Heterogeneous data feature fusion processing system based on artificial intelligence

Abstract

The invention relates to the technical field of data processing systems, and discloses an artificial intelligence-based heterogeneous data feature fusion processing system, which comprises an AI big data acquisition module, an AI preprocessing module, an AI big data storage management module, an AI feature extraction module, an AI association determination module, an AI fusion decision module, an AI monitoring module and an AI optimization module. The AI big data acquisition module comprises an acquisition unit and a storage unit, the acquisition unit integrates an AI scheduling model to dynamically adjust an acquisition strategy according to historical data requirements and real-time business scenes, and the storage unit performs preliminary intelligent classification on acquired original data to generate data labels. Through full-process AI driving, each link from acquisition to optimization is embedded into an AI model, strategy dynamic adjustment is realized, traditional manual configuration is replaced, and the system adaptability is improved. According to the method, a multi-level fusion mechanism is adopted, feature level and decision level fusion is combined, feature complementarity and decision reliability are considered, and the limitation of a single fusion level is broken through.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing systems, and in particular, relates to an artificial intelligence-based heterogeneous data feature fusion processing system. Background Art

[0002] With the rapid development of information technology, data generated by various industries (such as power systems, intelligent manufacturing, smart cities, and the Internet of Things) is becoming multimodal, heterogeneous, and rapidly growing. For example, power systems require the integration of equipment monitoring image data (infrared thermal imaging), sensor time series data (voltage / current waveforms), inspection voice recordings (equipment abnormal noise), and 3D point cloud data (transmission line spatial structure). Intelligent manufacturing requires the integration of industrial camera images, robotic arm motion trajectory data, and equipment vibration signals. These heterogeneous data contain complementary information, and feature fusion can significantly improve the accuracy of data analysis (e.g., fault diagnosis and situation prediction). Therefore, feature fusion processing of heterogeneous data has become a core technology in the field of data intelligence.

[0003] The current heterogeneous data fusion processing system generally has the following technical bottlenecks: First, it relies on manually preset acquisition frequency and range, and cannot be dynamically adjusted according to real-time business needs; the noise characteristics and missing patterns of different modal data (such as images, voice, and point clouds) are significantly different, but the existing system uses a unified cleaning and normalization method (such as fixed threshold noise filtering and manually set normalization parameters), resulting in data quality loss (such as sparse noise in point cloud data is not effectively removed, and missing frame repair of voice signals is distorted); Second, for modalities such as images, videos, and voice, it is necessary to manually configure the corresponding deep learning model (such as manually selecting R esNet extracts image features, LSTM processes time series data), lacks an automated model adaptation mechanism, increases development costs and makes it difficult to cope with new data modalities. Existing technologies mostly use statistical methods (such as correlation coefficients) or simple neural networks to analyze feature associations, which cannot capture complex semantic associations, resulting in insufficient feature synergy during fusion; third, existing systems mostly only perform feature-level fusion or decision-level fusion, without combining the advantages of both (feature-level fusion retains details but is computationally complex, while decision-level fusion is efficient but prone to losing cross-modal associations), resulting in insufficient decision-making accuracy in complex scenarios (such as difficulty in comprehensive judgment when multi-source data conflict).

[0004] In response to the above needs, this patent proposes a heterogeneous data feature fusion processing system based on artificial intelligence. By embedding AI models in the entire process of acquisition, preprocessing, feature extraction, association determination, fusion decision-making, monitoring optimization, etc., it solves the problems of insufficient intelligence, poor dynamic adaptability, weak noise robustness, and lack of closed-loop optimization in existing technologies, significantly improves the efficiency and accuracy of heterogeneous data processing, and is suitable for complex data fusion scenarios in multiple fields such as electricity, manufacturing, and the Internet of Things. Summary of the Invention

[0005] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0006] An artificial intelligence-based heterogeneous data feature fusion processing system includes an AI big data acquisition module, an AI preprocessing module, an AI big data storage management module, an AI feature extraction module, an AI association determination module, an AI fusion decision module, an AI monitoring module and an AI optimization module; the AI big data acquisition module includes an acquisition unit and a storage unit, the acquisition unit integrates an AI scheduling model to dynamically adjust the acquisition strategy according to historical data requirements and real-time business scenarios, the storage unit performs preliminary intelligent classification of the collected original data, generates data labels, compares and stores historical data and real-time data through data feature similarity analysis, automatically establishes data association indexes, and supports fast retrieval and historical data backtracking; the AI preprocessing module includes an intelligent cleaning submodule and an intelligent normalization submodule, the intelligent cleaning submodule identifies and filters noise data, outliers or duplicate data based on the isolation forest anomaly detection algorithm, the intelligent cleaning submodule adopts AI-driven missing value repair, and dynamically selects the repair strategy according to the data modal characteristics, and the intelligent normalization submodule automatically identifies data distribution characteristics and generates optimal normalization parameters through an adaptive normalization model.

