A method for constructing an intelligent one-stop service platform based on data fusion

Abstract

The application belongs to the field of intelligent service platform construction, and discloses a kind of intelligent one-stop service platform construction method based on data fusion.The method constructs four-layer architecture of data layer, fusion layer, service layer and application layer;Through unified access gateway, multi-source data is collected and stored hierarchically;Using configurable workflow engine, first-level cleaning standardization, second-level feature semantic alignment and third-level deep fusion processing are executed in turn, wherein multi-task learning model is used for joint training of aligned features in third-level fusion, and fusion knowledge graph is generated to mine hidden associations;Based on the fusion knowledge, business and intelligent services are packaged, and a service resource pool is constructed;Finally, personalized one-stop services are provided through a unified portal.The application converts multi-source data into structured knowledge through progressive deep fusion, solving the problem of insufficient shallow fusion and association mining of traditional platforms.

Description

Technical Field

[0001] This invention relates to the field of intelligent service platform construction, and in particular to a method for constructing an intelligent one-stop service platform based on data fusion. Background Technology

[0002] With the deepening of digital transformation, various information systems, IoT devices, Internet applications and business platforms generate massive amounts of multi-source heterogeneous data. This type of data is scattered across data sources in different systems, departments and technical architectures, exhibiting typical characteristics such as heterogeneous data types, inconsistent formats, non-interoperable semantics and fragmented business lines.

[0003] Traditional service platforms often limit data integration to simple ETL (Extract, Transform, Load) processes or rely on fixed rules for shallow correlations. Such one-off, static processes cannot automatically identify and construct deep, non-explicit hidden relationships from multi-source heterogeneous data, resulting in insufficient mining of the data's potential value and difficulty in supporting upper-level intelligent services and decision analysis. Summary of the Invention

[0004] The present invention aims to provide a method for constructing an intelligent one-stop service platform based on data fusion, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing an intelligent one-stop service platform based on data fusion includes the following steps: S1. Platform architecture construction: Construct a four-layer architecture including a data layer, an integration layer, a service layer, and an application layer. The data layer is built based on a distributed data lake, the integration layer has a built-in configurable multi-level integration workflow engine, the service layer adopts a microservice architecture, and the application layer deploys an adaptive rendering engine. S2. Based on the data layer, raw data is collected from multiple heterogeneous data sources through a unified access gateway, and the collected multi-source data is classified and stored in a hierarchical manner according to data type and business domain, and a metadata management index is established. S3. Based on the fusion layer, the multi-source data stored in S2 is configured and sequentially triggered to perform first-level fusion, second-level fusion, and third-level fusion processing through the multi-level fusion workflow engine; The first-level fusion is used to clean and standardize multi-source data, eliminating data redundancy and format differences; The secondary fusion is used to align the features and semantics of the data after the primary fusion, so as to achieve feature unification and spatiotemporal semantic matching of multi-source data. The three-level fusion is used to perform deep fusion and knowledge generation based on the data aligned by the two-level fusion, to uncover hidden relationships between data and form structured knowledge; S4. Based on the service layer, and based on the data and knowledge generated after the three-level fusion in S3, multiple business service modules and intelligent analysis services are encapsulated, and a service resource pool is constructed. The service resource pool supports service registration, discovery and dynamic invocation. S5. Based on the application layer, build a unified service portal, integrate and call the services in the service resource pool obtained in S4, and provide personalized intelligent one-stop service entry for different users according to user identity and business needs.

[0006] Preferably, the first-level fusion includes: The collected multi-source data is processed by a configurable rule engine and anomaly detection intelligent algorithm to remove duplicates, correct errors, and fill in missing values, resulting in clean data. The cleaned data in different formats will be converted into standardized data in a unified format.

[0007] Preferably, the secondary fusion includes: The standardized data is subjected to feature extraction and semantic representation; different data types are processed separately: natural language processing techniques are used to perform semantic analysis and word vector extraction on text data, and feature extraction algorithms are used to extract feature vectors for image, audio and video data; then, the different data types from multiple sources are converted into a unified feature vector representation. Based on spatiotemporal association rules and semantic mapping relationships, the feature vectors of the multi-source data are spatiotemporally and semantically aligned to generate aligned multi-source feature data.

