Intelligent perception method and system for overseas public opinion security situation based on big data fusion

Abstract

The application provides an overseas public opinion security situation intelligent perception method and system fusing big data, relates to the technical field of public opinion analysis, and comprises the following steps: obtaining overseas public opinion data and extracting multi-dimensional features, adopting a bidirectional cross attention mechanism to construct an adaptive weight matrix, generating a fusion feature matrix through multi-head time sequence convolution and a dynamic graph structure attention mechanism, constructing an evaluation index system, determining weights based on information entropy and evidence theory, determining an early warning level based on multi-criteria decision, and generating management and control suggestions and early warning information.

Description

Technical Field

[0001] The present invention relates to the technical field of public opinion analysis, and in particular to a method and system for intelligently perceiving overseas public opinion security situations by integrating big data. Background Art

[0002] With the rapid development of the Internet and social media, overseas public opinion information has shown big data characteristics. Traditional public opinion analysis methods are no longer able to cope with massive, diverse, and real-time public opinion data.

[0003] Currently, overseas public opinion security situation awareness relies primarily on manual analysis and simple statistical models, lacking the ability to deeply mine and intelligently analyze massive amounts of data. Existing technologies still suffer from incomplete and in-depth feature extraction, making it difficult to capture subtle changes and potential risks in the evolution of public opinion. There is a lack of effective modeling of the spatiotemporal correlations of public opinion data, making it impossible to accurately grasp the patterns of public opinion dissemination and the scope of its influence. Furthermore, the public opinion security situation assessment indicator system is incomplete, making it difficult to fully reflect key factors such as the spread, influence, and sensitivity of public opinion.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for intelligently perceiving overseas public opinion security situations by integrating big data, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for intelligently sensing overseas public opinion security trends by integrating big data, including:

[0007] Obtain public opinion data from overseas social media platforms and preprocess it to obtain standardized public opinion information and perform multi-dimensional feature extraction to construct a feature dataset;

[0008] A bidirectional cross-attention mechanism is used to calculate fine-grained correlations of feature vectors in the feature dataset and construct an adaptive weight matrix. The feature vectors in the feature dataset are weighted based on the adaptive weight matrix to obtain weighted feature vectors, which are then input into a multi-head temporal convolution unit with residual connections to extract temporal correlation features. The spatial dependency is obtained through a dynamic graph structure attention mechanism to generate a fused feature matrix.

[0009] Based on the fusion feature matrix, an evaluation index system is constructed, including a public opinion dissemination power index, a public opinion influence index, and a public opinion sensitivity index;

[0010] Based on information entropy and evidence theory, the information gain of indicators is calculated to determine the confidence interval, and the indicator weight vector is generated and combined with the quantitative score of the evaluation indicator. The multi-criteria decision-making method is used to determine the warning level of the public opinion security situation;

[0011] According to the public opinion security situation warning level, public opinion control suggestions and warning information are generated and pushed to the management terminal.

[0012] In an optional embodiment,

[0013] Obtain public opinion data from overseas social media platforms and preprocess it to obtain standardized public opinion information and perform multi-dimensional feature extraction. The feature dataset constructed includes:

[0014] Collect public opinion data from overseas social media platforms, including text data, image data, and video data posted by users;

[0015] Performing multilingual unified encoding on text data through cross-language aligned word vector spaces, extracting key information from image and video data using deep learning-based object detection methods, and integrating the multilingual unified encoding and key information to generate standardized public opinion information;

[0016] A hierarchical feature extraction network is used to extract features from the normalized public opinion information, wherein local features are extracted at the bottom layer, contextual relationship features are extracted at the middle layer, and global semantic features are extracted at the top layer. The extracted features are fused through jump connections to obtain a feature data set.

[0017] In an optional embodiment,

[0018] Using a bidirectional cross-attention mechanism to calculate fine-grained correlations on feature vectors in the feature dataset and construct an adaptive weight matrix includes:

[0019] Receiving feature vectors in a feature data set, mapping the feature vectors to generate a query vector, a key vector, and a value vector through different transformation matrices, wherein the transformation matrix is ​​a learnable parameter matrix;

[0020] Based on the bidirectional cross attention mechanism, the query vector and the key vector are matrix multiplied and scaled by a scaling factor before performing an exponential operation. The result of the exponential operation is divided by the sum of the exponentials corresponding to all key vectors to obtain a forward attention score. The transposed query vector and the transposed key vector are subjected to the same operation to obtain a reverse attention score.

[0021] The forward attention score and the reverse attention score are weightedly fused based on a learnable fusion weight parameter to obtain a comprehensive correlation score, and the comprehensive correlation score is input into a sigmoid activation function to obtain a normalized fine-grained correlation matrix;

[0022] The balance parameter, the fine-grained correlation matrix and the identity matrix are weighted and combined and then input into the softmax function to generate an adaptive weight matrix, wherein the adaptive weight matrix is ​​used to characterize the dynamic correlation strength between the eigenvectors, the eigenvector with the highest correlation corresponds to the largest weight value, and the eigenvector with the lowest correlation corresponds to the smallest weight value.

[0023] In an optional embodiment,

[0024] The weighted feature vector is obtained and input into a multi-head temporal convolution unit with residual connection to extract temporal correlation features. The spatial dependency is obtained through the dynamic graph structure attention mechanism to generate a fusion feature matrix including:

[0025] Inputting the weighted feature vector into a multi-head temporal convolution unit with residual connection, the multi-head temporal convolution unit performs channel segmentation on the weighted feature vector to obtain multiple feature sub-vectors, and inputting the feature sub-vectors into a convolution layer group with multiple one-dimensional convolution kernels of different scales to obtain a first convolution feature, wherein each one-dimensional convolution layer has an independent learnable weight matrix;

[0026] Adaptively weighting the first convolutional features to obtain temporal features, performing adaptive residual connection on the temporal features and corresponding feature subvectors, and performing layer normalization processing with learnable scaling parameters to obtain head output features, and concatenating multiple head output features in the channel dimension and performing nonlinear feature transformation to obtain multi-head fusion features;

[0027] Constructing a dynamic graph structure unit to perform a double nonlinear transformation on the multi-head fusion features to obtain a feature correlation matrix, constructing a dynamic adjacency matrix based on the feature correlation matrix, and performing spatial information aggregation on the multi-head fusion features based on the dynamic adjacency matrix to obtain spatial dependency features;

[0028] The multi-head fusion feature and the spatial dependency feature are input into an adaptive gated fusion unit to generate a fusion feature, the fusion feature is adaptively calibrated in terms of channel importance through a compression excitation module to obtain a calibration feature, the calibration feature is adaptively weighted residual connected with the weighted feature vector and subjected to layer normalization processing to generate a fusion feature matrix.

[0029] In an optional embodiment,

[0030] The evaluation index system based on the fusion feature matrix includes the public opinion communication power index, the public opinion influence index, and the public opinion sensitivity index, which includes:

[0031] The fusion feature matrix is ​​mapped to the public opinion communication power feature space through a multi-layer perceptron, and the public opinion communication power index is calculated based on the importance weight of the communication node and the length of the communication path;

[0032] Extracting time series features from the fusion feature matrix and combining them with user influence weights to construct a communication influence network that takes into account user interaction intensity, and calculating node influence scores to obtain a public opinion influence index;

[0033] Extracting text semantic features based on the fusion feature matrix, combining the preset sensitive dictionary and topic word vectors, and calculating the text sensitivity score through the attention mechanism to obtain the public opinion sensitivity index;

[0034] The evaluation index system is obtained by combining the public opinion communication power index, the public opinion influence index and the public opinion sensitivity index.

[0035] In an optional embodiment,

[0036] Based on information entropy and evidence theory, the information gain of indicators is calculated to determine the confidence interval, and the indicator weight vector is generated. The indicator weight vector is combined with the quantitative score of the evaluation indicator, and the multi-criteria decision-making method is used to determine the public opinion security situation warning level, including:

[0037] Collect sample data of the evaluation indicators and perform normalization processing to obtain a sample probability distribution, calculate the information entropy of the sample probability distribution, determine the information gain of each evaluation indicator based on the information entropy, construct an identification framework for the evaluation indicators, calculate the basic probability distribution value of each evaluation indicator based on the information gain, calculate the subset confidence function and the intersection likelihood function for the basic probability distribution value to obtain the confidence interval of the evaluation indicator;

[0038] Based on the confidence interval, the basic probability distribution values ​​of different evaluation indicators are time-series decomposed and long-range correlation features are extracted, a time-varying state transfer matrix and a dynamic evolution matrix are constructed, a dynamic conflict degree is calculated through singular value decomposition, an evidence conflict matrix is ​​established and a similarity matrix is ​​determined, the similarity matrix is ​​normalized to obtain a credibility score of the evaluation indicator, and the credibility score is weightedly combined with the basic probability distribution value to generate an indicator weight vector;

[0039] Construct a quantitative score matrix for evaluation indicators, perform weighted normalization operation on the indicator weight vector and the quantitative score matrix to obtain a decision matrix, extract the maximum and minimum values ​​of each evaluation indicator in the decision matrix based on a multi-criteria decision-making method to construct a positive ideal solution vector and a negative ideal solution vector, calculate the Euclidean distance from the decision matrix to the positive ideal solution vector and the negative ideal solution vector, calculate the relative proximity score based on the Euclidean distance, perform fuzzy clustering operation on the relative proximity score to divide the warning level interval, and determine the public opinion security situation warning level based on the warning level interval.

