A data fusion processing method and system based on edge computing

Abstract

The present application discloses a data fusion processing method and system based on edge computing, which relates to the field of data processing. The method includes: accessing multiple data sources at the edge node; performing preprocessing operations on the data of the above multiple data sources to obtain preprocessed data; performing feature extraction operations on the above preprocessed data to obtain feature information related to provident fund business; using a local fusion model based on the above feature information to perform local intelligent judgment operations at the edge node to obtain fusion processing results; sending the above fusion processing results and their associated tags to the central server for business collaborative decision-making. The present invention breaks through the architectural limitations of traditional centralized approval and realizes an innovative path of data fusion, intelligent judgment, and system collaboration. It is particularly suitable for large-scale, real-time business processing scenarios in provident fund intelligent service systems.

Description

Technical Field

[0001] This specification relates to the field of data processing. More specifically, this application relates to a data fusion processing method and system based on edge computing. Background Art

[0002] With the continued digitization of housing provident fund services, more and more users are submitting withdrawal applications, loan applications, and home purchase eligibility verification through online channels. These processes typically involve multiple steps, including user identity verification, provident fund contribution record verification, document consistency verification, and risk assessment. These processes are characterized by diverse data sources, complex document types, and a high reliance on human experience for decision-making.

[0003] Traditional provident fund business processing relies primarily on centralized servers to centrally receive and process all application data. This presents several significant issues: First, processing speed is slow. All images, forms, and behavioral data must be uploaded to the centralized server, which can easily cause network congestion and review delays. Second, computing pressure is high. The server must handle a large number of tasks such as image recognition and information comparison, resulting in high resource consumption. Third, risk identification is delayed. The inability to quickly identify risky users at the edge easily burdens back-end systems. Fourth, labor costs are high. Processes such as identity verification and document review still rely heavily on manual operations, which are inefficient and prone to errors. Finally, the system is not flexible enough. During peak business hours or when the network is unstable, the system's processing capacity decreases, impacting the user experience.

[0004] Therefore, there is an urgent need for an intelligent processing method that integrates edge computing capabilities and can complete data extraction, judgment and collaborative flow locally to improve the overall system efficiency and intelligence level. Summary of the Invention

[0005] The Summary of the Invention introduces a series of simplified concepts that will be further described in the Detailed Description of the Invention. The Summary of the Invention of this application is not intended to limit the key features and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0006] In the first aspect, this application proposes a data fusion processing method based on edge computing, including:

[0007] Access multiple data sources at the edge node, where the multiple data sources include user business processing data, identity authentication data, and image or document material data;

[0008] Performing preprocessing operations on the data from the multiple data sources to obtain preprocessed data, wherein the preprocessing operations include data format unification, outlier processing, time series alignment, and structured conversion;

[0009] Performing feature extraction on the pre-processed data to obtain feature information related to provident fund business, wherein the feature information includes user behavior feature information, identity feature information, and material consistency feature information;

[0010] Based on the above feature information, a local fusion model is used to perform local intelligent judgment operations at the edge node to obtain fusion processing results. The above local intelligent judgment operations include preliminary qualification review, loan risk assessment and abnormal behavior detection;

[0011] The above fusion processing results and their associated tags are sent to the central server for business collaborative decision-making.

[0012] In a feasible implementation, the specific steps of extracting the above-mentioned user behavior feature information include:

[0013] Based on the pre-processed user business processing data, the user behavior sequence is constructed, and the user access path pattern information is extracted by using position embedding coding combined with a sliding window mechanism;

[0014] Based on the above pre-processed user business processing data, a time-aware behavior clustering algorithm is used to extract the user's operating habit information, wherein the above operating habit information includes operation frequency and interaction time distribution;

[0015] The user behavior feature information is determined based on the user access path pattern information and the operation habit information.

[0016] In a feasible implementation manner, the determining of the user behavior characteristic information based on the user access path pattern information and the operation habit information includes:

[0017] Modeling the user access path pattern information as a weighted directed graph, wherein the nodes of the weighted directed graph represent operation pages or function entrances, the edges of the weighted directed graph represent page jumps, and the edge weights of the weighted directed graph represent jump frequencies;

[0018] Calculate structural indicators based on the weighted directed graph, where the structural indicators include node in-degree distribution, path entropy, and average jump length;

[0019] Based on the above operation habit information, the operation rhythm vector is constructed, and the temporal consistency score is obtained by calculating the cosine similarity with the historical rhythm pattern;

[0020] The above structural indicators and the above time consistency scores are used to construct the above user behavior feature information.

[0021] In a feasible implementation, the specific steps of extracting the identity feature information include:

[0022] Based on the pre-processed authentication data, the user authentication fields are standardized and the regional attributes are extracted;

[0023] Extract identity compliance features related to home purchase eligibility based on the pre-processed identity authentication data, where the identity compliance features include whether the applicant is a first-time homebuyer, marital status, and family co-loan suitability;

[0024] Extracting identity stability features related to provident fund contributions based on the pre-processed identity authentication data, wherein the identity stability features include contribution type, contribution history, and abnormal contribution amount information;

[0025] Based on the above-mentioned regional attributes, the above-mentioned identity compliance characteristics and the above-mentioned identity stability characteristics, a logical consistency comparison analysis is performed to generate integrated identity characteristic information.

[0026] In a feasible implementation manner, the specific steps of extracting the above-mentioned material consistency characteristic information include:

[0027] Perform field-level structured recognition on pre-processed image or document data to extract key field information related to provident fund business;

[0028] Based on the above key field information and user-filled information and identity authentication data, a one-to-one field comparison is performed to calculate the field consistency score;

[0029] Cross-check the same fields between multiple documents uploaded in the same business application, build a field coordination comparison matrix, and output a cross-material consistency score;

[0030] Use a lightweight image forgery detection algorithm to analyze the authenticity of image materials and obtain image authenticity scores;

[0031] The above-mentioned field consistency score, the above-mentioned cross-material consistency score and the above-mentioned image authenticity score are integrated to construct the above-mentioned material consistency feature information.

[0032] In a feasible implementation, the above-mentioned local intelligent judgment operation is performed at the edge node using a local fusion model based on the above-mentioned feature information to obtain a fusion processing result, including:

[0033] Get business type information;

[0034] Generating derived characteristic information based on the characteristic information and the business type information;

[0035] Performing a sparse coding operation on the derived feature information and the feature information to obtain edge node feature information;

[0036] The local fusion model is used to perform local intelligent judgment operations on the edge nodes based on the feature information of the edge nodes to obtain the fusion processing results.

