A unified enterprise financial file management system and method based on distributed storage

Abstract

The present invention discloses a unified enterprise financial file management system and method based on distributed storage. The present invention relates to the field of enterprise financial file management technology, including: a metadata processing module connected to a distributed storage node cluster, for converting the metadata attributes of the enterprise financial file into high-dimensional sparse vectors. The unified enterprise financial file management system and method based on distributed storage achieves efficient management of multi-dimensional financial file data through dynamic metadata topology reconstruction and dual-channel index optimization mechanism. Vectorized metadata encapsulation technology is used to convert traditional discrete attributes into high-dimensional correlation vectors. Combined with quantum annealing path optimization and three-stage aggregation verification strategy, the response time of multi-condition combination queries is reduced compared to traditional solutions, and network bandwidth consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise financial file management, and in particular to a unified enterprise financial file management system and method based on distributed storage. Background Art

[0002] As enterprises accelerate their digital transformation, financial record management is evolving towards electronic and intensive processes. Distributed storage-based financial record management systems have become the mainstream technology for tackling massive amounts of data and ensuring efficient access. However, existing systems have exposed a key flaw when supporting complex business scenarios: the need for dynamic correlation analysis of multi-dimensional attributes within enterprise financial records, such as time, department, project, and amount, is increasing. For example, in scenarios like tracking cross-year departmental budget execution or rapidly locating high-value contract documents, traditional distributed architectures, relying on a single-dimensional indexing mechanism, cannot effectively support multi-condition combination queries. This problem stems from the disconnect between data storage strategy and query logic. Specifically, the system partitions storage nodes solely by file creation time or basic identifiers, failing to model the multi-dimensional correlation characteristics of business metadata. Consequently, complex queries require traversing raw data across multiple nodes, triggering significant redundant data storage and computation, while also increasing response time due to inter-node communication latency. Furthermore, existing static indexing designs struggle to adapt to the dynamic demands of enterprises as new query dimensions are added. Each change in business rules requires a rebuild of the underlying storage structure, severely impairing system scalability. This contradiction is particularly prominent when the data scale continues to expand. Traditional optimization methods such as increasing the number of replicas or local caching strategies can only alleviate surface performance pressure, but cannot break through the efficiency bottleneck of multi-dimensional data correlation queries, restricting the development of enterprise financial file management towards real-time and intelligent directions. Summary of the Invention

[0003] (1) Technical problems solved

[0004] In response to the deficiencies of the existing technology, the present invention provides a unified enterprise financial file management system and method based on distributed storage, which solves the problem of efficiency optimization in multi-dimensional dynamic association query scenarios.

[0005] (2) Technical solution

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a unified enterprise financial file management system based on distributed storage, comprising:

[0007] The metadata processing module is connected to the distributed storage node cluster and is used to convert the metadata attributes of the enterprise financial files into high-dimensional sparse vectors, and generate a metadata association topology map containing cross-dimensional association strength coefficients through the dynamic dimension weight allocation module; in the implementation of the metadata processing module, the metadata attributes of the enterprise financial files are first structured and parsed and feature encoded, the timestamp is converted into a standardized time series vector, the department code is mapped into a hierarchical feature tree, and the project identifier and amount value are respectively embedded in the continuous vector space through normalization processing.

[0008] The query parsing module is in communication with the metadata processing module and is configured to decompose a multi-condition combination query into a semantic constraint core and a vector derivation domain through a semantic-vector dual-channel parser. During the implementation of the query parsing module, upon receiving a multi-condition combination query request from a user, the semantic-vector dual-channel parser first performs multi-level semantic parsing on the original query statement, including: extracting entity labels and logical relationships in the query conditions through natural language processing technology, classifying attributes with clear enumeration values ​​such as department codes and file types as semantic constraint cores, and generating a precisely matched query statement template; at the same time, mapping continuous or fuzzy association conditions such as amount intervals and time ranges to the vector derivation domain and converting them into hyperplane cutting operation parameters in the vector space.

[0009] The index dynamic optimization module is electrically connected to the metadata processing module and the query parsing module respectively, and includes an active evolution channel and a passive evolution channel, which are used to generate lightweight combined indexes based on reinforcement learning prediction and differential privacy analysis; in the implementation of the index dynamic optimization module, the active evolution channel uses a temporal convolution reinforcement learning model to perform time series analysis on historical query logs, extract periodic correlation features of time, department, and project dimensions, and predict high-frequency dimension combinations within a preset time period in the future: first, the query records in the continuous time window are converted into a three-dimensional feature tensor, and the convolution kernel is used to extract the dimension co-occurrence pattern across time steps. Then, the Q-learning algorithm is combined to calculate the expected query benefit value of different dimension combinations to generate a list of candidate combination indexes.

[0010] The distributed aggregation module is connected to the distributed storage node cluster through the data bus, and adopts a three-stage data aggregation strategy to perform phased screening and verification of cross-node computing flows. During the implementation of the distributed aggregation module, when a cross-node query request enters the execution phase, the node pre-aggregation process is first started: each storage node uses a dynamic similarity matching algorithm based on the locally stored vectorized metadata to perform preliminary screening of the query conditions, calculate the confidence score of each local data item, and only return a subset of results with a score higher than the first dynamic threshold. The compressed result set is transmitted to the central scheduler through the data bus, where the result set contains the data identifier, confidence score and dimension matching details.

[0011] The metadata processing module, query parsing module, index dynamic optimization module and distributed aggregation module are sequentially connected in series to form a closed-loop feedback link.

[0012] Preferably, the metadata processing module includes:

[0013] The dynamic dimension weight assignment module assigns a dynamic weight coefficient to each dimension of the metadata vector by analyzing the co-occurrence frequency of the time, department, project, and amount dimensions in historical query logs;

[0014] The incremental adversarial training submodule uses a generative adversarial network architecture to generate simulated query requests and conduct adversarial training on actual query patterns, outputting an optimized metadata topology distribution strategy.

[0015] The metadata association topology map is stored in a distributed hash table and includes a physical proximity mapping table for recording the physical node position relationship corresponding to the metadata vectors with high association.

[0016] Preferably, the query parsing module further includes a cross-layer path optimization engine. The implementation of the cross-layer path optimization engine is based on a vector-physical dual mapping verification mechanism, and the implementation process includes:

[0017] Based on the vector space characteristics of the metadata association topology graph, the vector boundaries of the candidate data clusters are located using the back-propagation gradient detection algorithm;

[0018] Combined with the spatial distribution topology of physical storage nodes, a quantum annealing simulator is used to calculate the minimum cross-node hop path;

[0019] The operating parameters of the quantum annealing simulator include the network delay weight factor between physical nodes, the storage node load balancing attenuation coefficient, and the thermodynamic simulation temperature parameter of the cross-dimensional query condition.

[0020] Preferably, in the index dynamic optimization module:

[0021] The active evolution channel uses a temporal convolutional reinforcement learning model to analyze the time series characteristics of query patterns to predict high-probability dimension combinations within a preset time period in the future;

[0022] When the query response time of the passive evolution channel exceeds the dynamic threshold, it triggers k-anonymization processing based on differential privacy protection, performs desensitization analysis on the query log and generates adversarial training samples;

[0023] The k-anonymization process includes generalizing the sensitive dimension values, ensuring that each set of query conditions contains at least k indistinguishable subject features, and injecting Laplace noise during the adversarial training sample generation process.

