Multi-source heterogeneous data fusion processing system and method based on artificial intelligence deep learning

Abstract

The invention relates to the technical field of data fusion, in particular to a multi-source heterogeneous data fusion processing system and method based on artificial intelligence deep learning, and the system comprises a data collection module, a data preprocessing module, a feature extraction module, a data fusion module, a result evaluation module and an application output module. The data collection module collects original data of different data sources and transmits the original data to the data preprocessing module, the data preprocessing module carries out space-time self-supervision training and data enhancement on the original data and then outputs preprocessed data, and the feature extraction module carries out feature extraction on the preprocessed data and outputs feature data. The data fusion module carries out fusion processing on the feature data to obtain fusion data, the result evaluation module evaluates the fusion data, the application output module outputs the evaluated fusion data, a Lasso regression model is adopted to carry out preliminary feature screening, and the accuracy and efficiency of feature extraction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data fusion, and specifically to a multi-source heterogeneous data fusion processing system and method based on artificial intelligence deep learning. Background Art

[0002] Multi-source heterogeneous data fusion processing is an important research direction in the field of artificial intelligence. It aims to integrate data from different data sources with different structures and characteristics to improve the accuracy and efficiency of data analysis. Traditional data processing methods often have the following defects when facing multi-source heterogeneous data: 1. When processing multi-source data, existing systems often have difficulty ensuring the temporal consistency of data. Deviations exist between timestamps of different data sources, resulting in errors when merging data. 2. In the feature extraction stage, most existing systems rely on traditional feature selection methods, which are ineffective when dealing with complex data, resulting in inadequate feature extraction; 3. When fusing data, existing systems often use fixed fusion algorithms and weight distribution strategies, lack dynamic adjustment mechanisms, and are difficult to adapt to the needs of different data and tasks. Summary of the invention

[0003] The purpose of the present invention is to overcome the above technical defects and provide a multi-source heterogeneous data fusion processing system and method based on artificial intelligence deep learning. First, through the data collection module, the data access, synchronization and storage units are used to obtain raw data from different data sources. By introducing the timestamp alignment algorithm, the temporal consistency of different data sources is ensured to improve the accuracy of data fusion. Subsequently, the data preprocessing module uses a mask generator, a spatiotemporal context encoder, a contrastive learning unit and a neural differential equation generator to perform spatiotemporal self-supervised training and data enhancement on the raw data. The mask generator, the spatiotemporal context encoder and the contrastive learning unit are used to deeply mine the spatiotemporal features of the data, and the data enhancement samples are generated by the neural differential equation generator to improve the diversity of the data and the generalization ability of the model. The feature extraction module uses a feature selection unit and a feature extraction unit to screen out key features from the preprocessed data, and uses the convolution operation to extract key features. The convolution neural network (CNN) deep learning model is used for feature extraction to provide high-quality feature data for subsequent fusion processing. The Lasso regression model is used for preliminary feature screening, combined with the convolution neural network (CNN) deep learning model for feature extraction to improve the accuracy and efficiency of feature extraction. The data fusion module uses feature fusion algorithm unit and fusion strategy unit, combined with hypergraph neural network and dynamic weight allocation strategy, as well as quantum annealing algorithm to optimize feature selection, to achieve effective fusion of heterogeneous feature data, improve the training effect and prediction accuracy of the model. The result evaluation module defines evaluation indicators such as accuracy, recall rate and F1 value, and uses K-fold cross validation to comprehensively analyze the system performance to ensure the stability and reliability of the model. The application output module uses visualization technology and API interface through the result display unit and interface output unit to achieve intuitive display and real-time output of fused data.

[0004] The multi-source heterogeneous data fusion processing system based on artificial intelligence deep learning provided by the present invention comprises a data collection module, a data preprocessing module, a feature extraction module, a data fusion module, a result evaluation module and an application output module. The data collection module collects raw data from different data sources and transmits it to the data preprocessing module. The data preprocessing module performs spatiotemporal self-supervised training and data enhancement on the raw data and then outputs preprocessed data. The feature extraction module performs feature extraction on the preprocessed data and outputs feature data. The data fusion module performs fusion processing on the feature data to obtain fused data. The result evaluation module evaluates the fused data. The application output module outputs the evaluated fused data. Further, the data collection module includes a data access unit, a data synchronization unit and a data storage unit, the data access unit communicates with different data sources to obtain raw data, the data synchronization unit ensures the temporal consistency of the raw data collected from different data sources, and the data storage unit temporarily stores the collected raw data; Furthermore, the data preprocessing module includes a mask generator, a spatiotemporal context encoder, a contrastive learning unit and a neural differential equation generator, wherein the mask generator randomly generates a mask for masking the spatiotemporal units in the original data, the spatiotemporal context encoder processes the unmasked original data, extracts spatiotemporal features, and predicts the masked original data, the contrastive learning unit implements a regularization term, compares and optimizes the model's representation of samples with different spatiotemporal contexts, and the neural differential equation generator defines and solves neural differential equations, generates data enhancement samples, and outputs preprocessed data after processing is completed; Further, the feature extraction module includes a feature selection unit and a feature extraction unit, the feature selection unit selects features from the preprocessed data, and the feature extraction unit extracts features from the preprocessed data to obtain feature data; Further, the data fusion module includes a feature fusion algorithm unit and a fusion strategy unit, the feature fusion algorithm unit performs feature fusion on feature data, and the fusion strategy unit defines fusion rules and weight allocation; Further, the result evaluation module includes an evaluation index unit and a performance analysis unit, wherein the evaluation index unit defines the evaluation criteria and the performance analysis unit analyzes the system performance; Furthermore, the application output module includes a result display unit and an interface output unit, the result display unit displays the fused data, and the interface output unit provides an API and an interface to output the fused data.