[0037] In a feasible embodiment, the above-mentioned local fusion model includes an input layer, a feature transformation layer, a multi-task output structure and an output layer. The above-mentioned feature transformation layer includes a dynamic feature cross-coding submodule, a local attention feature enhancement submodule and a multi-channel feature fusion submodule. The above-mentioned multi-task output structure includes a first output channel, a second output channel and a third output channel. The above-mentioned first output channel is used to perform the preliminary qualification judgment task, the above-mentioned second output channel is used to perform the loan risk scoring task, and the above-mentioned third output channel is used to perform the abnormal behavior detection task.

[0038] In a feasible implementation, the first output channel includes a shallow multi-layer perceptron model, the second output channel includes a lightweight regression network, and the third output channel includes a multi-label classifier.

[0039] In a feasible implementation, the specific steps of the central server performing business collaborative decision-making include:

[0040] Perform field parsing and rule matching on the above fusion processing results to obtain qualification judgment results, risk score values and abnormal behavior labels;

[0041] Execute decision path diversion based on the qualification judgment result, the risk score value, and the abnormal behavior label to obtain a decision structure, wherein the diversion includes any one of automatic flow, manual review triggering, risk warning registration, and user correction notification;

[0042] A feedback information package is generated based on the above decision results and pushed to the user terminal.

[0043] In the second aspect, this application proposes a data fusion processing system based on edge computing, including:

[0044] A receiving unit, configured to access multiple data sources at the edge node, wherein the multiple data sources include user service processing data, identity authentication data, and image or document material data;

[0045] A preprocessing unit, configured to perform preprocessing operations on the data from the plurality of data sources to obtain preprocessed data, wherein the preprocessing operations include data format unification, outlier processing, time series alignment, and structured conversion;

[0046] A feature extraction unit, configured to perform a feature extraction operation on the pre-processed data to obtain feature information related to provident fund business, wherein the feature information includes user behavior feature information, identity feature information, and material consistency feature information;

[0047] An acquisition unit is configured to perform a local intelligent judgment operation at the edge node using a local fusion model based on the feature information to obtain a fusion processing result, wherein the local intelligent judgment operation includes extracting preliminary qualification review, loan risk assessment, and abnormal behavior detection;

[0048] The decision-making unit is used to send the above fusion processing results and their associated tags to the central server for business collaborative decision-making.

[0049] In summary, the present invention proposes a data fusion processing method based on edge computing. Compared to existing technologies, the proposed method deploys data preprocessing, feature extraction, and preliminary intelligent judgment to edge nodes, significantly reducing central server load, bandwidth consumption, and network latency. This method supports the extraction of composite features from user behavior data, identity authentication data, and image / document material data, and constructs highly expressive behavior models, identity compliance models, and material consistency models, effectively improving risk identification capabilities and judgment accuracy. A local fusion model enables the parallel output of eligibility judgment, risk scoring, and anomaly detection, supporting the simultaneous triggering of multi-dimensional business paths, reducing duplicate calculations, and improving audit throughput. Edge nodes generate structured labeling results and push them to the central system, enabling rapid decision-making and feedback from the central side. It also supports continuous optimization of edge models and enhanced interpretability of results. This method can be deployed on any edge node with basic computing capabilities (such as provident fund self-service kiosks, business front-end terminals, and government cloud edge nodes). It does not rely on high-performance computing resources and has excellent adaptability and scalability. In summary, the present invention breaks through the architectural limitations of traditional centralized approval and realizes an innovative path of data fusion, intelligent judgment, and system collaboration. It is particularly suitable for large-scale, real-time business processing scenarios in provident fund intelligent service systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present description. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0051] Figure 1 A schematic diagram of a data fusion processing method based on edge computing provided in an embodiment of the present application;

[0052] Figure 2 A schematic diagram of the specific steps for extracting user behavior feature information provided in an embodiment of the present application;

[0053] Figure 3A schematic diagram of the process of determining user behavior feature information provided in an embodiment of the present application;

[0054] Figure 4 A schematic diagram illustrating the specific steps for extracting identity feature information provided in an embodiment of the present application;

[0055] Figure 5 A schematic diagram of the specific steps for extracting material consistency feature information provided in an embodiment of the present application;

[0056] Figure 6 A schematic diagram of a specific extraction operation process for obtaining a fusion processing result provided in an embodiment of the present application;

[0057] Figure 7 A schematic diagram of the structure of a data fusion processing system based on edge computing provided in an embodiment of the present application. DETAILED DESCRIPTION

[0058] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the embodiments described are only some embodiments of the present application, not all embodiments.

[0059] Figure 1 A schematic diagram of a data fusion processing method based on edge computing provided in an embodiment of the present application is provided. The method may specifically include:

[0060] S110. Access multiple data sources at the edge node, where the multiple data sources include user service processing data, identity authentication data, and image or document material data;

[0061] S120: performing preprocessing operations on the data from the plurality of data sources to obtain preprocessed data, wherein the preprocessing operations include data format unification, outlier processing, time series alignment, and structured conversion;

[0062] S130: Perform feature extraction on the pre-processed data to obtain feature information related to provident fund business, wherein the feature information includes user behavior feature information, identity feature information, and material consistency feature information;

[0063] S140. Performing a local intelligent judgment operation at the edge node using a local fusion model based on the feature information to obtain a fusion processing result, wherein the local intelligent judgment operation includes extracting preliminary qualification review, loan risk assessment, and abnormal behavior detection;

[0064] S150: Send the above fusion processing results and their associated tags to the central server for business collaborative decision-making.

[0065] Exemplarily, in step S110, multiple data sources are accessed at the edge node, and the multiple data sources include: (1) user business processing data, which is used to describe the user's interactive behavior, operation trajectory, click path and other information during the business processing process; (2) identity authentication data, including structured fields such as the user's real-name registration information, historical provident fund payment records, and unit reporting information; (3) image or document material data, which refers to image files or PDF documents of materials such as house purchase contracts, real estate certificates, provident fund payment certificates, and income certificates uploaded by users.

