Multi-source intelligence data intelligent analysis method based on deep learning

Abstract

The invention discloses a multi-source intelligence data intelligent analysis method based on deep learning. The method comprises the following steps of 1, collecting and standardizing original intelligence data; 2, constructing a multi-source information graph based on the multi-source information data set; 3, performing local neighborhood fuzzy reasoning on the multi-source information graph to obtain a node reasoning result; 4, performing fine-grained coding on the multi-source information graph through a relation enhancement encoder; 5, dynamically dividing the time sequence data through a self-adaptive time window mechanism, and carrying out fine-grained modeling to generate spatial-temporal feature representation; step 6, inputting the spatial-temporal feature representation into the improved GATv2 model to obtain a preliminary decision strategy; and step 7, adjusting a decision path to carry out strategy optimization on the preliminary decision strategy to obtain a final decision strategy. According to the method, efficient and accurate intelligence analysis and dynamic decision optimization are realized through the improved GATv2 model.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to an intelligent analysis method for multi-source intelligence data based on deep learning. Background Technology

[0002] With the rapid development of information technology, the sources and types of intelligence data are increasing. Traditional intelligence data analysis methods mainly rely on manually constructed rules and basic statistical methods, which often fail to fully utilize the potential of large-scale, multi-source, heterogeneous data. Specifically, existing technologies have several shortcomings in the intelligent analysis of multi-source intelligence data, especially in handling complex data relationships, spatiotemporal features, and decision optimization. Traditional intelligence data analysis methods are often limited to processing single data sources. For example, text-based intelligence methods mainly rely on natural language processing techniques, but they often cannot effectively integrate data from different modalities such as images, audio, and video, leading to limitations in the analysis results. With the increase in multi-source intelligence data, how to efficiently process different types of data has become an urgent technical challenge. Existing methods often struggle to achieve deep fusion of multimodal data and cannot fully uncover the complex implicit relationships between data.

[0003] Existing intelligence graph construction techniques largely rely on static, simplified graph models, often failing to consider the dynamic relationships and spatiotemporal dependencies between data points during graph construction. These traditional graph construction methods define connections between nodes using simple similarity metrics, often ignoring spatiotemporal changes and interactions between data sources. Therefore, existing graphs often fail to provide accurate results when processing complex, dynamically changing intelligence data. While traditional graph neural networks (GNNs) have made some progress in graph data analysis, they typically rely on graphs with fixed structures, lacking dynamic adjustment and optimization capabilities, resulting in poor performance when handling complex intelligence data. Furthermore, existing intelligence data analysis methods often rely on rule-based models or simple statistical decision models for decision optimization, which are ill-suited to the ever-changing and complex demands of intelligence analysis. Especially when facing multi-source data, existing technologies struggle to achieve real-time decision optimization and cannot flexibly adjust based on real-time intelligence data. Traditional optimization algorithms often remain at the level of static decision path selection, lacking the ability to dynamically adjust decision strategies, leading to limited effectiveness in complex and highly dynamic intelligence analysis scenarios.

[0004] Therefore, how to provide a method for intelligent analysis of multi-source intelligence data based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent analysis method for multi-source intelligence data based on deep learning. This invention dynamically constructs an intelligence graph by fusing multi-source heterogeneous data and an improved GATv2 model, optimizing relationships between nodes and accurately modeling spatiotemporal features. Combined with a relationship-enhancing encoder, it achieves fine-grained analysis and real-time decision optimization of multi-source intelligence data, overcoming the shortcomings of traditional methods in data fusion, spatiotemporal modeling, and decision optimization, and providing more efficient and accurate intelligence analysis and decision support capabilities.

[0006] A method for intelligent analysis of multi-source intelligence data based on deep learning according to an embodiment of the present invention includes the following steps: Step 1: Collect raw intelligence data from multiple heterogeneous data sources and standardize the raw intelligence data to obtain a multi-source intelligence dataset; Step 2: Based on the multi-source intelligence dataset, obtain a multi-source intelligence feature vector set, and use the feature vector of each data source as a node. Calculate the Pearson correlation coefficient between nodes to measure similarity and construct a multi-source intelligence graph. Step 3: Perform local neighborhood fuzzy inference on each node in the multi-source intelligence graph. The local neighborhood fuzzy inference is based on the local neighborhood information of each node and the fuzzy inference rules to obtain the node inference result. Step 4: Based on the node reasoning results, the features of the nodes in the multi-source intelligence graph are finely encoded using a relation enhancement encoder to obtain enhanced node feature representations; Step 5: Input the enhanced node feature representation into the multi-scale spatiotemporal convolution module. The multi-scale spatiotemporal convolution module dynamically divides the time series data through an adaptive time window mechanism and performs fine-grained modeling of spatiotemporal features to generate spatiotemporal feature representations. Step Six: Input the spatiotemporal feature representation into the improved GATv2 model, and obtain the preliminary decision strategy through the graph structure initialization module, local interaction modeling module, relationship optimization module, global information integration module, and decision output module; Step 7: Optimize the preliminary decision-making strategy by adjusting the decision-making path, optimize the intelligence analysis process, and obtain the final decision-making strategy.

[0007] Optionally, the raw intelligence data includes text data, image data, audio data, and video data; the standardization processing steps include outlier removal, missing value imputation, timestamp alignment, and normalization of different types of data in the raw intelligence data to obtain a multi-source intelligence dataset with a unified structure.

