Multi-source heterogeneous data acquisition system

Abstract

The invention discloses a multi-source heterogeneous data acquisition system. Comprising a multi-source access subsystem for generating an initial multi-source heterogeneous data set; the intelligent preprocessing subsystem is used for generating a preprocessing data set; the edge-cloud collaborative computing subsystem is used for generating a global processing data set; the security encryption subsystem is used for verifying the access authority of the user and encrypting and storing a global processing data set by adopting a differential privacy method; and the data adaptation output subsystem is used for acquiring the target feature of the user, and acquiring the data in the global processing data set corresponding to the target feature from the security encryption subsystem by using a homomorphic encryption protocol based on the user access authority and the target feature. The method supports user customized data output, adapts to different application scene requirements, realizes efficient, safe and flexible data acquisition and management, and provides high-quality data support for subsequent data fusion, analysis and application.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a multi-source heterogeneous data acquisition system. Background Technology

[0002] With the rapid development of information technology, data sources have become increasingly diversified, forming the typical characteristics of multi-source heterogeneous data. Multi-source heterogeneous data typically comes from IoT devices, enterprise databases, file systems (such as local documents and cloud storage files), and third-party API interfaces, and the data formats cover structured data (such as database table data), unstructured data (such as text, images, and voice), and time-series data (such as temperature and humidity data collected by sensors in real time).

[0003] In practical applications, the acquisition of multi-source heterogeneous data faces three core challenges: Poor access compatibility. Different data sources use significantly different communication protocols, and traditional data acquisition systems only support one or a few protocols, making it difficult to achieve unified access for multiple devices and data formats. At the same time, IoT devices come in a variety of models and lack a unified identity authentication and management mechanism, which can easily lead to chaotic device access or unauthorized devices illegally accessing the network.

[0004] There is an imbalance between processing efficiency and real-time performance. The volume of multi-source heterogeneous data is enormous, and some data (such as industrial equipment fault monitoring data) has extremely high real-time requirements. Traditional centralized cloud processing models suffer from data transmission delays and cannot meet real-time requirements; while simple edge processing models have limited computing power and are unable to complete complex multimodal data fusion analysis, resulting in a trade-off between data processing efficiency and real-time performance.

[0005] Insufficient security. Data is easily stolen or tampered with during transmission, and there is a risk of individual data privacy leakage during storage; traditional encryption methods (such as single symmetric encryption) have low security and lack fine-grained access control, failing to guarantee data security throughout its entire lifecycle.

[0006] In existing technologies, some multi-source data acquisition solutions only focus on data access and simple cleaning, failing to address the balance between real-time performance and processing efficiency. While some solutions incorporate edge computing or encryption technologies, they lack a fully integrated design encompassing the entire process from "access-processing-computation-encryption-output," making it difficult to meet the demands for efficiency, security, and flexibility in complex scenarios. Therefore, there is an urgent need for a multi-source heterogeneous data acquisition system capable of unified access, efficient processing, secure transmission, and customized output of multi-source data. Summary of the Invention

[0007] In view of this, the present invention provides a multi-source heterogeneous data acquisition system that can solve the above-mentioned technical problems.

[0008] To solve the above-mentioned technical problems, the present invention is implemented as follows.

[0009] A multi-source heterogeneous data acquisition system, comprising: The multi-source access subsystem connects to different types of data sources, uses the communication protocol corresponding to the data source to collect multi-source heterogeneous data, and generates an initial multi-source heterogeneous dataset. The intelligent preprocessing subsystem preprocesses the data in the initial multi-source heterogeneous dataset to generate a preprocessed dataset. The edge-cloud collaborative computing subsystem includes edge computing nodes and cloud computing nodes. The edge computing nodes perform local feature extraction and analysis on data in the preprocessed dataset whose real-time processing requirements exceed a first threshold, generating edge processing results. The cloud computing nodes use a multimodal neural network model to perform deep feature fusion and global analysis on data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold and the edge processing results, generating a global processing dataset. The security encryption subsystem is used to verify user access permissions and to encrypt and store the global processing dataset using a differential privacy method. The data adaptation and output subsystem obtains the user's target characteristics, and based on the user's access permissions and target characteristics, uses a homomorphic encryption protocol to obtain data from the global processing dataset corresponding to the target characteristics from the secure encryption subsystem.

