Network additional storage intelligent data management method and system

Abstract

The invention relates to the technical field of network additional storage, and discloses a network additional storage intelligent data management method and system, and the method comprises the steps: monitoring a file system, capturing a newly added file in real time, modifying a file, judging a data type, carrying out analysis, carrying out a fusion decision mechanism through a rule analyzer, and carrying out the classification of data of which the classification confidence is lower than a threshold value. After the pushing user confirms, the feedback is used for model incremental learning, the data popularity score is calculated, then the data is divided into hot data, warm data and cold data to be stored respectively, meanwhile, invalid data is cleared away regularly, natural language query is analyzed through a BERT model to generate a structured query statement, and the structured query statement is stored; entities, attributes and association relationships of the multi-modal data are extracted and imported into a graph database, a knowledge graph is constructed, personalized retrieval and association pushing are achieved, and full-process safety management and control and multi-terminal visual operation support are provided. Data classification management automation and retrieval intellectualization are realized, and the data asset management efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of network-attached storage technology, specifically to a network-attached storage intelligent data management method and system. Background Technology

[0002] Traditional NAS storage systems rely on manual file classification and keyword retrieval, which has the following drawbacks:

[0003] 1. Low data classification efficiency: Users need to manually create folders and classify files, resulting in extremely high management costs for multimodal databases containing images, videos, documents, etc.

[0004] 2. Insufficient search accuracy: Keyword-based searches cannot understand semantic relationships and do not support cross-modal searches;

[0005] 3. Lack of intelligent optimization: It is unable to dynamically adjust classification strategies and search rankings based on user behavior, and data lifecycle management relies on manual configuration.

[0006] While a few existing NAS systems have attempted to incorporate basic AI technologies, a complete multimodal data management system has not yet been formed, particularly in terms of semantic understanding, personalized learning, and cross-module collaboration. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a network-attached storage intelligent data management method and system, which has the advantages of automating data classification management and intelligent retrieval, and improving the efficiency of data asset management.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a network-attached storage intelligent data management method, comprising the following steps:

[0011] Step 1: Multimodal data access and metadata extraction: Monitor the NAS file system through SMB and NFS protocols, capture newly added and modified files in real time, and extract basic metadata and email metadata;

[0012] Step 2, Multimodal Data Parsing and Feature Extraction: Determine the data type, and then parse the data for different data types;

[0013] Step 3: Based on the rule parser, a fusion decision mechanism is implemented. For data with a classification confidence level below the threshold, the data is pushed to the user for confirmation and then used for incremental learning of the model.

[0014] Step 4: Data popularity assessment and hierarchical storage management: Calculate the data popularity score, and then divide the data into hot data, warm data, and cold data for separate storage, while regularly cleaning up invalid data;

[0015] Step 5: Knowledge Graph Construction and Retrieval Capabilities: The BERT model is used to parse natural language queries to generate structured query statements. Entities, attributes, and relationships of multimodal data are extracted and imported into a graph database. After constructing the knowledge graph, it supports accurate retrieval and related queries based on structured statements.

[0016] Step Six: Personalized Search Optimization and Intelligent Services: Record user search behavior, reorder search results to prioritize highly relevant files, and proactively push related files based on the current search content;

[0017] Step 7: Full-process security control and interactive support: Data is encrypted with AES-256, and web and APP operation interfaces are provided to support multiple interaction methods and data visualization.

[0018] Preferably, in step one, the basic metadata includes file name, file format, creation time, modification time, storage path, and author;

[0019] Email metadata includes email subject, sender information, recipient information, email sending time, attachment type and quantity.

[0020] Preferably, in step two, the formula for determining the data type definition is:

[0021] ;

[0022] in, Represents data type tags, Represents visual data. Represents text data. This represents audio data.

[0023] Preferably, in step two, parsing different data types includes:

[0024] S2.1 Visual Data: Visual features are extracted using a CNN convolutional neural network, and facial features are blurred using a privacy protection algorithm to generate visual labels;

[0025] S2.2 Text Data: Keywords and entity information are extracted using NLP (Natural Language Processing) technology, and semantic analysis is performed using a BERT pre-trained model to generate text labels;

[0026] S2.3 Audio data: After being converted into text using speech recognition technology, key information is extracted and combined with semantic analysis to generate text tags.