[0008] Optionally, step two specifically involves: The text data in the multi-source intelligence dataset is processed by word segmentation, stop word removal, and stemming to convert the text data into a fixed-length text feature vector, which represents the semantic information of the text. The audio data in the multi-source intelligence dataset is extracted in the time domain and frequency domain by using short-time Fourier transform and Mel frequency cepstral coefficients to obtain audio feature vectors; Video frames are separated from the video data in the multi-source intelligence dataset to obtain a set of video frames; Feature extraction is performed on each video frame in the video frame set and the image data in the multi-source intelligence dataset using a convolutional neural network to obtain video feature vectors and image feature vectors. A multi-source intelligence map is constructed based on a multi-source intelligence feature vector set, which includes text feature vectors, audio feature vectors, video feature vectors, and image feature vectors. The multi-source intelligence graph includes multiple nodes and edges, and each data source uses the feature vector corresponding to the multi-source intelligence feature vector set as a node in the multi-source intelligence graph. In the node feature space, the Pearson correlation coefficient between each pair of nodes is calculated to measure similarity. If the Pearson correlation coefficient is less than a preset similarity threshold, an edge connection is established between the corresponding nodes. The weight of the edge connection is the reciprocal of the corresponding Pearson correlation coefficient.

[0009] Optionally, step three specifically includes: For each node in the multi-source intelligence graph, local neighborhood fuzzy inference is performed. The local neighborhood fuzzy inference steps include: Based on the connection relationship of each node in the multi-source intelligence graph, a local neighborhood of a preset size is selected, wherein the local neighborhood is a set of nodes directly connected to the current node; Extract the feature information of adjacent nodes from the local neighborhood of each node as input for reasoning; Based on the characteristics of nodes and neighborhood information, fuzzy inference rules are defined as follows: If the Euclidean distance between nodes in the local neighborhood of the current node is greater than a preset distance threshold, then the two nodes are strongly associated. If the Euclidean distance between the current node and the nodes in their local neighborhood is less than or equal to a preset distance threshold, then the two nodes are weakly associated. Based on the fuzzy inference rules, inference is performed on each node and its local neighborhood, and defuzzification is performed using the centroid method to obtain the node inference result for each node.

[0010] Optionally, step four specifically includes: Based on the node inference results, the feature vector of each node is initialized with the corresponding inference results to obtain the initialized feature vector. Based on the multi-source intelligence graph, an adjacency matrix between nodes is constructed. The adjacency matrix construction step involves each node having an edge connection with each directly connected node, and the weight of the edge is assigned according to the Pearson correlation coefficient between the nodes. The initial feature vector of each node and the feature vectors of neighboring nodes are input into the relation enhancement encoder for encoding. The relation enhancement encoder combines the feature vector of each node with the feature vectors of neighboring nodes in a predetermined manner to obtain the interaction features between the node and its neighboring nodes. Inside the relation enhancement encoder, multiple coding layers are used to progressively encode the interaction features of nodes and neighboring nodes in a fine-grained manner. The fine-grained encoding involves each layer using convolution operations to process the features of nodes and neighboring nodes, and using the ReLU nonlinear activation function to enhance the expressive power of the features. In each coding layer, the node features are enhanced by combining the relationship information of nodes in the multi-source intelligence graph. After processing through multiple coding layers, the enhanced node feature representation of each node is obtained.

[0011] Optionally, step five specifically includes: The enhanced node feature representation is input into a multi-scale spatiotemporal convolution module, and the spatiotemporal feature modeling steps performed by the multi-scale spatiotemporal convolution module include: Based on the enhanced node feature representation, an adaptive time window mechanism is used to dynamically divide the time series data. The adaptive time window mechanism dynamically adjusts the size of the time window according to the time characteristics and time series information of the data to capture spatiotemporal changes at different time scales. By using multiple convolutional kernels to perform convolution operations at different temporal and spatial scales, fine-grained modeling of the spatiotemporal features of the data is achieved. Each convolutional kernel performs convolution operations on the data in both the temporal and spatial dimensions, extracting spatiotemporal features at different levels and capturing short-term and long-term spatiotemporal dependencies. The spatiotemporal dependency optimization module is used to optimize spatiotemporal feature modeling. The spatiotemporal dependency optimization module obtains the spatiotemporal feature representation by weighting and fusing the convolution results on both temporal and spatial scales.

[0012] Optionally, the improved GATv2 model is specifically as follows: The spatiotemporal feature representations are input into the graph structure initialization module, and each spatiotemporal feature representation is used as a node to obtain a node set. Calculate the Euclidean distance between every two nodes in the node set; If the Euclidean distance is less than a preset distance threshold, an edge connection is established between the corresponding nodes, and the weight of the edge connection is the reciprocal of the corresponding Euclidean distance. Traverse all node pairs in the node set to obtain the edge connection set; An initial structural graph is obtained by constructing a graph based on the node set and the edge connection set; The initial structure graph is input into the local interaction modeling module. Based on the node features and connection relationships in the initial structure graph, the interaction relationships between nodes are learned through the local neighborhood of the nodes, and the node features are modeled in a fine-grained manner to obtain local interaction features. The local interaction features are input into the relationship optimization module. Based on the local interaction features, the edge relationships between nodes are adjusted to optimize the information propagation process and obtain the optimized relationship features. The optimized relationship features are input into the global information integration module, and the optimized relationship features of all nodes in the entire initial structure graph are fused by weighted average to obtain the global node representation; The global node representation is input into the decision output module. The decision path in the intelligence data analysis process is formulated through the global node characteristics and the initial structure diagram, and a preliminary decision strategy is generated.

[0013] Optionally, the step of adjusting the edge relationships between nodes based on local interaction features to optimize the information propagation process and obtain optimized relationship features specifically involves: Based on local interaction features, the interaction influence values ​​between corresponding nodes are obtained by calculating the Euclidean distance between nodes; Based on the interaction influence values ​​between nodes, the relationships between edges in the initial structure graph are adjusted. Specific steps include: If the interaction impact value between two nodes is greater than the set interaction threshold, the edge weight between the corresponding two nodes will be adjusted to the current interaction impact value. If the interaction value between two nodes is less than or equal to the set interaction threshold, then the connection edge between the two nodes will be deleted. By traversing each pair of nodes in the initial structure graph and optimizing the information propagation process, the optimized relationship features are obtained.

