Archives management method and system based on multimodal data analysis technology

Abstract

The present invention relates to an archive management method and system based on multimodal data analysis technology, comprising the following steps: S1: collecting power-related archival data and using adaptive denoising technology to improve the processing quality of various file types; S2: based on the denoised power-related archival data, using a tokenizer to convert the text into integer index sequences, using padding to standardize the sequences to the same length, and using a text classification model to automatically classify the data and generate labels; S3: based on the classified text data, using a sequence-to-sequence generation model to automatically generate summaries for technical documents and reports; S4: formulating personalized archiving strategies based on user behavior analysis to meet the specific requirements of different departments or needs. This invention not only improves data accessibility and usability, but also effectively enhances the quality and efficiency of archive management.

Description

Technical Field

[0001] The present invention relates to the field of archive management, and in particular to an archive management method and system based on multimodal data analysis technology. Background Art

[0002] In today's information age, businesses and organizations generate and accumulate vast amounts of data during their operations. This data includes not only structured data but also a vast amount of unstructured data, such as text, images, and audio. With the deepening of digital transformation, particularly in the power industry, the volume and complexity of data are increasing. This data holds rich information value, but it also places higher demands on its storage, management, and analysis.

[0003] Traditional archive management methods typically rely on manual classification and storage, which is inefficient and prone to errors. With the development of artificial intelligence and data science technologies, intelligent archive management technology is gradually emerging. Summary of the Invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide an archive management method and system based on multimodal data analysis technology, which not only improves the accessibility and availability of data, but also effectively improves the quality and efficiency of archive management.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A file management method based on multimodal data analysis technology includes the following steps:

[0007] S1: Collect archival data related to electricity and use adaptive denoising technology to improve the processing quality of various types of files;

[0008] S2: Based on the denoised power-related archival data, a tokenizer is used to convert the text into integer index sequences, padding is used to normalize the sequences to the same length, and a text classification model is used to automatically classify the data and generate labels.

[0009] S3: Based on the classified text data, a sequence-to-sequence generation model is used to automatically generate summaries for technical documents and reports;

[0010] S4: Develop personalized archiving strategies based on user behavior analysis to meet the specific requirements of different departments or needs;

[0011] The S4 is specifically:

[0012] S41: Collecting user interaction data and user meta information, wherein the user interaction data includes usage logs, click streams, search queries, access frequency, and file modification records; the user meta information includes user role, department, and permission level;

[0013] S42: Extract features, including visit frequency, dwell time, visit sequence, and correlation between the most frequently searched keywords, and use cluster analysis to analyze behavioral patterns and classify users into different groups;

[0014] S43: Build a graph network model of user interaction, where nodes represent files and users, and edge weights represent interaction strength. Apply graph analysis techniques to identify highly connected groups or nodes.

[0015] S44: Apply association rule mining to identify co-occurrence behaviors and formulate targeted strategies; S45: Use reinforcement learning models to generate dynamic archiving strategies, adaptively adjust the organizational structure of archives according to user behavior, and define specific archiving rules based on user request speed and data storage efficiency. The archiving rules include priority and automatic archiving trigger conditions.

[0016] Furthermore, S1 is specifically:

[0017] Collecting archival data related to electricity, including equipment maintenance records, operation logs, sensor data, and contract documents;

[0018] Perform text data denoising on equipment maintenance records, operation logs, and contract documents. This includes using TextBlob for spelling and grammar correction, standardizing text formats, and removing redundant information. Regular expressions are used to remove unstructured noise, define common noise patterns, and automatically filter identified noise.

[0019] Sensor data is processed by filling missing values, smoothing, and detecting and correcting anomalies.

[0020] Furthermore, we use a tokenizer to convert the text into a sequence of integer indices, and use padding to normalize the sequences to the same length, as follows:

[0021] Use tokenizer to create a vocabulary and index mapping for the entire corpus, traverse the corpus, count the frequency of each word, create a unique vocabulary, and assign a unique integer index to each word in the vocabulary;

[0022] Convert each text sentence into an index sequence, replacing each word in the text with its index in the vocabulary according to the vocabulary mapping;

[0023] According to the task requirements and computing power, preset the maximum length of the sequence. For sequences that are too long, truncate them; for sequences that are too short, pad them. For sequences with a length less than the maximum length max_len, use padding tokens to pad them; for sequences that exceed the maximum length, truncate them.