[0005] Preferably, the multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning provided by the present invention comprises the following steps: Step S1: Data collection: The data access unit establishes connections with multiple data sources to collect raw data, the data synchronization unit ensures the temporal consistency of the raw data, and the data storage unit temporarily stores the collected raw data; Step S2: Data preprocessing: The mask generator randomly generates masks to mask the spatiotemporal units in the original data. The spatiotemporal context encoder processes the unmasked original data and extracts spatiotemporal features to predict the masked original data. The contrast learning unit implements the regularization term. The neural differential equation generator defines and solves the neural differential equation to produce data enhancement samples. After the processing is completed, the preprocessed data is output. Step S3: Feature extraction: The feature selection unit selects relevant features from the preprocessed data, and the feature extraction unit uses a convolutional neural network (CNN) deep learning model to extract features from the preprocessed data to obtain feature data; Step S4: Data fusion: Construct a multimodal dynamic fusion network, use the hypergraph neural network model to process feature data, and express the complex relationship between data, design a dynamic weight allocation strategy, realize real-time adjustment of feature weights, ensure the leading role of key features in the fusion process, introduce quantum annealing algorithm to optimize feature selection, screen out the feature combination that is most beneficial to system performance, output fusion data, and improve the training effect and prediction accuracy of the hypergraph neural network; Step S5: Result evaluation: the evaluation index unit defines the accuracy, recall rate and F1 value, and the performance analysis unit analyzes the system performance using cross-validation; Step S6: Application output: The result display unit uses visualization technology to display the fused data, and the interface output unit provides an API interface to achieve real-time output of the fused data.

[0006] Furthermore, in step S1, the data synchronization unit uses a timestamp alignment algorithm to ensure the temporal consistency of the original data. The formula used to calculate the absolute difference between timestamps is as follows: ; in, is the timestamp of data source A, is the timestamp of data source B, is the absolute difference between two timestamps.

[0007] Furthermore, the step S2 specifically includes the following steps: Step S21: Mask generation: A random generation strategy is used to create masks, which are used to cover up the spatiotemporal units in the original data. The purpose is to be able to predict these masked original data later, so as to train the robustness of the model. By masking part of the original data, the model learns to correctly predict the masked part based on the unmasked part during the training process, which helps the model capture the key dependencies in the data. Step S22: spatiotemporal context coding: processing the original data not covered by the mask, and extracting the spatiotemporal features of the original data; Step S23: Contrastive learning: Implement the regularization term to shorten the distance between the same original data points in the feature space and push the distance between different original data points to enhance the model's ability to learn the essential features of the original data. The formula used to calculate the loss function is as follows: ;

[0008] in, is the loss function, is the number of positive sample pairs in the original data, To measure the original data characteristics The similarity function between is the original data feature, For The corresponding original data features are is the temperature parameter, which is used to adjust the smoothness of the similarity; Step S24: Neural differential equation generation: Define a neural differential equation based on the characteristics of the original data. Such an equation can capture the dynamic change characteristics of the data, solve the neural differential equation, and generate data enhancement samples. These samples can increase the data diversity of model training and improve the generalization ability of the model. The preprocessed original data is output as preprocessed data; Furthermore, the step S24 specifically includes the following steps: Step S241: define a neural differential equation: according to the characteristics of the original data, define a neural network, the neural network takes the state and time of the original data as input, and outputs the rate of change of the state. The neural differential equation is expressed by the following formula: ; in, is the state variable of the original data, indicating the original data at time status, is a neural network, representing the dynamic changes of the original data. for Parameters, is the time variable of the original data; Step S242: solving the neural differential equation: using a numerical integration method to solve the neural differential equation; Step S243: Generate data enhancement samples: By integrating the neural differential equation in time, the states of the original data points at multiple time points can be obtained; Step S244: Output preprocessed data: The enhanced data generated by the neural differential equation is output as preprocessed data.

[0009] Furthermore, the step S3 specifically includes the following steps: Step S31: Preliminary feature screening: Reduce the dimension of preprocessed data, remove features that are irrelevant to the task or have very low relevance, use the Lasso regression model to evaluate feature importance, determine the feature importance threshold through cross-validation, select features whose feature importance is greater than the feature importance threshold to form a feature subset, and the objective function of the Lasso regression model is shown in the following formula: ; in, =( ) is the vector of regression coefficients, For the Preprocessed data, For the The first Features, is the intercept term, is the total number of preprocessed data, is the total number of features, is the penalty term coefficient; Step S32: Feature subset selection: using the forward stepwise selection method, add features in the order of the absolute value of the Lasso regression coefficient in step S31 from high to low, calculate the F1 value of the feature subset through cross-validation, and select the feature subset with the highest F1 value improvement; Step S33: Feature extraction: using the selected feature subset as input data, constructing a convolutional neural network (CNN) deep learning model, and extracting the feature subset; Furthermore, the step S33 specifically includes the following steps: Step S331: Convolutional neural network (CNN) deep learning model construction: designing the network architecture of the convolutional neural network (CNN) deep learning model, including input layer, convolution layer, pooling layer and fully connected layer; Step S332: Model training: Divide the feature subset into a training set, a validation set, and a test set, select a cross entropy loss function, and train a convolutional neural network (CNN) deep learning model on the training set. The formula used by the cross entropy loss function to evaluate the model performance is as follows: ; in, is the cross entropy loss, is the test set, The probability distribution predicted by the convolutional neural network CNN deep learning model; Step S333: Feature extraction: Select the output of the middle layer from the trained convolutional neural network CNN deep learning model as feature data; Step S334: Feature output: output the extracted feature data.

