A data management method and system based on artificial intelligence
By building an AI-based data management system, the problems of automatic classification, labeling, and trend prediction of multi-source heterogeneous data were solved, achieving unified cross-platform data management and efficient information retrieval, and improving the versatility and accuracy of data management.
Patent Information
- Application Number
- CN202611610068.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to handle multi-source heterogeneous data, resulting in high costs for manual annotation, insufficient data correlation mining, a lack of trend prediction capabilities, weak cross-domain migration capabilities, and an inability to achieve unified cross-platform data management.
An AI-based data management system is constructed, including modules for data acquisition and import, automatic classification and labeling, information extraction, data association and analysis, and trend prediction and analysis. It adopts deep learning neural networks and machine learning models to support automatic classification and labeling of multimodal data, construct cross-modal semantic graphs, perform data association and trend prediction, and introduce modality credibility assessment and causal invariance constraints.
It enables end-to-end multimodal data management, improves the classification and labeling accuracy and cross-modal generalization ability of complex unstructured data, supports the autonomous collection and management of multi-source data, and provides forward-looking data support and efficient information retrieval capabilities.
Smart Images

Figure CN123115267A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data management technology, specifically a data management method and system based on artificial intelligence. Background Technology
[0002] With the advent of the digital age, data is experiencing explosive growth, making the efficient management and value mining of multi-source heterogeneous data increasingly important. However, traditional data management systems struggle to handle unstructured data, suffering from problems such as high manual annotation costs, insufficient data correlation mining, and a lack of trend prediction capabilities.
[0003] In the prior art, Chinese patent application CN115374765A discloses a 5G data parsing system and method based on natural language processing computing power network. It constructs a three-layer architecture: a data access layer, an intelligent parsing layer, and an application service layer. The data access layer uniformly accesses and preprocesses text, image, audio, and video data in the 5G network. The intelligent parsing layer uses a machine learning intelligent annotation module to achieve semantic annotation, completes word segmentation, high-frequency word extraction, and topic clustering through a keyword classification and extraction module, and extracts entity and relation information based on semantic triplet technology. The application service layer relies on a knowledge graph query module to support natural language question input, realize spatiotemporal correlation annotation mining, and data correlation network construction. However, this solution is too specific to certain application scenarios, only adapting to specific data from 5G computing networks. It cannot universally process multi-source heterogeneous data from multiple fields such as industry, healthcare, and finance, and its cross-domain migration capability is weak. It lacks an independent predictive trend analysis module, making it unable to predict future trends, changes, and events based on historical data. The automatic classification and labeling uses Markov chains and Gaussian mixture models, which are insufficient for extracting deep features from complex unstructured data, resulting in low classification and labeling accuracy. Furthermore, it lacks an independent data acquisition and import module, relying on existing data transmission channels of the computing network. It does not support the independent acquisition, cleaning, transformation, and integration of multiple common data sources such as databases, sensors, and third-party APIs, making it difficult to achieve unified cross-platform data management.
[0004] Chinese patent application CN120725709A discloses a market trend analysis system based on multimodal data fusion. It collects and preprocesses three types of market-related data: text, numerical data, and images. It uses convolutional neural networks to extract image features and combines natural language processing (NLP) technology to extract text sentiment and key event information. A multimodal feature fusion module aligns and fuses feature vectors from different data types. Finally, it trains a machine learning prediction model based on the fused features to predict and warn of market price, sales, and other trends. However, this solution is functionally limited, focusing only on market trend prediction and failing to cover core data management aspects such as automatic data classification and labeling, intelligent natural language retrieval, and data association network construction. It also has limited data type support, only including text, numerical data, and images, and does not support the processing and analysis of audio and video data. Furthermore, it lacks a complete management chain from data collection, import, and storage to analysis and application, requiring reliance on external systems for data preprocessing and storage, resulting in poor system independence. Summary of the Invention
[0005] The purpose of this invention is to provide a data management method and system based on artificial intelligence to solve the problems raised in the prior art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-based data management system includes: a data acquisition and import module for acquiring and importing multimodal data from multiple data sources, and cleaning, transforming, and integrating the acquired data, wherein the multimodal data includes text data, image data, audio data, and video data; an automatic classification and labeling module connected to the data acquisition and import module for automatically classifying and labeling the multimodal data using a deep learning neural network; an information extraction module connected to the automatic classification and labeling module for extracting key information from the classified and labeled data using natural language processing and image analysis techniques, transforming the raw data into structured information; a data association and analysis module connected to the information extraction module for constructing a network of relationships between data; a trend prediction and analysis module connected to the data association and analysis module for using machine learning models to identify and learn patterns and regularities in the data, and predict future trends, changes, and events; and an intelligent retrieval module for enabling users to retrieve relevant information from the data management system through natural language queries using natural language processing techniques.
