Intelligent storage and analysis decision support system for archives based on cloud computing

Abstract

This invention provides a cloud-based intelligent archival storage and analysis decision support system, belonging to the field of cloud computing and archival management technology. The system comprises a cloud computing platform layer, an intelligent archival storage layer, a multi-dimensional analysis layer, a decision support layer, and a feedback optimization layer. This five-layer architecture forms a closed-loop system of "storage-analysis-decision-feedback-iteration." The cloud computing platform layer adopts an edge-cloud distributed architecture to achieve elastic resource scheduling. The intelligent archival storage layer, through heterogeneous storage and improved TF-IDF+CNN feature encoding, achieves intelligent storage and unique identification of multiple types of archives. This invention solves the technical problems of traditional archival management systems, such as scattered storage, singular analysis, unscientific decision-making, and lack of self-optimization capabilities. It achieves flexible archival storage, precise analysis, intelligent decision-making, and system self-optimization, significantly improving archival management efficiency and decision reliability. It is applicable to archival management and decision-making scenarios in multiple fields.

Description

Technical Field

[0001] This invention belongs to the technical fields of cloud computing technology, archive management technology, big data analysis technology and decision support system, and in particular relates to a cloud-based intelligent archive storage and analysis decision support system. Background Technology

[0002] As core information assets in government, business, and finance, archives directly impact the efficiency and scientific nature of decision-making for various stakeholders through their management, analysis, and utilization. With the rapid development of information technology, archival data is characterized by explosive growth in volume, diversification of data types (structured data such as archival metadata, semi-structured data such as tag information, and unstructured data such as documents and videos), low data value density, and diverse application scenarios. Traditional archival management models can no longer meet practical application needs, becoming a bottleneck restricting digital transformation in various fields.

[0003] Currently, most existing record management systems adopt local storage or simple cloud storage models, failing to form a systematic intelligent management and decision support system. This results in the following prominent technical shortcomings: First, the storage architecture is simplistic, unable to adapt to the categorized storage of diverse types of archival data, and lacks a collaborative processing mechanism between edge computing nodes and cloud core nodes. Local storage suffers from resource bottlenecks, unable to meet the storage needs of massive amounts of archival data, while pure cloud storage suffers from large data transmission latency and slow response speeds. Furthermore, data encryption and disaster recovery backup mechanisms are inadequate, making it difficult to guarantee the security and reliability of archival data. Second, the dimensions of archival data analysis are limited, mostly remaining at the level of basic statistical queries and simple searches, lacking machine learning capabilities. The deep integration and application of intelligent analysis methods such as deep learning cannot uncover the implicit relationships and potential value between archival data, making it difficult to provide comprehensive and accurate data analysis support for decision-making. Third, the decision support capability is weak, only able to simply display the analyzed data, without establishing a personalized decision-making solution generation mechanism, and lacking multi-scenario simulation verification of decision-making solutions, resulting in low scientificity and feasibility of decision-making solutions and poor implementation effects. Fourth, the system lacks a closed-loop feedback optimization mechanism, the analysis model and decision-making rules are fixed and cannot be dynamically adjusted according to the actual implementation effect of the decision-making solution, the system's adaptability and continuous service capability are insufficient, and it is difficult to adapt to the dynamic needs of different fields and scenarios.

[0004] Therefore, it is necessary to provide a new cloud-based intelligent archival storage and analysis decision support system to solve the above-mentioned technical problems. Summary of the Invention

[0005] The technical problem solved by this invention is to provide a cloud-based intelligent storage and analysis decision support system for archives that addresses the issues of scattered storage, simplistic analysis, unscientific decision-making, and lack of self-optimization capabilities in traditional archives management systems. This system achieves flexible archives storage, precise analysis, intelligent decision-making, and system self-optimization, significantly improving archives management efficiency and decision reliability. It is applicable to archives management and decision-making scenarios in multiple fields.

[0006] To address the aforementioned technical problems, the present invention provides a cloud-based intelligent archival storage and analysis decision support system, comprising: The five-layer architecture, consisting of a cloud computing platform layer, an intelligent archive storage layer, a multi-dimensional analysis layer, a decision support layer, and a feedback optimization layer, achieves bidirectional data interaction through cloud communication protocols, forming a closed-loop system of "storage-analysis-decision-feedback-iteration". The cloud computing platform layer adopts a distributed architecture of edge computing nodes and cloud core nodes. Edge computing nodes are used for preprocessing and local storage of raw archival data, while cloud core nodes are used for centralized storage of archival data, model training, and large-scale computing. Edge computing nodes and cloud core nodes achieve real-time interaction of heterogeneous data through a data synchronization protocol, and are configured with a resource elastic scheduling module to dynamically allocate computing and storage resources; The intelligent archive storage layer includes a heterogeneous storage module and an archive feature encoding module. The heterogeneous storage module is adapted to the classified storage of structured, semi-structured, and unstructured archive data. The archive feature encoding module extracts features and generates unique codes for archive data by combining an improved TF-IDF algorithm with a deep convolutional neural network, enabling rapid retrieval and association of archives. The multi-dimensional analysis layer has a built-in fusion analysis model, which is composed of a weighted fusion of a statistical analysis sub-model, a machine learning prediction sub-model, and a deep learning correlation analysis sub-model. The weight coefficients of each sub-model are determined by the analytic hierarchy process, so as to achieve multi-dimensional and high-precision archival data analysis. The decision support layer includes a personalized decision push module and a decision simulation verification module. The decision simulation verification module uses a Monte Carlo simulation engine to perform multi-scenario simulation verification of the generated decision scheme to ensure the feasibility of the decision. The feedback optimization layer includes a decision effect acquisition module and a model iteration optimization module. The model iteration optimization module dynamically iterates the parameters and weight coefficients of the fusion analysis model based on the gradient descent algorithm to achieve self-optimization of the system's analysis capability and decision support capability.

[0007] As a further embodiment of the present invention, the heterogeneous storage module includes a relational database, a non-relational database, and a distributed file system, which are used to store archive metadata, semi-structured archive tag data, and unstructured raw archive data, respectively. Each storage unit is equipped with a data encryption module and a disaster recovery backup module. The data encryption module uses the SM4 symmetric encryption algorithm to encrypt and store the archive data, and the disaster recovery backup module uses a multi-replica backup mechanism in different locations to ensure that the archive data is not lost or leaked.

[0008] As a further aspect of the present invention, the feature extraction process of the archive feature encoding module satisfies formulas (1) and (2), and the specific steps are as follows: Step 1: Extract shallow text features from the archive using the improved TF-IDF algorithm, and calculate the formula (1): ; in, For the first The first file The word frequency of each feature word The total number of archives. For including the first The number of files for each feature word; Step 2: Input the shallow text features into the deep convolutional neural network to extract the archive depth features and calculate formula (2): ; in, For convolution kernel weights, For bias terms, For convolution operations, The depth features extracted by the deep convolutional neural network; Step 3: Put and The features are concatenated to obtain the fused features. ,right Normalization is performed to generate a unique 128-dimensional vector code, which serves as a unique identifier for each document.

