A Financial Data Management System Based on a Large Language Model

Abstract

This invention provides a financial data management system based on a large language model, belonging to the field of financial data management technology. It performs semantic analysis and access feature extraction on financial data access requests based on a large language model, combines the fluctuation trend of data read popularity curves, and uses a spectral clustering algorithm to dynamically divide data access levels. Furthermore, it employs a dynamic statistical periodic model for time-segmented preloading. This improves the accuracy and efficiency of data access, optimizes server resource scheduling, facilitates intelligent management of financial data in high-concurrency scenarios, reduces server load, and enhances the overall performance and responsiveness of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial data management, and more specifically, to a financial data management system based on a large language model. BACKGROUND

[0002] With the rapid development of big data and information technology, financial data management systems play a crucial role in the management and decision-making of modern enterprises. Traditional financial data management systems usually rely on relational databases and predefined query statements to achieve data storage and access. Although such systems can meet the basic financial data operation needs, with the exponential growth of enterprise data volume and the increasing complexity of data structure, traditional systems gradually expose obvious shortcomings in data processing efficiency, user experience, and intelligent level. In recent years, artificial intelligence technology, especially the breakthroughs in the field of natural language processing (NLP), has provided new ideas for financial data management. Pre-trained large language models (such as the GPT series) have shown great potential in text analysis, semantic search, and intelligent question answering due to their strong semantic understanding and generation capabilities. However, the research on combining large language models with financial data management is still in its infancy, and how to efficiently handle high-frequency data access requests, dynamically optimize resource scheduling, and intelligently manage data levels are still core problems that need to be solved.

[0003] The existing technology mainly has the following shortcomings in financial data management: First, traditional systems cannot accurately capture the semantic features of data access requests, resulting in low data retrieval and response efficiency, especially when dealing with complex financial operation instructions, which often rely on manually defined rules or templates, lacking flexibility and scalability. Second, current financial data preloading strategies mostly use static rules or simple time series prediction models, failing to fully consider the real-time and diversity characteristics of data access, leading to uneven resource allocation, low cache hit rate, and high server load. In addition, existing technologies in financial data hierarchical management are mostly based on fixed grouping rules, which are difficult to dynamically adapt to changes in data access patterns, and cannot effectively improve data processing efficiency and resource utilization. In the practical application of advanced algorithms such as spectral clustering, semantic analysis, and dynamic statistical cycle models, existing technologies have not formed a systematic solution, resulting in obvious shortcomings in the mining and optimization of multi-dimensional data features. The above shortcomings seriously restrict the performance of financial data management systems in high-concurrency scenarios, and also make it difficult to meet the growing intelligent and efficient needs of enterprises for data management systems. SUMMARY

[0004] In order to solve the above technical problems, the present application is proposed. The present application provides a financial data management system based on a large language model, which solves the problem of how to count the period, improves the access efficiency of financial data in the high-frequency reading stage, and ensures that the performance of the server processing other instructions is not significantly affected.

[0005] According to one aspect of the present application, a financial data management system based on a large language model is provided, which comprises:

[0006] A data reading module is configured to obtain a financial data access request to be processed in the server, and based on the timestamp information in the financial data access request, use a sliding window method to count the data reading frequency within a unit time, and obtain a data reading heat curve.

[0007] A semantic analysis module is configured to perform semantic analysis on the operation instruction in the financial data access request based on a pre-trained large language model, and extract data access features.

[0008] An affinity matrix construction module is configured to generate a data access affinity matrix according to the similarity calculation of the data access features and historical access patterns.

[0009] An access level division module is configured to divide the financial data into multiple access levels by a spectral clustering algorithm according to the fluctuation trend of the data reading heat curve and in combination with the data access affinity matrix.

[0010] An equalization scheduling module is configured to establish a dynamic statistical period model by a time series analysis method based on the distribution characteristics of the access levels, and perform time period preloading of the financial data of different access levels according to the dynamic statistical period model, so as to realize equalization scheduling of server resources.

[0011] Further, the financial data access request contains a request identification number, an access timestamp, a data type label, and operation type information.

[0012] The financial data access request is sorted in the order of the access timestamp through a message queue.

[0013] Further, the semantic analysis module performs semantic analysis, which includes extracting an operation intention vector , an operation entity vector , and a time sequence association vector .