[0007] Furthermore, the data acquisition and import module is configured to acquire data from at least one data source among databases, files, sensors, and APIs, and supports periodic or real-time data synchronization and updates; and / or, the data acquisition and import module is also used to perform data encryption and access control during the acquisition and import process.
[0008] Furthermore, the automatic classification and annotation module is configured to: convert text data into semantic word vectors, convert image data into multi-channel pixel value matrices, convert audio data into Mel-spectrum feature vectors, and decompose video data into frame image sequences and audio sequences; and automatically match deep learning neural network structures according to the modality type of the input data, wherein a convolutional neural network with multi-scale feature fusion is selected for image and video frame-type visual data, and a recurrent neural network with long temporal dependency modeling is selected for text and audio transcription-type sequence data.
[0009] Furthermore, the automatic classification and labeling module includes a modality credibility evaluation sub-network. The modality credibility evaluation sub-network takes at least one of the following statistics as input: signal-to-noise ratio, missing rate, outlier ratio, cross-modal consistency score, and historical task error of each modality. It models the model using Bayesian inference or Dirichlet distribution, outputs the credibility weight of each modality in the current sample, current task, and current environment, and updates the credibility weight online.
[0010] Furthermore, the automatic classification and labeling module is configured to use differentiable neural architecture search and meta-learning algorithms to automatically match or generate deep learning neural network structures in the candidate network component space based on the modality type, data quality, credibility weight and task objective of the input data; wherein, for cross-modal associated data, graph neural networks and graph attention mechanisms are introduced to construct cross-modal semantic graphs.
[0011] Furthermore, the automatic classification and labeling module constructs an end-to-end multi-task joint classification and labeling network, which includes a shared multimodal encoder, modal branches, a shared classification layer, an information extraction alignment layer, an association graph constraint layer, and a causal invariance constraint layer. Classification and labeling, information extraction, association network construction, and trend prediction share the shared multimodal encoder and are jointly optimized through multi-task learning.
[0012] Furthermore, the automatic classification and labeling module is configured during training as follows: a multimodal weighted backpropagation algorithm is used, and the gradients of each modality branch are scaled according to the corresponding confidence weights; when the gradient directions of different modality branches conflict, gradient projection or Pareto optimization algorithm is used to resolve the conflict; when a certain modality is missing, a diffusion model or variational autoencoder is used to generate the semantic representation of the missing modality as a soft label to participate in training.
[0013] Furthermore, the intelligent retrieval module is configured to perform semantic intent parsing on the natural language query input by the user, convert the parsing results into structured data operation instructions, and support at least one of text query, image query and voice query; the information extraction module is configured to extract entities, relationships and features from text data, identify objects, scenes and features from image data, and identify speech content from audio data and convert it into text information.
[0014] Furthermore, the data association and analysis module is configured to construct a graph-structured association network with data items as nodes and different types of associations as edges, and to perform association rule mining, trend pattern analysis, and anomaly detection; the trend prediction analysis module is configured to perform time series analysis and statistical analysis on historical data, extract at least one feature from time, quantity, and influencing factors, train a deep learning neural network prediction model, output the prediction results of future trends, changes, and events, and display them visually.
[0015] This invention also provides an artificial intelligence-based data management method, using the system described in any of the above claims, comprising the following steps: S1, collecting multimodal data from at least one data source among databases, files, sensors, and APIs through the data acquisition and import module, and cleaning, transforming, and integrating the collected data; the multimodal data includes text data, image data, audio data, and video data; S2, automatically classifying and labeling the preprocessed multimodal data using a deep learning neural network through the automatic classification and labeling module; wherein, different modal data are mapped to a shared semantic space, constructing a multi-task joint classification and labeling network including a shared multimodal encoder, modal branches, a shared classification layer, an information extraction alignment layer, an association graph constraint layer, and a causal invariance constraint layer, and performing multimodal joint training based on the confidence weights of each modality; S3, using natural language processing and image analysis techniques through the information extraction module to extract data from the classified and labeled data... Extract key information from the data, including entities, relationships, and features, and transform the raw data into structured information; S4. Receive queries input by users in at least one of natural language, image, and voice through the intelligent retrieval module, perform semantic intent parsing on the queries and transform them into structured data operation instructions, retrieve and sort relevant information from the data management system, and output it; S5. Construct an association network using data items as nodes and different types of associations as edges through the data association and analysis module, and perform association rule mining, trend pattern analysis, and anomaly detection; S6. Based on historical data and / or the association network, use a machine learning model to perform time series analysis and pattern learning through the prediction trend analysis module to predict future trends, changes, and events; S7. Display at least one of the classification labeling results, information extraction results, association network, retrieval results, and prediction results in a visualization manner such as charts, heatmaps, association network diagrams, or prediction curves.