[0009] As a further aspect of the present invention, the process of determining the weight coefficients of each sub-model of the fusion analysis model using the analytic hierarchy process is as follows: Step 1: Construct a judgment matrix. Based on the analysis accuracy, applicable scenarios, and computational efficiency of each sub-model, determine the importance comparison relationship between each sub-model and construct a 3×3 judgment matrix. Step 2: Consistency check, calculate the consistency index; ,in, To determine the largest eigenvalue of a matrix, Query the average random consistency index for the number of sub-models. Calculate the random consistency ratio ,when When <0.1, the judgment matrix satisfies the consistency requirement; Step 3: Weight calculation. Calculate the weight coefficients of each sub-model using the eigenvalue method, ensuring that the sum of the weight coefficients is 1. Finally, determine the weights as follows: statistical analysis sub-model weight 0.2, machine learning prediction sub-model weight 0.3, and deep learning association analysis sub-model weight 0.5.

[0010] As a further aspect of the present invention, the statistical analysis sub-model is used to realize basic statistics, trend analysis and association rule mining of archival data, and the Apriori algorithm is used to mine explicit association relationships between archival data; The machine learning prediction sub-model uses the XGBoost algorithm to construct a trend prediction model for archival data and optimizes the model parameters using a grid search method. The deep learning association analysis sub-model uses graph neural networks (GNNs) to mine the implicit relationships between archival data. It constructs an archival association graph with unique vector codes of archives as nodes and the degree of association between archives as edges.

[0011] As a further aspect of the present invention, the personalized decision-making push module, based on the user profile module and the profile analysis results, uses a collaborative filtering algorithm to push customized decision-making schemes to different users. The specific steps are as follows: Step 1: The user profiling module collects user characteristics such as industry attributes, decision-making needs, and operating habits to construct user feature vectors; Step 2: Calculate the similarity between the user feature vector and the feature vector of the analysis results, and select the Top-N analysis results with the highest similarity; Step 3: Based on the user's historical decision-making preferences, generate at least 3 customized decision-making plans and push them out in order of suitability; the user profile module includes user industry attributes, decision-making needs, operating habits and other feature dimensions, and supports dynamic updates of user profiles.

[0012] As a further aspect of the present invention, the multi-scenario simulation verification of the decision simulation verification module includes a baseline scenario, an extreme scenario, and a random scenario, and the specific steps are as follows: Step 1: Set the parameter thresholds for each scenario; Step 2: Input the decision-making plan into the Monte Carlo simulation engine to simulate the implementation process of the plan under various scenarios and calculate the success rate of implementation. Profit Value Risk coefficient ; Step 3: Use the formula ; Quantitative scoring of decision-making options ( (For scoring, the maximum score is 100 points); Step 4: Select effective decision solutions with a score of ≥80 points and push them to the user; solutions with a score of less than 80 points are fed back to the multi-dimensional analysis layer, where the analysis model parameters are readjusted and new decision solutions are generated.

[0013] As a further aspect of the present invention, the decision-making effect collection module collects actual implementation data of the decision-making scheme through a combination of online data interface and offline manual input. The specific steps are as follows: Step 1: Collect real-time data on the implementation progress, effectiveness metrics, and resource consumption of the solution through online interfaces; Step 2: Collect unstructured data such as problem feedback and anomalies during the implementation of the solution through offline manual input; Step 3: Clean and standardize the collected data, remove outliers and missing values ​​to form a complete decision-making effect dataset; actual implementation data includes the progress of the solution implementation, implementation effect indicators, problem feedback information, etc.

[0014] As a further aspect of the present invention, the gradient descent algorithm of the model iterative optimization module satisfies formula (3), and the specific iterative steps are as follows: Formula (3):

[0015] in, For the first The parameters / weight coefficients of the fusion analysis model are used in the next iteration. For learning rate, loss function exist The gradient at the point is given, and the loss function is the mean square error between the predicted and actual values ​​of the decision scheme. Iteration steps: Step 1: Initialize model parameters and weight coefficients, set the learning rate η=0.01, and the number of iterations to 100; Step 2: Input the decision outcome dataset into the fusion analysis model and calculate the loss function. and gradient ; Step 3: Update the model parameters and weight coefficients according to formula (3) to complete one iteration; Step 4: After each iteration, use a test set to verify the model's analysis accuracy. When the improvement rate of analysis accuracy is <0.1%, stop the iteration and synchronize the optimized parameters to the edge and cloud models.

[0016] As a further aspect of the present invention, the resource elastic scheduling module of the cloud computing platform layer is based on Kubernetes container orchestration technology, and the specific scheduling steps are as follows: Step 1: Real-time collection of resource utilization rates of edge and cloud nodes; Step 2: Set resource allocation thresholds (edge ​​node utilization ≥ 80%, cloud node utilization ≥ 85%). Step 3: When the edge node utilization rate is ≥80%, automatically migrate some storage / computing tasks to the cloud; when the cloud node utilization rate is ≥85%, migrate non-core analysis tasks to the edge node. The threshold for resource allocation can be set by the user and can be dynamically adjusted according to the actual application scenario.

[0017] Compared with related technologies, the cloud computing-based intelligent archival storage and analysis decision support system provided by this invention has the following beneficial effects: This invention employs an edge-cloud distributed storage architecture, breaking the limitations of traditional single-storage models. Edge computing nodes enable local preprocessing of raw archival data and local storage for frequently accessed archives, while cloud core nodes achieve centralized management and large-scale computation of massive archives. A resource elastic scheduling module dynamically allocates computing and storage resources, effectively addressing the technical pain points of local storage resource bottlenecks and high latency in pure cloud storage. This achieves elastic scaling of archival storage, adapting to dynamic changes in the amount of archival data in different scenarios. Simultaneously, a heterogeneous storage module categorizes and stores structured, semi-structured, and unstructured archival data, coupled with the SM4 symmetric encryption algorithm and a multi-replica disaster recovery backup mechanism in different locations, ensuring that archival data is not lost or leaked from the storage perspective, thus building a solid data security defense. Furthermore, the unique vector encoding generated by the archival feature encoding module enables rapid retrieval and association of archives, improving retrieval efficiency by over 60% compared to traditional archival retrieval methods. This significantly reduces the manual costs of archival management and enhances the standardization and efficiency of archival management.

[0018] This invention overcomes the shortcomings of traditional, simplistic, and superficial archival analysis by constructing a fusion analysis model comprised of statistical analysis, machine learning prediction, and deep learning association analysis. Through the analytic hierarchy process (AHP), it scientifically allocates the weights of each sub-model, achieving a multi-dimensional combination of explicit association mining, trend prediction, and implicit association mining. The statistical analysis sub-model completes basic data statistics and explicit association rule mining, providing foundational data support for decision-making. The machine learning prediction sub-model, based on the XGBoost algorithm, accurately predicts the changing trends of archival data, providing a scientific basis for resource allocation and personnel scheduling. The deep learning association analysis sub-model breaks down data silos through graph neural networks (GNNs), uncovering implicit relationships between different types and scenarios of archives, fully releasing the potential value of archival data. Compared to existing analysis methods, this system improves analysis accuracy by more than 35%, providing users with more comprehensive and accurate data analysis results and contributing to more scientific decision-making.

[0019] This invention overcomes the weaknesses of traditional archival systems in decision support. By combining a user profiling module and a collaborative filtering algorithm, it pushes customized decision-making solutions based on different users' industry attributes, decision-making needs, and operating habits, achieving personalized services tailored to each user and adapting to the decision-making needs of multiple fields and positions. Simultaneously, the decision simulation verification module, based on a Monte Carlo simulation engine, simulates and verifies decision-making solutions in benchmark, extreme, and random scenarios. Quantitative scoring is used to filter effective solutions and eliminate decision suggestions with low feasibility, ensuring the scientific validity and implementability of decision-making solutions from the outset. Practical application verification shows that the success rate of the decision-making solutions pushed by this system reaches over 85%, significantly reducing the risk of decision-making errors and improving decision-making efficiency and quality.