[0024] Preferably, when the distributed aggregation module executes the three-stage data aggregation strategy:

[0025] In the node pre-aggregation phase, each storage node calculates a confidence score based on the local vectorized metadata and returns a subset of results with a score above a first threshold;

[0026] In the path reweighted aggregation phase, the central scheduler uses an adaptive ant colony algorithm to dynamically adjust the node access priority of the secondary query based on the confidence distribution of the node return results;

[0027] In the final precise aggregation phase, full data extraction is performed only on physical storage nodes with confidence scores higher than the second threshold, and data integrity is verified through a hash check chain.

[0028] Preferably, the three-stage data aggregation strategy further includes:

[0029] The logical consistency checking unit establishes the reverse mapping relationship between query conditions and result sets, and verifies whether the returned data fully satisfies all query constraints;

[0030] The spatiotemporal validity marking unit adds the timestamp signature of the data version and the geographic location code of the storage node to the query results to ensure compliance with timeliness and regional compliance requirements.

[0031] Preferably, the dynamic evolution process of the metadata association topology graph includes:

[0032] Record the incremental update trajectory of the topology graph as the query pattern changes;

[0033] When a new query dimension is added, the weight distribution of the prediction index of the active evolution channel and the correction index of the passive evolution channel is balanced through the game equilibrium controller;

[0034] The game balance controller uses the Nash equilibrium algorithm to calculate the index weight, and double-checks the validity of the newly created index through vector space coverage verification and physical storage load balancing verification.

[0035] Preferably, the implementation steps of the vector-physical dual mapping verification mechanism are:

[0036] The first verification is based on the vector space similarity threshold to filter candidate paths. The screening threshold is a dynamically adjusted value and is calculated using the dynamic threshold calculation formula:

[0037] Threshold = α (current node load / system average load) + β (number of query dimensions / maximum number of supported dimensions). Where Threshold is the dynamic threshold of response time that triggers the passive evolution channel, and α and β are dynamic adjustment coefficients.

[0038] Secondary verification eliminates paths whose delay exceeds the set threshold or whose bandwidth margin is lower than the safe value based on the physical connection delay and bandwidth margin of the actual node.

[0039] Preferably, the pheromone update rule of the adaptive ant colony algorithm satisfies the pheromone update formula:

[0040] ;

[0041] Where ρ is the pheromone volatility coefficient, : the pheromone increment of the kth ant on the path i→j, τ ij (t+1): pheromone concentration on the path from node i to node j at time step t+1, t: time step number, representing the number of algorithm iterations; i: node number of the starting point of the path, j: node number of the end point of the path, m: total number of ants, which is the number of computing agents in the path optimization, k: number of the ant currently computing, k∈[1,m];

[0042] The hash check chain adopts a Merkle tree structure, the leaf nodes are the data hash values ​​of each storage node, and the root hash value is synchronized to the central scheduler in real time.

[0043] A method for unified management of enterprise financial files based on the system comprises the following steps:

[0044] S1: Vectorize and encapsulate financial files through a self-organizing metadata decomposition algorithm to generate a dynamically evolving metadata association topology graph;

[0045] S2: Decompose query conditions using a semantic-vector dual-channel parser, and deduce the optimal query path by combining back-propagation gradient detection with quantum annealing simulation;

[0046] S3: Performs three-stage screening: node pre-aggregation, path reweighted aggregation, and final accurate aggregation, and outputs the results through a double verification mechanism;

[0047] S4: When a new query dimension is added, the dual-channel index evolution mechanism is triggered to generate a double-verified combined index and update the metadata topology.

[0048] (3) Beneficial effects

[0049] The present invention provides a unified enterprise financial file management system and method based on distributed storage. It has the following beneficial effects:

[0050] (1) This unified enterprise financial archive management system and method based on distributed storage achieves efficient management of multi-dimensional financial archive data through dynamic metadata topology reconstruction and dual-channel index optimization mechanisms. It uses vectorized metadata encapsulation technology to transform traditional discrete attributes into high-dimensional correlation vectors. Combined with quantum annealing path optimization and a three-stage aggregation verification strategy, the response time for multi-condition combination queries is reduced compared to traditional solutions, while also reducing network bandwidth consumption. Through a dynamic adversarial training model and an incremental index evolution mechanism, the system automatically completes index reconstruction when new query dimensions are added. This shortens the reconstruction time compared to traditional methods, while maintaining node load balancing within a reasonable fluctuation range, significantly improving the system's elastic scalability.

[0051] (2) This unified enterprise financial record management system and method based on distributed storage breaks through the bottleneck of real-time analysis and deep correlation mining in enterprise financial record management, supports millisecond-level multi-dimensional joint queries, and has high query accuracy. Through blockchain-style spatiotemporal validity marking and differential privacy protection mechanisms, it ensures the auditability and compliance of data throughout its life cycle, meeting data regulatory requirements such as GDPR. The system's adaptive dynamic evolution characteristics enhance the intelligence level of the financial record management process, providing a safe and reliable data governance infrastructure for enterprise digital transformation. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the overall framework of the present invention;

[0053] Figure 2 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] See also Figure 1 and Figure 2 The present invention provides a technical solution: a unified enterprise financial file management system based on distributed storage, comprising:

[0056] The metadata processing module, connected to the distributed storage node cluster, converts the metadata attributes of enterprise financial records into high-dimensional sparse vectors. The module then generates a metadata association topology map containing cross-dimensional association strength coefficients through a dynamic dimension weighting module. The metadata processing module first performs structured parsing and feature encoding on the metadata attributes of enterprise financial records. Timestamps are converted into standardized time series vectors, department codes are mapped into a hierarchical feature tree, and project identifiers and amount values ​​are normalized and embedded into a continuous vector space. The dynamic dimension weighting module assigns dynamic weights to each dimension of each metadata vector based on the co-occurrence frequency of time, department, project, and amount dimensions in historical query logs. This module automatically increases the weight of frequently co-occurring dimension combinations and performs sparse compression on less frequently associated dimensions. When generating the metadata association topology map, a multidimensional association matrix is ​​constructed based on the cross-dimensional association strength coefficients. The association strength coefficients are calculated by combining the co-occurrence probability between dimensions with the statistical weights of query condition combinations. A sliding window mechanism is used to dynamically weight the dimension combinations from the last N queries and record these values ​​in the physical proximity map of the distributed hash table.

[0057] The metadata processing module further optimizes the storage distribution strategy through an incremental adversarial training model. It uses a generative adversarial network to generate simulated query requests, which are then used together with real query logs to train the storage node distribution model. This allows metadata vectors with high correlation strength to be automatically clustered to adjacent node regions at the physical storage layer, while also maintaining incremental updates to the dynamically evolving log topology. The metadata correlation topology is stored in the distributed storage node cluster as key-value pairs, where the key is the feature hash value of the dimension combination, and the value is a composite data structure containing the physical node location, correlation strength coefficient, and version timestamp. This ensures rapid retrieval and dynamic updates of multi-dimensional relationships.

[0058] The query parsing module is in communication with the metadata processing module and is configured to decompose a multi-condition combination query into a semantic constraint core and a vector derivation domain through a semantic-vector dual-channel parser. During the implementation of the query parsing module, upon receiving a multi-condition combination query request from a user, the semantic-vector dual-channel parser first performs multi-level semantic parsing on the original query statement, including: extracting entity labels and logical relationships in the query conditions through natural language processing technology, classifying attributes with clear enumeration values ​​such as department codes and file types as semantic constraint cores, and generating a precisely matched query statement template; at the same time, mapping continuous or fuzzy association conditions such as amount intervals and time ranges to the vector derivation domain and converting them into hyperplane cutting operation parameters in the vector space.

