Open AccessArticle

Study on Public Perceptions and Disaster Prevention Framework of Tunnel Fires Based on Social Media and Artificial Intelligence

School of Communication, East China University of Political Science and Law, Shanghai 201620, China

Research Centre for Smart Urban Resilience and Firefighting, Department of Building Environment and Energy Engineering, The Hong Kong Polytechnic University, Hong Kong

School of Civil Engineering, Shanghai Normal University, Shanghai 201418, China

Author to whom correspondence should be addressed.

Fire 2024, 7(12), 462; https://doi.org/10.3390/fire7120462

Submission received: 20 September 2024 / Revised: 20 November 2024 / Accepted: 28 November 2024 / Published: 6 December 2024

(This article belongs to the Section Fire Social Science)

Download

Browse Figures

Figure 1
Tunnel fire safety research strategy based on social media and artificial intelligence. "> Figure 2
The number of short video posts and discussions related to tunnel fires from 2017 to 2023. "> Figure 3
The number of tunnel fire-related short videos published on the Douyin platform. "> Figure 4
Number of short video discussions related to tunnel fire in the Douyin platform. "> Figure 5
Topic modeling of tunnel fire short video on the Douyin platform. "> Figure 6
Changes in the proportion of emotional tendency of different themes in tunnel fires from 2017 to 2023. (a) the theme of tunnel fire alarm systems; (b) the theme of tunnel fire accidents; (c) the theme of tunnel fire emergency drills. "> Figure 7
Typical tunnel fire emergency treatment and evacuation plan. "> Figure 8
Dynamic disaster prevention framework for tunnel fire based on social media and artificial intelligence. "> Figure 9
The standard normalized confusion matrix of the tunnel fire critical target recognition model. "> Figure 10
Loss, precision, and recall during model training. (a) loss variation with the training epoch; (b) precision and recall variation with the training epoch. "> Figure 11
Real-time recognition results of tunnel CCTV surveillance video. ">

Versions Notes

Abstract

To investigate public perceptions regarding tunnel fire disasters and optimize the tunnel fire disaster prevention framework, this study takes the emerging social media platform Douyin as a case study, conducting an in-depth analysis of 2133 short videos related to tunnel fires on the platform. A computational communication method was used for analysis, Latent Dirichlet Allocation was used to cluster the discussion topics of these tunnel fire short videos, and a spatiotemporal evolution analysis of the number of videos posted, user comments, and emotional inclinations across different topics was performed. The findings reveal that there is a noticeable divergence in public opinion regarding emergency decision making in tunnel fires, related to the complexity of tunnel fire incidents, ethical dilemmas in tunnel fire escape scenarios, and insufficient knowledge popularization of fire safety practices. The study elucidates the public’s actual needs during tunnel fire incidents, and a dynamic disaster prevention framework for tunnel fires based on social media and artificial intelligence is proposed on this basis to enhance emergency response capabilities. Utilizing short videos on social media, the study constructs a critical target dataset under real tunnel fire scenarios. It proposes a computer vision-based model for identifying critical targets in tunnel fires. This model can accurately and in real-time identify key targets such as fires, smoke, vehicles, emergency exits, and people in real tunnel fire environments, achieving an average detection precision of 77.3%. This research bridges the cognitive differences between the general public and professionally knowledgeable tunnel engineers regarding tunnel fire evacuation, guiding tunnel fire emergency responses and personnel evacuation.

Keywords:

tunnel fires; social media; artificial intelligence; topic evolution; personnel evacuation

1. Introduction

Fire disasters are significant calamities during the operational phase of tunnel infrastructure, as fire incidents are likely to result in severe casualties and significant societal impact [1]. To address this issue, tunnel engineers and fire safety scholars have undertaken extensive work, including the installation of fire detectors for early warning [2], ventilation systems [3], and water-based firefighting equipment to control smoke spread and fire progression [4], as well as the construction of emergency escape passages for trapped individuals [5]. However, during real tunnel fire accidents, it remains challenging for the general public to follow the “idealized” emergency plans devised by tunnel designers, directly leading to casualties. To explore the causes of tunnel fire accidents and human uncertainties in such environments, previous research has typically involved methods such as evacuation simulations, VR simulations, interviews, and surveys [6]. These traditional data collection approaches are time-consuming and often lack effectiveness. The rapid development of emerging social media has transformed ordinary individuals from passive recipients in a linear communication model to prolific “prosumers” of content. On social media, they express their opinions and viewpoints on specific events in real-time through content production and commentary interactions.

Emerging social media platforms such as Twitter, TikTok, Douyin, and Weibo are characterized by diverse information modalities (text, audio, images, video), high communication efficiency (near-real-time information transmission), and low data production costs [7,8]. With the rapid advancement of artificial intelligence technologies such as natural language processing (NLP) and computer vision (CV), and the concept of viewing individuals as sensors, social media has gradually become a vital means of acquiring on-site disaster information and disseminating disaster relief knowledge [9,10,11,12,13,14]. Social media allows affected populations to seek assistance and share information, while also enabling decision makers and rescuers to understand real-time disaster information and respond accordingly. It facilitates public discussion about emergency escape plans under specific disaster scenarios. Therefore, analyzing social media data allows for a direct and effective understanding of the societal impact of tunnel fires from the public’s perspective, and it can optimize existing tunnel fire prevention methods and theories.

This study conducts an in-depth analysis of short video data related to “tunnel fire” on the Douyin platform, aiming to address the following research questions. First, what are the focal points and emotional trends of the public regarding “tunnel fire” short videos on social media? What evacuation and escape plans are considered correct by the public? Secondly, how can an intelligent tunnel fire disaster prevention framework be established or optimized based on the real needs of the public during tunnel fire incidents? These two questions are interrelated. To address the first question, we use Latent Dirichlet Allocation (LDA) to cluster discussion topics on “tunnel fire” within social media content. We then apply Python and Snownlp to analyze the spatiotemporal evolution of post volumes, user comments, and emotional inclinations for different topics. Based on this, a critical target dataset under real tunnel fire conditions was constructed using social media “tunnel fire” videos, and a computer vision-based model for identifying critical targets in tunnel fires was proposed, thus addressing the second question.