[0027] Preferably, S2.1 specifically includes:

[0028] S2.1a. Hierarchical feature extraction of visual data based on CNN convolutional neural network, assuming the input visual data is... ,in, Represents height, Represents width, Represents the number of channels;

[0029] S2.1b: Low-level visual features are extracted through the convolutional and pooling layers of the CNN and labeled as... ;

[0030] S2.1c: High-level visual features are extracted through the convolutional and pooling layers of a CNN and labeled as... ;

[0031] ;

[0032] in, This represents the composite operation of convolution, activation, and pooling, specifically:

[0033] ;

[0034] in, Representing the Layer convolution kernel weights, Representing the Layer bias terms;

[0035] S2.1d. For facial regions contained in the features, differential privacy is used for blurring, including:

[0036] Let the facial feature subspace be... Add to satisfy Gaussian noise for differential privacy ,in, Represents noise variance:

[0037] ;

[0038] in, Represents privacy budget, Represents the probability of failure. The global sensitivity representing facial features; the processed facial features are:

[0039] ;

[0040] Complete visual features after de-anonymization Replace here for The input is a fully connected layer and a classifier. The Softmax function generates probability distributions for each category, and finally, the category with the highest probability is taken as the classification label.

[0041] ;

[0042] ;

[0043] in, Representing the Visual categories The probability value, Represents the weight parameters of the fully connected layer, where represent Feature dimensions, The bias parameters represent the fully connected layer. Represents the activation function. Represents an exponential function. Represents the total number of category labels. This represents the final generated category label.

[0044] Preferably, S2.2 specifically includes:

[0045] S2.2a, Let the input text data be... ,in, Represents the first in the text One word, Represents text length;

[0046] S2.2b: Obtain the cleaned text sequence through word segmentation and stop word filtering. Then, keywords were extracted based on the TF-IDF algorithm, and the keyword weight calculation formula is as follows:

[0047] ;

[0048] in, Representative words In the text The number of times it appears in Represents the total number of documents in the text library. Representative words The number of documents, and through named entity recognition, the entity information in the text is identified. , Represents specific entity information;

[0049] Preprocessed The BERT pre-trained model is used for semantic feature extraction. BERT generates deep semantic features of the text through a bidirectional Transformer encoder. Let the input embedding of BERT be:

[0050] ;

[0051] in, Represents sentence-level feature tags. represent The Middle The word embedding vectors of each word are processed by the self-attention mechanism of an L-layer Transformer and a feedforward neural network to output the global semantic feature vector of the text.

[0052] ;

[0053] in, , Representing the BERT feature dimension, the semantic features are finally... Input a fully connected layer and a Softmax classifier to generate the text category probability distribution and determine the final label, using the following formula:

[0054] ;

[0055] ;

[0056] in, The weight matrix represents the weights of the fully connected layer. This represents the bias vector of the fully connected layer. Represents the total number of text categories. Representing the Text categories, This represents the final generated text label.

[0057] Preferably, S2.3 includes:

[0058] S2.3a, Let the input audio data be... ,in, Represents audio temporal sampling points. Represents the sampling length;

[0059] S2.3b. Convert audio signals into text sequences using automatic speech recognition technology, including pre-emphasis, framing, and windowing of the audio to obtain frame features. , The number of frames is then input into the acoustic model and language model, and the recognized text is output. , The first character representing the identified text One word;

[0060] S2.3c, Output the recognized text Using the S2.2 text-based semantic analysis method, the text tags corresponding to the audio were determined. BERT semantic features are labeled as ;

[0061] Preferably, step three includes:

[0062] S3.1. The basic metadata and email metadata extracted in step one are structured and encoded to form metadata features. , , Represents metadata feature dimensions;

[0063] S3.2, Combine the features and tags generated in step two for different data types to form content features. ,include:

[0064] Visual features: High-level characteristics after desensitization Visual labels ;

[0065] Text-based: BERT semantic features Text tags ;

[0066] Audio category: Semantic features of speech recognition technology after text conversion Text tags ;

[0067] S3.3, Mark the visual, text, and audio type identifiers determined in step two as data types. , ;

[0068] S3.4. Logical fusion is performed through a rule parser, including:

[0069] S3.4a. A pre-defined rule base is established based on business needs. The rule conditions include metadata conditions and tag conditions.