[0010] Furthermore, the step S4 specifically includes the following steps: Step S41: Use a hypergraph neural network to process feature data: The hypergraph neural network can better express the complex relationship between data. When processing heterogeneous data, it can effectively capture the association between different types of data. The formula used for the hypergraph convolution operation is as follows: ; in, For the Node in layer The characteristic data of and is the node of feature data, is the learnable weight matrix, is the activation function, For super edge The number of nodes, is a set of hyperedges; Step S42: Design a dynamic weight allocation strategy: adjust the feature weight in real time according to the feature importance of the feature data to ensure that the key features can play a leading role in the fusion process; Step S43: Introducing quantum annealing algorithm to optimize feature selection: using quantum annealing algorithm to find the optimal feature subset in the feature selection process, and screening out feature data that can improve the prediction accuracy of the hypergraph neural network; Step S44: Improving the training effect and prediction accuracy of the model: Through the above steps, the feature fusion process is optimized, thereby improving the training effect and prediction accuracy of the model; Furthermore, the step S43 specifically includes the following steps: Furthermore, the step S43 specifically includes the following steps: Step S431: define the parameters of the problem, including the feature set, the objective function, and the state of the quantum bit. The initial state of the quantum bit is usually a uniform superposition state, which is expressed by the following formula: ; in, is the initial state of the quantum bit, is the number of quantum bits, is the index variable used to traverse the quantum bits; Step S432: The quantum annealing process first defines the Hamiltonian, which is composed of the driving Hamiltonian and the problem Hamiltonian, and then performs time evolution. The evolution of the quantum state over time follows the Schrödinger equation, as shown in the following formula: ; in, is an imaginary unit, is the reduced Planck constant, is the Hamiltonian, is the state of the quantum bit; Step S433: After the quantum annealing process is completed, the state of the quantum bit is measured to obtain a classical solution; Step S434: Select a set of features based on the measurement results, and this set of features will be used for subsequent processing.

[0011] Preferably, the step S5 specifically comprises the following steps: Step S51: Evaluate the precision, recall and F1 value; Step S52: Select K-fold cross validation to evaluate the hypergraph neural network; Step S53: Use the trained hypergraph neural network to predict the fused data, and calculate the accuracy, recall and F1 value according to the defined evaluation indicators; Step S54: Record the evaluation results of each cross-validation, including the accuracy, recall and F1 value indicators.

[0012] Furthermore, the step S6 specifically includes the following steps: Step S61: Select visualization tool: select appropriate visualization tool output according to the characteristics of fused data and display requirements; Step S62: Interface output unit preparation: determine the API interface specification, write code to implement the API interface, and ensure the stability and security of the API interface.

[0013] The beneficial effects achieved by the present invention using the above scheme are as follows: (1) The present invention can efficiently collect raw data from different data sources through the data collection module, and perform spatiotemporal self-supervisory training and data enhancement on these data through the data preprocessing module, effectively solving the problem of integrating multi-source heterogeneous data. The data fusion module adopts a hypergraph neural network and a dynamic weight allocation strategy, which can handle complex heterogeneous data relationships and improve the accuracy and efficiency of data fusion. (2) The mask generator and spatiotemporal context encoder combined with the contrastive learning unit can improve the model's predictive ability and robustness to data. The neural differential equation generator generates data augmentation samples by defining and solving neural differential equations, which increases the data diversity of system training and improves the system's generalization ability. (3) The feature extraction module uses a convolutional neural network (CNN) deep learning model to efficiently extract key features from preprocessed data. The data fusion module introduces a quantum annealing algorithm to optimize feature selection and screen out the feature combination that is most beneficial to the performance of the hypergraph neural network, further improving the training effect and prediction accuracy of the hypergraph neural network. (4) The result evaluation module defines evaluation indicators such as accuracy, recall rate and F1 value, and conducts a comprehensive analysis of system performance through cross-validation to ensure the stability and reliability of the hypergraph neural network. The application output module provides visual display and API interface output, which makes it convenient for users to intuitively view the fused data, thus improving the availability and ease of use of the data. (5) The data synchronization unit adopts a timestamp alignment algorithm to ensure the temporal consistency of the original data, thereby improving the accuracy and efficiency of data processing. The various components in the data preprocessing module work together to efficiently process large-scale data and shorten the data processing cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a module schematic diagram of a multi-source heterogeneous data fusion processing system based on artificial intelligence deep learning of the present invention; Figure 2 is a unit schematic diagram of a data collection module; Figure 3 It is a unit schematic diagram of the data preprocessing module; Figure 4 It is a unit diagram of the feature extraction module; Figure 5 is a unit schematic diagram of the data fusion module; Figure 6 It is a unit diagram of the result evaluation module; Figure 7 is a unit diagram of the application output module; Figure 8 It is a workflow diagram of the multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning of the present invention; Fig. 9 is a flow chart of step S2; Fig.10 is a flow chart of step S3; Fig.11 is a flow chart of step S33; Fig.12 is a flow chart of step S4; Fig.13 This is a flowchart of step S6.