[0040] In an optional embodiment,

[0041] The basic probability distribution values ​​of different evaluation indicators are decomposed in time series and long-range correlation features are extracted. The time-varying state transfer matrix and dynamic evolution matrix are constructed. The dynamic conflict degree is calculated by singular value decomposition. The evidence conflict matrix is ​​established and the similarity matrix is ​​determined.

[0042] For each evaluation indicator, a fractional-order difference coefficient matrix is ​​constructed based on the gamma function, the fractional-order difference coefficient matrix is ​​convolved with the basic probability distribution value of the evaluation indicator to obtain a fractional-order difference sequence, and the fractional-order difference sequence is decomposed according to the time scale to obtain a multi-scale feature vector;

[0043] Performing autocorrelation analysis on the multi-scale feature vectors to extract long-range correlation features, inputting the long-range correlation features into a recursive neural network to extract time-dependent features, organizing the time-dependent features into a state vector, and using a least squares method to solve a time-varying state transfer matrix based on the transfer relationship between the state vectors at adjacent moments;

[0044] Performing a matrix multiplication operation on the time-varying state transfer matrix and the multi-scale eigenvector to obtain an evidence dynamic evolution matrix, performing singular value decomposition on the evidence dynamic evolution matrix to extract principal component features, and calculating the empty set probability between evaluation indicators based on the principal component features;

[0045] An evidence conflict matrix is ​​constructed by taking the empty set probability values ​​between each two evaluation indicators as matrix elements, wherein the empty set probability values ​​are obtained by the intersection operation of the basic probability distribution functions of the evaluation indicators. An adaptive weight network is constructed based on the evidence conflict matrix to obtain the evidence credibility weights. The basic probability distribution values ​​of the evaluation indicators are dynamically weighted and fused according to the evidence credibility weights to obtain a similarity matrix.

[0046] In an optional embodiment,

[0047] Generating public opinion control suggestions and warning information based on the public opinion security situation warning level and pushing them to the management terminal includes:

[0048] Based on the public opinion security situation warning level, a preset control rule tree is queried, and corresponding public opinion control suggestions are matched according to the path of the warning level in the control rule tree;

[0049] Generate warning information based on the warning level and the public opinion control suggestion, and push the warning information to the management terminal.

[0050] A second aspect of an embodiment of the present invention provides an intelligent perception system for overseas public opinion and security situation integrating big data, including:

[0051] The first unit is used to obtain public opinion data from overseas social media platforms and perform preprocessing to obtain standardized public opinion information and perform multi-dimensional feature extraction to construct a feature dataset;

[0052] The second unit is used to calculate the fine-grained correlation of the feature vectors in the feature data set using a bidirectional cross-attention mechanism and construct an adaptive weight matrix. The feature vectors in the feature data set are weighted based on the adaptive weight matrix to obtain weighted feature vectors and input them into a multi-head temporal convolution unit with residual connection to extract temporal correlation features. The spatial dependency is obtained through a dynamic graph structure attention mechanism to generate a fusion feature matrix.

[0053] The third unit is used to construct an evaluation index system including a public opinion dissemination power index, a public opinion influence index, and a public opinion sensitivity index based on the fusion feature matrix;

[0054] The fourth unit is used to calculate the information gain of indicators based on information entropy and evidence theory to determine the confidence interval, generate the indicator weight vector, and combine the indicator weight vector with the quantitative score of the evaluation indicator to determine the public opinion security situation warning level using a multi-criteria decision-making method;

[0055] The fifth unit is used to generate public opinion control suggestions and warning information based on the public opinion security situation warning level and push them to the management terminal.

[0056] According to a third aspect of the embodiments of the present invention,

[0057] An electronic device is provided, comprising:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] In the present invention, by adopting a bidirectional cross-attention mechanism and a multi-head temporal convolution unit combined with a dynamic graph structure attention mechanism, the spatiotemporal correlation characteristics of public opinion data can be effectively captured, the accuracy and comprehensiveness of feature extraction are improved, and a reliable data basis is provided for subsequent public opinion situation evaluation. A multidimensional evaluation index system including public opinion dissemination power, influence and sensitivity is constructed, and the index weights are calculated based on information entropy and evidence theory, thereby achieving a comprehensive and objective evaluation of the public opinion security situation, improving the accuracy and credibility of the early warning results, and providing timely and effective decision-making support for decision makers by generating public opinion control suggestions and early warning information and pushing them to the management terminal, which helps to quickly identify and respond to potential public opinion security risks and improve the efficiency and effectiveness of overseas public opinion management. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of an intelligent perception method for overseas public opinion security situation integrating big data according to an embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram of the multi-head time series convolution unit processing of the intelligent perception method of overseas public opinion security situation integrating big data according to an embodiment of the present invention;

[0064] Figure 3 This is a heat matrix diagram of the early warning accuracy of the intelligent perception method for overseas public opinion security situation integrating big data in an embodiment of the present invention;

[0065] Figure 4 This is a comparison chart of the changes in prediction accuracy of the intelligent perception method for overseas public opinion security situation that integrates big data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0067] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0068] Figure 1 This is a flow chart of an intelligent perception method for overseas public opinion security situation integrating big data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0069] Obtain public opinion data from overseas social media platforms and preprocess it to obtain standardized public opinion information and perform multi-dimensional feature extraction to construct a feature dataset;

[0070] A bidirectional cross-attention mechanism is used to calculate fine-grained correlations of feature vectors in the feature dataset and construct an adaptive weight matrix. The feature vectors in the feature dataset are weighted based on the adaptive weight matrix to obtain weighted feature vectors, which are then input into a multi-head temporal convolution unit with residual connections to extract temporal correlation features. The spatial dependency is obtained through a dynamic graph structure attention mechanism to generate a fused feature matrix.

[0071] Based on the fusion feature matrix, an evaluation index system is constructed, including a public opinion dissemination power index, a public opinion influence index, and a public opinion sensitivity index;

[0072] Based on information entropy and evidence theory, the information gain of indicators is calculated to determine the confidence interval, and the indicator weight vector is generated and combined with the quantitative score of the evaluation indicator. The multi-criteria decision-making method is used to determine the warning level of the public opinion security situation;

[0073] According to the public opinion security situation warning level, public opinion control suggestions and warning information are generated and pushed to the management terminal.

[0074] In an optional embodiment,

[0075] Obtain public opinion data from overseas social media platforms and preprocess it to obtain standardized public opinion information and perform multi-dimensional feature extraction. The feature dataset constructed includes:

[0076] Collect public opinion data from overseas social media platforms, including text data, image data, and video data posted by users;

[0077] Performing multilingual unified encoding on text data through cross-language aligned word vector spaces, extracting key information from image and video data using deep learning-based object detection methods, and integrating the multilingual unified encoding and key information to generate standardized public opinion information;

[0078] A hierarchical feature extraction network is used to extract features from the normalized public opinion information, wherein local features are extracted at the bottom layer, contextual relationship features are extracted at the middle layer, and global semantic features are extracted at the top layer. The extracted features are fused through jump connections to obtain a feature data set.

[0079] A data collection module was established to acquire data by configuring data interface parameters. The collected data included text, images, and videos posted by users on overseas social platforms. During the collection process, connections were established to the data source, the collection task execution cycle was set, and the collected raw data was initially cleaned and formatted. The data's timeliness and integrity were verified, and duplicate and invalid data records were removed. The cleaned data was then classified and stored according to the pre-set storage format.

[0080] The collected multimodal data is uniformly preprocessed and standardized. In the text data processing phase, the language type of multilingual text is identified, and the corresponding word segmentation algorithm is selected for processing based on the characteristics of different languages. The text after word segmentation is standardized, including removing interference information such as stop words and special characters, and the processed text is vectorized using a pre-trained multilingual model. To achieve cross-language semantic alignment, an adversarial training method is used to construct a cross-language word vector alignment module, mapping texts in different languages ​​into the same semantic space, so that texts with similar semantics have similar vector representations.

[0081] Image data is preprocessed, including size normalization and pixel value standardization. A deep learning object detection model is used to identify and locate key objects in the image. This object detection model employs a multi-scale feature extraction strategy, effectively identifying objects of varying sizes and locations within the image. For each detected object, the corresponding category information, location information, and related visual features are extracted and integrated into a unified feature vector.

[0082] Video data processing requires keyframe extraction. By calculating inter-frame differences, representative keyframes are selected and the same feature extraction process as image processing is applied to these keyframes. Simultaneously, temporal features are extracted to capture the dynamic changes in video content over time. Frame-level features and temporal features are fused to obtain a complete video feature representation.

[0083] After obtaining the feature representations of each modality, a unified feature fusion framework is established to map the features of different modalities into the same feature space. By designing a feature alignment module, semantic alignment between features of different modalities can be achieved. Using an attention mechanism, the features of different modalities are adaptively weighted to generate standardized public opinion information in a unified format.

[0084] For standardized public opinion information, a hierarchical feature extraction network is used for further feature learning. At the bottom-level feature extraction stage, we focus on extracting local features, using convolutional neural networks to capture local patterns and detailed features of the data. At the middle-level feature extraction stage, we analyze the contextual relationships between features and capture spatial dependencies between regions by constructing feature graph pyramids and feature association graphs. At the top-level feature extraction stage, we use an attention mechanism to model global semantic relationships and extract long-range dependency features.

[0085] By designing a skip connection structure, features extracted from different layers are fused. During the feature fusion process, a feature fusion weight matrix is ​​established to adaptively weight features at different layers. Through multi-scale feature alignment and residual connections, useful information at each layer is effectively retained, ultimately generating a feature dataset containing rich semantic information.

