A tunnel safety intelligent response system

Abstract

This invention relates to the field of intelligent safety and tunnel engineering safety technology, specifically to an intelligent tunnel safety response system, comprising a multimodal data acquisition module, a data feature extraction module, a risk prediction module, and a response module. The multimodal data acquisition module collects multi-dimensional data such as temperature, humidity, and harmful gas concentration within the tunnel; the data feature extraction module uses an improved CNN to extract multi-scale features from images and combines Fourier transform to extract frequency domain features of structural strain; the risk prediction module constructs a model based on GRU combined with an attention mechanism and outputs the risk probability; the response module automatically initiates emergency measures such as ventilation adjustment based on the risk probability. The purpose of this invention is to solve the problems of traditional tunnel safety monitoring systems, such as limited data, coarse feature extraction, delayed risk prediction, and passive emergency response.

Description

Technical Field

[0001] This invention belongs to the field of railway intelligent safety and tunnel engineering safety technology, and relates to a tunnel safety intelligent response system. Background Technology

[0002] With the rapid development of transportation infrastructure construction, tunnels, as important transportation nodes, play a crucial role in the stability and reliability of the transportation network through their safe operation. Tunnels present complex environments with various safety hazards, including structural aging, fires, accumulation of harmful gases, and traffic accidents. Traditional manual inspections and simple sensor monitoring are no longer sufficient to meet the needs of modern tunnel safety management. Therefore, constructing intelligent and automated tunnel safety monitoring and response systems has become an important research direction for ensuring the safe operation of tunnels.

[0003] In practical applications of tunnel safety monitoring, existing technologies have many problems. On the one hand, data acquisition is limited in scope, with most systems relying on only a few sensor data points such as temperature, humidity, and harmful gas concentrations, failing to comprehensively reflect the tunnel's operational status. For example, they cannot promptly capture potential risks caused by subtle structural deformations or abnormal vehicle movement. On the other hand, data processing and analysis capabilities are insufficient. Traditional methods often employ threshold judgments or simple statistical analysis, lacking the ability to deeply mine complex data and dynamically predict risks, making it difficult to provide effective early warnings before accidents occur. Furthermore, emergency response lacks intelligence and coordination; once a safety incident occurs, the optimal response strategy cannot be quickly activated based on the risk level, resulting in low rescue efficiency.

[0004] A search revealed several existing tunnel safety monitoring and response solutions. For example, patent CN119007413B, titled "A Tunnel Equipment and Facility Status Monitoring and Early Warning System Based on Machine Learning," collects equipment and facility data through sensors distributed throughout the tunnel and uses deep learning models for status analysis and early warning, thus improving the accuracy of equipment failure monitoring to some extent. Patent CN114582126A, "An Intelligent Management and Control Method and System Applicable to Extra-Long Tunnels that Balances Traffic Efficiency and Safety," achieves abnormal status assessment and graded early warning based on traffic operation data and road surface condition data, optimizing tunnel traffic management. However, these solutions still have limitations: the former mainly focuses on equipment and facility status monitoring, lacking the fusion analysis of environmental and traffic data; the latter emphasizes traffic control, without addressing the comprehensive prediction of structural safety and environmental risks. Neither has constructed a complete closed-loop system from multimodal data acquisition and deep feature extraction to intelligent risk prediction and response, making it difficult to meet the needs of multi-source data collaborative processing and dynamic risk proactive prevention and control in complex tunnel environments. Therefore, there is an urgent need for a tunnel safety intelligent response system that integrates multimodal data, improved feature extraction algorithms, and intelligent prediction models to address the technical bottlenecks in existing technologies. Summary of the Invention

[0005] This invention provides a tunnel safety intelligent response system that solves the problems of traditional tunnel safety monitoring data being limited, feature extraction being coarse, risk prediction being lagging, and emergency response being passive.