[0008] Preferably, the three-level fusion includes: Construct a multi-task learning model and input the aligned multi-source feature data into the multi-task learning model for joint training; Through joint training, a fused knowledge graph is generated, which is used to represent and mine hidden relationships between entities in multi-source feature data.

[0009] Preferably, the multi-task learning model includes entity recognition, relation extraction, and attribute completion; the fused knowledge graph is constructed from entity, relation, and attribute triples obtained through joint training.

[0010] Preferably, in step S2, the method of collecting raw data from multiple heterogeneous data sources includes at least two of the following: application programming interface calls, file upload and import, real-time data stream collection, and direct access by IoT devices.

[0011] Preferably, in step S3, a fusion quality assessment module is also provided to quantitatively assess the data quality of the fusion output at each level. The assessment indicators include completeness, consistency, accuracy, and timeliness. Based on the assessment results, if any assessment indicator fails to reach a preset threshold, the corresponding level of fusion process is re-executed or the fusion parameters are dynamically adjusted to achieve closed-loop optimization of the fusion process.

[0012] Preferably, in step S4, the business service module includes business processing, comprehensive query, analysis and early warning, and collaborative management modules; the intelligent analysis service integrates machine learning prediction models and anomaly detection algorithms to provide personalized recommendations, risk warnings, and decision support capabilities.

[0013] Preferably, in step S5, the unified service portal supports multi-channel access, including web browsers, mobile applications, smart terminals, and third-party system integration interfaces, and achieves cross-channel single sign-on and unified permission management through a unified identity authentication and authorization center.

[0014] Preferably, it also includes an operation and maintenance monitoring layer, which is used to monitor and visualize the operating status, performance indicators and security events of the data layer, fusion layer, service layer and application layer in real time, and predict and automatically alarm abnormal situations based on machine learning algorithms to ensure the high availability and security of the platform.

[0015] The beneficial effects of this technical solution compared to existing technologies are as follows: (1) This solution constructs a three-level progressive deep data fusion system, abandoning the traditional model that only stays at the shallow ETL processing and static association of fixed rules. It relies on a multi-level fusion workflow engine to realize a dynamic and configurable fusion process. In the three-level fusion, through the construction of a multi-task learning model and knowledge graph, it automatically identifies and mines the deep, non-explicit hidden relationships between multi-source heterogeneous data, transforms the original data into structured knowledge, fully releases the potential value of the data, and provides high-quality and high-value data and knowledge support for upper-level intelligent services and decision analysis.

[0016] (2) By adopting a multi-source heterogeneous data access and storage architecture with a distributed data lake and a unified access gateway, it supports multiple collection methods such as API calls, real-time stream acquisition, and direct connection of IoT devices, realizes unified access and hierarchical classification storage of multi-modal data such as text, images, audio and video, and establishes a global metadata management index, which solves the problems of scattered data access channels, chaotic storage structure, and difficulty in unified management of multiple types of data in traditional platforms, and improves the efficiency of data aggregation and management.

[0017] (3) By setting up a fusion quality assessment module in the fusion layer, the fusion results at each level are quantitatively assessed based on indicators such as completeness, consistency, accuracy, and timeliness. When the indicators do not reach the preset threshold, the corresponding level of fusion is automatically re-executed or the parameters are dynamically adjusted, forming a closed-loop mechanism of "fusion-assessment-optimization". This overcomes the defects of traditional data fusion process, such as no quality verification, inability to iterate and optimize, and poor data output stability, and ensures the reliability of data and knowledge after fusion.

[0018] (4) By adopting a service encapsulation and scheduling mode of microservice architecture + service mesh governance in the service layer, business capabilities and intelligent analysis capabilities are decoupled and encapsulated into discrete microservices, and a service resource pool that supports registration, discovery, routing and dynamic invocation is built. This changes the problems of poor service reusability, difficulty in expansion and low efficiency of cross-system collaboration caused by the monolithic architecture or tight coupling design of traditional platforms, and realizes flexible service orchestration and on-demand invocation.

[0019] (5) By deploying an adaptive rendering engine and a unified service portal at the application layer, combined with multi-channel access adaptation and a unified identity authentication and authorization center, personalized service interfaces are dynamically assembled and pushed according to user identity, context and business needs, realizing cross-channel single sign-on and unified permission management. This solves the problems of scattered entry points, fragmented experience and inability to provide personalized one-stop services in traditional service platforms, and improves user convenience and platform service adaptation capabilities.