[0066] In step S120, the system preprocesses the aforementioned multiple data sources to obtain a consistent, structured dataset. This preprocessing includes the following aspects: Standardizing data formats for data from different sources, such as performing OCR structured extraction on PDF documents and mapping and standardizing identity fields across different systems; performing outlier processing on missing items and extreme values in the original data to enhance data credibility; aligning data generated across multiple sources and at different times according to user dimensions to form a business event stream; and converting semi-structured data, such as image materials, into structured data, extracting them into key-value pairs or table fields for subsequent analysis.

[0067] In step S130, the system performs feature extraction based on the pre-processed data to construct a feature vector suitable for intelligent judgment of provident fund business. Specifically, it includes three types of features:

[0068] 1. User behavior feature information: Build jump graphs and operation rhythm vectors based on user access paths, and extract behavioral features such as access stability and behavior deviation index;

[0069] 2. Identity feature information: Extract identity compliance and trustworthiness indicators such as regional attributes, whether the applicant is a first-time homebuyer, marital status and co-loan compatibility, and deposit record stability from the user's identity authentication data;

[0070] 3. Material consistency feature information: Evaluate the consistency of user-submitted materials in semantic, image, and logical dimensions through field-level comparison, image forgery detection, and cross-material field comparison matrix.

[0071] In step S140, the system inputs the above features into the local fusion model deployed on the edge node to perform intelligent judgment operations. The fusion model supports a multi-task output structure and can complete the following tasks in parallel:

[0072] 1. Initial qualification review: Determine whether the user meets the basic entry requirements, such as payment month, first home identification, etc.;

[0073] 2. Loan risk assessment: Generate a loan risk score based on the intersection of behavior and identity;

[0074] 3. Abnormal behavior detection: Identify potential high-risk behaviors such as material forgery, field inconsistency, and access anomalies.

[0075] In step S150, the system packages the fusion processing results and their associated tags into a structured data package (e.g., JSON format), including information such as the judgment result, risk score, anomaly flags, and model version, and uploads it to the central server via a secure communication channel. Upon receiving this information, the central server can perform actions such as automatic approval, manual review, risk registration, and user notification based on business rules and platform policies, establishing a complete edge-center collaborative decision-making chain.

[0076] In summary, this embodiment proposes a data fusion processing method based on edge computing. Compared to existing technologies, the method proposed in this application deploys data preprocessing, feature extraction, and preliminary intelligent judgment to edge nodes, significantly reducing central server load, bandwidth consumption, and network latency. This method supports the extraction of composite features from user behavior data, identity authentication data, and image / document material data, and constructs highly expressive behavior models, identity compliance models, and material consistency models, effectively improving risk identification capabilities and judgment accuracy. A local fusion model enables the parallel output of eligibility judgment, risk scoring, and anomaly detection, supporting the simultaneous triggering of multi-dimensional business paths, reducing duplicate calculations, and improving review throughput. Edge nodes generate structured labeling results and push them to the central system, enabling rapid decision-making and feedback from the central side. It also supports continuous optimization of edge models and enhanced interpretability of results. This method can be deployed on any edge node with basic computing capabilities (such as provident fund self-service kiosks, business front-end terminals, and government cloud edge nodes). It does not rely on high-performance computing resources and has excellent adaptability and scalability. In summary, this embodiment breaks through the architectural limitations of traditional centralized approval and realizes an innovative path of data fusion, intelligent judgment, and system collaboration. It is particularly suitable for large-scale, real-time business processing scenarios in provident fund intelligent service systems.

[0077] In one possible implementation, Figure 2 As shown, Figure 2 This is a flow chart of specific steps for extracting user behavior feature information provided in an embodiment of the present application. The specific steps for extracting user behavior feature information include:

[0078] S210: Constructing a user behavior sequence based on the pre-processed user business handling data, and extracting user access path pattern information using position embedding coding combined with a sliding window mechanism;

[0079] S220: Extracting user operation habit information based on the pre-processed user service handling data using a time-aware behavior clustering algorithm, wherein the operation habit information includes operation frequency and interaction time distribution;

[0080] S230: Determine the user behavior feature information based on the user access path pattern information and the operation habit information.

[0081] For example, the process of extracting user behavior feature information mainly focuses on the user's operation path and behavior rhythm when handling provident fund business, and combines sequence modeling and time analysis methods to achieve automatic construction and expression of behavior features. Specifically, the process includes the following consecutive steps:

[0082] First, the system constructs a user behavior sequence based on the pre-processed user business processing data. The behavior sequence is arranged in chronological order and is recorded as , where each Indicates that the user The page or operation node visited at a time point, Indicates the total number of steps in this transaction. To effectively encode the user's access path, the system uses position embedding to vectorize each operation node and combines it with a sliding window mechanism to extract local subpath patterns, thereby forming the user's access path characteristics during this transaction.

[0083] Secondly, the system performs weighted aggregation on the operation node vectors in the behavior sequence to generate the access path pattern vector , the specific calculation formula is:

[0084]

[0085] in, Indicates the data processing area of the operation node Embedding vectors in the field of data processing, The data processing field is related to the time step data processing field The weight function related to the data processing field can be sine-cosine encoding or linear position bias, and the sequence level representation data processing field is obtained after fusion In the field of data processing, it is used to characterize the structural and sequential characteristics of user operation paths.

[0086] Subsequently, the system models the user's operation frequency in the time dimension, counts the user's operation frequency in each time period within 24 hours a day, and constructs the operation rhythm vector ,in Indicates that the user To further identify user behavior patterns, the system introduces a time-aware clustering algorithm (such as the time-weighted K-Means or T-BIRCH method) to perform cluster analysis on rhythm vectors, map users to specific operation behavior groups, and generate behavior category labels. .

[0087] Next, to quantify the consistency between the current behavior and the user's regular behavior, the system calculates the current rhythm vector The rhythm of its cluster center The cosine similarity of , as the temporal consistency score , the calculation formula is:

[0088]

[0089] Where, • represents the vector dot product, It is a 2-norm. The closer the score is to 1, the more consistent the user's current operation behavior is with their historical rhythm pattern, and the more credible the behavior is.

[0090] Finally, the system converts the path pattern vector , temporal consistency score , operation behavior category label path Path length and path selection entropy Combining multiple indicators to construct the final user behavior feature vector :

[0091]

[0092] This feature vector serves as input to subsequent edge judgment modules, supporting multiple tasks such as eligibility assessment, risk scoring, and abnormal behavior identification. This process is highly stable and versatile, adapting to complex user behavior scenarios and dynamic business path changes. It is particularly suitable for automated review and behavior anomaly detection scenarios in housing provident fund services.