[0020] This invention constructs a closed-loop optimization mechanism of "decision-feedback-iteration." Through a decision-making effect collection module, it comprehensively collects actual implementation data of decision-making schemes, including implementation progress, implementation effects, and problem feedback, providing real and comprehensive data support for model iteration. The model iteration optimization module, based on the gradient descent algorithm, uses the mean squared error between predicted and actual implementation values ​​as the loss function to dynamically update the parameters of the fusion analysis model and the weight coefficients of sub-models, achieving self-optimization of the system's analytical and decision support capabilities. Without manual intervention, the system can continuously iterate and upgrade according to changes in actual application scenarios, adapting to the needs of document management and decision-making in different fields and at different stages, extending the system's lifecycle, and enhancing its long-term service value and market adaptability. Attached Figure Description

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic diagram of the five-layer architecture of the cloud-based intelligent archival storage and analysis decision support system of the present invention; Figure 2 This is a schematic diagram of the closed-loop workflow of the system of the present invention; Figure 3 The feature extraction flowchart of the archive feature encoding module provided by this invention; Figure 4 A flowchart for determining the weight coefficients of each sub-model in the fusion analysis model using the analytic hierarchy process provided in this invention; Figure 5 The flowchart provided by this invention illustrates how a collaborative filtering algorithm can push customized content to different users. Figure 6 A flowchart of the multi-scenario simulation verification of the decision simulation verification module provided by the present invention; Figure 7 A flowchart of the decision-making effect acquisition module provided by the present invention; Figure 8The flowchart for model iteration provided by this invention; Figure 9 This invention provides a flowchart of the resource elastic scheduling process for the cloud computing platform layer. Detailed Implementation

[0023] Please refer to the following: Figure 1 and Figure 9 ,in, Figure 1 This is a schematic diagram of the five-layer architecture of the cloud-based intelligent archival storage and analysis decision support system of the present invention; Figure 2 This is a schematic diagram of the closed-loop workflow of the system of the present invention; Figure 3 The feature extraction flowchart of the archive feature encoding module provided by this invention; Figure 4 A flowchart for determining the weight coefficients of each sub-model in the fusion analysis model using the analytic hierarchy process provided in this invention; Figure 5 The flowchart provided by this invention illustrates how a collaborative filtering algorithm can push customized content to different users. Figure 6 A flowchart of the multi-scenario simulation verification of the decision simulation verification module provided by the present invention; Figure 7 A flowchart of the decision-making effect acquisition module provided by the present invention; Figure 8 The flowchart for model iteration provided by this invention; Figure 9 This invention provides a flowchart of the resource elastic scheduling process for the cloud computing platform layer. The cloud-based intelligent archival storage and analysis decision support system comprises a cloud computing platform layer, an intelligent archival storage layer, a multi-dimensional analysis layer, a decision support layer, and a feedback optimization layer. This five-layer architecture achieves bidirectional data interaction through cloud communication protocols, forming a closed-loop system of "storage-analysis-decision-feedback-iteration." To address the technical shortcomings of existing systems, such as scattered storage, simplistic analysis, unscientific decision-making, and lack of self-optimization capabilities; The cloud computing platform layer adopts a distributed architecture of edge computing nodes and cloud core nodes. Edge computing nodes are used for preprocessing and local storage of raw archival data, while cloud core nodes are used for centralized storage of archival data, model training, and large-scale computing. Edge computing nodes and cloud core nodes achieve real-time interaction of heterogeneous data through a data synchronization protocol, and are configured with a resource elastic scheduling module to dynamically allocate computing and storage resources; The intelligent archive storage layer includes a heterogeneous storage module and an archive feature encoding module. The heterogeneous storage module is adapted to the classified storage of structured, semi-structured, and unstructured archive data. The archive feature encoding module extracts features and generates unique codes for archive data by combining an improved TF-IDF algorithm with a deep convolutional neural network, enabling rapid retrieval and association of archives. The multi-dimensional analysis layer has a built-in fusion analysis model, which is composed of a weighted fusion of a statistical analysis sub-model, a machine learning prediction sub-model, and a deep learning correlation analysis sub-model. The weight coefficients of each sub-model are determined by the analytic hierarchy process, so as to achieve multi-dimensional and high-precision archival data analysis. The decision support layer includes a personalized decision push module and a decision simulation verification module. The decision simulation verification module uses a Monte Carlo simulation engine to perform multi-scenario simulation verification of the generated decision scheme to ensure the feasibility of the decision. The feedback optimization layer includes a decision effect acquisition module and a model iteration optimization module. The model iteration optimization module dynamically iterates the parameters and weight coefficients of the fusion analysis model based on the gradient descent algorithm to achieve self-optimization of the system's analysis capability and decision support capability.

[0024] The heterogeneous storage module includes a relational database, a non-relational database, and a distributed file system, which are used to store archive metadata, semi-structured archive tag data, and unstructured raw archive data (documents, pictures, audio, and video), respectively. Each storage unit is equipped with a data encryption module and a disaster recovery backup module. The data encryption module uses the SM4 symmetric encryption algorithm to encrypt and store the archive data, and the disaster recovery backup module uses an off-site multi-copy backup mechanism to ensure that the archive data is not lost or leaked.

[0025] The feature extraction process of the archive feature encoding module satisfies formulas (1) and (2), and the specific steps are as follows: Step 1: Extract shallow text features from the archive using the improved TF-IDF algorithm, and calculate the formula (1):

[0026] in, For the first The first file The word frequency of each feature word The total number of archives. For including the first The number of files for each feature word; Step 2: Input the shallow text features into the deep convolutional neural network to extract the archive depth features and calculate formula (2): ; in, For convolution kernel weights, For bias terms, For convolution operations, The depth features extracted by the deep convolutional neural network; Step 3: Put and The features are concatenated to obtain the fused features. ,right Normalization is performed to generate a unique 128-dimensional vector code, which serves as a unique identifier for each document.

[0027] The process of determining the weight coefficients of each sub-model in the fusion analysis model using the analytic hierarchy process is as follows: Step 1: Construct a judgment matrix. Based on the analysis accuracy, applicable scenarios, and computational efficiency of each sub-model, determine the importance comparison relationship between each sub-model and construct a 3×3 judgment matrix. Step 2: Consistency check, calculate the consistency index; ,in, To determine the largest eigenvalue of a matrix, Query the average random consistency index for the number of sub-models. Calculate the random consistency ratio ,when When <0.1, the judgment matrix satisfies the consistency requirement; Step 3: Weight calculation. Calculate the weight coefficients of each sub-model using the eigenvalue method, ensuring that the sum of the weight coefficients is 1. Finally, determine the weights as follows: statistical analysis sub-model weight 0.2, machine learning prediction sub-model weight 0.3, and deep learning association analysis sub-model weight 0.5.

[0028] The statistical analysis sub-model is used to realize basic statistics, trend analysis and association rule mining of archival data, and uses the Apriori algorithm to mine explicit association relationships between archival data; The machine learning prediction sub-model uses the XGBoost algorithm to build a trend prediction model for archival data and optimizes the model parameters (learning rate, tree depth, number of leaf nodes) through grid search. The deep learning association analysis sub-model uses graph neural networks (GNNs) to mine the implicit relationships between archival data. It constructs an archival association graph with unique vector codes of archives as nodes and the degree of association between archives as edges.