[0086] For example, assuming that we need to monitor overseas public opinion related to a certain electronic product, the system will first collect relevant multimodal data:

[0087] In terms of text, we collected user reviews in multiple languages, such as "The performance of this new product is excellent" and "The performance is outstanding". These texts in different languages ​​are uniformly mapped into the vector space to obtain feature vectors.

[0088] For product-related images, object detection is used to identify key elements such as the product itself and usage scenarios, extract location coordinates and category information, and generate feature vectors [0.82, 0.45, ..., 0.63].

[0089] For video content, key frames are extracted and temporal features are analyzed, and the feature vector [0.73, 0.28, ..., 0.91] is obtained by integration.

[0090] Through hierarchical feature extraction, the system captures specific features such as product appearance and price at the bottom layer, analyzes the emotional tendencies of user reviews at the middle layer, and grasps the overall market response at the top layer. The resulting feature data set contains complete information such as time, multimodal features, emotional scores, and influence scores.

[0091] In this embodiment, the attention mechanism and feature alignment module adopted realize the effective fusion of different modal features, can adaptively adjust the importance weights of different modal features, highlight the role of key information, and maintain the complementarity of each modal feature. The unified feature representation framework provides a standardized data basis for subsequent public opinion analysis. The design of the hierarchical feature extraction network enables the system to understand public opinion information from different abstract levels. The feature fusion mechanism realized by the jump connection structure effectively solves the gradient disappearance problem in the deep network and ensures the full utilization of the feature information of each layer.

[0092] In an optional embodiment,

[0093] Using a bidirectional cross-attention mechanism to calculate fine-grained correlations on feature vectors in the feature dataset and construct an adaptive weight matrix includes:

[0094] Receiving feature vectors in a feature data set, mapping the feature vectors to generate a query vector, a key vector, and a value vector through different transformation matrices, wherein the transformation matrix is ​​a learnable parameter matrix;

[0095] Based on the bidirectional cross attention mechanism, the query vector and the key vector are matrix multiplied and scaled by a scaling factor before performing an exponential operation. The result of the exponential operation is divided by the sum of the exponentials corresponding to all key vectors to obtain a forward attention score. The transposed query vector and the transposed key vector are subjected to the same operation to obtain a reverse attention score.

[0096] The forward attention score and the reverse attention score are weightedly fused based on a learnable fusion weight parameter to obtain a comprehensive correlation score, and the comprehensive correlation score is input into a sigmoid activation function to obtain a normalized fine-grained correlation matrix;

[0097] The balance parameter, the fine-grained correlation matrix and the identity matrix are weighted and combined and then input into the softmax function to generate an adaptive weight matrix, wherein the adaptive weight matrix is ​​used to characterize the dynamic correlation strength between the eigenvectors, the eigenvector with the highest correlation corresponds to the largest weight value, and the eigenvector with the lowest correlation corresponds to the smallest weight value.

[0098] The input feature dataset is preprocessed. Each feature vector is linearly transformed using three different learnable transformation matrices to obtain the query vector, key vector, and value vector, respectively. The parameters of the transformation matrices are continuously optimized during training to accommodate different types of feature mapping requirements. Specifically, the dimensions of each transformation matrix must match the input feature vector, and matrix multiplication is used to complete the mapping transformation in the feature space.

[0099] After obtaining the query vector, key vector, and value vector, a bidirectional cross-attention calculation is performed. During the forward attention calculation, the query vector and the key vector are matrix multiplied to obtain an initial attention score. To prevent vanishing or exploding gradients, this score is scaled by a scaling factor, typically set to the inverse of the square root of the key vector dimension. The scaled score is then exponentially increased to emphasize larger values. The result of the exponential operation is then normalized by dividing it by the sum of the exponentials corresponding to all key vectors to obtain the forward attention score.

[0100] In the reverse attention calculation, the query vector and key vector are transposed, and the same steps as the forward attention calculation are used, including matrix multiplication, scaling, exponential operation and normalization, to obtain the reverse attention score. This bidirectional calculation mechanism can capture the complementary relationship between features.

[0101] For the obtained forward and reverse attention scores, a learnable fusion weight parameter is introduced for weighted combination. The fusion weight parameter determines the contribution ratio of forward and reverse attention in the final result and is optimized through the backpropagation algorithm. The result of weighted fusion is the comprehensive relevance score;

[0102] The comprehensive correlation scores are input into the sigmoid activation function, and then nonlinearly transformed and normalized to obtain a fine-grained correlation matrix with values ​​ranging from [0 to 1]. Each element in the fine-grained correlation matrix represents the degree of correlation between the corresponding eigenvector pairs.

[0103] A weighted combination of a preset balance parameter, a fine-grained correlation matrix, and an identity matrix is ​​performed. The balance parameter adjusts the influence ratio between the correlation matrix and the identity matrix, while the inclusion of the identity matrix preserves the original information of the features. The combined result is then normalized using a softmax function to generate an adaptive weight matrix. In this weight matrix, feature vectors with higher correlations are assigned larger weights, while feature vectors with lower correlations are assigned smaller weights, thus achieving dynamic feature weighting.

[0104] For example, suppose a piece of public opinion data contains four key features: velocity, user interaction, sentiment, and topic relevance. The initial feature vectors are represented as [0.5, 0.3, 0.8, 0.2] and [0.6, 0.4, 0.1, 0.9], respectively. These features are mapped and transformed using three pre-trained transformation matrices: the query transformation matrix focuses on capturing the query properties of the features, the key transformation matrix on capturing the matching properties of the features, and the value transformation matrix on capturing the content properties of the features.

[0105] After this transformation, the resulting query vector emphasizes the query properties of the two features, velocity and sentiment, while the key vector captures the matching characteristics of user interaction and topic relevance. When calculating the attention score, a scaling factor of 0.5 is set, which is calculated by taking the square root of the feature dimension. Forward attention calculations show that the correlation between velocity and sentiment is the highest, reaching 0.75, while the correlation between user interaction and topic relevance is lower, at 0.35.

[0106] In reverse attention calculations, the attention scores calculated from the transposed feature vectors show that sentiment influences the speed of dissemination by 0.68, and topic relevance influences user engagement by 0.42. By setting the forward weight to 0.6 and the reverse weight to 0.4, the combined attention scores from both directions are combined to create a comprehensive relevance score that more comprehensively reflects the bidirectional correlation between features.

[0107] The comprehensive correlation score was mapped to the interval [0, 1] using the sigmoid function to obtain a fine-grained correlation matrix. The balance parameter was set to 0.7, and the fine-grained correlation matrix was weighted and combined with the identity matrix. The adaptive weight matrix was then generated using the softmax function. In the adaptive weight matrix, the feature pair of propagation speed and sentiment tendency received the highest weight of 0.82, indicating the strongest dynamic correlation between them.

[0108] In this embodiment, the learnable transformation matrix can adaptively adjust the feature mapping method, improve the flexibility and adaptability of feature representation, and capture the complementarity and dependency relationship between features by simultaneously considering the forward and reverse attention scores, avoiding the information loss that may be caused by unidirectional attention. The learnable fusion weight parameters enable the combination ratio of forward and reverse attention to be optimized according to actual needs, thereby enhancing the adaptability of the model.

[0109] In an optional embodiment,

[0110] The weighted feature vector is obtained and input into a multi-head temporal convolution unit with residual connection to extract temporal correlation features. The spatial dependency is obtained through the dynamic graph structure attention mechanism to generate a fusion feature matrix including:

[0111] Inputting the weighted feature vector into a multi-head temporal convolution unit with residual connection, the multi-head temporal convolution unit performs channel segmentation on the weighted feature vector to obtain multiple feature sub-vectors, and inputting the feature sub-vectors into a convolution layer group with multiple one-dimensional convolution kernels of different scales to obtain a first convolution feature, wherein each one-dimensional convolution layer has an independent learnable weight matrix;

[0112] Adaptively weighting the first convolutional features to obtain temporal features, performing adaptive residual connection on the temporal features and corresponding feature subvectors, and performing layer normalization processing with learnable scaling parameters to obtain head output features, and concatenating multiple head output features in the channel dimension and performing nonlinear feature transformation to obtain multi-head fusion features;

[0113] Constructing a dynamic graph structure unit to perform a double nonlinear transformation on the multi-head fusion features to obtain a feature correlation matrix, constructing a dynamic adjacency matrix based on the feature correlation matrix, and performing spatial information aggregation on the multi-head fusion features based on the dynamic adjacency matrix to obtain spatial dependency features;

[0114] The multi-head fusion feature and the spatial dependency feature are input into an adaptive gated fusion unit to generate a fusion feature, the fusion feature is adaptively calibrated in terms of channel importance through a compression excitation module to obtain a calibration feature, the calibration feature is adaptively weighted residual connected with the weighted feature vector and subjected to layer normalization processing to generate a fusion feature matrix.

[0115] Perform channel segmentation on the input weighted feature vector. Set the number of heads to 4 and evenly split the 64-dimensional feature vector into four sub-vectors along the channel dimension, each with a dimension of 16. The segmentation process maintains the temporal order of the features, ensuring that each sub-vector contains complete temporal information.

[0116] For each feature subvector, a processing unit consisting of multiple parallel one-dimensional convolutional layers is constructed. Each convolutional layer is configured with a convolution kernel of different sizes: a short-range convolution layer of length 3, a medium-range convolution layer of length 5, and a long-range convolution layer of length 7. Each convolutional layer is equipped with an independent weight matrix and bias term, and the initial value of the weight matrix is ​​initialized randomly. After performing the convolution operation, the first convolution feature is obtained.

