to calculate the forwarding utility value according to a preset utility function, which is based at least on the remaining power of the node that initiated the exchange and the predicted intersection probability with the destination of the data packet.

[0107] For example, a tourist's mobile phone with sufficient battery power and historical movement patterns indicating it's heading towards the scenic area service center (where the edge server is located) will have a high predicted rendezvous probability and will thus be considered a high-utility node. Both nodes evaluate whether forwarding the data packet to the other is better than continuing to carry it themselves. If the evaluation indicates the other party has higher delivery utility, they decide to exchange and forward the data packet; otherwise, they retain the data packet and wait for a better relay opportunity.

[0108] (3) Relay-style store-carry-forward: If calculations show that handing the data packet to a contact node has a higher expected delivery success rate than carrying it itself, then point-to-point transmission is performed. After transmission, the original node deletes its local copy, and the new node is responsible for carrying the packet and finding the next relay opportunity. This process is repeated until the data packet reaches a fixed gateway node that can communicate directly with the edge server. Optionally, nodes that communicate directly with the edge server are usually deployed at scenic area entrances and exits, visitor centers, and other locations with stable power supply and backhaul links.

[0109] In step S3, on the edge server, the guide request is understood using an artificial intelligence model library, and a guide answer is generated by combining it with a local knowledge base. In some specific embodiments, the artificial intelligence model library includes a speech recognition model, a natural language understanding model, a graph neural network model, and a speech synthesis model.

[0110] In some specific embodiments, generating the guide answer in step S3 includes: converting the voice data in the guide request into text using a speech recognition model; parsing the text using a natural language understanding model to identify the request intent and extract at least one entity and relationship; locating relevant entity nodes in a spatiotemporal knowledge graph built in a local knowledge base based on the extracted entities and relationships; using a graph neural network model, starting from the located entity nodes, iteratively passing messages and aggregating information along the edges of the spatiotemporal knowledge graph to infer implicit knowledge nodes and associated paths related to the request intent; integrating the node information on the associated paths, and generating the corresponding guide answer voice data using a speech synthesis model.

[0111] In step S4, on the edge server, a resource scheduling optimizer generates a resource regulation strategy based on the system's historical operating data and distributes the strategy to the visitor terminal through a dynamic communication network. The visitor terminal receives and stores the strategy and continuously monitors its own status locally. When the monitored status meets the preset conditions in the strategy, it autonomously executes the corresponding regulation action.

[0112] Optionally, the control actions in the resource control strategy include at least one of the following: adjusting the wireless transmission power of the tourist terminal; adjusting the confidence threshold of the local speech recognition module or the output sampling rate of the speech synthesis module in the tourist terminal; adjusting the local data caching strategy or data packet priority marking rules of the tourist terminal; wherein, when solving the macro-resource optimization problem, the resource scheduling optimizer controls the system's preference between long-term average energy consumption and service quality by adjusting the trade-off parameter V in the Lyapunov optimization framework.

[0113] In step S5, the guide answer is returned to the visitor's terminal via a dynamic communication network. Optionally, in some specific embodiments, the guide answer generated in step S5 is assigned the original request ID and the target terminal ID, and injected as a new data packet into the same dynamic opportunity network. The process of step S5 also follows the utility relay mechanism of step S2, searching for a path back to the visitor who initiated the request. Since the answer data packet may be larger than the request packet, the network may fragment it for transmission.

[0114] The method proposed in this invention is completed entirely in a local closed loop, without relying on any cloud services, thereby achieving a highly available, highly intelligent, and highly resilient tour guide experience in rural areas with weak infrastructure.

[0115] Figure 5 for Figure 4 A flowchart illustrating step S3 in Example 1000. (See attached diagram.) Figure 5 As shown, step S3 includes steps S31-S35. Optionally, the artificial intelligence model library includes speech recognition models, natural language understanding models, graph neural network models, and speech synthesis models.