[0029] The personalized decision-making push module, based on the user profile module and profile analysis results, uses a collaborative filtering algorithm to push customized decision-making solutions to different users. The specific steps are as follows: Step 1: The user profiling module collects user characteristics such as industry attributes, decision-making needs, and operating habits to construct user feature vectors; Step 2: Calculate the similarity between the user feature vector and the feature vector of the analysis results, and select the Top-N analysis results with the highest similarity; Step 3: Based on the user's historical decision-making preferences, generate at least 3 customized decision-making plans and push them out in order of suitability; the user profile module includes user industry attributes, decision-making needs, operating habits and other feature dimensions, and supports dynamic updates of user profiles.

[0030] The multi-scenario simulation verification of the decision simulation verification module includes a baseline scenario, an extreme scenario, and a random scenario. The specific steps are as follows: Step 1: Set parameter thresholds for each scenario (the baseline scenario uses regular business parameters, the extreme scenario uses extreme business load, and the random scenario uses random fluctuation parameters). Step 2: Input the decision-making plan into the Monte Carlo simulation engine to simulate the implementation process of the plan under various scenarios and calculate the success rate of implementation. Profit Value Risk coefficient ; Step 3: Use the formula ; Quantitative scoring of decision-making options ( (For scoring, the maximum score is 100 points); Step 4: Select effective decision solutions with a score of ≥80 points and push them to the user; solutions with a score of less than 80 points are fed back to the multi-dimensional analysis layer, where the analysis model parameters are readjusted and new decision solutions are generated.

[0031] The decision-making effect collection module collects actual implementation data of the decision-making scheme through a combination of online data interfaces and offline manual data entry. The specific steps are as follows: Step 1: Collect real-time data on the implementation progress, effectiveness metrics (efficiency improvement rate, satisfaction, etc.), and resource consumption of the solution through online interfaces; Step 2: Collect unstructured data such as problem feedback and anomalies during the implementation of the solution through offline manual input; Step 3: Clean and standardize the collected data, remove outliers and missing values ​​to form a complete decision-making effect dataset; actual implementation data includes the progress of the solution implementation, implementation effect indicators, problem feedback information, etc.

[0032] The gradient descent algorithm of the model iterative optimization module satisfies formula (3), and the specific iterative steps are as follows: Formula (3):

[0033] in, For the first The parameters / weight coefficients of the fusion analysis model are used in the next iteration. For learning rate, loss function exist The gradient at the point is given, and the loss function is the mean square error between the predicted and actual values ​​of the decision scheme. Iteration steps: Step 1: Initialize model parameters and weight coefficients, set the learning rate η=0.01, and the number of iterations to 100; Step 2: Input the decision outcome dataset into the fusion analysis model and calculate the loss function. and gradient ; Step 3: Update the model parameters and weight coefficients according to formula (3) to complete one iteration; Step 4: After each iteration, use a test set to verify the model's analysis accuracy. When the improvement rate of analysis accuracy is <0.1%, stop the iteration and synchronize the optimized parameters to the edge and cloud models.

[0034] The resource elastic scheduling module of the cloud computing platform layer is based on Kubernetes container orchestration technology. The specific scheduling steps are as follows: Step 1: Collect real-time resource utilization (CPU, memory, storage) of edge and cloud nodes. Step 2: Set resource allocation thresholds (edge ​​node utilization ≥ 80%, cloud node utilization ≥ 85%). Step 3: When the edge node utilization rate is ≥80%, automatically migrate some storage / computing tasks to the cloud; when the cloud node utilization rate is ≥85%, migrate non-core analysis tasks to the edge node. The threshold for resource allocation can be set by the user and can be dynamically adjusted according to the actual application scenario.

[0035] The cloud-based intelligent archival storage and analysis decision support system of this invention is deployed on a government cloud platform, adopting an "edge-cloud" distributed deployment mode to balance the needs of local data processing and centralized management. The specific deployment scheme is as follows: Edge computing node deployment: One edge computing node is deployed in each district and county government service center. Each node is equipped with four edge servers (CPU: Intel Xeon E5-2690, memory: 64GB, storage: 2TB SSD) to collect, preprocess, store locally and perform preliminary analysis of the original data of government archives in each district and county, reduce data transmission latency and improve local business response speed. Among them, the core nodes are deployed in the cloud: One cloud core node is deployed in the municipal government cloud data center, configured with 10 high-performance cloud servers (CPU: Intel Xeon Gold 6248, memory: 128GB, storage: 10TB SSD). It adopts a clustered deployment mode and is responsible for the centralized storage of all municipal government archives, full training and iteration of the fusion analysis model, large-scale data analysis and computing, and municipal government decision support. Among them, network communication deployment: Edge computing nodes and cloud core nodes achieve real-time data synchronization through the government intranet. The cloud communication protocol adopts the MQTT protocol (Message Transmission Protocol) to ensure the real-time performance, stability and security of data transmission, and the data synchronization delay is controlled within 1 second. Among them, hardware and software support: The system's hardware support includes storage arrays (for massive archive data backup), network switches (for inter-node communication), and terminal devices (computers, tablets, and mobile phones for user operation and decision-making viewing); software support includes a Linux operating system (CentOS 8.0), a Kubernetes container orchestration platform (for elastic resource scheduling), a Python data analysis environment (Python 3.9), a TensorFlow / PyTorch deep learning framework (for model training), a relational database MySQL (8.0, for storing archive metadata), a non-relational database MongoDB (5.0, for storing semi-structured labeled data), and a distributed file system HDFS (3.3, for storing unstructured raw data). The cloud computing platform layer is the foundational support layer of the system, responsible for resource scheduling, data transmission, and node coordination. It also includes a resource elastic scheduling module and a cloud communication module. The specific implementation is as follows: (1) Edge computing node: mainly responsible for the collection and preprocessing of original archival data. The collection methods include online interface docking (interfacing with the government business systems of various districts and counties) and offline import (USB flash drive, hard drive import). The preprocessing process includes deduplication (deleting duplicate archival data, using hash value comparison method), cleaning (removing archival data with missing value ratio ≥30% and too many outliers, using Z-score method to identify outliers), and standardization (converting archival data of different formats and units into a unified format, such as unifying documents to PDF format and time to YYYY-MM-DD format); at the same time, it is responsible for the local storage of core archival data of various districts and counties, running a lightweight version of the fusion analysis model (simplifying the calculation process, retaining the core analysis function), realizing the preliminary analysis of archival data, and reducing the computing pressure on the cloud.

[0036] (2) Cloud core node: responsible for the centralized storage of municipal government archives, receiving archive data synchronized from each edge node and managing it in a unified manner; running the full version of the fusion analysis model and carrying out large-scale data analysis operations (such as city-wide archive trend analysis and cross-district archive correlation mining); responsible for the training, iteration and updating of the fusion analysis model, and synchronizing the optimized model to each edge node; providing comprehensive data analysis and decision support for municipal government decision-making.

[0037] (3) Resource Elastic Scheduling Module: Based on Kubernetes container orchestration technology, it collects the resource utilization (CPU, memory, storage) of edge and cloud nodes in real time, and dynamically allocates computing and storage resources according to the preset resource allocation threshold. When the amount of archive data in a certain district or county suddenly increases (such as the centralized archiving of approval archives at the end of the quarter), and the resource utilization of the edge node is ≥80%, it automatically migrates some storage / computing tasks to the cloud. When the resource utilization of the cloud node is ≥85%, it migrates non-core analysis tasks (such as historical archive archiving analysis) to the edge node to achieve optimal resource allocation and avoid resource waste or resource bottlenecks.