[0020] (6) By setting up an operation and maintenance monitoring layer, the operation status, performance indicators and security events of the entire chain of the data layer, fusion layer, service layer and application layer are monitored and visualized in real time. Based on machine learning algorithms, anomaly prediction and automatic alarm are performed, which makes up for the shortcomings of traditional platforms that lack unified operation and maintenance monitoring, lack intelligent early warning, and have difficulty in guaranteeing high availability and security, and ensures that the platform operates stably and reliably under complex business scenarios and large-scale concurrency. Attached Figure Description

[0021] Figure 1 The system framework diagram provided for this invention; Detailed Implementation The present invention will now be described in further detail with reference to the accompanying drawings and embodiments: like Figure 1 The method for constructing an intelligent one-stop service platform based on data fusion, as shown, includes the following steps: S1. Platform architecture construction: A four-layer architecture is built, including a data layer, an integration layer, a service layer, and an application layer. The data layer adopts a distributed data lake based on Hadoop HDFS. This layer does not have a preset data model and aims to store raw data from various departments of the city in its original form and in its entirety.

[0022] The fusion layer has a built-in configurable multi-level fusion workflow engine based on the open-source workflow engine Apache Airflow. This engine allows operations and maintenance personnel to flexibly arrange task dependencies and execution parameters for "Level 1 Fusion → Level 2 Fusion → Level 3 Fusion → Quality Assessment" through a graphical interface or configuration file.

[0023] The service layer is built using the Spring Cloud microservice framework. Each business function or algorithm model is encapsulated as an independent microservice that can be deployed and scaled independently.

[0024] The application layer develops a unified portal built using the React front-end framework. It also integrates an adaptive rendering engine that automatically selects and renders the most suitable page template and interactive components based on the user-agent of the access terminal (such as a PC browser, mobile browser, or WeChat mini-program).

[0025] It also includes an operations and maintenance monitoring layer, which provides real-time monitoring and visualization of the operational status, performance metrics, and security events of the data layer, integration layer, service layer, and application layer. It integrates a machine learning-based anomaly detection plugin (Prophet model) to enable predictive alerts, ensuring the platform's high availability and security. Prometheus is deployed for metric collection, Grafana for visualization, and the ELK Stack (Elasticsearch, Logstash, Kibana) for centralized log management and analysis.

[0026] S2. Based on the data layer, raw data is collected from multiple heterogeneous data sources through a unified access gateway (Nginx extension), and the collected multi-source data is hierarchically classified and stored according to data type and business domain, and a metadata management index is established. For the government information disclosure API, the gateway periodically calls the interface and retrieves data in JSON / XML format.

[0027] A secure upload port is provided for Excel, PDF reports, and other files uploaded by various departments.

[0028] Traffic camera video streams and environmental monitoring sensor (IoT device) data are collected in real time using the Apache Kafka or MQTT protocol.

[0029] After the collected data enters the data lake, it is hierarchically categorized and stored according to data type (text / image / video / time series data) and business domain (transportation / environmental protection / healthcare / education). Simultaneously, a deployed metadata crawling service automatically extracts information such as the format, source, size, update time, and business tags of the data files, and builds a global metadata management index based on Apache Atlas to support subsequent data discovery and governance.

[0030] There are at least two ways to collect raw data from multiple heterogeneous data sources: application programming interface calls, file upload and import, real-time data stream acquisition, and direct access from IoT devices.

[0031] S3. Based on the fusion layer, the multi-source data stored in S2 is configured and sequentially triggered to perform first-level fusion, second-level fusion, and third-level fusion processing through a multi-level fusion workflow engine. The multi-level fusion workflow engine (Apache Airflow) allows operations personnel to define the execution order, dependencies, and runtime parameters of each fusion task through a graphical interface or YAML configuration file, forming a repeatable and monitorable automated data processing pipeline. The engine triggers the following fusion tasks sequentially according to the configuration: First-level fusion is used to clean and standardize multi-source data, eliminating data redundancy and format differences; first-level fusion includes: The configurable rule engine (Drools rule engine) executes business logic-based cleansing rules, such as verifying ID card number formats and unified administrative division codes. For fields that fail verification, error correction operations are performed: for example, when the ID card number verification code is incorrect but the birth date and sequence code are valid, the Luhn algorithm is used to recalculate and correct the verification code; when the administrative division code is an old code, it is automatically updated to the new national standard code according to a preset code mapping table. Simultaneously, an anomaly detection intelligent algorithm (Isolation Forest) is run to identify and handle outliers in numerical fields. During this process, duplicate records are deduplicated, and missing values ​​are filled with the mean, median, or marked as "unknown" depending on the business scenario.