[0024] For a sequence [x1, x2, …, xm], where m is the length of the sequence.

[0025] If the sequence length m < max_len, then pad the sequence to:

[0026] [x1, x2, …, xm, 0, 0, …, 0]

[0027] Where the padded sequence length is max_len.

[0028] If m > max_len, then truncate the sequence to:

[0029] [x1, x2, …, xmax_len].

[0030] Furthermore, the construction of the text classification model is as follows:

[0031] The input layer uses a pre-trained word embedding layer to convert words into vectors.

[0032] The input size is: (batch_size, sequence_length, embedding_dim);

[0033] Where batch_size is the batch size, sequence_length is the maximum number of words in each line of text, and embedding_dim is the dimension of the word vector.

[0034] Add multiple convolutional layers, each using a different kernel size, and add a corresponding pooling layer after each convolutional layer.

[0035] ;

[0036] Where, h , ,

[0036] , k ,

[0037] , , , i+k , i:j ,

[0038] represents a segment of the word embedding sequence starting at position i with a length of and represents any segment of the text; w k is the weight parameter of the convolutional kernel, b is the bias, and f is the activation function ReLU; E(x i+k ): represents the word embedding vector at the i + k-th position in the input sequence, and K represents the width of the convolutional kernel.

[0037] The pooling layer uses max pooling to extract the important features of the output of each convolutional layer:

[0038] p = max(h);

[0039] Among them, h is the result after convolution, and p is the feature after pooling;

[0040] The pooled features are flattened and used as the input of the fully connected layer;

[0041] Finally, the output of the fully connected layer passes through Softmax to produce the final category prediction;

[0042] The training process uses cross entropy loss as the loss function and combines the adaptive optimization algorithm AMS for weight optimization;

[0043] Deploy the trained text classification model to the production environment to achieve automatic classification of text data.

[0044] Furthermore, AMS is combined to optimize the weights, as follows:

[0045] Initialize the first-order momentum m0 and second-order momentum v0 for each parameter θ: m0←0, v0←0; set the learning rate α, the decay rates β1 and β2 of the first-order momentum and second-order momentum; set the small value To prevent division by zero errors, the subscript t indicates the corresponding time step;

[0046] Iterative update, at each time step t:

[0047] Calculate the gradient g t :

[0048] ;

[0049] Update the first-order momentum m t :

[0050] ;

[0051] Update the second-order momentum v t :

[0052] ;

[0053] Maintaining the max second-order momentum valuation :

[0054] ;

[0055] Bias-corrected first-order momentum:

[0056] ;

[0057] Parameter update:

[0058] ;

[0059] By keeping the maximum historical value of the second-order momentum, the over-ease problem in momentum update is avoided.

[0060] Furthermore, the sequence-to-sequence generation model is specifically:

[0061] In the sequence-to-sequence generation model, a generative language model is used to predict the sequence through an encoder-decoder architecture, sharing the information of the context representation;

[0062] Encoder: Converts input text into context-aware hidden states and generates an embedding representation of the input sequence using a transformer model;

[0063] Decoder: Generates the target sequence from the hidden state of the encoder, using the architecture of the Transformer model;

[0064] The decoder looks at the encoder's output according to its importance at each generation step, forms a dynamic context vector, and uses a multi-head self-attention mechanism to adjust the attention weights to focus on the key information of the input text;

[0065] For the attention mechanism, let the input matrix be X, which is projected into query Q, key K and value V matrices:

[0066] Q=XW Q ,K=XW K ,V=XW V ;

[0067] Among them, W Q 、W K 、W V Weight matrices for query, key, and value, respectively;

[0068] Attention calculation:

[0069] ;

[0070] Among them, d k is the dimension of the key vector;

[0071] During decoding, a beam search strategy is used to improve stability by exploring multiple sequences simultaneously. A copy mechanism is used to increase the ability to copy words directly from the input, which is particularly helpful for generating keywords in technical texts.

