A smart city comprehensive security information management platform

Abstract

The present invention discloses a smart city comprehensive security information management platform, which relates to the field of information technology and aims to solve the complexity of multimodal data processing and the challenge of energy management in urban security. The platform integrates multimodal data fusion technology, combines deep learning and attention mechanism, extracts emotional features from video, audio and text data, and realizes accurate emotional classification and behavior prediction. Through the emotional analysis module, the platform can generate security warning information to improve the foresight and responsiveness of public safety. At the same time, the platform optimizes task allocation and computing resources through load balancing to reduce energy consumption. The edge computing module dynamically adjusts computing resources according to the current load and environmental conditions to ensure the efficient operation of the platform. The environmental perception module monitors temperature and humidity data in real time to further optimize the allocation of computing resources. The present invention improves the level of public safety management and reduces environmental impact through these technologies.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a smart city comprehensive security information management platform. Background Art

[0002] With the development of smart cities, urban security monitoring systems need to process large amounts of video, audio, and text data to identify and predict individual emotional states and behaviors, which is essential for preventing crime, improving public safety, and responding to emergencies. However, how to accurately extract emotional features from these multimodal data and perform effective classification and prediction is a technical challenge. At the same time, as the number and computing needs of edge computing devices continue to increase, how to effectively manage and optimize the energy consumption of these devices to reduce overall operating costs and environmental impact is also an important technical issue.

[0003] Therefore, a smart city comprehensive security information management platform is proposed. By integrating multimodal data fusion module and sentiment analysis module, it can extract features from video, audio and text data, and generate fusion feature vectors through attention mechanism and feature alignment technology, and then perform sentiment classification and behavior prediction to generate security warning information. In addition, the platform's edge computing module and load balancing module can dynamically adjust computing resources, optimize task allocation and resource utilization, thereby reducing energy consumption and improving the energy efficiency and response speed of the system. Summary of the invention

[0004] The purpose of this invention is to solve the problems in the prior art of extracting emotional features from different types of data, making accurate classification and prediction, and how to optimize the energy usage of these devices as the number of deployed edge computing devices and computing needs increase, and to propose a smart city comprehensive security information management platform.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A smart city comprehensive security information management platform, comprising:

[0007] Multimodal data fusion module: extracts features from video, audio and text data, and generates a fused feature vector through the attention mechanism and feature alignment technology. The attention mechanism unit splices the video, audio and text feature vectors into a multimodal feature matrix, calculates the attention score of each modal feature, and applies the Softmax function to normalize the attention score to obtain the attention weight with a dimension of (1,3). The fusion layer unit uses the attention weight to weight the multimodal features and generates the final fused feature vector.

[0008] Sentiment analysis module: receives the fusion feature vector generated by the multimodal data fusion module, uses the deep learning model to perform sentiment classification and behavior prediction, and generates safety warning information. This is achieved through the following steps:

[0009] 1. Feature extraction: Use deep learning models to extract emotional features;

[0010] 2. Sentiment classification: Use the pre-trained BERT model to classify sentiment features, output sentiment labels with a dimension of 3 (positive, negative, neutral), and select the category with the highest probability as the final sentiment label. For each sentiment label (in can be positive, negative, or neutral), probability Given by: , is the raw score in the output layer of the BERT model for each sentiment category (positive, negative, neutral);

[0011] 3. Behavior prediction: Using a multi-task learning framework, sharing the underlying feature extraction layer, the output behavior prediction result dimension is 10 (different behavior categories);

[0012] 4. Safety warning: Combine emotion classification and behavior prediction results to generate detailed safety warning information;

[0013] Load balancing module: monitors the status of each security node, dynamically allocates tasks, and predicts future resource utilization to ensure efficient operation of the platform;

[0014] Edge computing module: dynamically adjusts computing resources according to current load and environmental conditions to optimize energy efficiency and response speed;

[0015] Environmental perception module: collects temperature and humidity data and feeds it back to the edge computing module to optimize the allocation of computing resources.