The main innovations and advantages of this work are as follows: First, a comprehensive analysis of short video information related to tunnel fires on social media was conducted, which explored public insights and needs regarding tunnel fires from the public’s perspective, thus overcoming the limitations of traditional tunnel fire research perspectives. Second, it integrates multiple disciplines, including tunnel fire science, computational communication, and computer science, establishing a knowledge–social–data-driven model for tunnel fire safety. The remainder of this paper is organized as follows. Section 2 reviews the application of social media in infrastructure disaster prevention and mitigation fields, as well as AI-based tunnel fire safety research. Section 3 introduces the research methodology and specific methods used in this study. In Section 4, the spatiotemporal evolution of tunnel fire-related short videos on social media, the results of topic modeling and sentiment analysis, and the social media-optimized tunnel fire critical target identification model are systematically analyzed. The conclusion is presented at the end.

2. Literature Review

2.1. Social Media for Disaster Prevention

With the rapid development of new social media recently, particularly of short video platforms, emerging social media such as Douyin and Weibo have gradually become important tools for smart disaster prevention and mitigation in infrastructure. This includes disseminating disaster prevention knowledge (before disasters occur), acquiring real-time disaster information (during disasters), and managing emergency response and evacuation (after disasters). Yan et al. proposed a method for urban flood disaster sensing and spatiotemporal evolution analysis based on social media [15], and utilized multimodal data fusion to achieve dynamic sensing of flood information in urban hotspots. Han et al. developed QuakeBERT, the first large language model in the earthquake disaster domain [9], based on historical earthquake-related data from social media. This model enables the analysis of public opinion trends, sentiment, and the physical and social impacts of earthquakes using real-time social media data. Du et al. integrated physical and social sensing data to estimate the probability distribution of flood inundation [16].

Currently, there is a lack of research on social media related to tunnel fire safety. Tunnel fire disasters develop rapidly, with high uncertainty associated with both the fire itself and the evacuation of trapped individuals, which differs from previous disaster prevention and mitigation research using social media.

2.2. Tunnel Fire Safety Based on Artificial Intelligence

Next-generation information technologies such as IoT, artificial intelligence, and big data have provided new means for proactive fire disaster sensing, and the trend toward intelligent disaster emergency response has become international [17,18]. Wang et al. proposed a smart fire calorimetry method based on computer vision and AI-driven fire identification [19,20], using fire images from mobile cameras and deep learning to quantify the heat release rate of fires. Zhang et al. achieved accurate backdraft prediction by integrating visual modal data and sensor temperature time-series data, with prediction accuracy reaching 84% [21].

It should be noted that AI-driven research on tunnel fire safety and personnel evacuation is still in its nascent stages. The effectiveness of using AI model predictions for tunnel fire scenarios to guide actual crowd evacuation is still limited. Furthermore, related research is mainly led by tunnel engineers or fire engineers, with a communication gap between “designers” and “users,” failing to adequately investigate the real needs of the general public during tunnel fire incidents.

In summary, emerging social media provides new opportunities for sensing sudden disaster information and managing emergency evacuations. Traditional tunnel fire safety research focuses on quantifying critical fire parameters and uncertainties using mathematical and physical methods, and on optimizing fire safety design (such as evacuation passage width and spacing) based on performance standards. This approach of formulating evacuation plans from a “fire perspective” has been extensively researched and practiced. Despite this, such an approach overlooks the uncertainties in human evacuation behavior during tunnel fires. Importantly, there exist significant knowledge and data barriers between the general public and tunnel engineers or operational decision makers, as critical data like tunnel surveillance footage are not publicly accessible. Current research from the public’s perspective on the social impact of tunnel fires and evacuation strategy formulation remains scarce. Therefore, this study proposes a novel tunnel fire analysis framework based on social media and artificial intelligence. It employs the Douyin platform, taking an individual public perspective to study tunnel fire evacuation safety. We believe that the methodological contributions and findings of this research provide valuable supplements to the field of tunnel fire safety.

3. Methodology and Data

3.1. Methodology

This study integrates knowledge from fire science and communication studies, employing artificial intelligence methods to connect the two disciplines. The proposed research strategy for tunnel fire safety based on social media and artificial intelligence is illustrated in Figure 1. This methodology primarily involves three steps: social media data acquisition, computational communication analysis, and artificial intelligence model training.

(a): The first step was data collection and cleaning, where web crawlers were employed to gather data on Douyin short video information and user comments related to the keyword “tunnel fire.” Initial data cleaning was conducted to remove duplicate posts from the raw data.
(b): The second step involved spatiotemporal evolution analysis of the collected social media data, which specifically included (1) analyzing the distribution of the number of “tunnel fire” short video posts and discussions across different regions; (2) performing LDA topic modeling on the short videos related to “tunnel fire” on social media, and visualizing the clustering results using the LDAvis tool; (3) conducting sentiment focus analysis on different themes of “tunnel fire” short videos and public discussions using Snownlp.
(c): Finally, we summarized the public’s views on tunnel fires, their perceived correct evacuation plans, and their needs. According to this, the recognition object of the tunnel fire model based on computer vision was determined. An image dataset under the real tunnel fire environment was constructed, and the tunnel fire key target identification model was optimized.

3.2. Data Collection and Cleaning

Since its official launch in 2017, Douyin has become one of the most popular short video social media platforms in China, with over 700 million daily active users and 989 million monthly active users [22]. Short video information from 28 October 2017 to 8 August 2024 was collected with “tunnel fire” as the keyword using web crawler methods. The collected information includes short video files and links, video titles, the number of likes, comments, and shares, publication time, publishing city, video-recognized text, user comments, and the city of the user comments. A total of 3275 short video basic information entries and 494,927 corresponding user comment data entries were collected. Duplicate data were removed based on short video titles and video IDs, and any personal privacy information in the data was anonymized. Subsequently, the short video titles, video-recognized text, and user comments were pre-processed, including the removal of special emojis and characters and filtering of stop words. Finally, a dataset comprising basic information on 2133 short videos and 351,693 corresponding user comments was sorted.