[0032] All cleaned data in different formats (CSV, JSON, XML) is converted into a unified standardized data format (Apache Parquet) within the platform, and the naming conventions, data types, and encoding of key fields (such as timestamps, geolocation identifiers, and subject IDs) are enforced to form standardized data tables.

[0033] Second-level fusion is used to align the features and semantics of the data processed by first-level fusion, achieving feature unification and spatiotemporal semantic matching of multi-source data; second-level fusion includes: The process involves feature extraction and semantic representation of standardized data. Different data types are processed separately: for text data, natural language processing techniques (including word segmentation, part-of-speech tagging, and named entity recognition) are used for semantic analysis, and a word embedding model (Word2Vec) is used to convert the text into dense word vector representations. For image, audio, and video data, feature extraction algorithms (using ResNet convolutional neural networks to extract image features and VGGish networks to extract audio features) are used to extract feature vectors. Finally, data from multiple sources and different data types are converted into a unified feature vector representation. Based on spatiotemporal association rules (e.g., associating events that occur within the same geofence and have overlapping time windows) and semantic mapping tables (e.g., establishing an equivalent mapping between the "company name" field in the business registration system and the "taxpayer name" field in the tax system), the similarity between feature vectors from different data sources is calculated or rule matching is performed. Through this process, spatiotemporal and semantic alignment of feature vectors from multi-source data is achieved, ultimately generating an aligned wide table of multi-source feature data, where each row represents an entity or event associated with multi-dimensional features.

[0034] Level 3 fusion is used for deep fusion and knowledge generation based on data aligned by Level 2 fusion, uncovering hidden relationships between data and forming structured knowledge; Level 3 fusion includes: A multi-task learning model is constructed, which uses a shared underlying neural network (multilayer perceptron) for general feature learning, with multiple task-specific output layers connected in parallel at the upper layers. The tasks include at least: entity recognition (identifying core entities in a feature wide table), relation extraction (determining the type of relationship between entities), and attribute completion (predicting or completing attribute values ​​for entities). To adapt to these tasks, the model's output layers are designed as follows: entity recognition uses sequence labeling or a classification head, relation extraction uses an entity-pair-based classification head, and attribute completion uses regression or a classification head.

[0035] The aligned multi-source feature data wide table of the secondary fusion output is used as the training dataset and input into the above multi-task learning model for joint training.

[0036] Training data preparation: To perform supervised joint training, a labeled training dataset needs to be constructed. Labels are obtained specifically through the following methods: Manual annotation and knowledge base mapping: For some core entities and relationships, domain experts manually annotate the sample data to form a high-quality seed tag set. Simultaneously, records in existing structured business databases (such as enterprise business information databases and geographic information databases) are automatically converted into (entity, attribute, value) triple tags through the semantic mapping relationships in the secondary fusion process.

[0037] Rule-based automatic labeling: This feature utilizes predefined domain rules and patterns to automatically label textual data. For example, it uses regular expressions to match patterns such as "located in" and "established in" in text, automatically extracting relationship tags <entity, located in, location>.

[0038] Remote supervision: Utilize existing large-scale knowledge bases (such as general knowledge graphs) to link entities within them to the feature data of this platform, thereby automatically generating relationship tags.

[0039] The labels obtained through one or more of the above methods are associated and aligned with the corresponding data rows in the "aligned multi-source feature data wide table" to jointly form a labeled training dataset for training multi-task learning models.

[0040] The joint training process is as follows: Loss function construction: Define the loss function (cross-entropy loss) for each of the three subtasks: entity recognition, relation extraction, and attribute completion.