[0014] Optionally, the step of fusing the optimized relationship features of all nodes in the entire initial structural graph through weighted averaging to obtain a global node representation is as follows: Based on the optimized relationship characteristics, a weighted value is assigned to each node; The weighted value is the connectivity degree of the node, which is the number of connections a node has with other nodes in the initial structure graph. The weighted average of the optimized relationship features of all nodes is calculated, and the specific steps include: Multiply the optimized relational features of each node by the corresponding weighted value, and sum over all nodes to obtain the global node representation of the entire initial structure graph; The global node representation includes the relationship information of all nodes in the initial structure graph and the interaction characteristics between nodes.

[0015] Optionally, step seven specifically includes: Based on the aforementioned preliminary decision-making strategy, the decision paths for each stage of the intelligence analysis process are initialized. Each decision path represents an operational process or analytical step in the intelligence analysis, including risk assessment and intelligence reasoning. The performance indicators of each decision path are statistically analyzed to evaluate the execution effect of each decision path, and the weight of each decision path is determined through error backpropagation. The performance indicators include the accuracy of the decision result and the processing time. The strategy optimization objective is set to adjust the choice of decision path to obtain the final decision strategy. The strategy optimization objective includes minimizing the time delay in the decision process, maximizing the inference accuracy, and minimizing the error rate.

[0016] The beneficial effects of this invention are: The deep learning-based intelligent analysis method for multi-source intelligence data provided by this invention has made significant progress in overcoming the limitations of existing multi-source intelligence data processing technologies. Compared with traditional methods, this invention can more efficiently integrate data from different modalities, including text, images, audio, and video data, and automatically extract features through a deep learning model. This multi-modal data fusion not only improves the comprehensiveness of data analysis but also effectively captures complex implicit relationships between different data sources. Especially in the graph construction process, this invention adopts an improved GATv2 model to dynamically adjust the edge relationships between nodes, fully considering the spatiotemporal dependencies and interactions between nodes, enabling the multi-source intelligence graph to more accurately reflect the actual situation when processing complex and dynamically changing data. Unlike traditional static graph structures, this invention introduces a spatiotemporal convolution module and an adaptive time window mechanism during graph construction, enabling dynamic partitioning and modeling of spatiotemporal data at different time scales, thereby effectively capturing short-term and long-term dependencies in the data. This optimization of spatiotemporal feature modeling not only improves the sensitivity to changes in spatiotemporal data but also ensures that the spatiotemporal features of the data can be accurately modeled with fine granularity, enhancing the accuracy and reliability of intelligence analysis.

[0017] This invention overcomes the shortcomings of traditional methods in the decision optimization process by integrating a relational enhancement encoder. Traditional intelligence analysis methods often rely on static decision paths, lacking flexibility and real-time adjustment capabilities. In contrast, the decision optimization module of this invention can dynamically adjust the decision path based on real-time data during intelligence analysis, resulting in more accurate and timely analysis results. This adaptive decision optimization mechanism enables this invention to maintain efficient decision support capabilities in complex and ever-changing intelligence environments. By innovatively combining deep learning, graph neural networks, and spatiotemporal modeling, this invention not only improves the processing capabilities and decision accuracy of multi-source intelligence data but also enhances flexibility and real-time performance in practical applications. This allows the analysis method to effectively cope with changing and complex intelligence data environments, providing more accurate and intelligent decision support, and has broad application prospects. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is an overall flowchart of a deep learning-based intelligent analysis method for multi-source intelligence data proposed in this invention. Figure 2 This is a schematic diagram illustrating the enhanced node feature representation generation steps of a deep learning-based intelligent analysis method for multi-source intelligence data proposed in this invention. Figure 3 This is a flowchart of the improved GATv2 model processing steps for an intelligent analysis method of multi-source intelligence data based on deep learning proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-3 A deep learning-based intelligent analysis method for multi-source intelligence data includes the following steps: Step 1: Collect raw intelligence data from multiple heterogeneous data sources and standardize the raw intelligence data to obtain a multi-source intelligence dataset; Step 2: Based on the multi-source intelligence dataset, obtain a multi-source intelligence feature vector set, and use the feature vector of each data source as a node. Calculate the Pearson correlation coefficient between nodes to measure similarity and construct a multi-source intelligence graph. Step 3: Perform local neighborhood fuzzy reasoning on each node in the multi-source intelligence graph. The local neighborhood fuzzy reasoning is based on the local neighborhood information of each node and fuzzy reasoning rules to obtain node reasoning results; Step 4: Based on the node reasoning results, use a relation enhancement encoder to perform fine-grained encoding on the features of the nodes in the multi-source intelligence graph to obtain enhanced node feature representations; Step 5: Input the enhanced node feature representations into a multi-scale spatio-temporal convolutional module. The multi-scale spatio-temporal convolutional module dynamically partitions time series data through an adaptive time window mechanism and performs fine-grained modeling on spatio-temporal features to generate spatio-temporal feature representations; Step 6: Input the spatio-temporal feature representations into an improved GATv2 model. Through a graph structure initialization module, a local interaction modeling module, a relation optimization module, a global information integration module, and a decision output module, obtain a preliminary decision-making strategy; Step 7: Optimize the preliminary decision-making strategy by adjusting the decision path, optimize the intelligence analysis process, and obtain a final decision-making strategy.

[0022] In this embodiment, the original intelligence data includes text data, image data, audio data, and video data; the standardization processing steps include performing outlier removal, missing value filling, timestamp alignment, and normalization processing on different types of data in the original intelligence data to obtain a multi-source intelligence data set with a unified structure.