[0093] In one possible implementation, Figure 3 As shown, Figure 3A schematic diagram of a process flow of steps for determining user behavior characteristic information provided in an embodiment of the present application, wherein determining the user behavior characteristic information based on the user access path pattern information and the operation habit information includes:

[0094] S310: Model the user access path pattern information as a weighted directed graph, wherein the nodes of the weighted directed graph represent operation pages or function entrances, the edges of the weighted directed graph represent page jumps, and the edge weights of the weighted directed graph represent jump frequencies;

[0095] S320: Calculate structural indicators based on the weighted directed graph, wherein the structural indicators include node in-degree distribution, path entropy, and average jump length;

[0096] S330: construct an operation rhythm vector based on the above operation habit information, and obtain a temporal consistency score by calculating the cosine similarity with the historical rhythm pattern;

[0097] S340: Construct the user behavior feature information using the structural index and the time consistency score.

[0098] For example, in one feasible implementation, in order to extract behavioral feature information strongly related to provident fund business from the user's operation process, the system integrates and analyzes the user's access path pattern and operation habit information, and constructs multi-dimensional behavioral features through graph modeling and time series mining. The specific process is as follows:

[0099] First, the system constructs the user's operation records during the business process into an access path sequence, where each element in the sequence represents a page access or function call. Based on this sequence, the system establishes a weighted directed graph. , used to represent the user's access path structure, where It is a set of nodes, each node represents an operation page or function entrance; is an edge set, each directed edge Indicates that the user has accessed the page Jump to page ; is the edge weight set, Indicates the frequency of the jump, reflecting the activity level of the path.

[0100] Specifically, the jump frequency The calculation is as follows:

[0101]

[0102] in: For users in behavioral data from the page Jump to page the number of times; For users from the page The sum of all jumps issued; It is the normalized jump weight, which is used to indicate the importance or frequency of jump.

[0103] Based on this graph, the system extracts the following structural indicators as the basis for determining access paths:

[0104] 1. Node in-degree distribution: out-degree For nodes The number of all outgoing edges; the in-degree For nodes The system can calculate the average in-degree and standard deviation of all nodes as the node in-degree distribution to reflect the complexity and dispersion of the operation process.

[0105] 2. Path Entropy: To measure the concentration of user paths, the following formula can be used for each page The information entropy of jump distribution is calculated as follows:

[0106]

[0107] The total path entropy is the average of all node entropies:

[0108]

[0109] in, Represents a slave node To Node The lower the path entropy, the more concentrated and predictable the user behavior is, which usually corresponds to a skilled and normal operation path.

[0110] 3. Average jump length :The system calculates the average distance of page jumps based on the jump path, which can be expressed as:

[0111]

[0112] in: : The logical distance between pages. If it is a direct jump, it can be set to 1; : Indicates the frequency of the jump; this indicator can reflect whether the operation path is reasonable. For example, a jump that is too far may indicate an abnormal process.

[0113] Secondly, the system counts the user's operation frequency within 24 hours based on the operation timestamp information and constructs the operation rhythm vector:

[0114]

[0115] in: For users in The number of operations within an hour. This vector reflects the user's temporal behavior pattern. For example, high frequency concentrated during working hours may represent compliance operations.

[0116] The system will change the current operating rhythm and historical rhythm mean Or the user group center Perform cosine similarity matching to obtain the temporal consistency score:

[0117]

[0118] in: is the vector dot product, indicating the direction similarity; : Represents the 2-norm (Euclidean length) of the vector; the higher the similarity (closer to 1), the more consistent the behavior time pattern is with historical / normal users, and the higher the credibility.

[0119] Finally, the system combines the above structural indicators with the time consistency score Concatenate into a set of vectors:

[0120]

[0121] in, is the path entropy, is the average jump length, is the average out-degree, is the average indegree, Score for temporal consistency.

[0122] The behavior feature vector This approach can serve as input to the intelligent judgment module within the provident fund system, enabling efficient initial qualification review, risk identification, and anomaly warning. It is particularly suitable for identifying potential violations caused by unreasonable process jumps, unusually dense time schedules, or frequent path switching. This approach combines structural and temporal modeling to enhance the system's ability to represent user behavior and enhance its robustness.

[0123] In one possible implementation, Figure 4 As shown, Figure 4 This is a flow chart of specific steps for extracting identity feature information provided in an embodiment of the present application. The specific steps for extracting identity feature information include:

[0124] S410, standardizing the user identity authentication field based on the preprocessed identity authentication data and extracting the regional attribute;

[0125] S420: Extracting identity compliance features related to home purchase eligibility based on the pre-processed identity authentication data, wherein the identity compliance features include whether the applicant is a first-time homebuyer, marital status, and suitability for family co-loan;

[0126] S430: Extracting identity stability features related to provident fund contributions based on the pre-processed identity authentication data, wherein the identity stability features include contribution type, contribution history, and abnormal contribution amount information;

[0127] S440: Perform a logical consistency comparison analysis based on the regional attributes, the identity compliance characteristics, and the identity stability characteristics to generate integrated identity characteristic information.

[0128] For example, first, based on the pre-processed authentication data, the user authentication fields are standardized, the field naming and encoding methods are unified, and the regional attributes are extracted. ,in Represents a regional vector, including province code, city level, and regional economic level.

[0129] Second, extract identity compliance features related to home purchase qualifications , including: whether the applicant is the first homebuyer ; Marital status code , representing single, married, and divorced respectively; Family co-loan suitability score , which is derived based on the matching degree between spouse’s identity and income. Therefore, the identity compliance vector can be constructed .

[0130] Next, extract the identity stability features related to provident fund contributions , including: deposit type , indicating that the individual unit pays on behalf of the individual; the payment history stability score , estimated based on the continuity and interruption of contributions in the past 24 months; volatility of contribution amount The coefficient of variation (CV) was used for calculation:

[0131]

[0132] in: is the average monthly payment amount; is the standard deviation; is a small constant to avoid division by zero. Finally, the identity stability vector .

[0133] Will As three-category vector input to the consistency mapping function In , logical comparison and fusion coding are performed, which are defined as follows:

[0134]

[0135] in: represents the fused identity feature vector; is the weight matrix used for linear projection of different feature dimensions; is the bias term; is the activation function (such as ReLU or tanh).