[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a general-purpose, full-process, multimodal data management system that covers the entire data lifecycle. For the first time, it deeply integrates six core modules: data collection and import, automatic classification and labeling, information extraction, data association and analysis, trend prediction and analysis, and intelligent retrieval. This forms a complete closed loop from data collection, preprocessing, classification and labeling, information extraction, association analysis to trend prediction. It is applicable to all industries such as industry, healthcare, and finance, and fundamentally solves the core pain points of existing technologies, such as poor universality and incomplete functional coverage.
[0017] 2. This invention adds an independent predictive trend analysis module, which learns patterns and rules from historical data based on machine learning models. It can predict future trends, changes and events in multiple fields, upgrading data management from post-analysis to pre-prediction, and providing forward-looking data support for business decisions.
[0018] 3. This invention adopts a standardized deep learning architecture combining CNN and RNN with backpropagation and cross-entropy loss to achieve automatic classification and labeling. It also introduces mechanisms such as modality credibility evaluation, multimodal joint training, and causal invariance constraints, which significantly improves the classification and labeling accuracy and cross-modal generalization ability of complex unstructured data, and solves the problems of low classification and labeling accuracy and insufficient deep feature extraction in existing technologies.
[0019] 4. This invention features an independent multi-source data acquisition and import module that supports independent access to multiple data sources such as databases, files, sensors, and third-party APIs. It automatically completes cleaning, conversion, integration, deduplication, and real-time synchronization, achieving unified acquisition and management of cross-platform and cross-format data, and providing a high-quality data foundation for end-to-end data processing.
[0020] 5. This invention deeply integrates a semantic-level natural language intelligent retrieval module, which supports multimodal queries such as text, images, and voice. It can parse the semantic intent of the user's natural language query and automatically convert it into data operation instructions, which greatly reduces the threshold for system use, improves the accuracy and efficiency of information retrieval, and realizes natural interaction between humans and the system. Attached Figure Description
[0021] Figure 1 This is a system architecture diagram of an artificial intelligence-based data management system according to the present invention; Figure 2 This is a schematic diagram illustrating the steps of a data management method based on artificial intelligence according to the present invention. Detailed Implementation
[0022] 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 skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] Example: Figures 1-2 As shown, the present invention provides a technical solution: like Figure 1As shown, this invention provides an artificial intelligence-based data management system, whose core modules include: a data acquisition and import module, an automatic classification and labeling module, an information extraction module, a data association and analysis module, a trend prediction and analysis module, and an intelligent retrieval module. The modules interact with each other via a data bus or internal interfaces. Multimodal data acquired by the data acquisition and import module is preprocessed and then pushed to the automatic classification and labeling module; the classification and labeling results output by the automatic classification and labeling module are simultaneously pushed to the information extraction module; the structured information output by the information extraction module is input to the data association and analysis module; the analysis results from the data association and analysis module are input to the trend prediction and analysis module and the intelligent retrieval module, respectively; the intelligent retrieval module supports interactive queries by users through natural language and other methods. The system as a whole forms a complete closed loop from data acquisition, preprocessing, classification and labeling, information extraction, association analysis to trend prediction.
[0024] The data acquisition and import module is responsible for collecting and importing data from various data sources, and performing data cleaning, transformation, and integration to ensure data quality. These data sources include, but are not limited to, databases, file systems, sensors, and third-party APIs. Different data sources store data in different formats and structures, and this module has the capability to process different data formats.
[0025] Specifically, before collecting data, this module adapts to and connects to different data sources, including configuring database connection strings, setting API access keys, and establishing sensor communication protocols to ensure orderly data acquisition. After connecting to the data source, the module retrieves data from the data source according to predetermined time intervals or trigger conditions. Data extraction methods include executing SQL queries, calling RESTful APIs, reading file streams, and receiving sensor push notifications.
[0026] The collected raw data may contain noise, outliers, or missing values. This module performs basic data preprocessing operations, including: removing duplicate data, filling missing values, smoothing noise, standardizing time format, and standardizing encoding format. The preprocessed data is imported into the corresponding data storage area in the data management system. The data storage area can be a relational database, NoSQL database, data warehouse, cloud storage, etc., selected according to actual needs.
[0027] For dynamically changing data sources, this module supports periodic or real-time data synchronization and updates to ensure the data in the system remains up-to-date. During data acquisition and import, this module also incorporates measures such as data encryption and access control to protect data privacy and security. The multimodal data includes text, image, audio, and video data.
[0028] The automatic classification and labeling module utilizes deep learning neural networks to automatically classify and label data, providing a foundation for subsequent analysis. For example... Figure 1 As shown, this module receives the preprocessed data output from the data acquisition and import module, and outputs the classified and labeled data to the information extraction module.