[0116] In step S31, the speech data in the tour guide request is converted into text using a speech recognition model. Specifically, step S31 receives the compressed and encoded audio data of the tour guide request from the network transport layer, decodes the audio and extracts its features, and then uses its deep learning network to map the continuous acoustic feature sequence into the corresponding text word sequence. This model has been optimized for background noise and possible dialect accents in rural environments to improve offline recognition accuracy, and finally outputs a plain text request corresponding to the speech content.

[0117] In step S32, the text is parsed using a natural language understanding model to identify the request intent and extract at least one entity and relationship.

[0118] In some specific embodiments, in step S32, the natural language understanding model performs deep analysis of the text, completing two core tasks: (1) Intent recognition: Determine the user's query purpose, such as belonging to the categories of "query history", "query function", "comparison difference" or "request explanation".

[0119] (2) Named entity recognition and relation extraction: Identify key entities in the text and extract semantic relations between entities. For example, for the entity "big locust tree", its semantic relations include "how many years old it is", indicating that the relation is "tree age".

[0120] Step S32 ultimately outputs a structured semantic understanding result, which typically includes an intent label and a set of (entity, relation) pairs. For example, the output is {intent: query attribute, entity relation: [(big locust tree, tree age)]}.

[0121] In step S33, based on the extracted entities and relationships, the relevant entity nodes are located in the spatiotemporal knowledge graph constructed in the local knowledge base.

[0122] In some specific embodiments, in step S33, the system uses the entity name in the input to perform a precise query in the spatiotemporal knowledge graph of the local knowledge base to find the unique graph node referred to by the name. For example, linking "big locust tree" to a node in the graph named "ancient locust (number 001, located at the village entrance)".

[0123] Spatiotemporal attributes play a crucial role here: if the system detects that the tourist terminal is located in the south of the village, and there are two ancient locust trees, one to the north and one to the south, in the knowledge graph, it will prioritize linking to the "ancient locust tree in the south of the village" node to achieve context awareness. Step S33 outputs one or more accurately located entity nodes in the spatiotemporal knowledge graph as a clear starting point for deep reasoning.

[0124] In step S34, using a graph neural network model, starting from the located entity node, the implicit knowledge nodes and associated paths related to the request intent are inferred by iterative message passing and information aggregation along the edges of the spatiotemporal knowledge graph.

[0125] In some specific embodiments, in step S34, the graph neural network model acts as the execution subject, and the data it processes includes entity nodes from S33, query intents from S32, and the structure and feature data of the entire spatiotemporal knowledge graph.

[0126] Specifically, step S34 starts with the located entity node and performs reasoning based on simulated cognition: (1) Message passing: The model initializes the features of the entity node. Then, multiple rounds of iteration are performed. In each round, each node aggregates the feature information of all its neighboring nodes.

[0127] (2) Information aggregation and updating: The aggregated information is processed through a learnable neural network layer to update the feature representation of the current node. After several iterations, the features of the starting node contain the association information of its multi-hop neighbors.

[0128] (3) Relevance path manifestation: By combining the query intent, the model can assess the importance of different propagation paths, thereby obtaining one or more of the most relevant inference paths. For example, for the query "tree age", the path "ancient locust tree -> planted in -> Ming Dynasty" may be activated; if the intent is "cultural significance", the paths "ancient locust tree -> witness -> historical event" and "ancient locust tree -> recorded in -> local chronicles" may be activated at the same time.

[0129] Finally, step S34 outputs a knowledge subgraph consisting of a set of highly relevant nodes and their relationships, which directly responds to the user's query intent.

[0130] In step S35, the node information on the associated path is integrated, and the corresponding navigation answer voice data is generated using a speech synthesis model.