3.3. Text and Image Mining

3.3.1. Topic Clustering

This study employed LDA for theme extraction from short videos related to tunnel fires. LDA is an unsupervised machine learning technique proposed by Blei et al. [23], used to identify latent thematic information within large collections of documents or corpora. It provides a probabilistic distribution of themes for each document, translating textual information into digital information. There are several software tools available for topic modeling and extraction in text data analysis.

It is usually necessary to set the number of extracted topics during the analysis process of an LDA topic model. The perplexity determines the optimal number of topics. Among them, perplexity is used to measure the goodness of a probability distribution or probability prediction sample, and it is also an important basis for calculating the optimal number of topics. The calculation equations are shown in Equation (1):

P e r p l e x i t y (D_{t e s t}) = \exp \{\frac{- \sum_{d = 1}^{M} \log p (w_{d})}{\sum_{d = 1}^{M} N_{d}}\}

(1)

where M denotes the size of the test corpus, N_d denotes the number of words, p(w_d) denotes the probability that the word, and w_d is the sequence of words.

To better understand and present the results of LDA topic modeling, this study used the pyLDAvis package in Python for visualization, which offers a clear representation of the core topics. Topic modeling helps us understand the representative opinions and insights of the public regarding typical tunnel fire scenarios.

3.3.2. Sentiment Analysis

Text sentiment analysis can quickly assess the sentiment trend in short video posts and comments related to tunnel fires, further exploring the concerns and real needs of the public during tunnel fire evacuation processes. Sentiment analysis is a crucial branch in the field of natural language processing (NLP), with automated sentiment analysis via computers being the current academic research trend. This study employs SnowNLP (a Python library suited for NLP tasks in Chinese) to conduct sentiment analysis of the short videos and quantify the sentiment scores [24]. The sentiment scores of SnowNLP are based on the theory of a plain Bayesian classification algorithm, and the calculation equation is as follows:

P (c a t e g o r y | w o r d) = \frac{P (c a t e g o r y | w o r d) P (c a t e g o r y)}{P (w o r d)}

(2)

The intensity of sentiment is expressed as values between 0 and 1. Sentiment scores above 0.6 are classified as positive, below 0.4 as negative, and between 0.4 and 0.6 as neutral. Time evolution analysis was conducted on the results of the sentiment analysis.

3.3.3. Real-Time Object Detection

With the advent and iteration of new algorithms in the era of artificial intelligence, as well as the enhancement and widespread availability of computer hardware, tunnel fire key object detection using computer vision has garnered widespread attention. Diverse and multi-perspective real-time video images (including existing tunnel CCTV surveillance, videos shot by the public’s mobile phones, and footage from vehicle dashcams or parking cameras) provide a reliable visual data foundation for the scientific handling of tunnel fires.

In the field of object detection, You Only Look Once (YOLO) offers advantages in speed (high FPS) and robustness, enabling real-time object localization and classification [25]. It is one of the most popular deep-learning frameworks. This study utilized YOLOv8, released by the paralytics community in 2023, as the detector for identifying key objects in tunnel fires. YOLOv8 was designed to be fast, accurate, and easy to use, which makes it an ideal choice for solving various object detection and tracking, instance segmentation, image classification, and pose estimation tasks. Meanwhile, YOLO v8 is open source, which makes it easy for users to train their own datasets to build new object detection models.

4. Results and Discussion

4.1. Time and Spatial Analysis Results

Metrics from the Douyin platform, such as the number of posts and comments, can directly reflect the public’s attention to a specific topic event. Figure 2 shows the temporal changes in the number of short video posts and discussions related to tunnel fires on Douyin from 2017 to 2023. It can be observed that the trends in the number of posts and discussions related to tunnel fires are closely aligned over time. From 2017 to 2019, the publication and discussion volumes of tunnel fire-related short videos on Douyin were relatively small because short video social media was in its early development stage at that time. As short video media became increasingly integrated into daily life, the number of related posts and discussions began to grow rapidly from 2020 onwards, indicating a marked increase in public attention toward tunnel fires. During this period, the number of tunnels in operation and the mileage of tunnel operations in China also grew rapidly, accompanied by an increase in fire incidents. Research data from Wan et al. [26] show that there were 53 tunnel fire incidents in China in 2021, which is 2.5 times the number recorded in 2017 (21 incidents).

By analyzing the IP addresses of short video publications and user comments, Figure 3 and Figure 4 depict the spatial distribution of short video posts and discussions related to tunnel fires on the Douyin platform, respectively. It is evident that users from economically developed eastern provinces, such as Guangdong, Zhejiang, and Jiangsu, present the highest numbers of tunnel fire-related short video posts and discussions. The central provinces have moderate numbers, while the western provinces, except for Sichuan, Yunnan, and Guizhou, show a significant decline in both. This phenomenon can be attributed to the more developed economies of the eastern provinces, with a higher number of highway and urban road tunnels leading to an increased number of tunnel fire incidents. Additionally, the complex geography of western provinces like Sichuan, Yunnan, and Guizhou makes tunnels an effective means of transportation, resulting in a higher number of tunnels and subsequent tunnel fire incidents.

Zhejiang Province ranks first in the number of posts and discussions. On the one hand, the frequency of tunnel fire incidents in this region is higher than in others, which is also confirmed by the study of Wan et al. On the other hand, this may relate to Douyin’s short video recommendation algorithm. Proximity is a crucial standard in news value theory. Events occurring closer to the audience are more likely to engage them, as such events might directly impact their lives. Thus, tunnel fire incidents occurring in a specific area may be preferentially shown to users in that region.

4.2. Topic Clustering and Sentiment Analysis

LDA topic modeling was performed on the short video topics after data cleaning, and pyLDAvis was used for visual analysis. The maximum number of iterations was set to 200, and based on the principle of minimizing the model’s perplexity, the number of topics was determined to be three. The visualization results are shown in Figure 5, where the size of the circles is positively correlated with the number of posts related to each topic, and the distance between circles represents the similarity between the topics. The topics related to tunnel fire short videos on the Douyin platform are categorized into “tunnel fire alarm systems,” “tunnel fire accidental incidents,” and “tunnel fire drills”. By analyzing the characteristic words under each topic in the horizontal bar chart, the key focus of public attention can be summarized. The details of representative short videos and related information for each topic, along with keywords, are provided in Table 1. These three topics and their keywords reflect users’ perceptions and expressions in response to sudden tunnel fire accidents, and further analysis can reveal the public’s cognitive preferences regarding tunnel fire accidents.