[0023] In this embodiment, Step 2 is specifically as follows: Perform word segmentation, stop word removal, and stemming on the text data in the multi-source intelligence data set, and convert the text data into a text feature vector of a fixed length. The text feature vector represents the semantic information of the text. Specifically, the text data is first segmented through natural language processing techniques (such as a tokenizer) into several words or phrases. Then, stop word removal is performed to remove common but irrelevant words for text analysis, such as "de", "shi", "zai", etc., to avoid noise affecting the analysis results. Subsequently, through stemming, the words are converted into their stem forms, such as converting "running" to "run", so as to unify the same words in different forms. Finally, through a pre-trained language model (such as Word2Vec, GloVe, BERT, etc.), each word is mapped to a high-dimensional space to generate a text feature vector of a fixed length. These vectors can effectively represent the semantic information of the text, making subsequent analysis and model training more efficient.

[0024] Audio data from multi-source intelligence datasets is processed using Short-Time Fourier Transform (STFT) and Mel-frequency cepstral coefficients (MFCC) to extract time-domain and frequency-domain features, resulting in an audio feature vector. Specifically, the STFT transforms the audio signal from the time domain to the frequency domain, revealing its frequency components. Then, MFCC is used to extract the time-frequency features of the audio signal, a method commonly used in speech recognition to effectively capture the timbre and rhythm information of the audio signal. After these processing steps, an audio feature vector is obtained, which integrates the frequency-domain features and time-domain information of the audio signal, and can be further used for audio analysis, classification, or recognition tasks.

[0025] Video frame separation is performed on video data from a multi-source intelligence dataset to obtain a set of video frames. Video data, as time-series data, typically consists of consecutive image frames. Video frame separation extracts each frame from the video, forming an independent image set. These image frames contain the spatiotemporal information of the video and are fundamental to analyzing its content. Each video frame can be considered an independent image; by processing these frames, dynamic information and time-series features can be extracted from the video.

[0026] Convolutional Neural Networks (CNNs) are used to extract features from each video frame in the video frame set and the image data in the multi-source intelligence dataset, resulting in video feature vectors and image feature vectors. Specifically, the CNN extracts spatial features from images or video frames through a series of convolutional and pooling layers. These features typically include low-level features such as edges, textures, and shapes, as well as high-level semantic features. For video frames, the features of each frame, after being processed by the CNN, yield a video feature vector reflecting the spatial and temporal characteristics of the video. Similarly, image data also has features extracted using CNNs to obtain image feature vectors, representing the spatial structure information of the image. These feature vectors provide the foundation for subsequent integration of multi-source intelligence data.

[0027] A multi-source intelligence graph is constructed based on a multi-source intelligence feature vector set, which includes text feature vectors, audio feature vectors, video feature vectors, and image feature vectors. By combining feature vectors from different data sources, a comprehensive intelligence graph is formed. Each data source corresponds to a node in the graph through its feature vector, and the relationships between nodes are represented by edges in the graph, reflecting the similarities and dependencies between data sources.

[0028] The multi-source intelligence graph comprises multiple nodes and edges. Each data source uses its corresponding feature vector from the multi-source intelligence feature vector set as a node in the multi-source intelligence graph. Nodes represent various data sources, while edges represent the relationships between nodes based on the similarity between data sources. This graph structure enables more effective data fusion, reasoning, and decision analysis in subsequent steps.

[0029] In the node feature space, the Pearson correlation coefficient is calculated for each pair of nodes to measure similarity. If the Pearson correlation coefficient is less than a preset similarity threshold, an edge connection is established between the corresponding nodes. The weight of the edge connection is the reciprocal of the corresponding Pearson correlation coefficient. The Pearson correlation coefficient measures the linear relationship between two nodes, reflecting their similarity. Through this similarity measurement, a graph reflecting the relationships between multi-source intelligence data can be constructed, where the edge weights represent the strength of similarity between nodes, thus providing a foundation for subsequent graph neural network processing.

[0030] In this embodiment, step three specifically includes: For each node in the multi-source intelligence graph, local neighborhood fuzzy inference is performed. The local neighborhood fuzzy inference steps include: Based on the connectivity of each node in the multi-source intelligence graph, a local neighborhood of a preset size is selected, which is the set of nodes directly connected to the current node. The size of the local neighborhood can be adjusted according to the actual situation, usually determined by a pre-set hyperparameter, with the aim of capturing the direct relationship between a node and its neighboring nodes.

[0031] The local neighborhood of a node is used for reasoning and calculating the similarity between adjacent nodes, ensuring that the information relied upon in the reasoning process can reflect the local environment of the node.

[0032] From the local neighborhood of each node, extract the feature information of the neighboring nodes directly connected to the current node. The features of the neighboring nodes can include their position in the graph, attribute information, credibility, risk assessment value, etc.

[0033] After feature extraction, the obtained neighbor node information will be used as input for reasoning, and will be used for the next step of similarity measurement and fuzzy reasoning rule calculation.

[0034] Based on the characteristics of nodes and neighborhood information, fuzzy inference rules are defined as follows: Strong association rule: If the Euclidean distance between the current node and a node in its local neighborhood is greater than a preset distance threshold, then the two nodes are considered to have a strong association. A strong association means that there is a high degree of similarity between the two nodes, reflecting their close connection in the graph.

[0035] Weak association rule: If the Euclidean distance between the current node and a node in its local neighborhood is less than or equal to a preset distance threshold, then the two nodes are considered to be weakly associated. A weak association indicates that the similarity between the two nodes is low, reflecting that their relationship is relatively loose or distant.

[0036] Based on the established fuzzy inference rules, inference is performed sequentially on each node and its local neighborhood. This process includes:

[0037] Node similarity calculation: First, based on node features and neighborhood information, the similarity between each pair of nodes is calculated, especially by evaluating the relationship between nodes through Euclidean distance or other similarity measures.

[0038] Fuzzy inference execution: Based on the preset fuzzy inference rules (i.e., the criteria for strong and weak associations), determine the strength of the relationship between each pair of nodes, and adjust the weight or credibility of the nodes accordingly.