[0007] As a preferred embodiment of the present invention, the AI big data storage management module classifies and stores the data processed by the AI preprocessing module. The AI big data storage management module adopts a multimodal vector database and generates data semantic vectors through self-supervised learning to realize the joint indexing and rapid retrieval of unstructured data and structured data. The AI big data storage management module constructs a spatiotemporal feature comparison library and uses a contrastive learning algorithm to generate a data feature similarity matrix to support intelligent matching and incremental updating of historical data.

[0008] As a preferred embodiment of the present invention, the AI feature extraction module extracts features from the data in the AI big data storage management module, integrates dedicated deep learning models for different modal data, and automatically matches the optimal extraction model through a model selector; the AI feature extraction module uses the AI model to evaluate feature integrity, enhance low-quality features, and then preliminarily extracts shallow correlation features between modalities through a lightweight association network to provide prior information for the subsequent association determination module.

[0009] As a preferred embodiment of the present invention, the AI association determination module adopts a deep association learning model to construct a multimodal feature association graph, quantifies the semantic association between features, and adjusts the association weight according to real-time fusion effect feedback through an online learning algorithm. The AI association determination module also includes a noise feature decoupling module, which combines the noise components in the separation features with the generative adversarial network. The generator fits the noise distribution and generates fake sample features, and the discriminator distinguishes between real features and fake samples to achieve noise purification and feature enhancement.

[0010] As a preferred embodiment of the present invention, the AI fusion decision module includes a feature-level AI fusion unit, a decision-level AI fusion unit and an optimization feedback unit, wherein the feature-level AI fusion unit integrates different modal features through an adaptive fusion layer for related features by designing a multimodal fusion network, and supports dynamic adjustment of the fusion strategy based on the degree of correlation; at the same time, adversarial feature alignment is introduced to align different modal feature spaces through adversarial training to reduce the fusion error caused by modal differences; the decision-level AI fusion unit performs secondary fusion through the decision voting neural network on the modal decisions output by each modal network and the fusion decision output by the feature-level fusion network, and uses a weighted voting mechanism to generate the final fusion result.

[0011] As a preferred embodiment of the present invention, the optimization feedback unit reversely optimizes the parameters of each modal network and the fusion network through the Q-Learning reinforcement learning algorithm to form a closed-loop self-optimization system.

[0012] As a preferred embodiment of the present invention, the AI monitoring module collects the computing resource occupancy, processing delay, and data throughput indicators of each module in real time through sensors or log analysis, and warns of potential faults based on anomaly detection models.

[0013] As a preferred embodiment of the present invention, the AI optimization module uses Bayesian optimization to automatically adjust AI model hyperparameters or reallocate computing resources based on monitoring data.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] 1. This invention is driven by AI throughout the entire process, embedding AI models in every link from acquisition to optimization to achieve dynamic strategy adjustment, replacing traditional manual configuration and improving system adaptability.

[0016] 2. The present invention adopts a multi-level fusion mechanism, combining the feature level and the decision level, taking into account both feature complementarity and decision reliability, breaking through the limitations of a single fusion level.

[0017] 3. The present invention realizes closed-loop monitoring and optimization, and the real-time feedback mechanism forms a complete closed loop of "collection-processing-fusion-optimization", ensuring the long-term and efficient operation of the system and adapting to dynamic business needs.

[0018] 4. This invention realizes full-link data quality assurance, purifies data input through noise decoupling, feature enhancement, adaptive normalization and other technologies, and improves fusion accuracy from the source.

[0019] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In the attached figure:

[0021] Figure 1 This is a flow chart of the heterogeneous data feature fusion processing system of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention.