[0072] The standard cross-entropy loss of the sequence-to-sequence model is used to drive the model to generate target sentences. A coverage loss is introduced and combined with the standard cross-entropy loss to adjust the relative strength weights so that the generated text has a certain degree of creativity while still maintaining consistency. The BLEU and ROUGE indicators are used to quantitatively analyze the n-gram or subsequence overlap between the generated summary and the reference summary to evaluate the accuracy and consistency of the generated text.

[0073] Furthermore, a beam search strategy is used during decoding to improve stability by exploring multiple sequences simultaneously, as follows:

[0074] Set e, the number of candidate sequences retained at each time step;

[0075] Initialize the candidate sequence set to the starting state;

[0076] For each time step t, the probability of each current sequence expanding to the next word is calculated. For a given candidate sequence S, the probability of its expansion is:

[0077] ;

[0078] Among them, y t is the generated word at time step t, X is the input sequence; T is the time step length;

[0079] Select the e×r words with the highest probability from the next word of each candidate sequence, where r is the vocabulary size;

[0080] Keep the e extended sequences with the highest probability as candidate sequences for the next time step;

[0081] Continue expanding until the maximum sequence length is reached or all sequences output end markers <eos>.

[0082] Furthermore, the Copy Mechanism is used to add the ability to copy words directly from the input, as follows:

[0083] Calculate probabilities for generating words and copying words from input separately;

[0084] Using a learnable parameter P gen Determines whether to generate or copy from the vocabulary:

[0085] ;

[0086] Among them, σ is the sigmoid function; h t and s t are the encoder hidden state and decoder hidden state at time step t respectively; is the input of the previous time step; w gen is the weight matrix;

[0087] Calculate the vocabulary distribution P of the generative model vocab :

[0088] ;

[0089] s t is the hidden state of the decoder at time step t; W out is the weight matrix of the output layer; b out is the bias of the output layer;

[0090] Calculate the input duplication probability P based on each input position copy ;

[0091] ;

[0092] in, Indicates the input location The input word Equal to the current output word y t ;

[0093] in, is the attention allocated to the input position The weight of

[0094] The combined generation and replication probabilities are used as the final output word probability P ( y t ):

[0095] .

[0096] An archive management system based on multimodal data analysis technology includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the archive management method based on multimodal data analysis technology as described above.

[0097] The present invention has the following beneficial effects:

[0098] 1. This invention not only improves the accessibility and usability of data, but also effectively improves the quality and efficiency of archive management;

[0099] 2. The present invention ensures that all text data are converted into integer index sequences of the same length, making them suitable for input into deep learning models for batch processing. It efficiently processes text inputs of different lengths, enables the text classification model to achieve accurate classification and label generation, and effectively improves the classification quality. It also uses the cross-entropy loss function and the AMSGrad optimization algorithm to enable the model to obtain a more stable learning process during training, while improving the modeling ability of complex classification tasks, effectively improving the stability and accuracy of the text classification model.

[0100] 3. By combining Beam Search and Copy Mechanism, the present invention can improve the quality of generated output when processing natural language. It combines the exploratory nature of Beam Search with the precision of Copy Mechanism to generate more fluent and appropriate text paragraphs, obtain more accurate summary content, and effectively improve the quality of archive management. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0102] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0103] refer to Figure 1 In this embodiment, a file management method based on multimodal data analysis technology is provided, comprising the following steps:

[0104] S1: Collect archival data related to electricity and use adaptive denoising technology to improve the processing quality of various types of files;

[0105] S2: Based on the denoised power-related archival data, a tokenizer is used to convert the text into integer index sequences, padding is used to normalize the sequences to the same length, and a text classification model is used to automatically classify the data and generate labels.

[0106] S3: Based on the classified text data, a sequence-to-sequence generation model is used to automatically generate summaries for technical documents and reports;

[0107] S4: Develop personalized archiving strategies based on user behavior analysis to meet the specific requirements of different departments or needs;

[0108] The S4 is specifically:

[0109] S41: Collecting user interaction data and user meta information, wherein the user interaction data includes usage logs, click streams, search queries, access frequency, and file modification records; the user meta information includes user role, department, and permission level;

[0110] S42: Extract features, including visit frequency, dwell time, visit sequence, and correlation between the most frequently searched keywords, and use cluster analysis to analyze behavioral patterns and classify users into different groups;

[0111] S43: Build a graph network model of user interaction, where nodes represent files and users, and edge weights represent interaction strength. Apply graph analysis techniques to identify highly connected groups or nodes.