[0016] Preferably, the multimodal data fusion module includes:

[0017] Feature extraction unit: Use the pre-trained ResNet-50 model to extract the features of video data, and the output feature vector dimension is 2048; use the pre-trained convolutional neural network (CNN) to extract the features of audio data, and the output feature vector dimension is 512; use the pre-trained BERT model to extract the features of text data, and the output feature vector dimension is 768; the extracted features are video feature vector, audio feature vector and text feature vector respectively;

[0018] Attention mechanism unit: concatenate video, audio and text feature vectors into a multimodal feature matrix with a dimension of (1,3328); calculate the attention score of each modal feature, and output the attention score with a dimension of (1,3); apply the Softmax function to normalize the attention score to obtain the attention weight with a dimension of (1,3);

[0019] Fusion layer unit: Use attention weights to weight multimodal features and generate the final fused feature vector with dimension (1,3328).

[0020] Preferably, the multimodal data fusion module further includes further processing the fused feature vector:

[0021] Feature alignment unit: uses linear transformation to align features of different modalities to the same feature space, and the dimension of the output aligned feature vector is (1,3328); dynamic feature selection unit: uses recursive feature elimination (RFE) algorithm to evaluate the importance of each feature, and the output feature importance score dimension is (1,3328); calculates the correlation matrix of the feature matrix, removes features with correlation exceeding the threshold of 0.8, and finally retains the feature vector dimension of (1,1664).

[0022] Preferably, the sentiment analysis module includes:

[0023] Emotional feature extraction: read the final retained feature vector dimension as (1,1664); use the LSTM model to extract emotional features, and output the emotional feature vector dimension as 128; normalize the emotional feature vector, and limit the value range to between 0 and 1; perform PCA dimensionality reduction to reduce the feature dimension to 64;

[0024] Sentiment classification: read the sentiment feature vector; use the pre-trained BERT model for classification, and output the sentiment label dimension of 3, namely positive, negative, and neutral, and the probability corresponding to each label is P1, P2, and P3 respectively; select the category with the highest probability as the final sentiment label;

[0025] Multi-task parallel processing unit: Build a multi-task learning framework and share the underlying feature extraction layer. The shared underlying feature extraction layer extracts common features through a pre-trained deep learning model and outputs a feature vector dimension of 2048. The behavior prediction task subnetwork outputs behavior prediction results with a dimension of 10 (different behavior categories), where:

[0026] receiving a fused feature vector or an output of a shared feature extraction layer;

[0027] Processed by a behavior prediction subnetwork specifically designed to recognize different behavior patterns;

[0028] Output behavior prediction results, with a dimension of 10, representing the original output values ​​of different behavior categories ;

[0029] The original output value The activation function converts it into a probability value to indicate the probability that the input sample belongs to each behavior category.

[0030] Preferably, the sentiment analysis module further includes:

[0031] A training and updating unit, used to regularly update the sentiment classification model;

[0032] Behavior analysis and early warning unit, used to discover the association rules between emotional states and behaviors.

[0033] Preferably, the sentiment analysis module further includes:

[0034] Anomaly detection and early warning unit, which is used to calculate anomaly scores of behavioral data;

[0035] The future behavior prediction unit uses the LSTM model to predict future behavior trends.

[0036] Preferably, the sentiment analysis module further includes:

[0037] A safety warning information generating unit, used to generate safety warning information;

[0038] Data collection and privacy protection unit to protect personal privacy.

[0039] Preferably, the sentiment analysis module further includes:

[0040] User interface and data usage instructions to ensure that users understand the purpose and method of data use.

[0041] Preferably, the load balancing module includes:

[0042] Status monitoring unit, used to record the status of security nodes;

[0043] Load distribution unit, which uses Q-learning algorithm to optimize task distribution;

[0044] Prediction model unit, using LSTM model to predict CPU utilization;

[0045] Failover unit, used for task redistribution when a node fails.