[0070] S3.4b, The metadata features from step S3.1 Content features and tags in S3.2 Data type tags in S3.3 As input, it is matched against the preset rule base in S3.4a;

[0071] If at least one rule is matched: take the target classification label corresponding to the rule with the highest matching degree as the preliminary classification result. ;

[0072] If no rule is matched: trigger the AI ​​semantic extraction process and proceed to S3.4c;

[0073] S3.4c, For data that does not match a rule, based on data type tags. Call the corresponding AI semantic extraction model to generate classification results and confidence scores:

[0074] Visual data: Classification labels output by the CNN model in step S2.1 As a preliminary classification result The maximum probability value of the model output is extracted as the confidence level. ;

[0075] Text data: Text labels output by the BERT model in step S2.2 As a preliminary classification result The maximum probability value of the model output is extracted as the confidence level. ;

[0076] Audio data: Labels output from the speech-to-text conversion in step S2.3 plus the BERT model. As a preliminary classification result The maximum probability value of the model output is extracted as the confidence level. ;

[0077] S3.4d, Set the classification confidence threshold Determine the final result:

[0078] If at least one rule is matched, the classification result is deemed valid, and the final label is output directly. , ;

[0079] Regarding confidence level If the classification result is deemed valid, the final label is output directly. , ;

[0080] Regarding confidence level If the classification result is questionable, a manual confirmation process will be triggered, and the final label will be determined after manual review.

[0081] The feedback from manual review is used for incremental learning and updating of the model. , , Optimize the classifier's ability to identify low-confidence samples.

[0082] Preferably, in step four, the formula for calculating the data popularity score is:

[0083] ;

[0084] in, Represents the data popularity score. Representative data Number of visits, Represents the weighting coefficient of the number of visits. Representative data The timestamp of the last access;

[0085] Set the heat threshold. and the hot-cold boundary threshold Then, scores are based on the data's popularity. The data is categorized into hot data, warm data, and cold data, specifically:

[0086] Hot data: ;

[0087] Temperature data: ;

[0088] Cold data: ;

[0089] Hot data is stored in high-performance storage with the highest read / write priority; warm data is stored in normal-performance storage with medium read / write priority; and cold data is archived and stored at low cost with the lowest read / write priority.

[0090] Data that has not been accessed within a fixed period of time is considered invalid data. Data with a popularity score of 0 and no business tags is also considered invalid data and is cleaned up.

[0091] Step five includes:

[0092] S5.1 Entity Extraction:

[0093] For text-based data: From BERT semantic features Named entity recognition results Extract the core entity from the middle;

[0094] For visual data: from visual labels and characteristics after desensitization Extract entities from the middle;

[0095] For audio data: Extracting entities from text tags converted from speech recognition technology. ;

[0096] S5.2 Supplement attributes for each entity, including file metadata. Storage location, data popularity score, and data type;

[0097] S5.3, Querying user natural language queries Perform word segmentation and embedding to generate BERT input:

[0098] ;

[0099] in, Represents the query term embedding vector. Represents the query length;

[0100] S5.4, BERT output feature A custom fully connected layer plus Softmax is used to classify query intent, while entity and relation keywords are extracted from the query using a BERT+CRF model.

[0101] S5.5 Map the parsed keywords into structured statements that can be executed by the graph database;

[0102] S5.6. The entities, attributes, and relationships between entities extracted in S5.1 are written into the graph database through the graph database API to form a complete knowledge graph. ,in, Represents entity nodes, Represents the relationship edge. Representative attribute;

[0103] Step six includes:

[0104] S6.1 Calculate user behavior weights The formula is:

[0105] ;

[0106] in, On behalf of users Click File Number of times, On behalf of users View file Total length of stay On behalf of users Maximum number of clicks On behalf of users Maximum stay duration;

[0107] S6.2 Calculate the personalized comprehensive score And sort:

[0108] ;

[0109] in, Representative document Compared with the current query Basic relevance score, Represents the correlation weighting coefficient. Represents the weighting coefficient of the behavior;

[0110] S6.3, Based on personalized comprehensive score The search results files are sorted in descending order from highest to lowest, generating a sorted result set. ;

[0111] S6.4, Entity Similarity Based on Knowledge Graph ;

[0112] ;

[0113] in, Representative document BERT semantic feature vectors Representative document BERT semantic feature vectors Represents the L2 norm;

[0114] S6.5, For the sorted result set Extract high-priority files from fixed locations Files that meet the following criteria will be selected for the push list:

[0115] ;

[0116] in, Represents the push notification list tags. This represents the similarity threshold for push notifications.