[0015] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0017] Embodiment 1, as Figure 1-Figure 7As shown, in this embodiment, a multi-source heterogeneous data fusion processing system based on artificial intelligence deep learning includes a data collection module, a data preprocessing module, a feature extraction module, a data fusion module, a result evaluation module and an application output module. The data collection module collects raw data from different data sources and transmits it to the data preprocessing module. The data preprocessing module performs spatiotemporal self-supervised training and data enhancement on the raw data and then outputs preprocessed data. The feature extraction module performs feature extraction on the preprocessed data and outputs feature data. The data fusion module performs fusion processing on the feature data to obtain fused data. The result evaluation module evaluates the fused data. The application output module outputs the evaluated fused data. The data collection module includes a data access unit, a data synchronization unit and a data storage unit. The data access unit communicates with different data sources to obtain raw data. The data synchronization unit ensures the temporal consistency of the raw data collected from different data sources. The data storage unit temporarily stores the collected raw data. The data preprocessing module includes a mask generator, a spatiotemporal context encoder, a contrastive learning unit and a neural differential equation generator. The mask generator randomly generates a mask for masking the spatiotemporal units in the original data. The spatiotemporal context encoder processes the unmasked original data and extracts spatiotemporal features for predicting the masked original data. The contrastive learning unit implements a regularization term, compares and optimizes the model's representation of samples with different spatiotemporal contexts. The neural differential equation generator defines and solves neural differential equations, generates data enhancement samples, and outputs preprocessed data after processing. The feature extraction module includes a feature selection unit and a feature extraction unit, wherein the feature selection unit selects features from the preprocessed data, and the feature extraction unit extracts features from the preprocessed data to obtain feature data; The data fusion module includes a feature fusion algorithm unit and a fusion strategy unit. The feature fusion algorithm unit performs feature fusion on feature data, and the fusion strategy unit defines fusion rules and weight allocation. The result evaluation module includes an evaluation index unit and a performance analysis unit, wherein the evaluation index unit defines the evaluation criteria and the performance analysis unit analyzes the system performance; The application output module includes a result display unit and an interface output unit. The result display unit displays the fused data, and the interface output unit provides an API and an interface to output the fused data. Through the data collection module, the system can efficiently obtain raw data from different data sources, and use the data synchronization unit to ensure the time consistency of the data, thereby enhancing the real-time and accuracy of data processing. The data preprocessing module improves the quality and diversity of the data through spatiotemporal self-supervised training and data enhancement technology, providing a solid foundation for subsequent feature extraction and fusion. The feature extraction module can accurately extract key features from the preprocessed data through the collaborative work of the feature selection unit and the feature extraction unit, thereby reducing the data dimension and improving the efficiency of subsequent data fusion. The data fusion module adopts advanced feature fusion algorithms and flexible fusion strategies, which can effectively integrate feature data from different data sources to generate a more comprehensive and accurate information representation. The result evaluation module can objectively and comprehensively evaluate the accuracy of the fused data and the system performance by defining clear evaluation criteria and performance analysis indicators, which helps to continuously optimize the system performance. The application output module not only provides an intuitive result display function, but also realizes the convenient output of fused data through APIs and interfaces, which is convenient for integration and interaction with other systems or applications.

[0018] Embodiment 2, as Figure 8 As shown, this embodiment is based on the above embodiment, and is a multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning, the method comprising the following steps: Step S1: Data collection: The data access unit establishes connections with multiple data sources to collect raw data, the data synchronization unit ensures the temporal consistency of the raw data, and the data storage unit temporarily stores the collected raw data; Step S2: Data preprocessing: The mask generator randomly generates masks to mask the spatiotemporal units in the original data. The spatiotemporal context encoder processes the unmasked original data and extracts spatiotemporal features to predict the masked original data. The contrast learning unit implements the regularization term. The neural differential equation generator defines and solves the neural differential equation to produce data enhancement samples. After the processing is completed, the preprocessed data is output. Step S3: Feature extraction: The feature selection unit selects relevant features from the preprocessed data, and the feature extraction unit uses a convolutional neural network (CNN) deep learning model to extract features from the preprocessed data to obtain feature data; Step S4: Data fusion: Construct a multimodal dynamic fusion network, use the hypergraph neural network model to process feature data, and express the complex relationship between data, design a dynamic weight allocation strategy, realize real-time adjustment of feature weights, ensure the leading role of key features in the fusion process, introduce quantum annealing algorithm to optimize feature selection, screen out the feature combination that is most beneficial to system performance, output fusion data, and improve the training effect and prediction accuracy of the hypergraph neural network; Step S5: Result evaluation: the evaluation index unit defines the accuracy, recall rate and F1 value, and the performance analysis unit analyzes the system performance using cross-validation; Step S6: Application output: The result display unit uses visualization technology to display the fused data, and the interface output unit provides an API interface to achieve real-time output of the fused data; In step S1, the data synchronization unit uses a timestamp alignment algorithm to ensure that the original data is consistent in time. The formula used to calculate the absolute difference between timestamps is as follows: ; in, is the timestamp of data source A, is the timestamp of data source B, is the absolute difference between two timestamps; Through the data collection step, the system can automatically collect raw data from multiple data sources, and use the timestamp alignment algorithm of the data synchronization unit to ensure the temporal consistency of the data, which greatly improves the efficiency and accuracy of data integration. In the data preprocessing step, the combined use of the mask generator and the spatiotemporal context encoder not only masks some spatiotemporal units in the data, but also predicts the masked data by extracting spatiotemporal features, which enhances the robustness and integrity of the data. At the same time, the application of the contrast learning unit and the neural differential equation generator further enriches the data samples and improves the effects of subsequent feature extraction and data fusion. The feature extraction step uses the convolutional neural network (CNN) deep learning model, which can accurately extract relevant features from the preprocessed data and provide high-quality input for subsequent data fusion. The data fusion step constructs a multimodal dynamic fusion network, uses a hypergraph neural network model to process heterogeneous data, and expresses the complex relationship between data. The application of the dynamic weight allocation strategy and the quantum annealing algorithm realizes the real-time adjustment of feature weights and the optimization of feature selection, ensuring the dominant role of key features in the fusion process, thereby improving the prediction accuracy and training effect of the system.

