Cross-platform public opinion information acquisition and analysis system based on deep learning

Abstract

The invention provides a cross-platform public opinion information collection and analysis system based on deep learning, and belongs to the technical field of public opinion information collection and analysis. A multi-platform data collector is deployed to capture social media data in real time to establish an original cache pool, and preprocessing and metadata extraction are performed on multi-modal data; a multi-modal deep fusion convolution algorithm is adopted to construct a parallel text, image and audio branch network for feature extraction, a cross-modal attention mechanism is adopted to parallelly calculate correlation weights of all modal features, and a gating fusion unit is used to perform efficient nonlinear transformation and information screening to generate a unified semantic vector. Semantic vectors are input into a self-adaptive cross-modal sentiment analysis recognition model, and the technical problem that the requirement for millisecond-level real-time processing of mass public opinion data cannot be met due to the fact that the inference time of a deep model is too long is solved through a collaborative mechanism of dynamic optimization, parallel processing and hierarchical response.

Description

Technical Field

[0001] This invention belongs to the field of public opinion information collection and analysis technology, specifically, it relates to a cross-platform public opinion information collection and analysis system based on deep learning. Background Technology

[0002] In the field of cross-platform public opinion information collection and analysis, traditional technologies mainly employ deep neural network models for sentiment analysis and public opinion judgment. These models utilize deep learning architectures such as convolutional neural networks, recurrent neural networks, and Transformers to extract features and classify sentiment from social media data. These traditional deep models are widely used in public opinion monitoring systems, online public opinion analysis platforms, and social media sentiment analysis tools, achieving high-precision sentiment recognition and semantic understanding. However, traditional deep models suffer from a serious drawback when processing massive amounts of public opinion data: excessively long inference times. Due to their complex model structure, large number of parameters, and enormous computational load, a single inference often takes hundreds of milliseconds or even seconds, far from meeting the millisecond-level response requirements of real-time public opinion monitoring. In current real-time public opinion analysis applications, the lack of effective model acceleration and dynamic optimization mechanisms makes the inference latency of deep models a bottleneck in system performance. When faced with massive data streams generated by sudden public opinion events, traditional technologies cannot complete the analysis and processing within the specified time, leading to delayed public opinion warnings and untimely emergency responses, severely impacting the effectiveness of public opinion management. Summary of the Invention

[0003] In view of this, the present invention provides a cross-platform public opinion information collection and analysis system based on deep learning, which can solve the technical problem in the prior art that the reasoning time of deep models is too long and cannot meet the requirements of millisecond-level real-time processing of massive public opinion data.

[0004] This invention is implemented as follows: It provides a deep learning-based cross-platform public opinion information collection and analysis system, comprising: deploying multi-platform data collectors to crawl social media platforms in real time; acquiring multimodal public opinion data such as text, images, videos, and audio through API interfaces and web crawler technology; establishing a raw data cache pool; preprocessing the multimodal data in the raw data cache pool, including data cleaning, format standardization, noise filtering, and missing value imputation, while extracting metadata information such as timestamps, user identifiers, and geographic locations; and using a multimodal deep fusion convolution algorithm to extract features from the preprocessed multimodal data, calculating the features of each modality through a cross-modal attention mechanism. Weights are assigned to different modalities using a gating fusion unit to generate a unified semantic representation vector. This unified semantic representation vector is then input into a cross-modal sentiment analysis and recognition model for sentiment bias analysis, outputting positive, negative, and neutral sentiment classifications and corresponding confidence scores. A two-layer game optimization model is used to calculate public opinion heat based on the sentiment analysis results. The upper-layer model aims to maximize the accuracy of public opinion influence assessment, while the lower-layer model aims to minimize computational resource consumption. The public opinion heat value is generated by solving the game equilibrium. A real-time monitoring and early warning mechanism is established, which automatically triggers an early warning signal when the public opinion heat value exceeds a preset threshold, and simultaneously initiates an emergency response process to generate risk assessment and handling suggestions.

[0005] Specifically, the steps of the multimodal deep fusion convolution algorithm are as follows: establishing text branch convolutional neural networks, image branch convolutional neural networks, and audio branch convolutional neural networks to extract features from the corresponding modal data; calculating the correlation weight matrix between text features, image features, and audio features through a cross-modal attention mechanism; weighting and fusing the weight matrix with the feature vectors of each modality; using a gated fusion unit to perform nonlinear transformation and information filtering on the fused multimodal features; and mapping the processed features to a fixed-dimensional unified semantic representation vector through a fully connected layer.

[0006] Specifically, the calculation steps of the cross-modal attention mechanism are as follows: the text feature vector, image feature vector, and audio feature vector are linearly transformed to obtain the query matrix, key matrix, and value matrix, respectively; the dot product of the query matrix and the key matrix is ​​calculated and normalized by the softmax function to obtain the attention weight; the attention weight is multiplied by the value matrix to obtain the weighted feature representation; and the weighted feature representations between different modalities are concatenated and dimensionality reduced.

[0007] Specifically, the processing mechanism of the gated fusion unit involves transforming the input multimodal fusion features using the sigmoid activation function to generate a gated signal; performing element-wise multiplication of the gated signal with the original fusion features to achieve information filtering; performing a nonlinear transformation on the filtered features using the tanh activation function to enhance expressive power; and adding the transformed features to the residual connection to maintain gradient propagation stability.

[0008] Specifically, the structure of the two-layer game optimization model is as follows: the upper-layer model aims to maximize the accuracy of public opinion influence assessment, with the objective function being the influence assessment accuracy maximization function, and the inputs including the propagation speed index, influence range coefficient, attention weight, emotional polarity intensity, and time decay factor; the lower-layer model aims to minimize computational resource consumption, with the objective function being the computational resource consumption minimization function, and the inputs including CPU utilization, memory usage, network bandwidth usage, and storage space usage; the two objective functions influence and constrain each other through resource allocation coupling terms.

[0009] Specifically, the triggering conditions for the real-time monitoring and early warning mechanism are as follows: if the fluctuation range of the public opinion heat value exceeds ±30% and the deviation of the emotional polarity intensity exceeds ±0.4, then the dual heat value abnormality regulator is activated; if only the fluctuation range of the public opinion heat value exceeds ±30%, then only the heat value fluctuation regulator is activated; if only the deviation of the emotional polarity intensity exceeds ±0.4, then only the emotional deviation regulator is activated.

[0010] The cross-modal sentiment analysis and recognition model refers to a multi-layer neural network based on the Transformer architecture, which includes an input embedding layer, a multi-head self-attention mechanism layer, a feedforward neural network layer, a layer normalization module, and an output classification layer. The number of heads in the multi-head attention mechanism is dynamically adjusted according to three system operating parameters: CPU utilization, memory usage, and GPU memory usage of the input data.

[0011] Specifically, the dynamic adjustment mechanism for the number of heads is set to 8 heads when the CPU utilization is in the range of [10%, 40%), 12 heads when the memory utilization is in the range of [20%, 60%), and 16 heads when the video memory utilization is in the range of [1000MB, 4000MB]. This dynamic adjustment mechanism allows the model to optimize the accuracy and efficiency of attention calculation based on the system's operating status, while avoiding the waste of computing resources and the occurrence of performance bottlenecks.

[0012] The influence assessment accuracy maximization function is used to calculate the objective function value of the upper-level game model to maximize the accuracy of public opinion influence assessment. The inputs include the propagation speed index, the influence range coefficient, the attention weight, the emotional polarity intensity, and the time decay factor. The output is the influence assessment accuracy value. The final influence assessment accuracy value is obtained by comprehensively calculating the logarithmic product, exponential weighted sum, and reciprocal transformation of each input parameter.

[0013] The attention head number adjustment function refers to an adaptive parameter control function based on the system's operating state, used to adjust the attention head number parameter of the multi-head attention mechanism in the cross-modal sentiment analysis and recognition model. A comprehensive performance score is calculated based on three key system parameters: CPU utilization, memory usage, and video memory usage. Different attention head number configurations are adopted when the comprehensive performance score falls into different ranges.