[0029] The operation steps for this module are as follows: S2-1. Receive the preprocessed data output from the data acquisition and import module, and organize the training dataset into corresponding feature vectors and category labels under a unified classification label system according to modality type. Specifically, text data is converted into semantic word vectors, image data into multi-channel pixel value matrices, audio data into Mel-spectrum feature vectors, and video data is decomposed into frame image sequences and audio sequences, which are then converted separately. The unified classification label system is aligned with the entity and relation extraction labels of the information extraction module.
[0030] Specifically, this module first performs modal encoding on text, images, audio, and video to obtain initial semantic representations for each modality. Then, it maps the data distribution of different modalities to a unified shared semantic space using an optimal transmission algorithm. The module treats samples from different modalities of the same event as positive sample pairs and samples from different events as negative sample pairs. It uses contrastive learning to bring positive sample pairs closer together and push negative sample pairs further apart in the shared semantic space. At the same time, it employs an information bottleneck algorithm to constrain the shared representation, ensuring that the representation retains as much information as possible that is relevant to the classification task, while compressing modal noise and redundant information that is irrelevant to the task.
[0031] Building upon this foundation, the module constructs a sub-network for modal credibility assessment. This sub-network takes as input statistics such as signal-to-noise ratio, missing rate, outlier ratio, cross-modal consistency score, and historical task error for each modality. Through Bayesian inference or Dirichlet distribution modeling, it outputs the credibility weight of each modality in the current sample, current task, and current environment. This weight is not a fixed hyperparameter but is updated online as data quality, task type, and environment change.
[0032] S2-2. Based on the modality type, data quality, confidence weights, and task objectives of the input data, automatically match or generate deep learning neural network structures. This process employs differentiable neural architecture search and meta-learning algorithms to search for the optimal structure in the candidate network component space.
[0033] For visual data such as images and video frames, the module searches the candidate space for combinations of components such as multi-scale convolution, dilated convolution, Transformer attention, and graph attention to form a visual encoder with multi-scale feature fusion. For sequence data such as text and audio transcription, the module searches for combinations of components such as gated recurrent units, temporal convolutional networks, and self-attention mechanisms to form a sequence encoder with long temporal dependency modeling. For cross-modal related data, the module introduces graph neural networks and graph attention mechanisms to construct a cross-modal semantic graph, enabling message passing between nodes of different modalities.
[0034] The module uses classification error, cross-modal consistency error, information bottleneck regularization term, and credibility regularization term as its meta-learning objective, and updates the network structure parameters and network weight parameters through two-layer optimization. When the data distribution changes, the meta-learner can quickly adjust the network structure or branch weights based on a small number of new samples, achieving cross-domain transfer.
[0035] S2-3. Based on the matched network structure, an end-to-end multi-task joint classification and labeling network is constructed. This network includes a shared multimodal encoder, modal branches, a shared classification layer, an information extraction and alignment layer, an association graph constraint layer, and a causal invariance constraint layer. Classification and labeling, information extraction, association network construction, and trend prediction share the underlying multimodal encoder and achieve joint optimization through multi-task learning.
[0036] The module employs a variational information bottleneck-constrained shared encoder, ensuring that the shared representation retains classification information while minimizing mode-specific noise and spurious correlations. The module further introduces a causal invariance risk minimization algorithm: training data is divided into multiple environments based on time, region, device, business, and user group. The classifier's prediction risk is required to remain consistent across all environments, and features that are effective only in specific environments but unstable across different environments are penalized. Through conditional independence tests and intervention stability scoring, the module selects causal invariant features from candidate features and suppresses spurious correlations.
[0037] Simultaneously, the module constructs multi-task consistency constraints: classification labels are aligned with entity and relation labels extracted from information; the edges of the association network constrain the classification results to avoid semantic contradictions; trend prediction errors are backpropagated to the classification and association modules to optimize the underlying representation. Each task automatically adjusts its loss weight through homoscedastic uncertainty weighting, with tasks with high uncertainty having lower weights and tasks with low uncertainty having higher weights.
[0038] S2-4. Jointly train the network constructed in S2-3 using multimodal training data. During training, training data from each modality are synchronously input into the corresponding network branch, and each branch performs forward propagation through its respective network layer. The outputs of each branch are integrated by a shared classification layer to obtain the final classification prediction result. Based on the real labels and prediction results of the unified classification label system, and combined with the confidence weights generated in S2-1, the module calculates the multimodal joint loss function value.