[0006] To solve the above problems, the technical solution adopted by the invention is as follows: A tunnel safety intelligent response system, comprising: The multimodal data acquisition module is used to collect data on temperature and humidity, concentration of harmful gases, structural strain, video images, and vehicle speed inside the tunnel. The data feature extraction module employs an improved convolutional neural network (CNN) to extract features from video image data. This improved CNN adds a spatial pyramid pooling layer after the traditional convolutional layer, extracting multi-scale image features through pooling operations at different scales. The calculation formula is as follows:

[0007] in, For the first The first pooling layer Line number The pooling result of the column, This represents the number of pixels in the corresponding pooling region. For the original image in The pixel value of the location, For the first The pooling regions of each pooling layer are used; simultaneously, Fourier transform is employed to extract frequency domain features from the structural strain data, as shown in the formula: in, It is a frequency domain function. For time-domain structural strain data, Angular frequency; The risk prediction module constructs a risk prediction model based on a gated recurrent unit (GRU) combined with an attention mechanism. It takes extracted multimodal data features as input, and the attention mechanism assigns weights based on the contribution of features at different time steps to the current risk prediction. The specific calculation process is as follows:

[0008] in, Let GRU be the hidden state at time t, with dimension . , This is the hidden state dimension, with a value range of 128-512. Let be the input data features at time t, with dimension . , The input feature dimension is determined by the feature dimension obtained after fusing multimodal data. , Let b be the weight matrix and b be the bias vector. Let K be the attention score at time t, and K be the time step index (ranging from 1 to T). Let T be the attention score at the Kth time step, and T be the total number of time steps (i.e., the length of the input sequence). Let be the attention weight at time t. Let K be the attention weight at time step K. The hidden state is the weighted summation; this model outputs the probability of tunnel safety risks over a future period. The response module automatically initiates corresponding emergency response measures, including adjusting ventilation equipment, traffic guidance, and dispatching emergency rescue resources, when the probability of the safety risk output by the risk prediction module exceeds a preset threshold.

[0009] The principle and advantages of this scheme are as follows: A multimodal sensor network deployed within the tunnel collects real-time multidimensional data, including temperature and humidity, harmful gas concentrations, structural strain, video images, and vehicle speed. For different data types, an improved convolutional neural network (CNN) combined with a spatial pyramid pooling (SPP) layer is used to extract multi-scale features from the video images, overcoming the limitations of traditional CNNs on fixed input sizes and enabling accurate identification of scenarios such as vehicle anomalies and personnel congestion within the tunnel. Simultaneously, Fourier transform is used to convert structural strain data to the frequency domain, capturing early structural damage features that are difficult to detect using traditional time-domain analysis. These extracted features are input into a risk prediction model based on a gated recurrent unit (GRU) and an attention mechanism. The model dynamically focuses on data features at key time steps (such as sudden temperature increases or abrupt vehicle stops) through attention weights to predict the probability of safety risks over a future period. When the risk probability exceeds a preset threshold, the response module automatically triggers tiered emergency measures.

[0010] Compared to existing technologies, some focus only on single-dimensional monitoring of equipment status, while this solution integrates multi-source data from the environment, structure, and traffic to achieve cross-validation and accurate identification of risks. For example, when video images detect traffic congestion, and temperature and humidity data show abnormal local temperature increases, combined with structural strain data revealing minute vibration changes, the system can predict potential fire hazards in advance. A single data source cannot achieve this multi-dimensional risk warning. Traditional CNNs require fixed input sizes, making it easy to overlook small targets (such as tunnel wall cracks). This solution extracts multi-scale features through SPP pooling, enabling the model to have a high detection rate for abnormal targets of different sizes (from millimeter-level cracks to vehicle accidents). Traditional time-domain analysis struggles to distinguish between normal vibrations and abnormal vibrations caused by structural damage. This solution successfully identified the resonant frequency shift phenomenon of a tunnel ventilation device through frequency domain analysis, providing early warning of equipment failure and preventing potential equipment collapse accidents inside the tunnel. Existing technologies use fixed-weight RNN models, which cannot dynamically focus on key abnormal events. The attention mechanism of this solution allows the model to automatically concentrate weights on data during the temperature index rise phase in fire prediction scenarios.

