Intelligent text travel integrated management method and system

Abstract

The invention discloses a smart text travel comprehensive management method and system, and relates to the technical field of smart text travel management. Comprising a semantic construction module, a passenger flow recommendation module, a priority evaluation module, a selection evaluation module, an access sequence module, a credit record module and a dynamic updating module. According to the method, tourists are modeled into intelligent agents, a multi-agent reinforcement learning algorithm is combined, adaptive optimization of tourist group decision is realized according to a cooperative behavior weight, interaction behaviors among tourists are simulated, and reasonability of scenic spot selection prediction and group intelligent feature expression ability are improved; a semantic attention distribution mechanism is introduced into a sequence-to-sequence model, a personalized access sequence is generated with selection probability distribution as a constraint, and dynamic sequence optimization based on cultural semantics and tourist portraits is achieved.

Description

Technical Field

[0001] This invention relates to the field of smart cultural tourism management technology, and in particular to a smart cultural tourism integrated management method and system. Background Technology

[0002] With the rapid development of smart tourism and the cultural industry, the integration of culture and tourism has become an important direction for promoting local economic development and cultural dissemination. Modern cultural tourism management systems are gradually introducing emerging technologies such as the Internet of Things, artificial intelligence, and big data to achieve efficient visitor flow management, optimized resource allocation, and improved visitor experience. However, cultural elements possess high semantic complexity and regional diversity. Management models that rely solely on statistics or superficial data analysis cannot fully reveal the deep relationship between cultural connotations and visitor behavior, resulting in limited intelligence in smart cultural tourism systems for precise recommendations and dynamic management.

[0003] Existing smart tourism management technologies mainly focus on single functional modules such as visitor flow prediction, route recommendation, or service scheduling. The integration of multi-source heterogeneous data is not high, and a unified semantic understanding framework is lacking. Although some systems have attempted to introduce knowledge graphs or deep learning models, they still have shortcomings in semantic layer association modeling, spatiotemporal dependency capture, and dynamic optimization of visitor behavior. Summary of the Invention

[0004] This invention proposes a smart cultural tourism integrated management method and system driven by cultural knowledge graphs and semantic weight matrices. By constructing a cultural knowledge graph to generate a semantic weight matrix, semantic twin models and spatiotemporal graph neural networks are used to achieve semantic-spatiotemporal joint prediction. Combined with multi-agent reinforcement learning and blockchain credit feedback mechanisms, dynamic intelligent management and personalized decision-making of the cultural tourism system are realized, solving the problem of insufficient semantic fusion of multi-source heterogeneous cultural tourism data, which leads to low accuracy in passenger flow prediction and attraction recommendation.

[0005] A smart cultural tourism integrated management method, characterized by comprising: Collect multi-source data from cultural and tourism areas, call cultural knowledge graphs and generate semantic weight matrices; determine the fusion weights of multi-source data based on semantic weight matrices, perform fusion processing on various types of data, and obtain multi-dimensional semantic feature vectors. Based on multidimensional semantic feature vectors and semantic weight matrices, semantic-spatiotemporal joint prediction is performed in the semantic twin model through a spatiotemporal graph neural network to obtain the passenger flow trend prediction results of each scenic spot node. Based on the predicted passenger flow trend, the initial recommendation weights of each scenic spot node are generated. The predicted visitor density and service load coefficient of each scenic spot node are obtained and integrated, and then further weighted in combination with the initial recommendation weight to obtain the dynamic priority score of each scenic spot node. Each tourist is modeled as an intelligent agent, and dynamic priority scores are used as input features. The collaborative behavior weights between agents are calculated based on the semantic weight matrix. Combined with a multi-agent reinforcement learning algorithm, the selection probability distribution of each scenic spot node is obtained. By using the selection probability distribution as a constraint, the sequence-to-sequence model is called, and the attention distribution obtained by combining the semantic weight matrix is ​​used to generate a personalized scenic spot visit sequence. When the intelligent agent is detected to have arrived at a scenic spot node in the scenic spot visit sequence, the interaction is triggered and the experience feedback data is recorded; the multi-dimensional semantic feature vector and experience feedback data are written into the blockchain to form a credit record, and the service quality is verified and a credit score is generated through on-chain smart contracts; Based on credit scores and experience feedback data, the cultural knowledge graph and multi-agent reinforcement learning algorithm are updated.

[0006] As a preferred technical solution of the present invention, the generation of the semantic weight matrix includes: calling a pre-constructed cultural knowledge graph, which is established based on publicly available and reliable cultural tourism semantic resources, using scenic spot nodes and tourist profile elements as semantic entities, and constructing semantic association relationships through relation triples and their attribute weights; the attribute weights are used to characterize the importance or association strength of the relationship between entities, and are determined comprehensively based on entity attribute similarity, relationship frequency, temporal relevance, and cultural theme relevance; performing semantic embedding processing on text, image, audio, and sensor data samples included in the cultural knowledge graph to obtain vectorized representations of semantic entities; constructing an initial semantic matrix by combining semantic similarity, spatial proximity, and cultural theme relevance between semantic entities; and performing normalization and hierarchical weighting processing on the initial semantic matrix to generate a semantic weight matrix using the weighted result of cultural theme weight, geographical adjacency weight, and behavioral preference weight.

[0007] As a preferred technical solution of the present invention, obtaining the multidimensional semantic feature vector includes: taking the scenic spot node as the target, extracting the association strength between the scenic spot node and other semantic entities from the semantic weight matrix; determining the fusion weight of the multi-source information related to the scenic spot node based on the association strength and the correlation of multi-source data in the semantic space; performing weighted fusion processing and semantic alignment processing on the multi-source data related to the scenic spot node according to the fusion weight to form a fusion feature; and performing feature unification and dimension mapping on the fusion feature to obtain a multidimensional semantic feature vector representing the semantic state of the scenic spot node.

[0008] As a preferred technical solution of the present invention, the joint prediction includes: using the multidimensional semantic feature vectors of scenic spot nodes as input, invoking the semantic association relationships between scenic spot nodes to establish a semantic association network; in the semantic twin model, constructing a spatiotemporal graph neural network structure based on the semantic association network, and controlling the information transmission weights between different scenic spot nodes according to the association strength between scenic spot nodes, and combining time series features to perform temporal propagation and update of the semantic feature vectors of scenic spot nodes; through the semantic propagation and time evolution process, calculating the passenger flow change trend of each scenic spot node in the target time period, and generating a predicted passenger flow trend.

[0009] As a preferred technical solution of the present invention, the generation of the initial recommendation weights for each scenic spot node includes: calculating the initial recommendation weights for each scenic spot node based on the predicted visitor flow trend; determining the reward and constraint parameters for recommendation optimization based on historical experience feedback data; and adaptively adjusting the initial recommendation weights under non-negative and normalized conditions to generate the initial recommendation weights for each scenic spot node.

