Rural industry service data aggregation and atlas construction method and system

Abstract

The invention provides a rural industry service data aggregation and atlas construction method and system, and the method comprises the steps: obtaining Internet text data and Internet of Things sensing data, extracting service entity information in the Internet text data through a semantic analysis model, and enabling the service entity information to comprise address text information; inputting the service entity information and the sensing data of the Internet of Things into the feature fusion model to generate a service data tuple; performing vectorization processing on the address text information in the service data tuple to generate an address semantic vector; a corresponding standard address is matched in a preset standard address base based on the address semantic vector, and geographic coordinates of the service data tuple are determined; constructing a knowledge graph based on the service data tuple and the geographic coordinates, and establishing a joint index structure to store the knowledge graph; the joint index structure comprises a spatial index dimension for dynamically adjusting the boundary based on the entity distribution density and a time index dimension for grading based on the entity attribute updating frequency.

Description

Technical Field

[0001] This application relates to the fields of data processing and knowledge graph technology, and in particular to a method and system for data aggregation and graph construction of rural industry services. Background Technology

[0002] Rural industry services (such as agricultural machinery scheduling, agricultural input distribution, and technical guidance) generate massive amounts of data. Building an efficient rural industry knowledge graph is crucial for achieving precise resource allocation and digital management of the industry.

[0003] Rural industry services are characterized by: extremely dispersed and complex data sources; highly non-standardized address descriptions in rural areas; and significant spatiotemporal dynamics. Traditional knowledge graph storage is mostly based on static nodes or simple attribute timestamps, lacking an efficient spatiotemporal indexing mechanism. When users need to perform complex spatiotemporal mixed queries (e.g., "find harvesters active within a county in the past week"), the system needs to perform a full graph scan or inefficient filtering, resulting in high response latency and difficulty in meeting real-time scheduling requirements.

[0004] Therefore, there is an urgent need for a method for rural industry service data aggregation and graph construction that can intelligently aggregate multi-source heterogeneous data, accurately resolve non-standard addresses, and support efficient spatiotemporal queries. Summary of the Invention

[0005] This application provides a method and system for rural industry service data aggregation and map construction to solve the shortcomings of existing technologies, such as difficulty in heterogeneous data fusion, inaccurate rural address positioning, and low spatiotemporal query efficiency.

[0006] Firstly, this application provides a method for rural industry service data aggregation and map construction, including: Acquire internet text data and IoT sensor data, and use a semantic analysis model to extract service entity information from the internet text data, including address text information. The service entity information and IoT sensor data are input into a feature fusion model to generate service data tuples. The address text information in the service data tuple is vectorized to generate address semantic vectors; Based on the address semantic vector, the corresponding standard address is matched in the pre-set standard address library to determine the geographic coordinates of the service data tuple; A knowledge graph is constructed based on service data tuples and geographic coordinates, and a joint index structure is established to store the knowledge graph; The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

[0007] Secondly, this application also provides a rural industry service data aggregation and mapping system, including: The acquisition module is used to acquire Internet text data and IoT sensor data, and to extract service entity information from Internet text data using a semantic analysis model. Service entity information includes address text information. The first generation module is used to fuse service entity information with IoT sensor data input features into a model to generate service data tuples. The second generation module is used to vectorize the address text information in the service data tuple to generate address semantic vectors. The determination module is used to match the corresponding standard address in a pre-set standard address library based on the address semantic vector to determine the geographic coordinates of the service data tuple; The building module is used to construct a knowledge graph based on service data tuples and geographic coordinates, and to establish a joint index structure to store the knowledge graph; The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

[0008] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described methods for rural industry service data aggregation and mapping.

[0009] Fourthly, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described methods for rural industry service data aggregation and map construction.

[0010] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements any of the above-described methods for rural industry service data aggregation and mapping.

[0011] The method and system for rural industry service data aggregation and graph construction provided in this application extract service entity information from Internet text data through a semantic analysis model, directly understanding the semantics of web page content rather than parsing its structure, thus achieving automated and highly robust data aggregation and improving the utilization and integrity of data sources. By integrating service entity information with IoT sensor data input feature fusion models, service data tuples are generated, achieving deep fusion of cross-modal data. By vectorizing address text information and matching corresponding standard addresses based on address semantic vectors, the success rate and accuracy of geocoding are improved, increasing the accuracy of non-standard address matching and enhancing the fault tolerance of address matching. By establishing a joint index to store the knowledge graph, the spatial index dimension that dynamically adjusts the boundary based on entity distribution density improves the efficiency of spatial queries. The time index dimension that is hierarchically based on the update frequency of entity attributes avoids storage waste caused by redundant indexes for static data, while ensuring high performance for time-slice queries on dynamic data, thus achieving real-time response to complex mixed queries. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the method for data aggregation and mapping of rural industry services provided in this application; Figure 2 This is a schematic diagram of the structure of the knowledge graph provided in an embodiment of this application; Figure 3 This is a schematic diagram of the architecture of the prototype system for the rural industry spatiotemporal map service provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the rural industry service data aggregation and mapping system provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] Figure 1 This is a flowchart illustrating the rural industry service data aggregation and map construction method provided in this application, such as... Figure 1 As shown, the method includes the following steps: S101. Obtain Internet text data and IoT sensor data, and use a semantic analysis model to extract service entity information from the Internet text data. The service entity information includes address text information. S102. Integrate service entity information with IoT sensor data input feature fusion model to generate service data tuples; S103. Vectorize the address text information in the service data tuple to generate an address semantic vector; S104. Based on the address semantic vector, match the corresponding standard address in the preset standard address library to determine the geographic coordinates of the service data tuple; S105. Construct a knowledge graph based on service data tuples and geographic coordinates, and establish a joint index structure to store the knowledge graph.

[0016] The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

[0017] The rural industry service data aggregation and graph construction method provided in this application extracts service entity information from Internet text data through a semantic analysis model, directly understanding the semantics of web page content rather than parsing its structure, achieving automated and highly robust data aggregation, and improving the utilization rate and integrity of data sources. By integrating service entity information with IoT sensor data input feature fusion models, service data tuples are generated, achieving deep fusion of cross-modal data. By vectorizing address text information and matching corresponding standard addresses based on address semantic vectors, the success rate and accuracy of geocoding are improved, increasing the accuracy of non-standard address matching and enhancing the fault tolerance of address matching. By establishing a joint index to store the knowledge graph, the spatial index dimension that dynamically adjusts the boundary based on entity distribution density improves the efficiency of spatial queries. The time index dimension that is graded based on the update frequency of entity attributes avoids storage waste caused by redundant indexes for static data, while ensuring high performance for time-slice queries of dynamic data, achieving real-time response to complex mixed queries.

[0018] S101. Obtain Internet text data and IoT sensor data, and use a semantic analysis model to extract service entity information from the Internet text data.