[0136] Next, based on the above regional attributes , Identity Compliance Vector and identity stability vector , perform logical consistency comparison analysis to identify potential conflicts or supportive relationships between identity information of different dimensions, and generate fused identity feature information based on this

[0137] First, we construct the joint interaction structure between the two types of features and define the consistency tensor mapping:

[0138]

[0139] in, Represents a third-order outer product operation, generating a tensor Used to capture the structural coupling and semantic relationships between different identity dimensions.

[0140] To simplify high-order tensor calculations while preserving core interaction information, an attention compression mechanism is introduced to map tensors into consistency kernel scores:

[0141]

[0142] in, The attention weights are trainable or rule-based, reflecting the importance of the interaction between the three fields, satisfying the following requirements:

[0143]

[0144] At the same time, in order to improve the interpretability of identity consistency scoring, the cross-domain cosine similarity indicator is further introduced:

[0145]

[0146] The three types of similarities are combined in the form of weighted average and defined as auxiliary consistency scoring items:

[0147]

[0148] in, , each coefficient can be set according to business priority.

[0149] Finally, the fused identity feature information representation is constructed , based on logical consistency scoring and original field concatenation, is defined as follows:

[0150]

[0151] in, represents vector concatenation, is the linear mapping weight matrix, is the bias term, For activation functions (such as ReLU or , income It represents the final fused identity feature information and has the ability to judge structural coupling, behavioral deviation and logical rationality.

[0152] This fusion feature can serve as the core input for subsequent intelligent judgment modules (such as loan eligibility assessment and abnormal risk screening), helping to improve the system's comprehensive identification capabilities of identity authenticity, compliance, and risk.

[0153] In one possible implementation, Figure 5 As shown, Figure 5 This is a schematic diagram of a specific process for extracting material consistency feature information provided in an embodiment of the present application. The specific process for extracting material consistency feature information includes:

[0154] Perform field-level structured recognition on pre-processed image or document data to extract key field information related to provident fund business;

[0155] Based on the above key field information and user-filled information and identity authentication data, a one-to-one field comparison is performed to calculate the field consistency score;

[0156] Cross-check the same fields between multiple documents uploaded in the same business application, build a field coordination comparison matrix, and output a cross-material consistency score;

[0157] Use a lightweight image forgery detection algorithm to analyze the authenticity of image materials and obtain image authenticity scores;

[0158] The above-mentioned field consistency score, the above-mentioned cross-material consistency score and the above-mentioned image authenticity score are integrated to construct the above-mentioned material consistency feature information.

[0159] For example, the process of extracting material consistency feature information includes multiple steps such as structured recognition of images and document materials, field consistency comparison, cross-material verification, and image authenticity assessment, in order to achieve a comprehensive analysis of the content, structure, and authenticity of user-uploaded materials. This process specifically includes the following sequential operations:

[0160] First, the system performs field-level structured recognition on pre-processed image or document material data. Based on OCR technology, layout template matching and natural language parsing algorithms, it identifies key content areas in various materials and extracts key field information sets for provident fund business review. , where each field Represented as a key-value pair , including field names and their corresponding values. The extracted fields include but are not limited to name, ID number, unit name, deposit amount, etc.

[0161] Next, the system compares the key field information with the user's information and identity authentication data at the field level to build a field consistency score. , comparing its value to the corresponding field value in the reference data Compare them. If they are exactly the same, the score is 1. If there are slight differences, the Levenshtein distance is used to calculate the similarity. The score calculation formula is:

[0162]

[0163] in, is the edit distance function, is the smoothing coefficient, which is used to control the error tolerance range. The scores of all fields are averaged to obtain the material field consistency score:

[0164]

[0165] Subsequently, the system cross-checks the same fields in multiple documents uploaded by users in the same business application. For each key field, such as "name", "ID number" and "work address", all documents are compared to build a collaborative field comparison matrix. , among which Rank Columns represent materials and The degree of match on this field is defined as the output of a similarity function, such as the normalized edit distance or cosine similarity after BERT encoding. The main diagonal elements are averaged to obtain the cross-material field consistency score:

[0166]

[0167] in, is the number of common fields in all materials, It represents the self-similarity score of the fields in each material and the same fields in other materials.

[0168] Furthermore, in order to prevent image security risks such as forgery and tampering, the system uses a lightweight image forgery detection algorithm to evaluate the authenticity of uploaded image materials. The algorithm judges the image content through color spectrum distribution analysis, edge continuity detection and frequency domain compression feature recognition, and outputs the forgery probability. , and then get the image authenticity score:

[0169]

[0170] The closer the image authenticity score is to 1, the more credible the image is.

[0171] Finally, the system scores the field consistency , cross-material consistency score and image authenticity score Fusion, building the final material consistency feature information vector :

[0172]

[0173] In one possible implementation, Figure 6 As shown, Figure 6 A schematic diagram of a specific extraction operation process for obtaining a fusion processing result provided in an embodiment of the present application. The above-mentioned local fusion model is used based on the above-mentioned feature information to perform a local intelligent judgment operation at the edge node to obtain the fusion processing result, including:

[0174] S1401. Obtain business type information;

[0175] S1402. Generate derived feature information based on the feature information and the service type information.

[0176] S1403: performing a sparse coding operation on the derived feature information and the feature information to obtain edge node feature information;

[0177] S1404: Perform local intelligent judgment operations at the edge node using a local fusion model based on the edge node feature information to obtain a fusion processing result.

[0178] In a feasible embodiment, the above-mentioned local fusion model includes an input layer, a feature transformation layer, a multi-task output structure and an output layer. The above-mentioned feature transformation layer includes a dynamic feature cross-coding submodule, a local attention feature enhancement submodule and a multi-channel feature fusion submodule. The above-mentioned multi-task output structure includes a first output channel, a second output channel and a third output channel. The above-mentioned first output channel is used to perform the preliminary qualification judgment task, the above-mentioned second output channel is used to perform the loan risk scoring task, and the above-mentioned third output channel is used to perform the abnormal behavior detection task.

[0179] In a feasible implementation, the first output channel includes a shallow multi-layer perceptron model, the second output channel includes a lightweight regression network, and the third output channel includes a multi-label classifier.