[0010] As a preferred technical solution of the present invention, the step of obtaining the dynamic priority score of each scenic spot node includes: determining the predicted passenger flow density based on the predicted passenger flow trend and historical passenger flow data of the scenic spot node; determining the service load coefficient based on the actual service capacity and real-time service load of the scenic spot node; weighting the predicted passenger flow density and service load coefficient based on preset weights to obtain the fusion value of the comprehensive status of the scenic spot node; and further weighting the fusion value according to the initial recommendation weight to generate the dynamic priority score of the corresponding scenic spot node.

[0011] As a preferred technical solution of the present invention, obtaining the selection probability distribution of each scenic spot node includes: calculating the collaborative behavior weights between agents based on the semantic associations between scenic spot nodes and between tourist profile elements; constructing a multi-agent interaction environment based on the collaborative behavior weights, and jointly modeling the state, actions, and rewards of each agent; in the multi-agent interaction environment, using a multi-agent reinforcement learning algorithm to iteratively train the agent's strategy, taking dynamic priority scoring as the state input, and combining predicted passenger flow density and service load constraints to optimize the objective function of each agent; and outputting the selection probability distribution of each scenic spot node after the strategy converges.

[0012] As a preferred technical solution of the present invention, the generation of personalized scenic spot visit sequences includes: using the selection probability distribution of each scenic spot node as a constraint, calling a sequence-to-sequence model to perform sequence modeling of tourist visit behavior; during the sequence modeling process, combining the semantic association between scenic spot nodes and tourist profile elements to generate an attention distribution to guide the sequence generation process; dynamically adjusting the decoding weights of the sequence-to-sequence model according to the attention distribution to control the output order and occurrence probability of candidate scenic spot nodes; and generating personalized scenic spot visit sequences under the premise of satisfying the constraints.

[0013] As a preferred technical solution of the present invention, the generation of credit scores includes: writing the multi-dimensional semantic feature vector of the scenic spot node and the corresponding experience feedback data into the blockchain to form an immutable credit record; calling the smart contract deployed on the chain to automatically verify and compare the service quality indicators in the experience feedback data to determine whether the service meets the preset quality standards; and generating a credit score for the corresponding scenic spot node or service provider based on the verification results of the smart contract, taking into account tourist satisfaction, service compliance and historical credit records.

[0014] A smart cultural tourism integrated management system includes: Semantic construction module: Collects multi-source data of cultural and tourism areas, calls cultural knowledge graph and generates semantic weight matrix; determines the fusion weight of multi-source data based on semantic weight matrix, performs fusion processing on various types of data, and obtains multi-dimensional semantic feature vector; Passenger flow recommendation module: Based on multi-dimensional semantic feature vectors and semantic weight matrices, semantic-spatiotemporal joint prediction is performed in the semantic twin model through spatiotemporal graph neural network to obtain the passenger flow trend prediction results of each scenic spot node. Based on the predicted passenger flow trend, the initial recommendation weights of each scenic spot node are generated. Priority assessment module: Obtain and integrate the predicted visitor density and service load coefficient of each scenic spot node, and then perform weighted processing in combination with the initial recommendation weight to obtain the dynamic priority score of each scenic spot node. Selection Evaluation Module: Each tourist is modeled as an agent, and dynamic priority scores are used as input features. The collaborative behavior weights between agents are calculated based on the semantic weight matrix. Combined with a multi-agent reinforcement learning algorithm, the selection probability distribution of each scenic spot node is obtained. The visit sequence module: with the selection probability distribution as a constraint, it calls the sequence to the sequence model and combines the attention distribution obtained by the semantic weight matrix to generate a personalized visit sequence for attractions; Credit Record Module: When the intelligent agent arrives at the scenic spot node in the scenic spot visit sequence, it triggers the interaction and records the experience feedback data; the multi-dimensional semantic feature vector and experience feedback data are written into the blockchain to form a credit record, and the service quality is verified and a credit score is generated through on-chain smart contracts; Dynamic update module: Updates the cultural knowledge graph and multi-agent reinforcement learning algorithm based on credit scores and experience feedback data.

[0015] The present invention has the following advantages: This invention achieves semantic unification and adaptive weight fusion of multi-source heterogeneous data by collecting multi-source data from cultural and tourism regions and combining it with cultural knowledge graphs to generate a semantic weight matrix. By introducing entity relationships and attribute weight calculations from cultural knowledge graphs into the semantic weight matrix generation process, and comprehensively considering entity similarity, relationship frequency, temporal relevance, and cultural theme relevance, the construction of the semantic matrix is ​​more in line with cultural semantic logic, thus improving the accuracy and interpretability of semantic association modeling.

[0016] This invention determines the fusion weights of multi-source data based on the semantic weight matrix and performs semantic alignment and feature mapping to obtain multi-dimensional semantic feature vectors that represent the semantic state of scenic spot nodes. This significantly enhances the discriminativeness and stability of the semantic features of scenic spots, providing high-quality input features for subsequent prediction and recommendation.

[0017] This invention constructs a semantic twin model and a spatiotemporal graph neural network structure to achieve spatiotemporal joint propagation and dynamic updating of semantic features, significantly improving the accuracy and timeliness of passenger flow trend prediction. By combining predicted passenger flow trends with historical feedback data, the recommendation weights are adaptively adjusted under non-negative and normalized conditions to achieve dynamic optimization of recommendation weights, effectively balancing the recommendation ratio of popular attractions and potential attractions, and improving the diversion of tourists and the balance of experience.

[0018] This invention integrates predicted visitor density and service load coefficient, and then performs further weighting based on the initial recommendation weights to obtain dynamic priority scores for scenic spot nodes. This enables a comprehensive quantitative assessment of the scenic area's operational status and enhances the dynamic response capability of resource scheduling and service management.

[0019] This invention models tourists as intelligent agents and combines them with multi-agent reinforcement learning algorithms to achieve adaptive optimization of tourist group decision-making based on collaborative behavior weights. It simulates the interactive behavior among tourists, improving the rationality of scenic spot selection prediction and the ability to express collective intelligence features. By introducing a semantic attention distribution mechanism into the sequence-to-sequence model, personalized access sequences are generated with selection probability distribution as a constraint. This achieves dynamic sequence optimization based on cultural semantics and tourist profiles, enhancing the personalization and cultural fit of tourism route recommendations.

[0020] This invention establishes a trustworthy credit scoring mechanism by writing multidimensional semantic feature vectors and experience feedback data into the blockchain and verifying service quality through smart contracts. This enables traceability and trustworthy management of cultural and tourism services, thereby enhancing the transparency and credibility of the cultural and tourism ecosystem. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of a smart cultural tourism integrated management system used in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings.