Topic 1 concerning tunnel fire alarm systems (574 short videos, 34,000 comments) primarily includes promotional videos and educational videos about the working principles of tunnel fire detectors and systems, such as the operational principles of fiber optic temperature sensors and smoke fire detectors in tunnels. Topic 3 focuses on tunnel fire emergency drills (671 short videos, 42,500 comments), mainly involving videos of regular tunnel fire emergency drills organized by various tunnel operating units and fire safety education videos shot by local fire departments.

Topic 2 addresses sudden tunnel fire accidents, featuring more short videos (887 videos, 238,700 comments) than the other two topics, becoming a hotspot for both video posting and discussion by the public. Compared to short video news on tunnel fire accidents released by official media or educational short videos on fire self-rescue and evacuation by fire departments, short videos filmed and shared by the public using mobile phones or dash cams during tunnel fire accidents tend to receive broader attention. These short videos are highly timely and are often published faster than traditional news media. Being released by users at the scene of tunnel fire accidents, they provide a strong sense of immediacy and offer the public substantial first-hand information, leading to widespread dissemination and discussion. The peak number of comments for individual short videos in Topic 2 ranges from 10,000 to 50,000, far exceeding the peak number of comments for videos in the other two topics.

Figure 6 presents the changes in emotional tendency proportions for different themes of tunnel fires on the Douyin platform from 2017 to 2023. There are significant differences in the proportion of emotional tendency of tunnel fire short videos with different themes. In the theme of tunnel fire alarm systems, neutral emotions were more prevalent than for the other two themes. Conversely, in the theme of tunnel fire emergency drills, positive emotions consistently dominated. However, in the theme of tunnel fire accidents, the proportion of negative emotions has risen annually and gradually surpassed positive emotions. This trend is linked to two factors. Firstly, as short video social media becomes increasingly integrated into the public’s daily life, the rapid growth in the number of short video posts means that emotional trends during this period align more closely with the public’s actual experiences. Secondly, as a typical sudden disaster, tunnel fires inevitably lead to negative experiences for the public, such as traffic congestion and even abandoning vehicles to evacuate, contributing to the rise in negative emotions.

4.3. Discussion and Implications

Section 4.2 reveals that the theme of tunnel fire accidents is a focal point of public attention. The authors analyzed popular short videos within this theme and the accompanying user comments to summarize the typical views and needs of the public regarding tunnel fire accidents.

4.3.1. Pass or Not? Wait or Evacuate?

Key questions include whether one should drive through or stop when a tunnel fire accident occurs ahead without congestion, and whether to wait or evacuate in case of a traffic jam. These decision-making behaviors in tunnel fire scenarios are major topics of discussion among users. Representative public opinions and corresponding comments are detailed in Table 2.

The main reasons for public support for driving out of the tunnel as soon as possible include the following: 1. Consideration of traffic regulations: most of the operating tunnels are highways and stopping would violate high highway traffic laws and risk rear-end collisions. 2. Moral considerations: stopping can cause traffic congestion within the tunnel, hindering the safe evacuation of subsequent vehicles. 3. People do not want to abandon their cars and financial investments. 4. Many users are not familiar with the tunnel disaster prevention facilities, especially escape routes.

The primary reasons for supporting vehicle abandonment and reverse evacuation are as follows: 1. Dense smoke ahead can greatly impair visibility, increasing the likelihood of traffic accidents and reducing safe evacuation chances. 2. Smoke can lead to vehicle stalling, posing a direct threat to the safety of drivers and passengers. 3. Some individuals are aware of basic tunnel disaster prevention facilities and believe that evacuation routes should be used.

Additionally, some public opinions do not fit the aforementioned categories, advocating for judgments based on the specific tunnel fire situation. Typical views include: 1. If the tunnel is short, keep going. If the tunnel is long, abandon the vehicle and escape. 2. If the fire site is close to the tunnel exit, it should be driven away from the tunnel as soon as possible. 3. The decision should be decisive, whether it is to quickly drive away or abandon the car, one should not hesitate.

The analysis indicates that there is significant public disagreement on the correct emergency decision-making behavior during tunnel fire accidents. This divergence is related to the complexity of tunnel fire accidents, ethical dilemmas in tunnel fire evacuations, and insufficient fire safety knowledge. Specific manifestations include the following:

(a): The complexity of tunnel fires: In the early stages of a tunnel fire accident, the affected vehicle may only be stalled or emitting light smoke. In contrast, during the fire stable stages, the fire may intensify, and dense smoke can completely obstruct visibility in tunnels. These scenarios require different emergency responses.
(b): Insufficient fire safety knowledge: Many people lack a basic understanding of how to respond to a tunnel fire, especially how to stay calm, quickly assess the fire, and choose the appropriate escape route in an emergency. In addition, in the face of emergencies, fast-moving vehicles leave very limited time for drivers to react and make decisions. This time pressure makes it easier for the public to make impulsive choices rather than rational analysis.
(c): Ethical dilemmas in tunnel fire accidents: Individuals often face ethical dilemmas, such as whether to accelerate through the fire site or stop. Accelerating hands over the decision to subsequent vehicles, while stopping means that all following vehicles in the tunnel cannot pass, potentially increasing the risk for oneself and others (e.g., rear-end collisions). This moral choice is not only about personal safety but also about the sense of responsibility for the safety of other people’s lives, which leads to public disagreement in emergency decision making. Individuals are often influenced by emotional and moral judgments in crisis situations, which further aggravates the uncertainty of decision making.