[0117] Then, an adaptive channel weighting module is constructed to perform global average pooling on the first convolution feature in the time dimension to obtain a channel description vector. The channel description vector is input into a two-layer fully connected network. The number of neurons in the first layer is half the number of channels, and a rectified linear activation function is used. The number of neurons in the second layer is equal to the number of channels, and a sigmoid activation function is used. The channel weight vector is output and multiplied by the first convolution feature to obtain the weighted time series feature.

[0118] An adaptive residual connection is performed to perform weighted summation of the time series features and the corresponding feature sub-vectors. The two parts of the features are multiplied by the learnable weight coefficients respectively, and the results are input into the layer normalization module. The layer normalization module contains learnable scaling parameters and translation parameters to standardize the features and obtain the head output features.

[0119] The output features of the four heads are concatenated in the channel dimension to restore the original 64-dimensional feature space. The concatenated features are passed through a nonlinear transformation layer, which includes a fully connected layer and a rectified linear activation function, to obtain multi-head fusion features.

[0120] Construct a dynamic graph structure unit, which contains two transformation matrices. Multiply the multi-head fusion features with these two matrices to obtain two transformed feature matrices. Perform matrix multiplication to obtain the feature correlation matrix. Apply a threshold function with a threshold of 0.5 to the feature correlation matrix to generate a binary dynamic adjacency matrix.

[0121] Spatial information aggregation is performed based on the dynamic adjacency matrix. For each feature node, its adjacent node set is determined according to the adjacency matrix, and the weighted sum of the adjacent node features is calculated. The weight is determined by the connection strength in the adjacency matrix to obtain the spatial dependency feature.

[0122] An adaptive gated fusion unit is designed, consisting of two gating mechanisms: an update gate and a reset gate. Each gate consists of a fully connected layer and a sigmoid activation function. Multi-head fusion features and spatial dependency features are input into the gating unit, and feature fusion is performed based on the gating signal to obtain preliminary fused features.

[0123] The fused features are optimized using a compression-excitation module, which performs global average pooling and learns channel importance weights through a two-layer fully connected network (the first layer halves the number of channels and uses rectified linear activation; the second layer restores the number of channels and uses sigmoid activation). The learned weights are multiplied by the fused features to obtain the calibrated features.

[0124] The calibration features are residually connected with the original weighted feature vectors, and a linear combination is performed using learnable weighting coefficients. The combined features are input into the layer normalization module, which performs mean normalization and variance normalization, and applies learnable scaling and translation parameters to generate a fused feature matrix.

[0125] For example, taking a 30-minute public opinion data sequence as an example, the initial feature vector contains 64 dimensions, representing different public opinion feature indicators. It is divided into four head processing units, each containing 16 feature dimensions.

[0126] The features of each head are input into three convolutional layers of different scales: the convolution kernel with a length of 3 captures short-term changes within 1 minute, the convolution kernel with a length of 5 captures medium-term changes of 2-3 minutes, and the convolution kernel with a length of 7 captures long-term changes of more than 5 minutes.

[0127] Adaptive channel weighting is used to calculate the importance of each feature dimension. For the residual connection of the head output features, the initial weight ratio between the original and transformed features is set to 1:1. After the features of the four heads are recombined into a 64-dimensional feature vector, the feature expression is enhanced through nonlinear transformation.

[0128] Dynamic graph structure analysis shows that in the 30-minute sequence, the feature correlation between adjacent time points is generally higher than the threshold of 0.5, forming a dense feature association network. The gated fusion unit adaptively adjusts the fusion ratio of different features according to the importance of the features, generating a fusion feature matrix that reflects the complete temporal evolution law.

[0129] In this embodiment, a multi-head segmentation mechanism is used to divide the feature vector into multiple sub-vectors for parallel processing, significantly improving the parallel efficiency of feature extraction. By learning the importance weight of each channel, dynamic feature importance assessment is achieved, allowing the model to focus on more representative feature dimensions. The correlation between features is learned through dual nonlinear transformations, and a data-driven dynamic adjacency relationship is established, making the spatial dependency modeling of features more flexible and accurate.

[0130] Existing technologies typically use a single convolutional structure or attention mechanism to process temporal features, making it difficult to simultaneously capture feature changes at different time scales. Fixed feature fusion weights are used, making it impossible to dynamically adjust the contribution of features based on their importance. Furthermore, when processing spatial dependencies between features, predefined static graph structures are often used, lacking the ability to dynamically model feature associations.

[0131] This embodiment improves the efficiency and accuracy of feature extraction, enhances the ability to capture multi-scale time series patterns, realizes dynamic evaluation of feature importance, improves the modeling accuracy of feature space dependencies, and ensures the adaptability and stability of the feature fusion process, providing more effective technical support for in-depth analysis of time series features.

[0132] Figure 2 This is a schematic diagram of the multi-head temporal convolution unit processing of the big data-integrated overseas public opinion security situation intelligent perception method according to an embodiment of the present invention, illustrating the processing flow of the multi-head temporal convolution unit in this technical solution. The input 64-dimensional weighted feature vector is evenly divided into four 16-dimensional feature sub-vectors, which are input into four independent head processing units. Each head contains three parallel one-dimensional convolution layers, each configured with convolution kernels of different scales (kernel widths of 3, 5, and 7, respectively) to capture feature changes at different time scales. The short-scale convolution kernel (width = 3) primarily extracts local, rapidly changing features; the medium-scale convolution kernel (width = 5) captures medium-term trends; and the long-scale convolution kernel (width = 7) identifies long-term evolution patterns. The convolution results of each head are adaptively weighted by the channel and then residually connected with the original feature sub-vectors to generate the head output features. The outputs of the four heads are concatenated along the channel dimension to restore the 64-dimensional fused feature vector. This fused feature is then input into the dynamic graph structure unit, where a feature correlation matrix and a dynamic adjacency matrix are constructed through dual nonlinear transformations to achieve adaptive aggregation of spatial information. After processing by the adaptive gated fusion unit and the compressed excitation module, combined with the global residual connection from the input (shown by the dotted line in the figure), the final fusion feature matrix is ​​generated.

[0133] The multi-head parallel structure design enables the model to extract information from different feature subspaces and different time scales simultaneously, significantly improving the modeling ability of complex public opinion data. The introduction of residual connections effectively alleviates the gradient vanishing problem of deep networks and ensures the stability and efficiency of model training.

[0134] In an optional embodiment,

[0135] The evaluation index system based on the fusion feature matrix includes the public opinion communication power index, the public opinion influence index, and the public opinion sensitivity index, which includes:

[0136] The fusion feature matrix is ​​mapped to the public opinion communication power feature space through a multi-layer perceptron, and the public opinion communication power index is calculated based on the importance weight of the communication node and the length of the communication path;

[0137] Extracting time series features from the fusion feature matrix and combining them with user influence weights to construct a communication influence network that takes into account user interaction intensity, and calculating node influence scores to obtain a public opinion influence index;

[0138] Extracting text semantic features based on the fusion feature matrix, combining the preset sensitive dictionary and topic word vectors, and calculating the text sensitivity score through the attention mechanism to obtain the public opinion sensitivity index;

[0139] The evaluation index system is obtained by combining the public opinion communication power index, the public opinion influence index and the public opinion sensitivity index.

[0140] The fused feature matrix is ​​input into the multi-layer perceptron, which consists of three layers: the input layer receives the fused feature matrix, the hidden layer is activated using a rectified linear function, and the output layer maps the features to the communication power feature space. For each communication node, based on its position and connection relationship in the communication network, the node importance weight is calculated, including: counting the number of first-order neighbors of the node, and then considering the position of the node on the communication path. The closer the node is to the source of the communication, the higher the weight. For each communication path, its length (number of forwarding levels) is recorded. A shorter communication path indicates that the information is transmitted more directly. The node importance weight is weighted and combined with the normalized value of the communication path length to obtain the public opinion communication power index.

[0141] Time series feature extraction is performed on the fused feature matrix, and a sliding window method is used to obtain feature changes in different time periods. A communication influence network is constructed by combining user influence weights (including factors such as the number of users' followers and the quality of their historical posts). In the communication influence network, the weight of the edges is determined by the intensity of interaction between users, including the frequency of interactive behaviors such as forwarding, commenting, and liking. For each node in the network, its influence score is calculated iteratively: the influence value of each node is initialized, and the influence of the current node is updated based on the influence and connection strength of adjacent nodes. Multiple rounds of iterations are performed until the score stabilizes, and the public opinion influence index is summarized.

[0142] The semantic features of the text are extracted based on the fusion feature matrix, and the text is converted into a semantic vector using a text encoder. In combination with a pre-built sensitivity dictionary, sensitive words appearing in the text are identified and weighted. The weights of sensitive words are pre-defined based on their sensitivity, and the topic vectors are used to represent the thematic tendencies of the text. The contribution of each part of the text to sensitivity is calculated through an attention mechanism: attention weights are assigned to text fragments, with a focus on fragments containing sensitive words and key topic words. The sensitivity scores of each fragment are weighted and summed to obtain the overall public opinion sensitivity index.

[0143] The three calculated indicators are combined, the relative importance of the indicators is considered, different combination weights are set, and indicator thresholds are established to divide different levels of public opinion risks and obtain a comprehensive evaluation indicator system.

[0144] For example, consider the public opinion spread of a hot topic that has sparked widespread discussion on social media platforms. First, analyze the characteristics of its dissemination network: Original posts are forwarded through multiple levels, forming a dissemination tree structure. The number of first-level dissemination nodes is 500, the number of second-level dissemination nodes is 2,000, and the number of third-level dissemination nodes and above is 5,000.