[0046] Preferably, the edge computing module includes:

[0047] Dynamic Voltage Frequency Scaling (DVFS) unit for adjusting CPU frequency and voltage;

[0048] Sleep mode management unit to save power;

[0049] The environment sensing unit is used to monitor environmental changes and adjust computing resources.

[0050] The present invention has the following beneficial effects:

[0051] 1. In the present invention, through the multimodal data fusion module, it is possible to efficiently process and analyze massive video, audio and text data from all corners of the city. Using a deep learning model combined with an attention mechanism and feature alignment technology, the platform can accurately extract and fuse key emotional features, quickly identify the emotional state of individuals, and provide real-time emotional analysis and behavior prediction for urban security, thereby effectively preventing crime, improving public safety, and responding to emergencies in a timely manner. In addition, the platform's cross-cultural adaptability unit further enhances the wide applicability of sentiment analysis and ensures the accuracy of analysis in different cultural backgrounds.

[0052] 2. In the present invention, the edge computing module is used to realize intelligent energy management of edge computing devices. The platform can dynamically adjust computing resources according to real-time load and environmental conditions to optimize the energy efficiency and response speed of the device. In addition, through load balancing and environmental perception technology, the platform further reduces the energy consumption of the device, reduces operating costs, and mitigates the impact on the environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is an architecture diagram of a smart city comprehensive security information management platform proposed by the present invention;

[0054] Figure 2 This is a data flow chart of a smart city comprehensive security information management platform in the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 creative work are within the scope of protection of the present invention.

[0056] In one embodiment, Figure 1-Figure 2 As shown, the present invention proposes a smart city comprehensive security information management platform, including:

[0057] Multimodal data fusion module: extracts features from video, audio and text data, and generates a fused feature vector through the attention mechanism and feature alignment technology. The attention mechanism unit splices the video, audio and text feature vectors into a multimodal feature matrix, calculates the attention score of each modal feature, and applies the Softmax function to normalize the attention score to obtain the attention weight with a dimension of (1,3). The fusion layer unit uses the attention weight to weight the multimodal features and generates the final fused feature vector.

[0058] Sentiment analysis module: receives the fusion feature vector generated by the multimodal data fusion module, uses the deep learning model to perform sentiment classification and behavior prediction, and generates safety warning information. This is achieved through the following steps:

[0059] 1. Feature extraction: Use deep learning models to extract emotional features;

[0060] 2. Sentiment classification: Use the pre-trained BERT model to classify sentiment features, output sentiment labels with a dimension of 3 (positive, negative, neutral), and select the category with the highest probability as the final sentiment label. For each sentiment label (in can be positive, negative, or neutral), probability Given by: , is the raw score in the output layer of the BERT model for each sentiment category (positive, negative, neutral);

[0061] 3. Behavior prediction: Using a multi-task learning framework, sharing the underlying feature extraction layer, the output behavior prediction result dimension is 10 (different behavior categories);

[0062] 4. Safety warning: Combine emotion classification and behavior prediction results to generate detailed safety warning information;

[0063] Load balancing module: monitors the status of each security node, dynamically allocates tasks, and predicts future resource utilization to ensure efficient operation of the platform;

[0064] Edge computing module: dynamically adjusts computing resources according to current load and environmental conditions to optimize energy efficiency and response speed;

[0065] Environmental perception module: collects temperature and humidity data and feeds it back to the edge computing module to optimize the allocation of computing resources.

[0066] Multimodal data fusion module

[0067] 1. Feature extraction unit

[0068] Video feature extraction: Use the pre-trained ResNet-50 model to extract the features of video data, and the output feature vector dimension is 2048.

[0069] For example, a camera captures video data inside a large shopping mall. The video frame size is 224x224 and the number of channels is 3. The ResNet-50 model is used to extract the features of each frame image. After removing the last fully connected layer, the feature vector dimension extracted is 2048. This step can efficiently extract meaningful visual features from the video for subsequent sentiment analysis and behavior prediction.