[0010] Preferably, the multi-source access subsystem includes a protocol adaptation module, a device management module, and a data receiving module; the protocol adaptation module interfaces with different IoT devices and stores MQTT, HTTP, OPC UA, and TCP / IP protocols, and matches the protocol corresponding to the data of each IoT device through an automatic protocol identification algorithm; the device management module performs identity registration management, status monitoring, and firmware updates for each IoT device, and also stores the identifier and public key of each IoT device; the data receiving module collects multi-source heterogeneous data through the protocol corresponding to the data of each IoT device and generates an initial multi-source heterogeneous dataset.

[0011] Preferably, the intelligent preprocessing subsystem includes an anomaly detection module, a missing value handling module, and a format standardization module. The anomaly detection module identifies anomalous data in the initial multi-source heterogeneous dataset through visual analysis combined with statistical methods, identifies duplicate data through hash comparison, and deletes both anomalous and duplicate data. The missing value handling module determines the imputation method for missing data based on the data type: mean or median for numerical data, mode for categorical data, and linear interpolation for time-series data. The format standardization module converts unstructured data into a unified format, converts text data into TXT format, converts image data into JPEG format and standardizes the resolution, converts audio data into WAV format and standardizes the sampling rate, and converts structured data and time-series data into JSON format.

[0012] Preferably, the edge-cloud collaborative computing subsystem includes a task allocation module, edge computing nodes, cloud computing nodes, and a model optimization module; the task allocation module determines the real-time requirements of the data in the preprocessed dataset, allocates data in the preprocessed dataset whose real-time processing requirements exceed a first threshold to the edge computing nodes, and allocates data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold to the cloud computing nodes; Edge computing nodes are equipped with edge feature extraction modules. For data in the preprocessed dataset whose real-time processing requirements exceed a first threshold, local feature extraction and analysis are performed using a lightweight convolutional neural network to generate edge processing results. These edge computing results are then sent to cloud computing nodes. For time-series data, the lightweight convolutional neural network performs trend analysis on the time-series data using a sliding window algorithm, and the trend analysis results are used as edge processing results. The cloud computing nodes are equipped with a cloud-based multimodal fusion module and a model optimization module. These modules are used to acquire edge processing results sent by edge computing nodes and to perform deep feature fusion and global analysis on data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold, as well as edge processing results, using a multimodal neural network model to generate a global processing dataset. The multimodal neural network model includes LSTM, CNN, and Transformer, and incorporates an attention mechanism. The model optimization module is used to evaluate the performance of each model in the multimodal neural network model and the overall performance of the multimodal neural network model. It uses a stochastic gradient descent optimizer to adjust the parameters of each model in the multimodal neural network model.

[0013] Preferably, the security encryption subsystem includes a key management module, an access verification module, a storage encryption module, and a transmission encryption module; The key management module is used to create several public-private key pairs and symmetric encryption keys. The first public key is used to verify the access permissions of IoT devices when they are connected, and the second public key is used to encrypt the global processing dataset. Both the first and second private keys are kept by the system administrator and used for decryption. The symmetric encryption key is used to encrypt the transmission of data whose access frequency exceeds a second preset threshold. The access verification module is used to verify the access rights of users and / or IoT devices. The verification of IoT device access rights includes: when an IoT device accesses the multi-source access subsystem, finding the public key of the IoT device through the IoT device's identifier and verifying the private key signature of the IoT device; when a user accesses the system, verifying the user's key and determining the user's access rights. The storage encryption module adds noise to the global processing dataset using differential privacy technology, and then stores the noise-added global processing dataset into an XML database; access permissions are set for the XML database. The transmission encryption module generates a session key using a homomorphic encryption protocol before sending the data in the global processing dataset corresponding to the target feature. The data in the global processing dataset corresponding to the target feature is then used as the transmission data, and the transmission data is encrypted.

[0014] Preferably, the data adaptation and output subsystem includes a requirement analysis module, a data filtering module, a format conversion module, and a data push module; The requirement parsing module receives the requirement information input by the user, determines the user's target characteristics based on the requirement information, and determines the information to be fed back to the user, including the user's email address, user IP address, feature direction, and target characteristics. The data filtering module, based on user access permissions and target characteristics, uses a homomorphic encryption protocol to obtain data from the global processing dataset corresponding to the target characteristics in the secure encryption subsystem, and uses this data as the target dataset. The format conversion module converts the target dataset into a suitable format based on the user's output requirements. Specifically, it converts the dataset to CSV or Excel format when outputting to enterprise applications, to TFRecord format when outputting to machine learning models, and to PDF report format when outputting to end users. The data push module pushes the converted target dataset in the adapted format to the user.