[0023] Example 1: A smart cultural tourism integrated management method, comprising the following steps: Step S1: Collect multi-source data of cultural and tourism areas, call the cultural knowledge graph and generate a semantic weight matrix; determine the fusion weight of multi-source data based on the semantic weight matrix, perform fusion processing on various types of data, and obtain multi-dimensional semantic feature vectors; In one embodiment of the present invention, the multi-source data in this step includes cultural knowledge data, operational and passenger flow data, service and carrying capacity data, tourist behavior and interaction data, environmental and IoT sensing data, as well as third-party and public data.

[0024] The cultural knowledge data primarily originates from the public knowledge base of the cultural and tourism departments, local chronicles, intangible cultural heritage databases, official websites of scenic spots, and semantically annotated documents, used to construct a cultural knowledge graph. Operational and visitor flow data is obtained through the scenic spot's gate system, mobile signal probes, Bluetooth / Wi-Fi detection devices, and ticketing platform interfaces. Service and carrying capacity data comes from the resource capacity archives and real-time monitoring platform of the scenic spot's operation and management system. Visitor behavior and interaction data is obtained through the official guide app, electronic explanation terminals, and online questionnaire system. Environmental and IoT sensing data is collected by IoT sensors deployed throughout the scenic spot, including temperature, humidity, noise, PM2.5, and visitor flow heat maps. Third-party and public data includes city event information, external platform reputation and popularity indices, and macro-level holiday schedules.

[0025] Taking the Forbidden City as an example, its data includes: the historical figure "Zhu Di" from the cultural knowledge graph, the cultural event "Ming Dynasty court etiquette", the theme element "ancient building complex", and the geographical location "Dongcheng District, Beijing"; visitor flow data includes the number of turnstiles entering and exiting, average stay time, and peak hours for entering the park; service data includes the coverage rate of guided tours on the day and hygiene guarantee capabilities; and environmental data includes real-time temperature, air quality, and noise monitoring values.

[0026] The generation of the semantic weight matrix includes: calling a pre-constructed cultural knowledge graph, which is built based on publicly available and credible cultural tourism semantic resources, using scenic spot nodes and tourist profile elements as semantic entities, and constructing semantic association relationships through relation triples and their attribute weights; the attribute weights are used to characterize the importance or association strength of the relationship between entities, and are determined comprehensively based on entity attribute similarity, relationship frequency, time relevance and cultural theme relevance. Cultural knowledge graphs refer to knowledge networks formed through structured and semantic modeling of cultural, tourism, and tourist behavior data. Semantic entities include attraction nodes and tourist profile elements: Attraction nodes are information units within a cultural and tourism region that possess independent cultural characteristics, geographical boundaries, and management attributes; their associated content includes historical figures, cultural events, thematic elements, and geographical locations. Tourist profile elements are data structures used to characterize tourist features and behavioral tendencies, including age, residential level, thematic preferences, consumption level, travel mode, and companion type; they may also include extended attributes such as holiday travel preferences, language ability, and physical limitations.

[0027] The relation triple is formalized as (E1, R, E2), where E1 and E2 are semantic entities, and R is the semantic relation between them (e.g., "located in", "related to a topic", "preference"). The attribute weight w(R) is calculated using a multi-factor weighting: w(R) = α1S attr +α2F rel +α3T corr +α4C topic ;where S attr For entity attribute similarity, F rel T represents the frequency of the relationship. corr For time-related factors, C topic The cultural theme relevance is represented by α1 to α4, which are statistically determined normalization coefficients.

[0028] Semantic embedding processing is performed on text, image, audio, and sensor data samples included in the cultural knowledge graph to obtain vectorized representations of semantic entities; Text data is processed by language models such as BERT or ERNIE to extract semantic embedding vectors; image data is processed by ResNet or VisionTransformer models to obtain visual feature representations; audio data is processed by MFCC feature extraction and convolutional network encoding; and sensor data undergoes time-series feature extraction and standardization. These features are uniformly mapped to the same semantic space through a multimodal fusion layer to obtain an entity vector set V={v1,v2,…,v...} n}

[0029] By combining semantic similarity, spatial proximity, and cultural theme relevance between semantic entities, an initial semantic matrix M0 is constructed, where M0(i,j) represents the semantic association strength between entities i and j. After matrix normalization, a weighted average is applied to generate the final semantic weight matrix M. s =β1M tpoic +β2M geo +β3M pref , of which M tpoic M geo M pref The weighted results correspond to the thematic semantic layer, geospatial layer, and tourist preference layer, respectively. β1 to β3 are empirical weights, and the initial settings for the three weights are equal and summed to 1. The semantic weight matrix is ​​used to determine the multi-source data fusion weights and semantic propagation paths in subsequent steps.

[0030] The process of obtaining the multidimensional semantic feature vector includes: taking the scenic spot node as the target, extracting the association strength between the scenic spot node and other semantic entities from the semantic weight matrix; determining the fusion weight of the multi-source information related to the scenic spot node based on the association strength and the correlation of multi-source data in the semantic space; performing weighted fusion processing and semantic alignment processing on the multi-source data related to the scenic spot node according to the fusion weight to form fusion features; and performing feature unification and dimension mapping on the fusion features to obtain a multidimensional semantic feature vector representing the semantic state of the scenic spot node.

[0031] Specifically, for each scenic spot node, its association weights with all semantic entities are extracted from the semantic weight matrix, and a fusion coefficient is calculated by combining the corresponding data modality similarity. Multi-source data fusion employs a hierarchical weighting strategy: first, normalization and semantic alignment are performed within each modality, and then feature fusion is performed between modalities. The resulting multi-dimensional semantic feature vector encompasses cultural semantic relationships and reflects real-time visitor flow, service status, and environmental context features.

[0032] Step S2: Based on the multidimensional semantic feature vector and semantic weight matrix, perform semantic-spatiotemporal joint prediction in the semantic twin model through a spatiotemporal graph neural network to obtain the passenger flow trend prediction results of each scenic spot node. Based on the predicted passenger flow trend, generate the initial recommendation weights of each scenic spot node. In one embodiment of the present invention, the semantic twin model refers to a dual-channel representation model that simultaneously expresses "cultural semantic relationships" and "passenger flow temporal evolution" within a unified semantic space. It includes an encoding channel oriented towards static semantics and an encoding channel oriented towards temporal dynamics. The two channels perform consistency constraints and feature alignment at a fusion layer to generate a fused representation for prediction. The spatiotemporal graph neural network refers to a representation learning framework based on a graph structure, where the "spatial" dimension is represented by weighted associations determined by a semantic weight matrix, and the "temporal" dimension is represented by temporal modeling of historical observation sequences. The network completes feature aggregation and time updates through multi-layer propagation to output passenger flow trends for future target time periods.