[0039] After fuzzy inference, defuzzification is performed using the centroid method. The centroid method is a commonly used defuzzification technique that calculates the centroid position of the fuzzy set to obtain the final node inference result. The specific operation is as follows:

[0040] By combining the results of fuzzy inference, the centroid method is used to calculate the inference value of the node, and the position of the node feature in the fuzzy set is converted into a clear numerical value.

[0041] The output of the centroid method will serve as the final inference result for a node, reflecting the node's credibility and risk assessment relative to other nodes in the graph.

[0042] Finally, after centroid defuzzification, the node inference results for each node are obtained, including information such as node credibility and risk assessment. These inference results will provide important basis for subsequent steps (such as relation enhancement coding, spatiotemporal modeling, etc.).

[0043] In this embodiment, step four specifically includes: Based on the node inference results, the feature vector of each node is initialized with the corresponding inference results to obtain the initialized feature vector. Based on the multi-source intelligence graph, an adjacency matrix between nodes is constructed. The adjacency matrix construction step involves each node having an edge connection with each directly connected node, and the weight of the edge is assigned according to the Pearson correlation coefficient between the nodes. The initial feature vector of each node and the feature vectors of neighboring nodes are input into the relation enhancement encoder for encoding. The relation enhancement encoder combines the feature vector of each node with the feature vectors of neighboring nodes in a predetermined manner to obtain the interaction features between the node and its neighboring nodes. Inside the relation enhancement encoder, multiple coding layers are used to progressively encode the interaction features of nodes and neighboring nodes in a fine-grained manner. The fine-grained encoding involves each layer using convolution operations to process the features of nodes and neighboring nodes, and using the ReLU nonlinear activation function to enhance the expressive power of the features. In each coding layer, the node features are enhanced by combining the relationship information of nodes in the multi-source intelligence graph. After processing through multiple coding layers, the enhanced node feature representation of each node is obtained.

[0044] In this embodiment, step five specifically includes: The enhanced node feature representation is input into a multi-scale spatiotemporal convolution module, and the spatiotemporal feature modeling steps performed by the multi-scale spatiotemporal convolution module include: Based on the enhanced node feature representation, an adaptive time window mechanism is used to dynamically divide the time series data. The adaptive time window mechanism dynamically adjusts the size of the time window according to the time characteristics and time series information of the data to capture spatiotemporal changes at different time scales. By using multiple convolutional kernels to perform convolution operations at different temporal and spatial scales, fine-grained modeling of the spatiotemporal features of the data is achieved. Each convolutional kernel performs convolution operations on the data in both the temporal and spatial dimensions, extracting spatiotemporal features at different levels and capturing short-term and long-term spatiotemporal dependencies. The spatiotemporal dependency optimization module is used to optimize spatiotemporal feature modeling. The spatiotemporal dependency optimization module obtains the spatiotemporal feature representation by weighting and fusing the convolution results on both temporal and spatial scales.

[0045] This step, by introducing a multi-scale spatiotemporal convolution module, effectively processes and models the spatiotemporal features in multi-source intelligence data. Through an adaptive time window mechanism, the model dynamically adjusts the size of the time window based on temporal features and time-series information, thereby capturing spatiotemporal changes at different time scales and ensuring accurate segmentation of time-series data. Utilizing multiple convolutional kernels for fine-grained modeling of spatiotemporal data allows for the extraction of rich spatiotemporal features from different time and spatial scales, accurately capturing short-term and long-term dependencies. Furthermore, the spatiotemporal dependency optimization module further optimizes the modeling of spatiotemporal features through weighted fusion of convolutional results, improving the model's performance in spatiotemporal data analysis and significantly enhancing the accuracy of intelligence analysis and decision support capabilities.

[0046] In this embodiment, the improved GATv2 model is specifically as follows: The spatiotemporal feature representations are input into the graph structure initialization module, and each spatiotemporal feature representation is used as a node to obtain a node set. Calculate the Euclidean distance between every two nodes in the node set; If the Euclidean distance is less than a preset distance threshold, an edge connection is established between the corresponding nodes, and the weight of the edge connection is the reciprocal of the corresponding Euclidean distance. Traverse all node pairs in the node set to obtain the edge connection set; An initial structural graph is obtained by constructing a graph based on the node set and the edge connection set; The initial structure graph is input into the local interaction modeling module. Based on the node features and connection relationships in the initial structure graph, the interaction relationships between nodes are learned through the local neighborhood of the nodes, and the node features are modeled in a fine-grained manner to obtain local interaction features. The local interaction features are input into the relationship optimization module. Based on the local interaction features, the edge relationships between nodes are adjusted to optimize the information propagation process and obtain the optimized relationship features. The optimized relationship features are input into the global information integration module, and the optimized relationship features of all nodes in the entire initial structure graph are fused by weighted average to obtain the global node representation; The global node representation is input into the decision output module. The decision path in the intelligence data analysis process is formulated through the global node characteristics and the initial structure diagram, and a preliminary decision strategy is generated.

[0047] This step utilizes a multi-layered graph neural network to process multi-source intelligence data, significantly improving the ability to model spatiotemporal features and express relationships between nodes. Through the graph structure initialization module and the local interaction modeling module, complex interrelationships between nodes can be accurately captured, and the relationship optimization module optimizes information propagation paths, ensuring the efficiency and accuracy of data flow. The global information integration module weightedly fuses node features to generate global node representations, providing more comprehensive data support for the decision output module and ultimately generating optimized decision-making strategies. This process improves the efficiency of intelligence data processing and the accuracy of decision-making, providing more intelligent decision support in complex intelligence analysis.