[0015] Beneficial effects: (1) This invention is applicable to scenarios such as Internet of Things, industrial production, and smart cities. It can solve the problems of poor access compatibility, low processing efficiency and insufficient security in the existing multi-source heterogeneous data acquisition process, and can efficiently acquire, process, securely transmit and customize the output of data from different sources and in different formats.

[0016] (2) This invention improves access compatibility, supports multiple communication protocols and device types, realizes unified access of multi-source data, and avoids unauthorized access through device identity management.

[0017] (3) The present invention balances processing efficiency and real-time performance, and adopts an edge-cloud collaborative computing mode, which not only meets the needs of rapid processing of real-time data, but also realizes deep fusion analysis of non-real-time data.

[0018] (4) This invention ensures data security. End-to-end encryption and fine-grained access control prevent data theft, tampering, and privacy leaks.

[0019] (5) This invention enhances flexibility. It supports customized data output by users, adapts to the needs of different application scenarios, and achieves efficient, secure, and flexible data collection and management, providing high-quality data support for subsequent data fusion, analysis, and application. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the multi-source heterogeneous data acquisition system of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] like Figure 1 As shown, this invention proposes a multi-source heterogeneous data acquisition system, comprising: The multi-source access subsystem connects to different types of data sources, uses the communication protocol corresponding to the data source to collect multi-source heterogeneous data, and generates an initial multi-source heterogeneous dataset. The intelligent preprocessing subsystem preprocesses the data in the initial multi-source heterogeneous dataset to generate a preprocessed dataset. The edge-cloud collaborative computing subsystem includes edge computing nodes and cloud computing nodes. The edge computing nodes perform local feature extraction and analysis on data in the preprocessed dataset whose real-time processing requirements exceed a first threshold, generating edge processing results. The cloud computing nodes use a multimodal neural network model to perform deep feature fusion and global analysis on data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold and the edge processing results, generating a global processing dataset. The security encryption subsystem is used to verify user access permissions and to encrypt and store the global processing dataset using a differential privacy method. The data adaptation and output subsystem obtains the user's target characteristics, and based on the user's access permissions and target characteristics, uses a homomorphic encryption protocol to obtain data from the global processing dataset corresponding to the target characteristics from the secure encryption subsystem.

[0023] Furthermore, the multi-source access subsystem includes a protocol adaptation module, a device management module, and a data receiving module. The protocol adaptation module interfaces with different IoT devices and stores MQTT, HTTP, OPC UA, and TCP / IP protocols. It matches the protocol corresponding to the data of each IoT device through an automatic protocol identification algorithm. The device management module performs identity registration management, status monitoring, and firmware updates for each IoT device. It also stores the identifier and public key of each IoT device. The data receiving module collects multi-source heterogeneous data through the protocol corresponding to the data of each IoT device and generates an initial multi-source heterogeneous dataset.

[0024] Furthermore, the intelligent preprocessing subsystem includes an anomaly detection module, a missing value handling module, and a format standardization module. The anomaly detection module identifies anomalous data in the initial multi-source heterogeneous dataset through visual analysis combined with statistical methods, identifies duplicate data through hash comparison, and deletes both anomalous and duplicate data. The missing value handling module determines the imputation method for missing data based on the data type: mean or median for numerical data, mode for categorical data, and linear interpolation for time-series data. The format standardization module converts unstructured data into a unified format, converts text data into TXT format, converts image data into JPEG format and standardizes the resolution, converts audio data into WAV format and standardizes the sampling rate, and converts structured data and time-series data into JSON format.