[0014] The steps for establishing the training dataset for the cross-modal sentiment analysis and recognition model specifically involve collecting multimodal public opinion sample data, including text, images, videos, and audio, from mainstream social media platforms such as Weibo, Douyin, Zhihu, and Tieba. The collected raw data is manually labeled to form sentiment polarity tags. Data augmentation techniques are used to expand the number of training samples. The labeled multimodal data is then divided into training, validation, and test sets in an 8:1:1 ratio. The training steps for the cross-modal sentiment analysis and recognition model specifically involve initializing model parameters using the Xavier normal distribution random initialization method, setting the learning rate to 0.001, and using the Adam optimizer for parameter updates. During training, a mini-batch gradient descent strategy with a batch size of 32 is employed, and the error between the model's prediction results and the true labels is calculated using the cross-entropy loss function.

[0015] The computational resource consumption minimization function is used to calculate the objective function value of the lower-level game model to minimize computational resource consumption. The inputs include CPU utilization, memory usage, network bandwidth usage, and storage space usage. The output is the computational resource consumption value, which is obtained by weighted summation of the square of CPU utilization, the product of memory usage and network bandwidth usage, and the square root of storage space usage.

[0016] Optionally, since network bandwidth and storage space are prepared in advance and are sufficient, the default setting for network bandwidth utilization and storage space utilization is 1. When there is significant network lag or hard drive read / write lag, you can decide based on experience whether to cancel the setting of 1 and use the actual values ​​of network bandwidth utilization and storage space utilization.

[0017] The propagation speed index is calculated by statistically analyzing the growth rate of forwards, comments, and likes per unit time. The influence range coefficient is measured by the number of users participating in the discussion, the breadth of geographical distribution, and the number of platforms covered. The attention weight is obtained by comprehensively evaluating search popularity, media coverage frequency, and the number of related topics.

[0018] The emotional polarity intensity is calculated from the confidence score of the cross-modal emotion analysis and recognition model, the time decay factor is calculated from the difference between the time of the public opinion event and the current time, the CPU utilization rate is obtained in real time through the system performance monitoring tool, and the memory occupancy rate is obtained through the memory management unit monitoring.

[0019] The network bandwidth utilization rate is obtained through network traffic monitoring statistics, the storage space utilization rate is obtained through disk usage statistics, the public opinion heat value is calculated by the equilibrium solution of the two-layer game optimization model, the influence assessment accuracy value is calculated and output by the influence assessment accuracy maximization function, and the computing resource consumption value is calculated and output by the computing resource consumption minimization function.

[0020] This invention employs a dynamic model optimization mechanism based on system operating status. It adaptively adjusts the number of heads in the multi-head attention mechanism of the cross-modal sentiment analysis and recognition model according to real-time hardware parameters such as CPU utilization, memory usage, and GPU memory usage. When system resources are sufficient, a 16-head attention configuration is used to improve analysis accuracy; when the system load is high, it automatically reduces to an 8-head or 12-head configuration to reduce computational complexity and effectively shorten model inference time. Simultaneously, it combines a two-layer game optimization model to minimize computational resource consumption while ensuring the accuracy of public opinion influence assessment, achieving a dynamic balance between inference speed and analysis accuracy. The attention head adjustment function of this invention achieves segmented parameter control through a comprehensive performance score, avoiding resource waste and performance bottlenecks caused by traditional fixed parameter configurations. The information filtering mechanism of the gated fusion unit further reduces redundant computation, and the parallel processing architecture of multi-modal features improves data processing efficiency. This allows the entire system to significantly shorten the single inference time while maintaining high analysis accuracy, meeting the millisecond-level processing requirements of massive public opinion data. In summary, this invention solves the technical problem mentioned in the background art, namely that the inference time of deep models is too long and cannot meet the millisecond-level real-time processing requirements of massive public opinion data, by establishing an adaptive dynamic optimization mechanism and an efficient parallel processing architecture. Attached Figure Description

[0021] Figure 1 This is a flowchart of the system execution steps of the present invention.

[0022] Figure 2 This is a real-time trend chart of the system hardware monitoring parameters in the embodiment.

[0023] Figure 3 This is a weight distribution diagram of multimodal feature fusion in the embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] like Figure 1 The diagram shown is a flowchart of a cross-platform public opinion information collection and analysis system based on deep learning provided by this invention. The system performs the following steps: S01. Deploy multi-platform data collectors to capture real-time data from social media platforms, and obtain multimodal public opinion data such as text, images, videos, and audio through API interfaces and web crawler technology to establish a raw data cache pool; S02. Preprocess the multimodal data in the original data cache pool, including data cleaning, format standardization, noise filtering and missing value imputation, and extract metadata information such as timestamps, user identifiers and geographic locations. S03. A multimodal deep fusion convolution algorithm is used to extract features from the preprocessed multimodal data. The cross-modal attention mechanism is used to calculate the feature weights of each modality. A gated fusion unit is used to integrate information from different modalities to generate a unified semantic representation vector. S04. Input the unified semantic representation vector into the cross-modal sentiment analysis and recognition model to perform sentiment tendency analysis, and output three sentiment classification results: positive, negative, and neutral, and corresponding confidence scores. S05. A two-layer game optimization model is used to calculate the public opinion heat index based on the sentiment analysis results. The upper-layer model aims to maximize the accuracy of the public opinion influence assessment, while the lower-layer model aims to minimize the consumption of computing resources. The public opinion heat index is generated by solving the game equilibrium. S06. Establish a real-time monitoring and early warning mechanism. When the public opinion heat value exceeds the preset threshold, an early warning signal will be automatically triggered, and an emergency response process will be initiated to conduct risk assessment and generate handling suggestions.

[0026] The specific steps of the multimodal deep fusion convolution algorithm include: establishing text branch convolutional neural networks, image branch convolutional neural networks, and audio branch convolutional neural networks to extract features from the corresponding modal data; calculating the correlation weight matrix between text features, image features, and audio features through a cross-modal attention mechanism; weighting and fusing the weight matrix with the feature vectors of each modality; using a gated fusion unit to perform nonlinear transformation and information filtering on the fused multimodal features; and mapping the processed features to a fixed-dimensional unified semantic representation vector through a fully connected layer.

[0027] The calculation steps of the cross-modal attention mechanism include: performing linear transformations on the text feature vector, image feature vector, and audio feature vector to obtain the query matrix, key matrix, and value matrix, respectively; calculating the dot product of the query matrix and the key matrix and normalizing it using the softmax function to obtain the attention weights; multiplying the attention weights by the value matrix to obtain the weighted feature representation; and concatenating and reducing the dimensionality of the weighted feature representations between different modalities.

[0028] The gating fusion unit's processing mechanism includes: transforming the input multimodal fusion features using the sigmoid activation function to generate a gating signal; performing element-wise multiplication of the gating signal and the original fusion features to achieve information filtering; performing nonlinear transformation on the filtered features using the tanh activation function to enhance expressive power; and adding the transformed features to the residual connection to maintain gradient propagation stability.

[0029] The specific structure of the two-layer game optimization model includes: the upper-layer model aims to maximize the accuracy of public opinion influence assessment, with the objective function being the function that maximizes the accuracy of influence assessment. The inputs include the propagation speed index, the influence range coefficient, the attention weight, the emotional polarity intensity, and the time decay factor. The constraints are that the computation time does not exceed 500ms and the memory usage does not exceed 1GB. The lower-layer model aims to minimize the consumption of computational resources, with the objective function being the function that minimizes the consumption of computational resources. The inputs include CPU utilization, memory usage, network bandwidth usage, and storage space usage. The constraints are that the accuracy of public opinion heat calculation is not less than 85%. The two objective functions influence and constrain each other through resource allocation coupling terms.