[0117] A network-attached storage intelligent data management system, applied to a network-attached storage intelligent data management method, includes a data access module, a dynamic classification strategy engine module, a storage management module, an intelligent retrieval module, and an interaction module;

[0118] The data access module is used to realize multimodal data access and metadata extraction;

[0119] The dynamic classification strategy engine module is used to complete intelligent data classification and incremental model optimization.

[0120] The storage management module is used to perform tiered storage of data based on data popularity and to clean up invalid data.

[0121] The intelligent retrieval module is used to construct a knowledge graph and realize personalized retrieval and related push notifications;

[0122] The interactive module is used to provide full-process security control and multi-terminal visual operation support.

[0123] Compared with the prior art, the present invention provides a network-attached storage intelligent data management method and system, which has the following beneficial effects:

[0124] 1. This invention achieves efficiency improvement:

[0125] By capturing newly added and modified files in real time through NAS file system monitoring, combined with hierarchical storage and intelligent retrieval, the cycle from data access to usability is significantly shortened; relying on dynamic policy engine to automatically execute classification rules, coupled with personalized search result reordering and related push, the cost of manual intervention is reduced; through multimodal data batch parsing and feature extraction technology, the processing efficiency of large-scale data and asset utilization efficiency are improved.

[0126] 2. This invention has high accuracy:

[0127] It combines models such as CNN and BERT to accurately extract multimodal features and generate labels; it transforms user rules into AI parameters to achieve intelligent classification, introduces human feedback to optimize the model for low-confidence results, and continuously improves classification accuracy; it builds entity relationships based on knowledge graphs and combines them with natural language parsing technology to ensure that search results are highly matched with user needs.

[0128] 3. This invention is safe and reliable:

[0129] Differential privacy technology is used to blur facial features in visual data, and AES-256 encryption is used throughout the process to protect data privacy; clear rules for determining invalid data are established and data is cleaned up regularly; at the same time, operation traceability is achieved through multi-terminal interactive interfaces to reduce data management risks; and a full lifecycle data management system is built by combining file metadata and business tags to ensure that data storage and use comply with compliance requirements. Attached Figure Description

[0130] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0131] Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0132] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. It is worth noting that the parts not disclosed in this application are all prior art, and will not be elaborated on here.

[0133] Please see Figures 1-2 A network-attached storage intelligent data management method includes the following steps:

[0134] Step 1: Multimodal Data Access and Metadata Extraction: Monitor the NAS file system via SMB and NFS protocols to capture newly added and modified files in real time, and extract basic metadata and email metadata. Basic metadata includes file name, file format, creation time, modification time, storage path, and author; email metadata includes email subject, sender information, recipient information, email sending time, attachment type, and quantity.

[0135] By monitoring the addition and modification of files in the NAS file system in real time through SMB and NFS protocols, basic metadata and email metadata are extracted to provide standardized basic information support for subsequent data parsing, classification and retrieval, and to ensure the integrity of the source data in the data management process.

[0136] Step 2, Multimodal Data Parsing and Feature Extraction: Determine the data type, and then parse different data types, including:

[0137] S2.1 Visual Data: Visual features are extracted using a CNN convolutional neural network, and facial features are blurred using a privacy protection algorithm to generate visual labels;

[0138] S2.1a. Hierarchical feature extraction of visual data based on CNN convolutional neural network, assuming the input visual data is... ,in, Represents height, Represents width, Represents the number of channels;

[0139] S2.1b: Low-level visual features are extracted through the convolutional and pooling layers of the CNN and labeled as... ;

[0140] S2.1c: High-level visual features are extracted through the convolutional and pooling layers of a CNN and labeled as... ;

[0141] S2.1d. For the face region contained in the feature, differential privacy is used for blurring.

[0142] First, multi-dimensional visual features are extracted from the raw visual data using CNN. Then, differential privacy noise blurring is introduced for sensitive facial features to protect privacy. Finally, the processed features are input into a classifier, and corresponding visual data classification labels are generated through probability calculation. This not only preserves the classification effectiveness of visual features but also achieves compliant protection of facial privacy.

[0143] S2.2 Text Data: Keywords and entity information are extracted using NLP (Natural Language Processing) technology, and semantic analysis is performed using a BERT pre-trained model to generate text labels;

[0144] S2.2a, Let the input text data be... ,in, Represents the first in the text One word, Represents text length;

[0145] S2.2b: Obtain the cleaned text sequence through word segmentation and stop word filtering. Then, keywords were extracted based on the TF-IDF algorithm, keyword weights were calculated, and the preprocessed keywords were... Input the BERT pre-trained model for semantic feature extraction. BERT generates deep semantic features of the text through a bidirectional Transformer encoder. After the self-attention mechanism of the L-layer Transformer and the operation of the feedforward neural network, the global semantic feature vector of the text is output.