[0131] In some specific embodiments, the execution process of step S35 includes: (1) Narrative Integration: First, a lightweight text generation logic organizes the structured knowledge subgraph into a coherent and natural spoken text according to the logic of human narrative. For example, independent facts such as "planted in the Ming Dynasty", "witnessed the XX battle", and "recorded in the county annals" are integrated into "This ancient locust tree was planted in the Ming Dynasty, more than 600 years ago. It has not only experienced vicissitudes, but also silently witnessed the famous XX battle in history, and therefore it is recorded in detail in the local county annals." (2) Speech Synthesis: Subsequently, the speech synthesis model receives the text and converts it into natural and fluent audio. The model can support specific timbres to enhance the friendliness of the explanation.

[0132] The final audio data for the guided tour is prepared and returned to the visitor via step S5.

[0133] Figure 6 for Figure 4 A flowchart illustrating step S4 in Example 1000. (See attached diagram.) Figure 6 As shown, step S4 includes steps S41-S46.

[0134] In step S41, the edge server collects historical system operation data, including the average battery decay rate of each tourist terminal, task backlog trends, and network encounter frequency. This invention addresses the limitation of DTN networks in real-time terminal connectivity by shifting data collection from real-time monitoring to the collection and aggregation of historical statistical data. Specifically, the edge server's data source is the locally recorded operation log uploaded when a tourist terminal encounters a fixed gateway. This log includes battery change curves, task processing records, and encounter timestamps with other nodes. This data is gradually aggregated to the edge server over several hours or days via the DTN network's store-carry-forward mechanism.

[0135] Optionally, the calculation of the statistical indicators in step S41 specifically includes: Average battery depletion rate: This is calculated by analyzing the terminal's battery level over time to determine the average battery consumption per unit time. This indicator reflects the terminal's energy consumption pattern, and the formula is as follows: ,in, and These are the start and end battery levels of the terminal's i-th recording cycle. It is the duration of the cycle.

[0136] Task backlog trend: This metric tracks the average latency from when a request is initiated to when a response is received, as well as the change in the queue length of local pending tasks. This indicator reflects fluctuations in service quality.

[0137] Network encounter frequency: The average interval and duration of encounters between a terminal and other nodes. This metric is used to infer the network connectivity of a region, and is expressed by the formula: ,in, In total time The number of encounters recorded internally.

[0138] In step S42, based on historical operational data, a virtual queue representing system task backlog and energy consumption deviation is constructed. Optionally, in this invention, the energy consumption virtual queue evolves according to the following formula: ,in, This is the actual energy consumption for the current cycle. This represents the target energy consumption value. Similarly, the task backlog queue evolves as follows: Q Where A(t) is the number of newly arrived tasks and D(t) is the number of completed tasks.

[0139] Optionally, the present invention also introduces a network connectivity queue to characterize deviations in the quality of service of a regional network: C .in, For the frequency of target encounters, This represents the average encounter frequency as actually observed.

[0140] In step S43, based on the Lyapunov optimization framework, with the goal of minimizing the weighted sum of the virtual queues, a macro-resource optimization problem is solved in each policy update cycle to generate a resource regulation policy containing at least one condition-action pair.

[0141] Classical Lyapunov optimization solves a drift-penalty minimization problem in each time slot. This invention innovatively adjusts the optimization object from real-time actions to macroscopic policy rules, solving the following macroscopic optimization problem in each policy update cycle T: ; The constraints include: That is, the generated policy rule set is a subset of the feasible rule space R; each rule It satisfies the pre-defined formal grammatical constraints; , where MM is the upper limit of the strategy size.

[0142] in: A set of regional indexes for the scenic area; The value of the virtual queue for task backlog in region k at the end of the period; The value of the virtual queue for energy consumption deviation in region k at the end of the cycle; The value of the virtual queue for network connectivity in region k at the end of the period; , , are the weight coefficients of each queue, reflecting the system's preference for different optimization objectives; V is the Lyapunov trade-off parameter, controlling the preference between stability and energy consumption; Let P be the average energy consumption expected to be generated in the future cycle; P is the macroeconomic resource regulation strategy set to be generated.