[0030] The triggering conditions for the real-time monitoring and early warning mechanism are as follows: if the fluctuation range of the public opinion heat value exceeds ±30% and the deviation of the emotional polarity intensity exceeds ±0.4, the dual heat value abnormality regulator will be activated; if only the fluctuation of the public opinion heat value exceeds ±30%, only the heat value fluctuation regulator will be activated; if only the emotional polarity intensity deviation exceeds ±0.4, only the emotional deviation regulator will be activated.

[0031] The system hardware monitoring parameters are obtained in the following ways: CPU core temperature is obtained in real time through temperature sensors, with a normal range of 35℃ to 75℃; CPU utilization is obtained through performance counter statistics, with a normal range of 10% to 80%; memory usage is obtained through memory management unit monitoring, with a normal range of 20% to 85%; memory access latency is obtained through memory controller measurement, in ns; memory bandwidth utilization is obtained through memory bus monitoring, in GB / s; video memory usage is obtained through graphics driver reading, in MB; video memory temperature is obtained through GPU temperature sensor monitoring, with a normal range of 40℃ to 85℃; network latency is obtained through ping tests, in ms; disk I / O rate is obtained through disk performance counter statistics, in MB / s; and power consumption is obtained through power management unit measurement, in W.

[0032] The data processing optimization conditions are as follows: when the CPU core temperature is below 45°C and the memory temperature is above 75°C, the brightness of the CPU performance regulator is increased and the memory cooling process is activated simultaneously; when only the CPU core temperature is below 45°C, only the brightness of the CPU performance regulator is increased; when only the memory temperature is above 75°C, only the memory cooling process is activated.

[0033] The specific structure of the cross-modal sentiment analysis and recognition model is a multi-layer neural network based on the Transformer architecture, including an input embedding layer, a multi-head self-attention mechanism layer, a feedforward neural network layer, a layer normalization module, and an output classification layer. The number of heads in the multi-head attention mechanism is dynamically adjusted according to three system operating parameters: CPU utilization, memory usage, and GPU memory usage. When the CPU utilization is in the range of [10%, 40%), it is set to 8 heads; when the memory usage is in the range of [20%, 60%), it is set to 12 heads; and when the GPU memory usage is in the range of [1000MB, 4000MB], it is set to 16 heads. Through the dynamic adjustment mechanism, the model optimizes the accuracy and efficiency of attention calculation according to the system operating status, while avoiding the waste of computing resources and the occurrence of performance bottlenecks. The steps for establishing the training dataset for the cross-modal sentiment analysis and recognition model specifically include collecting multimodal public opinion sample data containing text, images, videos, and audio from mainstream social platforms such as Weibo, Douyin, Zhihu, and Tieba; manually labeling the collected raw data to form sentiment polarity labels, with label categories including positive, negative, and neutral sentiment tendencies; expanding the number of training samples through data augmentation techniques, including text synonym replacement, image rotation and scaling, and audio noise addition; dividing the labeled multimodal data into training, validation, and test sets in an 8:1:1 ratio to ensure the balance and representativeness of the data distribution; and finally establishing a large-scale training dataset containing 500,000 multimodal public opinion samples. The specific steps for training the cross-modal sentiment analysis and recognition model include: initializing model parameters using the Xavier normal distribution random initialization method; setting the learning rate to 0.001 and using the Adam optimizer for parameter updates; employing a mini-batch gradient descent strategy with a batch size of 32 during training; calculating the error between the model's prediction results and the true labels using the cross-entropy loss function; setting an early stopping mechanism to stop training when the validation set loss does not decrease for 5 consecutive epochs to prevent overfitting; evaluating model performance on the validation set every 10 epochs and saving the optimal model parameters; continuing the entire training process until the model converges or reaches the maximum number of training epochs of 200; and evaluating the final performance metrics of the model on the test set after training, including accuracy, precision, recall, and F1 score.

[0034] The attention head number adjustment function is an adaptive parameter control function based on the system's operating state. It is used to adjust the attention head number parameters of the multi-head attention mechanism in the cross-modal sentiment analysis and recognition model. The function calculates a comprehensive performance score based on three key system parameters: CPU utilization, memory usage, and GPU memory usage. When the comprehensive performance score is in the range of [0, 0.3), an 8-head attention configuration is used to reduce computational complexity and ensure stable system operation. When the comprehensive performance score is in the range of [0.3, 0.7), a 12-head attention configuration is used to balance computational accuracy and efficiency requirements. When the comprehensive performance score is in the range of [0.7, 1.0], a 16-head attention configuration is used to maximize the model's expressive power and analytical accuracy. Through a segmented parameter adjustment strategy, the optimal balance between model performance and system resource consumption is achieved, ensuring that the public opinion analysis system can maintain a stable and efficient working state under different operating environments.

[0035] The function for maximizing the accuracy of influence assessment is used to calculate the objective function value of the upper-level game model in order to maximize the accuracy of public opinion influence assessment. The inputs include the propagation speed index, the influence range coefficient, the attention weight, the emotional polarity intensity, and the time decay factor. The output is the influence assessment accuracy value. The function obtains the final influence assessment accuracy value by comprehensively calculating the logarithmic product of the propagation speed index and the influence range coefficient, the exponential weighted sum of the attention weight and the emotional polarity intensity, and the reciprocal transformation of the time decay factor.

[0036] The computational resource consumption minimization function is used to calculate the objective function value of the lower-level game model to minimize computational resource consumption. The inputs include CPU utilization, memory usage, network bandwidth usage, and storage space usage. The output is the computational resource consumption value. The function obtains the final computational resource consumption value by weighted summing of the square of CPU utilization, the product of memory usage and network bandwidth usage, and the square root of storage space usage.

[0037] The propagation speed index is calculated by the growth rate of forwards, comments, and likes per unit time, derived from user behavior data statistics of social media platforms, and used as an input parameter for the influence assessment accuracy maximization function. The influence scope coefficient is measured by the number of users participating in the discussion, geographical distribution breadth, and platform coverage, derived from user geographic location information and platform coverage statistics, and used as an input parameter for the influence assessment accuracy maximization function. The attention weight is comprehensively evaluated by search popularity, media coverage frequency, and the number of related topics, derived from search engine data and media monitoring systems, and used as an input parameter for the influence assessment accuracy maximization function. The sentiment polarity intensity is calculated by the confidence score of the cross-modal sentiment analysis recognition model, derived from the sentiment analysis results in step S04, and used as an input parameter for the influence assessment accuracy maximization function and as a trigger condition for the real-time monitoring and early warning mechanism. The time decay factor is calculated by the difference between the time of the public opinion event and the current time, derived from the timestamp information extracted in step S02, and used as an input parameter for the influence assessment accuracy maximization function. The CPU utilization rate is obtained in real time through system performance monitoring tools, derived from the operating system's process manager statistics, and used as an input parameter for the resource consumption minimization function. The memory occupancy rate is obtained through monitoring by the memory management unit and is derived from system memory allocation statistics. It is used to calculate the input parameters of the resource consumption minimization function. The network bandwidth occupancy rate is obtained through network traffic monitoring statistics and is derived from data transmission statistics of network interfaces. It is used to calculate the input parameters of the resource consumption minimization function. The storage space occupancy rate is obtained through disk usage statistics and is derived from file system space allocation statistics. It is used to calculate the input parameters of the resource consumption minimization function. The public opinion heat value is calculated from the equilibrium solution of the two-layer game optimization model and is used for threshold comparison and early warning trigger judgment of the real-time monitoring and early warning mechanism. The influence assessment accuracy value is calculated and output by the influence assessment accuracy maximization function and is used for the objective function optimization of the upper-layer game model and the equilibrium solution of the two-layer game optimization model. The computational resource consumption value is calculated and output by the computational resource consumption minimization function and is used for the objective function optimization of the lower-layer game model and the equilibrium solution of the two-layer game optimization model.