[0048] In this embodiment, the step of adjusting the edge relationships between nodes based on local interaction features to optimize the information propagation process and obtain optimized relationship features specifically involves: Based on local interaction features, the interaction influence values ​​between corresponding nodes are obtained by calculating the Euclidean distance between nodes; Based on the interaction influence values ​​between nodes, the relationships between edges in the initial structure graph are adjusted. Specific steps include: If the interaction impact value between two nodes is greater than the set interaction threshold, the edge weight between the corresponding two nodes will be adjusted to the current interaction impact value. If the interaction value between two nodes is less than or equal to the set interaction threshold, then the connection edge between the two nodes will be deleted. By traversing each pair of nodes in the initial structure graph and optimizing the information propagation process, the optimized relationship features are obtained.

[0049] This step effectively improves the efficiency and accuracy of information dissemination by optimizing the edge relationships between nodes based on local interaction features. By adjusting edge weights and removing redundant connections, unnecessary information propagation paths are reduced, making information transmission between important nodes more precise. Dynamic calculation of interaction impact values ​​allows the graph structure to reflect the actual relationships between nodes in real time, ensuring more rational and effective interactions. This process not only optimizes the graph structure and information flow but also enhances the model's ability to capture key node relationships, ultimately improving the accuracy of multi-source intelligence data analysis and decision support effectiveness.

[0050] In this embodiment, the step of obtaining a global node representation by weighted averaging and fusing the optimized relationship features of all nodes in the entire initial structure graph is specifically as follows: Based on the optimized relationship characteristics, a weighted value is assigned to each node; The weighted value is the connectivity degree of the node, which is the number of connections a node has with other nodes in the initial structure graph. The weighted average of the optimized relationship features of all nodes is calculated, and the specific steps include: Multiply the optimized relational features of each node by the corresponding weighted value, and sum over all nodes to obtain the global node representation of the entire initial structure graph; The global node representation includes the relationship information of all nodes in the initial structure graph and the interaction characteristics between nodes.

[0051] This step significantly improves the graph's global information integration capability by fusing the optimized relationship features of all nodes in the initial structural graph through weighted averaging. The weight assigned to each node is based on its connectivity, i.e., the number of connections it has with other nodes in the graph. This ensures that nodes with high connectivity have a greater weight in the global features, effectively reflecting the influence of key nodes. Through weighted averaging, the optimized relationship features and interaction information of all nodes are integrated, generating a global node representation that comprehensively reflects the entire graph's relational structure and the interaction patterns between nodes. This method improves the graph's expressive power, more accurately capturing complex relational information and providing a more accurate global perspective for subsequent decision optimization.

[0052] In this embodiment, step seven specifically includes: Based on the aforementioned preliminary decision-making strategy, the decision paths for each stage of the intelligence analysis process are initialized. Each decision path represents an operational process or analytical step in the intelligence analysis, including risk assessment and intelligence reasoning. The performance indicators of each decision path are statistically analyzed to evaluate the execution effect of each decision path, and the weight of each decision path is determined through error backpropagation. The performance indicators include the accuracy of the decision result and the processing time. The strategy optimization objective is set to adjust the choice of decision path to obtain the final decision strategy. The strategy optimization objective includes minimizing the time delay in the decision process, maximizing the inference accuracy, and minimizing the error rate.

[0053] This step significantly improves decision-making efficiency and accuracy by optimizing the decision-making path in the intelligence analysis process. By initializing and evaluating the performance of each decision path, critical paths can be accurately identified and strengthened, effectively improving the execution of the decision-making process. The error backpropagation mechanism ensures dynamic optimization of the paths, adjusting path weights based on accuracy and processing time, so that important decision paths are prioritized. Simultaneously, by setting strategy optimization objectives, the time delay, inference accuracy, and error rate in the decision-making process are optimized, ultimately yielding the optimal decision strategy. This method enhances the intelligence level of intelligence analysis, providing more efficient, accurate, and real-time decision support.

[0054] Example 1: To verify the feasibility of this invention in practice, it was applied to an intelligent security monitoring system. The system comprehensively analyzes video, audio, and sensor data collected in real-time by multiple surveillance cameras to determine the presence of security risks or abnormal events. The monitoring equipment involved in this scenario includes video surveillance cameras, environmental sensors (such as temperature, humidity, and motion sensors), and microphones. The system aims to identify and handle potential security threats in real time, such as intrusions, fires, or equipment malfunctions.

[0055] In this application scenario, traditional security systems typically rely on a single type of data source for analysis, such as video surveillance data or sensor data alone. The main drawback of this approach is its limited information processing capabilities and susceptibility to anomalies in a single data source, leading to missed or false alarms in security incidents. To overcome this problem, the system employs a deep learning-based multi-source intelligence data intelligent analysis method. This method integrates information from different data sources (such as video, audio, and sensor data), comprehensively analyzing information from multiple data sources to improve the accuracy and response speed of event detection.

[0056] In this embodiment, the system first collects raw data from multiple surveillance cameras and sensors, including real-time video streams, environmental data (temperature, humidity), and audio data. Then, the system standardizes this data to eliminate differences between different data sources, resulting in a unified format multi-source intelligence dataset. Next, a pre-trained convolutional neural network (CNN) is used to extract features from the video data, natural language processing (NLP) is used for word segmentation and semantic analysis of the audio data, and Fourier transform is used to extract features from the sensor data, converting these data into corresponding feature vectors to construct a multi-source intelligence graph. After constructing the multi-source intelligence graph, the system performs local neighborhood fuzzy inference on each node in the graph to further evaluate the node's credibility and potential risk. Based on the node's similarity and distance, the inference results are used to enhance the node's feature representation, and then a spatiotemporal convolutional model is used to model the time-series data, capturing short-term and long-term spatiotemporal dependencies in the data. Through an improved GATv2 model, the system optimizes the information propagation process based on the relationships between nodes and generates preliminary decision-making strategies. Based on the initial decision-making strategy, the system dynamically adjusts the decision-making path, optimizes the weight of each decision-making path through error backpropagation, and finally arrives at the optimal decision-making strategy, generates security warning information, and notifies security personnel to take appropriate emergency measures.