[0033] The input of the semantic twin model is the multidimensional semantic feature vector of each scenic spot node and its correlation row in the semantic weight matrix; semantic propagation across nodes is completed in the semantic channel, trend extraction of historical sequence is completed in the time channel, weight coordination and attention selection are completed in the fusion layer, and the output is the quantitative result of passenger flow trend of the target time period (growth or decline trend, peak arrival time window, stable interval).

[0034] The joint prediction includes: using the multidimensional semantic feature vectors of scenic spot nodes as input, invoking the semantic association relationships between scenic spot nodes to establish a semantic association network; in the semantic twin model, constructing a spatiotemporal graph neural network structure based on the semantic association network, and controlling the information transmission weights between different scenic spot nodes according to the association strength between scenic spot nodes, and combining time series features to perform temporal propagation and update of the semantic feature vectors of scenic spot nodes; through the semantic propagation and time evolution process, calculating the passenger flow change trend of each scenic spot node in the target time period, and generating a predicted passenger flow trend.

[0035] The edge weights of the semantic association network come from the semantic weight matrix generated by S1. This matrix integrates semantic similarity, spatial proximity, and cultural theme relevance. Among them, scenic spot nodes and tourist profile elements are semantic entities, while historical figures, cultural events, theme elements, and geographical locations are the auxiliary contents of scenic spot nodes. Furthermore, tourist profile elements can include auxiliary contents (such as holiday travel preferences, language ability, and accessibility needs), which are consistent with the composition of the edge weights of this network segment.

[0036] Semantic association networks refer to graph structures with scenic spot nodes as vertices and strong relationships between entities given by semantic weight matrices as weighted edges. These networks are used to carry out semantic propagation paths and adjust the strength of relationships during the prediction phase. Temporal propagation and updating refer to the process of continuously updating the node state at discrete time steps. The update incorporates two types of information: a weighted aggregation from semantically adjacent nodes, and the node's own historical sequence and contextual factors (activities, weather, holidays).

[0037] The joint prediction process consists of three steps: First, generating the semantic adjacency weights for the current moment and completing a cross-node feature aggregation; second, extracting historical observations at multiple scales based on a sliding time window to form a composite representation of short-term and periodic trends; third, using an attention selection mechanism to weightedly fuse the features from "semantic aggregation" and "time extraction" to output the passenger flow trend results for the target time period. During the training phase, historical real passenger flow is used as the supervision signal, with added semantic consistency constraints to reduce the bias in trend patterns among semantically similar nodes.

[0038] The process of generating the initial recommendation weights for each attraction node includes: calculating the initial recommendation weights for each attraction node based on the predicted visitor flow trend; determining the reward and constraint parameters for recommendation optimization based on historical experience feedback data; and adaptively adjusting the initial recommendation weights under non-negative and normalized conditions to generate the initial recommendation weights for each attraction node.

[0039] User experience feedback data includes explicit ratings (1–5 points), textual tags (such as “clear explanations,” “crowded,” “value for money”), objective service indicators (average queue time, service response time, fault recovery time), complaint handling results, and willingness to revisit. Data is sourced from official navigation app interaction logs, on-site electronic evaluation terminals, and after-sales work order archives.

[0040] The initial recommendation weight refers to the basic quantitative value used for ranking and scheduling during the target period. Its setting follows the principle of "distribution and balance": attractions with higher predicted visitor flow are assigned lower weights, and attractions with lower predicted visitor flow are assigned higher weights, in order to achieve congestion management and balanced resource utilization. Reward parameters are derived from positive experience indicators (high satisfaction, low waiting time, high reputation stability), while constraint parameters are derived from negative experience or operational limitations (high complaint rate, SLA not meeting standards, facilities under maintenance). Both types of parameters are adjusted in a bounded manner to adjust the initial weights under the premise of non-negativity and normalization.

[0041] Initial ranking values ​​are generated for each attraction based on visitor flow trends; historical experience feedback is converted into reward and constraint signals, and the ranking values ​​are adaptively corrected in one go; normalization is performed to form a set of recommendation weights for downstream steps.

[0042] Step S3: Obtain and merge the predicted visitor density and service load coefficient of each scenic spot node, and then perform weighted processing in combination with the initial recommendation weight to obtain the dynamic priority score of each scenic spot node. The process of obtaining the dynamic priority score for each scenic spot node includes: determining the predicted passenger flow density based on the predicted passenger flow trend and historical passenger flow data of the scenic spot node; determining the service load coefficient based on the actual service capacity and real-time service load of the scenic spot node; weighting the predicted passenger flow density and service load coefficient based on preset weights to obtain the fusion value of the comprehensive status of the scenic spot node; and further weighting the fusion value according to the initial recommendation weights to generate the dynamic priority score of the corresponding scenic spot node.

[0043] In one embodiment of the present invention, this step aims to combine the passenger flow trend results obtained in the prediction stage with the real-time operating status of the scenic spot to form a dynamic priority score that reflects the comprehensive carrying capacity of the scenic spot and the quality of tourist experience, which is used for subsequent intelligent agent decision-making and personalized recommendations.

[0044] Predicted visitor density refers to the ratio of the number of visitors expected to enter or stay within a scenic spot node per unit time to the available space capacity, reflecting future congestion and traffic pressure. Service load coefficient refers to the ratio of the intensity of service resource usage at a scenic spot during the current time period to its maximum service capacity, used to quantify resource occupancy levels under operational conditions. Dynamic priority score refers to the result of comprehensively considering predicted visitor flow, service capacity, and recommendation weights, used to dynamically prioritize scenic spot visits; it is a core input parameter for visitor access decisions and reinforcement learning optimization.

[0045] The predicted passenger flow data comes from the output of the spatiotemporal graph neural network. After time sliding window smoothing and short-term anomaly detection, a predicted density sequence is generated. Service load data is accessed in real time from the operation logs and IoT device monitoring in the scenic area management system, such as camera passenger flow statistics, smart queuing terminals and facility health monitoring data. Historical passenger flow data is used to smooth the predicted values.

[0046] The predicted visitor density and service load coefficient are standardized within a unified range to form comparable quantitative indicators. The fusion ratio of the two is determined based on experience or training weights to obtain a comprehensive fusion value reflecting the real-time state of the attractions. The initial recommendation weights generated in step S2 are used as adjustment factors to further weight the fusion value, causing attractions with high recommendation weights but high loads to decrease in priority, while those with low recommendation weights and light loads to increase in priority. This process generates a priority score that can be dynamically adjusted over time. The generation of the dynamic priority score is performed cyclically at fixed time intervals, referencing the latest predicted visitor flow, real-time load, and feedback data each time it is updated, ensuring that the output priority sequence reflects the real-time state of the scenic area. The scoring results are stored in the semantic data space and used as input for subsequent multi-agent reinforcement learning (step S4) to drive the visitor agent's attraction selection decisions.