[0019] Specifically, the data sources include two dimensions: first, internet text data obtained through web crawlers from agricultural websites, e-commerce platforms, and other sources; and second, IoT sensor data such as meteorological and soil data obtained through sensors deployed in the fields.

[0020] Internet text data is characterized by its varied structure. Semantic analysis models are employed to extract service entity information from this data, enabling semantic perception and entity recognition. Optionally, the semantic analysis model is built upon the Qwen2.5-3B large language model deployed within the vLLM framework.

[0021] Through pre-configured prompt engineering, the large speech model can directly read the original HTML or text content of web pages. Utilizing its powerful natural language understanding capabilities, it can automatically identify and extract key service entity information such as "service entity name", "service price", "contact information", and "address text information" (e.g., "XX County XX Township XX Village 3 Group").

[0022] Even if the HTML tag structure of a webpage changes, as long as the readable semantic content remains the same, the large language model can still accurately extract service entity information from internet text data, achieving adaptive data collection.

[0023] In some embodiments, step S101 involves extracting service entity information from Internet text data using a semantic analysis model, including: Input Internet text data into a semantic analysis model, and identify and extract key fields of service organization, service price and job type according to preset instructions; Monitor the confidence level of the extraction results output by the semantic analysis model, and adjust the extraction strategy for Internet text data based on the confidence level feedback.

[0024] Specifically, pre-set prompt templates for rural industries are used as preset instructions. When unstructured internet text data (such as webpage source code containing HTML tags or plain text) is input into the semantic analysis model, preset instructions are attached, such as "Please analyze the following text and extract the name of the organization providing the service, the specific service price and unit, and the specific agricultural machinery operation type (such as deep tillage, plant protection, and harvesting)".

[0025] The semantic analysis model receives raw internet text data W according to preset instructions, uses an attention mechanism to focus on relevant regions in the internet text data W, identifies and extracts semantic units related to "industry status / parameters," "service organization information," "service type and description," "service price and unit specifications," and "evaluation feedback" from the internet text data W, and then labels and extracts the results D. In one possible implementation, the semantic analysis model extracts key fields such as service organization (e.g., "XX agricultural machinery professional cooperative"), service price (e.g., "60 yuan / mu"), and operation type (e.g., "wheat harvesting"), and outputs structured JSON format data.

[0026] For example, for internet text T, the service entity information output by the semantic analysis model is: {Industry, Service Provider, Price, Specification, Evaluation}, along with the position or weight of each key field in the original text.

[0027] During the extraction process, the confidence level of the extraction results output by the semantic analysis model is monitored in real time. The confidence level can be calculated based on the probability distribution of the tokens generated by the model, or it can be calculated based on the integrity verification of the extracted fields (for example, if the extracted "price" field is not numeric, the confidence level is reduced).

[0028] A confidence threshold (e.g., 0.85) is pre-set. When the confidence of the extracted result is higher than the threshold, the result is directly adopted and proceeds to the next step. When the confidence is lower than the threshold, it indicates that the current page structure is unusual or the text is obscure, and the model has failed to extract information with certainty. In this case, the system triggers a feedback adjustment mechanism to dynamically adjust the extraction strategy. Adjustment methods include, but are not limited to: Prompt Refinement: Automatically switches to a more detailed prompt template, such as adding a "few-shot" example, showing the model several similar web page texts and the correct extraction results, guiding the model to perform analogical reasoning.

[0029] Context window adjustment: Expand or shrink the text context window of the input model, remove irrelevant navigation bar or advertising noise, and refocus on the main text content for extraction again.

[0030] This confidence-based closed-loop feedback mechanism can adaptively respond to the ever-changing webpage structures on the Internet, significantly improving the recall and accuracy of data aggregation.

[0031] In one possible implementation, based on the inference capabilities provided by vLLM, and combined with a semantic analysis model fine-tuned for rural industries, the system automatically identifies and extracts different types of key fields related to the core business when encountering specific website or webpage content, and outputs service entity information. Optionally, the fine-tuning of the semantic analysis model can specifically employ Layerwise Importance Sampling (LISA) fine-tuning technology.

[0032] S102. Integrate service entity information with IoT sensor data input feature model to generate service data tuples.

[0033] Specifically, since internet text data consists of discrete symbols while IoT data consists of continuous numerical signals, they belong to different modalities. To achieve deep fusion, a feature fusion model was constructed. This model maps the service entity information extracted by S101 to predefined service data tuples and maps the service entity information to predefined specific data fields and relationships within the feature fusion model, thus realizing domain knowledge mapping and attribute association. In this process, a cross-modal semantic matching mechanism is introduced to match the textual service entity information with multimodal attribute descriptions in the predefined feature fusion model. For example, it associates textual descriptions with corresponding image feature standards in the feature fusion model, improving the accuracy of the mapping.

[0034] The feature fusion model employs a lightweight meta-learner, specifically a spatially aware lightweight meta-learner (SALM).

[0035] In some embodiments, the feature fusion model employs a meta-learner containing a two-stream network structure, and the meta-learner's processing includes: Extract textual features from service entity information and numerical features from IoT sensor data, respectively; The association weights between text features and numerical features are determined based on the attention mechanism, and the association weights are gating and adjusted based on the spatial sparsity of the data distribution to output service data tuples.

[0036] Specifically, the meta-learner SALM incorporates a bi-directional spatial attention mechanism. Textual and sensor data streams are processed by independent encoders to extract features, and then the attention module calculates the association weights between them. For example, the extracted textual description of "drought in a certain area" is aligned with the values ​​from soil moisture sensors in that area.

[0037] Dual-stream feature extraction extracts textual features of service entity information and numerical features of IoT sensor data, respectively.

[0038] The meta-learner SALM first constructs two parallel data processing channels (two-stream structure): Text Stream: The extracted service entity information (such as service descriptions and evaluation content) of the Internet text data is input into a text encoder (e.g., a Transformer-based encoding layer) to map the unstructured text into high-dimensional text features.

[0039] Numerical stream: The acquired IoT sensor data (such as soil moisture, air temperature, pH and other time-series data) is input into a numerical encoder (such as a multilayer perceptron MLP or a one-dimensional convolutional neural network) to extract numerical features that represent the environmental state.

[0040] The association weights between text features and numerical features are determined based on an attention mechanism. To integrate these two heterogeneous modalities, a dual-channel attention mechanism is introduced to calculate the alignment degree between text features and numerical features in the semantic space.

[0041] For example, if the text contains the description "drought" and the sensor value shows "soil moisture <20%", the attention mechanism will identify the high correlation between the two and assign a high association weight; conversely, if the text describes "paddy field" but the sensor shows dryness, the weight will be reduced or it will be marked as abnormal.