[0038] (4) Cloud communication module: The MQTT protocol is used to realize bidirectional data interaction between the edge and the cloud, and supports batch data synchronization and real-time data transmission; a data encryption transmission module is configured, and the SSL / TLS encryption protocol is used to encrypt the transmitted data to prevent the data from being stolen or tampered with during transmission; a data verification mechanism is set up to verify the integrity of the synchronized data and ensure the accuracy of data transmission.

[0039] The intelligent archival storage layer is responsible for the classification, storage, feature extraction, and unique identification of archival data. It includes a heterogeneous storage module and an archival feature encoding module. The specific implementation is as follows: (1) Heterogeneous storage module: It adopts a heterogeneous storage architecture of "relational database + non-relational database + distributed file system" to adapt to the storage needs of different types of archive data: ① Relational database MySQL: Used to store archival metadata, including structured data such as archival number, creation time, creating unit, archival type, archivist, and retention period. It uses a primary key index (archival number) to improve query efficiency. ② MongoDB (a non-relational database): Used to store semi-structured tag data of archives, including archive keywords, topics, associated archive numbers, classification tags, etc., supporting flexible tag expansion for easy classification and retrieval of archives; ③ HDFS (distributed file system): Used to store unstructured raw data of archives, including administrative approval documents, law enforcement videos, public service images, audio recordings, etc., using a block storage mode to support efficient storage and fast retrieval of massive amounts of data; Each storage unit is equipped with an SM4 symmetric encryption module to encrypt archive data (encryption keys are managed uniformly by the system administrator and changed regularly); An off-site disaster recovery backup module is configured, using a "local backup + off-site multi-copy backup" mechanism, with local backups performed daily and off-site backups performed weekly to ensure that archive data is not lost or leaked.

[0040] (2) Archive Feature Encoding Module: It is used to extract features and generate unique codes for the preprocessed archive data, enabling fast retrieval and correlation analysis of archives. The specific implementation steps are as follows: ① Shallow Feature Extraction: The text features of the archives are extracted through an improved TF-IDF algorithm. The calculation formula is , where is the th feature word frequency in the th archive, is the total number of archives, is the number of archives containing the th feature word; the improvement lies in filtering stop words (such as "de", "di", "de") to enhance the accuracy of feature extraction; ② Deep Feature Extraction: The shallow text features are input into a deep convolutional neural network ( ), and the deep features of the archives are extracted through convolutional operations. The calculation formula is where is the convolutional kernel weight (randomly generated initially and optimized through model training), is the bias term, is the convolutional operation, is the deep feature vector; ③ Feature Fusion and Encoding: The shallow features and the deep features are concatenated to obtain the fused feature . is normalized (mapping the feature values to the [0, 1] interval) to generate a 128-dimensional unique vector code, which serves as the unique identifier for each archive. The code is associated with the archive metadata for fast retrieval, correlation analysis, and uniqueness verification of the archives; Multi-dimensional Analysis Layer This layer is the core analysis layer of the system, responsible for multi-dimensional intelligent analysis of archive data. It incorporates a fusion analysis model, which is composed of a weighted fusion of a statistical analysis sub-model, a machine learning prediction sub-model, and a deep learning correlation analysis sub-model; The specific implementation is as follows: (1) Determination of the Weights of the Fusion Analysis Model: The weight coefficients of each sub-model are determined through the Analytic Hierarchy Process. The specific steps are as follows: ① Construction of the Judgment Matrix: Based on the analysis accuracy, applicable scenarios, and operation efficiency of each sub-model, the importance comparison relationship between each sub-model is determined, and a 3×3 judgment matrix (both rows and columns are the statistical analysis sub-model, the machine learning prediction sub-model, and the deep learning correlation analysis sub-model) is constructed. The judgment matrix is as follows: |Statistical Analysis Sub-model|Machine Learning Prediction|Deep Learning Association Analysis|Statistical Analysis|1|1 / 3|1 / 5||Machine Learning Prediction|3|1|1 / 3||Deep Learning Association Analysis|5|3|1| ② Consistency test: Calculate the largest eigenvalue of the judgment matrix. Consistency indicators Query the average random consistency index of the 3rd order judgment matrix. =0.58, calculate the random consistency ratio. =0.019 / 0.58≈0.033<0.1, therefore the matrix satisfies the consistency requirement; ③ Weight Calculation: The weight coefficients of each sub-model are calculated using the eigenvalue method. The final weights are: statistical analysis sub-model 0.2, machine learning prediction sub-model 0.3, and deep learning association analysis sub-model 0.5. The sum of the weight coefficients is 1.

[0041] (2) Specific implementation of each sub-model: ① Statistical Analysis Sub-model: The Apriori algorithm is used to mine explicit relationships between archival data, realizing basic statistics, trend analysis, and association rule mining of archival data. Basic statistics include the monthly number of administrative approvals, the proportion of public service types, and the number of archives filed in each district and county. Trend analysis uses a linear regression algorithm to analyze the quantitative change trends of law enforcement archives, public service archives, etc. in the past three years and predict future changes. Association rule mining uses the Apriori algorithm to mine relationships such as "the relationship between administrative approval archives and public service archives" and "the relationship between a certain type of law enforcement archive and a specific approval item", and outputs an association rule report.

[0042] ② Machine Learning Prediction Sub-model: The XGBoost algorithm is used to construct a trend prediction model for archival data. Historical archival data (archival data of various districts and counties and various types over the past 5 years) is used as the training set. Features such as the number of archives, archiving time, and business type are used. The model parameters are optimized by grid search method (learning rate = 0.05, tree depth = 6, number of leaf nodes = 32) to predict the amount of archives to be archived in a future period (such as quarter or year) and the growth trend of specific types of archives, so as to provide predictive support for the allocation of archival storage resources and personnel configuration.

[0043] ③ Deep Learning Association Analysis Sub-model: Graph Neural Network (GNN) is used to mine implicit relationships between archival data. The unique vector code of the archives is used as the node, and the degree of association between archives (calculated through feature similarity) is used as the edge to construct an archival association graph. Through the node embedding method of the GNN model, implicit relationships between archives of different districts, counties, types, and times are mined (such as the potential association between law enforcement archives of one district and county and livelihood archives of another district and county), breaking the island effect of archival data and fully mining the potential value of archival data.

[0044] (3) Output of fusion analysis results: The analysis results of the three sub-models are weighted and fused using the weight coefficients of each sub-model to obtain the final multi-dimensional analysis results. The fusion formula is as follows: ,in For statistical analysis results, For machine learning prediction results, The results of deep learning correlation analysis are presented in the form of charts (line charts, bar charts, correlation network diagrams) and text reports to facilitate subsequent decision support.

[0045] 4. Decision Support Layer: This layer is responsible for generating, simulating, verifying, and pushing personalized decision-making solutions. It includes a personalized decision push module, a decision simulation verification module, and a user profiling module, specifically implemented as follows: (1) User profile module: Collect user industry attributes (such as administrative approval positions, public service positions, and decision management positions in the government sector), decision needs (such as efficiency improvement, resource optimization, and risk prevention and control), and operating habits (such as commonly used analysis dimensions and decision preferences) to construct user feature vectors; support dynamic updates of user profiles, and adjust user feature vectors in real time based on user operating behavior and decision feedback to improve the adaptability of decision solutions.