[0011] Furthermore, the temperature and humidity sensor in the multimodal data acquisition module is a digital sensor, the harmful gas concentration sensor is an electrochemical sensor array, the structural strain sensor is a fiber optic strain sensor, the video image acquisition device is a 360-degree panoramic camera, and the vehicle speed data is acquired through millimeter-wave radar. Furthermore, before extracting features from the video image data, the data feature extraction module first uses a median filtering algorithm for noise reduction. The specific steps are as follows: replace the gray value of each pixel in the image with the median of the gray values ​​of its neighboring pixels. Furthermore, the risk prediction module's GRU-integrated attention-based risk prediction model is trained using historical tunnel safety data and simulated risk scenario data. During training, a cross-entropy loss function is used, with the following formula: Where N is the number of samples and C is the number of categories. For the sample Category The true label, Predict samples for the model Category The probability of. Furthermore, after the response module initiates emergency response measures, it transmits the response status to the remote monitoring center in real time via the 5G communication network. Furthermore, it also includes a data augmentation module, which is used to enhance the collected multimodal data. For video image data, random flipping, rotation, and brightness adjustment are used to enhance data diversity, and for numerical data, Gaussian noise is added to expand the dataset. Furthermore, after outputting the probability of safety risks, the risk prediction module combines digital twin technology to simulate risk occurrence scenarios in a virtual tunnel model and evaluate the effectiveness of different emergency response measures. Furthermore, when the response module initiates emergency rescue resource scheduling, it uses an improved genetic algorithm to optimize the rescue path, adding a real-time traffic congestion index and rescue time weight to the fitness function of the traditional genetic algorithm. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0013] Example 1 is basically as follows Figure 1 As shown, a tunnel safety intelligent response system includes: The multimodal data acquisition module is used to collect data on temperature and humidity, concentration of harmful gases, structural strain, video images, and vehicle speed inside the tunnel. The data feature extraction module employs an improved convolutional neural network (CNN) to extract features from video image data. This improved CNN adds a spatial pyramid pooling layer after the traditional convolutional layer, extracting multi-scale image features through pooling operations at different scales. The calculation formula is as follows:

[0014] in, For the first The first pooling layer Line number The pooling result of the column, This represents the number of pixels in the corresponding pooling region. For the original image in The pixel value of the location, For the first The pooling regions of each pooling layer are used; simultaneously, Fourier transform is employed to extract frequency domain features from the structural strain data, as shown in the formula: in, It is a frequency domain function. For time-domain structural strain data, Angular frequency; The risk prediction module constructs a risk prediction model based on a gated recurrent unit (GRU) combined with an attention mechanism. It takes extracted multimodal data features as input, and the attention mechanism assigns weights based on the contribution of features at different time steps to the current risk prediction. The specific calculation process is as follows:

[0015] in, Let GRU be the hidden state at time t, with dimension . , This is the hidden state dimension, with a value range of 128-512. Let be the input data features at time t, with dimension . , The input feature dimension is determined by the feature dimension obtained after fusing multimodal data. , Let b be the weight matrix and b be the bias vector. Let K be the attention score at time t, and K be the time step index (ranging from 1 to T). Let T be the attention score at the Kth time step, and T be the total number of time steps (i.e., the length of the input sequence). Let be the attention weight at time t. Let K be the attention weight at time step K. The hidden state is the weighted summation; this model outputs the probability of tunnel safety risks over a future period. The response module automatically initiates corresponding emergency response measures, including adjusting ventilation equipment, traffic guidance, and dispatching emergency rescue resources, when the probability of the safety risk output by the risk prediction module exceeds a preset threshold.

[0016] A multimodal sensor network deployed within the tunnel collects real-time multidimensional data, including temperature and humidity, harmful gas concentrations, structural strain, video images, and vehicle speed. For different data types, an improved convolutional neural network (CNN) combined with a spatial pyramid pooling (SPP) layer is used to extract multi-scale features from the video images, overcoming the limitations of traditional CNNs on fixed input sizes and enabling accurate identification of scenarios such as vehicle anomalies and personnel congestion within the tunnel. Simultaneously, Fourier transform is used to convert structural strain data to the frequency domain, capturing early structural damage features that are difficult to detect using traditional time-domain analysis. These extracted features are input into a risk prediction model based on a gated recurrent unit (GRU) and an attention mechanism. The model dynamically focuses on data features at key time steps, such as sudden temperature increases or sudden vehicle stops, using attention weights to predict the probability of safety risks over a future period. When the risk probability exceeds a preset threshold, the response module automatically triggers tiered emergency measures.

[0017] Compared to existing technologies, which focus only on single-dimensional monitoring of equipment status, this solution integrates multi-source data from the environment, structure, and traffic to achieve cross-validation and accurate identification of risks. For example, when video images detect traffic congestion, and temperature and humidity data show abnormal local temperature increases, combined with structural strain data revealing minute vibration changes, the system can predict potential fire hazards in advance. A single data source cannot achieve this multi-dimensional risk warning. Traditional CNNs require fixed input sizes, leading to the easy neglect of small targets, such as tunnel wall cracks. This solution extracts multi-scale features through SPP pooling, enabling the model to achieve high detection rates for anomalies of varying sizes, from millimeter-level cracks to vehicle accidents. Traditional time-domain analysis struggles to distinguish between normal vibrations and abnormal vibrations caused by structural damage. This solution, through frequency-domain analysis, successfully identified a resonant frequency shift in a tunnel ventilation system, providing early warning of equipment failure and preventing potential tunnel collapses. Existing technologies use fixed-weight RNN models, which cannot dynamically monitor key anomalies. The attention mechanism in this solution allows the model to automatically focus weights on data during the temperature rise phase in fire prediction scenarios.