[0112] S44: Apply association rule mining (such as the Apriori algorithm) to identify co-occurrence behaviors and develop targeted strategies; for example, users who frequently access a certain type of file often have organizational requirements for similar files.

[0113] S45: Use a reinforcement learning model to generate a dynamic archiving strategy, adaptively adjust the organizational structure of the archive according to user behavior, and define specific archiving rules based on user request speed and data storage efficiency. The archiving rules include priority and automatic archiving trigger conditions.

[0114] In this embodiment, the S1 is specifically:

[0115] Collecting archival data related to electricity, including equipment maintenance records, operation logs, sensor data, and contract documents;

[0116] Equipment maintenance records: Primarily structured text data that may contain formatting inconsistencies and spelling errors.

[0117] Operation logs: Semi-structured text data, which may contain redundant information and erroneous records. Sensor data: Numerical time series data, which may contain noise, missing values, and outliers. Contract text: Unstructured text data, which may contain formatting issues and semantic errors due to OCR recognition.

[0118] Perform text data denoising on equipment maintenance records, operation logs, and contract texts, including using TextBlob for spelling and grammar correction, unifying text formats (such as date formats, unit formats), and removing redundant information; use regular expressions to remove unstructured noise, define common noise patterns, and automatically filter out the identified noise;

[0119] For sensor data, perform missing value filling, smoothing, and anomaly detection and correction processing.

[0120] In this implementation, use a tokenizer to convert the text into an integer index sequence, and use padding to standardize the sequences to the same length, as follows:

[0121] Use a tokenizer to create a vocabulary and index mapping for the entire corpus. Traverse the corpus, count the frequency of each word, create a unique vocabulary, and assign a unique integer index to each word in the vocabulary;

[0122] Convert each text sentence into an index sequence. According to the vocabulary mapping, replace each word in the text with its index in the vocabulary;

[0123] Preset the maximum sequence length according to the task requirements and computing power. Truncate sequences that are too long and pad sequences that are too short; for sequences with a length less than the maximum length max_len, use padding tokens to pad them; for sequences that exceed the maximum length, truncate them;

[0124] For the sequence [x1,x2,…,xm]: m is the sequence length;

[0125] If the sequence length m < max_len, then pad the sequence to:

[0126] [x1,x2,…,xm,0,0,…,0]

[0127] Where the padded sequence length is max_len.

[0128] If m > max_len, then truncate the sequence to:

[0129] [x1,x2,…,xmax_len].

[0130] In this implementation, the construction of the text classification model is as follows:

[0131] The input layer uses a pre-trained word embedding layer to convert words into vectors;

[0132] The input size is: (batch_size, sequence_length, embedding_dim)

[0133] Where batch_size is the batch size, sequence_length is the maximum number of words per line of text, and embedding_dim is the dimension of the word vector;

[0134] Add multiple convolutional layers, each with a different kernel size, and add a corresponding pooling layer after each convolutional layer;

[0135] ;

[0136] Among them, h i:j It means starting at position i and having a length of A word embedding sequence represents any section of text; w k is the weight parameter of the convolution kernel, b is the bias, and f is the activation function ReLU; E(x i+k ): represents the word embedding vector at the i+kth position of the input sequence, and K represents the width of the convolution kernel;

[0137] The pooling layer uses maximum pooling to extract important features of each convolutional layer output:

[0138] p=max(h)

[0139] Among them, h is the result after convolution, and p is the feature after pooling;

[0140] The pooled features are flattened and used as input to the fully connected layer;

[0141] Finally, the output of the fully connected layer passes through Softmax to produce the final category prediction;

[0142] The training process uses cross entropy loss as the loss function and combines AMS for weight optimization;

[0143] Deploy the trained text classification model to the production environment to achieve automatic classification of text data.