[0014] When the operation intention vector is extracted, first, the outputs of each hidden layer of the model are weighted by inter-layer attention to highlight important features at different semantic levels, and high-dimensional semantic features are projected into an intention representation space through an intention mapping matrix.

[0015] Further, in extracting the operation entity vector , the semantic representation of the entity is fused with the type information, and the contribution weight of different entities is determined by the entity importance score;

[0016] In extracting the timing correlation vector , the dynamic fusion module is used to adaptively integrate the historical timing information and the current context information.

[0017] Further, the affinity matrix construction module retrieves the historical access pattern from the pre-constructed historical access feature library ; calculate the enhanced semantic similarity between the data access feature and the historical access pattern ;

[0018] Based on the enhanced semantic similarity , the data access affinity matrix is constructed.

[0019] Further, the data access affinity matrix is as follows:

[0020] ;

[0021] Where, and represent the i-th and j-th data access feature vectors, represents the total number of access requests in the current batch, represents the enhanced semantic similarity between the data access feature vector and the data access feature vector ;

[0022] Wherein, the calculation of the enhanced semantic similarity is as follows:

[0023] ;

[0024] Wherein, represents the cosine similarity, represents the similarity enhancement coefficient, represents the distance attenuation factor, represents the L2 norm, represents the data access feature vector.

[0025] Further, the access level division module divides the financial data into multiple access levels, including the following steps:

[0026] Feature extraction is performed according to a data reading heat curve in the data reading module to obtain a heat fluctuation feature vector; a time sequence correlation matrix is constructed based on the heat fluctuation feature vector ;

[0027] The time sequence correlation matrix is fused with the data access affinity matrix to construct a comprehensive affinity matrix ;

[0028] Based on the comprehensive affinity matrix , a Laplacian matrix is constructed , and the optimal clustering number is dynamically determined through eigenvalue decomposition , and a normalized feature matrix U is constructed by selecting the feature vector corresponding to the smallest eigenvalue

[0029] In the normalized feature space, an adaptive clustering method based on density peaks is used for clustering

[0030] Multi-dimensional statistical analysis is performed on each access level to construct a level feature vector .

[0031] Further, the adaptive clustering method based on density peaks is used for clustering, specifically including:

[0032] The local density and the minimum distance to the high-density point of each data point are calculated

[0033] The cluster center score is calculated in combination with the above information, the cluster center point is selected, and other data points are assigned to the corresponding access level through density reachability.

[0034] Further, the calculation of the cluster center score is as follows:

[0035] ;

[0036] wherein, represents the cluster center score, represents the average difference of point i and its K-neighbor density, is a density difference adjustment coefficient, used to enhance the cluster center candidate probability of the area with significant density change, is the global average density

[0037] wherein,

[0038] ;

[0039] ;

[0040] in, This represents the Euclidean distance between points i and j in the feature space. The adaptive cutoff distance is determined through K-nearest neighbor statistics. This is a kernel function for smooth truncation.

[0041] Furthermore, the balanced scheduling module divides data time periods according to the dynamic statistical periodic model, evaluates the access probability to filter the candidate set, dynamically adjusts the preloading order and timing in combination with the resource status, and optimizes the strategy to improve preloading efficiency.

[0042] Compared with existing technologies, the financial data management system based on a large language model provided by this invention performs semantic analysis and access feature extraction on financial data access requests based on a large language model. Combined with the fluctuation trend of data read popularity curves, it dynamically divides data access levels using a spectral clustering algorithm and employs a dynamic statistical periodic model for time-segmented preloading. This improves the accuracy and efficiency of data access, optimizes server resource scheduling, facilitates intelligent management of financial data in high-concurrency scenarios, reduces server load, and enhances the overall performance and responsiveness of the system. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0044] Figure 1 This is a system block diagram of a financial data management system based on a large language model according to an embodiment of the present invention. Detailed Implementation

[0045] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0046] Figure 1 This is a system block diagram of a financial data management system based on a large language model according to an embodiment of the present invention. Figure 1 As shown, the financial data management system based on a large language model includes:

[0047] The data reading module is used to acquire financial data access requests to be processed in the server, and based on the timestamp information in the financial data access requests, to use the sliding window method to count the data reading frequency per unit time and obtain the data reading popularity curve.