[0023] An artificial intelligence-based heterogeneous data feature fusion processing system, such as Figure 1 As shown, it includes AI big data acquisition module, AI preprocessing module, AI big data storage management module, AI feature extraction module, AI association determination module, AI fusion decision module, AI monitoring module and AI optimization module; the AI big data acquisition module includes an acquisition unit and a storage unit. The acquisition unit integrates the AI scheduling model to dynamically adjust the acquisition strategy according to historical data requirements and real-time business scenarios. The storage unit performs preliminary intelligent classification on the collected raw data, generates data labels, and compares and stores historical data with real-time data through data feature similarity analysis, automatically establishes data association indexes, and supports fast retrieval and historical data backtracking; the AI preprocessing module includes an intelligent cleaning submodule and an intelligent normalization submodule. The intelligent cleaning submodule identifies and filters noise data, outliers or duplicate data based on the anomaly detection algorithm of the isolation forest. The intelligent cleaning submodule adopts AI-driven missing value repair and dynamically selects the repair strategy according to the modal characteristics of the data. The intelligent normalization submodule automatically identifies data distribution characteristics and generates optimal normalization parameters through an adaptive normalization model.

[0024] The AI big data acquisition module starts working. The AI scheduling model integrated in the acquisition unit analyzes historical data requirements (such as the peak load period of the power system) and real-time business scenarios (such as equipment abnormality warning signals), and dynamically adjusts the acquisition strategy (such as increasing the frequency of sensor data acquisition in areas with frequent faults); parses user queries through NLP and automatically connects to the corresponding data source (image sensor, SCADA system, etc.); the storage unit performs preliminary intelligent classification of the original data (based on the deep learning classification model) and generates data labels (such as "image-transmission line-20250422" and "voice-device abnormal sound-abnormal"); calculates the similarity of data features (such as cosine similarity), compares and stores real-time data with historical data, and establishes associated indexes (such as associating point cloud data of the same device at different time periods), supporting second-level retrieval and historical data backtracking.

[0025] The intelligent cleaning submodule detects outliers (such as sudden out-of-range data from sensors) based on the isolation forest algorithm and automatically filters out noise and duplicate data. For missing data, it selects a repair strategy based on modal characteristics (such as bilinear interpolation for images and LSTM prediction for time series data). The intelligent normalization submodule analyzes data distribution (such as the image pixel value range [0,255] and sensor value [-100,100]) through an adaptive normalization model to generate optimal normalization parameters (such as standardization or normalization formulas) and unify the scale of multimodal data.

[0026] The dynamic collection strategy (AI scheduling model) of the present invention improves the targetedness of data collection and reduces invalid data redundancy; through intelligent classification and comparative storage (storage unit), rapid data retrieval (response in seconds) is achieved, and the efficiency of historical data backtracking is improved by more than 80%, providing an efficient data foundation for subsequent analysis.

[0027] The AI big data storage management module classifies and stores the data processed by the AI preprocessing module. The AI big data storage management module adopts a multimodal vector database and generates data semantic vectors through self-supervised learning to realize the joint indexing and rapid retrieval of unstructured data and structured data. The AI big data storage management module builds a spatiotemporal feature comparison library and uses a contrastive learning algorithm to generate a data feature similarity matrix to support intelligent matching and incremental updating of historical data.

[0028] The AI big data storage management module classifies and stores pre-processed data, uses a multimodal vector database to generate data semantic vectors through self-supervised learning, and realizes the joint indexing of structured data and unstructured data; uses a contrastive learning algorithm to calculate the data feature similarity matrix (such as the cosine similarity of equipment vibration data in different time periods), supports intelligent matching of historical data (such as automatically associating historical cases with similar fault scenarios) and incremental updates (real-time data is automatically integrated into the existing feature library).

[0029] The present invention supports joint retrieval of structured and unstructured data through a dynamic multimodal vector database, improving retrieval efficiency by 50%; it realizes intelligent matching of similar historical data (such as automatically associating historical failure cases of similar equipment) through a spatiotemporal feature comparison library, assists incremental learning, and reduces repeated calculations.

[0030] The AI feature extraction module extracts features from the data in the AI big data storage management module, integrates dedicated deep learning models for different modal data, and automatically matches the optimal extraction model through a model selector; the AI feature extraction module uses the AI model to evaluate feature integrity, enhance low-quality features, and then preliminarily extracts shallow correlation features between modalities through a lightweight association network to provide prior information for the subsequent association determination module.

[0031] The model selector automatically calls a dedicated model (ResNet extracts image features, VGG extracts video features) based on data labels (such as images and speech) and outputs a high-dimensional feature vector; the feature integrity is evaluated through the autoencoder, and low-quality features (such as edge features of blurred images) are enhanced using GAN.