[0070] Audio feature extraction: A pre-trained convolutional neural network (CNN) is used to extract the features of audio data, and the output feature vector dimension is 512.

[0071] For example, in the same scenario, the microphone records audio data. The size of the audio spectrogram is 128x128, and the number of channels is 1. The features of the spectrogram are extracted using a pre-trained CNN model, and the dimension of the extracted feature vector is 512. This step can extract the features of the sound from the audio, which is used to identify emotions and behaviors in the speech.

[0072] Text feature extraction: Use the pre-trained BERT model to extract the features of text data, and the output feature vector dimension is 768.

[0073] For example, if there is text data in the scene (such as social media posts or chat records), the text is encoded using the tokenizer of the BERT model with a maximum length of 128. The encoded text features are extracted using the BERT model, and the dimension of the extracted feature vector is 768. This step can extract semantic features from the text for sentiment analysis.

[0074] 2. Attention Mechanism Unit

[0075] Feature concatenation: Concatenate the video feature vector, audio feature vector, and text feature vector into a large matrix by column with a dimension of (1, 3328).

[0076] For example, the video feature vector (2048 dimensions), audio feature vector (512 dimensions), and text feature vector (768 dimensions) extracted above are concatenated into a large matrix with a dimension of (1, 3328). This step integrates the features of different modalities together, providing a basis for subsequent fusion processing.

[0077] Attention score: A single-layer neural network is used to calculate the attention score of each modal feature. The activation function is ReLU and the output attention score dimension is (1, 3).

[0078] For example, define a single-layer neural network with an input dimension of 3328 and an output dimension of 3. Use the ReLU activation function to calculate the attention score, and the output attention score dimension is (1, 3). This step can determine the importance of different modal features and provide a basis for subsequent weighted fusion.

[0079] Normalization: Apply the Softmax function to normalize the attention scores to obtain the attention weights with a dimension of (1, 3).

[0080] For example, the Softmax function is used to normalize the attention scores, and the dimension of the normalized attention weights is (1, 3). This step ensures that the sum of the attention weights is 1, making the weighted fusion more reasonable.

[0081] Weighted fusion: Use attention weights to weight multimodal features and obtain a weighted feature matrix with a dimension of (1, 3328).

[0082] For example, multiplying the attention weight by the concatenated feature matrix yields a weighted feature matrix with a dimension of (1,3328). This step generates a feature vector that comprehensively considers multiple modal information through weighted fusion.

[0083] 3. Feature Alignment Unit

[0084] Linear transformation: Linear transformation is used to align the features of different modalities into the same feature space, and the dimension of the output aligned feature vector is (1, 3328).

[0085] For example, linear transformation is used to align video, audio, and text features to the same feature space, and the dimension of the output aligned feature vector is (1, 3328). This step ensures that features of different modalities are compared and processed in the same feature space.

[0086] Dynamic feature selection: The recursive feature elimination (RFE) algorithm is used to evaluate the importance of each feature, and the output feature importance score dimension is (1, 3328). The correlation matrix of the feature matrix is ​​calculated, and the features with correlation exceeding the threshold of 0.8 are removed. The final feature vector dimension retained is (1, 1664).

[0087] For example, the RFE algorithm is used to evaluate the importance of each feature, and the output feature importance score dimension is (1, 3328). The correlation matrix of the feature matrix is ​​calculated, and the features with correlation exceeding 0.8 are removed. The final feature vector dimension retained is (1, 1664). This step removes redundant features, improves the purity of features and the efficiency of the model.

[0088] Sentiment Analysis Module

[0089] 1. Emotional feature extraction

[0090] Read feature vector: The dimension of the feature vector finally retained is (1, 1664).

[0091] For example: Read the above retained feature vector, the dimension is (1, 1664). This step provides input data for sentiment feature extraction.

[0092] Use LSTM model to extract sentiment features: the output sentiment feature vector dimension is 128.