[0145] In calculating the spreadability index, each transmission node is weighted according to its level in the transmission tree, with the first-level nodes having the highest weight. The transmission path length is also taken into account, as it is found that most effective transmission occurs within three levels, which is used to calculate the spreadability index.

[0146] Influence metrics calculations extract user interaction data, including forwarding volume, comment volume, and like volume. The constructed communication influence network reveals that several key node users have large follower bases, whose voices trigger large-scale interactive communication. Through iterative calculations, highly influential user groups are identified.

[0147] When calculating sensitivity indices, we analyze text content and discover that it involves multiple sensitive topics. We use an attention mechanism to focus on text segments containing sensitive words, and then analyze the overall textual content using the topic word vectors to ultimately determine the sensitivity of the topic.

[0148] A weighted combination of the three indicators yields a comprehensive assessment of the hot topic, which is used to guide subsequent public opinion guidance and handling efforts. By continuously monitoring changes in these indicators, we can keep abreast of public opinion trends and provide a basis for decision-making.

[0149] In this embodiment, a method of temporal feature extraction combined with user influence weights is adopted to achieve an accurate characterization of the dynamic characteristics of the propagation process, integrate information from multiple dimensions of text semantic features, sensitive dictionaries, and keyword vectors, and achieve accurate understanding of text content through the attention mechanism, thereby improving the comprehensiveness and accuracy of sensitivity assessment, comprehensively reflecting the propagation characteristics, scope of influence, and content attributes of public opinion events, and providing a reliable quantitative basis for the overall grasp of the public opinion situation.

[0150] In an optional embodiment,

[0151] Based on information entropy and evidence theory, the information gain of indicators is calculated to determine the confidence interval, and the indicator weight vector is generated. The indicator weight vector is combined with the quantitative score of the evaluation indicator, and the multi-criteria decision-making method is used to determine the public opinion security situation warning level, including:

[0152] Collect sample data of the evaluation indicators and perform normalization processing to obtain a sample probability distribution, calculate the information entropy of the sample probability distribution, determine the information gain of each evaluation indicator based on the information entropy, construct an identification framework for the evaluation indicators, calculate the basic probability distribution value of each evaluation indicator based on the information gain, calculate the subset confidence function and the intersection likelihood function for the basic probability distribution value to obtain the confidence interval of the evaluation indicator;

[0153] Based on the confidence interval, the basic probability distribution values ​​of different evaluation indicators are time-series decomposed and long-range correlation features are extracted, a time-varying state transfer matrix and a dynamic evolution matrix are constructed, a dynamic conflict degree is calculated through singular value decomposition, an evidence conflict matrix is ​​established and a similarity matrix is ​​determined, the similarity matrix is ​​normalized to obtain a credibility score of the evaluation indicator, and the credibility score is weightedly combined with the basic probability distribution value to generate an indicator weight vector;

[0154] Construct a quantitative score matrix for evaluation indicators, perform weighted normalization operation on the indicator weight vector and the quantitative score matrix to obtain a decision matrix, extract the maximum and minimum values ​​of each evaluation indicator in the decision matrix based on a multi-criteria decision-making method to construct a positive ideal solution vector and a negative ideal solution vector, calculate the Euclidean distance from the decision matrix to the positive ideal solution vector and the negative ideal solution vector, calculate the relative proximity score based on the Euclidean distance, perform fuzzy clustering operation on the relative proximity score to divide the warning level interval, and determine the public opinion security situation warning level based on the warning level interval.

[0155] For each evaluation indicator, the raw data is normalized, mapping data of different dimensions and value ranges to a standard interval. Statistical analysis is performed on the normalized data, and the sample probability distribution of each indicator is obtained by calculating the frequency of samples within each interval.

[0156] After obtaining the sample probability distribution, we perform information entropy analysis on each indicator. The process of information entropy analysis quantifies the uncertainty of the data, statistically analyzing the distribution of samples within each value interval, and calculating the proportion of each interval in the population to obtain the overall information entropy value. The greater the information entropy, the greater the uncertainty of the indicator.

[0157] Based on the information entropy analysis results, we further analyze the contribution of each indicator to the system, known as information gain. By comparing the change in system uncertainty before and after adding a particular indicator, we can determine the information gain value of that indicator. The larger the information gain value, the greater the contribution of that indicator to system identification. This process provides an important basis for the subsequent construction of the identification framework.

[0158] The system stratifies the indicators based on their information gain. Indicators with higher information gain are placed at the top of the identification framework, serving as the primary basis for identification. Indicators with lower information gain are placed at the bottom, serving as auxiliary basis for identification. This hierarchical structure makes the identification process more targeted and structured.

[0159] Based on the identification framework, a basic probability distribution value is calculated for each indicator. This process takes into account the indicator's information gain and its position in the identification framework. Indicators at higher levels generally receive larger basic probability distribution values, while indicators at lower levels receive smaller distribution values. This distribution reflects the varying importance of different indicators in the identification process.

[0160] The obtained basic probability distribution values ​​require further combinatorial analysis. By analyzing the combined effects of different indicator combinations, a subset confidence function is calculated. Simultaneously, the interactions between different indicators need to be analyzed to calculate the intersection likelihood function. These two functions together determine the confidence interval for each evaluation indicator, reflecting the range of the indicator's credibility.

[0161] Perform time series decomposition on the basic probability distribution values. By setting a time window, perform sliding analysis on the continuous monitoring data to extract long-range correlation features from the data. These long-range correlation features reflect the stable patterns and long-term trends of indicator values ​​over time. A time-varying state transition matrix is ​​constructed to describe the state transition patterns of the indicator at different time points. A dynamic evolution matrix is ​​also constructed to record the evolutionary characteristics of the indicator values.

[0162] The dynamic evolution matrix is ​​decomposed to extract key characteristic information and calculate the dynamic conflict degree. The dynamic conflict degree reflects the dynamic conflict relationship between different indicators. Based on this conflict degree, an evidence conflict matrix is ​​established. The similarity relationship between different indicators is further analyzed to form a similarity matrix. The similarity matrix is ​​normalized to obtain the credibility score for each evaluation indicator.

[0163] The credibility score and the basic probability distribution value are weighted and combined to generate the final indicator weight vector. The weight vector comprehensively considers both the importance and credibility of the indicator. At the same time, a quantitative score matrix for the evaluation indicators is constructed to record the specific scores of each indicator in different evaluation dimensions.

[0164] The indicator weight vector is combined with the quantitative score matrix and a weighted normalization operation is performed to obtain a decision matrix. The optimal and worst-case values ​​of each indicator are extracted from the decision matrix to construct the positive and negative ideal solution vectors, respectively. The distance from each evaluation object in the decision matrix to these two ideal solutions is calculated to obtain a relative proximity score.

[0165] Fuzzy cluster analysis is performed on the relative proximity scores, dividing them into different warning level intervals. The most reasonable interval division scheme is determined by optimizing the location of the cluster center. The warning level of the public opinion security situation is ultimately determined based on the interval in which the current evaluation indicator score falls.

[0166] For example, taking an online public opinion event that occurred on a social platform, real-time monitoring data is collected on the event's communication power indicators (such as the number of reposts, comments, and topic popularity), influence indicators (such as the number of participating users and the participation of opinion leaders), and sensitivity indicators (such as emotional tendency and the frequency of sensitive words).

[0167] Normalize the raw data, for example, normalize the forwarding volume from 0 to 1 million to the range of 0 to 1. Calculate the frequency distribution of the normalized data to obtain the sample probability distribution. The information entropy of the forwarding volume indicator is calculated to be 0.85, the information entropy of the comment volume indicator is 0.76, and so on.

[0168] Through comparative analysis, we calculated that the information gain of the forwarding volume metric is 0.32, and the information gain of the comment volume metric is 0.28. Based on this, we constructed a three-tiered identification framework: the first tier includes metrics with an information gain greater than 0.3, the second tier includes metrics with an information gain between 0.2 and 0.3, and the third tier includes metrics with an information gain less than 0.2.

[0169] Based on the information gain value, the basic probability distribution value for the forwarding volume indicator is calculated to be 0.15, and the basic probability distribution value for the comment volume indicator is calculated to be 0.12. Through the indicator combination analysis, the confidence of the subset containing forwarding volume and comment volume is calculated to be 0.23, and the intersection likelihood is 0.18, which determines the confidence interval of the forwarding volume indicator to be [0.12, 0.18].

[0170] Seven consecutive days of monitoring data were time-series decomposed, and long-range correlation features were extracted using a 24-hour sliding window. A 15×15 state transition matrix and dynamic evolution matrix were constructed. Singular value decomposition yielded a principal eigenvalue of 0.82, and a calculated dynamic conflict degree of 0.35.

[0171] After establishing the evidence conflict matrix, the calculated similarity between forwarding volume and comment volume was 0.78. After normalization, the credibility score for the forwarding volume metric was 0.82. Combining the credibility score with the basic probability distribution value at a weight of 0.6:0.4 yielded a final weight of 0.17 for the forwarding volume metric.

[0172] A quantitative score matrix containing all indicator scores was constructed, and the weight vector was combined with the score matrix to form a decision matrix. The maximum and minimum values ​​in the decision matrix were extracted to construct positive and negative ideal solution vectors. After calculating the Euclidean distance, a relative proximity score of 0.75 was obtained.

[0173] Through fuzzy cluster analysis, the relative proximity scores are divided into four warning level intervals: [0, 0.3] for low warning, [0.3, 0.6] for moderate warning, [0.6, 0.8] for high warning, and [0.8, 1] for severe warning. The current score of 0.75 falls into the high warning range, so the warning level for this public opinion event is determined to be high warning.