[0032] The model selector of the present invention automatically matches the optimal extraction model, avoiding manual parameter adjustment, and improving feature extraction efficiency by 30%; low-quality feature enhancement improves the quality of subsequent fusion input; shallow correlation pre-analysis provides priors for deep correlation, reducing computational complexity, and improving correlation analysis speed by 20%.

[0033] The AI association determination module uses a deep association learning model to construct a multimodal feature association graph, quantify the semantic association between features, and adjust the association weight based on real-time fusion effect feedback through an online learning algorithm. The AI association determination module also includes a noise feature decoupling module, which combines the generative adversarial network to separate the noise components in the features. The generator fits the noise distribution and generates fake sample features, and the discriminator distinguishes between real features and fake samples to achieve noise purification and feature enhancement.

[0034] The deep association learning model constructs a multimodal feature association graph to quantify semantic associations. Through online learning (gradient descent), it dynamically adjusts association weights based on fusion feedback (e.g., the association weights between enhanced images and vibration data in fault scenarios). The GAN generator fits the noise distribution to generate fake sample features, while the discriminator distinguishes between real and fake samples, purging noise from features and improving input quality.

[0035] The present invention quantifies semantic relevance through deep association learning, improves feature synergy, and increases fusion accuracy by 15%. It removes data noise through noise decoupling (GAN), improves the signal-to-noise ratio of the purified features by 40%, and reduces fusion error.

[0036] The AI fusion decision module includes a feature-level AI fusion unit, a decision-level AI fusion unit and an optimization feedback unit. The feature-level AI fusion unit integrates different modal features through an adaptive fusion layer for related features by designing a multimodal fusion network, and supports dynamic adjustment of the fusion strategy based on the degree of correlation; at the same time, adversarial feature alignment is introduced to align different modal feature spaces through adversarial training to reduce the fusion error caused by modal differences; the decision-level AI fusion unit performs secondary fusion through the decision voting neural network by combining the modal decisions output by each modal network and the fusion decision output by the feature-level fusion network, and generates the final fusion result using a weighted voting mechanism.

[0037] The multimodal fusion network integrates related features through an adaptive fusion layer, dynamically adjusting strategies based on correlation (deep fusion of highly correlated features, while retaining independent modal features with low correlation). Adversarial feature alignment (such as domain adaptation algorithms) is introduced to narrow the gap between the image and point cloud feature spaces, reducing fusion errors. The outputs of each modality network (e.g., CNN outputs image classification decisions, RNN outputs speech recognition decisions) and the feature-level fusion network are input into a decision-voting neural network, where a weighted vote is performed to generate the final result (weights are dynamically determined by feature correlation and historical decision accuracy, such that modalities with higher historical accuracy are given higher weights).

[0038] The present invention dynamically adjusts strategies through feature-level fusion (such as deep fusion of highly correlated features) to adapt to complex scenarios (such as multi-source fault data fusion), improving decision-making accuracy by 20%; through decision-level weighted voting combined with modal historical accuracy (such as giving visual recognition models a higher weight in image-dominated scenarios), it reduces the risk of single-modal misjudgment and improves robustness by 30%; Q-Learning closed-loop optimization continuously iterates model parameters, enhancing the system's adaptability and maintaining a stable long-term operating accuracy of over 95%.

[0039] The optimization feedback unit reversely optimizes the parameters of each modal network and the fusion network using a Q-Learning reinforcement learning algorithm, forming a closed-loop self-optimization system. Based on the accuracy of the fusion results, the Q-Learning algorithm reversely optimizes the parameters of each network (such as adjusting the fusion layer weights), forming a closed-loop self-optimization (such as continuously improving the accuracy of equipment fault identification).

[0040] The AI monitoring module collects real-time metrics such as computing resource usage, processing latency, and data throughput from each module through sensors or log analysis, and uses anomaly detection models to warn of potential failures. The AI monitoring module collects metrics such as CPU / GPU utilization and processing latency in real time through sensors or logs, and uses anomaly detection models (such as the Isolation Forest model) to warn of hardware overloads or algorithm failures (such as a sudden drop in feature extraction model accuracy).