[0146] S2.3 Audio data: After being converted into text using speech recognition technology, key information is extracted and combined with semantic analysis to generate text tags;

[0147] S2.3a, Let the input audio data be... ,in, Represents audio temporal sampling points. Represents the sampling length;

[0148] S2.3b. Convert audio signals into text sequences using automatic speech recognition technology, including pre-emphasis, framing, and windowing of the audio to obtain frame features. , The number of frames is then input into the acoustic model and language model, and the recognized text is output. , The first character representing the identified text One word;

[0149] S2.3c, Output the recognized text Using the S2.2 text-based semantic analysis method, the text tags corresponding to the audio were determined. BERT semantic features are labeled as ;

[0150] We use CNN, BERT, and ASR technologies to extract features and generate labels for visual, text, and audio data, respectively. At the same time, we perform differential privacy protection on facial features in visual data, which achieves both structured representation of multimodal data and privacy compliance requirements.

[0151] Step 3: Based on the rule parser, a fusion decision mechanism is implemented. For data with a classification confidence level below the threshold, the data is pushed to the user for confirmation and then used for incremental learning of the model.

[0152] S3.1. The basic metadata and email metadata extracted in step one are structured and encoded to form metadata features. , , Represents metadata feature dimensions;

[0153] S3.2, Combine the features and tags generated in step two for different data types to form content features. ;

[0154] S3.3, Mark the visual, text, and audio type identifiers determined in step two as data types. , ;

[0155] S3.4. Logical fusion is performed through a rule parser, including:

[0156] S3.4a. A pre-defined rule base is established based on business needs. The rule conditions include metadata conditions and tag conditions.

[0157] S3.4b, The metadata features from step S3.1 Content features and tags in S3.2 Data type tags in S3.3 As input, it is matched against the preset rule base in S3.4a;

[0158] S3.4c, For data that does not match a rule, based on data type tags. Call the corresponding AI semantic extraction model to generate classification results and confidence scores:

[0159] S3.4d, Set the classification confidence threshold Determine the final result;

[0160] By adopting a fusion decision-making mechanism that prioritizes rules and supplements them with AI, comprehensive coverage of the classification scope is achieved. In addition, the feedback results of manual review can be used simultaneously for rule base iteration and AI model incremental learning, forming a closed loop from rule optimization to model upgrade, continuously improving classification accuracy, and taking into account both the practicality of business implementation and the long-term iteration capability of the system.

[0161] Step 4: Data Popularity Assessment and Tiered Storage Management: Calculate the data popularity score, and then set the popularity threshold. and the hot-cold boundary threshold Based on the data popularity score, the data is divided into hot data, warm data, and cold data. Hot data is stored in high-performance storage with the highest read and write priority; warm data is stored in normal performance storage with medium read and write priority; and cold data is archived and stored at low cost with the lowest read and write priority. Data that has not been accessed within a fixed period of time is judged as invalid data. At the same time, data with a popularity score of 0 and no business tag is also judged as invalid data and cleaned up.

[0162] By calculating data popularity scores and setting thresholds to classify data into hot, warm, and cold categories, and matching differentiated storage strategies, while cleaning up invalid data that has no access, no popularity, and no business tags, storage costs are reduced and storage resource utilization is improved while ensuring data access efficiency.

[0163] Step 5: Knowledge Graph Construction and Retrieval Capabilities: The BERT model is used to parse natural language queries to generate structured query statements. Entities, attributes, and relationships from multimodal data are extracted and imported into a graph database. After constructing the knowledge graph, it supports precise retrieval and related queries based on structured statements, including:

[0164] S5.1 Entity Extraction:

[0165] S5.2 Supplement attributes for each entity, including file metadata. Storage location, data popularity score, and data type;

[0166] S5.3, Querying user natural language queries Perform word segmentation and embedding to generate BERT input:

[0167] S5.4, BERT output feature A custom fully connected layer plus Softmax is used to classify query intent, while entity and relation keywords are extracted from the query using a BERT+CRF model.