[0057] In this embodiment, the system can process and analyze data from multiple monitoring devices in real time, enabling more accurate identification of abnormal events. For example, when a video camera detects an intruder, an audio sensor can detect abnormal sounds (such as the sound of a window breaking), while an environmental sensor detects a temperature change. Based on the fusion of these multi-source data, the system reacts quickly, issuing a security alarm and initiating the corresponding emergency response procedures. To verify the effectiveness of this invention, a comparative experimental data table is provided below, showing the performance comparison between the method of this invention and traditional methods (based on a single data source analysis method) in handling different security events:

[0058] Table 1 Comparison of Security Incident Detection Accuracy and Response Time

[0059] As can be seen from the data in Table 1, the accuracy of traditional methods in detecting multiple security events is generally lower than that of the method of this invention, especially in critical events such as intrusion detection and fire alarms, where the accuracy of traditional methods is significantly lower, while the method of this invention can significantly improve the detection accuracy. More significantly, the method of this invention also has a clear advantage in response time, significantly shortening the time from event occurrence to response, particularly in fire alarm and intrusion detection scenarios, where the response time is reduced by more than 50%. These results demonstrate that this invention, through multi-source data fusion and deep learning model optimization, can effectively improve the detection accuracy and real-time performance of security systems, providing strong support for timely response and handling of security events. This embodiment demonstrates the advantages of the method of this invention in practical applications, especially in terms of accuracy, real-time performance, and decision optimization in multi-source intelligence data analysis. Compared with traditional methods, this invention can more effectively integrate multimodal data, improve the intelligence level of security monitoring systems, and enhance the system's ability to handle complex situations.