[0025] Furthermore, the edge-cloud collaborative computing subsystem includes a task allocation module, edge computing nodes, cloud computing nodes, and a model optimization module; the task allocation module determines the real-time requirements of the data in the preprocessed dataset, allocates data in the preprocessed dataset whose real-time processing requirements exceed a first threshold to the edge computing nodes, and allocates data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold to the cloud computing nodes; Edge computing nodes are equipped with edge feature extraction modules. For data in the preprocessed dataset whose real-time processing requirements exceed a first threshold, local feature extraction and analysis are performed using a lightweight convolutional neural network to generate edge processing results. These edge computing results are then sent to cloud computing nodes. For time-series data, the lightweight convolutional neural network performs trend analysis on the time-series data using a sliding window algorithm, and the trend analysis results are used as edge processing results. The cloud computing nodes are equipped with a cloud-based multimodal fusion module and a model optimization module. These modules are used to acquire edge processing results sent by edge computing nodes and to perform deep feature fusion and global analysis on data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold, as well as edge processing results, using a multimodal neural network model to generate a global processing dataset. The multimodal neural network model includes LSTM, CNN, and Transformer, and incorporates an attention mechanism. The model optimization module is used to evaluate the performance of each model in the multimodal neural network model and the overall performance of the multimodal neural network model. It uses a stochastic gradient descent optimizer to adjust the parameters of each model in the multimodal neural network model.

[0026] Furthermore, the security encryption subsystem includes a key management module, an access verification module, a storage encryption module, and a transmission encryption module; The key management module is used to create several public-private key pairs and symmetric encryption keys. The first public key is used to verify the access permissions of IoT devices when they are connected, and the second public key is used to encrypt the global processing dataset. Both the first and second private keys are kept by the system administrator and used for decryption. The symmetric encryption key is used to encrypt the transmission of data whose access frequency exceeds a second preset threshold. The access verification module is used to verify the access rights of users and / or IoT devices. Verifying the access rights of IoT devices includes: when an IoT device accesses the multi-source access subsystem, finding the public key of the IoT device through the IoT device's identifier and verifying the signature of the IoT device's private key; when a user accesses the system, verifying the user's key and determining the user's access rights.

[0027] The storage encryption module adds noise to the global processing dataset using differential privacy technology, and then stores the noise-added global processing dataset into an XML database; access permissions are set for the XML database. The transmission encryption module generates a session key using a homomorphic encryption protocol before sending the data in the global processing dataset corresponding to the target feature. The data in the global processing dataset corresponding to the target feature is then used as the transmission data, and the transmission data is encrypted.

[0028] Furthermore, the data adaptation and output subsystem includes a requirement analysis module, a data filtering module, a format conversion module, and a data push module; The requirement parsing module receives the requirement information input by the user, determines the user's target characteristics based on the requirement information, and determines the information to be fed back to the user, including the user's email address, user IP address, feature direction, and target characteristics. The data filtering module, based on user access permissions and target characteristics, uses a homomorphic encryption protocol to obtain data from the global processing dataset corresponding to the target characteristics in the secure encryption subsystem, and uses this data as the target dataset. The format conversion module converts the target dataset into a suitable format based on the user's output requirements. Specifically, it converts the dataset to CSV or Excel format when outputting to enterprise applications, to TFRecord format when outputting to machine learning models, and to PDF report format when outputting to end users. The data push module pushes the converted target dataset in the adapted format to the user.

[0029] In this invention, the training process of the multimodal neural network model includes: Data preparation steps: Collect text sample sets (such as news corpora), image sample sets (such as ImageNet subsets), and speech sample sets (such as the TIMIT speech library). After labeling the samples, divide them into training set (70%), validation set (15%), and test set (15%). Model building steps: Build an LSTM network to process temporal features, a CNN to process image features, and a Transformer model to process text and speech features. Connect the feature vectors output by the three networks through a fully connected layer to form a basic multimodal model. Training optimization steps: Train the basic multimodal model using the training set, calculate the difference between the model's predicted values ​​and the sample labels using the cross-entropy loss function, adjust the model parameters using the stochastic gradient descent optimizer (learning rate 0.001, momentum 0.9), monitor the model for overfitting using the validation set, and adjust the model regularization parameters (such as Dropout probability). Model evaluation steps: Use the test set to evaluate the trained model, calculate the feature extraction accuracy of each modality and the overall accuracy after multimodal fusion. When the overall accuracy reaches 90% or more, the model training is considered complete and the model can be put into use.