[0041] Joint optimization: The loss functions of the above subtasks are combined into a global multi-task loss function using a weighted summation method. During training, this global loss function is simultaneously optimized through backpropagation. This joint optimization mechanism allows the model to fully consider the inherent coupling and consistency constraints between entities, relationships, and attributes in multi-source feature data when learning general feature representations across tasks using shared underlying parameters. This forces the model to uncover deeper hidden association patterns that are beneficial to multiple tasks.

[0042] Training objective: The goal of training is to enable the model to accurately and collaboratively extract the entity, relation, and attribute information required to construct the knowledge graph from the aligned multi-source feature data wide table.

[0043] After training, the full set of feature data is input into the model for inference. The model will collaboratively and simultaneously output predicted entity, relation, and attribute triples (e.g., <Company A, located in, Park B>, <Company A, risk level, high>). These triples can be imported into a graph database (Neo4j) to automatically construct a fused knowledge graph. This graph, with its network structure, intuitively represents and mines the deep, hidden relationships between various entities in multi-source feature data, providing structured knowledge support for intelligent services.

[0044] The fusion layer also includes a fusion quality assessment module, which is used to quantitatively assess the data quality of the fusion output at each level. The assessment indicators include completeness, consistency, accuracy, and timeliness. Based on the assessment results, if any assessment indicator fails to reach a preset threshold, the corresponding level of fusion process will be re-executed or the fusion parameters will be dynamically adjusted to achieve closed-loop optimization of the fusion process.

[0045] The specific implementation process of quantitative assessment is as follows: Integrity check: The system automatically calculates the percentage of null values ​​in key fields. For example, it calculates the percentage of non-null records in the "Company Name" field; if it is below 95%, it is considered unacceptable.

[0046] Consistency verification: The system executes pre-defined logical rules for checks. For example, to verify that the "Company Deregistration Date" is not empty, the "Operating Status" must be "Deregistered". The system counts the number of records violating the rules; records exceeding the allowed percentage are deemed unqualified.

[0047] Accuracy verification: The system will sample and compare the merged results with authoritative data sources. For example, 100 pieces of enterprise information will be randomly selected and compared with the database of the State Administration for Industry and Commerce. If the accuracy rate is less than 98%, the enterprise will be deemed unqualified.

[0048] Timeliness monitoring: The system calculates the time difference between the generation of data and the completion of data fusion. For example, if sensor data is not processed within 5 minutes, it is considered a timeout.

[0049] S4. Based on the service layer, based on the data and knowledge generated after the three-level fusion in S3, multiple business service modules and intelligent analysis services are encapsulated, and a service resource pool is built. The service resource pool supports service registration, discovery and dynamic invocation. Business service module encapsulation: Based on integrated knowledge graphs and standardized data, develop independent microservices to form reusable business service modules. These modules should include at least: Business processing module: Encapsulates cross-departmental business processes, such as "One-stop service for business establishment".

[0050] Comprehensive query module: Provides knowledge graph-based related queries, such as "enterprise panoramic profile query".

[0051] Analysis and early warning module: encapsulates real-time risk scanning and early warning rules based on graph patterns.

[0052] Collaboration Management Module: Enables automatic creation, distribution, and status tracking of tasks across organizations.

[0053] Intelligent analytics service integration: Machine learning prediction models (XGBoost), anomaly detection algorithms (Temporal Convolutional Network TCN), and other technologies are encapsulated as independent intelligent services. These services leverage the interconnected information of the knowledge graph to provide personalized recommendations, risk warnings, and decision support capabilities.

[0054] Service Resource Pool Construction: All the aforementioned business services and intelligent analytics services are provided externally via standardized API interfaces (RESTful API or gRPC interface). Upon startup, these service instances automatically register their network address, service name, and metadata with the Service Registry and Discovery Center (Nacos). Service consumers (such as application-layer portals) discover available services by querying this center and dynamically invoke services through technologies such as client-side load balancing, thus forming a loosely coupled and scalable service resource pool.

[0055] S5. Based on the application layer, a unified service portal is built, which integrates and calls the services in the service resource pool obtained in S4, and provides personalized intelligent one-stop service entry for different users according to user identity and business needs.

[0056] Build a unified service portal: Portal Development and Multi-Channel Adaptation: The core interactive interface of the unified service portal was developed using the front-end framework (React). The portal supports multi-channel access, including: Provide responsive web pages for web browsers.

[0057] Provide Hybrid Apps or Mini Programs for mobile applications.