[0093] For example, use the pre-trained LSTM model to extract sentiment features and output a sentiment feature vector with a dimension of 128. This step can extract sentiment trends from time series data.

[0094] Normalization: Normalize the sentiment feature vector and limit the value range to between 0 and 1.

[0095] For example, the sentiment feature vector is normalized and its value range is limited to between 0 and 1. This step ensures that the value range of the sentiment feature is consistent, which facilitates subsequent processing and analysis.

[0096] PCA dimensionality reduction: Reduce feature dimensions to 64.

[0097] Example: Perform PCA dimensionality reduction to reduce the feature dimension to 64. This step reduces the feature dimension and improves the computational efficiency of the model.

[0098] 2. Sentiment Classification

[0099] Read the sentiment feature vector: Use the pre-trained BERT model for classification and output the sentiment label dimension as 3 (positive, negative, neutral).

[0100] For example: read the sentiment feature vector, use the pre-trained BERT model for classification, and output the sentiment label dimension as 3 (positive, negative, neutral). The probabilities corresponding to each label are P1, P2, and P3 respectively. Select the category with the highest probability as the final sentiment label. This step can accurately identify the sentiment category.

[0101] 3. Behavior prediction

[0102] Multi-task parallel processing unit: Build a multi-task learning framework, share the underlying feature extraction layer, and output feature vector dimension of 2048. The behavior prediction task subnetwork outputs behavior prediction result dimension of 10 (different behavior categories).

[0103] For example: build a multi-task learning framework, share the underlying feature extraction layer, and output feature vector dimensions of 2048. The behavior prediction task subnetwork receives the fused feature vector or the output of the shared feature extraction layer, processes it through the behavior prediction subnetwork, and outputs the behavior prediction result dimension of 10 (different behavior categories). This step can identify different behavior patterns and provide a basis for safety warnings.

[0104] 3. Load balancing module

[0105] 1. Status monitoring unit

[0106] Record status data: Record the CPU utilization and network latency of each security node every 10 minutes, use scripts to obtain status data from the nodes, and store it in the central database.

[0107] For example, the CPU utilization and network latency of each security node are recorded every 10 minutes, and the status data is obtained from the node using a script and stored in a central database. This step can monitor the status of each node in real time and provide data support for dynamic task allocation.

[0108] 2. Load distribution unit

[0109] Optimize task allocation: Use the Q-learning algorithm to optimize task allocation, calculate task completion time and resource utilization, and update the Q value.

[0110] For example, using the Q-learning algorithm to optimize task allocation, calculate task completion time and resource utilization, and update the Q value. This step can dynamically adjust task allocation to ensure efficient operation of the system.

[0111] 3. Prediction Model Unit

[0112] Predict future resource utilization: Use a two-layer LSTM model to predict CPU utilization within the next hour. If the predicted value exceeds 70%, assign tasks in advance to avoid resource bottlenecks.

[0113] For example, a two-layer LSTM model is used to predict the CPU utilization within the next hour. If the predicted value exceeds 70%, tasks are assigned in advance to avoid resource bottlenecks. This step can prevent resource shortages in advance and ensure system stability.

[0114] 4. Failover Unit

[0115] Heartbeat packet check: Each node sends a heartbeat packet to the central monitoring platform every 1 second. A scheduled task is used to check the heartbeat time. If no heartbeat packet is received within 5 seconds, the node is considered to be faulty and the task is reallocated to ensure the continuity and stability of the platform.

[0116] For example, each node sends a heartbeat packet to the central monitoring platform every 1 second, and a scheduled task is used to check the heartbeat time. If no heartbeat packet is received within 5 seconds, the node is considered to be faulty and the task is reallocated to ensure the continuity and stability of the platform. This step can detect and handle node failures in a timely manner to ensure the reliable operation of the system.