4.3.2. What Is the Correct Solution to a Tunnel Fire Accident?

Figure 7 illustrates a typical tunnel fire emergency response and evacuation plan designed by tunnel engineers. Longitudinal ventilation is the most commonly used emergency ventilation method. People located upstream of the fire source should use evacuation passages within the tunnel to quickly escape the fire site, and the traffic signals at the tunnel entrance should be promptly be closed to prevent vehicles from entering. On the other hand, once all vehicles downstream of the fire source have exited the tunnel, fire trucks should enter the tunnel in reverse from the exit to reach the fire site for firefighting.

It is evident that although tunnel engineers have devised differentiated evacuation plans for people in different areas, the emergency response strategy still fails to fully account for the uncertainties and the individual differences in people in fire scenarios. Many tunnel fire accident investigation reports reveal that even when tunnel control room personnel adjust jet fans, traffic signals, broadcasts, and information boards according to the established plan, the public often cannot fully comply with the set evacuation procedures. The main reasons are as follows: 1. Uncertainty and complexity of tunnel fires: real tunnel fire scenarios dynamically change. For instance, the heat release rate increases rapidly in the early stages of a fire, and the speed and spread of smoke change once longitudinal ventilation is activated. 2. Fixed cognition of tunnel fires: both tunnel engineers and the public have a fixed perception of tunnel fire disasters, believing there are set methods to handle such accidents. However, in the face of complex and ever-changing tunnel fire disasters, existing fixed emergency plans are no longer applicable.

Therefore, it is necessary to adopt a dynamic perspective, analyze fire disasters from the viewpoint of each trapped individual, and adjust evacuation and escape plans based on the latest real-time information from the fire site.

4.4. Tunnel Fire Disaster Prevention Framework Based on Social Media Optimization

4.4.1. General Framework

The authors proposed a dynamic tunnel fire disaster prevention framework based on artificial intelligence in previous work [27], integrating modules for fire state inversion, fire scenario reconstruction, and personnel evacuation. This framework was primarily designed from the perspective of tunnel engineers and insufficiently considered the real needs of the public. Building on prior research, the authors further propose a dynamic tunnel fire disaster prevention framework that incorporates insights from social media and artificial intelligence, as shown in Figure 8.

This framework optimizes the previous study in the following ways. First, by using computational communication methods to analyze tunnel fire-related information on social media, the current perspectives and needs of users regarding tunnel fire accidents are identified. This understanding is integrated with tunnel fire scientific knowledge to optimize the dynamic prediction algorithm module. For example, since users are often unfamiliar with the escape route facilities inside tunnels, escape doors are set as identified targets in the tunnel fire critical target identification model. Additionally, short video data shared by users on social media can be utilized to expand and supplement the tunnel fire database, further enhancing the accuracy of tunnel fire prediction algorithms. New social media platforms are utilized for timely feedback of real-time disaster information, eliminating information barriers between tunnel control room personnel, firefighters, and the public. This facilitates collaboration between rescue groups and evacuating individuals.

4.4.2. Case Study—Tunnel Fire Object Detection Based on Social Media

Based on the aforementioned framework, this section presents a specific algorithmic case study. Currently, researchers have proposed fire and smoke early warning models based on computer vision, which enable early fire warnings and predictions of fire development trends [22,23]. However, these models may face challenges in tunnel fire scenarios. For example, the relatively low-lighting environment inside tunnels and the possibility of red vehicle lights triggering false alarms for fire warnings can be problematic. Additionally, reflections from tunnel walls or other background objects within the tunnel can interfere with the identification of critical targets in tunnel fires using computer vision.

The reason for these challenges lies in the fact that the image datasets used to train the fire warning models primarily consist of publicly available fire datasets from the internet, such as the Kaggle-FIRE Dataset, BoWFire, Mivia Fire & Smoke Detection Dataset, etc. [28]. These datasets mainly focus on forest fires, building fires, and roadside vehicle fires, with few datasets pertinent to real tunnel fire scenarios. This study collected a substantial number of short videos depicting real tunnel fire incidents from the Douyin platform, providing diverse camera angles, including footage shot from the tunnel bottom by the general public using mobile phones or dash cams, and footage captured from the tunnel ceiling by CCTV systems. Compared with public fire datasets, these short video data shared on social media are more suitable for tunnel fire scenarios, and the data acquisition cost is lower.

In this case study, YOLOv8 was used as the detector and classifier of critical targets for tunnel fires. By collecting short videos of tunnel fire on social media in Section 3.2, a large number of real tunnel fire accident image data were obtained. In detail, 3286 images from various views were finally obtained by using the dataset augmentation method. Then, the image annotation tool LabelImg was used to annotate all the images, and the annotation categories included “fire”, “smoke”, “emergency exit”, “car” and “person”. To facilitate model training, the size of all images was resized to 640×640. The YOLOv8 model was customized to suit the tunnel fire object detection task in the present work. The input comprised video frames from social media, and the output was updated to five nodes, which correspond to five categories, i.e., fire, smoke, emergency exit, car and person. The 3286 images were divided into a training dataset and validation dataset at a ratio of 8:1:1.

The training of the proposed tunnel fire recognition model was run on the PyTorch framework under the Windows system. The CPU and GPU used for the training dataset were Intel Core(R) i5-13600KF CPU @5.10 GHz and GeForce RTX 4060ti 16 GB (Santa Clara, CA, USA), respectively. The software environments were CUDA11.8, CUDNN8.8.2, and Python3.8. The optimization method adopted by the YOLO v8s model was end-to-end stochastic gradient Descent (SGD), the task mode was detected, the input image size was 640 × 640 after preprocessing, and the momentum and weight-decay values were set to 0.937 and 0.0005, respectively. Considering that the training time was too long due to the large number of samples and the number of layers, the batch number of images was set to −1 during training, that is, the number of batches was dynamically adjusted. The total number of training epochs was 200.

The key target identification model of the tunnel fire normalized confusion matrix is shown in Figure 9. It can be seen that all categories were predicted relatively well, and the “emergency exit “category was relatively lacking, which may be related to the small number of escape door images in the training dataset; more training data are needed in the future to optimize the model.