[0058] Provide customized interfaces for smart terminals (such as government service self-service machines).

[0059] Provide standard integration interfaces (such as OpenAPI) for third-party systems.

[0060] The portal's backend adaptive rendering engine can automatically select and render the most suitable template based on the terminal type from which the request originates.

[0061] Unified Identity Management and Personalization: The platform establishes a unified identity authentication and authorization center (implemented based on the OAuth 2.0 protocol). User logins from all channels are completed through this center, enabling single sign-on across channels. This center centrally manages user accounts, roles, and permissions. Based on the logged-in user's identity and role, the portal dynamically loads and presents service modules and data within the user's authorized scope from the service resource pool, achieving unified personalized service recommendations and interface customization, truly providing a "one-person-one-interface" intelligent one-stop service experience.

[0062] The specific implementation process is as follows: To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and do not limit the scope of protection of this invention.

[0063] This embodiment takes the construction of a "smart city one-stop public service platform" as an example to illustrate the implementation process of the present invention in detail.

[0064] Step 1: Platform Architecture Setup Reference Figure 1 The system architecture shown first constructs a four-layer logical architecture for the platform and a comprehensive operation and maintenance monitoring layer.

[0065] 1. Data Layer: A distributed data lake is built based on the Hadoop Distributed File System (HDFS) to serve as a unified storage platform for raw data.

[0066] 2. Fusion Layer: As the core processing layer, it is a custom-developed, configurable, multi-level fusion workflow engine based on the Apache Airflow workflow engine. Operations personnel can use its web interface to drag and drop workflows containing task nodes such as "data cleaning," "feature alignment," "model training," and "quality assessment," and configure the execution parameters and dependencies of each node.

[0067] 3. Service Layer: Built using the Spring Cloud microservice framework to ensure that each business function unit (service) can be developed, deployed and extended independently.

[0068] 4. Application Layer: A unified service portal is developed using the React front-end framework and integrated with an adaptive rendering engine. This engine can automatically identify the access terminal (such as PC browser, mobile device, WeChat mini program) and load the corresponding interface templates and component libraries to achieve a consistent experience across channels.

[0069] 5. Operation and Maintenance Monitoring Layer: Deploy Prometheus as a performance metric collector, Grafana as a visualization dashboard, and the ELK (Elasticsearch, Logstash, Kibana) stack as a log management and analysis center, which together form the infrastructure for full-coverage monitoring of the following layers.

[0070] Step 2: Multi-source data acquisition and standardized storage The data layer receives data through a "unified access gateway" (developed based on Nginx extensions and integrating multiple protocol adapters): For government APIs: The gateway's HTTP adapter calls the interface at preset intervals to obtain JSON / XML format data.

[0071] For file data: The gateway provides an HTTPS-based encrypted file upload interface to receive Excel reports and PDF documents from various departments.

[0072] For real-time streaming data: The gateway's Kafka client subscribes to video stream tags from urban traffic cameras and MQTT topic messages reported by environmental sensors (Internet of Things) for real-time data collection.

[0073] After all the collected raw data enters the data lake, it is stored hierarchically according to "data type" (text / image / time series) and "business domain" (transportation / environmental protection / medical). At the same time, a metadata crawling service is launched to automatically scan new data, extract its technical attributes (format, size, location) and business attributes (source department, update cycle), and register this information with the central metadata management system built on Apache Atlas to form a global data asset catalog.

[0074] Step 3: Configurable multi-level data fusion processing The Airflow workflow engine in the fusion layer triggers the following tasks in sequence: 1. Primary Integration: Cleaning and Standardization The "Data Cleaning Microservice" is invoked. This service embeds the Drools rule engine, executing business rules such as "ID number validity verification" and "administrative division code standardization." Simultaneously, an anomaly detection algorithm based on Isolation Forest is run to identify and handle numerical outliers. Missing fields are filled using the mean or median of similar data.

[0075] After cleaning, all data in different formats are uniformly converted into the Apache Parquet columnar storage format, and the naming and type of key fields (such as timestamp, location_id) are standardized, outputting a "standardized data table".

[0076] 2. Second-level fusion: Feature and semantic alignment Feature extraction: For the text of "Citizen Hotline Work Orders", NLP microservices are used for word segmentation and entity recognition, and word vectors are generated using the Word2Vec model. For "Traffic Incident Images", CV microservices are used to extract deep feature vectors using a pre-trained ResNet-50 model.