[0046] (2) Personalized Decision Push Module: Based on user profiles and the analysis results of multi-dimensional analysis layers, a collaborative filtering algorithm is used to push customized decision-making solutions to different users. The specific steps are as follows: ① Calculate the similarity between the user feature vector and the feature vector of the multi-dimensional analysis results (using the cosine similarity algorithm). ② Filter the Top-N analysis results with the highest similarity (N=5), and generate at least 3 customized decision-making schemes based on the user's historical decision-making preferences; ③ Sort the solutions from highest to lowest suitability and push them to the corresponding users; for example, push decision-making management positions with solutions to improve the efficiency of municipal government services and optimize the allocation of archival resources, and push administrative approval positions with solutions to optimize certain approval business processes and improve the efficiency of archival retrieval.

[0047] (3) Decision simulation verification module: Based on the Monte Carlo simulation engine, the generated decision scheme is simulated and verified in multiple scenarios to ensure the scientificity and feasibility of the decision scheme. The specific steps are as follows: ① Set parameter thresholds for three scenarios: baseline scenario (normal business load, such as daily file archiving volume and business processing efficiency), extreme scenario (extreme business load, such as annual centralized file archiving and sudden business surge), and random scenario (random fluctuation parameters, such as random fluctuations in file archiving volume and random changes in business processing efficiency). ② Input the decision-making plan into the Monte Carlo simulation engine, simulate the implementation process of the plan in various scenarios, run 1000 simulations, and calculate the success rate of the plan implementation in each scenario. Profit Value (e.g., efficiency improvement rate, resource savings), risk coefficient (e.g., difficulty in implementing the plan, potential risks); ③ Use formula Quantitative scoring of decision-making options ( (For scoring, out of 100 points). ④ Filter valid decision solutions with a score of ≥80 and push them to users in descending order of score; solutions with a score of less than 80 are fed back to the multi-dimensional analysis layer, the parameters of the fusion analysis model (such as sub-model weights and algorithm parameters) are readjusted, and decision solutions are regenerated until the score requirements are met.

[0048] 5. Feedback Optimization Layer: This layer is the core layer for the system to achieve self-optimization. It is responsible for collecting decision-making results and iteratively optimizing the model. It includes a decision-making results collection module and a model iterative optimization module, which are implemented as follows: (1) Decision-making effect collection module: The actual implementation data of the decision-making plan is collected through a combination of online data interface and offline manual data entry. The specific steps are as follows: ① Online data collection: Through the online data interface of the government business system, collect in real time the implementation progress of decision-making plans (such as the completion rate of plan implementation and the achievement of phased goals), implementation effect indicators (such as the improvement rate of government service efficiency, public satisfaction, and reduction in resource consumption), and resource consumption (such as CPU utilization and storage usage). ② Offline data collection: Collect unstructured data such as feedback on problems during the implementation of the solution (e.g., difficulty in implementation, existing vulnerabilities) and abnormal situations (e.g., data anomalies, system failures) through offline manual data entry. ③ Data processing: Clean the collected data (remove outliers and missing values) and standardize it (unify data format and units) to form a complete decision effect dataset for model iteration and optimization.

[0049] (2) Model Iteration Optimization Module: Based on the gradient descent algorithm, the mean square error between the predicted and actual values ​​of the decision scheme is used as the loss function to dynamically iterate the parameters of the fusion analysis model and the weight coefficients of each sub-model, thereby achieving self-optimization of the system's analytical and decision support capabilities. The specific iteration steps are as follows: ① Initialization parameters: Set the initial parameters of the fusion analysis model (such as the convolution kernel weights of CNN and the model parameters of XGBoost) and the weight coefficients of each sub-model. The learning rate η=0.01, the number of iterations is 100, and the stopping iteration threshold is (analysis accuracy improvement rate <0.1%). ② Calculate the loss function: Input the decision outcome dataset into the fusion analysis model, calculate the mean squared error between the predicted and actual values ​​of the decision scheme, and use this as the loss function. The formula is (in For data volume, This is the actual implemented value. (Predicted value); ③ Gradient calculation and parameter update: Calculate the loss function In the current parameters gradient at According to the gradient descent algorithm formula 5. Update the model parameters and the weight coefficients of each sub-model to complete one iteration; ④ Validation and Iteration Cessation: After each iteration, the analytical accuracy of the fusion analysis model is validated using a test set (30% of the decision effect dataset), and the accuracy improvement rate is calculated. When the accuracy improvement rate is <0.1%, the iteration is stopped, and the optimized model parameters and weight coefficients are synchronized to the fusion analysis model at the edge and in the cloud to achieve system self-optimization. If the stopping threshold is not reached, steps ②-④ are repeated until the requirements are met.

[0050] The working principle of the cloud-based intelligent archival storage and analysis decision support system is as follows: The working principle of the cloud-based intelligent archival storage and analysis decision support system of this invention is a closed-loop process of "archival access - feature encoding - cloud-based elastic storage - multi-dimensional intelligent analysis - personalized decision generation - simulation verification - decision push - effect collection - model iteration". Each link is closely connected and works collaboratively to ensure the system's intelligence, efficiency and self-optimization capabilities. The specific detailed steps are as follows: S1: Access and Preprocessing of Raw Archive Data The archive data (structured, semi-structured, and unstructured) of the government service centers in each district and county are connected to the system's edge computing nodes through online interfaces (connecting with government business systems) and offline import (USB flash drives, hard drives); Structured data includes archival metadata, semi-structured data includes archival tag information, and unstructured data includes documents, videos, images, etc. The preprocessing module of the edge computing node performs deduplication, cleaning, and standardization on the raw data: (1) Deduplication: The hash value comparison method is used to calculate the hash value of each file, compare the hash values, delete duplicate file data, and avoid data redundancy; (2) Cleaning: The Z-score method is used to identify outliers and remove archive data with missing values ​​≥30% or too many outliers; for archive data with missing values ​​<30%, the missing values ​​are filled by means such as mean filling and median filling. (3) Standardization: Convert archival data of different formats and units into a unified format, such as documents in PDF format, images in JPG format, time in YYYY-MM-DD format, and numerical units in standard units; The preprocessed archival data is divided into edge local storage data and cloud synchronized data. Edge local storage data consists of core archival data from each district and county (such as approval archives that are frequently accessed daily), while cloud synchronized data consists of archival data that needs to be centrally analyzed throughout the city (such as cross-district and county related archives and annual statistical archives).

[0051] S2: Archival Characterization Coding and Unique Identifier Generation The archive feature encoding module in the intelligent archive storage layer extracts text features from the preprocessed archive data using an improved TF-IDF algorithm. The calculation formula is as follows: ,in, For the first The first file The word frequency of each feature word The total number of archives. For including the first The number of files for each feature word; filtering stop words improves the accuracy of feature extraction, resulting in shallow text features. ; Deep feature extraction: extracting shallow text features Input a deep convolutional neural network (CNN), and extract the depth features of the file through convolution operations. The calculation formula is as follows: ,in For convolution kernel weights, For bias terms, Convolution operations are performed to obtain deep features.

[0052] Feature fusion and encoding: converting shallow features with depth features Feature concatenation is performed to obtain fused features. ;right\ Normalization is performed to map the feature values ​​to the [0,1] interval; a 128-dimensional unique vector code is generated as a unique identifier for each archive. The code is associated with the archive metadata, thus completing the feature encoding of the archive.