[0042] In one possible implementation, the SALM module achieves deep coupling between the Visual Transformer (ViT) and the Text Transformer by replacing the traditional cross-attention layer of BLIP. The input directly uses visual features extracted from the BLIP ViT backbone network. Text features generated by the BERT encoder .

[0043] When textual and numerical features are received, the association weight is calculated using the following formula. : in, and These are visual features and text features, respectively; The projection matrix is ​​learnable; For bias terms; Use the sigmoid activation function. Output feature map. The calculation is performed using the following formula: in, This is for the splicing operation. This design achieves adaptive feature fusion under spatial correlation while preserving modality-specific semantic information.

[0044] The association weights are gating based on the spatial sparsity of the data distribution. Considering the extremely uneven spatial distribution of rural data (dense data in some demonstration areas and sparse data in remote mountainous areas), the meta-learner introduces a heterogeneous data stream gating fusion mechanism (Gating Attention for Cross-Domain Fusion) to dynamically adjust the weights of different modalities according to data quality.

[0045] The meta-learner SALM calculates a sparsity index based on the geographic location corresponding to the currently processed service data tuples, combined with pre-statistical regional data density. A pre-defined gating network is used. When spatial sparsity is high (i.e., few data samples), the gating network automatically adjusts its parameters, increasing the weight of prior knowledge in the meta-learner and reducing overfitting to the few current samples, thus ensuring generalization ability in cases with few samples. When spatial sparsity is low (i.e., abundant data), the gating network allows more observational data features to pass through, relying primarily on the current samples for accurate fusion.

[0046] The gating coefficient is calculated using the following formula. : in, Indicates the main feature and auxiliary features Piecing them together, As weight, This is the bias term. Final fusion result. It is expressed as follows: A dual-weighting mechanism enables balanced cross-domain information integration.

[0047] Furthermore, to address the data sparsity problem in the agricultural field, the meta-learner SALM also introduces a meta-learning sub-network based on a Transformer encoder with dimensionality reduction parameters, utilizing a limited auxiliary dataset. Adaptive tuning parameters for training specific tasks : in, It adopts a two-layer structure; Aggregate support set embeddings. These parameters dynamically adjust spatial attention weights during inference, enabling the model to quickly adapt to unseen agricultural scenarios.

[0048] Exquisite features The multimodal fusion head passed to BLIP replaces the original cross-modal features. The meta-learner SALM adaptively adjusts its parameters. Update during training to ensure context-aware attention weight adjustments.

[0049] After attention-weighting and gating, the feature vectors are reconstructed by the decoder, ultimately outputting a service data tuple in a unified format (e.g., <timestamp, source ID, entity type, field 1: value 1, field 2: value 2, ...>). This tuple not only contains the fused attribute values ​​but also implicitly includes confidence information of the multimodal data, providing a high-quality data foundation for subsequent graph construction and solving the problem of unified representation of heterogeneous data. S103. Vectorize the address text information in the service data tuple to generate address semantic vectors.

[0050] Specifically, to address the issue of non-standard rural address descriptions, direct string matching is avoided. The original address text (e.g., "Hongxing Farm Branch 2") undergoes word segmentation and noise reduction preprocessing, followed by vectorization. This maps the address text to a high-dimensional numerical vector (e.g., 300-dimensional), i.e., an address semantic vector. This vector captures the semantic similarity behind the address names. For example, in semantic space, the vector distance between "Hongxing Farm" and "Hongxing Agricultural Demonstration Park" will be very close, thus overcoming the matching obstacles caused by literal differences.

[0051] In some embodiments, S103 performs vectorization processing on the address text information in the service data tuple to generate an address semantic vector, including: Use regular expressions to clean up redundant prefix characters in address text information; The cleaned address text information is segmented to obtain an address word sequence; Input the address word sequence into a pre-trained word embedding model, and map each address word into a word vector in a high-dimensional vector space; We perform weighted aggregation calculations on all word vectors in the address word sequence to generate an address semantic vector that represents the overall semantic features of the address text information.

[0052] Specifically, data standardization preprocessing is a fundamental step in address matching, aiming to eliminate format differences, redundant information, and multilingual ambiguity in the original address data, providing a unified and standardized input for subsequent semantic modeling and matching. In this embodiment, techniques such as regular expression matching, rule mapping, and multilingual word segmentation optimization are used to convert unstructured address data from service orders and the address database into a structured format.

[0053] Regular expressions are used to clean redundant prefix characters in address text information. To address the issue of redundant administrative division prefixes in original addresses, a regular expression replacement method is employed. Taking Xinjiang as an example, original addresses often contain redundant prefix characters such as "Xinjiang Uygur Autonomous Region" and "Xinjiang". These redundant characters can be automatically removed using regular expression matching rules, retaining the core address elements. The process can be represented as follows: in, The original address string. This is the normalized address after regularization.

[0054] The cleaned address text information is segmented into word sequences to obtain address word sequences. For multilingual addresses (such as bilingual (Chinese and Uyghur) addresses in Xinjiang), a word segmentation optimization mechanism is introduced, which maps transliterated words to standard Chinese words by establishing a standard address vocabulary library. For example, Uyghur transliterated address words are converted into corresponding standard Chinese words to avoid matching errors caused by language differences. Specifically, for each word in the original address... Through mapping function This enables unified processing of multilingual addresses.

[0055] The address word sequence is input into a pre-trained word embedding model, which maps each address word to a word vector in a high-dimensional vector space. Even after standardization and preprocessing, the address data still exists in text form, making it difficult for traditional string matching methods to capture semantic relationships between addresses (such as the similarity between "Beijing Road" and "Beijing East Road"). Therefore, the Word2Vec word embedding model is introduced for semantic modeling, converting the address text into high-dimensional vectors, enabling address matching to be calculated at the semantic level, overcoming the limitations of literal matching.

[0056] Semantic modeling employs the Skip-gram model, whose core principle is to learn distributed representations of words by predicting contextual words. The objective function of the Skip-gram model is: in, The training sample pairs consist of word w and its corresponding context word c. For model parameters, To predict the probability of context word c given word w, the softmax function is used: in, Let w be the input vector for the word w. Let V be the output vector of the context vocabulary c, and V be the vocabulary list.

[0057] To improve model training efficiency, GPU acceleration technology is used to train 300-dimensional word vectors, enabling address fragments to be transformed into vector representations with semantic features. For address fragments composed of multiple words (such as "Urumqi City, Tianshan District"), their semantic vectors are calculated by averaging the vectors of each word. Where n is the number of words contained in the address segment. Let be the vector representation of the i-th word.

[0058] S104. Based on the address semantic vector, match the corresponding standard address in the preset standard address library to determine the geographic coordinates of the service data tuple.

[0059] Specifically, address semantic vectors can only reflect the semantic associations of addresses, while addresses themselves have a clear administrative division hierarchy. To further improve matching accuracy, a hierarchical matching strategy was designed, combining administrative division code filtering with semantic vector similarity calculation to achieve multi-dimensional address matching.