[0041] The AI Optimization Module uses Bayesian optimization to automatically adjust AI model hyperparameters or reallocate computing resources based on monitoring data. This ensures efficient system operation by automatically adjusting hyperparameters (such as learning rate and number of network layers) through Bayesian optimization or reallocating computing resources through Neural Architecture Search (NAS) (for example, allocating more GPU computing power to the fusion module during peak hours).

[0042] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. An artificial intelligence-based heterogeneous data feature fusion processing system, characterized by: It includes AI big data acquisition module, AI preprocessing module, AI big data storage management module, AI feature extraction module, AI association determination module, AI fusion decision module, AI monitoring module and AI optimization module; the AI big data acquisition module includes acquisition unit and storage unit. The acquisition unit integrates AI scheduling model to dynamically adjust the acquisition strategy according to historical data requirements and real-time business scenarios. The storage unit performs preliminary intelligent classification of the collected raw data, generates data labels, and compares and stores historical data with real-time data through data feature similarity analysis. It automatically establishes data association indexes to support fast retrieval and historical data backtracking; the AI preprocessing module includes intelligent cleaning submodule and intelligent normalization submodule. The intelligent cleaning submodule identifies and filters noise data, outliers or duplicate data based on the anomaly detection algorithm of isolation forest. The intelligent cleaning submodule adopts AI-driven missing value repair and dynamically selects repair strategies according to data modal characteristics. The intelligent normalization submodule automatically identifies data distribution characteristics and generates optimal normalization parameters through adaptive normalization model.

2. The artificial intelligence-based heterogeneous data feature fusion processing system according to claim 1 is characterized in that: The AI big data storage management module classifies and stores the data processed by the AI preprocessing module. The AI big data storage management module adopts a multimodal vector database and generates data semantic vectors through self-supervised learning to realize the joint indexing and rapid retrieval of unstructured data and structured data. The AI big data storage management module builds a spatiotemporal feature comparison library and uses a contrastive learning algorithm to generate a data feature similarity matrix to support intelligent matching and incremental updating of historical data.

3. The artificial intelligence-based heterogeneous data feature fusion processing system according to claim 1 is characterized in that: The AI feature extraction module extracts features from the data in the AI big data storage management module, integrates dedicated deep learning models for different modal data, and automatically matches the optimal extraction model through a model selector; the AI feature extraction module uses the AI model to evaluate feature integrity, enhance low-quality features, and then preliminarily extracts shallow correlation features between modalities through a lightweight association network to provide prior information for the subsequent association determination module.

4. The artificial intelligence-based heterogeneous data feature fusion processing system according to claim 1 is characterized in that: The AI association determination module uses a deep association learning model to construct a multimodal feature association graph, quantify the semantic association between features, and adjust the association weight based on real-time fusion effect feedback through an online learning algorithm. The AI association determination module also includes a noise feature decoupling module, which combines the generative adversarial network to separate the noise components in the features. The generator fits the noise distribution and generates fake sample features, and the discriminator distinguishes between real features and fake samples to achieve noise purification and feature enhancement.

5. The artificial intelligence-based heterogeneous data feature fusion processing system according to claim 1 is characterized in that: The AI fusion decision module includes a feature-level AI fusion unit, a decision-level AI fusion unit and an optimization feedback unit. The feature-level AI fusion unit integrates different modal features through an adaptive fusion layer for related features by designing a multimodal fusion network, and supports dynamic adjustment of the fusion strategy based on the degree of correlation; at the same time, adversarial feature alignment is introduced to align different modal feature spaces through adversarial training to reduce the fusion error caused by modal differences; the decision-level AI fusion unit performs secondary fusion through the decision voting neural network by combining the modal decisions output by each modal network and the fusion decision output by the feature-level fusion network, and generates the final fusion result using a weighted voting mechanism.

6. The artificial intelligence-based heterogeneous data feature fusion processing system according to claim 5 is characterized in that: The optimization feedback unit reversely optimizes the parameters of each modal network and the fusion network through the Q-Learning reinforcement learning algorithm to form a closed-loop self-optimization system.

7. The artificial intelligence-based heterogeneous data feature fusion processing system according to claim 1 is characterized in that: The AI monitoring module collects the computing resource usage, processing delay, and data throughput indicators of each module in real time through sensors or log analysis, and warns of potential failures based on anomaly detection models.

8. The artificial intelligence-based heterogeneous data feature fusion processing system according to claim 1 is characterized in that: The AI optimization module uses Bayesian optimization to automatically adjust AI model hyperparameters or reallocate computing resources based on monitoring data.

Images