[0144] In this implementation, weight optimization is performed in conjunction with AMS, as follows:

[0145] Initialize the first-order momentum m0 and second-order momentum v0 for each parameter θ: m0←0, v0←0; set the learning rate α, the decay rates β1 and β2 of the first-order momentum and second-order momentum; set the small value To prevent division by zero errors, the subscript t represents the time step t;

[0146] Iterative update, at each time step t:

[0147] Calculate the gradient g t :

[0148] ;

[0149] Update the first-order momentum m t :

[0150] ;

[0151] Update the second-order momentum v t :

[0152] ;

[0153] Maintaining the max second-order momentum valuation :

[0154] ;

[0155] Bias-corrected first-order momentum:

[0156] ;

[0157] Parameter update:

[0158] ;

[0159] By keeping the maximum historical value of the second-order momentum, the over-ease problem in momentum update is avoided.

[0160] In this implementation, the sequence-to-sequence generation model is specifically:

[0161] In the sequence-to-sequence generation model, a generative language model is used to predict the sequence through an encoder-decoder architecture, sharing the information of the context representation;

[0162] Encoder: Converts input text into context-aware hidden states and generates an embedding representation of the input sequence using a transformer model;

[0163] Decoder: Generates the target sequence from the hidden state of the encoder, using a Transformer architecture design, including an input layer, a multi-layer Transformer decoder layer, and an output layer; the input layer inputs the target sequence generated from the hidden state of the encoder, and the multi-layer Transformer decoder layer includes masked multi-head self-attention, encoder-decoder attention, a feedforward network, residual and normalization layers; the output layer outputs the vocabulary probability distribution through softmax; decoding generates each word of the target sequence in sequence, and the output of the current step is used as the input of the next step until a terminator is encountered.

[0164] The decoder looks at the encoder's output according to its importance at each generation step, forms a dynamic context vector, and uses a multi-head self-attention mechanism to adjust the attention weights to focus on the key information of the input text;

[0165] For the attention mechanism, let the input matrix be X, which is projected into query Q, key K and value V matrices:

[0166] Q=XW Q ,K=XW K ,V=XW V ;

[0167] Among them, W Q 、W K 、W V Weight matrices for query, key, and value, respectively;

[0168] Attention calculation:

[0169] ;

[0170] Among them, d k is the dimension of the key vector;

[0171] A beam search strategy is used during decoding to improve stability by exploring multiple sequences simultaneously. A copy mechanism is used to add the ability to copy words directly from the input, which is particularly helpful for generating keywords in technical texts.

[0172] The standard cross-entropy loss of the sequence-to-sequence model is used to drive the model to generate target sentences. Diversified losses (such as those that incentivize coverage of new n-grams, avoid repetition, and penalize patterned output) are introduced. Combined with the weighted standard cross-entropy loss, the relative strength weights are adjusted to ensure that the generated text has a certain degree of creativity while still maintaining consistency. The BLEU and ROUGE metrics are used to quantitatively analyze the degree of n-gram or subsequence overlap between the generated summary and the reference summary, thereby evaluating the accuracy and consistency of the generated text.

[0173] In this implementation, a beam search strategy is used during decoding to improve stability by exploring multiple sequences simultaneously, as follows:

[0174] Set e, the number of candidate sequences retained at each time step;

[0175] Initialize the candidate sequence set to the starting state;

[0176] For each time step t, the probability of each current sequence expanding to the next word is calculated. For a given candidate sequence S, the probability of its expansion is:

[0177] ;

[0178] Among them, y t is the generated word at time step t, X is the input sequence; T is the time step length;

[0179] Select the e×r words with the highest probability from the next word of each candidate sequence, where r is the vocabulary size;

[0180] Keep the e extended sequences with the highest probability as candidate sequences for the next time step;

[0181] Continue expanding until the maximum sequence length is reached or all sequences output end markers <eos>.