[0039] Specifically, the gradients of each modality branch are scaled according to their corresponding confidence weights. When the confidence of a modality falls below a dynamic threshold, its gradient contribution is automatically reduced, triggering missing modality completion or resampling. When gradient directions of different modality branches conflict, gradient projection or Pareto optimization algorithms are used to project conflicting gradients to non-conflicting directions, preventing one modality from dominating training. Meta-learning algorithms are used to optimize the confidence evaluation network and task weights, enabling the model to quickly adapt to a small number of new domain samples. Classification risk is simultaneously optimized across multiple training environments, requiring consistent risk across environments and penalizing environment-related features. Counterfactual samples are generated through counterfactual reasoning, and counterfactual consistency constraints are used to enhance model robustness. When a modality is missing, a diffusion model or variational autoencoder is used to generate a semantic representation of the missing modality, which is then used as a soft label in training.
[0040] The module uses the gradient descent optimization algorithm to synchronously update the weights and bias parameters of all network branches based on the calculated gradients of each layer, in order to minimize the value of the multimodal joint loss function. It repeatedly performs the steps of multimodal data forward propagation, joint loss calculation, weighted backpropagation, gradient conflict resolution, and parameter synchronization update, adjusting the network parameters through multiple iterations until the multimodal joint loss function gradually converges. The training process stops when the loss function converges to a preset threshold or reaches a preset number of iterations.
[0041] The loss function used is cross-entropy loss, and the calculation steps are as follows: A training set is provided, and each sample contains a set of features x and corresponding class labels. Design a deep learning neural network architecture, including an input layer, hidden layers, and an output layer; input a set of features x through the neural network and perform forward propagation to obtain the probability distribution of the output nodes. The cross-entropy loss is used to calculate the loss for each sample. For a single sample... The formula for calculating cross-entropy loss is: ; in It is the cth type of the real label. This represents the probability of class c predicted by the model. The overall loss function L is obtained by averaging or summing the losses of all samples. The average cross-entropy loss can be expressed as: ; Where N is the number of training samples.
[0042] The key mathematical formulas of the backpropagation algorithm are as follows: For each node j in the output layer, its gradient... Calculated based on the partial derivative of the loss function with respect to the node output: ; in L is the gradient, and L is the loss function. This is the input to node j. For the hidden layer, its gradient is calculated layer by layer using the chain rule: ; in, Here, L is the gradient, L is the loss function, l represents the current layer, and k represents the next layer. It is the input of the node. This represents the weighted sum of inputs from the k-th node in the next layer (l+1). These are the activation values of the nodes. Finally, gradient descent is used to update the weights in the network. and bias Their gradients are calculated using the following formulas: ; ; in It is the activation value of node i in the input layer or the previous hidden layer.
[0043] S2-5. Cross-validate the model during training using multimodal test data. Validation includes not only conventional random partitioning validation but also multi-environment out-of-systems validation: dividing the model into multiple test environments based on time, region, device, and business, and checking the model's classification accuracy, uncertainty, and cross-modal consistency in each environment. Based on the classification accuracy, prediction uncertainty, cross-modal consistency score, and causal invariance test results for each modality, the module dynamically adjusts the modality weights, network regularization parameters, and network structure.
[0044] Specifically, the module employs Bayesian neural networks or Monte Carlo dropout to estimate prediction uncertainty; it uses causal invariance tests to screen for environmentally invariant features; and it uses counterfactual augmentation to generate counterfactual verification samples to check the model's stability under intervention conditions. When a modality's performance deteriorates in an out-of-distribution environment, the module automatically reduces its credibility weight and triggers a meta-learner to quickly fine-tune that modality branch. The module also uses topological data analysis or manifold learning to detect anomalous distributions in the cross-modal semantic space, suppressing overfitting and improving cross-modal generalization ability.
[0045] S2-6. Deploy the trained deep learning model as an online inference engine to automatically classify and semantically label the unlabeled multimodal data input in real time from the data acquisition and import module. During online inference, the module first calculates the confidence weights of each modality of the current sample through the confidence evaluation sub-network in S2-1; then it performs forward propagation through the network structure matched in S2-2; finally, it combines the multi-task joint network in S2-3 to output the classification labels, confidence scores, and semantic annotation results.
[0046] The output results are synchronously pushed to the information extraction module. For samples with confidence levels below the threshold, the module automatically triggers manual review or proactive sampling requests; for high-confidence samples, the module adds them as pseudo-labels to the supervised training set, and updates the confidence evaluation network, cross-modal alignment network, and classification network through continuous meta-learning. The module also employs an online causal discovery algorithm to continuously update the time-delay causal graph based on new data, ensuring consistency between classification labeling results and causal association analysis.
[0047] Furthermore, the module supports interpretable output: for each classification labeling result, it outputs the primary modalities it depends on, confidence weights, causal features, and environment invariance scores, making it easier for users to understand the model's decision-making basis. The module integrates manually corrected results, prediction errors, and retrieval feedback into a closed loop, enabling continuous model evolution.