[0018] The multimodal data acquisition module employs digital temperature and humidity sensors, an electrochemical sensor array for harmful gas concentration sensors, and a fiber optic strain sensor for structural strain. A 360-degree panoramic camera is used for video image acquisition, and vehicle speed data is collected via millimeter-wave radar. The digital temperature and humidity sensors can output digital signals in real time, avoiding noise interference during analog signal transmission and ensuring the accuracy of temperature and humidity data within the tunnel. The electrochemical sensor array can simultaneously detect multiple harmful gases, such as… , , Through the collaborative work of multiple sensors, not only can the accurate measurement of single gas concentration be achieved, but also the identification of complex gas components through cross-sensitivity characteristics, solving the problem of limited monitoring range of traditional single gas sensors; Fiber optic strain sensors, based on optical principles, have advantages such as anti-electromagnetic interference, corrosion resistance, and long-distance transmission. They can be embedded inside tunnel structures to achieve distributed monitoring of minute deformations such as 0.001mm-level cracks. Compared with traditional resistance strain gauges, they have better durability and long-term stability, making them suitable for long-term tracking of tunnel structural health; 360-degree panoramic cameras provide blind-spot-free video monitoring through multi-lens stitching technology. Combined with improved CNN algorithms, they can capture panoramic information such as vehicle driving trajectories and personnel activities in the tunnel in real time, significantly improving the detection efficiency of abnormal events such as vehicles driving in the wrong direction and personnel lingering, avoiding the monitoring blind spots of traditional fixed-view cameras; Millimeter-wave radar is unaffected by environmental factors such as light and smoke, and can accurately measure vehicle speed, position, and traffic flow data in all weather conditions. Especially in extreme scenarios such as heavy rain and dense smoke in tunnels, it can still stably output traffic flow information, making up for the detection defects of video images in harsh environments. Before extracting features from the video image data, the data feature extraction module first uses a median filtering algorithm for noise reduction. Specifically, the grayscale value of each pixel in the image is replaced with the median of the grayscale values ​​of its neighboring pixels. Median filtering, as a non-linear filtering method, works by sorting the grayscale values ​​in the neighborhood of a pixel and taking the median. This effectively suppresses salt-and-pepper noise in the image, such as random sensor interference and black-and-white noise caused by transmission errors. Compared to linear methods like mean filtering, median filtering, while removing noise, can retain image edge and detail information to the maximum extent, such as key features like vehicle outlines in tunnels and road markings, avoiding feature blurring or distortion caused by filtering. In practice, this is achieved by defining an appropriate filtering window, such as... or The neighborhood method traverses the image pixel by pixel, replacing the gray value of each pixel with the median value of its neighborhood. This method can specifically eliminate discrete noise points and also has a certain effect on suppressing Gaussian noise.

[0019] Complex lighting conditions inside tunnels, such as strong light contrast at entrances and exits, uneven internal lighting, and environmental factors like dust and smoke, can easily lead to noise in video images. Directly inputting these noises into a CNN model may cause feature misjudgments, such as misidentifying noise as abnormal targets. Median filtering preprocessing can improve image quality, enabling the subsequent improved CNN model, combined with spatial pyramid pooling, to more accurately extract effective features such as vehicle trajectories, personnel movements, and structural cracks.

[0020] The risk prediction module's GRU-integrated attention-based risk prediction model is trained using historical tunnel safety data and simulated risk scenario data. During training, a cross-entropy loss function is used, with the following formula: Where N is the number of samples and C is the number of categories. For the sample Category The true label, Predict samples for the model Category The probability; GRU, as a variant of recurrent neural networks (RNNs), effectively solves the gradient vanishing problem of traditional RNNs through a gating mechanism, and can capture long-term time-dependent features of tunnel safety data, such as the cumulative changes in temperature and humidity over time and the gradual development of structural strain. Combined with an attention mechanism, the model can automatically learn the contribution of data at different time steps to the current risk prediction. For example, in fire risk prediction, the model will assign higher weights to features during periods of rapid temperature increases, rather than uniformly processing data at all times, thereby improving sensitivity to sudden risks such as fires and traffic accidents.