[0048] The semantic analysis module is used to perform semantic analysis on the operation instructions in the financial data access request based on a pre-trained large language model, and extract data access features.

[0049] The affinity matrix construction module is used to generate a data access affinity matrix based on the similarity calculation between the data access features and historical access patterns.

[0050] The access level division module is used to divide financial data into multiple access levels based on the fluctuation trend of the data reading heat curve and the data access affinity matrix, using a spectral clustering algorithm.

[0051] The balanced scheduling module is used to establish a dynamic statistical periodic model based on the distribution characteristics of the access levels using time-series analysis methods, and to preload financial data of different access levels in different time periods according to the dynamic statistical periodic model, so as to achieve balanced scheduling of server resources.

[0052] The data reading module acquires pending financial data access requests from the server. Each financial data access request includes a request identifier, access timestamp, data type label, and operation type information. These requests are sorted via a message queue according to the access timestamps. The sorted financial data access requests are then grouped according to the data type label, and a corresponding sliding window is set for each group. The initial time span of the sliding window is T minutes, and the sliding step size is t minutes. Within each sliding window, the number of financial data access requests is cumulatively counted to obtain the data reading frequency within the window period. By continuously sliding the sliding window along the time axis, the data reading frequencies within each time window are connected to form a data reading heat curve, where the horizontal axis of the data reading heat curve represents the time axis, and the vertical axis represents the data reading frequency.

[0053] The semantic analysis module performs semantic analysis on the operation instructions in the financial data access request based on a pre-trained large language model. The operation instructions include database query statements, financial operation types, and descriptions of the operation objects. The operation instructions are input into the pre-trained large language model, and an attention mechanism is used to extract the operation intent vector. , Operation entity vector Time-series correlation vector Among them, in extracting the operation intention vector First, inter-layer attention weighting is applied to the outputs of each hidden layer of the model to highlight important features at different semantic levels. Then, the high-dimensional semantic features are projected onto the intent representation space through an intent mapping matrix. The operation intent vector... The specific calculation method is as follows:

[0054] ;

[0055] in, This represents the output of the k-th hidden layer of the large language model. This represents the attention weights of the k-th layer. Intent mapping matrix, Indicates the number of model layers.

[0056] Extracting the operation entity vector At the same time, an entity type awareness mechanism is introduced to fuse the semantic representation of entities with type information. The contribution weight of different entities is determined through entity importance scoring, and the operation entity vector... The specific calculation method is as follows:

[0057] ;

[0058] in, This represents the embedding representation of the j-th operation entity. This represents the entity weight coefficient calculated based on the entity importance scoring module. This represents the corresponding entity type mask vector. Represents the element-wise multiplication operation of vectors. This indicates the total number of entities included in the current operation instruction.

[0059] Extracting the time-series correlation vector At that time, a dynamic fusion module based on a gating mechanism is designed to adaptively integrate historical time series information with current context information. The specific calculation method of the time series correlation vector Vt is as follows:

[0060] ;

[0061] in, This represents the temporal vector state of the system's memory at the previous moment. The context semantic vector representing the current operation instruction. This represents the time-series information fusion coefficients calculated through a gating network, used to balance the importance of historical and current information. This indicates that recurrent neural network operations are used to capture long-term dependencies.

[0062] The three types of semantic vectors are integrated through an adaptive feature fusion module to obtain a unified data access feature representation. The specific calculation method is as follows:

[0063] ;

[0064] in, Represents the fusion weight matrix. Indicates the bias term. This represents the activation function. This represents vector concatenation. and These represent the fusion coefficients of the entity vector and the temporal vector, respectively.

[0065] On the other hand, the affinity matrix construction module retrieves historical access patterns from a pre-built historical access feature library. The historical access feature database stores typical access feature combinations over different time periods, including access patterns at multiple time scales such as daytime, weektime, and monthtime; the data access features are calculated. With the aforementioned historical access pattern Enhanced semantic similarity between An adaptive weighting mechanism based on feature distance is introduced:

[0066] ;

[0067] in, Represents cosine similarity. Represents the similarity enhancement coefficient. Represents the distance attenuation factor. This represents the L2 norm.