[0047] Step S4: Model each tourist as an intelligent agent, and use dynamic priority score as input feature. Calculate the collaborative behavior weights between intelligent agents based on the semantic weight matrix, and combine with a multi-agent reinforcement learning algorithm to obtain the selection probability distribution of each scenic spot node. In one embodiment of the present invention, in this step, individual tourists are abstracted as "intelligent agents," that is, computing units with independent decision-making capabilities. The decision objective of each intelligent agent is to select the optimal sequence of attractions to visit under given constraints, in order to maximize experience benefits and reduce the risk of congestion. This modeling method does not rely on external system calls, but is completed based on the data generated in the preceding steps. Its inputs include: dynamic priority scores of attraction nodes obtained in step S3; semantic relationships between attractions provided by the semantic weight matrix; tourist profile elements (interest themes, consumption preferences, travel types, time budgets, etc.); and real-time operational status (service load and passenger flow prediction results). The decision variable for each intelligent agent is the "next proposed attraction node," and the objective is to optimally select the access path under semantic and spatial constraints.

[0048] The process of obtaining the selection probability distribution of each scenic spot node includes: calculating the collaborative behavior weights among agents based on the semantic relationships between scenic spot nodes and between tourist profile elements; constructing a multi-agent interaction environment based on the collaborative behavior weights, and jointly modeling the state, actions, and rewards of each agent; in the multi-agent interaction environment, using a multi-agent reinforcement learning algorithm to iteratively train the agent's policy, taking dynamic priority scoring as the state input, and combining predicted passenger flow density and service load constraints to optimize the objective function of each agent; and outputting the selection probability distribution of each scenic spot node after the policy converges.

[0049] Collaborative behavior weights refer to the degree of behavioral correlation between different tourist agents, used to measure the similarity of their preferences, travel patterns, or interest topics in the semantic space. This weight is calculated based on the semantic similarity between tourist profile elements in the semantic weight matrix, and corrected by incorporating the spatial proximity of tourists within the same time window. The multi-agent interaction environment refers to a behavioral decision-making field composed of all tourist agents, where each agent independently makes action decisions based on the environmental state (i.e., current passenger flow, service load, and attraction priority). The environment updates its state after each round of decisions (e.g., changes in passenger flow distribution and attraction load), forming a cyclical feedback process. The reinforcement learning process refers to a training method that gradually adjusts the agent's decision-making strategy through trial-and-error interaction and reward feedback. Its essence is an iterative optimization process of "input state—select action—receive reward—update strategy," rather than a single algorithm call.

[0050] Each agent's input for decision-making includes three parts: (1) the global state of the environment: the dynamic priority scores and service load status of each attraction; (2) its own individual state: the tourist profile features and current itinerary progress; and (3) the weights of collaborative behaviors among agents: reflecting the group's behavioral trends and mutual influences. The output is a probability distribution of the visit selection of a set of attraction nodes, which describes the likelihood of the tourist selecting each candidate attraction in the current state.

[0051] The process of obtaining the selection probability distribution is as follows: All tourist agents read the initial state data, including the dynamic priority scoring table and semantic weight matrix generated in the previous step. Each agent calculates a set of candidate attractions based on the current visitor density and its own preferences, and uses the weight of collaborative behavior as an influencing factor. After an agent selects a visit action, it updates the environmental state, i.e., the real-time load of the corresponding attraction, the correction of visitor flow prediction, and the experience feedback parameters are updated accordingly. Each agent receives a reward value based on the results (such as queuing time, experience score, and cultural compatibility), which is used to update its strategy. After multiple rounds of interaction, the strategies of each agent tend to stabilize, and the group behavior converges to a set of selection probability distributions.

[0052] The experience feedback data comes from visitor logs and real-time interaction records; the weight of collaborative behavior is calculated by the similarity of visitor profile elements in the knowledge graph and verified by historical visit sequences; the reward items are mainly based on quantitative indicators such as visitor satisfaction rating, service waiting time and cultural preference matching degree.

[0053] This step does not directly call external algorithm models. Instead, within the existing data semantic structure and spatiotemporal constraints, it achieves dynamic optimal distribution learning of tourist behavior through a process of iterative training and state updates. Reinforcement learning is used only as a mathematical descriptive tool, its function being equivalent to an adaptive optimization process. The final generated selection probability distribution is jointly determined by the balance of cooperative behavior among agents and individual experience feedback, reflecting the dynamic decision-making trends of real tourist groups under the constraints of cultural themes and visitor flow distribution.

[0054] By abstracting individual tourists into interactive intelligent agents and constructing collaborative behavioral relationships under the constraints of a semantic weight matrix, this step achieves joint decision optimization based on three dimensions: cultural semantics, tourist preferences, and spatial status. The generated selection probability distribution not only reflects the group's travel patterns but also takes into account individual interest differences, making the allocation of scenic area resources more dynamically adaptable. This result will serve as the direct input for the next step of generating personalized visit sequences (S5), providing a quantitative basis for intelligent recommendation and experience guidance of cultural tourism services.

[0055] Step S5: Using the selection probability distribution as a constraint, call the sequence-to-sequence model and combine it with the attention distribution obtained from the semantic weight matrix to generate a personalized scenic spot visit sequence; In one embodiment of the present invention, in this step, based on the probability distribution of attraction selection output in the previous stage (S4), the visitor behavior is sequentially planned to form a personalized visit path that conforms to the individual characteristics and cultural preferences of the visitor.

[0056] The inputs for this step include: the probability distribution of each attraction node, used to limit the probability constraints of the candidate attraction set and its order of appearance; a semantic weight matrix, used to calculate the semantic relevance between attractions and between attractions and tourist profile elements; tourist profile elements (including theme interests, consumption level, travel time budget, companion type, etc.), used to guide the personalized adjustment of the visit sequence; and semantic structure information of the cultural knowledge graph, used to ensure that the recommended path conforms to cultural logic and geographical order, such as theme continuity and spatial accessibility.

[0057] The process of generating personalized attraction visit sequences includes: using the selection probability distribution of each attraction node as a constraint, calling a sequence-to-sequence model to perform sequence modeling of tourist visit behavior; during the sequence modeling process, combining the semantic association between attraction nodes and tourist profile elements to generate an attention distribution to guide the sequence generation process; dynamically adjusting the decoding weights of the sequence-to-sequence model according to the attention distribution to control the output order and occurrence probability of candidate attraction nodes; and generating personalized attraction visit sequences under the premise of satisfying the constraints.