[0038] The multimodal deep fusion convolutional algorithm applied to a cross-platform public opinion information collection and analysis system can achieve deep semantic alignment and information complementarity between different data modalities. By using a multi-branch convolutional neural network architecture to process heterogeneous data such as text, images, and audio, it effectively solves the problems of incomplete information and semantic understanding bias in traditional single-modal analysis methods when dealing with complex public opinion scenarios. The cross-modal attention mechanism can automatically learn the relevance weights between different modalities, enabling the system to dynamically adjust the importance of each modality's information according to the specific public opinion content. The gated fusion unit ensures the retention of key features and the filtering of redundant information during the fusion process through nonlinear transformation and information filtering mechanisms. The final generated unified semantic representation vector contains comprehensive semantic information of multimodal data, significantly improving the accuracy and robustness of public opinion sentiment analysis. It is particularly suitable for the public opinion analysis needs of diversified content expression methods in the current social media environment. The algorithm is particularly suitable for cross-platform public opinion analysis systems because user expressions on modern social media platforms exhibit obvious multimodal characteristics. Single text analysis cannot accurately capture rich emotional information such as image expressions, video emotions, and audio tone. However, by constructing a unified cross-modal feature space, the algorithm can comprehensively utilize complementary information from various modalities, effectively addressing the diversity and complexity of public opinion expression. At the same time, the attention mechanism and gating unit design in the algorithm take into account the differences in content features across different platforms, and can adaptively adjust the feature fusion strategy to ensure consistent analytical accuracy and stability when processing heterogeneous data from different platforms. The algorithm extracts lexical semantic features and syntactic structure information through a text branch convolutional neural network. It controls the granularity and abstraction of text features using convolution kernel size and pooling stride parameters. The image branch network extracts visual features through multi-scale convolution operations, balancing computational efficiency and detail preservation using feature map channel number and spatial resolution parameters. The audio branch network extracts spectral features using one-dimensional convolution, controlling the capture of temporal and frequency domain information through sampling rate and window length parameters. A cross-modal attention mechanism calculates an attention weight matrix, which quantifies the correlation strength between different modal features. A gated fusion unit outputs a gate signal value, controlling the passage ratio of each modality. The fully connected layer ultimately outputs a fixed-dimensional semantic vector containing the fused multimodal semantic representation. Through a series of parameter calculations and data transformation processes, the algorithm can convert the original heterogeneous multimodal data into a unified high-level semantic representation, providing reliable feature input for subsequent sentiment analysis and public opinion judgment.

[0039] The specific implementation methods of the above steps are described in detail below.

[0040] The specific implementation of step S01 involves establishing a distributed multi-platform data collector architecture and configuring corresponding collection modules for the data interfaces and webpage structures of different social media platforms. The system first connects to the open data interfaces of mainstream platforms such as Weibo, Douyin, Zhihu, and Tieba via application programming interfaces (APIs), setting an access frequency limit of 300 requests per minute to avoid triggering the platforms' anti-crawler mechanisms. Simultaneously, a web crawler module is deployed, employing a crawler framework to scrape data from platforms without open interfaces. The crawler latency is set to 2-5 seconds to simulate normal user behavior. The data collection module uses multi-threaded concurrent processing technology to simultaneously process multimodal data types such as text, images, video, and audio. A thread pool size of 16 threads is set to balance collection efficiency and system resource consumption. A buffer queue mechanism temporarily stores the collected raw data in a memory cache pool with a capacity of 1GB to ensure the continuity and stability of data collection. The purpose of this step is to provide a comprehensive data source foundation for the entire public opinion analysis system, ensuring the acquisition of real-time public opinion information from various platforms.

[0041] The specific implementation of step S02 involves performing standardized preprocessing operations on the multimodal data in the original data cache pool. The data cleaning module first uses regular expression matching to remove noise information such as special characters, URL links, and emoticons from the text data, setting a text length filtering threshold of 10-5000 characters to exclude invalid content that is too short or too long. Format standardization processing uses a unified encoding conversion algorithm to convert all text data to UTF-8 encoding format, image data is uniformly adjusted to JPEG format with a resolution parameter set to 800×600 pixels, and audio data is converted to WAV format with a sampling rate set to 16kHz. The noise filtering module uses a signal-to-noise ratio (SNR) calculation method to assess the quality of audio data, setting a SNR threshold of 20dB to filter low-quality audio. Image quality assessment uses a blur detection algorithm with a sharpness threshold of 0.8 to ensure image usability. Missing value imputation processing uses an interpolation algorithm to complete the time series data, and uses the median imputation method to handle numerical missing data. The metadata extraction module parses key attributes such as timestamps, user identifiers, geographic coordinates, and platform source identifiers from the raw data. It then uses a hash algorithm to anonymize the user identifiers to protect privacy. This step aims to ensure data quality and consistency in subsequent analysis.

[0042] The specific implementation of step S03 involves using a multimodal deep fusion convolutional algorithm to achieve feature extraction and fusion of heterogeneous data. The text branch convolutional neural network uses a word embedding layer to convert text into a dense vector representation, setting the vocabulary size to 50,000 words and the embedding dimension to 300. The convolutional layers use multi-scale convolutional kernels to extract n-gram features, with kernel sizes of 3, 4, and 5 to capture text patterns of different lengths, and the number of convolutional kernels is set to 128. The image branch convolutional neural network uses a combination of convolutional and pooling layers to extract visual features. The convolutional layers are set to three stages, each containing two convolutional layers, with kernel sizes of 3×3 and channel numbers of 64, 128, and 256 respectively. The pooling layers use max pooling with a pooling window size of 2×2. The audio branch convolutional neural network uses one-dimensional convolutional operations to process spectral features, setting the kernel length to 5, the stride to 2, and the number of channels to 64. The cross-modal attention mechanism maps the feature vectors of each modality to query, key, and value matrices through linear transformation. Attention weights are calculated using a dot-product attention mechanism and normalized using a softmax function. A temperature parameter is set to 0.1 to enhance attention concentration. The gated fusion unit uses a sigmoid activation function to generate a gate signal with a threshold of 0.5. Information filtering is achieved through element-wise multiplication, followed by a nonlinear transformation using a tanh activation function. The residual connection weight parameter is set to 0.8. The fully connected layer maps the fused features to a 1024-dimensional unified semantic representation vector. This step transforms multimodal data into a unified high-dimensional feature representation, providing rich semantic information for sentiment analysis.

[0043] The specific implementation of step S04 involves inputting the unified semantic representation vector into a cross-modal sentiment analysis and recognition model based on the Transformer architecture for sentiment classification. The model structure includes an input embedding layer that converts the 1024-dimensional semantic vector into a 512-dimensional internal representation, with positional encoding using sine and cosine coding to add positional information. The multi-head self-attention mechanism layer dynamically adjusts the number of attention heads based on system performance: 8 heads when CPU utilization is between 10% and 40%, 12 heads when memory usage is between 20% and 60%, and 16 heads when GPU memory usage is between 1000MB and 4000MB. The feedforward neural network layer uses a two-layer fully connected structure with a hidden layer dimension of 2048 and the ReLU activation function. The layer normalization module uses standardized parameters with a mean of 0 and a variance of 1 for feature standardization. The output classification layer uses a softmax function to convert the model output into probability distributions for positive, negative, and neutral sentiment categories, while also outputting corresponding confidence scores. The confidence threshold is set to 0.7 to ensure the reliability of the prediction results. The model training employed the cross-entropy loss function, with a learning rate of 0.001, a batch size of 32, and 200 training epochs. An early stopping mechanism was implemented, halting training if the loss on the validation set did not decrease after five consecutive epochs. The purpose of this step was to accurately identify the sentiment bias in the public opinion data, providing a foundation for sentiment analysis in subsequent popularity calculations.