[0117] 4. Edge computing module

[0118] 1. Dynamic Voltage Frequency Scaling (DVFS) Unit

[0119] Real-time monitoring of CPU utilization: Read CPU utilization every second and monitor the current CPU utilization in real time. When CPU utilization is less than 30%, reduce the frequency to 800MHz; when CPU utilization is higher than 70%, adjust the frequency to 2.4GHz. Adjust the voltage accordingly to ensure energy efficiency and response speed.

[0120] For example, read the CPU usage rate every second and monitor the current CPU utilization rate in real time. When the CPU utilization rate is less than 30%, reduce the frequency to 800MHz; when the CPU utilization rate is higher than 70%, adjust the frequency to 2.4GHz. Adjust the voltage accordingly to ensure energy efficiency and response speed. This step can dynamically adjust computing resources to optimize energy efficiency and response speed.

[0121] 2. Sleep mode management unit

[0122] Check platform activity: Check platform activity every second and enter memory sleep mode after 10 seconds of inactivity. Set GPIO interrupt or RTC timer as wakeup source to ensure fast response.

[0123] For example: Check platform activity every second and enter memory sleep mode after 10 seconds of inactivity. Set GPIO interrupt or RTC timer as the wakeup source to ensure a quick response. This step can save power during low activity periods while ensuring a quick response when needed.

[0124] 3. Environmental perception unit

[0125] Collect environmental data: Environmental sensors send temperature and humidity data every 1 minute to monitor environmental changes in real time. When the temperature exceeds 35°C or the humidity exceeds 60%, the CPU frequency is adjusted and the fan is started for heat dissipation. The temperature, humidity and adjustment results are recorded to ensure that the device operates within a safe range.

[0126] For example, the environmental sensor sends temperature and humidity data every 1 minute to monitor environmental changes in real time. When the temperature exceeds 35°C or the humidity exceeds 60%, the CPU frequency is adjusted and the fan is started for heat dissipation. The temperature, humidity and adjustment results are recorded to ensure that the device operates within a safe range. This step can ensure that the device operates normally in harsh environments and extend the life of the device.