[0058] In this method, the sequence-to-sequence model does not refer to a specific algorithmic framework, but rather to a mapping mechanism of "input behavior sequence—output visit sequence." Its core function is to generate the optimal visit order under probabilistic constraints and semantic context guidance. The model consists of two processes: encoding and decoding. The encoding process represents and compresses the input selection probability distribution and semantic features; the decoding process progressively outputs a sequence of attraction nodes under attention guidance. The attention distribution refers to the weight distribution used in the sequence generation process to measure the semantic matching degree between the current candidate attraction and the tourist profile elements. Its role is to dynamically adjust the focus position of the generation order, making the output sequence more consistent with cultural themes and individual preferences. The personalized attraction visit sequence refers to the set of visit paths generated under semantic, spatial, cultural theme, and tourist feature constraints, and the output result is a set of attraction node sequences ordered by time.

[0059] The personalized attraction visit sequence process includes: using the selection probability distribution obtained in step S4 as the initial set of candidate attractions, each attraction node carries a corresponding probability value to control its order and frequency of appearance in the visit sequence. The correlation strength between attraction nodes and tourist profile elements in the semantic weight matrix is ​​read, identifying attractions with "high semantic relevance" at the cultural theme level, such as locations associated with historical figures, exhibition halls with the same theme, or adjacent cultural event nodes. Under the above semantic constraints, the tourist's current state (visited nodes, time budget, interest shift, etc.) is represented as a semantic vector through an encoding process, and the weight ranking of the next candidate attractions is calculated in conjunction with the attention distribution. At each output round, the decoding weights are updated based on the real-time attention distribution, causing the model to favor attraction nodes with high semantic matching with tourist profile features, moderate geographical distance, and low service load, avoiding congestion or cultural theme jumps. When the cumulative visit duration or the number of candidate nodes meets the constraints set by the tourist, the final visit sequence result is output.

[0060] The probability distribution is output from the reinforcement learning stage, reflecting the balance between group behavior trends and individual preferences; the semantic relationships and cultural theme structures of attractions come from the entity relationship table of the knowledge graph; the geographical location data is provided by the scenic area's geographic information system and is used to calculate spatial continuity; the tourist profile data comes from tourist registration or historical visit logs to ensure that the recommendation results have individual consistency.

[0061] The sequence-to-sequence generation process is a dynamic optimization process. During generation, each step does not involve a fixed algorithm output, but rather a decision based on a combination of factors including the current access state, semantic association weights, and probabilistic constraints. This process iterates continuously until the access sequence satisfies three types of constraints: time budget, cultural theme continuity, and service load balancing.

[0062] This step achieves personalized planning of tourist access routes by using a sequence generation mechanism constrained by a selection probability distribution and guided by a semantic weight matrix. The generated access sequences balance the logical continuity of cultural themes with the rationality of spatial itineraries, reducing overcrowding at attractions while enhancing the coherence and cultural immersion of the tourist experience. Compared with traditional recommendation methods based on distance or ratings, this method achieves intelligent guidance at the cultural semantic level, enabling tourist routes to have dynamic self-adaptation and personalized optimization capabilities.

[0063] Step S6: When the intelligent agent arrives at the scenic spot node in the scenic spot visit sequence, the interaction is triggered and the experience feedback data is recorded; the multi-dimensional semantic feature vector and experience feedback data are written into the blockchain to form a credit record, and the service quality is verified and a credit score is generated through the on-chain smart contract; In one embodiment of the present invention, in this step, after a tourist (i.e., an intelligent agent) arrives at any scenic spot node in the visit sequence, an interactive feedback event is triggered to collect tourist on-site experience data and service quality information, and the information is matched and stored with the multidimensional semantic features of the aforementioned scenic spot node to form a credible and verifiable cultural service quality credit record.

[0064] The process of generating a credit score includes: writing the multidimensional semantic feature vector of the attraction node and the corresponding experience feedback data into the blockchain to form an immutable credit record; calling the smart contract deployed on the chain to automatically verify and compare the service quality indicators in the experience feedback data to determine whether the service meets the preset quality standards; and generating a credit score for the corresponding attraction node or service provider based on the verification results of the smart contract, taking into account tourist satisfaction, service compliance, and historical credit records.

[0065] The experience feedback data primarily comes from three categories: subjective feedback data, including visitor satisfaction ratings, text comment tags, voice evaluation content, and interactive emoticons; objective operational data, including average visitor dwell time at the attraction, queuing time, equipment usage frequency, guided tour call records, and venue congestion index; and service delivery records, automatically submitted to the blockchain by the service provider, including service timestamps, facility operational status, complaint and response records, etc. The multidimensional semantic feature vector comes from the semantic feature space constructed in step S1, representing the attraction's cultural theme attributes, spatial location characteristics, historical event associations, and visitor group characteristics. After the feedback data is bound to the feature vector, a complete on-chain record structure of "experience event—semantic entity—service record" is formed.

[0066] Blockchain credit records refer to a multi-dimensional data storage mechanism based on decentralized ledger technology. Each credit record includes a scenic spot node identifier, multi-dimensional semantic features, experience feedback data, and a timestamp. A unique index is generated through a hash algorithm to ensure data immutability and full-chain traceability. Smart contracts are sets of automatically executed rules deployed on the blockchain network, used to verify the authenticity of feedback data, calculate indicators, and generate scores without human intervention. Smart contracts incorporate cultural tourism service quality standard parameters (such as response time thresholds, complaint handling timeliness, and equipment availability standards), automatically comparing indicator values ​​in the feedback data to determine whether the service meets the standards. Credit scores are quantitative evaluation values ​​based on on-chain verification results, used to measure the service reliability of scenic spots or service providers. This score comprehensively considers three indicators: tourist satisfaction, service quality compliance, and historical credit stability, and is publicly verifiable on the blockchain in the form of a hash index.

[0067] When a tourist agent arrives at a scenic spot node, the positioning system and access sequence matching module automatically identify the arrival event, triggering a data collection interface to collect the tourist's experience feedback data at that node. The collected experience feedback data is matched with the semantic feature vector corresponding to that node in the semantic weight matrix, and then encrypted using a hash algorithm to generate an on-chain data structure. The generated data is written to the blockchain storage node via a distributed ledger, triggering the execution of an on-chain smart contract. The smart contract automatically reads service quality-related fields, performs verification and consistency comparison against preset cultural tourism service standards, and generates a credit confirmation record upon successful verification. After verification, the smart contract calculates the credit score for the scenic spot or service provider by combining the average tourist satisfaction rate, service compliance rate, and complaint closure rate from historical records. The score result is bound to the corresponding transaction hash, forming a complete on-chain credit trajectory.

[0068] The experience feedback data comes from a two-way channel between active submissions from tourists' devices and automatic recordings from the server; semantic feature vectors correspond one-to-one with entities in the knowledge graph, ensuring the consistency of feedback information with cultural themes; blockchain ledger nodes are set up among different management institutions (such as scenic area management, cultural departments, and service provider alliances) to ensure the openness and mutual trust of credit records; the scoring rules of smart contracts can be dynamically updated according to industry standards, making the scoring system evolvable.