[0180] For example, the system uses a local fusion model to intelligently analyze user feature information at the edge node to achieve rapid, task-based judgment of provident fund business requests, improving edge-side response efficiency and local autonomy. The processing flow includes four stages: business type identification, feature derivation, encoding processing, and fusion model reasoning, and ultimately outputs multi-dimensional business processing results. Specifically, it includes the following consecutive steps:

[0181] First, the system obtains the business type information of the current business request at the edge node. This identifies whether the business being processed belongs to a specific business scenario, such as "Provident Fund Loan Application," "Provident Fund Withdrawal," "Account Information Change," or "Deposit Status Verification." This business type information is passed in as a business request parameter and represented as a one-hot encoding or embedded vector, serving as auxiliary input for subsequent feature derivation.

[0182] Subsequently, the system generates derived feature information based on the extracted user feature information (including user behavior features, identity features, and material consistency features) and the current business type information. This process captures the semantic interaction relationship between features under different business types by constructing combined features and conditional logic features. For example: in the loan application scenario, the system crosses the "co-loan adaptability score" and "marital status" to form a "co-loan structure risk feature" to identify potential credit compensation relationships; in withdrawal-related businesses, the system focuses on activating fields related to account status, deposit duration, and unit information changes to determine user withdrawal eligibility; in the anomaly detection scenario, the system combines the "user access path complexity" with the "image authenticity score of uploaded materials" to construct anomaly recognition features to assist in identifying forged materials and process circumvention behaviors.

[0183] After the derived feature vector obtained through the above process is concatenated with the original feature, the system performs sparse coding on it to form a unified input format for edge nodes. Sparse coding includes: performing one-hot encoding and linear projection on categorical fields, normalizing or standardizing numeric fields, and filtering and compressing sparse feature dimensions, and finally constructing the edge node feature vector:

[0184]

[0185] in, Represents the original three types of feature information, represents the derived eigenvector, Represents a splicing operation.

[0186] Next, the system inputs the feature vector into the local fusion model deployed on the edge for inference processing. The fusion model structure includes the following components:

[0187] 1. Input layer: receives sparsely encoded edge feature vectors;

[0188] 2. Feature transformation layer, including the following three sub-modules:

[0189] Dynamic feature cross-coding submodule: automatically learns the nonlinear interaction between different feature combinations to improve feature expression capabilities;

[0190] Local attention feature enhancement submodule: This module introduces a local attention mechanism to adjust the weights of key fields of interest for different tasks, thus enabling task-related feature selection.

[0191] Multi-channel feature fusion submodule: parallel processing of behavior, identity, and material feature channels, and unified fusion mapping to form a shared feature representation;

[0192] 3. Multi-task output structure: The shared feature representation is input into three independent task channels:

[0193] The first output channel is used to perform the initial loan qualification assessment, using a shallow multi-layer perceptron (MLP) model to output the classification result of whether the loan meets the admission requirements.

[0194] The second output channel is used to perform loan risk scoring tasks. It uses a lightweight regression network structure to score user risk with a continuous value (e.g., between 0 and 1).

[0195] The third output channel is used to perform behavioral anomaly detection tasks. It adopts a multi-label classifier structure and simultaneously outputs multiple anomaly labels, such as "forged application", "abnormal operation" and "frequent attempts".

[0196] 4. Output layer: Integrate the three-channel results to form the final fusion processing result output , which respectively represent the preliminary qualification review results, risk score values and abnormal behavior labels.

[0197] Through local fusion model inference processing, edge nodes can complete intelligent responses to business requests locally, eliminating the need to upload all data to the central server. This effectively improves response efficiency, data security, and system scalability. This structure is particularly suitable for public finance business scenarios that are sensitive to processing delays and have high data privacy requirements.

[0198] In a feasible implementation, the specific steps of the central server performing business collaborative decision-making include:

[0199] Perform field parsing and rule matching on the above fusion processing results to obtain qualification judgment results, risk score values and abnormal behavior labels;

[0200] Execute decision path diversion based on the qualification judgment result, the risk score value, and the abnormal behavior label to obtain a decision structure, wherein the diversion includes any one of automatic flow, manual review triggering, risk warning registration, and user correction notification;

[0201] A feedback information package is generated based on the above decision results and pushed to the user terminal.

[0202] For example, after receiving the fusion processing results from the edge nodes, the central server executes a collaborative business decision-making process to achieve unified response, collaborative distribution, and personalized feedback for provident fund business requests. This process involves field parsing, rule matching, path diversion, and feedback generation, ensuring that the system achieves automated, intelligent, and auditable business processing capabilities while meeting compliance requirements.

[0203] First, the central server performs field parsing and rule matching operations on the fusion processing results uploaded by the edge nodes. Specifically, the fusion processing results usually contain three core fields: qualification judgment results It is a binary or multi-class result, indicating whether the user meets the provident fund business access conditions; risk score value A floating point number between 0 and 1, used to quantify the user's potential credit or behavioral risk; abnormal behavior label It is a multi-label set that identifies multiple anomaly types that may exist in the current business.

[0204] After the field is parsed, the server matches it according to the preset business rule table. For example: and , then they are considered as low-risk qualified users; if and ,or , then the manual review process is triggered; if , then determine whether compliance can be restored through material correction.

[0205] Next, the server performs decision routing based on the above results to generate a clear business decision structure. The main routing strategies include:

[0206] 1. Auto-Pass: When a user's qualifications are determined to be acceptable, their risk score is low, and there are no abnormal tags, the next business step (such as loan approval) is automatically completed.

[0207] 2. Manual-Review Trigger: When a high-risk score or abnormal behavior is detected, the system pushes the business request to the manual backend for human review.

[0208] 3. Risk Alert Registration: If the risk score exceeds the warning threshold (e.g., 0.85) or the abnormal tag involves high-risk behavior (e.g., forged materials, identity conflict), an internal risk control event registration is generated and linked to the risk control model for further tracking;

[0209] 4. User Correction Notice: If the qualification is not passed but there are fields that can be corrected (such as missing materials or incomplete information), the system will automatically identify the required materials and generate a prompt for the user to submit them.

[0210] Finally, based on the above decision results, the server constructs a feedback information package and pushes it to the user terminal. The content includes: processing results (such as "passed", "pending review", "supplementary materials required"); risk description (such as "the score is 0.82, there are frequent operational abnormalities"); correction instructions and operation instructions (such as "please upload a complete income certificate, click here to upload"); review process status and estimated processing time.