[0044] The specific implementation of step S05 involves using a two-layer game optimization model to calculate the public opinion heat index based on the sentiment analysis results. The upper-layer model aims to maximize the accuracy of public opinion influence assessment. Input parameters include the propagation speed index, influence range coefficient, attention weight, sentiment polarity intensity, and time decay factor. The propagation speed index is calculated statistically based on the growth rates of forwards, comments, and likes, with a value range of 0-100. The influence range coefficient is comprehensively evaluated based on the number of participating users, geographical distribution breadth, and platform coverage, with a value range of 0-10. The attention weight is calculated based on search popularity and media coverage frequency, with a weight range of 0-1. The lower-layer model aims to minimize computational resource consumption. Input parameters include CPU utilization, memory usage, network bandwidth usage, and storage space usage, with each parameter ranging from 0% to 100%. The game equilibrium solution uses an iterative optimization algorithm, setting an upper limit of 100 iterations, a convergence accuracy threshold of 0.001, and constraints including a computation time not exceeding 500 milliseconds, memory usage not exceeding 1GB, and a public opinion heat calculation accuracy not less than 85%. The resource allocation coupling term uses a weighted average method to realize the mutual influence and constraint between the two-layer model, with the coupling weight parameter set to 0.6. The purpose of this step is to optimize system resource consumption while ensuring computational accuracy, thereby generating accurate public opinion heat values.

[0045] The specific implementation of step S06 involves establishing a real-time monitoring and early warning mechanism based on threshold judgment. The monitoring module continuously detects fluctuations in public opinion heat and changes in emotional polarity intensity. The heat fluctuation threshold is set to ±30%, and the emotional polarity intensity deviation threshold is set to ±0.4. When the public opinion heat fluctuation exceeds ±30% and the emotional polarity intensity deviation exceeds ±0.4, the system automatically triggers a dual heat anomaly regulator. The regulator uses a proportional-integral-derivative (PID) control algorithm to handle the anomaly, with a proportional coefficient of 0.8, an integral coefficient of 0.2, and a derivative coefficient of 0.1. The heat fluctuation regulator is activated when only the public opinion heat fluctuation exceeds the threshold, and the emotional polarity intensity deviation regulator is activated when only the emotional polarity intensity deviation exceeds the threshold. The early warning signal generation module employs a multi-level early warning mechanism, setting three warning levels: minor, moderate, and severe, with corresponding thresholds of 20%, 50%, and 80%, respectively. The emergency response process includes risk assessment calculation, generation of handling suggestions, and notification to relevant departments. Risk assessment uses a risk matrix method to calculate the risk level, and handling suggestions are generated by matching with an expert knowledge base to produce corresponding handling plans. Hardware monitoring parameters include key indicators such as CPU core temperature, CPU utilization, memory usage, and video memory temperature. The normal range for CPU core temperature is set at 35℃~75℃, CPU utilization at 10%~80%, memory usage at 20%~85%, and video memory temperature at 40℃~85℃. The data processing optimization mechanism dynamically adjusts based on hardware status. When the CPU core temperature is below 45℃ and the video memory temperature is above 75℃, both CPU performance enhancement and video memory cooling are activated simultaneously. The purpose of this step is to achieve real-time hardware monitoring and timely early warning, ensuring the system can quickly respond to abnormal situations.

[0046] It should be explained that the key technical ideas of this invention include a multimodal deep fusion convolution algorithm, a game theory-based public opinion heat optimization model, and a dynamic adaptive attention mechanism adjustment technique. The multimodal deep fusion convolution algorithm has significant advantages over traditional single-modal analysis methods. Traditional methods can only process single data types such as text or images, failing to fully utilize the diverse expressions of social media content. This algorithm, however, constructs three independent feature extraction branches for text, images, and audio, and combines a cross-modal attention mechanism to achieve semantic alignment of heterogeneous data. It can comprehensively utilize complementary information between different modalities, significantly improving the accuracy and robustness of sentiment recognition, and is particularly suitable for the diverse expression methods of users in the current social media environment. The game theory-based public opinion heat optimization model balances computational accuracy and resource consumption through a two-layer game structure. The upper-layer model aims to maximize the accuracy of influence assessment, while the lower-layer model aims to minimize computational resource consumption. The two achieve an optimal balance point through a game equilibrium mechanism. Compared to traditional fixed-weight evaluation methods, this model can dynamically adjust the computational strategy according to the actual system operating state, ensuring both the accuracy of the analysis results and optimizing computational efficiency. The dynamic adaptive attention mechanism adjusts the number of attention heads in real time according to the system's operating status, overcoming the limitations of traditional fixed parameter settings that cannot adapt to different operating environments. By comprehensively considering key system indicators such as CPU utilization, memory usage, and GPU memory usage, it achieves intelligent matching between model complexity and system performance, ensuring high analysis accuracy even in resource-constrained environments. The synergistic effect of these key technologies forms a complete intelligent public opinion analysis solution. The multimodal fusion algorithm provides the system with a rich and accurate feature representation foundation, the game optimization model achieves the optimal balance between computational efficiency and analysis accuracy, and the dynamic adjustment mechanism ensures stable operation of the system in different environments. The three work together to form a public opinion analysis system with high adaptability, high accuracy, and high efficiency, demonstrating stronger practicality and reliability compared to existing technologies in handling complex and ever-changing online public opinion environments.

[0047] It should be noted that in existing public opinion analysis systems, different modalities such as text, images, and audio data have different data structures, feature dimensions, and semantic representations. Direct fusion often leads to problems such as feature space mismatch, severe information redundancy, and low computational efficiency. This invention designs a specialized cross-modal attention mechanism, which performs linear transformations on the feature vectors of each modality to obtain query matrices, key matrices, and value matrices. Through dot product operations and softmax normalization, it automatically learns the relevance weights between different modalities, achieving semantic alignment and effective fusion of heterogeneous features. The gated fusion unit further controls the information throughput ratio through sigmoid gating signals, enhances feature expressiveness through tanh nonlinear transformation, and maintains gradient propagation stability through residual connections, ensuring efficient integration of multimodal information and intelligent filtering of redundant information. Traditional public opinion analysis systems lack real-time monitoring and intelligent allocation mechanisms for the usage status of hardware resources such as CPU, memory, and video memory, often resulting in some resources being overloaded while others are idle, causing the overall system performance to fall short of optimal levels. This invention establishes a complete hardware monitoring parameter acquisition system. It acquires key parameters such as CPU core temperature, utilization rate, memory usage rate, and video memory temperature in real time through temperature sensors, performance counters, and memory management units. Different processing strategies are adopted according to different hardware states. When the CPU core temperature is below 45°C, the CPU performance regulator is enhanced. When the video memory temperature is above 75°C, heat dissipation is initiated. Through dynamic resource allocation and load balancing mechanisms, the coordinated operation of various hardware resources and the maximum utilization of the overall system performance are achieved.

[0048] Specifically, the principle of this invention is as follows: The key to solving the deep learning model inference latency problem lies in constructing a complete dynamic optimization and parallel acceleration mechanism. First, the attention head number adjustment function dynamically controls model complexity by monitoring system operating parameters in real time. When the overall performance score is low, the number of attention heads is automatically reduced to decrease computation; when system resources are sufficient, the number of attention heads is increased to improve accuracy. This adaptive adjustment mechanism avoids the performance waste or resource shortage problems of traditional fixed configurations under different load conditions, ensuring that the model always runs at optimal computational complexity. Second, the multimodal deep fusion convolution algorithm adopts a parallel processing architecture. Text, image, and audio branches perform feature extraction simultaneously, avoiding time accumulation caused by serial processing. The cross-modal attention mechanism significantly improves the efficiency of correlation calculation through matrix parallel operations. The sigmoid and tanh activation function calculations of the gated fusion unit can fully utilize the parallel computing capabilities of the GPU, significantly shortening the overall inference time. Furthermore, the two-layer game optimization model maximizes the accuracy of influence assessment while ensuring a computation time of no more than 500ms through the collaborative optimization of the objective functions of the upper and lower layers. Simultaneously, it minimizes resource consumption while maintaining an accuracy of at least 85%. This multi-objective optimization strategy ensures that the system maintains sufficient analytical accuracy while meeting real-time requirements. Finally, the tiered triggering strategy of the real-time monitoring and early warning mechanism avoids unnecessary complex calculations. When only heat fluctuation adjustment is needed, the complete dual regulator is not activated, reducing system response latency. The logical consistency of the technical solution of this invention lies in its adherence to a layered acceleration approach, from dynamic optimization to parallel acceleration to tiered processing. Each technical step is optimized for different sources of latency, forming a complete real-time assurance system.