[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1.A deep learning-based multi-source intelligence data intelligent analysis method, characterized in that, The method comprises the following steps: Step 1: Collecting original intelligence data from multiple heterogeneous data sources, and performing standardization processing on the original intelligence data to obtain a multi-source intelligence data set; Step 2: Obtaining a multi-source intelligence feature vector set based on the multi-source intelligence data set, taking the feature vector of each data source as a node, calculating the Pearson correlation coefficient between the nodes for similarity measurement, and constructing a multi-source intelligence graph; Step 3: Performing local neighborhood fuzzy reasoning on each node in the multi-source intelligence graph, obtaining a node reasoning result based on the local neighborhood information and fuzzy reasoning rules of each node; Step 4: Based on the node reasoning result, performing fine-grained coding on the features of the nodes in the multi-source intelligence graph through a relationship enhancement encoder to obtain enhanced node feature representations; Step 5: Inputting the enhanced node feature representations into a multi-scale spatio-temporal convolution module, which dynamically divides the time series data through an adaptive time window mechanism and performs fine-grained modeling on the spatio-temporal features to generate spatio-temporal feature representations; Step 6: Inputting the spatio-temporal feature representations into an improved GATv2 model through a graph structure initialization module, a local interaction modeling module, a relationship optimization module, a global information integration module, and a decision output module to obtain a preliminary decision strategy; Step 7: Optimizing the preliminary decision strategy by adjusting the decision path to optimize the intelligence analysis process and obtain a final decision strategy. 2.The deep learning-based multi-source intelligence data intelligent analysis method according to claim 1, characterized in that, The original intelligence data includes text data, image data, audio data, and video data; the standardization processing step includes outlier rejection, missing value filling, timestamp alignment, and normalization processing on different types of data in the original intelligence data to obtain a multi-source intelligence data set with a unified structure. 3.The deep learning-based multi-source intelligence data intelligent analysis method of claim 1, wherein, The step 2 is specifically: Performing word segmentation, stop word removal, and stemming processing on the text data in the multi-source intelligence data set to convert the text data into fixed-length text feature vectors, which represent the semantic information of the text; Extracting the time domain and frequency domain features of the audio data in the multi-source intelligence data set through short-time Fourier transform and mel frequency cepstrum coefficients to obtain audio feature vectors; Separating the video frames in the video data in the multi-source intelligence data set to obtain a video frame set; Extracting features from each video frame in the video frame set and the image data in the multi-source intelligence data set through a convolutional neural network to obtain video feature vectors and image feature vectors; Based on the multi-source intelligence feature vector set, a multi-source intelligence graph is constructed, which includes text feature vectors, audio feature vectors, video feature vectors, and image feature vectors; The multi-source intelligence graph includes multiple nodes and edges, and each data source is represented as a node in the multi-source intelligence graph through the corresponding feature vector in the multi-source intelligence feature vector set; In the node feature space, the Pearson correlation coefficient between each pair of nodes is calculated for similarity measurement, and if the Pearson correlation coefficient is less than a preset similarity threshold, an edge connection is established between the corresponding nodes, and the weight of the edge connection is the inverse of the corresponding Pearson correlation coefficient. 4.The deep learning-based multi-source intelligence data intelligent analysis method of claim 1, wherein, The step 3 is specifically: locally adjacent fuzzy reasoning is performed on each node in the multi-source information graph, the locally adjacent fuzzy reasoning step comprising: a local neighborhood of a preset size is selected according to the connection relationship of each node in the multi-source information graph, the local neighborhood being a node set directly connected to the current node; feature information of adjacent nodes is extracted from the local neighborhood of each node as input for reasoning; a fuzzy reasoning rule is set according to the features and neighborhood information of the nodes, the fuzzy reasoning rule being: if the Euclidean distance between nodes in the local neighborhood of the current node and is greater than a preset distance threshold, then the two corresponding nodes are strongly associated; if the Euclidean distance between nodes in the local neighborhood of the current node and is less than or equal to a preset distance threshold, then the two corresponding nodes are weakly associated; based on the fuzzy reasoning rule, reasoning is performed on each node and its local neighborhood, and de-fuzzification is performed by the centroid method to obtain a node reasoning result for each node. 5.The deep learning based multi-source intelligence data intelligent analysis method according to claim 1, wherein, Step four is specifically: based on the node reasoning result, the feature vector of each node is initialized as the corresponding reasoning result to obtain an initialized feature vector; an adjacent relationship matrix between nodes is constructed according to the multi-source information graph, the adjacent relationship matrix construction step being that each node and the directly connected node have an edge connection, and the weight of the edge is assigned according to the Pearson correlation coefficient between the nodes; the initialized feature vector of each node and the feature vector of the adjacent node are input into a relationship-enhanced encoder for encoding, the relationship-enhanced encoder combining the feature vector of each node with the feature vector of the adjacent node in a predetermined manner to obtain the interaction features between the node and the adjacent node; within the relationship-enhanced encoder, multiple encoding layers are used to gradually perform fine-grained encoding on the interaction features of the node and the adjacent node, the fine-grained encoding being that each layer uses convolution operations to process the features of the node and the adjacent node, and a ReLU nonlinear activation function is used to enhance the expression ability of the features; in each encoding layer, the node features are enhanced in combination with the relationship information of the nodes in the multi-source information graph, and after processing by multiple encoding layers, an enhanced node feature representation of each node is obtained. 6.The deep learning based multi-source intelligence data intelligent analysis method according to claim 1, wherein, Step five is specifically: the enhanced node feature representation is input into a multi-scale spatio-temporal convolution module, and the spatio-temporal feature modeling step of the multi-scale spatio-temporal convolution module comprises: based on the enhanced node feature representation, an adaptive time window mechanism is used to dynamically divide the time series data, the adaptive time window mechanism dynamically adjusting the size of the time window according to the time features and time sequence information of the data to capture spatio-temporal changes at different time scales; by using multiple convolution kernels, convolution operations are performed for different time and spatial scales to perform fine-grained modeling of the spatio-temporal features of the data, each convolution kernel performing convolution operations on the data in the time and spatial dimensions to extract spatio-temporal features at different levels and capture short-term and long-term spatio-temporal dependencies; a spatio-temporal dependency optimization module is used to optimize the spatio-temporal feature modeling, the spatio-temporal dependency optimization module weighting and fusing the convolution results in the time and spatial scales to obtain a spatio-temporal feature representation. 7.The deep learning based multi-source intelligence data intelligent analysis method of claim 1, wherein, The improved GATv2 model is specifically: Input the spatio-temporal feature representation into a graph structure initialization module, taking each spatio-temporal feature representation as a node to obtain a node set; Calculate the Euclidean distance between each two nodes in the node set; If the Euclidean distance is less than a preset distance threshold, an edge connection is established between the corresponding nodes, and the weight of the edge connection is the inverse of the corresponding Euclidean distance; Iterate through all node pairs in the node set to obtain an edge connection set; Based on the node set and the edge connection set, an initial structure graph is constructed; Input the initial structure graph into a local interaction modeling module, learn the interaction relationship between nodes through the local neighborhood of the nodes based on the node features and connection relationships in the initial structure graph, and perform fine-grained modeling on the node features to obtain local interaction features; Input the local interaction features into a relationship optimization module, adjust the edge relationship between nodes based on the local interaction features, and optimize the information propagation process to obtain optimized relationship features; Input the optimized relationship features into a global information integration module, and obtain global node representations by weighted average fusion of the optimized relationship features of all nodes in the entire initial structure graph; Input the global node representations into a decision output module, and generate a preliminary decision strategy by formulating a decision path in the intelligence data analysis process based on the global node features and the initial structure graph. 8.The deep learning based multi-source intelligence data intelligent analysis method of claim 1, wherein, The step of adjusting the edge relationship between nodes based on the local interaction features to optimize the information propagation process to obtain the optimized relationship features is specifically: Based on the local interaction features, the Euclidean distance between nodes is calculated to obtain the interaction influence value between the corresponding nodes; Based on the interaction influence value between nodes, the relationship of the edges in the initial structure graph is adjusted, and the specific steps include: If the interaction influence value between two nodes is greater than a set interaction threshold, the edge weight between the corresponding two nodes is adjusted to the current interaction influence value; If the interaction value between two nodes is less than or equal to the set interaction threshold, the connection edge between the corresponding two nodes is deleted; Iterate through each pair of nodes in the initial structure graph to optimize the information propagation process to obtain the optimized relationship features. 9.The deep learning based multi-source intelligence data intelligent analysis method of claim 1, wherein, The step of obtaining the global node representations by weighted average fusion of the optimized relationship features of all nodes in the entire initial structure graph is specifically: Based on the optimized relationship features, a weighting value is assigned to each node; The weighting value is the connectivity of the node, and the connectivity is the number of connections of the node with other nodes in the initial structure graph; The weighted average of the optimized relationship features of all nodes includes the following steps: Multiply the optimized relationship features of each node by the corresponding weighting value, and sum all nodes to obtain the global node representations of the entire initial structure graph; The global node representations include the relationship information of all nodes in the initial structure graph and the interaction features between nodes. 10.The deep learning based multi-source intelligence data intelligent analysis method according to claim 1, wherein, The step seven is specifically: Based on the preliminary decision strategy, initialize the decision paths of each stage in the intelligence analysis process, and each decision path represents an operation process or analysis step in the intelligence analysis, including risk assessment and intelligence reasoning; The performance index of each decision path is counted to evaluate the execution effect of each decision path, and the weight of each decision path is determined through error back propagation, and the performance index includes the accuracy of the decision result and the processing time; A strategy optimization target is set to adjust the selection of the decision path, and a final decision strategy is obtained, and the strategy optimization target includes minimizing the time delay in the decision process, maximizing the inference accuracy, and minimizing the error rate.

Images