[0060] In some embodiments, matching the corresponding standard address in a preset standard address library based on the address semantic vector includes: Extract administrative division features from address text information, and use these features to perform hierarchical filtering in a standard address database to obtain a set of candidate addresses; Determine the semantic similarity between the address semantic vector and each standard address vector in the candidate address set; The standard address with the highest semantic similarity in the candidate address set is selected as the initial matching result; Determine the Euclidean distance between the geographic coordinates of the initial matching result and the original location information contained in the service data tuple. If the Euclidean distance is less than a preset threshold, the initial matching result is deemed valid.

[0061] Specifically, the system extracts administrative division features from address text information and uses these features to perform hierarchical filtering in a standard address database to obtain a candidate address set. Natural language processing techniques (such as Named Entity Recognition, NER) are used to parse the address text information in the service data tuples to extract administrative division features. These extracted features are then used to perform hierarchical filtering in a pre-defined hierarchical standard address database. For example, if the address text contains "Pujiang County," the system will only search within the standard address range of "Pujiang County" and its subordinate townships, thereby quickly eliminating irrelevant data and obtaining a narrowed candidate address set.

[0062] In one possible implementation, hierarchical matching follows a descending hierarchy logic. Address text information is initially filtered using administrative division features, retaining only candidate addresses whose administrative division features are completely identical across different levels. The filtering rule can be expressed as: address A and address B are considered to be from the same source if and only if their administrative division features are identical across different levels; otherwise, they are filtered. Specifically, the following conditions must be met: in, The result of the filtering operation on address A and address B, This indicates the administrative division characteristics of address x at level y.

[0063] Determine the semantic similarity between the address semantic vector and each standard address vector in the candidate address set, and select the standard address with the highest semantic similarity in the candidate address set as the initial matching result. Call the generated address semantic vector (i.e., the vector representation of the address to be matched) and retrieve the pre-trained vector for each standard address in the candidate address set. Calculate the cosine similarity between the address vector to be matched and each standard address vector in the candidate set, and use this as the semantic similarity, specifically satisfying: in, and These are the semantic vectors for address A and address B, respectively. Higher semantic similarity indicates a stronger semantic connection between the addresses.

[0064] The matching process employs a strategy that combines hierarchical matching with semantic similarity. It not only compares the degree of character overlap but also identifies semantic connections (for example, "village committee" and "brigade headquarters" are semantically close). The results are iterated through, and the standard address with the highest semantic similarity (e.g., a score of 0.98) in the candidate address set is selected as the initial matching result.

[0065] The Euclidean distance between the geographic coordinates of the initial matching result and the original location information contained in the service data tuple is determined. If the Euclidean distance is less than a preset threshold, the initial matching result is considered valid. Addresses that have undergone hierarchical and semantic similarity matching may still have cross-regional mismatches. To prevent errors caused by the same name in different locations or semantic misunderstandings, such as a text description of "Village A" but the actual data collection device is located in "Village B", or matching an address in Beijing to Hebei Province, further optimization is performed through geographic distance verification.

[0066] The service data tuple typically contains raw location information (such as GPS coordinates from the phone during data collection, latitude and longitude from image EXIF ​​information, or location information from the sensor itself). The Euclidean distance between the geographic coordinates of the initial matching result (i.e., the precise latitude and longitude of the standard address in the standard address database) and the raw location information is calculated. A preset threshold is set. If the calculated Euclidean distance is less than the preset threshold, the text description is considered consistent with the physical location, the initial matching result is deemed valid, and it is used as the final geographic coordinates to ensure the spatial accuracy of the matching result. If the calculated Euclidean distance is greater than the preset threshold, there may be a description error or location drift, requiring manual review or proceeding to the next round of suboptimal matching.

[0067] Geographic distance calculation is based on latitude and longitude coordinates on the Earth's surface, using a spherical approximation of the Euclidean distance formula. Specifically, the following conditions are met: Where r is the Earth's radius (approximately 6371 kilometers). , These are the latitude and longitude of address A, respectively. , These are the latitude and longitude of address B, respectively.

[0068] In one possible implementation, based on the map service's address accuracy requirements, the system sets a preset threshold of 500 meters to verify the matching results: when the Euclidean distance d between two addresses is ≤ 500 meters, it is considered a valid match; when d > 500 meters, it is considered an ambiguous match, requiring further manual confirmation or automatic correction. The specific rules for this determination are as follows: Geographic distance verification can effectively filter out mismatched results across regions and mark fuzzy matching results, ensuring that the accuracy of the final output address coordinates meets the actual needs of rural industry services.

[0069] In some embodiments, matching the corresponding standard address in a preset standard address library based on the address semantic vector further includes: Detect whether the address text information contains preset special administrative unit identifier characters; If the address text information is detected to contain a preset special administrative unit identifier character, then the target candidate address containing the same identifier character in the candidate address set is identified. The semantic similarity of the target candidate addresses is weighted using preset weighting coefficients.

[0070] Specifically, when calculating semantic similarity, the address text information is checked to see if it contains preset special administrative unit identifier characters. In some embodiments, the preset special identifier character library is mainly for the agricultural reclamation system or the Xinjiang Production and Construction Corps system, including but not limited to characters such as "division", "regiment", "company", and "farm" (such as farm, forest farm).

[0071] If special identifier characters are detected (e.g., address text "Red Star Second Field Third Company"), the system will specifically identify target candidate addresses containing the same special identifier characters when filtering the candidate address set. Priority will be given to standard addresses that also have a "field" or "company" hierarchical structure, rather than ordinary "township" or "village" structure addresses.

[0072] To improve the accuracy of matching specific areas and prevent confusion with incorrect general administrative addresses (e.g., preventing "Hongxing Erchang" from being incorrectly matched with the similarly named "Hongxing Village"), the semantic similarity of target candidate addresses is weighted using a preset weighting coefficient. Specifically, the system will increase the similarity score of these target candidate addresses. For example: in, and These are the weighting coefficients. This indicates the number of words from set S contained in address A. This weighted mechanism prioritizes matching addresses containing special administrative division units, thus improving the recall rate of addresses with special administrative divisions.

[0073] S105. Construct a knowledge graph based on service data tuples and geographic coordinates, and establish a joint index structure to store the knowledge graph.

[0074] Specifically, a knowledge graph is constructed using graph databases (such as Neo4j and Neo4j Spatial), with service subjects as nodes, service relationships as edges, and latitude and longitude as spatial attributes of the nodes.

[0075] Figure 2 This is a schematic diagram of the structure of the knowledge graph provided in the embodiments of this application, such as... Figure 2 As shown, knowledge graphs integrate and model the three basic elements of time, space, and attributes, and explicitly express spatiotemporal characteristics semantically.