[0059] During the conversion process of the vector-derived domain, the parser dynamically adjusts the normalization coefficients and weight distributions of each dimension based on the metadata association topology generated by the metadata processing module. For example, the monetary value is linearly scaled to the maximum-minimum range of the current enterprise financial file and then embedded in the high-dimensional space. Gaussian weights are then added based on the distribution characteristics of the monetary dimension in historical queries. During the dual-channel collaborative verification phase, the exact matching conditions generated by the semantic constraint core and the hyperplane parameters of the vector-derived domain are checked for logical consistency through a cross-validation mechanism. If the department code in the semantic constraint core and the monetary range in the vector-derived domain do not overlap in historical data, an abnormal condition warning is triggered and user confirmation is requested. The parsed semantic constraint core and vector-derived domain are transmitted to the cross-layer path optimization engine via a standardized interface. The semantic constraint core is used to generate preliminary filtering conditions for distributed storage nodes, while the vector-derived domain drives the back-propagation gradient detection algorithm to locate the vector boundaries of candidate data clusters in the metadata association topology. Simultaneously, a quantum annealing simulator is used to optimize the network topology of the physical storage nodes, ultimately outputting an optimal query path set that meets the dual verification conditions.

[0060] The index dynamic optimization module is electrically connected to the metadata processing module and the query parsing module respectively, and includes an active evolution channel and a passive evolution channel, which are used to generate lightweight combined indexes based on reinforcement learning prediction and differential privacy analysis; in the implementation of the index dynamic optimization module, the active evolution channel uses a temporal convolution reinforcement learning model to perform time series analysis on historical query logs, extract periodic correlation features of time, department, and project dimensions, and predict high-frequency dimension combinations within a preset time period in the future: first, the query records in the continuous time window are converted into a three-dimensional feature tensor, and the convolution kernel is used to extract the dimension co-occurrence pattern across time steps. Then, the Q-learning algorithm is combined to calculate the expected query benefit value of different dimension combinations to generate a list of candidate combination indexes.

[0061] The passive evolution channel monitors query response time in real time. When a dynamic threshold is exceeded, a differential privacy mechanism is triggered to process the original query logs. The k-anonymization algorithm is used to generalize sensitive dimension values, ensuring that each query condition contains at least k indistinguishable subject features. Frequent itemset mining is then performed on the anonymized logs to generate adversarial training samples, which are then fed into a generative adversarial network. A game equilibrium controller coordinates the outputs of the active and passive channels using a Nash equilibrium algorithm to calculate the weighted distribution of the predicted and corrected indexes. Specifically, a dual-matrix game model is employed, using the active channel's prediction accuracy and the passive channel's corrected coverage as parameters of the payoff function to output the optimal weighted combination. New indexes undergo a dual verification mechanism: vector space coverage verification calculates the coverage radius of the index dimension in the metadata association topology graph and eliminates redundant indexes whose coverage radius exceeds a set threshold. Physical storage load balancing verification simulates the node access pressure distribution after index deployment. Any node load discrepancy exceeding 20% ​​triggers an index rebuild. The final generated lightweight combined index is stored in the dynamic evolution log of the metadata association topology graph in bitmap compression format. Each index entry contains the dimension combination hash value, effective time range and version identifier, and is synchronized to each storage node through the heartbeat mechanism.

[0062] The distributed aggregation module is connected to the distributed storage node cluster through the data bus, and adopts a three-stage data aggregation strategy to perform phased screening and verification of cross-node computing flows. During the implementation of the distributed aggregation module, when a cross-node query request enters the execution phase, the node pre-aggregation process is first started: each storage node uses a dynamic similarity matching algorithm based on the locally stored vectorized metadata to perform preliminary screening of the query conditions, calculate the confidence score of each local data item, and only return a subset of results with a score higher than the first dynamic threshold. The compressed result set is transmitted to the central scheduler through the data bus, where the result set contains the data identifier, confidence score and dimension matching details.

[0063] The confidence score formula is:

[0064] , where Score is the comprehensive score of the matching degree between the data item and the query condition. The higher the value, the higher the matching degree. o : The dynamic weight coefficient of the oth dimension, the weight distribution result from the metadata processing module, the query dimension o : The value range of the oth dimension in the query condition, data dimension o : The actual value of the oth dimension in the data item, n: The total number of dimensions, o: The dimension number currently being calculated.

[0065] Then it enters the path reweighted aggregation stage: the central scheduler uses a load-aware ant colony optimization algorithm to dynamically adjust the secondary query path based on the distribution characteristics of the received node results, uses the confidence score as the pheromone concentration indicator, and combines the real-time load rate of the storage node to generate a node access priority queue and send precise query instructions to high-priority nodes.

[0066] The real-time load rate is calculated using the node priority formula:

[0067] , where, Priority: the access priority value of the storage node. The larger the value, the higher the priority. It is used for path optimization decision. γ: the influence coefficient of CPU utilization, which is 0.6. δ: the influence coefficient of bandwidth margin, which is 0.4.

[0068] The final precise aggregation stage is executed on the selected nodes: only data items with confidence scores higher than the second threshold are fully extracted. During the extraction process, streaming hash verification technology is used to aggregate the hash values ​​of the data blocks step by step according to the Merkle tree structure to generate a root hash value, which is compared and verified with the original hash value recorded in the metadata association topology map. The intermediate results generated in each stage must pass a double verification mechanism: the logical consistency verification unit establishes a reverse mapping relationship between the query conditions and the result set to verify whether the returned data fully meets all constraints; the spatiotemporal validity marking unit adds a timestamp signature of the data version and the geographic location code of the storage node to the final result to ensure that the result meets the timeliness requirements and regional compliance policies. The entire aggregation process adopts a pipeline parallel processing mechanism. The path optimization calculation is started when the pre-aggregation stage is not completed, and the data processing rate of each stage is dynamically adjusted through the back pressure mechanism to avoid node overload.

[0069] The metadata processing module, query parsing module, index dynamic optimization module, and distributed aggregation module are connected in series, forming a closed-loop feedback loop. During the operation of this closed-loop feedback loop, the metadata association topology generated by the metadata processing module is transmitted via a data bus to the query parsing module, triggering the semantic-vector dual-channel parser to decompose the query conditions. The parsed semantic constraint kernel and vector derivation domain are simultaneously input into the index dynamic optimization module, driving the active evolution channel to generate predictive indexes and monitoring abnormal query patterns in the passive evolution channel in real time. The distributed aggregation module generates node load data, query path efficiency metrics, and verification results during the three-stage aggregation process, which are then transmitted back to the incremental adversarial training model in the metadata processing module via a feedback bus. The deviation between the actual query path and the theoretical optimal path is converted into an adversarial training loss function, dynamically adjusting the mapping between dimension weight coefficients and physical proximity in the metadata topology. The game balance controller in the index dynamic optimization module also receives the index verification results from the aggregation module, modifies the reward function parameters of the reinforcement learning model through backpropagation, and injects the adjusted index weight distribution strategy into the cross-layer path optimization engine of the query parsing module. The index verification results include vector space coverage error and physical load balance.

[0070] A dual-verification mechanism is implemented at key nodes in the closed-loop chain. This means that after each topology update, a cross-module consistency check is triggered to compare the metadata processing module's vector space mapping with the distributed aggregation module's actual storage location. If the deviation exceeds the fault tolerance threshold, a topology rollback mechanism is initiated, and adversarial training is re-executed to generate a corrected version. Feedback data streams are managed using a priority queue, marking core metrics such as query response latency, index hit rate, and aggregation verification failure rate as high-priority feedback signals, ensuring that the system completes closed-loop adjustments to key parameters within 10 milliseconds.