[0030] In this invention, the homomorphic encryption protocol implementation process of the transmission encryption module includes: Key negotiation steps: The data sender and receiver negotiate and generate a session key using the Diffie-Hellman key exchange algorithm, ensuring that the session key is known only to the two parties; Data encryption steps: The Paillier homomorphic encryption algorithm is used to encrypt the transmitted data using the session key, generating ciphertext data; Data transmission steps: Transmit encrypted data via TCP communication protocol, and simultaneously send a data verification value (generated using SHA-256 hash algorithm); Data decryption steps: After receiving the ciphertext data, the receiver decrypts the ciphertext using the session key to generate plaintext data. The receiver then calculates the plaintext data checksum using a hash algorithm and compares it with the checksum provided by the sender to verify the data integrity.

[0031] The present invention also provides a specific embodiment of a multi-source heterogeneous data acquisition system.

[0032] The multi-source access subsystem is used to connect to different types of data sources such as IoT devices, databases, file systems, and API interfaces. It adapts to different communication protocols through an adaptive protocol conversion module, collects structured data, unstructured data, and time-series data, and generates an initial multi-source heterogeneous dataset. The intelligent preprocessing subsystem is connected to the multi-source access subsystem and is used to perform outlier deletion, duplicate value removal, missing value filling and data format standardization on the initial multi-source heterogeneous dataset to generate a preprocessed dataset. The edge-cloud collaborative computing subsystem is connected to the intelligent preprocessing subsystem and includes edge computing nodes and cloud computing nodes. The edge computing nodes are used to perform local feature extraction and preliminary analysis on data with high real-time requirements in the preprocessing dataset to generate edge processing results. The cloud computing nodes are used to receive the edge processing results and non-real-time data in the preprocessing dataset, and perform deep feature fusion and global analysis through a multimodal neural network model to generate a global processing dataset. The security encryption subsystem is connected to the multi-source access subsystem, the edge-cloud collaborative computing subsystem, and the data adaptation and output subsystem, respectively. It is used to ensure the security of data transmission through homomorphic encryption protocol, encrypt stored data through differential privacy technology, and control user access permissions through key management module. The data adaptation and output subsystem is connected to the edge-cloud collaborative computing subsystem. It is used to obtain target features according to user needs, filter target data subsets from the global processing dataset based on the target features, adapt different output formats through the data format conversion module, and output customized datasets to users.

[0033] In this invention, the multi-source access subsystem achieves unified access to different data sources such as IoT devices, databases, file systems, and API interfaces through adaptive protocol conversion and device management, and collects multiple types of data and performs preliminary classification and temporary storage. Intelligent preprocessing subsystem: Cleans and standardizes the initial data through anomaly detection, missing value handling, and format standardization to generate a high-quality preprocessed dataset; Edge-cloud collaborative computing subsystem: Based on the real-time data requirements, processing tasks are allocated. Edge nodes process real-time data, while cloud nodes perform multimodal data fusion analysis, balancing processing efficiency and real-time performance. Security encryption subsystem: Ensures data security throughout its entire lifecycle through transmission encryption, storage encryption, and key management; controls data access permissions through access verification. Data adaptation and output subsystem: Filters data and converts formats according to user needs, providing customized data output services.

[0034] Another embodiment of the present invention provides a multi-source heterogeneous data acquisition system, including a multi-source access subsystem, an intelligent preprocessing subsystem, an edge-cloud collaborative computing subsystem, a security encryption subsystem, and a data adaptation output subsystem. Each subsystem interacts with data via Ethernet. The specific implementation process is as follows: Implementation of multi-source access subsystem Protocol Adaptation: After system startup, the protocol adaptation module loads the MQTT, HTTP, OPC UA, and TCP / IP protocol libraries. When an IoT sensor (such as a temperature sensor) sends an access request, the module analyzes the request data packet format using an automatic protocol identification algorithm, matches the MQTT protocol, and establishes an MQTT connection between the sensor and the system. When the enterprise MySQL database sends data, the module matches the JDBC protocol and establishes a connection after verifying the database account and password.

[0035] Device Management: Newly connected temperature sensors are registered in the device management module, and their unique MAC address is entered. The system generates a public key and a private key for the device. The public key is stored in the public key library of the device management module, and the private key is sent to the sensor locally. The device management module monitors the online status of the sensors by periodically sending heartbeat packets. When the sensor firmware needs to be updated, a firmware update package is pushed via the MQTT protocol.