[0068] A data access affinity matrix is ​​constructed based on the enhanced semantic similarity. :

[0069] ;

[0070] in, and These represent the i-th and j-th data access feature vectors, respectively. This represents the total number of access requests in the current batch. By considering the global similarity distribution, the data access affinity matrix A is normalized to ensure that it satisfies the symmetry property, and the matrix element values ​​reflect the degree of similarity in access patterns between different access requests.

[0071] It is worth noting that although the above description outlines the technical process for generating the data access affinity matrix, for ease of understanding, the following specific examples will illustrate this:

[0072] For example, in a bank's core accounting system, when multiple tellers are processing accounting transactions simultaneously, the operation instructions might be as follows:

[0073] Teller A executes: "To retrieve customer X's transaction details for the past three months for loan approval."

[0074] Teller B executes: "Export customer Y's balance sheet for financial analysis."

[0075] Teller C executes: "Check customer X's credit record for loan approval."

[0076] When analyzing these operational instructions using a large language model: during the extraction of the operational intent vector Vi, the model can identify that tellers A and C have similar operational intents (both related to loan approval), while teller B's operational intent is significantly different (for financial analysis purposes).

[0077] When extracting the operation entity vector Ve, the model identifies key entities: "Customer X", "Transaction Details", "Loan Approval" (Teller A), "Customer Y", "Balance Sheet", "Financial Analysis" (Teller B), "Customer X", "Credit Record", and "Loan Approval" (Teller C). This reveals that Tellers A and C are operating on the same customer entity.

[0078] In the time-series correlation vector Vt analysis, if these operations occur within a short time window (e.g., within 10 minutes), and historical data indicates that loan approval typically requires querying transaction details and credit records, the system will establish a time-series correlation between teller A and teller C's operations. The final generated data access affinity matrix will show:

[0079] A_{AC} (the affinity between tellers A and C) will get a high value (e.g., 0.85);

[0080] A_{AB} and A_{BC} (affinity associated with teller B) will receive lower values ​​(e.g., 0.2).

[0081] This difference in affinity indicates that the operations of tellers A and C are highly business-related and may require joint preloading or caching optimization of related data; while the operations of teller B can be processed relatively independently.

[0082] For example, in a company's financial system, finance personnel perform the following operations during the month-end settlement period: "Export the monthly payroll details for department P", "Query the overtime pay statistics for department Q", and "Statistics on social security contributions for department P".

[0083] The system's analysis reveals that these operations all fall under the human resource cost accounting scenario, and that the operations related to department P exhibit stronger data access correlations, thus reflecting this business relationship pattern in the affinity matrix. Ultimately, the goal of this technical solution is to accurately characterize the relationships between data access patterns, providing a reliable basis for subsequent data stratification and preloading strategies, thereby improving the system's data access efficiency. This semantic understanding-based correlation analysis method better grasps the inherent connections within business scenarios than traditional methods based solely on access frequency, leading to more intelligent data management decisions.

[0084] The access level division module reads the fluctuation trend of the popularity curve from the data and first performs feature extraction and analysis of the popularity curve. Specifically, the popularity curve is divided into multiple consecutive observation windows in the time dimension, with each observation window having a time span of Tw and an overlap rate of 50% between adjacent windows to ensure the continuity of feature extraction. The curve sequence within each observation window is decomposed in the time-frequency domain by combining wavelet transform and Fourier transform to extract multi-scale frequency features, and noise components are filtered out by adaptive thresholding to obtain a popularity fluctuation feature vector that reflects the periodicity of the access pattern.

[0085] Based on the aforementioned heat fluctuation feature vector, a temporal correlation matrix R considering temporal locality is constructed, and a double Gaussian kernel is introduced for similarity measurement:

[0086] ;

[0087] in, and Let i and j represent the heat fluctuation feature vectors at time i and j, respectively. and Indicates the corresponding timestamp, The smoothing parameter representing the feature space is dynamically adjusted based on the standard deviation of the feature vector distribution. The smoothing parameter represents the temporal space and is used to control the rate of time decay. Two Gaussian kernels capture feature similarity and temporal correlation, respectively.

[0088] To comprehensively utilize temporal and access pattern information, the temporal correlation matrix R and the data access affinity matrix A obtained in step S2 are fused using multimodal features to construct a comprehensive affinity matrix C. The fusion process considers the interaction enhancement effect between matrices.