[0071] The metadata processing module includes:

[0072] The dynamic dimension weight assignment module assigns a dynamic weight coefficient to each dimension of the metadata vector by analyzing the co-occurrence frequency of the time, department, project, and amount dimensions in historical query logs;

[0073] The incremental adversarial training submodule uses a generative adversarial network architecture to generate simulated query requests and conduct adversarial training on actual query patterns, outputting an optimized metadata topology distribution strategy.

[0074] The metadata association topology is stored in a distributed hash table and includes a physical proximity mapping table for recording the physical node position relationship corresponding to the metadata vectors with high association.

[0075] It should be further explained that, in the specific implementation process, in the implementation process of the unified management system of enterprise financial files, the metadata processing module conducts multi-dimensional correlation analysis on the historical query logs through the dynamic dimension weight allocation module: first, the co-occurrence frequency data of the time, department, project, and amount dimensions are extracted, and a dimension correlation matrix is ​​constructed. The sliding window mechanism is used to perform dynamic weighted statistics on the dimension combinations in the last N queries, and a dynamic weight coefficient is assigned to the dimension of each metadata vector. Among them, the weight value of the high-frequency co-occurrence dimension combination is automatically increased through the exponential decay function, among which the high-frequency co-occurrence dimension combination includes department-time and project-amount;

[0076] The weight value of the high-frequency co-occurrence dimension combination is automatically increased by the exponential decay function and the dynamic weight calculation formula is calculated. The dynamic weight calculation formula is as follows:

[0077] , where w o : The normalized weight value of the oth dimension combination, n: The total number of optional dimension combinations, f o is the co-occurrence frequency of the dimension combination, λ is the dynamic adjustment factor, o is the dimension number currently calculated, and p is the temporary dimension variable during the summation process.

[0078] The incremental adversarial training submodule adopts a generative adversarial network architecture. The generator synthesizes simulated query requests based on historical query patterns. The simulated query requests include forged department codes, time ranges, and amount ranges. The discriminator combines real query logs with metadata association topology maps for adversarial training. The storage node distribution strategy is optimized through gradient backpropagation, so that metadata vectors with high correlation strength form spatial proximity clusters in the physical storage layer.

[0079] The incremental adversarial training submodule stores the generated metadata association topology map in a distributed hash table. The physical proximity mapping table adopts a four-layer nested structure, including: the first layer is the dimension combination hash value, the second layer records the association strength coefficient calculated by jointly calculating the conditional probability between dimensions and the query weight, the third layer stores the IP address and port number of the target physical node, and the fourth layer maintains version control information, which includes the creation timestamp and the last update timestamp.

[0080] When executing a query, the system prioritizes accessing node clusters in the physical proximity mapping table whose correlation strength coefficient exceeds a preset threshold of 0.7. It synchronizes node status data every 5 seconds through a heartbeat mechanism. If a node is detected to be offline or its load exceeds the warning value, the dynamic expansion mechanism of the neighboring node group is triggered, migrating copies of the associated data to the backup node group. The incremental update mechanism of the incremental adversarial training submodule uses write-time copy technology, generating a new version branch when updating the metadata topology map. After passing double verification, the new version is marked as effective, and the old version is retained in the three most recent historical version pools for rollback. This double verification includes vector space consistency verification and physical storage stress testing.

[0081] The query parsing module also includes a cross-layer path optimization engine. The implementation of the cross-layer path optimization engine is based on a vector-physical dual mapping verification mechanism. The implementation process includes:

[0082] Based on the vector space characteristics of the metadata association topology graph, the vector boundaries of the candidate data clusters are located using the back-propagation gradient detection algorithm;

[0083] Combined with the spatial distribution topology of physical storage nodes, a quantum annealing simulator is used to calculate the minimum cross-node hop path;

[0084] The operating parameters of the quantum annealing simulator include the network delay weight factor between physical nodes, the storage node load balancing attenuation coefficient, and the thermodynamic simulation temperature parameter of the cross-dimensional query condition.

[0085] It should be further explained that, in the specific implementation process, in the implementation process of the cross-layer path optimization engine, after receiving the semantic constraint kernel and vector derivation domain transmitted by the query parsing module, the back-propagation gradient detection algorithm first locates the candidate data cluster based on the vector space characteristics of the metadata association topology graph, including: inputting the hyperplane parameters of the vector derivation domain into the pre-trained multi-layer perceptron model, calculating the projected gradient value of each metadata vector in the direction of the hyperplane normal vector, and dynamically adjusting the gradient descent step size to determine the vector boundary range of the candidate data cluster, wherein the boundary threshold is dynamically adjusted according to the dynamic boundary adjustment formula based on the average query load rate of the current system; the dynamic boundary adjustment formula is as follows:

[0086]

[0087] Among them, Boundary adj : The candidate data cluster boundary threshold is dynamically adjusted according to the number of concurrent queries. The baseline boundary value is the default boundary threshold set by system initialization.

[0088] The quantum annealing simulator simultaneously starts the physical path optimization calculation, abstracting the network topology of the storage nodes into the Ising model. Specifically, the network delay weight factor between nodes is converted into the spin coupling strength, the storage node load balancing attenuation coefficient is mapped to the external magnetic field parameter, and the thermodynamic simulation temperature parameter of the cross-dimensional query condition is calculated according to the number of query dimensions according to the following formula:

[0089] Dynamic settings are performed to find the node access path with the lowest energy through a simulated annealing process; where T0 is the basic temperature constant and T is the thermodynamic temperature parameter in the quantum annealing simulation, which is used to control the randomness of the optimization process.

[0090] The vector-physical dual mapping verification mechanism is executed in two stages. The process includes: the first verification stage calculates the spatial similarity of the metadata vectors in the candidate path, using the dynamic threshold calculation formula in the dynamic threshold screening algorithm. The specific formula is:

[0091] Threshold = α (current node load / system average load) + β (number of query dimensions / maximum number of supported dimensions). Where Threshold is the dynamic threshold of response time that triggers the passive evolution channel, and α and β are dynamic adjustment coefficients.

[0092] The secondary verification phase, based on real-time status data from physical nodes, eliminates candidate paths whose network latency exceeds a set threshold of 200ms or whose current bandwidth margin is below a safe level. A path scoring weighting algorithm is then used to generate a final optimized path queue, with the safe level preset at 30% of the total bandwidth. After the path is determined, the engine triggers an incremental update of the metadata-associated topology graph and writes the characteristic parameters of the query path into the dynamic evolution log, which serves as the input feature vector for the subsequent adversarial training model. The characteristic parameters include dimension combination pattern, number of node hops, and response latency.

[0093] In the index dynamic optimization module:

[0094] The active evolution channel uses a temporal convolutional reinforcement learning model to analyze the time series characteristics of query patterns to predict high-probability dimension combinations within a preset time period in the future;

[0095] When the query response time of the passive evolution channel exceeds the dynamic threshold, it triggers k-anonymization processing based on differential privacy protection, performs desensitization analysis on the query log and generates adversarial training samples;

[0096] The k-anonymization process involves generalizing the values ​​of sensitive dimensions, ensuring that each set of query conditions contains at least k indistinguishable subject features, and injecting Laplace noise during the generation of adversarial training samples.