[0021] Training with the cross-entropy loss function offers the advantage of directly optimizing the model's classification probability output and minimizing the difference between the predicted distribution and the true label distribution for multi-classification tasks such as low-risk, medium-risk, and high-risk level classification. Cross-entropy penalizes the probability of misclassification, especially misjudgment of the high-risk category, forcing the model to improve its ability to identify minority class samples. For example, in training with simulated fire scenario data, the model's prediction accuracy for the high-risk category increased from 68% when trained using mean squared error (MSE) to 89%, significantly reducing the risk of missed detections. Furthermore, the mathematical properties of the cross-entropy loss function facilitate rapid parameter updates during backpropagation. Combined with a mixed training set of historical data and simulated scenarios, it enhances the model's generalization ability to diverse risk scenarios within tunnels, such as structural defects, traffic congestion, and environmental anomalies. This allows the model to accurately predict risk levels even for complex risk combinations not previously trained on, such as abnormal structural strain + hazardous gas leakage + vehicle stagnation, providing a reliable decision-making basis for the response module.

[0022] After the response module initiates emergency response measures, it transmits the response status to the remote monitoring center in real time via the 5G communication network. The 5G network, with its high bandwidth, low latency, and massive connectivity, ensures that real-time data on emergency response status within the tunnel, such as the on / off status of ventilation equipment, personnel evacuation progress, and the location of rescue resources, is rapidly uploaded to the remote monitoring center in the form of high-definition video and structured data. For example, in a tunnel fire scenario, the 5G network can transmit 4K video streams and sensor data from the scene within one second, enabling command personnel at the monitoring center to monitor the fire's spread and the operational status of emergency equipment in real time.

[0023] Meanwhile, 5G slicing technology can create dedicated communication channels for tunnel emergency scenarios, ensuring the reliability and security of data transmission and avoiding information delays or loss due to network congestion. Based on real-time data, the remote monitoring center can dynamically adjust emergency strategies, such as optimizing escape routes based on real-time personnel evacuation trajectories or remotely intervening in ventilation system parameters to improve smoke extraction efficiency. Furthermore, 5G networks support multi-node collaborative communication, enabling real-time audio and video interaction between on-site rescue personnel, smart devices, and remote experts within the tunnel. For example, experts can guide on-site personnel in equipment repair using AR glasses, breaking through the spatial limitations of traditional emergency command and improving the scientific rigor and precision of rescue plans. It also includes a data augmentation module for enhancing the acquired multimodal data. For video image data, it uses random flipping, rotation, and brightness adjustment to enhance data diversity. For numerical data, it adds Gaussian noise to expand the dataset. For video image data, it performs random flipping, horizontal / vertical flipping, and rotation. And brightness adjustment Operations such as brightness variations can simulate monitoring scenarios under different perspectives and lighting conditions within a tunnel. For example, by flipping the image to simulate vehicles entering the tunnel from different directions, and by adjusting the brightness to simulate low-light environments such as heavy rain and fog, the model is exposed to a wider variety of image variations during training, avoiding overfitting caused by limited data. Experiments show that after image enhancement, the model's accuracy in recognizing vehicles in high-contrast lighting conditions at tunnel entrances and exits increases significantly. Upgraded to This significantly improves the stability of detection in complex environments.

[0024] For numerical data such as temperature, humidity, structural strain, and harmful gas concentrations, adding Gaussian noise with a mean of 0 and adaptively adjusted standard deviation can simulate data fluctuations in real-world scenarios, such as sensor measurement errors and environmental interference. For example, injecting a small amount of noise into structural strain data forces the model to learn to distinguish between normal fluctuations and abnormal deformation features, improving its sensitivity to early structural damage. Data augmentation expands the training dataset, alleviating the problem of scarce labeled data in the field of tunnel safety, while also enhancing the model's robustness to noisy data. After outputting the safety risk probability, the risk prediction module, combined with digital twin technology, simulates risk scenarios in a virtual tunnel model to evaluate the effectiveness of different emergency response measures. Digital twin technology constructs a virtual model that maps 1:1 to the physical tunnel, allowing real-time synchronization of multimodal data such as temperature and humidity fields, personnel locations, and equipment status. After outputting the risk probability, it simulates the risk evolution process based on historical data and physical laws, such as fire spread paths and structural collapse trends. For example, when the model predicts an explosion risk in a certain area due to excessive concentrations of harmful gases, the digital twin system can dynamically simulate different ventilation strategies. By comparing simulation results, such as the rate of gas concentration decrease within 30 minutes and changes in the safe area, it automatically recommends the optimal ventilation plan. Compared to traditional experience-based decision-making, the effectiveness of the response strategy is improved by more than 50%.