[0019] Embodiment three, as Figure 8-Figure 9 As shown, this embodiment is based on the above embodiment, and step S2 specifically includes the following steps: Step S21: Mask generation: A random generation strategy is used to create masks, which are used to cover up the spatiotemporal units in the original data. The purpose is to be able to predict these masked data later, so as to train the robustness of the model. By masking part of the original data, the model learns to correctly predict the masked part based on the unmasked part during the training process, which helps the model capture the key dependencies in the data. Step S22: spatiotemporal context coding: processing the original data not covered by the mask, and extracting the spatiotemporal features of the original data; Step S23: Contrastive learning: Implement the regularization term to shorten the distance between the same original data points in the feature space and push the distance between different original data points to enhance the model's ability to learn the essential features of the original data. The formula used to calculate the loss function is as follows: : in, is the loss function, is the number of positive sample pairs in the original data, To measure the original data characteristics The similarity function between is the original data feature, For The corresponding original data features are is the temperature parameter, which is used to adjust the smoothness of the similarity; Step S24: Neural differential equation generation: Define a neural differential equation based on the characteristics of the original data. Such an equation can capture the dynamic change characteristics of the data, solve the neural differential equation, and generate data enhancement samples. These samples can increase the data diversity of model training and improve the generalization ability of the model. The preprocessed original data is output as preprocessed data; The step S24 specifically includes the following steps: Step S241: define a neural differential equation: according to the characteristics of the original data, define a neural network, the neural network takes the state and time of the original data as input, and outputs the rate of change of the state. The neural differential equation is expressed by the following formula: ; in, is the state variable of the original data, indicating the original data at time status, is a neural network, representing the dynamic changes of the original data. for Parameters, is the time variable of the original data; Step S242: solving the neural differential equation: using a numerical integration method to solve the neural differential equation; Step S243: Generate data enhancement samples: By integrating the neural differential equation in time, the states of the original data points at multiple time points can be obtained, and these states are enhanced versions of the original data; Step S244: outputting preprocessed data: outputting enhanced data generated by the neural differential equation as preprocessed data; Through the mask generation strategy of step S21, the spatiotemporal units in the original data are randomly masked, forcing the model to learn how to predict the masked parts based on the unmasked parts during the training process. This strategy helps the model capture the key dependencies in the data, thereby improving the model's ability to handle incomplete or missing data and enhancing the robustness of the model. The spatiotemporal context encoding of step S22 processes the original data that is not masked, extracts its spatiotemporal features, and provides rich feature information for subsequent model training. The contrastive learning of step S23 effectively enhances the model's ability to learn the essential characteristics of the original data by shortening the distance between the same original data points in the feature space and pushing the distance between different original data points. This regularization term helps the model better understand and distinguish the differences between different data points, thereby improving the classification or prediction accuracy of the model. The neural differential equation generation of step S24 defines and solves the neural differential equation according to the characteristics of the original data. Differential equations are used to capture the dynamic changing characteristics of data. By solving neural differential equations, the states of original data at multiple time points can be generated. These states are used as data enhancement samples to increase the data diversity of model training. This data enhancement strategy helps to improve the generalization ability of the model and enables the model to better adapt to different data distributions and scenarios. The entire step S2 constitutes a complete preprocessing flow, from mask generation to spatiotemporal context encoding, to contrastive learning and neural differential equation generation. Each step provides useful information and enhancement for subsequent model training. The optimization of this preprocessing flow helps to improve the performance and efficiency of the entire system. This embodiment combines multiple deep learning techniques such as random masking strategies, contrastive learning and neural differential equations, and demonstrates the potential of these technologies in processing complex data, enhancing model robustness and improving feature learning capabilities. This innovative application provides new ideas and methods for the development of deep learning technology.