[0168] S5.5 Map the parsed keywords into structured statements that can be executed by the graph database;

[0169] S5.6. The entities, attributes, and relationships between entities extracted in S5.1 are written into the graph database through the graph database API to form a complete knowledge graph. ;

[0170] Knowledge graphs are constructed based on entities, attributes, and relationships in multimodal data. BERT is used to parse natural language queries and generate structured statements to achieve accurate retrieval, breaking down data silos and improving the efficiency of mining information related to data and the level of intelligence in retrieval.

[0171] Step Six: Personalized Search Optimization and Intelligent Services: Record user search behavior, reorder search results to prioritize highly relevant documents, and proactively push related documents based on the current search content, including calculating the weight of user behavior. And calculate a personalized comprehensive score. Then based on the personalized comprehensive score The search results files are sorted in descending order from highest to lowest, generating a sorted result set. Based on knowledge graph entity similarity For the sorted result set Extract high-priority files from fixed locations Filter files that meet the criteria to create a push list;

[0172] Record user search behavior and calculate personalized comprehensive scores, re-rank search results, and push related files based on knowledge graph entity similarity to meet users' personalized search needs and improve the relevance of search results and user experience;

[0173] Step 7: Full-process security control and interactive support: Data is encrypted with AES-256, and web and APP operation interfaces are provided to support multiple interaction methods and data visualization.

[0174] It adopts AES-256 encryption to ensure data security throughout the entire process, and provides multi-terminal interactive interfaces and data visualization functions such as Web and APP. It not only ensures the security and compliance of the data management process, but also reduces the user's operating threshold and improves the ease of use of the system.

[0175] A network-attached storage intelligent data management system, applied to a network-attached storage intelligent data management method, includes a data access module, a dynamic classification strategy engine module, a storage management module, an intelligent retrieval module, and an interaction module;

[0176] The data access module is used to realize multimodal data access and metadata extraction;

[0177] The dynamic classification strategy engine module is used to complete intelligent data classification and incremental model optimization;

[0178] The storage management module is used to perform tiered storage based on data popularity and to clean up invalid data;

[0179] The intelligent retrieval module is used to build knowledge graphs and enable personalized retrieval and related push notifications;

[0180] The interaction module is used to provide full-process security control and multi-terminal visual operation support.

[0181] 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 alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A network-attached storage intelligent data management method, characterized in that, Includes the following steps: Step 1: Multimodal data access and metadata extraction: Monitor the NAS file system through SMB and NFS protocols, capture newly added and modified files in real time, and extract basic metadata and email metadata; Step 2, Multimodal Data Parsing and Feature Extraction: Determine the data type, and then parse the data for different data types; Step 3: Based on the rule parser, a fusion decision mechanism is implemented. For data with a classification confidence level below the threshold, the data is pushed to the user for confirmation and then used for incremental learning of the model. Step 4: Data popularity assessment and hierarchical storage management: Calculate the data popularity score, and then divide the data into hot data, warm data, and cold data for separate storage, while regularly cleaning up invalid data; Step 5: Knowledge Graph Construction and Retrieval Capabilities: The BERT model is used to parse natural language queries to generate structured query statements. Entities, attributes, and relationships of multimodal data are extracted and imported into a graph database. After constructing the knowledge graph, it supports accurate retrieval and related queries based on structured statements. Step Six: Personalized Search Optimization and Intelligent Services: Record user search behavior, reorder search results to prioritize highly relevant files, and proactively push related files based on the current search content; Step 7: Full-process security control and interactive support: Data is encrypted with AES-256, and web and APP operation interfaces are provided to support multiple interaction methods and data visualization.

2. The intelligent data management method for network-attached storage according to claim 1, characterized in that, In step one, the basic metadata includes file name, file format, creation time, modification time, storage path, and author; Email metadata includes email subject, sender information, recipient information, email sending time, attachment type and quantity.

3. The intelligent data management method for network-attached storage according to claim 1, characterized in that, In step two, the formula for determining the data type definition is as follows: ； in, Represents data type tags, Represents visual data. Represents text data. This represents audio data.

4. The intelligent data management method for network-attached storage according to claim 3, characterized in that, Step two, which involves parsing different data types, includes: S2.1 Visual Data: Visual features are extracted using a CNN convolutional neural network, and facial features are blurred using a privacy protection algorithm to generate visual labels; S2.2 Text Data: Keywords and entity information are extracted using NLP (Natural Language Processing) technology, and semantic analysis is performed using a BERT pre-trained model to generate text labels; S2.3 Audio data: After being converted into text using speech recognition technology, key information is extracted and combined with semantic analysis to generate text tags.