[0053] S3: Cloud Computing Elastic Storage and Data Management The heterogeneous storage module of the intelligent archival storage layer classifies and stores the encoded archival data into corresponding storage units at the edge / cloud according to the type of archival data: (1) Store the archive metadata (structured) in a MySQL database and create a primary key index to improve query efficiency; (2) The semi-structured tag data of the archives is stored in the MongoDB database, which supports flexible tag expansion; (3) The unstructured raw data of the archives is stored in the HDFS distributed file system, which adopts block storage and supports the efficient storage and reading of massive data; 2. Data security guarantee: All archive data is encrypted by the SM4 symmetric encryption algorithm and stored. The encryption key is managed by the system administrator and changed regularly; an off-site disaster recovery backup module is configured to realize local backup and off-site multi-copy backup to ensure that the archive data is not lost or leaked; The resource elastic scheduling module at the cloud computing platform layer, based on Kubernetes container orchestration technology, collects the resource utilization (CPU, memory, storage) of edge and cloud nodes in real time. When the resource utilization of the edge node is ≥80%, it automatically migrates some storage / computing tasks to the cloud. When the resource utilization of the cloud node is ≥85%, it migrates non-core analysis tasks to the edge node. The resource allocation threshold can be set by the user to achieve optimal resource configuration. Edge computing nodes and cloud core nodes use the MQTT protocol to achieve real-time synchronization of heterogeneous data, with data synchronization latency controlled within 1 second, ensuring data consistency between the edge and the cloud.

[0054] S4: Multi-dimensional Intelligent Analysis and Feature Mining Sub-model analysis: A multi-dimensional analysis layer fusion analysis model that calls three sub-models to analyze the stored archive data respectively: (1) Statistical analysis sub-model: The Apriori algorithm is used to perform basic statistics on the archive data (monthly / quarterly / annual archive filing volume, proportion of each type of archive, etc.), trend analysis (linear regression algorithm to analyze the trend of archive quantity change), and association rule mining (mining explicit association relationships between archives), and output the basic analysis results Fstat; (2) Machine learning prediction sub-model: The XGBoost algorithm is used to construct a trend prediction model with historical archive data as the training set to predict the amount of archives to be archived in a certain period in the future, the growth trend of specific types of archives, etc., and output the prediction result Fml; (3) Deep learning association analysis sub-model: Using graph neural network GNN, with unique vector encoding of archives as nodes and the degree of association between archives as edges, an archive association graph is constructed to mine the implicit association between archives and output the association analysis results Fdnn; Fusion Analysis: Using weighted coefficients determined by the Analytic Hierarchy Process (statistical analysis 0.2, machine learning prediction 0.3, deep learning association analysis 0.5), the analysis results of the three sub-models are weighted and fused. The fusion formula is as follows: This yields the final multi-dimensional analysis results; The results of the fusion analysis are output in the form of charts (line charts, bar charts, and relational network diagrams) and text reports, which are convenient for subsequent decision support modules to call.

[0055] S5: Personalized Decision-Making Scheme Generation and Multi-Scenario Simulation Verification 1. User Profile Construction: The user profile module in the decision support layer collects user characteristics such as industry attributes, decision-making needs, and operating habits to construct user feature vectors, which support dynamic updates; 2. Personalized Decision Generation: The personalized decision push module, based on user profiles and multi-dimensional analysis results, uses a collaborative filtering algorithm to calculate the similarity between the user's feature vector and the feature vector of the analysis results, selects the Top-N analysis results with the highest similarity, and generates at least 3 customized decision schemes by combining the user's historical decision preferences. 3. Multi-scenario simulation verification: The decision simulation verification module, based on the Monte Carlo simulation engine, performs multi-scenario simulation verification for each alternative decision scheme. (1) Set parameter thresholds for baseline scenario, extreme scenario, and random scenario; (2) Input the decision-making scheme, run 1000 simulations, and calculate the success rate of the scheme implementation under each scenario. Profit Value Risk coefficient ; (3) Using the formula

[0056] The decision-making options are quantitatively scored (out of 100 points). 4. Solution Screening: Valid decision solutions with a score of ≥80 are screened and sorted from highest to lowest score; solutions with a score below 80 are fed back to the multi-dimensional analysis layer, the parameters of the fusion analysis model are readjusted, and decision solutions are regenerated until the score requirements are met.

[0057] S6: Decision-making plan delivery and actual implementation effect collection 1. Decision Push: The system pushes effective decision solutions to corresponding users through terminal devices (computers, tablets, mobile phones), and users select and implement the decision solutions according to their actual needs; the system records the user's selection results and the implementation time of the solutions; 2. Results Collection: The decision-making results collection module of the feedback optimization layer collects actual implementation data of the decision-making plan through a combination of online data interfaces and offline manual data entry. (1) Online data collection: Real-time collection of the implementation progress of the plan, implementation effect indicators (efficiency improvement rate, public satisfaction, etc.), and resource consumption; (2) Offline data collection: Manually inputting unstructured data such as problem feedback and abnormal situations during the implementation of the solution; 3. Data processing: Clean and standardize the collected data, remove outliers and missing values ​​to form a complete decision effect dataset for model iteration and optimization.

[0058] S7: Model Iterative Optimization and System Self-Upgrading 1. Loss Function Construction: The model iterative optimization module in the feedback optimization layer uses the mean squared error between the predicted and actual values ​​of the decision scheme as the loss function. The formula is

[0059] Gradient calculation and parameter update: Calculate the loss function based on the gradient descent algorithm. In the current parameters gradient at According to the formula

[0060] (Learning rate η=0.01), update the parameters of the fusion analysis model and the weight coefficients of each sub-model; 3. Iterative verification: After each iteration, the analytical accuracy of the fusion analysis model is verified using a test set, and the improvement rate of analytical accuracy is calculated; 4. Stop Iteration and Synchronize: When the improvement rate of analysis accuracy is <0.1%, stop the iteration and synchronize the optimized model parameters and weight coefficients to the fusion analysis model of edge and cloud to realize the self-optimization of system analysis capability and decision support capability; if the stopping threshold is not reached, repeat steps 2-3 until the requirements are met. 5. Closed-loop formation: After completing the model iteration, the system enters the next round of the closed-loop process of "storage-analysis-decision-feedback-iteration" to continuously improve system performance and adapt to dynamic application scenarios.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments, or they can be used directly or indirectly, without departing from the principles and spirit of the invention. In other related technical fields, the scope of the invention is defined by the appended claims and their equivalents, and they are similarly included within the scope of patent protection of the invention.

Claims

1. A cloud computing-based intelligent archival storage and analysis decision support system, characterized in that, include: The five-layer architecture, consisting of a cloud computing platform layer, an intelligent archive storage layer, a multi-dimensional analysis layer, a decision support layer, and a feedback optimization layer, achieves bidirectional data interaction through cloud communication protocols, forming a closed-loop system of "storage-analysis-decision-feedback-iteration". The cloud computing platform layer adopts a distributed architecture of edge computing nodes and cloud core nodes. Edge computing nodes are used for preprocessing and local storage of raw archival data, while cloud core nodes are used for centralized storage of archival data, model training, and large-scale computing. Edge computing nodes and cloud core nodes achieve real-time interaction of heterogeneous data through a data synchronization protocol, and are configured with a resource elastic scheduling module to dynamically allocate computing and storage resources; The intelligent archival storage layer includes a heterogeneous storage module and an archival feature encoding module. The heterogeneous storage module is adapted to the classified storage of structured, semi-structured, and unstructured archival data. The archival feature encoding module extracts features and generates unique codes for archival data by combining an improved TF-IDF algorithm with a deep convolutional neural network, enabling rapid retrieval and association of archives. The multi-dimensional analysis layer has a built-in fusion analysis model, which is composed of a weighted fusion of a statistical analysis sub-model, a machine learning prediction sub-model, and a deep learning correlation analysis sub-model. The weight coefficients of each sub-model are determined by the analytic hierarchy process, enabling multi-dimensional and high-precision archival data analysis. The decision support layer includes a personalized decision push module and a decision simulation verification module. The decision simulation verification module uses a Monte Carlo simulation engine to perform multi-scenario simulation verification of the generated decision schemes to ensure the feasibility of the decisions. The feedback optimization layer includes a decision effect acquisition module and a model iteration optimization module. The model iteration optimization module dynamically iterates the parameters and weight coefficients of the fusion analysis model based on the gradient descent algorithm, thereby realizing the self-optimization of the system's analytical and decision support capabilities.