To evaluate the performance of the model, loss function, precision, recall, and average detection accuracy [email protected] were used as evaluation metrics. Precision refers to the proportion of all boxes that are correctly detected by the model, reflecting the ability of “good detection”, and recall refers to the proportion of all boxes that are correctly predicted, reflecting the ability of “no missed detection”. The loss curve, accuracy, and recall situation of the model training are shown in Figure 10. The loss of the model decreases sharply with the increase in training cycles, and after 30 cycles of training, the model converges.

Average precision is one of the important indicators to evaluate the performance of the tunnel fire early warning model. The average precision of a single category is represented by AP, which was obtained by integrating the values of precision and recall. Its physical meaning is the area enclosed by plotting P and R as horizontal and vertical coordinates, as shown in Equation (3). For the multi-category detection task, the mean average precision mAP was used to characterize the overall performance of the model, and the more its value approached 1, the better recognition effect of the model was indicated, as shown in Equation (4).

AP = \int_{0}^{1} P_{obtru} (R) dR

(3)

mAP = \frac{\sum_{i = 1}^{c} {AP}_{i}}{c}

(4)

AP_i denotes the average precision of different categories; c represents the total number of detection categories. The mAP50 denotes the mean average precision calculated at an intersection over a union (IOUobstru) threshold of 0.50. In the 200th epoch, all kinds of losses were reduced to the minimum, and the final convergence value of mAP50 of the tunnel fire recognition model reached 0.773.

Figure 11 shows the recognition of the trained model in real videos. The trained model can realize the real-time tunnel fire key target monitoring of CCTV surveillance video. Although the case study is simple, it demonstrated the power of social media data in addressing issues related to tunnel fire safety.

4.5. Challenges, Limitations, and Ways Forward

Considering the large user base of the Douyin platform and the wealth of information in tunnel fire short videos, this study is primarily based on a single platform and a Chinese language environment, which presents certain limitations. It is important to note that scraping tunnel fire-related videos from social media for training artificial intelligence models may raise issues of data privacy and informed consent. When conducting related research, ethical guidelines and legal regulations must be followed, and measures should be taken to protect individuals’ data privacy. In the future, multimodal fusion learning and other methods can be combined to further update related research.

5. Conclusions

This study integrates knowledge from fire science and communication studies, employing artificial intelligence methods to connect the two disciplines. It focuses on the genuine thoughts and evacuation needs of individuals during real tunnel fire incidents. Additionally, this research provides new perspectives for traditional tunnel fire science studies.

On one hand, by mining the textual information of short videos on this social media platform, numerous user opinions on tunnel fire accidents were extracted and summarized, and the tunnel fire prediction algorithm was optimized according to public escape needs. On the other hand, we have made full use of the short video information shared by different users on Douyin to construct a key target dataset for tunnel fires, enhancing the predictive accuracy of AI models. The mAP50 of the tunnel fire recognition model reached up to 77.3%. Additionally, the analysis reveals the cognitive differences between the general public and tunnel designers and operators regarding tunnel fire evacuation. The findings of this study are of significant importance for further enhancing the disaster prevention and mitigation capabilities of tunnel infrastructure and improving emergency response to sudden fire disasters.

Author Contributions

Conceptualization, C.G.; Formal analysis, C.L.; Funding acquisition, C.G.; Investigation, C.G.; Methodology, C.L. and C.G.; Resources, C.L.; Writing—original draft, C.L. and C.G.; Writing—review and editing, C.L., Y.Z., X.T. and C.G. All authors have read and agreed to the published version of the manuscript.

Funding

This work was sponsored by the National Natural Science Foundation of China (grant number 52408429) and the Natural Science Foundation of Shanghai (grant number 24ZR1455800).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The code used in this study and trained models are available on the website: https://github.com/ChaoGuosuper/Tunnel-fire-safety-based-on-Social-media (accessed on 20 September 2024).

Conflicts of Interest

The authors declare no conflicts of interest.