[0076] To meet the needs of spatiotemporal data modeling, storage, and fine-grained time node association for rural industry knowledge graphs, a joint index structure is adopted. The joint index structure includes a spatial index dimension with dynamically adjusted boundaries based on entity distribution density, and a time index dimension with hierarchical classification based on entity attribute update frequency.

[0077] In one possible implementation, the composite index structure is specifically an Adaptive R-Tree with Multi-Granularity Time Partitioning (ART-MGTP) index. This composite index combines the adaptive R-tree used for spatial indexing with dynamic time partitioning, dividing time into variable-granularity slices (e.g., daily, seasonal, annual) based on entity distribution density and entity attribute update frequency to reduce storage and computational overhead. By adaptively adjusting the time partitioning granularity, redundant spatiotemporal entries are avoided, while supporting efficient range queries and relational queries. The ART-MGTP index is further linked to a time tree structure, enabling accurate modeling of temporal relationships (e.g., planting, growth, and harvesting cycles) and establishing associations between time nodes and spatial entities (e.g., villages, industrial bases) at multiple granularities, thereby enhancing the spatiotemporal representation and query capabilities of the graph.

[0078] With this spatiotemporal joint index, when a user initiates a request to "query all service records of a certain county (space) last week (time)," the corresponding spatiotemporal shard can be quickly located, achieving a millisecond-level query response.

[0079] In some embodiments, the spatial index dimension adopts an adaptive R-tree structure, which is constructed through the following steps: Based on the entity distribution density in the current space, the node splitting operation is triggered to generate multiple candidate schemes for splitting the plane; For each candidate solution, the overlap cost function value is calculated. The overlap cost function value is positively correlated with the sum of the minimum outer rectangle areas of the child nodes generated after splitting, and also positively correlated with the area of ​​the overlapping region between the child nodes. The candidate scheme with the minimum overlap cost function value is selected as the optimal splitting strategy, and the minimum outer rectangle range of the split nodes is determined based on the optimal splitting strategy.

[0080] Specifically, given the extremely uneven spatial distribution of rural industrial entities (e.g., highly dense service points in industrial demonstration parks, while sparse service points in field planting areas), traditional R-trees, if fixedly split based solely on node capacity, often result in a large number of dead spaces or overlapping areas in the generated boundary rectangles, severely reducing query efficiency.

[0081] In this embodiment, the spatial index dynamically adjusts its boundaries based on the spatial distribution density of rural industrial entities, employing an adaptive R-tree structure. In densely populated areas (such as industrial parks), the minimum bounding rectangle (MBR) of the index automatically splits to become smaller, providing higher resolution; in sparsely populated mountainous areas, the MBR automatically merges. This mechanism is achieved by minimizing the overlap cost function of nodes, significantly reducing invalid search paths during spatial queries. The construction and updating process of the adaptive R-tree structure is described below.

[0082] Node splitting is triggered based on entity distribution density, and the number of entities within each leaf node of the R-tree is monitored in real time. Unlike traditional methods that force splitting only when the number of entities within a node reaches a preset maximum capacity, this embodiment introduces entity distribution density as a trigger condition. Specifically, the variance of entity distribution density within the current space (MBR, minimum bounding rectangle) is calculated. When it is detected that the entity distribution density of a certain area (such as the town center) is significantly higher than that of the surrounding areas, causing a decrease in the query response time of the current node, a node splitting operation is triggered.

[0083] For a set of spatial entities The coordinates are The MBR of node N is defined as follows: Candidate schemes for splitting the plane are generated. Once a split is triggered, it's necessary to decide how to divide the set of entities within a parent node into two child nodes. The entities are sorted according to their position coordinates along the spatial dimensions of the current node (longitude axis X and latitude axis Y). Different cutting points are tried along the coordinate axes to generate multiple candidate splitting plane schemes. For example, scheme A might divide the region into north and south parts along latitude 30.5 degrees; scheme B might divide the region into east and west parts along longitude 120.2 degrees.

[0084] For each candidate solution, simulate the state after classification and calculate the overlap cost function value. The overlap cost function value is positively correlated with the sum of the minimum outer rectangle areas of the child nodes generated after splitting, and also positively correlated with the area of ​​the overlapping region between child nodes, specifically satisfying: in, For overlapping cost functions, and As a weighting factor, and Quantified separately The overlap ratio and area are considered. A smaller overlap cost function value indicates a smaller minimum outer rectangle area for the child nodes generated after splitting, resulting in tighter enclosure between child nodes and less blank space. Conversely, a larger overlap cost function value increases the probability of searching both child nodes simultaneously during a query, leading to lower query efficiency. The overlap cost function balances query efficiency and storage overhead.

[0085] The optimal splitting strategy and MBR determination are performed by selecting the candidate scheme with the minimum overlap cost function value as the optimal splitting strategy. Based on this strategy, the minimum outer rectangle range of the split nodes is determined. All candidate schemes are traversed, and their overlap cost function values ​​are compared. The scheme with the minimum overlap cost function value is selected as the optimal splitting strategy, minimizing spatial overlap while ensuring the compactness of the split child nodes. Finally, physical separation is performed based on the optimal splitting strategy, redistributing the entities within the original node to the two child nodes and calculating the new boundary to determine the minimum outer rectangle range of the split nodes.

[0086] The spatial index in this embodiment can adaptively fit the actual geographical distribution of rural industrial entities, generating fine grids in densely populated areas and larger grids in sparse areas, thereby significantly improving the efficiency of spatial range queries and k-nearest neighbor queries.

[0087] In some embodiments, the time index dimension employs a multi-granularity time slice structure, constructed through the following steps: Pre-built time index trees containing different time levels; Detect the update frequency of entity attributes in the knowledge graph; if the update frequency is higher than the first threshold, map the entity index to a fine-grained time slice; if the update frequency is lower than the second threshold, map the entity index to a coarse-grained time slice. Monitor the data storage density within each time slice, merge time slices with adjacent time slices whose data storage density is lower than a preset merging threshold, and split time slices whose data storage density is higher than a preset splitting threshold.

[0088] Specifically, to address the extreme imbalance in the temporal distribution of rural industry service data (e.g., data explodes during busy farming seasons and is sparse during off-seasons), and the significant differences in the update frequency of different entity attributes (e.g., the location of agricultural machinery changes in real time, while the base address remains almost unchanged), a multi-granularity time slicing structure is adopted for the time index dimension. The construction process of the multi-granularity time slicing structure is described below.

[0089] A pre-built time index tree with different time levels is provided. This time index tree adopts a hierarchical structure similar to a B+ tree, containing multiple levels such as "year", "month", "day", and "hour". For example, the root node represents the "year", its child nodes represent the "month", and so on. This hierarchical structure provides a basic framework for data storage at different granularities.

[0090] Detect the update frequency of entity attributes in the knowledge graph. When data is written to the knowledge graph, the update frequency of entity attributes in the knowledge graph is automatically detected.