[0143] In step S44, the resource control policy is encapsulated into a policy data packet and distributed to the visitor terminal via the dynamic communication network using an opportunistic relay method. Optionally, in step S44, each policy rule in the policy data packet is encapsulated into a special DTN data packet containing the following fields: policy ID, version number, effective area, effective time period, condition-action list, time to live, and priority.

[0144] In step S45, the visitor terminal receives and stores the policy data packet and continuously monitors its own status locally. In some specific embodiments, after receiving the policy data packet, the terminal in this invention parses it and stores it in the policy rule base in the local non-volatile storage area. Multiple policies can coexist, with the latest version number taking precedence. Optionally, in step S45, the terminal starts a lightweight monitoring thread in the background to continuously sample the following status variables: Battery level: Reads battery percentage once per second. Location information: GPS coordinates are retrieved every 30 seconds (if enabled). Time information: Current system time Local task queue: Number of pending tour requests Network status: Number of currently connected nodes, signal strength After each state update, the monitoring thread iterates through the local policy rule base to check if any rule conditions are met. Rule matching is based on priority sorting, with higher priority rules being matched first.

[0145] In step S46, when the visitor terminal detects that its own state meets the preset conditions in the condition-action pair, it autonomously executes the corresponding control action. Compared to traditional terminals where action execution is triggered by cloud commands or local hard-coded rules, the special processing of this invention includes: Autonomy in action execution: When the rule matching engine in step S45 finds a rule that meets the conditions, it immediately triggers the corresponding control action without needing confirmation from the server. The execution process includes: Communication power consumption control: Adjust Bluetooth / Wi-Fi transmit power parameters by calling the operating system API.

[0146] Accuracy adjustment: Send configuration commands to the local speech recognition module to adjust the confidence threshold; send commands to the speech synthesis module to adjust the output sampling rate.

[0147] Data strategy adjustment: Adjust local caching strategies or data packet priority marking rules.

[0148] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0149] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A resilient edge navigation system for rural scenic areas, characterized in that, include: Multiple visitor terminals are used to initiate guided tour requests and receive feedback; Multiple network nodes are fixedly deployed in the scenic area or acted by the tourist terminals. The network nodes form a dynamic communication network through self-organization, which is used to transmit data in an environment without a public network by opportunistic relay. An edge server, deployed locally within the scenic area, communicates with at least one of the aforementioned network nodes; The edge server includes: A local knowledge base stores structured scenic area guide information; An artificial intelligence model library is used to perform semantic understanding of the tour request and to retrieve and reason about the tour knowledge in order to generate tour answers; The resource scheduling optimizer is used to dynamically adjust the calculation accuracy of the artificial intelligence model and the communication power consumption of the network nodes based on the real-time resource status of the tourist terminal and the edge server, so as to maintain the long-term stable operation of the system under resource constraints.

2. The system according to claim 1, characterized in that, The artificial intelligence model library includes speech recognition models, natural language understanding models, and speech synthesis models. All models have undergone pruning and quantization processing and are stored in the edge server for independent end-to-end processing from voice request to voice answer in an environment without a public network.

3. The system according to claim 2, characterized in that, The artificial intelligence model library also includes a graph neural network model; the data in the local knowledge base is organized into a spatiotemporal knowledge graph; the graph neural network model is used to perform multi-hop reasoning on the spatiotemporal knowledge graph to generate a related guided narrative.

4. The system according to claim 1, characterized in that, The dynamic communication network is a delay-tolerant network; the network nodes are configured as DTN agents, and when two network nodes enter the communication range, they evaluate and decide whether to exchange and forward data packets based on a preset utility function. The utility function is calculated based at least on the remaining battery power of the network node that initiated the exchange, the predicted intersection probability of its movement trajectory and the destination of the data packet.

5. The system according to claim 4, characterized in that, The utility function is calculated according to the following formula: ； in: Indicates the remaining power factor. This represents the node's current remaining battery power. Fully charged; To predict the probability factor of intersection; This is a data packet urgency factor. It is the maximum lifespan of the data packet. This indicates the time the device has been in existence. This is a factor representing the historical delivery success rate. , , , These are weighting coefficients, all of which are positive numbers, used to adjust the relative importance of the four factors.