[0025] Furthermore, digital twins can test emergency response procedures for various extreme scenarios, such as multi-vehicle pileups and tunnel power outages, in a virtual environment without interrupting actual tunnel operations. For example, by simulating the evacuation process, the layout of LED indicator screens and the frequency of voice broadcasts can be optimized to shorten evacuation time; by simulating the routes of rescue vehicles, defects in the setting of traffic control nodes within the tunnel can be identified in advance to avoid congestion on rescue routes.

[0026] When the response module initiates emergency rescue resource scheduling, it employs an improved genetic algorithm to optimize the rescue route. The improved algorithm incorporates real-time traffic congestion indices and rescue time weights into the fitness function of traditional genetic algorithms. Traditional genetic algorithms, which optimize for fixed distances or times, struggle to handle real-time changes in traffic conditions within tunnels, such as congestion caused by accidents or temporary road closures. The improved algorithm, by integrating real-time traffic congestion indices and calculating based on millimeter-wave radar and video data, enables the route planning model to dynamically avoid congested areas. For example, when a rear-end collision occurs on a section of the tunnel, causing a decrease in traffic efficiency... The algorithm can identify congestion points within 2 seconds and automatically adjust the routes of rescue vehicles, detouring through adjacent tunnels or alternative routes. Compared to traditional fixed-route planning, this reduces travel time. above.

[0027] The introduction of rescue time weighting enables differentiated responses to different risk levels: for high-risk scenarios such as fires and people trapped, the algorithm forcibly prioritizes the "shortest rescue time" objective by sacrificing some path distance, such as allowing brief reverse driving or occupying emergency lanes, to ensure that rescue resources arrive within the golden rescue time. For example, in a simulated tunnel fire scenario, the improved algorithm reduced the arrival time of fire trucks from 12 minutes in the traditional method to 7 minutes, gaining crucial time for firefighting and personnel search and rescue. Furthermore, the dynamic adjustment mechanism of the fitness function, such as automatically adjusting the ratio of congestion index to time weight according to the severity of the accident, can adapt to diverse emergency scenarios, prioritizing traffic safety in general traffic accidents and ensuring response speed in major disasters, making the dispatch strategy more flexible and intelligent.

[0028] In actual use, System Architecture and Module Deployment 1. Multimodal data acquisition module Sensor deployment: DHT22 digital temperature and humidity sensors and MQ-135 electrochemical gas sensor arrays are distributed and installed every 50 meters on the inner wall of the tunnel for detection. , , And fiber optic strain sensors. Among them, the fiber optic strain sensors are embedded in tunnel lining structures such as arches and sidewalls to achieve deformation monitoring at the 0.001mm level.

[0029] Video and traffic data collection: 360-degree panoramic cameras with 4K resolution are installed every 100 meters on the tunnel ceiling, supporting multi-lens stitching. These cameras, along with millimeter-wave radar deployed at tunnel entrances, exits, and curves, collect real-time vehicle speed data with high accuracy. Location and traffic flow data.

[0030] Data transmission: All sensors aggregate data to edge computing nodes via an industrial-grade ZigBee network or 5G gateway, achieving synchronization within seconds.

[0031] 2. Data Feature Extraction Module Image preprocessing: For the panoramic camera video stream, first apply a median filtering algorithm. The window removes salt and pepper noise while preserving key features such as vehicle outlines and road markings.

[0032] Multi-scale feature extraction: Video images: An improved CNN, such as ResNet-50+SPP, is used, with spatial pyramid pooling layers added after the convolutional layers. , , Pooling windows are used to extract cross-scale features such as vehicle details and overall tunnel structural deformation, outputting a 2048-dimensional feature vector.

[0033] Structural strain: The time-domain strain data collected by the fiber optic grating sensor is sampled at a rate of 100Hz and converted into a frequency-domain signal through Fourier transform. Features such as the dominant vibration frequency and harmonic components are extracted to form a 50-dimensional frequency-domain feature vector.