[0097] It should be further explained that, in the specific implementation process, during the implementation of the index dynamic optimization module, the active evolution channel uses a temporal convolution reinforcement learning model to extract features from query records within a continuous time window. This includes: discretizing timestamps, department levels, and project categories according to a preset time granularity to construct a three-dimensional feature tensor, capturing dimensional correlation patterns across time steps through dilated convolution kernels, and using an attention mechanism to strengthen the feature weights of high-frequency dimension combinations. This outputs a high-probability dimension combination prediction list for the next two hours, where the preset time granularity is 15 minutes.

[0098] The prediction results are evaluated using a Q-learning algorithm, using query response time reduction and index hit rate as reward function parameters. The expected returns for different dimensional combinations are calculated, and a candidate combination index bitmap is generated and preloaded into a memory buffer. The passive evolution channel features a dynamic trigger mechanism. When the real-time monitored query response time exceeds a dynamic threshold, the differential privacy protection process is immediately initiated.

[0099] The process of initiating the differential privacy protection process includes: performing interval generalization processing on the amount range and department code in the original query log, including converting the precise amount value into an interval of 100,000 yuan, and retaining the first three levels of the tree structure of the department code, ensuring that each set of query conditions contains at least k indistinguishable feature entities after generalization; then using the improved FP-Growth algorithm to mine frequent item sets on the desensitized log, extract abnormal dimension combination patterns and inject them into the generative adversarial network to generate an adversarial training sample set.

[0100] The game equilibrium controller adopts a dual-matrix game model for weight distribution, takes the prediction accuracy of the active channel and the correction coverage of the passive channel as the profit matrix parameters, calculates the optimal weight ratio through the Nash equilibrium point, and dynamically adjusts the effective weight value of each dimension combination in the index bitmap.

[0101] The prediction accuracy calculation formula is:

[0102] , where A p : The hit ratio of the prediction index generated by the active evolution channel in the actual query, which measures the effectiveness of the active prediction mechanism. The higher the value, the better the adaptability of the prediction model to the business scenario, which guides the reward calculation of the reinforcement learning model. p The higher the value, the greater the reward value obtained by the model; the trigger index preloading strategy is: when A p When ≥0.7, the predicted index is loaded into the memory cache in advance.

[0103] The formula for calculating the modified coverage is:

[0104] , where C r: The coverage ratio of the modified index generated by the passive evolution channel to abnormal queries reflects the system's dynamic adaptability to unforeseen query patterns and drives the generation frequency of adversarial training samples. r The lower it is, the stronger the strength of the generated adversarial samples is. The index reconstruction threshold is controlled as follows: when Cr≤0.6, the full index reconstruction process is triggered.

[0105] New indexes undergo two stages of verification: a vector space coverage verification phase, in which the candidate index's coverage radius within the metadata topology is calculated. If the coverage radius exceeds 1.5 times the average of similar historical indexes, an index reconstruction is triggered. A physical storage load balancing verification phase uses a discrete event simulator to predict the access pressure distribution across nodes after index deployment. When the load difference between nodes exceeds 20%, virtual shards are automatically added for load balancing compensation. The final, effective index entries are bound to the metadata topology via a versioning protocol. Each entry contains a valid time window, a dimension combination fingerprint, and a checksum. These entries are then synchronized to the local index cache of each storage node using an incremental update mechanism.

[0106] The calculation formula for coverage radius is:

[0107] , where Δd o : standard deviation of the oth dimension, w o : Dimension weight coefficient, R: coverage radius, n: total number of optional dimension combinations, o: dimension number currently being calculated.

[0108] When the distributed aggregation module implements the three-stage data aggregation strategy:

[0109] In the node pre-aggregation phase, each storage node calculates a confidence score based on the local vectorized metadata and returns a subset of results with a score above a first threshold;

[0110] In the path reweighted aggregation phase, the central scheduler uses an adaptive ant colony algorithm to dynamically adjust the node access priority of the secondary query based on the confidence distribution of the node return results;

[0111] In the final precise aggregation phase, full data extraction is performed only on physical storage nodes with confidence scores higher than the second threshold, and data integrity is verified through a hash check chain.

[0112] It should be further explained that in the specific implementation process, when the distributed aggregation module executes the three-stage data aggregation strategy, the node pre-aggregation phase first initiates the vectorized metadata matching calculation locally on each storage node. This includes: based on the semantic constraint core and vector derivation domain output by the query parsing module, a dynamic sliding window algorithm is used to perform similarity matching on the locally stored metadata vectors;

[0113] Each node only returns the subset of results with scores higher than the dynamic first threshold, and transmits the compressed data fingerprint and score matrix through the data bus. The dynamic first threshold calculation formula is:

[0114] , where T1 is the dynamic first threshold, μ is the mean of historical scores, and σ is the variance coefficient.

[0115] During the path reweighting aggregation phase, the central scheduler uses a load-aware ant colony algorithm to optimize the query path. This involves using the average confidence score returned by the node as the pheromone concentration benchmark, combining it with the real-time load indicators of the storage nodes, building a node access priority queue, and injecting the priority weights into the quantum annealing simulator through a backpropagation mechanism for path optimization verification. The real-time load indicator calculation formula is:

[0116] , where Node priority : The access priority quantification value of the node in the current query path optimization, which is used to dynamically adjust the query path selection. The larger the value, the node should be accessed first. CPU usage : The current CPU resource usage percentage of the node, Score avg : The average confidence score of the results returned by the node in the pre-aggregation stage, Bandwidth free : The network bandwidth capacity currently unused by the node, Bandwidth total : The maximum bandwidth capacity of the node's physical network interface, γ, δ: dynamic adjustment factors, γ=0.6, δ=0.4, used to balance the impact of CPU and bandwidth on priority.

[0117] In the final precise aggregation stage, a complete data extraction request is initiated for the top N nodes in the priority queue. During the data transmission process, streaming block verification technology is used to generate a hierarchical hash value for each data block according to the Merkle tree structure. The leaf node is the data block hash, and the parent node is generated by the cascade calculation of the child node hash. The final root hash value is compared and verified with the pre-stored benchmark hash in the metadata association topology map. If three consecutive data blocks fail to be verified, the node data repair process is triggered.

[0118] The intermediate results generated at each aggregation stage are processed through a dual verification mechanism. These include: a logical consistency check unit establishes a reverse mapping index for query conditions and employs a Bloom filter to rapidly verify that the returned data fully satisfies all constraints; a spatiotemporal validity tagging unit adds a blockchain-style timestamp and geocoding identifier to the final result, and performs cross-version consistency verification using the version number and metadata topology evolution log. The aggregation process utilizes a pipelined parallel architecture, initiating path optimization calculations when a node pre-aggregates 80% of the data. A backpressure feedback mechanism dynamically adjusts the data transmission rate of each node, automatically reducing the query priority of a node if its response latency exceeds a threshold.

[0119] The three-stage data aggregation strategy also includes:

[0120] The logical consistency checking unit establishes the reverse mapping relationship between query conditions and result sets, and verifies whether the returned data fully satisfies all query constraints;

[0121] The spatiotemporal validity marking unit adds the timestamp signature of the data version and the geographic location code of the storage node to the query results to ensure compliance with timeliness and regional compliance requirements.