[0036] Data reception: The data receiving module receives time-series data from the temperature sensor via the MQTT protocol and structured data (such as production order data, including order number, product model, and quantity) from the MySQL database via the JDBC protocol. The time-series data is temporarily stored in the time-series buffer of the edge node, and the structured data is temporarily stored in the structured data buffer in the cloud.

[0037] Implementation of intelligent preprocessing subsystem Anomaly detection: The anomaly detection module reads the temperature data from the time-series buffer, calculates the mean (25℃) and standard deviation (2℃) of the temperature data using the 3σ principle, and identifies and deletes data that exceeds the range of [19℃, 31℃] as outliers; it also calculates the hash value of each structured data item using a hash algorithm and deletes order data with duplicate hash values.

[0038] Missing value handling: The missing value handling module found that some temperature data was missing and used linear interpolation to fill the missing values; it also found that the "product model" field in the order data was missing and used the mode to fill the missing values.

[0039] Format standardization: The format standardization module converts temperature time-series data into JSON format; converts order structured data into JSON format; and converts unstructured image data uploaded by cameras (originally in BMP format) into JPEG format, with a unified resolution of 1920×1080.

[0040] Implementation of the Edge-Cloud Collaborative Computing Subsystem Task allocation: The task allocation module determines that temperature data is real-time data and allocates it to edge computing nodes; it determines that order data and image data are non-real-time data and allocates them to cloud computing nodes.

[0041] Edge feature extraction: The edge feature extraction module is deployed on edge computing nodes. It uses MobileNet lightweight CNN to extract features from temperature data and identify temperature change trends. It also uses a sliding window algorithm to calculate the average temperature and generate edge processing results.

[0042] Cloud-based multimodal fusion: The cloud-based multimodal fusion module constructs multimodal neural network models (e.g., LSTM for temperature edge results, CNN for image data, and Transformer for order data), inputting edge processing results, image data, and order data into the model; and weightedly fusing multimodal features through an attention mechanism to generate a global processing dataset.

[0043] Model optimization: The model optimization module uses a test set to evaluate model performance. The initial model image quality score accuracy is 85%. By adjusting the convolution kernel parameters of the CNN network through the stochastic gradient descent optimizer, the accuracy is improved to 92%, which meets the requirements of practical applications.

[0044] The specific embodiments described above only illustrate the design principles of the present invention. The shapes and names of the components in this description may differ and are not limited. Therefore, those skilled in the art can modify or make equivalent substitutions to the technical solutions described in the foregoing embodiments; and these modifications and substitutions do not depart from the inventive spirit and technical solutions of the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A multi-source heterogeneous data acquisition system, characterized in that, include: The multi-source access subsystem connects to different types of data sources, uses the communication protocol corresponding to the data source to collect multi-source heterogeneous data, and generates an initial multi-source heterogeneous dataset. The intelligent preprocessing subsystem preprocesses the data in the initial multi-source heterogeneous dataset to generate a preprocessed dataset. The edge-cloud collaborative computing subsystem includes edge computing nodes and cloud computing nodes; Edge computing nodes perform local feature extraction and analysis on data in the preprocessed dataset whose real-time processing requirements exceed the first threshold, and generate edge processing results. The cloud computing nodes use a multimodal neural network model to perform deep feature fusion and global analysis on data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold and edge processing results, generating a global processing dataset. The security encryption subsystem is used to verify user access permissions and to encrypt and store the global processing dataset using a differential privacy method. The data adaptation and output subsystem obtains the user's target characteristics, and based on the user's access permissions and target characteristics, uses a homomorphic encryption protocol to obtain data from the global processing dataset corresponding to the target characteristics from the secure encryption subsystem.

2. The system as described in claim 1, characterized in that, The multi-source access subsystem includes a protocol adaptation module, a device management module, and a data receiving module. The protocol adaptation module interfaces with different IoT devices and stores MQTT, HTTP, OPCUA, and TCP / IP protocols. It matches the protocol corresponding to the data of each IoT device through an automatic protocol identification algorithm. The device management module performs identity registration management, status monitoring, and firmware updates for each IoT device. It also stores the identifier and public key of each IoT device. The data receiving module collects multi-source heterogeneous data through the protocols corresponding to the data of each IoT device, and generates an initial multi-source heterogeneous dataset.