[0091] If the update frequency is higher than the first threshold, the entity index is mapped to a fine-grained time slice; if the update frequency is lower than the second threshold, the entity index is mapped to a coarse-grained time slice. For entities with high-frequency updates (such as combine harvesters or drones in operation, or fluctuating market prices), when an update frequency higher than the first threshold is detected (e.g., update interval less than 1 hour), the entity index is directly mapped to a fine-grained time slice (e.g., hourly or daily nodes). This means users can precisely query the status of an entity within a specific hour. For entities with low-frequency updates (such as the location of an agricultural supply store, the planting area of ​​a demonstration base, or long-term effective service policies), when an update frequency lower than the second threshold is detected (e.g., update interval greater than 1 month), the entity index is mapped to a coarse-grained time slice (e.g., yearly or monthly nodes). This avoids repeatedly creating a large number of daily indexes for long-term unchanging data, saving storage space. This hierarchical strategy based on attribute update frequency avoids the storage redundancy and query overhead caused by creating a uniform fine-grained index for all data.

[0092] Time slice In particle size Defined as: in, and These are the start and end times of the time slice, respectively. This refers to the granularity or length of the time slice.

[0093] Monitor the data storage density within each time slice. Instead of stopping at the initial mapping, dynamically optimize based on the actual data volume, continuously monitoring the data storage density within each time slice (i.e., the number of graph nodes and relationships contained within that time period).

[0094] Time slices with data storage density below a preset merging threshold are merged with adjacent time slices. For example, during the off-season for agriculture (such as winter), there is less service order and agricultural machinery scheduling data. If the data storage density within a continuous time period (such as November 1st to November 15th) is detected to be below the preset merging threshold, the merging mechanism is automatically triggered. Low-density time slices are merged with adjacent time slices; for example, 15 "day" level slices are merged into one "half-month" level slice. In this way, when querying data for that time period, only one index node needs to be accessed, significantly reducing I / O overhead.

[0095] Time slices with data storage density exceeding a preset splitting threshold are split. Specifically, during busy farming seasons (such as summer and autumn harvests), massive amounts of cross-regional operation scheduling data are generated in a short period. If the data storage density of a certain time node (such as June 15th) exceeds the preset splitting threshold, it indicates that the node is overloaded, leading to slower query responses. In this case, the time slice is split, for example, by refining the original "day" level slice into 24 "hour" level slices, or by splitting it according to morning / afternoon, distributing the data for storage, thereby ensuring query performance under high concurrency.

[0096] In this embodiment, the time index is automatically scaled according to the peak and off-peak season characteristics of rural industrial activities, and sparse time slices are merged and dense time slices are split, achieving the best balance between storage cost and query efficiency.

[0097] In some embodiments, the composite index structure is represented in a graph database as a composite index key, which is composed of spatial index node encoding and time slice identifier. Specifically, at the physical storage level of graph databases (such as database systems based on Native Graph Store), in order to achieve efficient retrieval of spatiotemporal data, simply relying on the single attribute index provided by the graph database is often inefficient (for example, first querying all nodes with "time = 2023" and then traversing and filtering nodes with "location in County A" will cause huge I / O overhead).

[0098] Therefore, the composite index structure provided in this application embodiment is specifically implemented as a composite index key in a graph database, and the specific construction method is as follows: Spatial index node encoding: Based on the constructed adaptive R tree structure, a unique spatial index node encoding is assigned to each leaf node in the tree (i.e., the minimum bounding rectangle MBR containing the actual data object). This node encoding can be a space fill curve (such as the Hilbert Curve) value or a unique identifier that has been hashed.

[0099] Time slice identifier: Based on the constructed multi-granularity time slice structure, a standardized time slice identifier is generated for each time slice. This identifier contains granularity information and time value.

[0100] Generation and storage of composite index keys: Spatial index node codes and time slice identifiers are combined using specific connectors or bitwise operations to generate composite index keys.

[0101] In one possible implementation, the ART-MGTP index uses a composite index key. Connecting the spatial and temporal dimensions, where N is the R-Tree node encoding. This identifies the time slice. Each entry in the composite index key maps to a subgraph in Neo4j, storing spatiotemporal relationships, for example: Composite index keys support mixed queries through cross-space (R-Tree) and time (partition slicing) constraints, such as "find all tea gardens within 50 kilometers of County X during the 2022 harvest season".

[0102] The ART-MGTP index replaces Neo4jSpatial's static R-tree by injecting adaptive splitting logic and coupling it with the hierarchical nodes of the temporal tree. This is beneficial for spanning both spatial regions and temporal ranges in R-trees. The query uses adaptive R-tree parsing to retrieve candidate nodes {N}; partitioning. Convert to slices via dynamic granularity selection Use composite index keys Connect the results to obtain the relevant subgraphs.

[0103] Time relationship (e.g., The edges are modeled as edges between nodes in a time tree, with edge weights derived from historical patterns. The ART-MGTP index also supports probabilistic queries, such as "predict the harvest date of farm Y based on the sowing time from 2021 to 2023".

[0104] In one possible implementation, GraphAware TimeTree 3.0 is used for time hierarchy management, and Neo4j 5.0's native vector indexes are used to accelerate spatial queries. The adaptive R tree uses the Java Topology Suite (JTS) for geometry, while time partitioning is optimized for real-time event density analysis using Apache Kafka Streams.

[0105] By combining adaptive spatial indexing with dynamic temporal partitioning, accurate spatiotemporal modeling of rural industries is achieved, surpassing static graph methods in terms of scalability and query flexibility.

[0106] In some embodiments, the method further includes: Receive service requests, which include the request type, time window, and geographic location range; Generate spatiotemporal query instructions based on service requests; Utilize a composite index structure to quickly retrieve service entity nodes within a geographical location range and time window in a knowledge graph; Calculate the matching degree between the retrieved service entity nodes and the service requests, and generate a resource scheduling scheme based on the matching degree.

[0107] Specifically, service resource scheduling is performed based on the constructed graph, and the specific process is as follows: Receiving service requests: Service requests are received through user interfaces (such as farmers' apps or government-mandated dashboards). For example, a farmer may post a request: "Request to use a wheat harvester to harvest wheat in a wheat field located in XX Township, XX County (geographical location range, automatically resolved by the system to the center coordinates and a radius of 5 kilometers) between October 1st and October 3rd, 2023 (time window) (request type)."

[0108] Generate spatiotemporal query instructions: Parse the user's service request and transform it into a spatiotemporal query instruction that can be executed by the database. First, convert the natural language time description into the corresponding time slice identifier range (e.g., [T_D_20231001, T_D_20231003]), and convert the geographic location into a spatial query rectangle (Query MBR).