[0127] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart city comprehensive security information management platform, characterized in that: include: Multimodal data fusion module: extracts features from video, audio and text data, and generates a fused feature vector through the attention mechanism and feature alignment technology. The attention mechanism unit splices the video, audio and text feature vectors into a multimodal feature matrix, calculates the attention score of each modal feature, and applies the Softmax function to normalize the attention score to obtain the attention weight with a dimension of (1,3). The fusion layer unit uses the attention weight to weight the multimodal features and generates the final fused feature vector. Sentiment analysis module: receives the fusion feature vector generated by the multimodal data fusion module, uses the deep learning model to perform sentiment classification and behavior prediction, and generates safety warning information. This is achieved through the following steps: Step 1: Feature extraction: Use deep learning models to extract emotional features; Step 2: Sentiment classification: Use the pre-trained BERT model to classify the sentiment features, output the sentiment label dimension as 3, and the sentiment categories include: positive, negative, and neutral. Select the category with the highest probability as the final sentiment label. For each sentiment label Is one of positive, negative, or neutral, with a probability of Given by: , is the raw score in the output layer of the BERT model for each sentiment category; Step 3: Behavior prediction: Use a multi-task learning framework, share the underlying feature extraction layer, and output behavior prediction results, including 10 different behavior categories with a dimension of 10; Step 4: Safety warning: Combine the emotion classification and behavior prediction results to generate detailed safety warning information; Load balancing module: monitors the status of each security node, dynamically allocates tasks, and predicts future resource utilization to ensure efficient operation of the platform. The load balancing module checks the node status in conjunction with the heartbeat packet and reallocates tasks when a node fails to ensure the reliability and continuity of the platform; Edge computing module: dynamically adjusts computing resources according to current load and environmental conditions, optimizes energy efficiency and response speed, and dynamically adjusts frequency and voltage according to CPU utilization to ensure a balance between energy efficiency and response speed; Environmental perception module: collects temperature and humidity data and feeds it back to the edge computing module to optimize the allocation of computing resources; The multimodal data fusion module includes: Feature extraction unit: Use the pre-trained ResNet-50 model to extract the features of video data, and the output feature vector dimension is 2048; use the pre-trained convolutional neural network CNN to extract the features of audio data, and the output feature vector dimension is 512; use the pre-trained BERT model to extract the features of text data, and the output feature vector dimension is 768; the extracted features are video feature vector, audio feature vector and text feature vector respectively; Attention mechanism unit: concatenate video, audio and text feature vectors into a multimodal feature matrix with a dimension of (1,3328); calculate the attention score of each modal feature, and output the attention score with a dimension of (1,3); apply the Softmax function to normalize the attention score to obtain the attention weight with a dimension of (1,3); Fusion layer unit: Use attention weights to weight multimodal features and generate the final fusion feature vector with a dimension of (1,3328); The multimodal data fusion module further includes further processing of the fused feature vector: Feature alignment unit: uses linear transformation to align features of different modalities to the same feature space, and the dimension of the output aligned feature vector is (1,3328); Dynamic feature selection unit: uses recursive feature elimination (RFE) algorithm to evaluate the importance of each feature, and the dimension of the output feature importance score is (1,3328); calculates the correlation matrix of the feature matrix, removes features with correlation exceeding the threshold of 0.8, and the dimension of the feature vector retained is (1,1664); The sentiment analysis module includes: Emotional feature extraction: read the final retained feature vector dimension as (1,1664); use the LSTM model to extract emotional features, and output the emotional feature vector dimension as 128; normalize the emotional feature vector, and limit the value range to between 0 and 1; perform PCA dimensionality reduction to reduce the feature dimension to 64; Sentiment classification: read the sentiment feature vector; use the pre-trained BERT model for classification, and output the sentiment label dimension of 3, namely positive, negative, and neutral, and the probability corresponding to each label is P1, P2, and P3 respectively; select the category with the highest probability as the final sentiment label; Multi-task parallel processing unit: Build a multi-task learning framework and share the underlying feature extraction layer. The shared underlying feature extraction layer extracts common features through a pre-trained deep learning model, and the output feature vector dimension is 2048; the behavior prediction task subnetwork outputs the behavior prediction result dimension of 10, and the behavior categories include 10 different behavior categories, among which: receiving a fused feature vector or an output of a shared feature extraction layer; Processed by a behavior prediction subnetwork specifically designed to recognize different behavior patterns; Output behavior prediction results, with a dimension of 10, representing the original output values ​​of different behavior categories ; The original output value Converted into probability values ​​through activation functions to represent the probability that the input sample belongs to each behavior category; The load balancing module includes: Status monitoring unit, used to record the status of security nodes; Load distribution unit, which uses Q-learning algorithm to optimize task distribution; Prediction model unit, using LSTM model to predict CPU utilization; Failover unit, used for task redistribution when a node fails; The edge computing module includes: Dynamic voltage and frequency scaling DVFS unit, used to adjust CPU frequency and voltage; Sleep mode management unit to save power; The environment sensing unit is used to monitor environmental changes and adjust computing resources.

2. A smart city comprehensive security information management platform according to claim 1, characterized in that: The sentiment analysis module also includes: A training and updating unit, used to regularly update the sentiment classification model; Behavior analysis and early warning unit, used to discover the association rules between emotional states and behaviors.

3. A smart city comprehensive security information management platform according to claim 1, characterized in that: The sentiment analysis module also includes: Anomaly detection and early warning unit, which is used to calculate anomaly scores of behavioral data; The future behavior prediction unit uses the LSTM model to predict future behavior trends.

4. A smart city comprehensive security information management platform according to claim 1, characterized in that: The sentiment analysis module also includes: A safety warning information generating unit, used to generate safety warning information; Data collection and privacy protection unit to protect personal privacy.

5. A smart city comprehensive security information management platform according to claim 1, characterized in that: The sentiment analysis module also includes: User interface and data usage instructions to ensure that users understand the purpose and method of data use.

Images