[0020] Embodiment 4, as Figure 8 , Fig.10 , Fig.11 As shown, this embodiment is based on the above embodiment, and the step S3 specifically includes the following steps: Step S31: Preliminary feature screening: Reduce the dimension of preprocessed data, remove features that are irrelevant to the task or have very low relevance, use the Lasso regression model to evaluate feature importance, determine the feature importance threshold through cross-validation, select features whose feature importance is greater than the feature importance threshold to form a feature subset, and the objective function of the Lasso regression model is shown in the following formula: ; in, =( ) is the vector of regression coefficients, For the Preprocessed data, For the The first Features, is the intercept term, is the total number of preprocessed data, is the total number of features, is the penalty term coefficient; Step S32: Feature subset selection: Use the forward stepwise selection method to add features in the order of the absolute value of the Lasso regression coefficient in step S31 from high to low, calculate the F1 value of the feature subset through cross-validation, and select the feature subset with the highest F1 value improvement. The code used for cross-validation is as follows: for each feature subset S do performance_sum = 0 for a = 1 to k do train_set = combine all but the a-th fold as the training set validation_set = the a-th fold train model on train_set performance = evaluate model on validation_set performance_sum = performance_sum + performance end for average_performance = performance_sum / p if average_performance is the best so far then best_feature_subset = S end if end for; Step S33: Feature extraction: using the selected feature subset as input data, constructing a convolutional neural network (CNN) deep learning model, and extracting the feature subset; The step S33 specifically includes the following steps: Step S331: Convolutional neural network (CNN) deep learning model construction: designing the network architecture of the convolutional neural network (CNN) deep learning model, including input layer, convolution layer, pooling layer and fully connected layer; Step S332: Model training: Divide the feature subset into a training set, a validation set, and a test set, select a cross entropy loss function, and train a convolutional neural network (CNN) deep learning model on the training set. The formula used by the cross entropy loss function to evaluate the model performance is as follows: ; in, is the cross entropy loss, is the test set, The probability distribution predicted by the convolutional neural network CNN deep learning model; Step S333: Feature extraction: Select the output of the middle layer from the trained convolutional neural network CNN deep learning model as feature data; Step S334: feature output: output the extracted feature data; The feature screening and subset selection process reduces the dimension of the data and improves the training efficiency and prediction speed of the model. The construction and training of the CNN deep learning model provides a powerful tool for feature extraction, which helps to improve the performance of the model. The use of the intermediate layer output retains the useful features of the data and lays a good foundation for subsequent fusion processing.

[0021] Embodiment 5, as Figure 8 , Fig.12 As shown, this embodiment is based on the above embodiment, and the step S4 specifically includes the following steps: Step S41: Use a hypergraph neural network to process feature data: The hypergraph neural network can better express the complex relationship between data. When processing heterogeneous data, it can effectively capture the association between different types of data. The formula used for the hypergraph convolution operation is as follows: ; in, For the Node in layer The characteristic data of and is the node of feature data, is the learnable weight matrix, is the activation function, For super edge The number of nodes, is a set of hyperedges; Step S42: Design a dynamic weight allocation strategy: adjust the feature weight in real time according to the feature importance of the feature data to ensure that the key features can play a leading role in the fusion process; Step S43: Introducing quantum annealing algorithm to optimize feature selection: using quantum annealing algorithm to find the optimal feature subset in the feature selection process, and screening out feature data that can improve the prediction accuracy of the hypergraph neural network; Step S44: Improving the training effect and prediction accuracy of the model: Through the above steps, the feature fusion process is optimized, thereby improving the training effect and prediction accuracy of the model; The step S43 specifically includes the following steps: Step S431: define the parameters of the problem, including the feature set, the objective function, and the state of the quantum bit. The initial state of the quantum bit is usually a uniform superposition state, which is expressed by the following formula: ; in, is the initial state of the quantum bit, is the number of quantum bits, is the index variable used to traverse the quantum bits; Step S432: The quantum annealing process first defines the Hamiltonian, which is composed of the driving Hamiltonian and the problem Hamiltonian, and then performs time evolution. The evolution of the quantum state over time follows the Schrödinger equation, as shown in the following formula: ; in, is an imaginary unit, is the reduced Planck constant, is the Hamiltonian, is the state of the quantum bit; Step S433: After the quantum annealing process is completed, the state of the quantum bit is measured to obtain a classical solution; Step S434: Select a set of features according to the measurement results, and this set of features will be used for subsequent processing; By using a hypergraph neural network to process feature data, this embodiment can more effectively express the complex relationships between data. In particular, when processing heterogeneous data, the hypergraph neural network can capture the associations between different types of data, which is difficult to achieve with traditional graph neural networks. This improvement in capability helps the model understand the data more accurately, thereby improving the accuracy of subsequent processing and analysis. The design of a dynamic weight allocation strategy can adjust the feature weights in real time according to the importance of the feature data, which means that in the feature fusion process, key features can play a greater role, while secondary features will be appropriately weakened. This strategy helps the model focus more on features that have an important impact on the prediction results, thereby improving the prediction accuracy of the model. The introduction of a quantum annealing algorithm for feature selection is another significant advantage of this embodiment. The quantum annealing algorithm can find the optimal solution in the feature selection process and screen out features that are important to the hypergraph neural network. The advantage of this algorithm is that it can quickly find the optimal solution in a huge feature space, thereby avoiding the computational complexity and time cost problems that traditional feature selection methods may face. Through the comprehensive application of the above steps, this embodiment can optimize the feature fusion process, thereby improving the training effect and prediction accuracy of the model. Specifically, the use of hypergraph neural network, dynamic weight allocation strategy and quantum annealing algorithm can help the model understand data more accurately, process information more efficiently, and ultimately improve the accuracy of prediction results. This embodiment combines multiple advanced technologies such as hypergraph neural network, dynamic weight allocation strategy and quantum annealing algorithm, which not only improves the performance of the algorithm, but also expands the application scope of the algorithm. This innovative combination makes this embodiment have significant practical value in processing complex data, optimizing feature selection and improving prediction accuracy.