[0049] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0050] In this embodiment, the specific implementation methods of steps S01 to S02 are the same as those described above, and will not be repeated in detail here.

[0051] The specific implementation of step S03 involves using a multimodal deep fusion convolution algorithm to achieve feature extraction and fusion of heterogeneous data. The calculation process of the cross-modal attention mechanism is specifically represented as follows: ; In the formula, For the first A query matrix for each modality; For the first The bond matrix of each mode; For the first A value matrix for each modality; For the first Feature vectors of each modality; For the first Feature vectors of each modality; and These represent indices for text, image, and audio modalities, respectively. Represents text modality, Represents image modality, Indicates audio modality; , , The matrix represents the learnable linear transformation parameters, each with a dimension of 512×512. The attention weight matrix is ​​calculated as follows: ; In the formula, For modality For modes Attention weight matrix; The dimension of the key vector is set to 512; This represents the intermodal correlation bias term, ranging from -0.5 to 0.5. The weighted feature representation is calculated as follows: ; ; In the formula, For modality For modes The weighted feature representation matrix; This is the fused feature vector after dimensionality reduction and concatenation. For feature splicing operations; , , These represent the cross-modal weighted features of text-image, text-audio, and image-audio, respectively. Subscript 12 represents the weighted feature of text modality 1 on image modality 2, subscript 13 represents the weighted feature of text modality 1 on audio modality 3, and subscript 23 represents the weighted feature of image modality 2 on audio modality 3. The dimension-reduced projection matrix has a dimension of 1536×1024. The processing mechanism of the gated fusion unit is specifically represented as follows: ; ; ; In the formula, This is the gating signal vector; It is the sigmoid activation function; This is a gated weight matrix with a dimension of 1024×1024; The input is the fused feature vector; This is the gated bias vector; This is an element-wise multiplication operation; These are the filtered feature vectors; Output vector for the gating unit; The output weight matrix has a dimension of 1024×1024; This is the output bias vector; This is the residual connectivity coefficient, with a value of 0.8.

[0052] The specific implementation of step S04 involves inputting the unified semantic representation vector into the cross-modal sentiment analysis and recognition model for sentiment classification. The attention head number adjustment function is specifically expressed as follows: ; ; In the formula, This is the overall performance score. , , These are weighting coefficients, set to 0.4, 0.3, and 0.3 respectively. This represents the percentage of CPU usage. This represents the percentage of memory usage. This represents video memory usage, in MB. The maximum video memory capacity is set to 8192MB; The number of attention heads is dynamically adjusted. The parameter is obtained as follows: The system performance is monitored in real time using system performance monitoring tools, including step 1: reading the processor performance counters; and step 2: calculating the average utilization within the current time window. The data is obtained through monitoring by the memory management unit, including step 1: querying the system memory allocation table; and step 2: calculating the ratio of used memory to total memory. Obtained through the graphics driver interface, including step 1: calling the graphics card status query function; step 2: reading the current video memory usage value.

[0053] The specific implementation of step S05 involves using a two-layer game optimization model to calculate the public opinion heat value based on the sentiment analysis results. The function for maximizing the accuracy of influence assessment is specifically expressed as follows: ; In the formula, This is the accuracy value for impact assessment; , , The weighting parameters are set to 0.35, 0.40, and 0.25 respectively. The propagation speed index; This is the influence range coefficient; Weighted by attention level; The intensity of emotional polarity; This is the time decay factor; To prevent extremely small constants from being divided by zero, a value of 0.001 is used. The specific expression of the resource consumption minimization function is as follows: ; In the formula, To calculate resource consumption values; , , The weighting parameters are set to 0.4, 0.35, and 0.25 respectively. This represents the percentage of network bandwidth utilization. This represents the percentage of storage space utilization. The parameter is obtained as follows: The data was obtained through an experimental method, including step 1: counting the number of reposts, comments, and likes per unit time; and step 2: calculating the growth rate of each indicator and performing a weighted average. The data was obtained through an experimental approach, including step 1: counting the number of users participating in the discussion; and step 2: calculating the geographical distribution breadth and platform coverage. The data was obtained through an experimental approach, including step 1: obtaining search popularity data; and step 2: statistically analyzing the frequency of media reports and performing normalization processing. The confidence score is calculated from the sentiment analysis model output in step S04. The formula is calculated by the difference between the time the public opinion event occurred and the current time. ,in This is the current timestamp, in seconds. This is the timestamp of the public opinion event, in seconds. The time normalization constant is set to 86400 seconds. The terms in the influence assessment accuracy maximization function achieve dimensional uniformity through the normalization of weight coefficients. This makes the function output These are dimensionless values. The terms in the resource consumption minimization function are unified through dimension adjustment of the weighting coefficients. The dimensions are , The dimensions are , The dimensions are This makes the function output It has a unified unit of measurement for resource consumption.

[0054] In this step, as an optional setting, since sufficient network bandwidth and storage space are prepared in advance, the default setting for network bandwidth utilization and storage space utilization is 1. When significant network lag or hard drive read / write slowdowns occur, you can determine based on experience whether to cancel the setting of 1 and use the actual values ​​of network bandwidth utilization and storage space utilization instead.

[0055] The specific implementation of step S06 is to establish a real-time monitoring and early warning mechanism based on threshold judgment. The specific judgment conditions for abnormal public opinion heat detection are as follows: ; ; ; In the formula, This refers to the volatility of public opinion. This represents the current level of public opinion attention. This serves as the baseline for public opinion sentiment intensity. This indicates a deviation in the intensity of emotional polarity. This represents the current intensity of emotional polarity. Baseline emotional polarity intensity; The warning levels are 0 for normal, 1 for single abnormality, and 2 for double abnormality.

[0056] It should be explained that the cross-modal attention mechanism calculation formula achieves the quantitative calculation of semantic relevance between different modalities through linear transformation and dot product operation, where the softmax normalization function is expressed as: ; In the formula, For the first One input element; For the first One input element; The total number of elements in the input vector; Compared to traditional fixed-weight fusion methods, this mechanism can dynamically adjust the importance of each modality based on the semantic content of the input data, significantly improving the adaptability and accuracy of multimodal information fusion, especially demonstrating stronger robustness when processing public opinion data with semantic imbalance between modalities. The gating fusion unit processing formula uses the sigmoid function to generate the gating signal, which is expressed as: ; In the formula, The input value is the sigmoid function; information filtering is achieved through element-wise multiplication, combined with the tanh nonlinear transformation function: ; In the formula, This is the input value for the tanh function; By maintaining gradient propagation stability through residual connections, this mechanism effectively filters redundant information and retains key features compared to simple feature stitching methods, avoiding information redundancy and noise interference problems in multimodal fusion. The attention head number adjustment function dynamically controls model complexity by comprehensively considering system operating parameters. Compared to fixed parameter configurations, this function optimizes computational resource consumption while ensuring analytical accuracy, achieving an intelligent balance between model performance and system efficiency. The influence assessment accuracy maximization function employs a logarithmic transformation. ; Exponential weighted transformation: ; and reciprocal transformation: ; The combination of these functions fully considers the nonlinear characteristics and time decay effect of public opinion dissemination. Compared with the linear weighting method, this function can more accurately capture the dynamic changes in public opinion influence, improving the accuracy and reliability of heat assessment. The resource consumption minimization function is achieved through squared terms: ; Product term: ; Sum of square roots: ; The weighted combination reflects the consumption characteristics and interrelationships of different resource types. Compared to a simple linear summation method, this function can more realistically reflect the actual consumption of system resources, providing a more accurate objective function basis for game optimization. The formula for detecting and judging abnormal public opinion trends employs a combination of relative change rate and absolute difference judgment mechanism. Through a tiered early warning strategy, it achieves accurate identification of different degrees of anomalies. Compared to a single threshold judgment method, this mechanism can reduce false alarms and missed alarms, improving the reliability and practicality of the early warning system.