[0077] Alignment operation: Based on pre-configured spatiotemporal association rules (such as "event association within a 1-kilometer range and a 10-minute time window") and semantic mapping tables (such as "business registration.company name = taxation.taxpayer name"), the similarity of feature vectors from different sources is calculated to achieve spatiotemporal and semantic alignment across data sources and generate an "aligned feature wide table".

[0078] 3. Three-level integration: Knowledge graph generation Model and Training: Construct a multi-task learning model, which shares a fully connected network at the bottom layer and has three task heads at the top layer: entity recognition (sequence labeling), relation extraction (entity pair-based classification), and attribute completion (classification / regression).

[0079] Training data preparation: Combining the feature wide table produced by "secondary fusion", labeled training samples are generated through remote supervision (using existing knowledge base for entity linking) and rule-based automatic annotation; the labeled content of the training samples corresponds to the entity, relation and attribute extraction targets of the multi-task learning model, ensuring that the model training direction is consistent with the knowledge graph construction requirements.

[0080] Joint Training and Inference: Labeled data is input into the model, and joint training is performed using a weighted summation loss function. After training, inference is performed on the full feature wide table, and triples in the form of <entity, relation, entity> and <entity, attribute, value> are output in batches.

[0081] Graph Construction: Import triples into the Neo4j graph database to automatically construct a "city operation fusion knowledge graph".

[0082] 4. Integration of quality closed-loop optimization (corresponding to claim 7) The workflow includes a "quality assessment node" that evaluates the completeness, consistency, accuracy, and timeliness of the fusion output at each level.

[0083] If the consistency of the secondary fusion output is lower than the threshold (e.g., 95%), the evaluation module will automatically trigger the re-execution of the secondary fusion task through the Airflow API, or send instructions to dynamically adjust the similarity threshold of feature alignment, forming a closed loop of "processing-evaluation-feedback-optimization".

[0084] Step 4: Intelligent Service Packaging and Governance The service layer encapsulates services based on the aforementioned "fusion knowledge graph" and standardized data: Business service module: Develop independent microservices, such as "Enterprise Panoramic Profile Query" (comprehensive query), "High-Risk Industrial Chain Early Warning" (analysis and early warning), and "Cross-departmental Joint Handling of Matters" (collaborative management).

[0085] Intelligent analysis service: It encapsulates XGBoost prediction model, Temporal Convolutional Network (TCN) anomaly detection algorithm and other technologies into services, and provides capabilities such as "personalized policy push" and "infrastructure risk warning".

[0086] Service resource pool: All services are published as RESTful APIs and registered with the Nacos service registry. Consumers such as portals discover services through Nacos, enabling dynamic service invocation and load balancing.

[0087] Step 5: One-stop portal and unified access 1. Multi-channel portal: The unified service portal provides adaptive interactive interfaces for the Web, mobile App, and self-service terminals. The backend adaptive rendering engine automatically adapts the interface based on the User-Agent of the request.

[0088] 2. Unified Identity Management: Establish a unified authentication and authorization center based on the OAuth 2.0 protocol. Users can log in once and access services within their authorized scope on all channels, achieving true single sign-on and personalized service delivery.

[0089] Step 6: Full-Stack Intelligent Operation and Maintenance Support The operations and maintenance monitoring layer continues to work: Prometheus collects real-time metrics such as QPS, latency, and error rate for each microservice.

[0090] The Grafana dashboard centrally displays the system's health status.

[0091] The Prophet time-series prediction model, integrated into the monitoring stack, can predict the future trends of key indicators and provide early warnings before potential anomalies occur.

[0092] When the system detects that "the memory usage of the data fusion task continues to rise", it can automatically issue an alert and coordinate with Kubernetes to expand resources for the task Pod, ensuring high availability of the platform.

[0093] Through the implementation of the above steps, this solution has successfully built a platform that can deeply mine knowledge from multi-source heterogeneous data and provide intelligent one-stop services, effectively solving the traditional problems of data silos and service fragmentation.