References

Ding, Z.; Xu, S.; Xie, X.; Zheng, K.; Wang, D.; Fan, J.; Li, H.; Liao, L. A Building Information Modeling-Based Fire Emergency Evacuation Simulation System for Large Infrastructures. Reliab. Eng. Syst. Saf. 2024, 244, 109917. [Google Scholar] [CrossRef]
Sun, B.; Guo, T. A Physics-Informed Artificial Fish Swarm Algorithm for Multiple Tunnel Fire Source Locations Prediction. Int. J. Therm. Sci. 2024, 199, 108939. [Google Scholar] [CrossRef]
Ying, H.; Xu, Z.; Zhao, W.; Liu, Q.; Tao, H.; Yan, L. Full-Scale Experiment on the Ventilation Performance of Vehicle-Mounted Mobile Fans in a Road Tunnel. Tunn. Undergr. Sp. Technol. 2024, 148, 105737. [Google Scholar] [CrossRef]
Liu, Y.; Fang, Z.; Tang, Z.; Beji, T.; Merci, B. The Combined Effect of a Water Mist System and Longitudinal Ventilation on the Fire and Smoke Dynamics in a Tunnel. Fire Saf. J. 2021, 122, 103351. [Google Scholar] [CrossRef]
Zhang, Y.; Yan, Z.; Zhu, H.; Shen, Y.; Guo, Q.; Guo, Q. Experimental Investigation of Pedestrian Evacuation Using an Extra-Long Steep-Slope Evacuation Path in a High Altitude Tunnel Fire. Sustain. Cities Soc. 2019, 46, 101423. [Google Scholar] [CrossRef]
Li, W.; Seike, M.; Fujiwara, A.; Chikaraishi, M. Negative Emotion Degree in Smoke Filled Tunnel Evacuation. Tunn. Undergr. Sp. Technol. 2024, 153, 106010. [Google Scholar] [CrossRef]
Florath, J.; Chanussot, J.; Keller, S. Utilizing Volunteered Geographic Information for Real-Time Analysis of Fire Hazards: Investigating the Potential of Twitter Data in Assessing the Impacted Areas. Fire 2024, 7, 6. [Google Scholar] [CrossRef]
Lai, C. The making of a livestreaming village: Algorithmic practices and place-making in North Xiazhu. Chin. J. Commun. 2022, 15, 489–511. [Google Scholar] [CrossRef]
Han, J.; Zheng, Z.; Lu, X.Z.; Chen, K.Y.; Lin, J.R. Enhanced Earthquake Impact Analysis Based on Social Media Texts via Large Language Model. Int. J. Disaster Risk Reduct. 2024, 109, 104574. [Google Scholar] [CrossRef]
Li, L.; Bensi, M.; Baecher, G. Exploring the Potential of Social Media Crowdsourcing for Post-Earthquake Damage Assessment. Int. J. Disaster Risk Reduct. 2023, 98, 104062. [Google Scholar] [CrossRef]
Hodorog, A.; Petri, I.; Rezgui, Y. Machine Learning and Natural Language Processing of Social Media Data for Event Detection in Smart Cities. Sustain. Cities Soc. 2022, 85, 104026. [Google Scholar] [CrossRef]
Roy, P.K.; Kumar, A.; Singh, J.P.; Dwivedi, Y.K.; Rana, N.P.; Raman, R. Disaster Related Social Media Content Processing for Sustainable Cities. Sustain. Cities Soc. 2021, 75, 103363. [Google Scholar] [CrossRef]
Yao, F.; Wang, Y. Towards Resilient and Smart Cities: A Real-Time Urban Analytical and Geo-Visual System for Social Media Streaming Data. Sustain. Cities Soc. 2020, 63, 102448. [Google Scholar] [CrossRef]
Veeraragavan, A.J.; Shanmugavel, R.; Abraham, N.; Subramanian, D.; Pandian, S. Kinetic Studies Validated by Artificial Neural Network Simulation for the Removal of Dye from Simulated Waste Water by the Activated Carbon Produced from Acalypha Indica Leaves. Environ. Technol. Innov. 2021, 21, 101244. [Google Scholar] [CrossRef]
Yan, Z.; Guo, X.; Zhao, Z.; Tang, L. Achieving Fine-Grained Urban Flood Perception and Spatio-Temporal Evolution Analysis Based on Social Media. Sustain. Cities Soc. 2024, 101, 105077. [Google Scholar] [CrossRef]
Du, W.; Xia, Q.; Cheng, B.; Xu, L.; Chen, Z.; Zhang, X.; Huang, M.; Chen, N. Flood Inundation Probability Estimation by Integrating Physical and Social Sensing Data: Case Study of 2021 Heavy Rainfall in Henan, China. Remote Sens. 2024, 16, 2734. [Google Scholar] [CrossRef]
Yabe, T.; Rao, P.S.C.; Ukkusuri, S.V.; Cutter, S.L. Toward Data-Driven, Dynamical Complex Systems Approaches to Disaster Resilience. Proc. Natl. Acad. Sci. USA 2022, 119, e2111997119. [Google Scholar] [CrossRef]
Shuman, J.K.; Schwilk, D.W.; Balch, J.K.; Barnes, R.T.; Higuera, P.E.; Roos, C.I.; Stavros, E.N.; Bendix, J.; Buma, B.; Bertolino, S.; et al. Reimagine Fire Science for the Anthropocene. PNAS Nexus 2022, 1, 1–14. [Google Scholar] [CrossRef]
Wang, Z.; Zhang, T.; Huang, X. Predicting Real-Time Fire Heat Release Rate by Flame Images and Deep Learning. Proc. Combust. Inst. 2022, 39, 4115–4123. [Google Scholar] [CrossRef]
Wang, Z.; Zhang, T.; Wu, X.; Huang, X. Predicting Transient Building Fire Based on External Smoke Images and Deep Learning Fast Fourier Transform. J. Build. Eng. 2022, 47, 103823. [Google Scholar] [CrossRef]
Zhang, T.; Ding, F.; Wang, Z.; Xiao, F.; Lu, C.X.; Huang, X. Forecasting Backdraft with Multimodal Method: Fusion of Fire Image and Sensor Data. Eng. Appl. Artif. Intell. 2024, 132, 107939. [Google Scholar] [CrossRef]
Lai, L. 2024 Monthly Active Users of Douyin in China 2022–2024. Available online: https://www.statista.com/statistics/1361354/china-monthly-active-users-of-douyin-chinese-tiktok/#:~:text=Bite%2Dsized%20video%20format%20has,active%20users%20in%20the%20country (accessed on 25 August 2024).
Blei, D.M.; Ng, A.Y.; Jordan, M.I. Latent dirichlet allocation. J. Mach. Learn. Res. 2003, 3, 993–1022. [Google Scholar]
Chen, C.; Chen, J.; Shi, C. Research on Credit Evaluation Model of Online Store Based on SnowNLP. E3S Web Conf. EDP Sci. 2018, 53, 03039. [Google Scholar] [CrossRef]
Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You Only Look Once: Unified, Real-Time Object Detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016, Las Vegas, NV, USA, 27–30 June 2016; pp. 779–788. [Google Scholar] [CrossRef]
Wan, H.; Jiang, Y.; Jiang, J. A Survey of Fire Accidents during the Process of Highway Tunnel Operation in China from 2010 to 2021: Characteristics and Countermeasures. Tunn. Undergr. Sp. Technol. 2023, 139, 105237. [Google Scholar] [CrossRef]
Guo, C.; Guo, Q.; Zhang, T.; Li, W.; Zhu, H.; Yan, Z. Study on the General Framework for Real-Time Heat Release Rate Inversion of Tunnel Fires with Deep Learning and Transfer Learning. Tunn. Undergr. Sp. Technol. 2024, 148, 105751. [Google Scholar] [CrossRef]
Chino, D.Y.; Avalhais, L.P.; Rodrigues, J.F.; Traina, A.J. Bowfire: Detection of fire in still images by integrating pixel color and texture analysis. In Proceedings of the 2015 28th SIBGRAPI conference on graphics, patterns and images IEEE, Salvador, Brazil, 26–29 August 2015; pp. 95–102. [Google Scholar]

Figure 1. Tunnel fire safety research strategy based on social media and artificial intelligence.

Figure 2. The number of short video posts and discussions related to tunnel fires from 2017 to 2023.

Figure 3. The number of tunnel fire-related short videos published on the Douyin platform.