5. The network-attached storage intelligent data management method according to claim 3, characterized in that, S2.1 specifically includes: S2.1a. Hierarchical feature extraction of visual data based on CNN convolutional neural network, assuming the input visual data is... ,in, Represents height, Represents width, Represents the number of channels; S2.1b: Low-level visual features are extracted through the convolutional and pooling layers of the CNN and labeled as... ; S2.1c: High-level visual features are extracted through the convolutional and pooling layers of a CNN and labeled as... ; ； in, This represents the composite operation of convolution, activation, and pooling, specifically: ； in, Representing the Layer convolution kernel weights, Representing the Layer bias terms; S2.1d. For facial regions contained in the features, differential privacy is used for blurring, including: Let the facial feature subspace be... Add to satisfy Gaussian noise for differential privacy ,in, Represents noise variance: ； in, Represents privacy budget, Represents the probability of failure. The global sensitivity representing facial features; the processed facial features are: ； Complete visual features after de-anonymization Replace here for The input is a fully connected layer and a classifier. The Softmax function generates probability distributions for each category, and finally, the category with the highest probability is taken as the classification label. ； ； in, Representing the Visual categories The probability value, Represents the weight parameters of the fully connected layer, where represent Feature dimensions, The bias parameters represent the fully connected layer. Represents the activation function. Represents an exponential function. Represents the total number of category labels. This represents the final generated category label.

6. The network-attached storage intelligent data management method according to claim 3, characterized in that, S2.2 specifically includes: S2.2a, Let the input text data be... ,in, Represents the first in the text One word, Represents text length; S2.2b: Obtain the cleaned text sequence through word segmentation and stop word filtering. Then, keywords were extracted based on the TF-IDF algorithm, and the keyword weight calculation formula is as follows: ； in, Representative words In the text The number of times it appears in Represents the total number of documents in the text library. Representative words The number of documents, and through named entity recognition, the entity information in the text is identified. , Represents specific entity information; Preprocessed The BERT pre-trained model is used for semantic feature extraction. BERT generates deep semantic features of the text through a bidirectional Transformer encoder. Let the input embedding of BERT be: ； in, Represents sentence-level feature tags. represent The Middle The word embedding vectors of each word are processed by the self-attention mechanism of an L-layer Transformer and a feedforward neural network to output the global semantic feature vector of the text. ； in, , Representing the BERT feature dimension, the semantic features are finally... Input a fully connected layer and a Softmax classifier to generate the text category probability distribution and determine the final label, using the following formula: ； ； in, The weight matrix represents the weights of the fully connected layer. This represents the bias vector of the fully connected layer. Represents the total number of text categories. Representing the Text categories, This represents the final generated text label.

7. The network-attached storage intelligent data management method according to claim 3, characterized in that, S2.3 includes: S2.3a, Let the input audio data be... ,in, Represents audio temporal sampling points. Represents the sampling length; S2.3b. Convert audio signals into text sequences using automatic speech recognition technology, including pre-emphasis, framing, and windowing of the audio to obtain frame features. , The number of frames is then input into the acoustic model and language model, and the recognized text is output. , The first character representing the identified text One word; S2.3c, Output the recognized text Using the S2.2 text-based semantic analysis method, the text tags corresponding to the audio were determined. BERT semantic features are labeled as .

8. The intelligent data management method for network-attached storage according to claim 1, characterized in that, Step three includes: S3.

1. The basic metadata and email metadata extracted in step one are structured and encoded to form metadata features. , , Represents metadata feature dimensions; S3.2, Combine the features and tags generated in step two for different data types to form content features. ,include: Visual features: High-level characteristics after desensitization Visual labels ; Text-based: BERT semantic features Text tags ; Audio category: Semantic features of speech recognition technology after text conversion Text tags ; S3.3, Mark the visual, text, and audio type identifiers determined in step two as data types. , ; S3.