[0034] Data fusion: Normalizing image features, frequency domain features, and numerical data such as temperature, humidity, and gas concentration to... The vectors are concatenated into a 3000-dimensional comprehensive feature vector and then input into the risk prediction module.

[0035] 3. Risk Prediction Module Model architecture: The model adopts a GRU + attention mechanism. The input layer is a 3000-dimensional feature vector, the hidden layer contains two GRU layers with 256 neurons each, the attention layer calculates the weights of the time steps using a formula, and the output layer is a 3-class classification of low risk, medium risk and high risk, with the activation function being Softmax.

[0036] Training strategy: Dataset: This dataset contains three years of historical data from a tunnel, including scenarios such as normal operation, fire simulation, and structural cracks. It has been expanded using data augmentation modules, including random image flipping / rotation / brightness adjustment, and the addition of numerical data. Gaussian noise was used to generate 100,000 training samples.

[0037] Optimization objective: Employ the cross-entropy loss function. Optimizer, learning rate 0.001, training to validation set loss converges to approximately Ultimately, high-risk category It reached 0.92.

[0038] Real-time prediction: Edge computing nodes receive a comprehensive feature vector every 10 seconds, and the model outputs the risk probability for the next 30 minutes (e.g., triggering an early warning when the high risk probability is >0.7).

[0039] 4. Response Module Tiered response strategy: Low risk, probability ): Only records data, and displays "normal operation" on the LED screen inside the tunnel.

[0040] Medium risk Initiate local ventilation, such as turning on three sets of fans near the accident site, and push early warning information (including real-time video stream, risk location, and characteristic data) to the remote monitoring center via 5G network. At the same time, trigger traffic guidance such as speed limits. The lane indicator light turned yellow.

[0041] High risk probability : Emergency Response: All tunnel ventilation fans are running at full capacity, tunnel entrances are closed, and personnel are guided to evacuate along the nearest evacuation route via voice broadcasts and LED screens, based on the optimal path simulated by digital twins.

[0042] Rescue dispatch: An improved genetic algorithm is used to optimize the rescue route, with the fitness function being: Where f is the fitness value, t is the estimated rescue time (in minutes), λ is the time weight (ranging from 0.6 to 0.9), c is the congestion index (ranging from 0 to 1), and μ is the congestion weight (ranging from 0.1 to 0.4). Traditional genetic algorithms, with fixed distance or time as the optimization objective, struggle to cope with real-time changes in traffic conditions within tunnels, such as congestion caused by accidents or temporary road closures. The improved algorithm, by integrating real-time traffic congestion indices and calculating in real-time based on millimeter-wave radar and video data, enables the path planning model to dynamically avoid congested areas. For example, when a rear-end collision occurs on a section of the tunnel, causing a 70% decrease in traffic efficiency, the algorithm can identify the congestion node within 2 seconds and automatically adjust the rescue vehicle route, detouring through adjacent tunnels or alternative routes. Compared to traditional fixed-route planning, this reduces travel time by more than 40%.

[0043] 5G communication and digital twins: Response status, such as wind turbine start-up and shutdown, and personnel evacuation progress, is transmitted via 5G slicing network, reducing latency. The data is transmitted in real time to a remote monitoring center, which uses a digital twin model to dynamically simulate the evolution of risks, such as the rate of smoke diffusion and changes in structural stress, to verify the effectiveness of emergency measures. For example, the simulation shows that a certain ventilation strategy can... If the concentration drops below the safe threshold within 15 minutes, the command center can remotely adjust the fan power accordingly.

[0044] 5. Auxiliary Module Data augmentation module: Performs real-time online enhancement on the collected data, with an image enhancement ratio of [missing information]. Each time, one of the following is randomly selected: flip / rotate / brightness adjustment. The numerical data enhancement ratio is [value missing]. Random noise is added to ensure that the model continuously adapts to environmental changes.

[0045] Self-optimization module: Automatically collects deviation data between actual risk events and prediction results every week, and fine-tunes the GRU model parameters through the PPO algorithm to improve long-term prediction accuracy.

[0046] II. System Implementation and Verification 1. Hardware Deployment Edge computing nodes: Utilize NVIDIA Jetson Xavier NX, integrating a GPU acceleration module to support real-time deep learning inference.