[0057] To better understand and implement this invention, the following is a specific application scenario of Example 2: A technical team, addressing the need for monitoring public opinion during online emergencies, adopted the cross-platform public opinion information collection and analysis system of this invention for real-time public opinion analysis. The team needed to perform millisecond-level real-time monitoring of public opinion information on mainstream social media platforms such as Weibo, Douyin, Zhihu, and Tieba, especially to quickly identify public opinion hotspots and conduct sentiment analysis during emergencies.

[0058] In step S01, the technical team deployed a multi-platform data collector, connecting to various social media platforms via API interfaces, setting the data scraping frequency to 100 requests per second, and establishing a capacity of... The raw data cache pool for each data item. The multimodal data acquired by the collector includes text content, image files, video clips, and audio recordings, with text data accounting for 65%, image data accounting for 25%, video data accounting for 7%, and audio data accounting for 3%. Detailed metadata information is recorded during the data collection process, including publication timestamps, user identifiers, geographic location coordinates, and platform source identifiers.

[0059] In step S02, preprocessing operations are performed on the multimodal data in the original data cache pool. During data cleaning, duplicate data, junk information, and illegal content are removed, achieving a data purification rate of 92%. Format standardization processes unify text to UTF-8 encoding, images to RGB format, and audio to WAV format with a sampling rate of 16kHz. Noise filtering uses a signal-to-noise ratio threshold of 15dB for audio denoising, and a Gaussian filter with a kernel size of 5×5 for image denoising. Missing value imputation uses interpolation to fill in missing timestamps and uses the mean to fill in missing numerical metadata.

[0060] In step S03, a multimodal deep fusion convolutional algorithm is used for feature extraction. The text branch convolutional neural network uses a word embedding dimension of 256, with convolutional kernel sizes of 3, 4, and 5, each with 128 kernels, and a maximum pooling stride of 2. The image branch network uses a ResNet-50 architecture, with the input image size adjusted to 224×224 pixels and the feature map channel number of 512. The audio branch network uses one-dimensional convolution, with a window length of 25ms, an overlap rate of 50%, and a Mel-spectrum coefficient dimension of 80. During the cross-modal attention mechanism calculation, the query matrix, key matrix, and value matrix are all set to a dimension of 512, and the attention weight matrix, after softmax normalization, has a value range between 0 and 1. The average gate signal value generated by the gated fusion unit is 0.73, indicating that most feature information is effectively preserved. The unified semantic representation vector output by the fully connected layer has a dimension of 1024.

[0061] In step S04, the unified semantic representation vector is input into the cross-modal sentiment analysis and recognition model. This model is based on the Transformer architecture, containing a 12-layer encoder, with 8 attention heads per layer and a hidden layer dimension of 768. The model dynamically adjusts the number of attention heads based on system operating parameters. Currently, the system's CPU utilization is 35%, memory usage is 45%, and GPU memory usage is 2500MB, with a comprehensive performance score of 0.52; therefore, a 12-head attention head configuration is adopted. The sentiment analysis results output a positive sentiment confidence score of 0.82, a negative sentiment confidence score of 0.15, and a neutral sentiment confidence score of 0.03, ultimately determining a positive sentiment tendency.

[0062] In step S05, a two-layer game optimization model is used to calculate the public opinion heat index. The input parameters of the upper-layer model include a propagation speed index of 3.2, an influence range coefficient of 1.8, a attention weight of 0.75, an emotional polarity intensity of 0.82, and a time decay factor of 0.95. The input parameters of the lower-layer model include CPU utilization of 35%, memory utilization of 45%, network bandwidth utilization of 25%, and storage space utilization of 60%. After solving the game equilibrium, the influence assessment accuracy is 0.89, the computational resource consumption is 2.3, and the final generated public opinion heat index is 7.6.

[0063] In step S06, the real-time monitoring and early warning mechanism performs threshold judgment on the public opinion heat value. The preset heat threshold is 8.0, and the current heat value of 7.6 does not exceed the threshold, so the system maintains normal monitoring status. However, the sentiment polarity intensity deviation value is 0.12, which exceeds the threshold range of ±0.4, so the sentiment offset adjuster is activated. The adjuster fine-tunes the weight parameters of the sentiment analysis model, and the corrected sentiment polarity intensity is recalculated to be 0.78.

[0064] like Figure 2 As shown, the real-time changes in system hardware monitoring parameters display the dynamic trends of key indicators such as CPU core temperature and memory usage. Monitoring data indicates that the CPU core temperature remains at 42℃, below the optimization threshold of 45℃, and the system automatically increases the brightness of the CPU performance regulator. The video memory temperature remains at 72℃, below the 75℃ cooling activation threshold, and video memory cooling remains in standby mode. The average memory access latency is 85ns, the memory bandwidth utilization is 12GB / s, the video memory usage is 2500MB, the network latency is 15ms, the disk I / O rate is 450MB / s, and the power consumption is 280W.

[0065] The technical team established a detailed performance monitoring table during the implementation process, as shown in Table 1.

[0066] Table 1 System Operation Performance Monitoring Data Table

[0067] Table 1 records in detail the real-time monitoring data of various hardware parameters during system operation. All parameters are within the normal range, ensuring the stable operation of the system.

[0068] During the dynamic adjustment of the number of attention heads, the system optimizes parameters based on the comprehensive performance score. As shown in Table 2, the attention head configuration strategies corresponding to different performance score ranges demonstrate the working principle of the adaptive adjustment mechanism.

[0069] Table 2 Dynamic Adjustment Configuration Table for Number of Attention Points

[0070] Table 2 shows that the system selects a 12-head attention configuration based on the current performance score of 0.52, with an expected inference time of 180ms, which meets the millisecond-level response requirement.

[0071] like Figure 3 As shown, the distribution of feature weights for each modality during multimodal feature fusion reflects the effectiveness of the cross-modal attention mechanism. The text feature weight is 0.45, the image feature weight is 0.35, and the audio feature weight is 0.20, indicating that text information carries the most significant emotional semantics in the current public opinion content, image information provides important emotional enhancement, and audio information plays a supporting verification role. The processing results of the gated fusion unit show that the average activation value of the sigmoid-gated signal is 0.73, the amplitude of the feature vector after tanh nonlinear transformation ranges from -0.95 to 0.98, and the residual connection maintains the stability of gradient propagation.

[0072] In solving the two-layer game optimization model, the upper-layer model maximizes the objective function of influence assessment accuracy, resulting in a logarithmic product of the propagation speed index and the influence range coefficient of 1.75, an exponentially weighted sum of attention weight and emotional polarity intensity of 1.34, and a reciprocal transformation of the time decay factor of 1.05. The lower-layer model minimizes the objective function of computational resource consumption, resulting in a CPU utilization square of 0.12, a memory utilization rate and network bandwidth utilization product of 0.11, and a storage space utilization square root of 0.77. The game equilibrium solution process achieves mutual constraints between the two objective functions through resource allocation coupling terms, ultimately reaching an equilibrium solution with a public opinion heat value of 7.6.

[0073] The real-time monitoring and early warning mechanism demonstrated good responsiveness during operation. When the sentiment polarity deviation exceeded the threshold, the system activated the sentiment shift adjuster within 50ms. The adjuster fine-tuned the model parameters using a gradient descent algorithm with a learning rate of 0.0001, converging to the optimal parameter configuration after 5 iterations. The adjusted sentiment analysis results showed that the confidence level for positive sentiment was adjusted to 0.78, the confidence level for negative sentiment was 0.18, and the confidence level for neutral sentiment was 0.04. The sentiment polarity deviation was reduced to 0.08, meeting the system stability requirements.