4. Logical fusion is performed through a rule parser, including: S3.4a. A pre-defined rule base is established based on business needs. The rule conditions include metadata conditions and tag conditions. S3.4b, The metadata features from step S3.1 Content features and tags in S3.2 Data type tags in S3.3 As input, it is matched against the preset rule base in S3.4a; If at least one rule is matched: take the target classification label corresponding to the rule with the highest matching degree as the preliminary classification result. ; If no rule is matched: trigger the AI ​​semantic extraction process and proceed to S3.4c; S3.4c, For data that does not match a rule, based on data type tags. Call the corresponding AI semantic extraction model to generate classification results and confidence scores: Visual data: Classification labels output by the CNN model in step S2.1 As a preliminary classification result The maximum probability value of the model output is extracted as the confidence level. ; Text data: Text labels output by the BERT model in step S2.2 As a preliminary classification result The maximum probability value of the model output is extracted as the confidence level. ; Audio data: Labels output from the speech-to-text conversion in step S2.3 plus the BERT model. As a preliminary classification result The maximum probability value of the model output is extracted as the confidence level. ; S3.4d, Set the classification confidence threshold Determine the final result: If at least one rule is matched, the classification result is deemed valid, and the final label is output directly. , ; Regarding confidence level If the classification result is deemed valid, the final label is output directly. , ; Regarding confidence level If the classification result is questionable, a manual confirmation process will be triggered, and the final label will be determined after manual review. The feedback from manual review is used for incremental learning and updating of the model. , , Optimize the classifier's ability to identify low-confidence samples.

9. The intelligent data management method for network-attached storage according to claim 8, characterized in that, In step four, the formula for calculating the data popularity score is as follows: ； in, Represents the data popularity score. Representative data Number of visits, Represents the weighting coefficient of the number of visits. Representative data The timestamp of the last access; Set the heat threshold. and the hot-cold boundary threshold Then, scores are based on the data's popularity. The data is categorized into hot data, warm data, and cold data, specifically: Hot data: ; Temperature data: ; Cold data: ; Hot data is stored in high-performance storage with the highest read / write priority; warm data is stored in normal-performance storage with medium read / write priority; and cold data is archived and stored at low cost with the lowest read / write priority. Data that has not been accessed within a fixed period of time is considered invalid data. Data with a popularity score of 0 and no business tags is also considered invalid data and is cleaned up. Step five includes: S5.1 Entity Extraction: For text-based data: From BERT semantic features Named entity recognition results Extract the core entity from the middle; For visual data: from visual labels and characteristics after desensitization Extract entities from the middle; For audio data: Extracting entities from text tags converted from speech recognition technology. ; S5.2 Supplement attributes for each entity, including file metadata. Storage location, data popularity score, and data type; S5.3, Querying user natural language queries Perform word segmentation and embedding to generate BERT input: ； in, Represents the query term embedding vector. Represents the query length; S5.4, BERT output feature A custom fully connected layer plus Softmax is used to classify query intent, while entity and relation keywords are extracted from the query using a BERT+CRF model. S5.5 Map the parsed keywords into structured statements that can be executed by the graph database; S5.

6. The entities, attributes, and relationships between entities extracted in S5.1 are written into the graph database through the graph database API to form a complete knowledge graph. ,in, Represents entity nodes, Represents the relationship edge. Representative attribute; Step six includes: S6.1 Calculate user behavior weights The formula is: ； in, On behalf of users Click File Number of times, On behalf of users View file Total length of stay On behalf of users Maximum number of clicks On behalf of users Maximum stay duration; S6.2 Calculate the personalized comprehensive score And sort: ； in, Representative document Compared with the current query Basic relevance score, Represents the correlation weighting coefficient. Represents the weighting coefficient of the behavior; S6.3, Based on personalized comprehensive score The search results files are sorted in descending order from highest to lowest, generating a sorted result set. ; S6.4, Entity Similarity Based on Knowledge Graph ; ； in, Representative document BERT semantic feature vectors Representative document BERT semantic feature vectors Represents the L2 norm; S6.5, For the sorted result set Extract high-priority files from fixed locations Files that meet the following criteria will be selected for the push list: ； in, Represents the push notification list tags. This represents the similarity threshold for push notifications.

10. A network-attached storage intelligent data management system, applied to the network-attached storage intelligent data management method according to any one of claims 1 to 9, characterized in that, It includes a data access module, a dynamic classification strategy engine module, a storage management module, an intelligent retrieval module, and an interaction module; The data access module is used to realize multimodal data access and metadata extraction; The dynamic classification strategy engine module is used to complete intelligent data classification and incremental model optimization. The storage management module is used to perform tiered storage of data based on data popularity and to clean up invalid data. The intelligent retrieval module is used to construct a knowledge graph and realize personalized retrieval and related push notifications; The interactive module is used to provide full-process security control and multi-terminal visual operation support.

Images