[0048] The information extraction module can handle various data types, including text, images, and audio. For text data, the module identifies and extracts keywords, entities, sentence relationships, and other information. For example, it can extract entities such as people, locations, and times, as well as the relationships between them, from a news article. When processing image data, the module identifies objects, scenes, and features in the image and extracts this information. For example, it can extract information such as mountains, lakes, and buildings from a landscape photograph. For audio data, the module identifies sounds and speech content in the audio and converts them into usable text information. This is valuable for processing scenarios such as telephone recordings and voice commands.
[0049] During the information extraction process, the module utilizes natural language processing and image analysis techniques to transform raw data into machine-understandable feature vectors. These feature vectors capture important information and patterns within the data. Simultaneously, the module identifies relationships within the data, such as semantic relationships in text and connections between objects in images. The information extraction module can extract information from different data sources in real time, maintaining the data's up-to-date state and providing an accurate information foundation for subsequent analysis and applications.
[0050] The Natural Language Processing (NLP) technology is used to process text data. The data management system utilizes NLP technology to allow users to perform intelligent searches and queries using natural language. The data management system uses NLP technology to extract key information from text data, including entities, relationships, and features. Through NLP technology, the system automatically classifies and labels text data, categorizing it into different classes. When processing large amounts of text data, NLP technology can generate text summaries and perform automatic translation. The image analysis technology is used to process image data. The data management system uses image analysis technology to extract key information from image data, automatically identify different objects and structures in the images, and perform labeling and classification, thereby achieving intelligent analysis and management of image data. Image analysis technology can help the system automatically classify and label image data and can be used for sentiment analysis.
[0051] The data association and analysis module constructs a network of relationships between data based on their attributes and characteristics. This network can be a graph structure, where data items act as nodes and different types of associations form edges, thus revealing the connections and dependencies between data. This module uses association rule mining techniques to discover frequently occurring sets of related items from large-scale datasets. For example, it can mine association rules for purchasing goods from transaction data, helping merchants understand user purchasing habits. The module can analyze data trends and patterns, helping users understand data changes and key patterns. This helps in discovering market trends, changes in user behavior, and other information. The module can detect anomalies in the data, i.e., data that deviates from conventional patterns. Anomaly detection helps uncover potential problems and risks, such as financial fraud detection. Analysis results can be displayed visually, such as charts, heatmaps, and network graphs. The analysis results can provide decision support for users, helping them make more informed decisions.
[0052] The predictive trend analysis module first analyzes historical data to identify patterns, periodicity, and trends. This involves techniques such as time series analysis and statistical analysis to obtain the data's changing patterns. From the historical data, the module extracts important features, such as time, quantity, and influencing factors, as input features for the predictive model. The predictive trend analysis module uses a deep learning neural network to build a predictive model. The neural network model is trained using historical data to predict future trends. During training, the model adjusts the network weights based on the error between actual observations and predicted values, gradually optimizing it. The trained neural network model can be used to predict future data trends. By inputting data from previous time steps, the model can output a predicted value for a future time step. The model is evaluated using a validation dataset to check its predictive accuracy and performance. If necessary, the model can be adjusted and optimized. The prediction results are visualized and presented to users in the form of trend charts, prediction curves, etc., to help them better understand the future development trends of the data.
[0053] The intelligent retrieval module allows users to submit queries to the system without using complex query language. They can simply ask or describe the query in a manner similar to natural language. For example, a user can directly ask, "What was the product with the highest sales volume in the past year?" without writing complex SQL statements. This module possesses natural language understanding capabilities, analyzing the semantics of the user's query and translating it into machine-understandable query statements. The module parses the user's query request into specific data operations, such as retrieval, sorting, and filtering. Based on the user's intent, the system can determine which information needs to be extracted from the data storage. After parsing the query, the module matches and retrieves data from the data storage according to the user's needs, obtaining data relevant to the query. Retrieval can be based on keyword matching, semantic similarity, and other methods. The intelligent retrieval module ranks the search results, presenting the most relevant data results to the user first. Furthermore, the module supports not only text queries but also image queries, voice queries, and other query modes. Query results can reflect data changes in real time, keeping the data up-to-date.
[0054] like Figure 2 As shown, the present invention also provides a data management method based on artificial intelligence, the steps of which are as follows: S8-1. Collect data; the data includes text data, image data, audio data, and video data; S8-2. Perform preprocessing operations on the collected data; S8-3. Use machine learning algorithms to automatically classify and label data; S8-4. Using natural language processing and image analysis techniques, extract key information from the data, including entities, relationships, and features; S8-5. Implement intelligent search and query functions based on natural language; S8-6. Utilize data association analysis techniques to construct a network of relationships between data; S8-7. Based on the results of data correlation analysis, use machine learning models for prediction and trend analysis; S8-8. Present the analysis results in a visual format using charts and images.