[0109] This approach leverages a composite index structure to quickly retrieve service entity nodes within the requested geographic location and time window in the knowledge graph. An adaptive R-tree structure is used to quickly locate spatial nodes intersecting the query rectangle, eliminating the vast majority of areas outside the service range. Under the located spatial nodes, multi-granularity time slices are used to directly locate the corresponding date composite index key.

[0110] Service entity nodes within the corresponding time and space range (such as harvesters in an idle state or local agricultural machinery cooperatives) can be directly read using composite index keys. Retrieval can be completed by scanning only a very small number of index leaf nodes, eliminating the need to traverse the entire table and achieving millisecond-level response times.

[0111] After retrieving service entity nodes (multiple nodes may be retrieved), the matching degree between each service entity node and the service request is calculated. The matching degree calculation can comprehensively consider multiple factors, such as the distance between the entity's current location and the demand location (the closer the distance, the higher the score); the service entity's credit score; and price matching degree. Finally, a resource scheduling plan is generated based on the service entity node with the highest calculated matching degree, and this plan is pushed to the user, achieving precise and efficient allocation of rural industry resources.

[0112] In one possible implementation, a prototype system for rural industry spatiotemporal map services was designed based on data aggregation and graph construction. Figure 3 This is a schematic diagram of the architecture of the prototype system for rural industry spatiotemporal map services provided in the embodiments of this application, such as... Figure 3 As shown, the system is developed using open-source tools such as GeoServer. Targeting the service needs of rural industries, it manages resources such as orders, service organizations, service demands, and demonstration bases through spatiotemporal mapping. The system integrates the Neo4j graph database through the Timetree / Sptiail middleware and uses Java to implement interaction with the Neo4j graph database and GeoServer, providing spatiotemporal map management services and technical support for rural industry services.

[0113] The following describes the rural industry service data aggregation and mapping system provided in this application. The rural industry service data aggregation and mapping system described below can be referred to in correspondence with the rural industry service data aggregation and mapping method described above.

[0114] Figure 4 This is a schematic diagram of the structure of the rural industry service data aggregation and map construction system provided in the embodiments of this application, such as... Figure 4 As shown, the system includes: The acquisition module 601 is used to acquire Internet text data and IoT sensor data, and to extract service entity information from Internet text data using a semantic analysis model. The service entity information includes address text information. The first generation module 602 is used to integrate service entity information with IoT sensor data input feature fusion model to generate service data tuples. The second generation module 603 is used to perform vectorization processing on the address text information in the service data tuple to generate an address semantic vector. The determination module 604 is used to match the corresponding standard address in a preset standard address library based on the address semantic vector to determine the geographic coordinates of the service data tuple; Module 605 is used to build a knowledge graph based on service data tuples and geographic coordinates, and to establish a joint index structure to store the knowledge graph; The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

[0115] In some embodiments, the service entity information in Internet text data is extracted using a semantic analysis model, including: Input Internet text data into a semantic analysis model, and identify and extract key fields of service organization, service price and job type according to preset instructions; Monitor the confidence level of the extraction results output by the semantic analysis model, and adjust the extraction strategy for Internet text data based on the confidence level feedback.

[0116] In some embodiments, the feature fusion model employs a meta-learner containing a two-stream network structure, and the meta-learner's processing includes: Extract textual features from service entity information and numerical features from IoT sensor data, respectively; The association weights between text features and numerical features are determined based on the attention mechanism, and the association weights are gating and adjusted based on the spatial sparsity of the data distribution to output service data tuples.

[0117] In some embodiments, matching the corresponding standard address in a preset standard address library based on the address semantic vector includes: Extract administrative division features from address text information, and use these features to perform hierarchical filtering in a standard address database to obtain a set of candidate addresses; Determine the semantic similarity between the address semantic vector and each standard address vector in the candidate address set; The standard address with the highest semantic similarity in the candidate address set is selected as the initial matching result; Determine the Euclidean distance between the geographic coordinates of the initial matching result and the original location information contained in the service data tuple. If the Euclidean distance is less than a preset threshold, the initial matching result is deemed valid.

[0118] In some embodiments, matching the corresponding standard address in a preset standard address library based on the address semantic vector further includes: Detect whether the address text information contains preset special administrative unit identifier characters; If the address text information is detected to contain a preset special administrative unit identifier character, then the target candidate address containing the same identifier character in the candidate address set is identified. The semantic similarity of the target candidate addresses is weighted using preset weighting coefficients.

[0119] In some embodiments, the spatial index dimension adopts an adaptive R-tree structure, which is constructed through the following steps: Based on the entity distribution density in the current space, the node splitting operation is triggered to generate multiple candidate schemes for splitting the plane; For each candidate solution, the overlap cost function value is calculated. The overlap cost function value is positively correlated with the sum of the minimum outer rectangle areas of the child nodes generated after splitting, and also positively correlated with the area of ​​the overlapping region between the child nodes. The candidate scheme with the minimum overlap cost function value is selected as the optimal splitting strategy, and the minimum outer rectangle range of the split nodes is determined based on the optimal splitting strategy.

[0120] In some embodiments, the time index dimension employs a multi-granularity time slice structure, constructed through the following steps: Pre-built time index trees containing different time levels; Detect the update frequency of entity attributes in the knowledge graph; if the update frequency is higher than the first threshold, map the entity index to a fine-grained time slice; if the update frequency is lower than the second threshold, map the entity index to a coarse-grained time slice. Monitor the data storage density within each time slice, merge time slices with adjacent time slices whose data storage density is lower than a preset merging threshold, and split time slices whose data storage density is higher than a preset splitting threshold.

[0121] In some embodiments, the composite index structure is represented in the graph database as a composite index key, which is composed of spatial index node encoding and time slice identifier.

[0122] In some embodiments, the system further includes: The receiving module is used to receive service requests, which include the request type, time window, and geographical location range. The third generation module is used to generate spatiotemporal query instructions based on service requests; The retrieval module is used to quickly retrieve service entity nodes within a geographical location and time window in the knowledge graph using a composite index structure; The fourth generation module is used to calculate the matching degree between the retrieved service entity nodes and the service requests, and generate a resource scheduling scheme based on the matching degree.