[0211] The entire process is coordinated by a central server, with logs recorded for auditing and model optimization, achieving closed-loop management of the entire provident fund business approval chain. This collaborative processing mechanism significantly improves the system's ability to handle high-concurrency, multi-heterogeneous business requests, while simultaneously meeting the dual requirements of regulatory compliance and user experience.

[0212] Second, as Figure 7 As shown, Figure 7 This is a structural diagram of a data fusion processing system based on edge computing provided in an embodiment of the present application. This application proposes a data fusion processing system based on edge computing, including:

[0213] A receiving unit 21 is configured to access multiple data sources at the edge node, wherein the multiple data sources include user service processing data, identity authentication data, and image or document material data;

[0214] A preprocessing unit 22 is configured to perform preprocessing operations on the data from the plurality of data sources to obtain preprocessed data, wherein the preprocessing operations include data format unification, outlier processing, time series alignment, and structured conversion;

[0215] A feature extraction unit 23 is configured to perform a feature extraction operation on the pre-processed data to obtain feature information related to provident fund business, wherein the feature information includes user behavior feature information, identity feature information, and material consistency feature information;

[0216] An acquisition unit 24 is configured to perform a local intelligent judgment operation at the edge node using a local fusion model based on the feature information to obtain a fusion processing result, wherein the local intelligent judgment operation includes extracting preliminary qualification review, loan risk assessment, and abnormal behavior detection;

[0217] The decision unit 25 is used to send the above fusion processing results and their associated tags to the central server for business collaborative decision-making.

[0218] In a feasible implementation, the specific steps of extracting the above-mentioned user behavior feature information include:

[0219] Based on the pre-processed user business processing data, the user behavior sequence is constructed, and the user access path pattern information is extracted by using position embedding coding combined with a sliding window mechanism;

[0220] Based on the above pre-processed user business processing data, a time-aware behavior clustering algorithm is used to extract the user's operating habit information, wherein the above operating habit information includes operation frequency and interaction time distribution;

[0221] The user behavior feature information is determined based on the user access path pattern information and the operation habit information.

[0222] In a feasible implementation manner, the determining of the user behavior characteristic information based on the user access path pattern information and the operation habit information includes:

[0223] Modeling the user access path pattern information as a weighted directed graph, wherein the nodes of the weighted directed graph represent operation pages or function entrances, the edges of the weighted directed graph represent page jumps, and the edge weights of the weighted directed graph represent jump frequencies;

[0224] Calculate structural indicators based on the weighted directed graph, where the structural indicators include node in-degree distribution, path entropy, and average jump length;

[0225] Based on the above operation habit information, the operation rhythm vector is constructed, and the temporal consistency score is obtained by calculating the cosine similarity with the historical rhythm pattern;

[0226] The above structural indicators and the above time consistency scores are used to construct the above user behavior feature information.

[0227] In a feasible implementation, the specific steps of extracting the identity feature information include:

[0228] Based on the pre-processed authentication data, the user authentication fields are standardized and the regional attributes are extracted;

[0229] Extract identity compliance features related to home purchase eligibility based on the pre-processed identity authentication data, where the identity compliance features include whether the applicant is a first-time homebuyer, marital status, and family co-loan suitability;

[0230] Extracting identity stability features related to provident fund contributions based on the pre-processed identity authentication data, wherein the identity stability features include contribution type, contribution history, and abnormal contribution amount information;

[0231] Based on the above-mentioned regional attributes, the above-mentioned identity compliance characteristics and the above-mentioned identity stability characteristics, a logical consistency comparison analysis is performed to generate integrated identity characteristic information.

[0232] In a feasible implementation manner, the specific steps of extracting the above-mentioned material consistency characteristic information include:

[0233] Perform field-level structured recognition on pre-processed image or document data to extract key field information related to provident fund business;

[0234] Based on the above key field information and user-filled information and identity authentication data, a one-to-one field comparison is performed to calculate the field consistency score;

[0235] Cross-check the same fields between multiple documents uploaded in the same business application, build a field coordination comparison matrix, and output a cross-material consistency score;

[0236] Use a lightweight image forgery detection algorithm to analyze the authenticity of image materials and obtain image authenticity scores;

[0237] The above-mentioned field consistency score, the above-mentioned cross-material consistency score and the above-mentioned image authenticity score are integrated to construct the above-mentioned material consistency feature information.

[0238] In a feasible implementation, the above-mentioned local intelligent judgment operation is performed at the edge node using a local fusion model based on the above-mentioned feature information to obtain a fusion processing result, including:

[0239] Get business type information;

[0240] Generating derived characteristic information based on the characteristic information and the business type information;

[0241] Performing a sparse coding operation on the derived feature information and the feature information to obtain edge node feature information;

[0242] The local fusion model is used to perform local intelligent judgment operations on the edge nodes based on the feature information of the edge nodes to obtain the fusion processing results.

[0243] In a feasible embodiment, the above-mentioned local fusion model includes an input layer, a feature transformation layer, a multi-task output structure and an output layer. The above-mentioned feature transformation layer includes a dynamic feature cross-coding submodule, a local attention feature enhancement submodule and a multi-channel feature fusion submodule. The above-mentioned multi-task output structure includes a first output channel, a second output channel and a third output channel. The above-mentioned first output channel is used to perform the preliminary qualification judgment task, the above-mentioned second output channel is used to perform the loan risk scoring task, and the above-mentioned third output channel is used to perform the abnormal behavior detection task.

[0244] In a feasible implementation, the first output channel includes a shallow multi-layer perceptron model, the second output channel includes a lightweight regression network, and the third output channel includes a multi-label classifier.

[0245] In a feasible implementation, the specific steps of the central server performing business collaborative decision-making include:

[0246] Perform field parsing and rule matching on the above fusion processing results to obtain qualification judgment results, risk score values and abnormal behavior labels;

[0247] Execute decision path diversion based on the qualification judgment result, the risk score value, and the abnormal behavior label to obtain a decision structure, wherein the diversion includes any one of automatic flow, manual review triggering, risk warning registration, and user correction notification;

[0248] A feedback information package is generated based on the above decision results and pushed to the user terminal.