3. The system as described in claim 1, characterized in that, The intelligent preprocessing subsystem includes an anomaly detection module, a missing value handling module, and a format standardization module. The anomaly detection module identifies anomalous data in the initial multi-source heterogeneous dataset through visualization analysis combined with statistical methods, identifies duplicate data through hash comparison, and deletes both anomalous and duplicate data. The missing value handling module determines the imputation method for missing data based on the data type: mean or median for numerical data, mode for categorical data, and linear interpolation for time-series data. The format standardization module transforms unstructured data into a unified format, text data into TXT format, image data into JPEG format and standardizes the resolution, audio data into WAV format and standardizes the sampling rate, and structured data and time series data into JSON format.

4. The system as described in any one of claims 1-3, characterized in that, The edge-cloud collaborative computing subsystem includes a task allocation module, edge computing nodes, cloud computing nodes, and a model optimization module. The task allocation module determines the real-time requirements of the data in the preprocessed dataset and allocates data in the preprocessed dataset whose real-time processing requirements exceed a first threshold to the edge computing nodes. Data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold will be allocated to cloud computing nodes. Edge computing nodes are equipped with edge feature extraction modules. For data in the preprocessed dataset whose real-time processing requirements exceed a first threshold, local feature extraction and analysis are performed using a lightweight convolutional neural network to generate edge processing results. These edge computing results are then sent to cloud computing nodes. For time-series data, the lightweight convolutional neural network performs trend analysis on the time-series data using a sliding window algorithm, and the trend analysis results are used as edge processing results. The cloud computing nodes are equipped with a cloud-based multimodal fusion module and a model optimization module. These modules are used to acquire edge processing results sent by edge computing nodes and to perform deep feature fusion and global analysis on data in the preprocessed dataset whose real-time processing requirements do not exceed the first threshold, as well as edge processing results, using a multimodal neural network model to generate a global processing dataset. The multimodal neural network model includes LSTM, CNN, and Transformer, and incorporates an attention mechanism. The model optimization module is used to evaluate the performance of each model in the multimodal neural network model and the overall performance of the multimodal neural network model. It uses a stochastic gradient descent optimizer to adjust the parameters of each model in the multimodal neural network model.

5. The system as described in claim 4, characterized in that, The security encryption subsystem includes a key management module, an access verification module, a storage encryption module, and a transmission encryption module; The key management module is used to create several public-private key pairs and symmetric encryption keys. The first public key is used to verify the access permissions of IoT devices when they are connected, and the second public key is used to encrypt the global processing dataset. Both the first and second private keys are kept by the system administrator and used for decryption. The symmetric encryption key is used to encrypt the transmission of data whose access frequency exceeds a second preset threshold. The access verification module is used to verify the access rights of users and / or IoT devices. The verification of IoT device access rights includes: when an IoT device accesses the multi-source access subsystem, finding the public key of the IoT device through the IoT device's identifier and verifying the private key signature of the IoT device; when a user accesses the system, verifying the user's key and determining the user's access rights. The storage encryption module adds noise to the global processing dataset using differential privacy technology, and then stores the noise-added global processing dataset into an XML database; access permissions are set for the XML database. The transmission encryption module generates a session key using a homomorphic encryption protocol before sending the data in the global processing dataset corresponding to the target feature. The data in the global processing dataset corresponding to the target feature is then used as the transmission data, and the transmission data is encrypted.

6. The system as described in claim 5, characterized in that, The data adaptation and output subsystem includes a requirements analysis module, a data filtering module, a format conversion module, and a data push module; The requirement parsing module receives the requirement information input by the user and determines the user's target characteristics based on the requirement information; The information to be fed back to the user includes the user's email address, IP address, feature direction, and target features. The data filtering module, based on user access permissions and target characteristics, uses a homomorphic encryption protocol to obtain data from the global processing dataset corresponding to the target characteristics in the secure encryption subsystem, and uses this data as the target dataset. The format conversion module converts the target dataset into a suitable format based on the user's output requirements. Specifically, it converts the dataset to CSV or Excel format when outputting to enterprise applications, to TFRecord format when outputting to machine learning models, and to PDF report format when outputting to end users. The data push module pushes the converted target dataset in the adapted format to the user.

Images