[0069] Step S7: Update the cultural knowledge graph and multi-agent reinforcement learning algorithm based on credit scores and experience feedback data.

[0070] In one embodiment of the present invention, this step constitutes the self-learning and evolutionary stage of the entire smart cultural tourism integrated management method. The core of this stage lies in using the credit score and experience feedback data obtained from the previous stage of blockchain notarization to dynamically optimize the attribute weights of semantic entities and multi-agent decision-making strategies in the knowledge graph.

[0071] The cultural knowledge graph is updated by automatically reducing the semantic weight between a scenic spot node and thematic entities such as "high-quality service" and "high satisfaction" when the credit score of the node drops significantly; conversely, increasing the association weight when the score improves. The similarity weight is dynamically adjusted based on statistical learning of the matching frequency in feedback data to determine the relationships between tourist profile elements and scenic spot nodes. For example, if a large number of tourists give a scenic spot high ratings in the "family-themed" scenario, the association weight between the "family-themed" category and that node is increased. Geographical adjacency weights and cultural theme similarity are re-normalized using the latest spatiotemporal statistical data to ensure the timeliness and topological stability of the graph. The updated cultural knowledge graph more accurately reflects the current state of cultural and tourism resources and tourist preference trends at the semantic level, forming a dynamically expandable cultural semantic network that provides more timely semantic support for subsequent prediction and recommendation processes.

[0072] The updated multi-agent reinforcement learning algorithm includes: using blockchain-recorded credit scores as positive reward signals and low-scoring nodes or negative feedback events as penalty signals, thereby reshaping the agent's reward function. The agent's experience replay pool is updated using experiential feedback data, with new samples including tourist visit paths, feedback scores, and behavioral outcomes to improve the diversity and realism of training samples. While maintaining the stability of historical strategies, iterative training adjusts the agent's policy parameters, making the agent more inclined to choose high-credit attractions or service providers in future decisions, thus forming a positive ecological feedback loop. The updated multi-agent strategy possesses self-learning characteristics, dynamically optimizing its decisions based on real tourist behavior and feedback, making the recommendation results adaptive in the time dimension and maintaining cultural orientation consistency in the spatial dimension.

[0073] All data required for updates comes from on-chain credit records and knowledge graph entities generated in the preceding steps, with no external dependencies; the update process follows the principles of weight smoothing and time decay to avoid structural instability caused by short-term abnormal data; the comprehensive adjustment of cultural theme weights, geographical adjacency weights, and behavioral preference weights is consistent with the semantic weight matrix generated in step S1; the update frequency can be automatically triggered based on the amount of data accumulated or a set period (such as daily or weekly) to ensure the real-time performance and computational controllability of the semantic system.

[0074] This step establishes a cyclical, self-evolving mechanism between cultural knowledge and tourist behavior. The introduction of credit scoring provides an objective basis for semantic weight updates, while the application of experience feedback data enables the model to continuously learn. This update mechanism forms a closed-loop feedback system from tourist experience to knowledge structure and then to decision optimization, enabling smart cultural tourism management to transition from static recommendation to dynamic optimization, significantly improving adaptability, transparency, and the quality of intelligent decision-making in complex cultural tourism environments.

[0075] Example 2, a smart cultural tourism integrated management system, see [link / reference] Figure 1 As shown, it includes the following modules: Semantic construction module: Collects multi-source data of cultural and tourism areas, calls cultural knowledge graph and generates semantic weight matrix; determines the fusion weight of multi-source data based on semantic weight matrix, performs fusion processing on various types of data, and obtains multi-dimensional semantic feature vector; Passenger flow recommendation module: Based on multi-dimensional semantic feature vectors and semantic weight matrices, semantic-spatiotemporal joint prediction is performed in the semantic twin model through spatiotemporal graph neural network to obtain the passenger flow trend prediction results of each scenic spot node. Based on the predicted passenger flow trend, the initial recommendation weights of each scenic spot node are generated. Priority assessment module: Obtain and integrate the predicted visitor density and service load coefficient of each scenic spot node, and then perform weighted processing in combination with the initial recommendation weight to obtain the dynamic priority score of each scenic spot node. Selection Evaluation Module: Each tourist is modeled as an agent, and dynamic priority scores are used as input features. The collaborative behavior weights between agents are calculated based on the semantic weight matrix. Combined with a multi-agent reinforcement learning algorithm, the selection probability distribution of each scenic spot node is obtained. The visit sequence module: with the selection probability distribution as a constraint, it calls the sequence to the sequence model and combines the attention distribution obtained by the semantic weight matrix to generate a personalized visit sequence for attractions; Credit Record Module: When the intelligent agent arrives at the scenic spot node in the scenic spot visit sequence, it triggers the interaction and records the experience feedback data; the multi-dimensional semantic feature vector and experience feedback data are written into the blockchain to form a credit record, and the service quality is verified and a credit score is generated through on-chain smart contracts; Dynamic update module: Updates the cultural knowledge graph and multi-agent reinforcement learning algorithm based on credit scores and experience feedback data.

[0076] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart cultural tourism integrated management method, characterized in that, include: Collect multi-source data from cultural and tourism areas, call up cultural knowledge graphs, and generate semantic weight matrices; Based on the semantic weight matrix, the fusion weights of multi-source data are determined, and various types of data are fused to obtain multi-dimensional semantic feature vectors. Based on multidimensional semantic feature vectors and semantic weight matrices, semantic-spatiotemporal joint prediction is performed in the semantic twin model through a spatiotemporal graph neural network to obtain the passenger flow trend prediction results of each scenic spot node. Based on the predicted passenger flow trend, the initial recommendation weights of each scenic spot node are generated. The predicted visitor density and service load coefficient of each scenic spot node are obtained and integrated, and then further weighted in combination with the initial recommendation weight to obtain the dynamic priority score of each scenic spot node. Each tourist is modeled as an intelligent agent, and dynamic priority scores are used as input features. The collaborative behavior weights between agents are calculated based on the semantic weight matrix. Combined with a multi-agent reinforcement learning algorithm, the selection probability distribution of each scenic spot node is obtained. By using the selection probability distribution as a constraint, the sequence-to-sequence model is called, and the attention distribution obtained by combining the semantic weight matrix is ​​used to generate a personalized scenic spot visit sequence. When the agent is detected to have arrived at a scenic spot node in the scenic spot visit sequence, the interaction is triggered and the experience feedback data is recorded. Multidimensional semantic feature vectors and experience feedback data are written into the blockchain to form credit records, and service quality is verified and credit scores are generated through on-chain smart contracts. Based on credit scores and experience feedback data, the cultural knowledge graph and multi-agent reinforcement learning algorithm are updated.