[0055] In Example 1, the system interfaces with the network management system, customer service platform, spectrum acquisition equipment, and data center monitoring equipment to collect four types of multimodal data: text alarm work orders, spectrum waveform images, user complaint recordings, and data center monitoring videos. During data collection, basic preprocessing such as deduplication, noise reduction, and format standardization is automatically performed. After processing, the data is pushed to the automatic classification and labeling module.
[0056] The automatic classification and labeling module employs a multimodal deep learning network. Text work orders are converted into semantic vectors using a recurrent network, while waveform images utilize a multi-scale convolutional network. Model training is completed using a weighted backpropagation algorithm. The system automatically distinguishes between labels such as transmission faults, base station interference, and user-perceived anomalies. Labeling results are simultaneously sent to the information extraction module, replacing manual labeling. The information extraction module combines NLP and image analysis techniques to process various maintenance data: for alarm texts, it extracts entity and causal relationships such as network element number, fault time, and coverage area; for spectrum images, it identifies interference values and waveform anomalies; and for complaint audio transcribed into text, it extracts fault location and user issues, all converted into standardized structured data.
[0057] Operations and maintenance personnel can directly initiate queries using natural language, without needing to write professional query statements. For example, by entering "alarms and corresponding spectrum diagrams of high-interference cells in the western part of the city in the past half month", the system can parse the query semantics to complete cross-text and image data retrieval, display the results sorted by relevance, and also support voice and image upload retrieval.
[0058] The data association and analysis module constructs an association network using structured maintenance events as nodes and fault causality and spatial correlation as edges. It mines and analyzes fault propagation links through association rules, detects anomalies in indicators, and visualizes equipment dependencies and fault impact ranges using network graphs and heatmaps, assisting staff in quickly locating the root cause of faults. The prediction and trend analysis module extracts features such as historical alarms, equipment load, and complaint volume to train a deep learning model for time series analysis. It can predict base station outages, high-interference periods and areas for the next week, displaying the prediction results in trend charts and providing maintenance decision-making suggestions such as advance inspections and cell optimization.
[0059] In Example 2, in step S8-1, the user collects personal health data, including steps, heart rate, and sleep quality, through a health app or sensor device. In step S8-2, the collected medical data is denoised, normalized, and standardized to ensure data consistency and analyzability. In step S8-3, machine learning algorithms are used to automatically classify and label patient diagnoses and symptoms, such as tumor type and disease severity. In step S8-4, natural language processing and image analysis technologies are used to extract key information from medical literature and medical images, such as symptom descriptions and image features. In step S8-5, doctors can perform intelligent searches and queries using natural language, such as "find cases related to a certain disease," and the system returns relevant data and analysis results. In step S8-6, data association analysis technology is used to analyze the correlation between medical data, such as the relationship between diseases and genetic factors, constructing a correlation network between data. In step S8-7, based on the correlation network between data, machine learning models are used for prediction, such as predicting the future development trend of a certain disease, to help doctors make better diagnostic and treatment decisions. In step S8-8, the medical data analysis results are visualized in the form of charts, images, etc., so that doctors can have a clearer understanding of the patient's condition and data trends.
[0060] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A data management system based on artificial intelligence, characterized in that, include: The data acquisition and import module is used to acquire and import multimodal data from multiple data sources, and to clean, transform, and integrate the acquired data; the multimodal data includes text data, image data, audio data, and video data. An automatic classification and labeling module, connected to the data acquisition and import module, is used to automatically classify and label the multimodal data using a deep learning neural network; The information extraction module, connected to the automatic classification and labeling module, is used to extract key information from the classified and labeled data using natural language processing and image analysis techniques, and to transform the raw data into structured information. The data association and analysis module, connected to the information extraction module, is used to construct a network of associations between data. The trend prediction and analysis module is connected to the data association and analysis module and is used to identify and learn patterns and regularities in the data using machine learning models to predict future trends, changes and events. The intelligent retrieval module is used to enable users to retrieve relevant information from the data management system through natural language queries, utilizing natural language processing technology.
2. The data management system based on artificial intelligence according to claim 1, characterized in that, The data acquisition and import module is configured to acquire data from at least one data source among databases, files, sensors, and APIs, and supports periodic or real-time data synchronization and updates; the data acquisition and import module is also used to perform data encryption and access control during the acquisition and import process.
3. The data management system based on artificial intelligence according to claim 2, characterized in that, The automatic classification and annotation module is configured to: convert text data into semantic word vectors, image data into multi-channel pixel value matrices, audio data into Mel-spectrum feature vectors, and video data into frame image sequences and audio sequences; and automatically match deep learning neural network structures according to the modality type of the input data, wherein a convolutional neural network with multi-scale feature fusion is selected for image and video frame-type visual data, and a recurrent neural network with long temporal dependency modeling is selected for text and audio transcription-type sequence data.