[0123] It should be noted that the rural industry service data aggregation and map construction system provided in this application embodiment can realize all the method steps implemented in the above method embodiment and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0124] Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 5As shown, the electronic device may include: a processor 701, a communications interface 702, a memory 703, and a communication bus 704. The processor 701, communications interface 702, and memory 703 communicate with each other via the communication bus 704. The processor 701 can call logical instructions from the memory 703 to execute a method for rural industry service data aggregation and mapping, which includes: Acquire internet text data and IoT sensor data, and use a semantic analysis model to extract service entity information from the internet text data, including address text information. The service entity information and IoT sensor data are input into a feature fusion model to generate service data tuples. The address text information in the service data tuple is vectorized to generate address semantic vectors; Based on the address semantic vector, the corresponding standard address is matched in the pre-set standard address library to determine the geographic coordinates of the service data tuple; A knowledge graph is constructed based on service data tuples and geographic coordinates, and a joint index structure is established to store the knowledge graph; The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

[0125] Furthermore, the logical instructions in the aforementioned memory 703 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0126] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the rural industry service data aggregation and map construction method provided by the above methods, the method including: Acquire internet text data and IoT sensor data, and use a semantic analysis model to extract service entity information from the internet text data, including address text information. The service entity information and IoT sensor data are input into a feature fusion model to generate service data tuples. The address text information in the service data tuple is vectorized to generate address semantic vectors; Based on the address semantic vector, the corresponding standard address is matched in the pre-set standard address library to determine the geographic coordinates of the service data tuple; A knowledge graph is constructed based on service data tuples and geographic coordinates, and a joint index structure is established to store the knowledge graph; The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

[0127] Furthermore, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the rural industry service data aggregation and map construction method provided by the methods described above, the method comprising: Acquire internet text data and IoT sensor data, and use a semantic analysis model to extract service entity information from the internet text data, including address text information. The service entity information and IoT sensor data are input into a feature fusion model to generate service data tuples. The address text information in the service data tuple is vectorized to generate address semantic vectors; Based on the address semantic vector, the corresponding standard address is matched in the pre-set standard address library to determine the geographic coordinates of the service data tuple; A knowledge graph is constructed based on service data tuples and geographic coordinates, and a joint index structure is established to store the knowledge graph; The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for rural industry service data aggregation and graph construction, characterized in that, include: Acquire Internet text data and IoT sensor data, and use a semantic analysis model to extract service entity information from the Internet text data, including address text information; The service entity information and the IoT sensor data are input into a feature fusion model to generate service data tuples. The address text information in the service data tuple is vectorized to generate an address semantic vector; Based on the address semantic vector, the corresponding standard address is matched in a preset standard address library to determine the geographic coordinates of the service data tuple; A knowledge graph is constructed based on the service data tuples and the geographic coordinates, and a joint index structure is established to store the knowledge graph; The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

2. The method for rural industry service data aggregation and map construction according to claim 1, characterized in that, The step of extracting service entity information from the Internet text data using a semantic analysis model includes: The Internet text data is input into a semantic analysis model, and key fields such as service organization, service price, and job type are identified and extracted according to preset instructions. Monitor the confidence level of the extraction results output by the semantic analysis model, and adjust the extraction strategy for Internet text data based on the confidence level feedback.

3. The method for rural industry service data aggregation and map construction according to claim 1, characterized in that, The feature fusion model employs a meta-learner containing a two-stream network structure, and the processing procedure of the meta-learner includes: Extract the text features of the service entity information and the numerical features of the IoT sensor data respectively; The association weights between the text features and the numerical features are determined based on the attention mechanism, and the association weights are gating and adjusted based on the spatial sparsity of the data distribution to output the service data tuple.

4. The method for rural industry service data aggregation and map construction according to claim 1, characterized in that, The step of matching the corresponding standard address in a preset standard address library based on the address semantic vector includes: The administrative division features are extracted from the address text information, and the administrative division features are used to perform hierarchical filtering in the standard address database to obtain a set of candidate addresses; Determine the semantic similarity between the address semantic vector and each standard address vector in the candidate address set; The standard address with the highest semantic similarity in the candidate address set is selected as the initial matching result; The Euclidean distance between the geographic coordinates of the initial matching result and the original location information contained in the service data tuple is determined. If the Euclidean distance is less than a preset threshold, the initial matching result is determined to be valid.

5. The method for rural industry service data aggregation and map construction according to claim 4, characterized in that, The step of matching the corresponding standard address in a preset standard address library based on the address semantic vector also includes: Detect whether the address text information contains preset special administrative unit identification characters; If it is detected whether the address text information contains a preset special administrative unit identifier character, then the target candidate address containing the same identifier character in the candidate address set is identified; The semantic similarity of the target candidate addresses is weighted using preset weighting coefficients.

6. The method for rural industry service data aggregation and map construction according to claim 1, characterized in that, The spatial index dimension adopts an adaptive R-tree structure and is constructed through the following steps: Based on the entity distribution density in the current space, the node splitting operation is triggered to generate multiple candidate schemes for splitting the plane; For each candidate scheme, the overlap cost function value is calculated. The overlap cost function value is positively correlated with the sum of the minimum outer rectangle areas of the child nodes generated after splitting, and is also positively correlated with the area of ​​the overlapping region between the child nodes. The candidate scheme with the minimum overlap cost function value is selected as the optimal splitting strategy, and the minimum outer rectangle range of the split nodes is determined based on the optimal splitting strategy.

7. The method for rural industry service data aggregation and map construction according to claim 1, characterized in that, The time index dimension adopts a multi-granularity time slice structure, which is constructed through the following steps: Pre-built time index trees containing different time levels; The update frequency of entity attributes in the knowledge graph is detected; if the update frequency is higher than a first threshold, the entity index is mapped to a fine-grained time slice; if the update frequency is lower than a second threshold, the entity index is mapped to a coarse-grained time slice. Monitor the data storage density within each time slice, merge time slices with adjacent time slices whose data storage density is lower than a preset merging threshold, and split time slices whose data storage density is higher than a preset splitting threshold.

8. The method for rural industry service data aggregation and map construction according to claim 1, characterized in that, The composite index structure is represented in the graph database as a composite index key, which is composed of spatial index node encoding and time slice identifier.

9. The method for rural industry service data aggregation and map construction according to claim 1, characterized in that, The method further includes: Receive a service request, the service request including the request type, time window, and geographical location range; Generate a spatiotemporal query instruction based on the service request; The aforementioned composite index structure is used to quickly retrieve service entity nodes within the aforementioned geographical location range and time window in the knowledge graph; Calculate the matching degree between the retrieved service entity nodes and the service requests, and generate a resource scheduling scheme based on the matching degree.

10. A system for data aggregation and mapping of rural industry services, characterized in that, include: The acquisition module is used to acquire Internet text data and IoT sensor data, and use a semantic analysis model to extract service entity information from the Internet text data, including address text information. The first generation module is used to combine the service entity information with the IoT sensor data input feature fusion model to generate service data tuples. The second generation module is used to vectorize the address text information in the service data tuple to generate an address semantic vector. The determination module is used to match the corresponding standard address in a preset standard address library based on the address semantic vector, and determine the geographic coordinates of the service data tuple; A construction module is used to construct a knowledge graph based on the service data tuples and the geographic coordinates, and to establish a joint index structure to store the knowledge graph; The composite index structure includes a spatial index dimension that dynamically adjusts the boundary based on entity distribution density, and a time index dimension that is graded based on entity attribute update frequency.

Images