[0094] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for constructing an intelligent one-stop service platform based on data fusion, characterized in that, Includes the following steps: S1. Platform architecture construction: Construct a four-layer architecture including a data layer, an integration layer, a service layer, and an application layer. The data layer is built based on a distributed data lake, the integration layer has a built-in configurable multi-level integration workflow engine, the service layer adopts a microservice architecture, and the application layer deploys an adaptive rendering engine. S2. Based on the data layer, raw data is collected from multiple heterogeneous data sources through a unified access gateway, and the collected multi-source data is classified and stored in a hierarchical manner according to data type and business domain, and a metadata management index is established. S3. Based on the fusion layer, the multi-source data stored in S2 is configured and sequentially triggered to perform first-level fusion, second-level fusion, and third-level fusion processing through the multi-level fusion workflow engine; The first-level fusion is used to clean and standardize multi-source data, eliminating data redundancy and format differences; The secondary fusion is used to align the features and semantics of the data after the primary fusion, so as to achieve feature unification and spatiotemporal semantic matching of multi-source data. The three-level fusion is used to perform deep fusion and knowledge generation based on the data aligned by the two-level fusion, to uncover hidden relationships between data and form structured knowledge; S4. Based on the service layer, and based on the data and knowledge generated after the three-level fusion in S3, multiple business service modules and intelligent analysis services are encapsulated, and a service resource pool is constructed. The service resource pool supports service registration, discovery and dynamic invocation. S5. Based on the application layer, build a unified service portal, integrate and call the services in the service resource pool obtained in S4, and provide personalized intelligent one-stop service entry for different users according to user identity and business needs.

2. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 1, characterized in that, The first-level fusion includes: The collected multi-source data is processed by a configurable rule engine and anomaly detection intelligent algorithm to remove duplicates, correct errors, and fill in missing values, resulting in clean data. The cleaned data in different formats will be converted into standardized data in a unified format.

3. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 2, characterized in that, The secondary fusion includes: The standardized data is subjected to feature extraction and semantic representation; different data types are processed separately: natural language processing techniques are used to perform semantic analysis and word vector extraction on text data, and feature extraction algorithms are used to extract feature vectors for image, audio and video data; then, the different data types from multiple sources are converted into a unified feature vector representation. Based on spatiotemporal association rules and semantic mapping relationships, the feature vectors of the multi-source data are spatiotemporally and semantically aligned to generate aligned multi-source feature data.

4. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 3, characterized in that, The three-level fusion includes: Construct a multi-task learning model and input the aligned multi-source feature data into the multi-task learning model for joint training; Through joint training, a fused knowledge graph is generated, which is used to represent and mine hidden relationships between entities in multi-source feature data.

5. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 4, characterized in that: The multi-task learning model includes entity recognition, relation extraction, and attribute completion; the fused knowledge graph is constructed from the structured triples of entities, relations, and attributes obtained through joint training.

6. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 1, characterized in that, In step S2, the method of collecting raw data from multiple heterogeneous data sources includes at least two of the following: application programming interface calls, file upload and import, real-time data stream collection, and direct access by IoT devices.

7. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 1, characterized in that: The S3 also includes a fusion quality assessment module, which is used to quantitatively assess the data quality of the fusion output at each level. The assessment indicators include completeness, consistency, accuracy, and timeliness. Based on the assessment results, if any assessment indicator fails to reach a preset threshold, the corresponding level of fusion process will be re-executed or the fusion parameters will be dynamically adjusted to achieve closed-loop optimization of the fusion process.

8. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 1, characterized in that: In S4, the business service module includes business processing, comprehensive query, analysis and early warning, and collaborative management modules; the intelligent analysis service integrates machine learning prediction models and anomaly detection algorithms to provide personalized recommendations, risk warnings, and decision support capabilities.

9. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 1, characterized in that, In S5, the unified service portal supports multi-channel access, including web browsers, mobile applications, smart terminals and third-party system integration interfaces, and realizes cross-channel single sign-on and unified permission management through a unified identity authentication and authorization center.

10. The method for constructing an intelligent one-stop service platform based on data fusion as described in claim 1, characterized in that, It also includes an operation and maintenance monitoring layer, which is used to monitor and visualize the operation status, performance indicators and security events of the data layer, fusion layer, service layer and application layer in real time, and predict and automatically alarm abnormal situations based on machine learning algorithms to ensure the high availability and security of the platform.

Images