2. The cloud computing-based intelligent archival storage and analysis decision support system according to claim 1, characterized in that: The heterogeneous storage module includes a relational database, a non-relational database, and a distributed file system, which are used to store archive metadata, semi-structured archive tag data, and unstructured raw archive data, respectively. Each storage unit is equipped with a data encryption module and a disaster recovery backup module. The data encryption module uses the SM4 symmetric encryption algorithm to encrypt and store archive data, and the disaster recovery backup module uses an off-site multi-copy backup mechanism to ensure that archive data is not lost or leaked.

3. The cloud computing-based intelligent archival storage and analysis decision support system according to claim 1, characterized in that: The feature extraction process of the archive feature encoding module satisfies formulas (1) and (2), and the specific steps are as follows: Step 1: Extract shallow text features from the archive using the improved TF-IDF algorithm, and calculate the features using formula (1): ; in, For the first The first file The word frequency of each feature word The total number of archives. For including the first The number of files for each feature word; Step 2: Input the shallow text features into the deep convolutional neural network to extract the archive depth features and calculate formula (2): ; in, For convolution kernel weights, For bias terms, For convolution operations, The depth features extracted by the deep convolutional neural network; Step 3: Put and The features are concatenated to obtain the fused features. ,right Normalization is performed to generate a unique 128-dimensional vector code, which serves as a unique identifier for each document.

4. The cloud computing-based intelligent archival storage and analysis decision support system according to claim 1, characterized in that: The process of determining the weight coefficients of each sub-model in the fusion analysis model using the analytic hierarchy process is as follows: Step 1: Construct a judgment matrix. Based on the analysis accuracy, applicable scenarios, and computational efficiency of each sub-model, determine the importance comparison relationship between each sub-model and construct a 3×3 judgment matrix. Step 2: Consistency check, calculate the consistency index; ,in, To determine the largest eigenvalue of a matrix, Query the average random consistency index for the number of sub-models. Calculate the random consistency ratio ,when When <0.1, the judgment matrix satisfies the consistency requirement; Step 3: Weight calculation. Calculate the weight coefficients of each sub-model using the eigenvalue method, ensuring that the sum of the weight coefficients is 1. Finally, determine the weights as follows: statistical analysis sub-model weight 0.2, machine learning prediction sub-model weight 0.3, and deep learning association analysis sub-model weight 0.

5.

5. The cloud computing-based intelligent archival storage and analysis decision support system according to claim 1, characterized in that: The statistical analysis sub-model is used to perform basic statistics, trend analysis, and association rule mining of archival data, and the Apriori algorithm is used to mine explicit associations between archival data. The machine learning prediction sub-model uses the XGBoost algorithm to construct a trend prediction model for archival data and optimizes the model parameters using a grid search method. The deep learning association analysis sub-model uses graph neural networks (GNNs) to mine the implicit relationships between archival data. It constructs an archival association graph with unique vector codes of archives as nodes and the degree of association between archives as edges.

6. The cloud computing-based intelligent archival storage and analysis decision support system according to claim 1, characterized in that: The personalized decision-making recommendation module, based on user profiles and profile analysis results, uses a collaborative filtering algorithm to push customized decision-making solutions to different users. The specific steps are as follows: Step 1: The user profiling module collects user characteristics such as industry attributes, decision-making needs, and operating habits to construct user feature vectors; Step 2: Calculate the similarity between the user feature vector and the feature vector of the analysis results, and select the Top-N analysis results with the highest similarity; Step 3: Based on the user's historical decision-making preferences, generate at least 3 customized decision-making plans and push them out in order of suitability; the user profile module includes the user's industry attributes, decision-making needs, and operating habit characteristics, and supports dynamic updates of the user profile.

7. The cloud computing-based intelligent archival storage and analysis decision support system according to claim 1, characterized in that: The multi-scenario simulation verification module for decision-making simulation verification includes baseline scenarios, extreme scenarios, and random scenarios. The specific steps are as follows: Step 1: Set the parameter thresholds for each scenario; Step 2: Input the decision-making plan into the Monte Carlo simulation engine to simulate the implementation process of the plan under various scenarios and calculate the success rate of implementation. Profit Value Risk coefficient ; Step 3: Use the formula ; Quantitative scoring of decision-making options ( (For scoring, the maximum score is 100 points); Step 4: Select effective decision-making solutions with a score of ≥80 points and push them to the user; solutions with a score of less than 80 points are fed back to the multi-dimensional analysis layer, where the analysis model parameters are readjusted and new decision-making solutions are generated.

8. The cloud computing-based intelligent archival storage and analysis decision support system according to claim 1, characterized in that: The decision-making effect collection module collects actual implementation data of decision-making schemes through a combination of online data interfaces and offline manual data entry. The specific steps are as follows: Step 1: Collect real-time data on the implementation progress, effectiveness metrics, and resource consumption of the solution through online interfaces; Step 2: Collect unstructured data such as feedback on problems and anomalies during the implementation of the solution through offline manual input; Step 3: Clean and standardize the collected data, remove outliers and missing values, and form a complete decision effect dataset; actual implementation data includes the progress of solution implementation, implementation effect indicators, and problem feedback information.

9. The cloud computing-based intelligent archival storage and analysis decision support system according to claim 1, characterized in that: The gradient descent algorithm of the model iterative optimization module satisfies formula (3), and the specific iterative steps are as follows: Formula (3): ； in, For the first The parameters / weight coefficients of the fusion analysis model are used in the next iteration. For learning rate, For loss function exist The gradient at the point is given, and the loss function is the mean square error between the predicted and actual values ​​of the decision scheme. Iteration steps: Step 1: Initialize model parameters and weight coefficients, set the learning rate η=0.01, and the number of iterations to 100; Step 2: Input the decision outcome dataset into the fusion analysis model and calculate the loss function. and gradient ; Step 3: Update the model parameters and weight coefficients according to formula (3) to complete one iteration; Step 4: After each iteration, use a test set to verify the model's analysis accuracy. When the improvement rate of analysis accuracy is <0.1%, stop the iteration and synchronize the optimized parameters to the edge and cloud models.

10. The cloud computing-based intelligent archival storage and analysis decision support system according to any one of claims 1-9, characterized in that: The resource elastic scheduling module at the cloud computing platform layer is based on Kubernetes container orchestration technology. The specific scheduling steps are as follows: Step 1: Real-time collection of resource utilization rates of edge and cloud nodes; Step 2: Set resource allocation thresholds (edge ​​node utilization ≥ 80%, cloud node utilization ≥ 85%). Step 3: When the edge node utilization rate is ≥80%, automatically migrate some storage / computing tasks to the cloud; when the cloud node utilization rate is ≥85%, migrate non-core analysis tasks to the edge nodes. The threshold for resource allocation can be set by the user and can be dynamically adjusted according to the actual application scenario.

Images