Figure 4. Number of short video discussions related to tunnel fire in the Douyin platform.

Figure 5. Topic modeling of tunnel fire short video on the Douyin platform.

Figure 6. Changes in the proportion of emotional tendency of different themes in tunnel fires from 2017 to 2023. (a) the theme of tunnel fire alarm systems; (b) the theme of tunnel fire accidents; (c) the theme of tunnel fire emergency drills.

Figure 7. Typical tunnel fire emergency treatment and evacuation plan.

Figure 8. Dynamic disaster prevention framework for tunnel fire based on social media and artificial intelligence.

Figure 9. The standard normalized confusion matrix of the tunnel fire critical target recognition model.

Figure 10. Loss, precision, and recall during model training. (a) loss variation with the training epoch; (b) precision and recall variation with the training epoch.

Figure 11. Real-time recognition results of tunnel CCTV surveillance video.

Table 1. Details of representative short videos and related information for each topic.

No	Topic	Representative Keywords	Representative Douyin Example
1	Tunnel Fire Alarm Systems	Firefighting; Fire detector; Fire; Detector; Linear; Automotive; Temperature-sensing; Fire protection equipment; Professional; Flame; Alarm; Image; Operator; Smoke-sensing; Fire engineering; Fire accident; Pipes; Beam; Testing; Construction; Educational film; Technology; Detection; Cable; Daily; Education; Coatings; Construction site; Fault; Subway	Hikvision 4 G fire detector test # Hikvision # security monitoring
1	Tunnel Fire Alarm Systems		Highway tunnel fire alarm system grating fiber linkage test # Fire detection # tunnel engineering # Fire safety
2	Tunnel Fire accidents	Vehicles; System; Fire; Truck; Alarm; Site; Automatic; Assistant; Gas; Fire site; Firefighter; Driver; Personnel; Large fire; Traffic accident; Alarm device; Traffic; Road; Monitoring; Driving; Highway; Surveillance; Smoke; Dense smoke; Energy; Direction; Safety precautions; Event; Tunnel entrance; Construction	A car caught fire in the tunnel of Yan ‘an East Road in Shanghai # autocombustion # Shanghai.
2	Tunnel Fire accidents		A white car rear-end fire in Chongqing ZhenWushan tunnel # salute to all firefighters may peace # safe travel # people mountain # people sea # everyone pay attention to safety
3	Tunnel Fire Drills	Firefighting; Fire; Emergency; Rescue; Drill; Alarm; System; Accident; Tunnel fire; Coal mine; Product; Smart; Production; Lighting; Systems engineering; Highway; Area; Debugging; Engineering; Sales; Road; Counties and districts; Video; Warning light; Commercial; Hours; Life; Construction; Nationwide; Emergencies	Hong Kong-Zhuhai-Macao Bridge Tunnel Fire suppression system testing # Fire Engineering # Fire testing
3	Tunnel Fire Drills		Hengyang City fire rescue detachment to carry out large-scale tunnel fire rescue combat training # fire # tunnel rescue

Table 2. Representative views and corresponding comments of users on Douyin.

Representative Opinion	Basis of Opinion	Examples of User Reviews	Number of Likes on Comments
Drive out of the tunnel as quickly as possible.	It will violate traffic laws	This tunnel is a highway. It’s better to go straight ahead	598
		This is a highway. If you drive or run back, you will be rear-ended by a car.	110
		There’s a penalty for parking on the highway	86
	Stopping will hinder the following vehicles to escape	This is a tunnel. Get out of the car and run. The cars behind are all blocked.	15
	Distrust of tunnel evacuation routes	I thought it was the Han River outside the tunnel exit door.	151
		Can I get into the escape route no matter which way the smoke goes?	24
		Is there a charge for using the escape route?	7
		I just knew there was an escape route. Is an escape route standard?	13
		Is this door usually open?	24
	Don’t want to give up the car and personal belongings	No car? Just run away?	4238
	Don’t want to give up the car and personal belongings	I want to live with the car. How else can I tell dad?	431
Abandon the vehicle reverse evacuation.	Heavy smoke can cause traffic accidents.	The level of smoke in the tunnel leads to poor vision, and if there is an accident in front of it, it will also lead to traffic jams, so it is more difficult to escape at this time.	29
	Heavy smoke can cause traffic accidents.	The smoke in the tunnel gets thicker as it goes forward.	1262
	Heavy smoke can cause the car to stall	The car would stall due to the heavy smoke, high temperature, and lack of oxygen in the tunnel, and the passengers would be suffocated by the smoke.	41
It depends on the situation.	According to the tunnel length	Unless you are familiar with the tunnel, go to the tunnel exit immediately. Otherwise, don’t hesitate to leave the car and run for the emergency exit or back.	2278
	According to the fire location in the tunnel	Short tunnels of several hundred meters can be rushed. The right thing to do is to pull over with a double flash and run back.	1262

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Share and Cite

MDPI and ACS Style

Lai, C.; Zhang, Y.; Tang, X.; Guo, C. Study on Public Perceptions and Disaster Prevention Framework of Tunnel Fires Based on Social Media and Artificial Intelligence. Fire 2024, 7, 462. https://doi.org/10.3390/fire7120462

AMA Style

Lai C, Zhang Y, Tang X, Guo C. Study on Public Perceptions and Disaster Prevention Framework of Tunnel Fires Based on Social Media and Artificial Intelligence. Fire. 2024; 7(12):462. https://doi.org/10.3390/fire7120462

Chicago/Turabian Style

Lai, Chuyao, Yuxin Zhang, Xiaofan Tang, and Chao Guo. 2024. "Study on Public Perceptions and Disaster Prevention Framework of Tunnel Fires Based on Social Media and Artificial Intelligence" Fire 7, no. 12: 462. https://doi.org/10.3390/fire7120462

APA Style

Lai, C., Zhang, Y., Tang, X., & Guo, C. (2024). Study on Public Perceptions and Disaster Prevention Framework of Tunnel Fires Based on Social Media and Artificial Intelligence. Fire, 7(12), 462. https://doi.org/10.3390/fire7120462

Article Menu