6. The system according to claim 1, characterized in that, The resource scheduling optimizer is specifically used for: Based on the historical system operation data collected by the edge server, a virtual queue representing system task backlog and energy consumption deviation is constructed; Based on Lyapunov optimization theory, with the goal of minimizing the weighted sum of the virtual queues, a macro-resource optimization problem is solved in each policy update cycle to generate a resource regulation policy containing at least one condition-action pair. The resource regulation strategy is encapsulated into a strategy data packet and distributed to the tourist terminal via the dynamic communication network in an opportunistic relay manner. The tourist terminal receives and stores the policy data packet, and continuously monitors its own status locally. When the monitored status meets the preset conditions in the condition-action pair, it autonomously executes the corresponding control action.

7. The system according to claim 6, characterized in that, The control actions in the resource control strategy include at least one of the following: Adjust the wireless transmission power of the tourist terminal; Adjust the confidence threshold of the local speech recognition module or the output sampling rate of the speech synthesis module in the tourist terminal; Adjust the local data caching strategy or data packet priority marking rules of the tourist terminal; When solving the macro-level resource optimization problem, the resource scheduling optimizer controls the system's preference between long-term average energy consumption and service quality by adjusting the trade-off parameter V in the Lyapunov optimization framework.

8. A resilient edge-based intelligent tour guide method for rural scenic areas, characterized in that, The method, implemented in rural scenic areas without a stable public network connection, includes: S1. Initiate a guided tour request through the visitor terminal; S2. The tour request is transmitted to a locally deployed edge server via an opportunistic relay method through a dynamic communication network composed of multiple self-organizing network nodes. S3. In the edge server, the guide request is understood using an artificial intelligence model library, and a guide answer is generated by combining it with a local knowledge base; S4. In the edge server, a resource scheduling optimizer generates a resource regulation strategy based on historical system operation data and distributes the strategy to the visitor terminal through the dynamic communication network. The visitor terminal receives and stores the strategy and continuously monitors its own status locally. When the monitored status meets the preset conditions in the strategy, it autonomously executes the corresponding regulation action. S5. The guide answers are returned to the visitor terminal via the dynamic communication network.

9. The method according to claim 8, characterized in that, Step S4 specifically includes: S41. The edge server collects historical operating data of the system, including the average power consumption rate of each visitor terminal, task backlog trend, and network encounter frequency; S42. Based on the historical operating data, construct a virtual queue representing the system task backlog and energy consumption deviation; S43. Based on the Lyapunov optimization framework, with the goal of minimizing the weighted sum of the virtual queues, a macro-resource optimization problem is solved in each policy update cycle to generate a resource regulation policy containing at least one condition-action pair; S44. The resource regulation strategy is encapsulated into a strategy data packet and distributed to the visitor terminal via the dynamic communication network in an opportunistic relay manner; S45. The tourist terminal receives and stores the policy data packet, and continuously monitors its own status locally; S46. When the tourist terminal detects that its own state meets the preset conditions in the condition-action pair, it autonomously executes the corresponding control action.

10. The method according to claim 8, characterized in that, The artificial intelligence model library includes speech recognition models, natural language understanding models, graph neural network models, and speech synthesis models; the generation of guide answers in step S3 includes: S31. Use the speech recognition model to convert the speech data in the tour guide request into text; S32. Use the natural language understanding model to parse the text to identify the request intent and extract at least one entity and relationship; S33. Based on the extracted entities and relationships, locate the relevant entity nodes in the spatiotemporal knowledge graph constructed in the local knowledge base; S34. Using the graph neural network model, starting from the located entity node, iterative message passing and information aggregation are performed along the edges of the spatiotemporal knowledge graph to infer the implicit knowledge nodes and associated paths related to the request intent. S35. Integrate the node information on the associated path and use the speech synthesis model to generate corresponding navigation answer speech data.

Images