[0074] Throughout the implementation process, the system processed a large-scale multimodal public opinion dataset containing 50,000 text data entries, 12,000 image data entries, 3,500 video data entries, and 1,500 audio files. Data preprocessing took an average of 15ms per data entry, feature extraction took an average of 85ms per data entry, sentiment analysis took an average of 65ms per data entry, and popularity calculation took an average of 25ms per data entry. The overall processing time per data entry was controlled within 190ms, meeting the technical requirements for millisecond-level real-time processing.

[0075] The technological advancements brought by this invention compared to traditional methods are mainly reflected in three aspects. First, a unified semantic representation of heterogeneous data is achieved through a multimodal deep fusion convolution algorithm, solving the problem of incomplete information in traditional single-modal analysis methods. The cross-modal attention mechanism can automatically learn the correlation weights between different modalities, and the gated fusion unit ensures the preservation of key features and the filtering of redundant information, making the sentiment analysis results more accurate and robust. Second, a dynamic model optimization mechanism based on the system's operating state achieves an adaptive balance between inference speed and analysis accuracy. The attention head number adjustment function optimizes the model complexity in real time according to the hardware resource status, avoiding resource waste or performance bottlenecks caused by traditional fixed parameter configurations, and ensuring stable and efficient operation of the system under different operating environments. Finally, the two-layer game optimization model achieves the optimal trade-off between influence assessment accuracy and computational resource consumption through the collaborative optimization of the upper and lower layer objective functions. Compared with traditional single optimization objective methods, the game equilibrium mechanism can maximize analysis accuracy while meeting real-time constraints, and minimize resource consumption while ensuring accuracy requirements, significantly improving the overall performance and practicality of the system.

[0076] It should be noted that the variables involved in this invention are explained in detail in Table 3 below.

[0077] Table 3. Variable Explanation Table

[0078] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-platform public opinion information collection and analysis system based on deep learning, characterized in that, The application relates to a multi-modal public opinion monitoring method based on multi-platform data collector. The method comprises the following steps: deploying a multi-platform data collector to perform real-time data crawling on a social media platform, obtaining multi-modal public opinion data through an API interface and a webpage crawler technology, and establishing an original data cache pool; preprocessing the multi-modal data in the original data cache pool; adopting a multi-modal deep fusion convolution algorithm to extract features of the preprocessed multi-modal data, calculating feature weights of each mode through a cross-modal attention mechanism, using a gate fusion unit to integrate different modal information to generate a unified semantic representation vector; inputting the unified semantic representation vector into a cross-modal sentiment analysis and recognition model to perform sentiment polarity analysis, and outputting positive, negative and neutral sentiment classification results and corresponding confidence scores; adopting a double-layer game optimization model to calculate public opinion heat based on the sentiment analysis result, wherein an upper-layer model takes maximizing public opinion influence evaluation accuracy as a target, a target function is an influence evaluation accuracy maximization function, and inputs include a propagation speed index, an influence range coefficient, an attention weight, a sentiment polarity intensity and a time attenuation factor; a lower-layer model takes minimizing calculation resource consumption as a target, a target function is a calculation resource consumption minimization function, and inputs include CPU usage, memory occupancy, network bandwidth occupancy and storage space occupancy. A real-time monitoring and early warning mechanism is established, and a warning signal is automatically triggered when the public opinion heat value exceeds a preset threshold. The multi-modal deep fusion convolution algorithm comprises the following steps: establishing a text branch convolutional neural network, an image branch convolutional neural network and an audio branch convolutional neural network to extract features of corresponding modal data respectively; calculating a correlation weight matrix among text features, image features and audio features through a cross-modal attention mechanism; weighting and fusing the weight matrix and the modal feature vectors; using a gate fusion unit to perform nonlinear transformation and information screening on the fused multi-modal features; and mapping the processed features to a fixed-dimension unified semantic representation vector through a full connection layer. 2.The deep learning-based cross-platform public opinion information collection and analysis system according to claim 1, characterized in that, The calculation steps of the cross-modal attention mechanism are as follows: linearly transforming a text feature vector, an image feature vector and an audio feature vector to obtain a query matrix, a key matrix and a value matrix respectively; calculating a dot product of the query matrix and the key matrix, and normalizing the dot product through a softmax function to obtain attention weights; multiplying the attention weights and the value matrix to obtain weighted feature representation; and splicing and reducing the dimension of the weighted feature representation among different modes. 3.The deep learning-based cross-platform public opinion information collection and analysis system according to claim 2, characterized in that, The processing mechanism of the gate fusion unit is as follows: performing sigmoid activation function transformation on the input multi-modal fusion features to generate a gate signal; performing element-level multiplication operation on the gate signal and the original fusion features to realize information screening; performing nonlinear transformation on the screened features through a tanh activation function to enhance expression capability; and adding the transformed features and a residual connection to maintain gradient propagation stability. 4.The deep learning-based cross-platform public opinion information collection and analysis system according to claim 3, characterized in that, The structure of the double-layer game optimization model is as follows: the upper-layer model takes maximizing public opinion influence evaluation accuracy as a target, a target function is an influence evaluation accuracy maximization function, and inputs include a propagation speed index, an influence range coefficient, an attention weight, a sentiment polarity intensity and a time attenuation factor; the lower-layer model takes minimizing calculation resource consumption as a target, a target function is a calculation resource consumption minimization function, and inputs include CPU usage, memory occupancy, network bandwidth occupancy and storage space occupancy. 5.The cross-platform public opinion information collection and analysis system based on deep learning according to claim 4, characterized in that, ​ 6.The cross-platform public opinion information collection and analysis system based on deep learning according to claim 5, wherein, The trigger condition of the real-time monitoring and early warning mechanism is specifically that if the fluctuation range of the public opinion heat value exceeds ± 30% and the deviation of the emotional polarity intensity exceeds ± 0.4, the heat abnormality double regulator is started; if only the public opinion heat value fluctuation exceeds ± 30%, only the heat fluctuation regulator is started; and if only the deviation of the emotional polarity intensity exceeds ± 0.4, only the emotional deviation regulator is started. 7.The cross-platform public opinion information collection and analysis system based on deep learning according to claim 6, characterized in that, The cross-modal sentiment analysis and recognition model refers to a multi-layer neural network based on a Transformer architecture, including an input embedding layer, a multi-head self-attention mechanism layer, a feedforward neural network layer, a layer normalization module and an output classification layer, wherein the number of heads of the multi-head attention mechanism is dynamically adjusted according to three system running parameters of CPU usage, memory occupancy and video memory occupancy. 8.The cross-platform public opinion information collection and analysis system based on deep learning according to claim 7, wherein, The head number dynamic adjustment mechanism is specifically that when the CPU usage is in the interval [10%, 40%), the number of heads is set to 8, when the memory occupancy is in the interval [20%, 60%), the number of heads is set to 12, and when the video memory occupancy is in the interval [1000MB, 4000MB], the number of heads is set to 16. 9.The cross-platform public opinion information collection and analysis system based on deep learning according to claim 8, characterized in that, The influence evaluation accuracy maximization function is used to calculate the objective function value of the upper game model to maximize the evaluation accuracy of public opinion influence, and the input includes the propagation speed index, the influence range coefficient, the attention weight, the emotional polarity intensity and the time decay factor. The output is the influence evaluation accuracy value, and the final influence evaluation accuracy value is obtained by comprehensive operation of the logarithmic product, exponential weighted sum and reciprocal transformation of each input parameter. 10.The cross-platform public opinion information collection and analysis system based on deep learning according to claim 9, wherein, The attention head number adjustment function refers to an adaptive parameter control function based on the system running state, which is used to adjust the attention head number parameter of the multi-head attention mechanism in the cross-modal sentiment analysis and recognition model. A comprehensive performance score value is calculated based on three key system parameters of CPU usage, memory occupancy and video memory occupancy. When the comprehensive performance score value belongs to different intervals, different head number attention configurations are adopted.

Images