[0182] In this implementation, the Copy Mechanism is used to add the ability to copy words directly from the input, as follows:

[0183] Calculate probabilities for generating words and copying words from input separately;

[0184] Using a learnable parameter P gen (Generation probability) decides whether to generate from vocabulary or copy:

[0185] ;

[0186] Among them, σ is the sigmoid function; h t and s t are the encoder hidden state and decoder hidden state at time step t respectively; is the input of the previous time step; w gen is the weight matrix;

[0187] Calculate the vocabulary distribution P of the generative model vocab :

[0188] ;

[0189] s t is the hidden state of the decoder at time step t; W out is the weight matrix of the output layer; b out is the bias of the output layer;

[0190] Calculate the input duplication probability P based on each input position copy ;

[0191] ;

[0192] in, Indicates the input location The input word Equal to the current output word y t ;

[0193] in, is the attention allocated to the input position The weight of

[0194] The combined generation and replication probabilities are used as the final output word probability P ( y t ):

[0195] .

[0196] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0197] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0198] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0199] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0200] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.< / eos> < / eos>

Claims

1. A file management method based on multimodal data analysis technology, characterized in that: The following steps are included: S1: Collect archival data related to electricity and use adaptive denoising technology to improve the processing quality of various types of files; S2: Based on the denoised power-related archival data, a tokenizer is used to convert the text into integer index sequences, padding is used to normalize the sequences to the same length, and a text classification model is used to automatically classify the data and generate labels. S3: Based on the classified text data, a sequence-to-sequence generation model is used to automatically generate summaries for technical documents and reports; S4: Develop personalized archiving strategies based on user behavior analysis to meet the requirements of different departments or needs; The S4 is specifically: S41: Collecting user interaction data and user meta information, wherein the user interaction data includes usage logs, click streams, search queries, access frequency, and file modification records; the user meta information includes user role, department, and permission level; S42: Extract features, including visit frequency, dwell time, visit sequence, and correlation between the most frequently searched keywords, and use cluster analysis to analyze behavioral patterns and classify users into different groups; S43: Build a graph network model of user interaction, where nodes represent files and users, and edge weights represent interaction strength. Apply graph analysis techniques to identify highly connected groups or nodes. S44: Apply association rule mining to identify co-occurrence behaviors and formulate targeted strategies; S45: Using a reinforcement learning model to generate a dynamic archiving strategy, adaptively adjusting the organizational structure of archives based on user behavior, and defining specific archiving rules based on user request speed and data storage efficiency. The archiving rules include priorities and automatic archiving trigger conditions. The sequence-to-sequence generation model specifically employs an encoder-decoder architecture, where the encoder uses a transformer model to generate an embedded representation of the input sequence, and the decoder uses a transformer architecture combined with a multi-head self-attention mechanism to generate the target sequence. During decoding, a beam search strategy and a copy mechanism are combined to improve the stability and accuracy of generated summaries. The beam search strategy is as follows: set the number of candidate sequences e to be retained at each time step; at each time step t, calculate the probability of each candidate sequence being extended with the next word, and retain the e extended sequences with the highest probability as candidate sequences for the next time step, until the maximum sequence length is reached or all sequences have output end markers; The Copy Mechanism specifically uses a learnable parameter Pgen to dynamically decide whether to generate words or copy words from the input text, and combines the generation probability and the copy probability as the final output word probability.

2. The archive management method based on multimodal data analysis technology according to claim 1, characterized in that: Specifically, S1 includes: collecting archival data related to electricity, including equipment maintenance records, operation logs, sensor data, and contract texts; performing text data denoising on equipment maintenance records, operation logs, and contract texts, including using TextBlob for spelling and grammar correction, unifying text formats, and removing redundant information; and performing missing value filling, smoothing, and anomaly detection and correction on sensor data.

3. The archive management method based on multimodal data analysis technology according to claim 1, characterized in that: The text classification model is built based on a convolutional neural network. Its construction process includes: using a pre-trained word embedding layer in the input layer to convert words into vectors; adding multiple convolutional layers and pooling layers to extract text features; outputting category predictions through a fully connected layer and a softmax function; and using cross-entropy loss as the loss function during training, combined with the AMSGrad optimization algorithm for weight optimization.

4. An archive management system based on multimodal data analysis technology, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory, wherein when the processor executes the computer program, the method according to any one of claims 1 to 3 is performed.

Images