[0089] ;

[0090] in, The time-series correlation weighting coefficients are determined through grid search optimization on the validation set. This represents the interaction enhancement coefficient, used to adjust the influence strength of matrix interaction terms. This represents the Hadamard product operation, used to capture two non-linear interaction patterns of correlation.

[0091] Based on the comprehensive affinity matrix, a normalized Laplacian matrix L is constructed, which contains global structural information between data points:

[0092] ;

[0093] in, The normalization method is a degree matrix, where the diagonal element Dii is equal to the sum of the elements in the i-th row of matrix C. This normalization method can balance the differences in connection strength between different data points; for the Laplace matrix... Perform eigenvalue decomposition and determine the optimal number of clusters m using dynamic programming analysis of eigenvalue differences.

[0094] ;

[0095] Furthermore, the eigenvectors corresponding to the m smallest eigenvalues ​​are selected to construct a feature matrix V, and each row of the feature matrix V is normalized using the L2 norm to obtain a normalized feature matrix U. In the normalized feature space, using an adaptive clustering method based on density peaks, the local density of each data point i is first calculated. Then calculate the minimum distance from each data point to the high-density points. Combined with local density and distance Calculate the cluster center score of the data points. :

[0096]

[0097] in, This represents the average difference in density between point i and its K nearest neighbors. This is a density difference adjustment coefficient, used to enhance the candidate probability of cluster centers in regions with significant density changes. The global average density is used. Cluster centers are selected based on the ranking of cluster center scores and an adaptive interval threshold. The remaining data points are then assigned to the corresponding access levels according to the density reachability principle, where density reachability is determined by the continuity of the density path.

[0098] in,

[0099] ;

[0100] ;

[0101] in, This represents the Euclidean distance between points i and j in the feature space. The adaptive cutoff distance is determined through K-nearest neighbor statistics. To smoothly truncate the kernel function, a Gaussian form is used to ensure the continuity of the density estimation.

[0102] Finally, multi-dimensional statistical analysis was performed on the data characteristics of each access level to construct the hierarchical feature vector Q:

[0103] ;

[0104] in, This represents the set of data points at the l-th access level. This indicates the number of data points in the set. This indicates a vector concatenation operation, where the feature vector integrates the spatial distribution, density characteristics, and cluster significance information of the data points.

[0105] It's worth noting that while the above description details the technical solution for dividing financial data into multiple access levels, the specific implementation details need to be understood in conjunction with the actual business scenario. For example, taking "accounts receivable details" data as an example, if data point i represents access records from Monday morning and data point j represents access records from Tuesday morning, the calculated Rij value will be larger due to the close time and both being weekday mornings. However, if data point j is changed to access records from Sunday, even with similar characteristics, the final relevance value will decrease accordingly due to the difference in time attributes. Regardless, the ultimate goal is to group financial data with similar access characteristics and business relevance into the same access level to facilitate targeted resource scheduling and performance optimization by the system. This hierarchical division is not static but dynamically adjusted as business models change and access characteristics evolve, thereby continuously maintaining high system response efficiency.

[0106] The specific implementation process of the balanced scheduling module using time series analysis to establish a dynamic statistical periodic model is as follows: First, the system extracts time series feature sequences from the data access logs, including historical data in three dimensions: access frequency, data volume, and resource consumption. For each dimension of data, a sliding window method is used for preprocessing. When the window size is set to T, the system calculates statistical features within the window, including basic statistics such as mean, variance, peak, and trough. Next, the preprocessed time series data is decomposed into multiple scales. The system considers the variation characteristics at three time scales: hourly, daily, and weekly. When a significant periodic pattern is found at a certain time scale, the system marks that feature as a candidate periodic feature. Then, autocorrelation analysis and cross-correlation analysis are applied to each candidate periodic feature to calculate the correlation coefficient at different time delays. If the correlation coefficient shows a significant peak at a certain delay value, that delay value is taken as a potential statistical period. The system maintains multiple potential statistical periods simultaneously and assigns a confidence score to each period. When the prediction accuracy of a certain period remains high, its confidence score will increase accordingly.