[0122] It should be further explained that, in the specific implementation process, during the logical consistency check and spatiotemporal validity marking implementation of the three-stage data aggregation strategy, the logical consistency check unit initiates the reverse mapping verification after the node pre-aggregation is completed. This includes: first, establishing a bidirectional index relationship between the query conditions and the result data, converting the time range, amount range, department code and other constraints in the original query into a Bloom filter bitmap structure, and performing a fast matching test on the returned result set. For amount range verification, the system extracts the amount field of all records in the result set and performs an interval inclusion check using a range tree structure. If an abnormal value outside the query range is detected, the data repair process is triggered. The data repair process includes: automatically initiating a data re-verification request to the corresponding storage node and temporarily freezing the query permission of the node until the repair is complete.

[0123] The spatiotemporal validity marking unit performs multi-dimensional verification in the final aggregation stage, including: the timestamp signature adopts a blockchain-style chain hash structure, and the timestamp of each result set is cascaded with the hash value of the previous block to form an unalterable time series chain; the geographic location coding converts the latitude and longitude coordinates of the storage node into a 12-bit string code based on the GeoHash algorithm, and matches it with the regional compliance library in the enterprise financial file management strategy. If it is detected that the data storage location does not comply with the preset compliance area, the data desensitization reprocessing process is automatically triggered. Among them, the cascade calculation formula is:

[0124] Among them, Hashcurrent : The encrypted hash value generated by the current data block, SHA256: Secure Hash Algorithm, Hash prev : The hash value of the previous data block in the chain structure, Timestamp: The precise time record of the data operation, Data fingerprint : The summary identifier of the current data block.

[0125] The version consistency check compares the result set version number with the dynamic evolution log version sequence of the metadata-associated topology map, and uses a binary search method to locate the nearest valid version interval. If the current result version differs from the latest version of the topology map by more than three iteration cycles, the asynchronous data synchronization mechanism is started. Abnormal data found during the verification process will generate a diagnostic report containing the error type, node location, and dimension deviation value, which will be transmitted back to the adversarial training model of the metadata processing module in real time through a closed-loop feedback link to drive the dynamic optimization of the storage distribution strategy. A double-layer encryption mechanism is used for spatiotemporal tag data during transmission, including: the timestamp signature is protected by an asymmetric encryption algorithm, and the geographic location code is encrypted by a lightweight stream cipher to ensure the security of the tag information on the transmission link. The result set that finally passes the verification will be attached with a metadata header containing the verification pass identifier, version number, and spatiotemporal stamp, and will be encapsulated in a compressed binary format and returned to the user end.

[0126] The dynamic evolution process of the metadata association topology graph includes:

[0127] Record the incremental update trajectory of the topology graph as the query pattern changes;

[0128] When a new query dimension is added, the weight distribution of the prediction index of the active evolution channel and the correction index of the passive evolution channel is balanced through the game equilibrium controller;

[0129] The game equilibrium controller uses the Nash equilibrium algorithm to calculate the index weight, and double-checks the validity of the newly created index through vector space coverage verification and physical storage load balancing verification.

[0130] It should be further explained that in the specific implementation process, during the dynamic evolution of the metadata association topology, when a new query dimension is added or the business rules are changed, the system first triggers the dimension compatibility check, including: calculating the cosine similarity between the new dimension and the existing dimension set. If the similarity is lower than the preset threshold of 0.4, the topology branch evolution mechanism is activated. The cosine similarity calculation formula is:

[0131] , where o: the dimension number of the current calculation, n: the total number of dimensions, w o old : The weight value of the oth dimension in the old dimension set, w o new: The weight value of the oth dimension after the new dimension is added.

[0132] The game equilibrium controller performs dual-channel index weight balancing in this process, including: inputting the dimension combination index predicted by the active evolution channel and the modified index generated by the passive evolution channel into a two-layer decision network. The first layer uses the gradient boosting decision tree to calculate the expected query benefits of each index. The second layer solves the optimal weight distribution matrix through the Nash equilibrium algorithm and outputs the index fusion strategy. The new index needs to go through a double verification process, which includes: the vector space coverage verification stage, by calculating the coverage density of the candidate index in the topology map, if the density value is lower than 80% of the historical benchmark value, the index reconstruction is triggered; the physical storage load balancing verification stage, using a discrete event simulator to predict the node access pressure distribution after the index is deployed, and automatically injecting virtual shards for load compensation when the load difference between nodes exceeds 20%. Among them, the coverage density calculation formula is:

[0133] ; Among them, Density: the coverage density of the candidate index in the topology graph, d: the dimension identifier of the current calculation, Dimensions: the dimension set supported by the system, the number of index hits d : The number of queries that successfully matched the index on dimension d, the total number of queries d : The total number of queries that occurred on dimension d.

[0134] The dynamic evolution log adopts an incremental recording method. Each time the topology is updated, a triple log item is generated, which includes the change type, the vector space mapping relationship before and after the change, and the physical storage location offset. It forms a chain structure with the historical version through the version control protocol, including the addition of dimensions, weight adjustment, and node migration.

[0135] Abnormal states generated during the evolution process will trigger a rollback mechanism, including: reverse parsing the three most recent valid change records based on the evolution log, recalculating dimension weights and verifying physical distribution until the system returns to a stable state; among them, abnormal states include sudden changes in node load or vector space fractures.

[0136] The final, effective topology version is broadcast and synchronized to each node via a heartbeat protocol. This synchronization process utilizes a differential transmission mechanism, sending only the modified metadata vector blocks and their association matrices. Local vector space consistency checks are then performed on the receiving nodes. This dynamic evolution mechanism reduces topology reconstruction time compared to traditional solutions and improves index verification pass rates when three new query dimensions are added.

[0137] The implementation steps of the vector-physical dual mapping verification mechanism are:

[0138] The first verification is based on the vector space similarity threshold to filter candidate paths. The screening threshold is a dynamically adjusted value and is calculated using the dynamic threshold calculation formula:

[0139] Threshold = α (current node load / system average load) + β (number of query dimensions / maximum number of supported dimensions). Where Threshold is the dynamic threshold of response time that triggers the passive evolution channel, and α and β are dynamic adjustment coefficients.

[0140] Secondary verification eliminates paths whose delay exceeds the set threshold or whose bandwidth margin is lower than the safe value based on the physical connection delay and bandwidth margin of the actual node.

[0141] It should be further explained that, in the specific implementation process, during the execution of the vector-physical dual mapping verification mechanism, the path screening is performed based on the dynamically adjusted similarity threshold in the first verification phase, including: the dynamic adjustment coefficient α in the threshold calculation formula is calculated according to the current node load rate according to the formula:

[0142] Dynamic setting: The dynamic adjustment coefficient β is obtained from the preset parameter matrix through a table lookup method based on the query dimension complexity. When the similarity score of a candidate path is lower than the calculation threshold, the path degradation mechanism is triggered and it is moved to the queue for review. Among them, the query dimension complexity is defined as the product of the number of dimensions and their association strength.

[0143] During the secondary verification phase, hard elimination is performed based on real-time monitoring data from physical nodes. These include: The network delay threshold is calculated using a dynamic baseline method. When the inter-node delay exceeds the threshold, the corresponding path is automatically marked as failed. The baseline value in the dynamic baseline calculation method is 1.3 times the median historical delay. Bandwidth margin detection uses a sliding window prediction model, combining the current bandwidth usage with the predicted usage within the next three seconds. If the predicted value falls below the safe level, the path is immediately eliminated.

[0144] The candidate paths that have passed the verification enter the scoring weighting stage, and finally the paths with the top 20% comprehensive scores are selected to form the optimized path queue. The formula for calculating the comprehensive score of the path is:

[0145] , where S sim Score the vector space similarity, S load Score the node load balancing score, S total : The overall quality score of the path.