[0249] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data fusion processing method based on edge computing, characterized in that: include: Accessing multiple data sources at the edge node, wherein the multiple data sources include user business processing data, identity authentication data, and image or document material data; Performing preprocessing operations on the data from the multiple data sources to obtain preprocessed data, wherein the preprocessing operations include data format unification, outlier processing, time series alignment, and structured conversion; Performing a feature extraction operation on the pre-processed data to obtain feature information related to provident fund business, wherein the feature information includes user behavior feature information, identity feature information, and material consistency feature information; Based on the feature information, a local fusion model is used to perform local intelligent judgment operations at the edge node to obtain a fusion processing result, wherein the local intelligent judgment operations include extracting preliminary qualification review, loan risk assessment, and abnormal behavior detection; Sending the fusion processing results and their associated tags to the central server for business collaborative decision-making; The adopting of a local fusion model based on the feature information to perform a local intelligent judgment operation at the edge node to obtain a fusion processing result includes: Get business type information; generating derived characteristic information according to the characteristic information and the service type information; Performing a sparse coding operation on the derived feature information and the feature information to obtain edge node feature information; The local fusion model is used to perform local intelligent judgment operations on the edge nodes based on the feature information of the edge nodes to obtain the fusion processing results.

2. The data fusion processing method based on edge computing according to claim 1 is characterized in that: The specific steps of extracting the user behavior feature information include: Based on the pre-processed user business processing data, the user behavior sequence is constructed, and the user access path pattern information is extracted by using position embedding coding combined with a sliding window mechanism; Extracting user operation habit information based on the pre-processed user business handling data using a time-aware behavior clustering algorithm, wherein the operation habit information includes operation frequency and interaction time distribution; The user behavior feature information is determined based on the user access path pattern information and the operation habit information.

3. The data fusion processing method based on edge computing according to claim 2 is characterized in that: The determining of the user behavior characteristic information based on the user access path pattern information and the operation habit information includes: Modeling the user access path pattern information as a weighted directed graph, wherein the nodes of the weighted directed graph represent operation pages or function entrances, the edges of the weighted directed graph represent page jumps, and the edge weights of the weighted directed graph represent jump frequencies; Calculating structural indicators based on the weighted directed graph, wherein the structural indicators include node in-degree distribution, path entropy, and average jump length; constructing an operation rhythm vector based on the operation habit information, and obtaining a temporal consistency score by calculating the cosine similarity with the historical rhythm pattern; The user behavior feature information is constructed by combining the structural index and the time consistency score.

4. The data fusion processing method based on edge computing according to claim 1 is characterized in that: The specific steps of extracting the identity feature information include: Based on the pre-processed authentication data, the user authentication fields are standardized and the regional attributes are extracted; Extracting identity compliance features related to home purchase eligibility based on the pre-processed identity authentication data, wherein the identity compliance features include whether the applicant is a first-time homebuyer, marital status, and family co-loan suitability; Extracting identity stability features related to provident fund contributions based on the pre-processed identity authentication data, wherein the identity stability features include contribution type, contribution history, and abnormal contribution amount information; A logical consistency comparison analysis is performed based on the regional attributes, the identity compliance features, and the identity stability features to generate fused identity feature information.

5. The data fusion processing method based on edge computing according to claim 1 is characterized in that: The specific steps of extracting the material consistency feature information include: Perform field-level structured recognition on pre-processed image or document data to extract key field information related to provident fund business; Perform one-to-one field comparison based on the key field information, user-filled information, and identity authentication data, and calculate the field consistency score; Cross-check the same fields between multiple documents uploaded in the same business application, build a field coordination comparison matrix, and output a cross-material consistency score; Use a lightweight image forgery detection algorithm to analyze the authenticity of image materials and obtain image authenticity scores; The field consistency score, the cross-material consistency score, and the image authenticity score are fused to construct the material consistency feature information.

6. The data fusion processing method based on edge computing according to claim 1 is characterized in that: The local fusion model includes an input layer, a feature transformation layer, a multi-task output structure and an output layer. The feature transformation layer includes a dynamic feature cross-coding submodule, a local attention feature enhancement submodule and a multi-channel feature fusion submodule. The multi-task output structure includes a first output channel, a second output channel and a third output channel. The first output channel is used to perform the preliminary qualification judgment task, the second output channel is used to perform the loan risk scoring task, and the third output channel is used to perform the abnormal behavior detection task.

7. The data fusion processing method based on edge computing according to claim 6 is characterized in that: The first output channel includes a shallow multi-layer perceptron model, the second output channel includes a lightweight regression network, and the third output channel includes a multi-label classifier.

8. The data fusion processing method based on edge computing according to claim 7 is characterized in that: The specific steps of the central server performing business collaborative decision-making include: Performing field parsing and rule matching on the fusion processing results to obtain qualification judgment results, risk score values and abnormal behavior labels; Perform decision path diversion based on the qualification judgment result, the risk score value, and the abnormal behavior label to obtain a decision structure, wherein the diversion includes any one of automatic flow, manual review triggering, risk warning registration, and user correction notification; Generate a feedback information package based on the decision result and push it to the user terminal.

9. A data fusion processing system based on edge computing, characterized in that: include: A receiving unit, configured to access multiple data sources at the edge node, wherein the multiple data sources include user service processing data, identity authentication data, and image or document material data; a preprocessing unit, configured to perform a preprocessing operation on the data from the plurality of data sources to obtain preprocessed data, wherein the preprocessing operation includes data format unification, outlier processing, time series alignment, and structured conversion; a feature extraction unit, configured to perform a feature extraction operation on the pre-processed data to obtain feature information related to provident fund business, wherein the feature information includes user behavior feature information, identity feature information, and material consistency feature information; an acquisition unit, configured to perform a local intelligent judgment operation at an edge node using a local fusion model based on the feature information to obtain a fusion processing result, wherein the local intelligent judgment operation includes extracting preliminary qualification review, loan risk assessment, and abnormal behavior detection; A decision-making unit, configured to send the fusion processing result and its associated tags to a central server for business collaborative decision-making; The adopting of a local fusion model based on the feature information to perform a local intelligent judgment operation at the edge node to obtain a fusion processing result includes: Get business type information; generating derived characteristic information according to the characteristic information and the service type information; Performing a sparse coding operation on the derived feature information and the feature information to obtain edge node feature information; The local fusion model is used to perform local intelligent judgment operations on the edge nodes based on the feature information of the edge nodes to obtain the fusion processing results.

Images