[0107] During model operation, the system updates the state vectors at various time scales in real time. Each state vector contains the feature value at the current moment and the historical evolution trend. When new access data arrives, the system predicts the next statistical period based on these state vectors. If a sudden change in the access pattern is detected, the system initiates a rapid adjustment mechanism, updating the state estimate using an exponentially weighted moving average method to enable the model to quickly adapt to the new access pattern. Simultaneously, the system maintains an adaptive learning rate. When the prediction error is large, the learning rate is increased to accelerate model adjustment; when prediction performance is stable, the learning rate is decreased to maintain model stability. Furthermore, the system introduces a seasonal adjustment mechanism. By detecting periodic changes in data at different time scales, it identifies and compensates for the impact of seasonal factors. When a strong seasonal pattern is detected, the system adjusts the prediction strategy accordingly. To improve the model's robustness, the system also detects and handles outliers. When data at a certain point in time deviates significantly from historical patterns, it corrects this through local smoothing or interpolation methods to ensure that outliers do not excessively affect the model's predictive performance. Finally, the system will periodically evaluate the model's prediction accuracy and dynamically adjust the model parameters based on the evaluation results. If the prediction effect at a certain time scale is found to be consistently poor, the weight of that scale in the final prediction will be reduced, thereby ensuring the overall prediction performance of the model.

[0108] The specific implementation process of the balanced scheduling module for time-segmented preloading of financial data at different access levels based on the dynamic statistical period model is as follows: When the system receives a preloading instruction, it first obtains the predicted statistical period and preloading priority of each access level at the current time, and divides the data of each level into several time periods according to the predicted period. For each time period, the system evaluates the expected access probability and required resource amount of the data. If the expected access probability of a certain time period is higher than the dynamic threshold, and the current server resources are sufficient, then the data corresponding to that time period is marked as a preloading candidate set.

[0109] Next, the system sorts the candidate set according to preloading priority and calculates the optimal preloading time for each data block. When the system detects idle resources, it prioritizes loading higher-priority data blocks while monitoring actual access patterns. If a significant deviation is detected between the actual access pattern and the prediction, the system dynamically adjusts the preloading strategy for subsequent time periods. During preloading, the system continuously monitors the server load. When the server load exceeds a safety threshold, it pauses preloading and releases some cached resources. For preloaded data, the system sets different cache expiration times based on its timeliness and manages the cache space using the LRU-K algorithm. When cache space is insufficient, the system prioritizes cleaning up data with low access frequency and nearing expiration. If the preloading hit rate at a certain level consistently falls below expectations, the system re-evaluates the preloading strategy for that level and adjusts the prediction model parameters if necessary. Through this dynamic adjustment mechanism, the system can intelligently preload data at different access levels while ensuring server performance, thereby improving overall data access efficiency.

[0110] Meanwhile, the system maintains a preloading operation log, recording detailed information for each preloading operation, including preloading time, data range, resource consumption, and cache hit rate. This log data will be used for subsequent strategy optimization and model tuning. During actual operation, the system dynamically adjusts the aggressiveness of preloading based on the characteristics of peak and off-peak business periods. A more aggressive preloading strategy can be adopted during off-peak periods, while a more conservative approach is used during peak periods to ensure the normal operation of core business processes. Furthermore, the system considers data correlation. When preloading a data block, it simultaneously evaluates the preloading value of highly correlated data blocks, achieving collaborative preloading of related data and further improving cache hit rate.

[0111] In summary, the financial data management system based on a large language model, as described in this invention, is explained. By performing semantic analysis and access feature extraction on financial data access requests using a large language model, and combining this with the fluctuation trend of data read popularity curves, a spectral clustering algorithm is used to dynamically divide data access levels, and a dynamic statistical periodic model is employed for time-segmented preloading. This improves the accuracy and efficiency of data access, optimizes server resource scheduling, facilitates intelligent management of financial data in high-concurrency scenarios, reduces server load, and enhances the overall performance and responsiveness of the system.

[0112] Here, those skilled in the art will understand that the specific operations of each step in the above-described image-based PICC intravenous therapy information monitoring method have been referenced above. Figure 1 The description of the financial data management system based on the large language model is detailed here, and therefore, its repeated description will be omitted.