[0174] In this embodiment, the analysis method based on information entropy and information gain can accurately identify and quantify the contribution of each evaluation indicator to the system, providing a reliable basis for building a multi-level identification framework. The calculation of basic probability distribution value, subset trust function and intersection likelihood function is introduced to realize the quantitative expression of the credibility of the evaluation indicators. Through the construction of state transition matrix and dynamic evolution matrix, the dynamic change characteristics of the indicators are accurately described. In summary, this embodiment has strong systematicity and practicality, and can provide effective technical support for public opinion security situation warning.

[0175] In an optional embodiment,

[0176] The basic probability distribution values ​​of different evaluation indicators are decomposed in time series and long-range correlation features are extracted. The time-varying state transfer matrix and dynamic evolution matrix are constructed. The dynamic conflict degree is calculated by singular value decomposition. The evidence conflict matrix is ​​established and the similarity matrix is ​​determined.

[0177] For each evaluation indicator, a fractional-order difference coefficient matrix is ​​constructed based on the gamma function, the fractional-order difference coefficient matrix is ​​convolved with the basic probability distribution value of the evaluation indicator to obtain a fractional-order difference sequence, and the fractional-order difference sequence is decomposed according to the time scale to obtain a multi-scale feature vector;

[0178] Performing autocorrelation analysis on the multi-scale feature vectors to extract long-range correlation features, inputting the long-range correlation features into a recursive neural network to extract time-dependent features, organizing the time-dependent features into a state vector, and using a least squares method to solve a time-varying state transfer matrix based on the transfer relationship between the state vectors at adjacent moments;

[0179] Performing a matrix multiplication operation on the time-varying state transfer matrix and the multi-scale eigenvector to obtain an evidence dynamic evolution matrix, performing singular value decomposition on the evidence dynamic evolution matrix to extract principal component features, and calculating the empty set probability between evaluation indicators based on the principal component features;

[0180] An evidence conflict matrix is ​​constructed by taking the empty set probability values ​​between each two evaluation indicators as matrix elements, wherein the empty set probability values ​​are obtained by the intersection operation of the basic probability distribution functions of the evaluation indicators. An adaptive weight network is constructed based on the evidence conflict matrix to obtain the evidence credibility weights. The basic probability distribution values ​​of the evaluation indicators are dynamically weighted and fused according to the evidence credibility weights to obtain a similarity matrix.

[0181] When performing fractional differencing on each evaluation metric, a set of consecutive differencing order parameter values ​​is selected, such as 0.1, 0.2, 0.3, and so on, up to 0.9, forming nine differencing orders. A gamma function is used to generate a corresponding coefficient sequence for each differencing order. For each order r, the coefficients are calculated based on the gamma function. The time series is divided into n equal points, and the corresponding coefficient value is calculated for each time point t, resulting in an n-dimensional coefficient sequence. The coefficient sequences corresponding to all orders are combined into a matrix, forming a 9×n-dimensional fractional-order difference coefficient matrix.

[0182] The convolution operation is performed on the basic probability distribution value sequence of the evaluation indicator and the fractional-order difference coefficient matrix. The basic probability distribution value sequence is slidingly multiplied with each row of the coefficient matrix and the sum is calculated. For example, at a certain time point t, its fractional-order difference value is equal to the weighted sum of the basic probability distribution value at that point and the k previous time points and the corresponding coefficient value. The result is a sequence containing multiple difference order features.

[0183] Perform multi-scale decomposition on the difference sequence. Set multiple time scales, such as hourly (1 hour), daily (24 hours), and weekly (168 hours). For each scale, set a time window of corresponding length. The window slides across the sequence, extracting features from the data within the window each time. The extracted features include statistics such as mean, variance, kurtosis, and skewness, as well as frequency domain features obtained through wavelet transform. Features extracted at different scales are combined to form a multi-scale feature vector.

[0184] When performing autocorrelation analysis on multiscale eigenvectors, set the time delay sequence τ = [1, 2, 3, ..., k]. For each delay value, calculate the correlation coefficient between the eigenvector and the delayed sequence. Align the original sequence with the delayed sequence, calculate the covariance of the elements at corresponding positions, and then divide it by the product of the standard deviations of the two sequences. This yields a series of coefficients reflecting the correlation at different delays.

[0185] Design a recursive neural network architecture consisting of an input layer, multiple LSTM layers, and an output layer. The input layer receives long-range correlation features. The LSTM layer contains multiple memory cells, each of which includes an input gate, a forget gate, and an output gate. These gated cells control the flow of information to extract temporal features. During the network backpropagation process, gradient descent is used to optimize network parameters until the loss function converges. The network output features are organized in chronological order to form a state vector.

[0186] Analyze the relationship between state vectors at adjacent moments and establish a state transition equation. The current state vector is a linear combination of the previous state vector plus a random perturbation term. Use the least squares method to solve this equation, construct a sum-of-squares error function, and calculate the partial derivatives of the parameters and set them to zero. The optimal parameter values ​​are obtained through iterative solution, and these parameters form the time-varying state transition matrix.

[0187] The time-varying state transition matrix is ​​multiplied by the multiscale eigenvectors to obtain the evidence dynamic evolution matrix. Singular value decomposition is then performed to decompose the matrix into the product of three matrices: the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix. The singular values ​​are sorted by size, and the three singular values ​​with the largest contribution and their corresponding singular vectors are selected as principal component features.

[0188] To calculate the null set probability between evaluation metrics, obtain the basic probability distribution functions for the two metrics, calculate the intersection, find the common focal elements of the two distribution functions, and calculate the product of the probabilities corresponding to these focal elements. Sum all the intersection probabilities to obtain the null set probability. Repeat this calculation process for all metric pairs, and fill the corresponding positions in the matrix with the results to construct the evidence conflict matrix.

[0189] An adaptive weighted network is constructed based on the evidence conflict matrix. The network nodes represent the various evaluation metrics, and the connection weights between nodes are initialized to equal values. The weights are adjusted through an iterative optimization process: the weighted centrality of each node is calculated, and the connection weights are updated based on the centrality value. This process is repeated until the weight values ​​stabilize, resulting in the evidence credibility weight.

[0190] The evidence credibility weight is weighted with the basic probability distribution value of each evaluation indicator, and the basic probability distribution value of each indicator is multiplied by its corresponding credibility weight. The weighted results are normalized to ensure that the final sum of the probability values ​​is 1, and the similarity matrix is ​​obtained.

[0191] For example, assuming that a public opinion situation warning is conducted for a hot event, three evaluation indicators are selected: information dissemination speed (X1), user participation (X2) and emotional tendency (X3).

[0192] Construct a fractional-order difference coefficient matrix, selecting the difference orders [0.1, 0.3, 0.5, 0.7, 0.9]. A 5 × 24 coefficient matrix is ​​calculated using the gamma function (assuming hourly sampling and a 24-hour observation period). The initial basic probability distribution values ​​for the three indicators are: X1 = 0.35, X2 = 0.40, and X3 = 0.25. Convolve the basic probability distribution values ​​with the coefficient matrix to obtain a fractional-order difference sequence for each indicator.

[0193] Perform multi-scale decomposition on the difference series, setting time scales of 1 hour, 6 hours, and 12 hours. Statistical features are extracted at each time scale. For example, for X1 at the 1-hour scale, the calculated eigenvalues ​​are 0.42, 0.15, kurtosis 1.8, and 0.6, forming the eigenvector.

[0194] When performing autocorrelation analysis, we set the delay value τ to [1, 2, 3, 4, 5] hours. The calculated autocorrelation coefficients of X1 at different delays were [0.85, 0.72, 0.60, 0.45, 0.32], forming long-range correlation features. These features were then fed into the designed LSTM network (consisting of an input layer, two LSTM layers, and an output layer) to extract temporal dependency features and form a state vector.

[0195] A state transition equation is constructed based on adjacent state vectors. The least squares method is used to solve the resulting 3×3 time-varying state transition matrix. This matrix is ​​multiplied by the multiscale eigenvectors to obtain the evidence dynamic evolution matrix. Singular value decomposition is performed on the matrix, and the eigenvectors corresponding to the two largest singular values ​​are selected as principal component features.

[0196] Calculate the null set probabilities between the indicators: the null set probability between X1 and X2 is 0.15, the null set probability between X1 and X3 is 0.25, and the null set probability between X2 and X3 is 0.20. Construct a 3×3 evidence conflict matrix and build an adaptive weight network based on this matrix. The resulting credibility weights for the three indicators are 0.38, 0.42, and 0.20, respectively.

[0197] The credibility weights are weighted and fused with the basic probability distribution values ​​to obtain the final similarity matrix.

[0198] In this embodiment, a fractional-order difference processing method is introduced. By constructing a fractional-order difference coefficient matrix based on the gamma function, more detailed feature extraction of the evaluation index data is achieved. In combination with the multi-scale decomposition technology, data characteristics are analyzed at different time scales, overcoming the problem of incomplete information at a single time scale. By constructing a time-varying state transition matrix, an accurate description of the dynamic characteristics of the system is achieved. An evidence conflict matrix is ​​constructed based on the probability of the empty set, and an adaptive weight network is established. The weights of the evaluation indicators can be dynamically adjusted according to the credibility of the evidence, overcoming the limitations of the traditional fixed weight method.

[0199] Existing public opinion security situation early warning solutions typically use a single integer-order difference method to process evaluation indicator data, which makes it difficult to effectively capture nonlinear features in the data. They often use a fixed time window for feature extraction of time series data, which can easily lose important information at different time scales. They also use fixed weights or simple dynamic weighting methods, which fail to fully consider the dynamic correlations and conflict characteristics between indicators.