[0146] The characteristic parameters of all eliminated paths are written into the log database of the incremental adversarial training submodule and used as a negative sample input source for the generative adversarial network to optimize the subsequent path inference model. The characteristic parameters include dimension combination patterns, failure reason codes, and historical scoring records. The verification results are synchronized to the metadata processing module in real time via a closed-loop feedback bus, triggering the dynamic update of the physical proximity mapping table. This includes: for frequently passed path-associated node pairs, their association strength coefficients are increased, and the physical proximity weight update formula between nodes is set as follows:

[0147] , where W new : Updated physical proximity weight value, W old : The physical proximity weight value before updating, η is the learning rate coefficient.

[0148] This verification mechanism can improve the accuracy of filtering invalid paths in a stress test of thousands of queries per second.

[0149] The pheromone update rule of the adaptive ant colony algorithm satisfies the pheromone update formula:

[0150] ;

[0151] Where ρ is the pheromone volatility coefficient, : the pheromone increment of the kth ant on the path i→j, τ ij (t+1): pheromone concentration on the path from node i to node j at time step t+1, t: time step number, representing the number of algorithm iterations; i: node number of the starting point of the path, j: node number of the end point of the path, m: total number of ants, which is the number of computing agents in the path optimization, k: number of the ant currently computing, k∈[1,m];

[0152] The hash check chain adopts the Merkle tree structure, the leaf nodes are the data hash values ​​of each storage node, and the root hash value is synchronized to the central scheduler in real time.

[0153] It should be further explained that, in the specific implementation process, in the implementation of the pheromone update rule of the adaptive ant colony algorithm, the initial value of the pheromone concentration between nodes is dynamically set according to the success rate of the historical query path, including: for the node pairs with a query success rate higher than 90% in the past hour, the initial pheromone value is set according to the formula:

[0154] Calculate, where Success count : The number of successful queries on this path in the past hour, Total count : the total number of queries on the path in the past hour, τ init : path initial pheromone concentration value, τ max: The maximum pheromone threshold set by the system, τ max The maximum pheromone concentration threshold set for the system.

[0155] The pheromone volatility coefficient ρ adopts a dynamic adjustment strategy, which is adjusted according to the current network congestion level. Adaptive changes are made. When a sudden increase in transmission delay between nodes is detected, the volatile rate is automatically increased to accelerate path updates.

[0156] The calculation of is deeply coupled with the path quality index: when ant k passes through the path from node i to j, the pheromone increment formula is:

[0157] ,in, : The pheromone increment of the kth ant on the path i→j, Score diff : The difference in confidence scores between nodes, Distance ij : Network topology distance between physical nodes, Bandwidth total : The maximum bandwidth capacity of the node's physical network interface, : CPU utilization of target node j, τ: pheromone concentration value on the path, reflecting the historical quality of the path, i: node number of the starting point of the path, j: node number of the end point of the path, k: ant number currently being calculated, Bandwidth avail : The real-time available bandwidth of the path from node i to j.

[0158] After the path priority queue is generated, a sliding window mechanism is used to retain the optimal N paths, where N is rounded to the square root of the current number of active nodes, and its characteristic parameters are injected into the Ising model parameter library of the quantum annealing simulator for subsequent path optimization iterations.

[0159] During the Merkle tree construction process of the hash check chain, the leaf node hash value is generated using a dual calculation mechanism, including: first calculating the SHA-256 hash value of the data block content, and then adding the geographical location code of the storage node for secondary hash calculation to form a composite hash value that is resistant to geographical location tampering. The intermediate node hash generation rule is set as follows: when the node level is an odd number of layers, cascade hashing is used. , XOR hash is used for even layers , thereby enhancing the anti-collision ability of the tree structure.

[0160] Synchronous updates of the root hash value are achieved through a two-phase commit protocol: each storage node generates a temporary root hash when data changes and submits it to the central scheduler. Only after more than two-thirds of the nodes have passed verification will the new root hash value be written to the version control log of the metadata-related topology map. If a hash chain break is found during the verification process, the system automatically triggers a three-level repair mechanism. The process includes: first, recalculating the data block on the local node. If failure occurs, the data is restored from the neighboring node replica. Finally, if the repair is still not possible, the full data synchronization process is initiated. All hash verification records are timestamped and signed, and an unalterable audit log is formed through a blockchain-style chain storage structure. Each log entry contains a three-tuple structure of the previous block hash, the current operation type code, and the data fingerprint. This mechanism achieves a verification response within seconds in a 10TB data scale test, with high data consistency verification accuracy.

[0161] A method for unified management of enterprise financial files based on a distributed storage-based unified management system for enterprise financial files comprises the following steps:

[0162] S1: Vectorize and encapsulate financial files through a self-organizing metadata decomposition algorithm to generate a dynamically evolving metadata association topology graph;

[0163] S2: Decompose query conditions using a semantic-vector dual-channel parser, and deduce the optimal query path by combining back-propagation gradient detection with quantum annealing simulation;

[0164] S3: Performs three-stage screening: node pre-aggregation, path reweighted aggregation, and final accurate aggregation, and outputs the results through a double verification mechanism;

[0165] S4: When a new query dimension is added, the dual-channel index evolution mechanism is triggered to generate a double-verified combined index and update the metadata topology.

[0166] Through innovative dynamic metadata topology reconstruction and dual-channel index optimization mechanisms, efficient management of multi-dimensional financial archive data is achieved. Vectorized metadata encapsulation technology is used to transform traditional discrete attributes into high-dimensional correlation vectors. Combined with quantum annealing path optimization and a three-stage aggregation verification strategy, the response time for multi-condition combination queries is reduced compared to traditional solutions, while also reducing network bandwidth consumption. Through a dynamic adversarial training model and incremental index evolution mechanism, the system automatically completes index reconstruction when new query dimensions are added. This reduces reconstruction time compared to traditional methods, while maintaining node load balancing within a reasonable fluctuation range, significantly improving the system's elastic scalability.

[0167] It breaks through the bottleneck of real-time analysis and deep correlation mining in enterprise financial record management, supports millisecond-level multi-dimensional joint queries with high query accuracy; through blockchain-style spatiotemporal validity marking and differential privacy protection mechanism, it ensures the auditability and compliance of data throughout its life cycle, meeting data supervision requirements such as GDPR; the system's adaptive dynamic evolution characteristics enhance the intelligence level of financial record management processes, providing a safe and reliable data governance infrastructure for enterprise digital transformation.