Claims

1. A large language model-based financial data management system, characterized by, The method comprises the following steps: The data reading module is used to obtain the financial data access request to be processed in the server, and based on the timestamp information in the financial data access request, the data reading frequency in a unit time is counted by using the sliding window method to obtain the data reading heat curve; The semantic analysis module is used to perform semantic analysis on the operation instruction in the financial data access request based on the pre-trained large language model to extract the data access feature; The affinity matrix construction module is used to generate the data access affinity matrix according to the similarity calculation of the data access feature and the historical access mode; The access level division module is used to divide the financial data into multiple access levels by the spectral clustering algorithm according to the fluctuation trend of the data reading heat curve and in combination with the data access affinity matrix; The balanced scheduling module is used to establish a dynamic statistical period model by using the time sequence analysis method based on the distribution characteristics of the access level, and to perform time period preloading on the financial data of different access levels according to the dynamic statistical period model to realize the balanced scheduling of the server resources; The access level division module divides the financial data into multiple access levels, which comprises the following steps: According to the data reading hotness curve in the data reading module, feature extraction is performed to obtain a hotness fluctuation feature vector; and based on the hotness fluctuation feature vector, a time sequence correlation matrix is constructed ; The time correlation matrix is constructed The data access affinity matrix is constructed The multi-modal feature fusion is performed to construct a comprehensive affinity matrix ; based on the integrated affinity matrix constructing a laplacian matrix and dynamically determining the optimal clustering number by eigenvalue decomposition selecting eigenvectors corresponding to the smallest eigenvalues to construct a normalized feature matrix U In the normalized feature space, an adaptive clustering method based on density peaks is used for clustering; Perform multi-dimensional statistical analysis on each access level to build level feature vector ; The adaptive clustering method based on density peaks for clustering specifically comprises the following steps: The local density and the minimum distance to the high-density point of each data point are calculated; The cluster center score is calculated in combination with the above information, the cluster center point is selected, and other data points are assigned to the corresponding access level through density reachability; The calculation of the cluster center score is shown in the following formula: ； wherein, denotes the cluster center score, denotes the average difference of point i to its K-neighbor density, is a density difference adjustment coefficient to enhance the cluster center candidate probability in regions where the density changes significantly, is the global average density; Wherein, ； ； wherein, denotes the Euclidean distance between points i and j in the feature space, denotes the adaptive cut-off distance determined by K-Nearest Neighbor statistics, is a smooth cut-off kernel function. 2.The large language model-based financial data management system of claim 1, wherein, The financial data access request contains a request identification number, an access timestamp, a data type label and operation type information; The financial data access request is sorted in the order of the access timestamp through the message queue. 3.The large language model-based financial data management system of claim 1, wherein, The semantic analysis module performs semantic analysis including extracting an operation intention vector , an operation entity vector , and a timing association vector ; extracting the operation intention vector At this time, first, the inter-layer attention is weighted to the output of each hidden layer of the model, so as to highlight the important features of different semantic levels, and the high-dimensional semantic features are projected to the intention representation space through the intention mapping matrix. 4.The large language model-based financial data management system of claim 3, wherein, When extracting the operation entity vector The semantic representation of the entity is fused with the type information, and the contribution weight of different entities is determined by the entity importance score. in extracting the time-series correlation vector At the time, the dynamic fusion module is used to adaptively integrate the historical time-series information and the current context information. 5.The large language model-based financial data management system of claim 1, wherein, The affinity matrix construction module retrieves historical access patterns from a pre-constructed historical access feature library ; calculates the data access features between the historical access patterns and the enhanced semantic similarity ; based on the enhanced semantic similarity constructing the data access affinity matrix. 6.The large language model-based financial data management system of claim 5, wherein, The data access affinity matrix As shown in the following formula: ； wherein, and denote the i-th and j-th data access feature vector, denotes the total number of access requests in the current batch, denotes the data access feature vector and the enhanced semantic similarity of the data access feature vectors . The calculation of the enhanced semantic similarity is shown in the following formula: ； wherein, denotes a cosine similarity, denotes a similarity enhancement coefficient, denotes a distance decay factor, denotes an L2 norm, denotes a data access feature vector. 7.The large language model-based financial data management system of claim 1, wherein, The balanced scheduling module divides the data time period by predicting the period according to the dynamic statistical period model, evaluates the access probability to screen the candidate set, dynamically adjusts the preloading order and timing in combination with the resource state, and optimizes the strategy to improve the preloading efficiency.

Images