[0200] This embodiment significantly improves the ability to extract nonlinear features, enhances the ability to capture long-term dependencies of time series data, realizes the adaptive dynamic fusion of evaluation indicators, can more accurately describe the evolution characteristics of public opinion, provide more reliable early warning basis, has stronger practicality and adaptability, and significantly improves the accuracy and reliability of public opinion security situation early warning.

[0201] Figure 3This is a heat matrix diagram of the early warning accuracy of the intelligent perception method for overseas public opinion security situation that integrates big data according to an embodiment of the present invention. It shows the early warning accuracy performance of four methods at different levels of public opinion complexity. The grayscale value in the figure reflects the accuracy level, with lighter colors indicating higher accuracy. This technical solution performs best at all levels of complexity, with a particularly significant advantage in high-complexity areas. Specifically, when the public opinion complexity is 1.0, the early warning accuracy of this technical solution reaches 96.8%. Even in the extreme case of complexity increasing to 6.0, it maintains a high level of 89.7%, a decrease of only 7.1 percentage points. In contrast, the accuracy of the traditional DS evidence theory method (Dempster-Shafer evidence theory) drops from 93.0% to 81.2% as the complexity changes from 1.0 to 6.0, a drop of 11.8 percentage points. The ordinary Bayesian network and SVM classification methods show drops of 15.3 and 16.5 percentage points, respectively. The color change gradient in the thermal matrix clearly demonstrates this performance difference: the color change gradient of this technical solution is the slowest, while the other methods become significantly darker in high-complexity areas.

[0202] The fractional-order differential coefficient matrix design in this technical solution can effectively capture the long-term memory characteristics of the data. In addition, the multi-scale feature decomposition extracts features at different time granularities, so that the system can still maintain stability and accuracy when facing complex and drastically changing public opinion data.

[0203] Figure 4 This chart compares the prediction accuracy of the intelligent perception method for overseas public opinion security situation integrated with big data, according to an embodiment of the present invention. It shows the prediction accuracy of different public opinion situation warning technology solutions in different time windows. The present technology solution (marked by diamonds) exhibits the highest prediction accuracy across all time windows, particularly in the 24-hour and 48-hour time windows, with accuracy rates reaching 91.2% and 88.7%, respectively. This represents a significant improvement over the DS evidence theory method (marked by triangles, with accuracy rates of 82.3% and 79.5%, respectively) and the BP neural network method (marked by squares, with accuracy rates of 78.9% and 75.2%, respectively). The traditional SVM method (marked by crosses) performs the worst, with accuracy rates not exceeding 75% across all time windows. While the accuracy of all methods decreases with increasing time windows, the present technology solution experiences the smallest decrease, dropping by only 9.4 percentage points, from 93.5% in the 6-hour window to 84.1% in the 72-hour window. Under the same conditions, the other methods all experience decreases of over 12 percentage points.

[0204] This demonstrates that this technical solution, through the integration of fractional-order differencing and evidence theory, effectively captures the long-term dependencies of public opinion data and enhances the model's robustness for long-term prediction windows. Data shows that across all test scenarios, this technical solution improves accuracy by an average of 8.7 percentage points, with a maximum improvement of 13.1 percentage points (compared to the SVM method for a 48-hour window).

[0205] In an optional embodiment,

[0206] Generating public opinion control suggestions and warning information based on the public opinion security situation warning level and pushing them to the management terminal includes:

[0207] Based on the public opinion security situation warning level, a preset control rule tree is queried, and corresponding public opinion control suggestions are matched according to the path of the warning level in the control rule tree;

[0208] Generate warning information based on the warning level and the public opinion control suggestion, and push the warning information to the management terminal.

[0209] Build a pre-set control rule tree. This tree uses a multi-level tree structure. The root node represents the initial state. The first-level nodes represent different warning levels. The second-level nodes correspond to the key influencing factors at each warning level. The third-level nodes record specific control measures. Under each warning level node, multiple branch paths are set, each path containing specific judgment conditions and corresponding control recommendations.

[0210] Once the warning level is determined, the system matches paths within the control rule tree, locating the node corresponding to the warning level. Based on the specific circumstances of that warning level, such as the scope of impact, development trends, and transmission speed, the system then matches down through the subtree of that node layer by layer. At each node, the system determines whether the current situation meets the conditions of that node until it finds the most matching leaf node and obtains the corresponding public opinion control recommendations.

[0211] The warning level and the corresponding control recommendations are integrated to generate warning information. This information includes multiple dimensions, including warning time, warning level, event type, impact scope, development trends, propagation characteristics, potential risks, and control recommendations. This information is structured and organized to form a standard warning report.

[0212] When pushing warning information, we identify recipients and set push rules based on the responsibilities and permissions of different management terminals. We push warning information with varying levels of detail to different management terminals. A real-time push mechanism ensures timely delivery of warning information. Push logs are also recorded, including push time, receiving terminal, and information content, for subsequent tracking and statistics.

[0213] For example, taking a negative public opinion incident at an overseas university as an example, the assessment system determined that the incident was a Level 3 warning (moderate warning). The Level 3 warning node was located in the management and control rule tree, and path matching was performed based on the specific manifestations of the current public opinion. The incident currently shows regional dissemination characteristics, mainly fermenting in Region A. The speed of dissemination on social media platforms is at a medium level, with approximately 5,000 new related discussions added every hour. Netizens' emotions are clearly divided, with supporters and opponents each occupying one side, of which more than 60% are comments with obvious emotional tendencies. The media has followed up on the incident, but it has not yet formed a large-scale dissemination. In terms of sensitivity, since it involves the pre-set second category of issues, it is a moderately sensitive incident.

[0214] Based on the above characteristics, the rule tree is matched layer by layer to locate the corresponding control recommendation path. The control recommendations generated based on this path include: organizing relevant persons in charge to hold a press conference to promptly announce the progress of the incident investigation; coordinating media coverage to ensure objective coverage and guide rational discussion; entrusting experts in the field to interpret and resolve doubts; establishing a special Q&A area on relevant social platforms to respond to public concerns; and strengthening communication with representatives to understand their demands and provide timely feedback on treatment plans.

[0215] The system integrates warning levels and control recommendations to generate warning information, including basic information about the incident, its scope of impact, development trends, analysis of online sentiment, mainstream media attitudes, and potential risk assessments. Warning information is distributed to management terminals at all levels according to a hierarchical authorization principle: a complete warning analysis report is distributed to key personnel at relevant institutions; specific tasks in different areas are distributed to grassroots terminals, and the information reception status of each terminal is recorded to ensure timely processing and feedback of warning information.

[0216] In this embodiment, by establishing a multi-level management and control rule tree, accurate matching of management and control suggestions under different warning levels is achieved, and path matching is performed based on preset judgment conditions to ensure the pertinence and operability of management and control suggestions. The hierarchical push mechanism of warning information ensures the timeliness and accuracy of information transmission, avoids redundancy and omissions in the information transmission process, forms a closed-loop warning response mechanism, and improves the efficiency and accuracy of public opinion handling.

[0217] A second aspect of an embodiment of the present invention provides an intelligent perception system for overseas public opinion and security situation integrating big data, including:

[0218] The first unit is used to obtain public opinion data from overseas social media platforms and perform preprocessing to obtain standardized public opinion information and perform multi-dimensional feature extraction to construct a feature dataset;

[0219] The second unit is used to calculate the fine-grained correlation of the feature vectors in the feature data set using a bidirectional cross-attention mechanism and construct an adaptive weight matrix. The feature vectors in the feature data set are weighted based on the adaptive weight matrix to obtain weighted feature vectors and input them into a multi-head temporal convolution unit with residual connection to extract temporal correlation features. The spatial dependency is obtained through a dynamic graph structure attention mechanism to generate a fusion feature matrix.

[0220] The third unit is used to construct an evaluation index system including a public opinion dissemination power index, a public opinion influence index, and a public opinion sensitivity index based on the fusion feature matrix;

[0221] The fourth unit is used to calculate the information gain of indicators based on information entropy and evidence theory to determine the confidence interval, generate the indicator weight vector, and combine the indicator weight vector with the quantitative score of the evaluation indicator to determine the public opinion security situation warning level using a multi-criteria decision-making method;

[0222] The fifth unit is used to generate public opinion control suggestions and warning information based on the public opinion security situation warning level and push them to the management terminal.