4. The data management system based on artificial intelligence according to claim 3, characterized in that, The automatic classification and labeling module includes a modality credibility evaluation sub-network. The modality credibility evaluation sub-network takes at least one of the following statistics as input: signal-to-noise ratio, missing rate, outlier ratio, cross-modal consistency score, and historical task error for each modality. It models the modality using Bayesian inference or Dirichlet distribution, outputs the credibility weight of each modality in the current sample, current task, and current environment, and updates the credibility weight online.
5. A data management system based on artificial intelligence according to claim 3, characterized in that, The automatic classification and labeling module is configured to use differentiable neural architecture search and meta-learning algorithms to automatically match or generate deep learning neural network structures in the candidate network component space based on the modality type, data quality, credibility weight and task objective of the input data; wherein, for cross-modal associated data, graph neural networks and graph attention mechanisms are introduced to construct cross-modal semantic graphs.
6. A data management system based on artificial intelligence according to claim 3, characterized in that, The automatic classification and labeling module constructs an end-to-end multi-task joint classification and labeling network, which includes a shared multimodal encoder, modal branches, a shared classification layer, an information extraction alignment layer, an association graph constraint layer, and a causal invariance constraint layer. Classification and labeling, information extraction, association network construction, and trend prediction share the shared multimodal encoder and are jointly optimized through multi-task learning.
7. A data management system based on artificial intelligence according to claim 6, characterized in that, The automatic classification and labeling module is configured during training as follows: a multimodal weighted backpropagation algorithm is used, and the gradients of each modality branch are scaled according to the corresponding confidence weights; when the gradient directions of different modality branches conflict, gradient projection or Pareto optimization algorithm is used to resolve the conflict; when a certain modality is missing, a diffusion model or variational autoencoder is used to generate the semantic representation of the missing modality as a soft label to participate in training.
8. A data management system based on artificial intelligence according to claim 1, characterized in that, The intelligent retrieval module is configured to perform semantic intent parsing on natural language queries input by users, convert the parsing results into structured data operation instructions, and support at least one of text queries, image queries, and voice queries; the information extraction module is configured to extract entities, relationships, and features from text data, identify objects, scenes, and features from image data, and identify speech content from audio data and convert it into text information.
9. A data management system based on artificial intelligence according to claim 1, characterized in that, The data association and analysis module is configured to construct a graph-structured association network with data items as nodes and different types of associations as edges, and to perform association rule mining, trend pattern analysis, and anomaly detection. The trend prediction and analysis module is configured to perform time series analysis and statistical analysis on historical data, extract at least one feature from time, quantity, and influencing factors, train a deep learning neural network prediction model, and output the prediction results of future trends, changes, and events for visualization.
10. A data management method based on artificial intelligence, using the data management system based on artificial intelligence according to any one of claims 1-9, characterized in that, Includes the following steps: S1. The data acquisition and import module acquires multimodal data from at least one data source among databases, files, sensors, and APIs, and cleans, transforms, and integrates the acquired data; the multimodal data includes text data, image data, audio data, and video data. S2. The preprocessed multimodal data is automatically classified and labeled using a deep learning neural network through the automatic classification and labeling module. Specifically, different modal data are mapped to a shared semantic space to construct a multi-task joint classification and labeling network that includes a shared multimodal encoder, modal branches, a shared classification layer, an information extraction alignment layer, an association graph constraint layer, and a causal invariance constraint layer. Multimodal joint training is then performed based on the confidence weights of each modality. S3. The information extraction module uses natural language processing and image analysis technology to extract key information from the classified and labeled data. The key information includes entities, relationships and features, and transforms the raw data into structured information. S4. The intelligent retrieval module receives queries input by users in at least one of natural language, image, and voice, performs semantic intent parsing on the queries and converts them into structured data operation instructions, and retrieves and sorts relevant information from the data management system and outputs it. S5. Construct an association network using data items as nodes and different types of associations as edges through the data association and analysis module, and perform association rule mining, trend pattern analysis and anomaly detection. S6. Based on historical data and / or the correlation network, the prediction trend analysis module uses a machine learning model to perform time series analysis and pattern learning to predict future trends, changes and events. S7. Display at least one of the following: classification labeling results, information extraction results, association network, retrieval results, and prediction results, in a visual manner such as a chart, heatmap, association network diagram, or prediction curve.
Citation Information
Patent Citations
Computing power network 5G data analysis system and method based on natural language processing
CN115374765A
Market trend analysis system based on multi-modal data fusion
CN120725709A