2. The intelligent cultural tourism integrated management method according to claim 1, characterized in that, The generation of the semantic weight matrix includes: calling a pre-constructed cultural knowledge graph, which is built based on publicly available and reliable cultural tourism semantic resources, using scenic spot nodes and tourist profile elements as semantic entities, and constructing semantic associations through relation triples and their attribute weights; the attribute weights are used to characterize the importance or association strength of the relationship between entities, and are determined comprehensively based on entity attribute similarity, relationship frequency, temporal relevance, and cultural theme relevance; performing semantic embedding processing on text, image, audio, and sensor data samples included in the cultural knowledge graph to obtain vectorized representations of semantic entities; constructing an initial semantic matrix by combining semantic similarity, spatial proximity, and cultural theme relevance between semantic entities; and normalizing and hierarchically weighting the initial semantic matrix to generate a semantic weight matrix by weighting the cultural theme weight, geographical adjacency weight, and behavioral preference weight.

3. The intelligent cultural tourism integrated management method according to claim 1, characterized in that, The process of obtaining the multidimensional semantic feature vector includes: taking the scenic spot node as the target, extracting the association strength between the scenic spot node and other semantic entities from the semantic weight matrix; determining the fusion weight of the multi-source information related to the scenic spot node based on the association strength and the correlation of multi-source data in the semantic space; performing weighted fusion processing and semantic alignment processing on the multi-source data related to the scenic spot node according to the fusion weight to form fusion features; and performing feature unification and dimension mapping on the fusion features to obtain a multidimensional semantic feature vector representing the semantic state of the scenic spot node.

4. The intelligent cultural tourism integrated management method according to claim 1, characterized in that, The joint prediction includes: using the multidimensional semantic feature vectors of scenic spot nodes as input, invoking the semantic association relationships between scenic spot nodes to establish a semantic association network; in the semantic twin model, constructing a spatiotemporal graph neural network structure based on the semantic association network, and controlling the information transmission weights between different scenic spot nodes according to the association strength between scenic spot nodes, and combining time series features to perform temporal propagation and update of the semantic feature vectors of scenic spot nodes; through the semantic propagation and time evolution process, calculating the passenger flow change trend of each scenic spot node in the target time period, and generating a predicted passenger flow trend.

5. The intelligent cultural tourism integrated management method according to claim 1, characterized in that, The process of generating the initial recommendation weights for each attraction node includes: calculating the initial recommendation weights for each attraction node based on the predicted visitor flow trend; determining the reward and constraint parameters for recommendation optimization based on historical experience feedback data; and adaptively adjusting the initial recommendation weights under non-negative and normalized conditions to generate the initial recommendation weights for each attraction node.

6. The intelligent cultural tourism integrated management method according to claim 1, characterized in that, The process of obtaining the dynamic priority score for each scenic spot node includes: determining the predicted passenger flow density based on the predicted passenger flow trend and historical passenger flow data of the scenic spot node; determining the service load coefficient based on the actual service capacity and real-time service load of the scenic spot node; weighting the predicted passenger flow density and service load coefficient based on preset weights to obtain the fusion value of the comprehensive status of the scenic spot node; and further weighting the fusion value according to the initial recommendation weights to generate the dynamic priority score of the corresponding scenic spot node.

7. The intelligent cultural tourism integrated management method according to claim 1, characterized in that, The process of obtaining the selection probability distribution of each scenic spot node includes: calculating the collaborative behavior weights among agents based on the semantic relationships between scenic spot nodes and between tourist profile elements; constructing a multi-agent interaction environment based on the collaborative behavior weights, and jointly modeling the state, actions, and rewards of each agent; in the multi-agent interaction environment, using a multi-agent reinforcement learning algorithm to iteratively train the agent's policy, taking dynamic priority scoring as the state input, and combining predicted passenger flow density and service load constraints to optimize the objective function of each agent; and outputting the selection probability distribution of each scenic spot node after the policy converges.

8. The intelligent cultural tourism integrated management method according to claim 1, characterized in that, The process of generating personalized attraction visit sequences includes: using the selection probability distribution of each attraction node as a constraint, calling a sequence-to-sequence model to perform sequence modeling of tourist visit behavior; during the sequence modeling process, combining the semantic association between attraction nodes and tourist profile elements to generate an attention distribution to guide the sequence generation process; dynamically adjusting the decoding weights of the sequence-to-sequence model according to the attention distribution to control the output order and occurrence probability of candidate attraction nodes; and generating personalized attraction visit sequences under the premise of satisfying the constraints.

9. The intelligent cultural tourism integrated management method according to claim 1, characterized in that, The process of generating a credit score includes: writing the multidimensional semantic feature vector of the attraction node and the corresponding experience feedback data into the blockchain to form an immutable credit record; calling the smart contract deployed on the chain to automatically verify and compare the service quality indicators in the experience feedback data to determine whether the service meets the preset quality standards; and generating a credit score for the corresponding attraction node or service provider based on the verification results of the smart contract, taking into account tourist satisfaction, service compliance, and historical credit records.

10. A smart cultural tourism integrated management system, characterized in that, The system applies a smart cultural tourism integrated management method according to any one of claims 1 to 9, including: Semantic construction module: Collects multi-source data of cultural and tourism areas, calls cultural knowledge graph and generates semantic weight matrix; determines the fusion weight of multi-source data based on semantic weight matrix, performs fusion processing on various types of data, and obtains multi-dimensional semantic feature vector; Passenger flow recommendation module: Based on multi-dimensional semantic feature vectors and semantic weight matrices, semantic-spatiotemporal joint prediction is performed in the semantic twin model through spatiotemporal graph neural network to obtain the passenger flow trend prediction results of each scenic spot node. Based on the predicted passenger flow trend, the initial recommendation weights of each scenic spot node are generated. Priority assessment module: Obtain and integrate the predicted visitor density and service load coefficient of each scenic spot node, and then perform weighted processing in combination with the initial recommendation weight to obtain the dynamic priority score of each scenic spot node. Selection Evaluation Module: Each tourist is modeled as an agent, and dynamic priority scores are used as input features. The collaborative behavior weights between agents are calculated based on the semantic weight matrix. Combined with a multi-agent reinforcement learning algorithm, the selection probability distribution of each scenic spot node is obtained. The visit sequence module: with the selection probability distribution as a constraint, it calls the sequence to the sequence model and combines the attention distribution obtained by the semantic weight matrix to generate a personalized visit sequence for attractions; Credit Record Module: When the intelligent agent arrives at the scenic spot node in the scenic spot visit sequence, it triggers the interaction and records the experience feedback data; the multi-dimensional semantic feature vector and experience feedback data are written into the blockchain to form a credit record, and the service quality is verified and a credit score is generated through on-chain smart contracts; Dynamic update module: Updates the cultural knowledge graph and multi-agent reinforcement learning algorithm based on credit scores and experience feedback data.

Images