[0022] Embodiment six, as Fig.13 As shown, this embodiment is based on the above embodiment, and step S5 specifically includes the following steps: Step S51: Evaluate the precision, recall and F1 value; Step S52: Select K-fold cross validation to evaluate the hypergraph neural network; Step S53: Use the trained hypergraph neural network to predict the fused data, and calculate the accuracy, recall and F1 value according to the defined evaluation indicators; Step S54: Record the evaluation results of each cross-validation, including accuracy, recall and F1 value indicators; The step S6 specifically comprises the following steps: Step S61: Select visualization tool: select appropriate visualization tool output according to the characteristics of fused data and display requirements; Step S62: Interface output unit preparation: determine the API interface specification, write code to implement the API interface, and ensure the stability and security of the API interface; By evaluating the three key indicators of accuracy, recall and F1 value, the performance of the hypergraph neural network on the fusion data can be comprehensively measured. The accuracy reflects the proportion of correct predictions made by the model, the recall reflects the model's ability to identify positive samples, and the F1 value is the harmonic average of the accuracy and recall. It combines the advantages of these two indicators and provides a more comprehensive evaluation of the model performance. Using K-fold cross-validation to evaluate the hypergraph neural network can effectively reduce the risk of overfitting and improve the reliability and stability of the evaluation results. By dividing the data set into K parts, using K-1 parts as the training set each time and the remaining 1 part as the test set, K-fold cross-validation is performed. By training and testing again, we can make full use of the data set and obtain more accurate evaluation results. We can use the trained hypergraph neural network to predict the fused data and calculate the accuracy, recall and F1 value according to the defined evaluation indicators. This can accurately evaluate the performance of the model in practical applications. This helps to promptly discover problems with the model and make targeted optimizations and improvements. Recording the evaluation results of each cross-validation, including accuracy, recall and F1 value indicators, can provide detailed data support for subsequent analysis and optimization. By analyzing these data, we can gain an in-depth understanding of the performance of the model on different data sets and further guide the optimization direction of the model.

[0023] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A multi-source heterogeneous data fusion processing system based on artificial intelligence deep learning, characterized by: The system includes a data collection module, a data preprocessing module, a feature extraction module, a data fusion module, a result evaluation module and an application output module. The data collection module collects raw data from different data sources and transmits it to the data preprocessing module. The data preprocessing module performs spatiotemporal self-supervised training and data enhancement on the raw data and then outputs preprocessed data. The feature extraction module performs feature extraction on the preprocessed data and outputs feature data. The data fusion module performs fusion processing on the feature data to obtain fused data. The result evaluation module evaluates the fused data. The application output module outputs the evaluated fused data. The data collection module includes a data access unit, a data synchronization unit and a data storage unit; the data preprocessing module includes a mask generator, a spatiotemporal context encoder, a contrastive learning unit and a neural differential equation generator; the feature extraction module includes a feature selection unit and a feature extraction unit; the data fusion module includes a feature fusion algorithm unit and a fusion strategy unit; the result evaluation module includes an evaluation index unit and a performance analysis unit; and the application output module includes a result display unit and an interface output unit.

2. A multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning is applied to the multi-source heterogeneous data fusion processing system based on artificial intelligence deep learning according to claim 1, characterized in that: The method comprises the following steps: Step S1: Data collection: The data access unit establishes connections with multiple data sources to collect raw data, the data synchronization unit ensures the temporal consistency of the raw data, and the data storage unit temporarily stores the collected raw data; Step S2: Data preprocessing: The mask generator randomly generates masks to mask the spatiotemporal units in the original data. The spatiotemporal context encoder processes the unmasked original data and extracts spatiotemporal features to predict the masked original data. The contrast learning unit implements the regularization term. The neural differential equation generator defines and solves the neural differential equation to produce data enhancement samples. After the processing is completed, the preprocessed data is output. Step S3: Feature extraction: The feature selection unit selects relevant features from the preprocessed data, and the feature extraction unit uses a convolutional neural network (CNN) deep learning model to extract features from the preprocessed data to obtain feature data; Step S4: Data fusion: Construct a multimodal dynamic fusion network, use the hypergraph neural network to process feature data, and express the complex relationship between data, design a dynamic weight allocation strategy, realize real-time adjustment of feature weights, introduce quantum annealing algorithm to optimize feature selection, screen out the feature combination that is most beneficial to the performance of the hypergraph neural network, and output fusion data; Step S5: Result evaluation: the evaluation index unit defines the accuracy, recall rate and F1 value, and the performance analysis unit analyzes the system performance using cross-validation; Step S6: Application output: The result display unit uses visualization technology to display the fused data, and the interface output unit provides an API interface to achieve real-time output of the fused data.

3. The multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning according to claim 2 is characterized in that: In step S1, the data synchronization unit uses a timestamp alignment algorithm to ensure that the original data is consistent in time. The formula used to calculate the absolute difference between timestamps is as follows: ； in, is the timestamp of data source A, is the timestamp of data source B, is the absolute difference between two timestamps.