[0168] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0169] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A unified enterprise financial file management system based on distributed storage, characterized in that: include: The metadata processing module is connected to the distributed storage node cluster and is used to convert the metadata attributes of the enterprise financial files into high-dimensional sparse vectors, and generate a metadata association topology map containing cross-dimensional association strength coefficients through the dynamic dimension weight allocation module; a query parsing module, in communication with the metadata processing module, configured to decompose a multi-condition combination query into a semantic constraint core and a vector derivation domain through a semantic-vector dual-channel parser; The index dynamic optimization module is electrically connected to the metadata processing module and the query parsing module, and includes an active evolution channel and a passive evolution channel. It is used to generate a lightweight combined index based on reinforcement learning prediction and differential privacy analysis. The distributed aggregation module is connected to the distributed storage node cluster through the data bus and uses a three-stage data aggregation strategy to perform phased screening and verification of cross-node computing flows; The metadata processing module, query parsing module, index dynamic optimization module and distributed aggregation module are connected in series to form a closed-loop feedback link; In the index dynamic optimization module: The active evolution channel uses a temporal convolutional reinforcement learning model to analyze the time series characteristics of query patterns to predict high-probability dimension combinations within a preset time period in the future; When the query response time of the passive evolution channel exceeds the dynamic threshold, it triggers k-anonymization processing based on differential privacy protection, performs desensitization analysis on the query log and generates adversarial training samples; The k-anonymization process includes generalizing the sensitive dimension values ​​to ensure that each set of query conditions contains at least k indistinguishable subject features, and injecting Laplace noise during the adversarial training sample generation process; When the distributed aggregation module executes the three-stage data aggregation strategy: In the node pre-aggregation phase, each storage node calculates a confidence score based on the local vectorized metadata and returns a subset of results with a score above a first threshold; In the path reweighted aggregation phase, the central scheduler uses an adaptive ant colony algorithm to dynamically adjust the node access priority of the secondary query based on the confidence distribution of the node return results; In the final precise aggregation phase, full data extraction is performed only on physical storage nodes with confidence scores higher than the second threshold, and data integrity is verified through a hash check chain; When generating the metadata association topology, a multidimensional association matrix is ​​constructed based on the cross-dimensional association strength coefficient. The association strength coefficient is calculated by combining the co-occurrence probability between dimensions with the statistical weight of the query condition combination. A sliding window mechanism is used to dynamically perform weighted averaging on the dimension combinations in the query and record the results in the physical proximity mapping table of the distributed hash table. Through natural language processing technology, entity labels and logical relationships in query conditions are extracted, attributes with clear enumeration values ​​are classified as semantic constraint cores, and precise matching query statement templates are generated; at the same time, continuity or fuzzy association conditions are mapped to vector derivation domains and converted into hyperplane cutting operation parameters in vector space.

2. The enterprise financial file unified management system based on distributed storage according to claim 1, characterized in that: The metadata processing module includes: The dynamic dimension weight assignment module assigns a dynamic weight coefficient to each dimension of the metadata vector by analyzing the co-occurrence frequency of the time, department, project, and amount dimensions in historical query logs; The incremental adversarial training submodule uses a generative adversarial network architecture to generate simulated query requests and conduct adversarial training on actual query patterns, outputting an optimized metadata topology distribution strategy. The metadata association topology map is stored in a distributed hash table and includes a physical proximity mapping table for recording the physical node position relationship corresponding to the metadata vectors with high association.

3. The enterprise financial file unified management system based on distributed storage according to claim 2, characterized in that: The query parsing module also includes a cross-layer path optimization engine. The implementation of the cross-layer path optimization engine is based on a vector-physical dual mapping verification mechanism. The implementation process includes: Based on the vector space characteristics of the metadata association topology graph, the vector boundaries of the candidate data clusters are located using the back-propagation gradient detection algorithm; Combined with the spatial distribution topology of physical storage nodes, a quantum annealing simulator is used to calculate the minimum cross-node hop path; The operating parameters of the quantum annealing simulator include a network delay weight factor between physical nodes, a storage node load balancing attenuation coefficient, and a thermodynamic simulation temperature parameter of a cross-dimensional query condition.

4. The enterprise financial file unified management system based on distributed storage according to claim 3 is characterized by: The three-stage data aggregation strategy also includes: The logical consistency checking unit establishes the reverse mapping relationship between query conditions and result sets, and verifies whether the returned data fully satisfies all query constraints; The spatiotemporal validity marking unit adds the timestamp signature of the data version and the geographic location code of the storage node to the query results to ensure compliance with timeliness and regional compliance requirements.

5. The enterprise financial file unified management system based on distributed storage according to claim 4 is characterized by: The dynamic evolution process of the metadata association topology graph includes: Record the incremental update trajectory of the topology graph as the query pattern changes; When a new query dimension is added, the weight distribution of the prediction index of the active evolution channel and the correction index of the passive evolution channel is balanced through the game equilibrium controller; The game balance controller uses the Nash equilibrium algorithm to calculate the index weight, and double-checks the validity of the newly created index through vector space coverage verification and physical storage load balancing verification.

6. The enterprise financial file unified management system based on distributed storage according to claim 5, characterized in that: The implementation steps of the vector-physical dual mapping verification mechanism are as follows: The first verification is based on the vector space similarity threshold to filter candidate paths. The screening threshold is a dynamically adjusted value and is calculated using the dynamic threshold calculation formula: ; Among them, Threshold: the dynamic threshold of the response time that triggers the passive evolution channel, α, β: dynamic adjustment coefficients; Secondary verification eliminates paths whose delay exceeds the set threshold or whose bandwidth margin is lower than the safe value based on the physical connection delay and bandwidth margin of the actual node.

7. The enterprise financial file unified management system based on distributed storage according to claim 6, characterized in that: The pheromone update rule of the adaptive ant colony algorithm satisfies the pheromone update formula: ； Where ρ is the pheromone volatility coefficient, : pheromone increment of the kth ant on path i→j, τ ij (t+1): pheromone concentration on the path from node i to node j at time step t+1, t: time step number, representing the number of algorithm iterations; i: node number of the starting point of the path, j: node number of the end point of the path, m: total number of ants, which is the number of computing agents in the path optimization, k: number of the ant currently computing, k∈[1,m]; The hash check chain adopts a Merkle tree structure, the leaf nodes are the data hash values ​​of each storage node, and the root hash value is synchronized to the central scheduler in real time.

8. A method for unified management of enterprise financial files based on the system according to any one of claims 1 to 7, characterized in that: The following steps are involved: S1: Vectorize and encapsulate financial files through a self-organizing metadata decomposition algorithm to generate a dynamically evolving metadata association topology graph; S2: Decompose query conditions using a semantic-vector dual-channel parser, and deduce the optimal query path by combining back-propagation gradient detection with quantum annealing simulation; The query parsing module also includes a cross-layer path optimization engine. The implementation of the cross-layer path optimization engine is based on a vector-physical dual mapping verification mechanism. The implementation process includes: locating the vector boundaries of candidate data clusters using a back-propagation gradient detection algorithm based on the vector space characteristics of the metadata association topology graph; calculating the minimum cross-node hop path using a quantum annealing simulator in combination with the spatial distribution topology of physical storage nodes; the operating parameters of the quantum annealing simulator include a network delay weight factor between physical nodes, a storage node load balancing attenuation coefficient, and a thermodynamic simulation temperature parameter of the cross-dimensional query condition; S3: Executes a three-stage screening process: node pre-aggregation, path reweighted aggregation, and final precise aggregation, and outputs the results through a dual verification mechanism. This three-stage data aggregation strategy also includes: a logical consistency check unit, which establishes an inverse mapping relationship between query conditions and result sets to verify whether the returned data fully satisfies all query constraints; and a spatiotemporal validity marking unit, which adds a timestamp signature of the data version and the geographic location code of the storage node to the query results to ensure compliance with timeliness and regional compliance requirements. S4: When a new query dimension is added, the dual-channel index evolution mechanism is triggered to generate a double-verified combined index and update the metadata topology; the dynamic evolution process of the metadata association topology graph includes: recording the incremental update trajectory of the topology graph as the query pattern changes; when a new query dimension is added, the weight distribution of the predicted index of the active evolution channel and the corrected index of the passive evolution channel is balanced through the game balance controller; the game balance controller uses the Nash equilibrium algorithm to calculate the index weight, and double-checks the validity of the newly created index through vector space coverage verification and physical storage load balancing verification.

Images