[0047] Communication network: 5G macro base stations are deployed inside the tunnel with a coverage radius of 200 meters, and ZigBee relay nodes are used to ensure that sensor data is transmitted without dead zones.

[0048] 2. Software Implementation Data processing and model inference: Based on Python, CNN and GRU models are built using the PyTorch framework, image preprocessing is performed using OpenCV, and Fourier transform is completed using NumPy.

[0049] Digital Twin: A virtual tunnel model is built using Unity3D and synchronized with the physical system in real time via the MQTT protocol, achieving a simulated frame rate of 60fps.

[0050] The above are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A tunnel safety intelligent response system, characterized in that, include: The multimodal data acquisition module is used to collect data on temperature and humidity, concentration of harmful gases, structural strain, video images, and vehicle speed inside the tunnel. The data feature extraction module employs an improved convolutional neural network (CNN) to extract features from video image data. This CNN adds a spatial pyramid pooling layer after the traditional convolutional layers, extracting multi-scale image features through pooling operations at different scales. The calculation formula is as follows: in, For the first The first pooling layer Line number The pooling result of the column, This represents the number of pixels in the corresponding pooling region. For the original image in The pixel value of the location, For the first The pooling regions of each pooling layer are used; simultaneously, Fourier transform is employed to extract frequency domain features from the structural strain data, as shown in the formula: in, It is a frequency domain function. For time-domain structural strain data, Angular frequency; The risk prediction module constructs a risk prediction model based on a gated recurrent unit (GRU) combined with an attention mechanism. It takes extracted multimodal data features as input, and the attention mechanism assigns weights based on the contribution of features at different time steps to the current risk prediction. The specific calculation process is as follows: in, Let GRU be the hidden state at time t, with dimension . , This is the hidden state dimension, with a value range of 128-512. Let be the input data features at time t, with dimension . , The input feature dimension is determined by the feature dimension obtained after fusing multimodal data. , Let b be the weight matrix and b be the bias vector. Let K be the attention score at time t, and K be the time step index (ranging from 1 to T). Let T be the attention score at the Kth time step, and T be the total number of time steps (i.e., the length of the input sequence). Let be the attention weight at time t. Let K be the attention weight at time step K. The hidden state is the weighted summation; this model outputs the probability of tunnel safety risks over a future period. The response module automatically initiates corresponding emergency response measures, including adjusting ventilation equipment, traffic guidance, and dispatching emergency rescue resources, when the probability of the safety risk output by the risk prediction module exceeds a preset threshold.

2. The tunnel safety intelligent response system according to claim 1, characterized in that, The temperature and humidity sensor in the multimodal data acquisition module is a digital sensor, the harmful gas concentration sensor is an electrochemical sensor array, the structural strain sensor is a fiber optic strain sensor, the video image acquisition device is a 360-degree panoramic camera, and the vehicle speed data is acquired through millimeter-wave radar.

3. The tunnel safety intelligent response system according to claim 1, characterized in that, Before extracting features from the video image data, the data feature extraction module first uses a median filtering algorithm for noise reduction. The specific steps are: replacing the gray value of each pixel in the image with the median of the gray values ​​of its neighboring pixels.

4. The tunnel safety intelligent response system according to claim 1, characterized in that, The risk prediction module's GRU-integrated attention-based risk prediction model is trained using historical tunnel safety data and simulated risk scenario data. During training, a cross-entropy loss function is used, with the following formula: in, For the sample size, For the number of categories, For the sample Category The true label, Predict samples for the model Category The probability of.

5. The tunnel safety intelligent response system according to claim 1, characterized in that, After the response module initiates emergency response measures, it transmits the response status to the remote monitoring center in real time via the 5G communication network.

6. The tunnel safety intelligent response system according to claim 1, characterized in that, It also includes a data augmentation module, which is used to enhance the collected multimodal data. For video image data, random flipping, rotation, and brightness adjustment are used to enhance data diversity, and Gaussian noise is added to expand the dataset for numerical data.

7. The tunnel safety intelligent response system according to claim 1, characterized in that, After outputting the probability of safety risks, the risk prediction module uses digital twin technology to simulate risk scenarios in a virtual tunnel model and evaluate the effectiveness of different emergency response measures.

8. The tunnel safety intelligent response system according to claim 1, characterized in that, When the response module initiates emergency rescue resource scheduling, it uses an improved genetic algorithm to optimize the rescue path, adding real-time traffic congestion index and rescue time weight to the fitness function of the traditional genetic algorithm.

Images