4. The multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning according to claim 2 is characterized in that: The step S2 specifically comprises the following steps: Step S21: Mask generation: A random generation strategy is used to create masks, which are used to cover up the spatiotemporal units in the original data. By covering up part of the original data, the model learns to correctly predict the masked part based on the unmasked part during the training process; Step S22: spatiotemporal context coding: processing the original data not covered by the mask, and extracting the spatiotemporal features of the original data; Step S23: Contrastive learning: Implement the regularization term to shorten the distance between the same original data points in the feature space and push the distance between different original data points. The formula used to calculate the loss function is as follows: ： in, is the loss function, is the number of positive sample pairs in the original data, To measure the original data characteristics The similarity function between is the original data feature, For The corresponding original data features are is the temperature parameter, which is used to adjust the smoothness of the similarity; Step S24: Neural differential equation generation: define the neural differential equation according to the characteristics of the original data, solve the neural differential equation, generate data enhancement samples, and output the preprocessed original data as preprocessed data.

5. The multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning according to claim 4 is characterized in that: The step S24 specifically includes the following steps: Step S241: define a neural differential equation: according to the characteristics of the original data, define a neural network, the neural network takes the state and time of the original data as input, and outputs the rate of change of the state. The neural differential equation is expressed by the following formula: ； in, is the state variable of the original data, indicating the original data at time status, is a neural network, representing the dynamic changes of the original data. for Parameters, is the time variable of the original data; Step S242: solving the neural differential equation: using a numerical integration method to solve the neural differential equation; Step S243: Generate data enhancement samples: By integrating the neural differential equation in time, the states of the original data points at multiple time points can be obtained; Step S244: Output preprocessed data: The enhanced data generated by the neural differential equation is output as preprocessed data.

6. The multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning according to claim 2 is characterized in that: The step S3 specifically comprises the following steps: Step S31: Preliminary feature screening: Reduce the dimension of preprocessed data, remove features that are irrelevant to the task or have very low relevance, use the Lasso regression model to evaluate feature importance, determine the feature importance threshold through cross-validation, select features whose feature importance is greater than the feature importance threshold to form a feature subset, and the objective function of the Lasso regression model is shown in the following formula: ； in, =( ) is the vector of regression coefficients, For the Preprocessed data, For the The th feature of the preprocessed data, is the intercept term , is the total number of preprocessed data, is the total number of features, is the penalty term coefficient; Step S32: Feature subset selection: using the forward stepwise selection method, add features in the order of the absolute value of the Lasso regression coefficient in step S31 from high to low, calculate the F1 value of the feature subset through cross-validation, and select the feature subset with the highest F1 value improvement; Step S33: Feature extraction: Use the selected feature subset as input data, build a convolutional neural network (CNN) deep learning model, and extract the feature subset.

7. The multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning according to claim 6 is characterized in that: The step S33 specifically includes the following steps: Step S331: Convolutional neural network (CNN) deep learning model construction: designing the network architecture of the convolutional neural network (CNN) deep learning model, including input layer, convolution layer, pooling layer and fully connected layer; Step S332: Model training: Divide the feature subset into a training set, a validation set, and a test set, select a cross entropy loss function, and train a convolutional neural network (CNN) deep learning model on the training set. The formula used by the cross entropy loss function to evaluate the model performance is as follows: ； in, is the cross entropy loss, is the test set, The probability distribution predicted by the convolutional neural network CNN deep learning model; Step S333: Feature extraction: Select the output of the middle layer from the trained convolutional neural network CNN deep learning model as feature data; Step S334: Feature output: output the extracted feature data.

8. The multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning according to claim 2 is characterized in that: The step S4 specifically comprises the following steps: Step S41: Use a hypergraph neural network to process feature data: The hypergraph neural network can better express the complex relationship between data. The formula used for the hypergraph convolution operation is as follows: ； in, For the Node in layer The characteristic data of and is the node of feature data, is the learnable weight matrix, is the activation function, For super edge The number of nodes, is a set of hyperedges; Step S42: Design a dynamic weight allocation strategy: adjust the feature weight in real time according to the feature importance of the feature data; Step S43: Introducing quantum annealing algorithm to optimize feature selection: using quantum annealing algorithm to find the optimal feature subset in the feature selection process, and screening out feature data that can improve the prediction accuracy of the hypergraph neural network; Step S44: Improving the training effect and prediction accuracy of the hypergraph neural network: Through the above steps, the feature fusion process is optimized.

9. The multi-source heterogeneous data fusion processing method based on artificial intelligence deep learning according to claim 8 is characterized in that: The step S43 specifically includes the following steps: Step S431: define the parameters of the problem, including the feature set, the objective function, and the state of the quantum bit. The initial state of the quantum bit is usually a uniform superposition state, which is expressed by the following formula: ； in, is the initial state of the quantum bit, is the number of quantum bits, is the index variable used to traverse the quantum bits; Step S432: The quantum annealing process first defines the Hamiltonian, which is composed of the driving Hamiltonian and the problem Hamiltonian, and then performs time evolution. The evolution of the quantum state over time follows the Schrödinger equation, as shown in the following formula: ； in, is an imaginary unit, is the reduced Planck constant, is the Hamiltonian, is the state of the quantum bit; Step S433: After the quantum annealing process is completed, the state of the quantum bit is measured to obtain a classical solution; Step S434: Select a set of features based on the measurement results, and this set of features will be used for subsequent processing.

Images