[0223] According to a third aspect of the embodiments of the present invention,

[0224] An electronic device is provided, comprising:

[0225] processor;

[0226] a memory for storing processor-executable instructions;

[0227] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0228] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0229] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent perception method for overseas public opinion security situation integrating big data, characterized by: include: Obtain public opinion data from overseas social media platforms and preprocess it to obtain standardized public opinion information and perform multi-dimensional feature extraction to construct a feature dataset; A bidirectional cross-attention mechanism is used to calculate the fine-grained correlation of the feature vectors in the feature data set and construct an adaptive weight matrix. The feature vectors in the feature data set are weighted based on the adaptive weight matrix to obtain weighted feature vectors and input them into a multi-head temporal convolution unit with residual connection to extract temporal correlation features. The spatial dependency is obtained through the dynamic graph structure attention mechanism to generate a fusion feature matrix, including: Inputting the weighted feature vector into a multi-head temporal convolution unit with residual connection, the multi-head temporal convolution unit performs channel segmentation on the weighted feature vector to obtain multiple feature sub-vectors, and inputting the feature sub-vectors into a convolution layer group with multiple one-dimensional convolution kernels of different scales to obtain a first convolution feature, wherein each one-dimensional convolution layer has an independent learnable weight matrix; Adaptively weighting the first convolutional features to obtain temporal features, performing adaptive residual connection on the temporal features and corresponding feature subvectors, and performing layer normalization processing with learnable scaling parameters to obtain head output features, and concatenating multiple head output features in the channel dimension and performing nonlinear feature transformation to obtain multi-head fusion features; Constructing a dynamic graph structure unit to perform a double nonlinear transformation on the multi-head fusion features to obtain a feature correlation matrix, constructing a dynamic adjacency matrix based on the feature correlation matrix, and performing spatial information aggregation on the multi-head fusion features based on the dynamic adjacency matrix to obtain spatial dependency features; Inputting the multi-head fusion feature and the spatial dependency feature into an adaptive gated fusion unit to generate a fusion feature, performing channel importance adaptive calibration on the fusion feature through a compression excitation module to obtain a calibration feature, performing an adaptive weighted residual connection on the calibration feature and the weighted feature vector and performing layer normalization processing to generate a fusion feature matrix; Based on the fusion feature matrix, an evaluation index system is constructed, including a public opinion dissemination power index, a public opinion influence index, and a public opinion sensitivity index; Based on information entropy and evidence theory, the information gain of indicators is calculated to determine the confidence interval, and the indicator weight vector is generated and combined with the quantitative score of the evaluation indicator. The multi-criteria decision-making method is used to determine the warning level of the public opinion security situation; According to the public opinion security situation warning level, public opinion control suggestions and warning information are generated and pushed to the management terminal.

2. The method according to claim 1, characterized in that Obtain public opinion data from overseas social media platforms and preprocess it to obtain standardized public opinion information and perform multi-dimensional feature extraction. The feature dataset constructed includes: Collect public opinion data from overseas social media platforms, including text data, image data, and video data posted by users; Performing multilingual unified encoding on text data through cross-language aligned word vector spaces, extracting key information from image and video data using deep learning-based object detection methods, and integrating the multilingual unified encoding and key information to generate standardized public opinion information; A hierarchical feature extraction network is used to extract features from the normalized public opinion information, wherein local features are extracted at the bottom layer, contextual relationship features are extracted at the middle layer, and global semantic features are extracted at the top layer. The extracted features are fused through jump connections to obtain a feature data set.

3. The method according to claim 1, characterized in that Using a bidirectional cross-attention mechanism to calculate fine-grained correlations on feature vectors in the feature dataset and construct an adaptive weight matrix includes: Receiving feature vectors in a feature data set, mapping the feature vectors to generate a query vector, a key vector, and a value vector through different transformation matrices, wherein the transformation matrix is ​​a learnable parameter matrix; Based on the bidirectional cross attention mechanism, the query vector and the key vector are matrix multiplied and scaled by a scaling factor before performing an exponential operation. The result of the exponential operation is divided by the sum of the exponentials corresponding to all key vectors to obtain a forward attention score. The transposed query vector and the transposed key vector are subjected to the same operation to obtain a reverse attention score. The forward attention score and the reverse attention score are weightedly fused based on a learnable fusion weight parameter to obtain a comprehensive correlation score, and the comprehensive correlation score is input into a sigmoid activation function to obtain a normalized fine-grained correlation matrix; The balance parameter, the fine-grained correlation matrix and the identity matrix are weighted and combined and then input into the softmax function to generate an adaptive weight matrix, wherein the adaptive weight matrix is ​​used to characterize the dynamic correlation strength between the eigenvectors, the eigenvector with the highest correlation corresponds to the largest weight value, and the eigenvector with the lowest correlation corresponds to the smallest weight value.

4. The method according to claim 1, wherein The evaluation index system based on the fusion feature matrix includes the public opinion communication power index, the public opinion influence index, and the public opinion sensitivity index, which includes: The fusion feature matrix is ​​mapped to the public opinion communication power feature space through a multi-layer perceptron, and the public opinion communication power index is calculated based on the importance weight of the communication node and the length of the communication path; Extracting time series features from the fusion feature matrix and combining them with user influence weights to construct a communication influence network that takes into account user interaction intensity, and calculating node influence scores to obtain a public opinion influence index; Extracting text semantic features based on the fusion feature matrix, combining the preset sensitive dictionary and topic word vectors, and calculating the text sensitivity score through the attention mechanism to obtain the public opinion sensitivity index; The evaluation index system is obtained by combining the public opinion communication power index, the public opinion influence index and the public opinion sensitivity index.

5. The method according to claim 1, wherein Based on information entropy and evidence theory, the information gain of indicators is calculated to determine the confidence interval, and the indicator weight vector is generated. The indicator weight vector is combined with the quantitative score of the evaluation indicator, and the multi-criteria decision-making method is used to determine the public opinion security situation warning level, including: Collect sample data of the evaluation indicators and perform normalization processing to obtain a sample probability distribution, calculate the information entropy of the sample probability distribution, determine the information gain of each evaluation indicator based on the information entropy, construct an identification framework for the evaluation indicators, calculate the basic probability distribution value of each evaluation indicator based on the information gain, calculate the subset confidence function and the intersection likelihood function for the basic probability distribution value to obtain the confidence interval of the evaluation indicator; Based on the confidence interval, the basic probability distribution values ​​of different evaluation indicators are time-series decomposed and long-range correlation features are extracted, a time-varying state transfer matrix and a dynamic evolution matrix are constructed, a dynamic conflict degree is calculated through singular value decomposition, an evidence conflict matrix is ​​established and a similarity matrix is ​​determined, the similarity matrix is ​​normalized to obtain a credibility score of the evaluation indicator, and the credibility score is weightedly combined with the basic probability distribution value to generate an indicator weight vector; Construct a quantitative score matrix for evaluation indicators, perform weighted normalization operation on the indicator weight vector and the quantitative score matrix to obtain a decision matrix, extract the maximum and minimum values ​​of each evaluation indicator in the decision matrix based on a multi-criteria decision-making method to construct a positive ideal solution vector and a negative ideal solution vector, calculate the Euclidean distance from the decision matrix to the positive ideal solution vector and the negative ideal solution vector, calculate the relative proximity score based on the Euclidean distance, perform fuzzy clustering operation on the relative proximity score to divide the warning level interval, and determine the public opinion security situation warning level based on the warning level interval.

6. The method according to claim 5, characterized in that The basic probability distribution values ​​of different evaluation indicators are decomposed in time series and long-range correlation features are extracted. The time-varying state transfer matrix and dynamic evolution matrix are constructed. The dynamic conflict degree is calculated by singular value decomposition. The evidence conflict matrix is ​​established and the similarity matrix is ​​determined. For each evaluation indicator, a fractional-order difference coefficient matrix is ​​constructed based on the gamma function, the fractional-order difference coefficient matrix is ​​convolved with the basic probability distribution value of the evaluation indicator to obtain a fractional-order difference sequence, and the fractional-order difference sequence is decomposed according to the time scale to obtain a multi-scale feature vector; Performing autocorrelation analysis on the multi-scale feature vectors to extract long-range correlation features, inputting the long-range correlation features into a recursive neural network to extract time-dependent features, organizing the time-dependent features into a state vector, and using a least squares method to solve a time-varying state transfer matrix based on the transfer relationship between the state vectors at adjacent moments; Performing a matrix multiplication operation on the time-varying state transfer matrix and the multi-scale eigenvector to obtain an evidence dynamic evolution matrix, performing singular value decomposition on the evidence dynamic evolution matrix to extract principal component features, and calculating the empty set probability between evaluation indicators based on the principal component features; An evidence conflict matrix is ​​constructed by taking the empty set probability values ​​between each two evaluation indicators as matrix elements, wherein the empty set probability values ​​are obtained by the intersection operation of the basic probability distribution functions of the evaluation indicators. An adaptive weight network is constructed based on the evidence conflict matrix to obtain the evidence credibility weights. The basic probability distribution values ​​of the evaluation indicators are dynamically weighted and fused according to the evidence credibility weights to obtain a similarity matrix.

7. The method according to claim 1, characterized in that Generating public opinion control suggestions and warning information based on the public opinion security situation warning level and pushing them to the management terminal includes: Based on the public opinion security situation warning level, a preset control rule tree is queried, and corresponding public opinion control suggestions are matched according to the path of the warning level in the control rule tree; Generate warning information based on the warning level and the public opinion control suggestion, and push the warning information to the management terminal.

8. An intelligent perception system for overseas public opinion security situation integrating big data, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain public opinion data from overseas social media platforms and perform preprocessing to obtain standardized public opinion information and perform multi-dimensional feature extraction to construct a feature dataset; The second unit is used to calculate the fine-grained correlation of the feature vectors in the feature data set using a bidirectional cross-attention mechanism and construct an adaptive weight matrix. The feature vectors in the feature data set are weighted based on the adaptive weight matrix to obtain weighted feature vectors and input them into a multi-head temporal convolution unit with residual connection to extract temporal correlation features. The spatial dependency is obtained through a dynamic graph structure attention mechanism to generate a fusion feature matrix. The third unit is used to construct an evaluation index system including a public opinion dissemination power index, a public opinion influence index, and a public opinion sensitivity index based on the fusion feature matrix; The fourth unit is used to calculate the information gain of indicators based on information entropy and evidence theory to determine the confidence interval, generate the indicator weight vector, and combine the indicator weight vector with the quantitative score of the evaluation indicator to determine the public opinion security situation warning level using a multi-criteria decision-making method; The fifth unit is used to generate public opinion control suggestions and warning information based on the public opinion security situation warning level and push them to the management terminal.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

Images