A tunnel fire intelligent prediction method and system based on external flue gas image
By capturing smoke images outside the tunnel and utilizing deep learning algorithms, the problem of real-time transmission of fire source information in tunnel fires has been solved, enabling accurate prediction of fire source power and location, and supporting safety assessment and rescue efforts in tunnel fires.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2024-06-11
- Publication Date
- 2026-07-21
AI Technical Summary
Existing tunnel fire detection technologies are unable to transmit real-time information on fire source power and location due to dense smoke and high temperatures, making it difficult to deploy personnel evacuation and firefighting and rescue operations within tunnels.
By capturing images of external smoke overflowing from the tunnel exit, and analyzing these images using a camera array and deep learning algorithms, the power and location of fire sources inside the tunnel can be predicted in real time.
It enables accurate and real-time prediction of the power and location of fire sources inside tunnels without entering the fire scene, providing a basis for safety assessment of personnel evacuation and fire fighting and rescue in tunnel fires, and reducing equipment installation and maintenance costs.
Smart Images
Figure CN118609051B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fire safety systems, and in particular to a method and system for intelligent prediction of tunnel fires based on external smoke images. Background Technology
[0002] When a fire occurs in a tunnel, the semi-enclosed structure allows hot, dense smoke to rapidly fill the tunnel space. Due to the obstruction of the smoke, monitoring cameras inside the tunnel cannot continuously observe the fire's development. Existing technologies mostly employ sensor arrays within the tunnel to transmit overall information about the fire scene to the outside (Wu X, Zhang X, Jiang Y, et al. Anintelligent tunnel firefighting system and small-scale demonstration[J]. Tunnelling and Underground Space Technology, 2022, 120:104301.). However, due to the obstruction of dense smoke and the effects of high temperatures, detectors such as heat sensors, smoke sensors, and cameras within the tunnel cannot operate continuously and cannot transmit crucial fire scene information, including fire source power and location, in real time. The power and location information of the fire source are critical for the deployment of personnel evacuation and firefighting and rescue operations within the tunnel. Existing tunnel fire detection technologies still need improvement and development; therefore, it is necessary to explore a new method and system for predicting tunnel fires. Summary of the Invention
[0003] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and system for intelligent prediction of tunnel fires based on external smoke images. By taking images of external smoke overflowing from the tunnel exit in a safe area away from the fire outside the tunnel, and analyzing the external smoke images through a deep learning algorithm, the power and location of the fire source inside the tunnel can be predicted in real time.
[0004] The present invention is achieved by at least one of the following technical solutions.
[0005] A method for intelligent prediction of tunnel fires based on external smoke images includes the following steps:
[0006] Images of external smoke overflowing from each end of the tunnel were captured from multiple perspectives using a camera array.
[0007] The camera's data acquisition module receives external smoke images and transmits them to a server, which includes a field computer terminal and a cloud server.
[0008] The server segments, denoises, and performs grayscale processing on the external smoke image, extracting the effective computational region of the external smoke image;
[0009] By processing external smoke images using deep learning algorithms, the fire source power and location of tunnel fires can be predicted in real time based on the shape, color, and texture of the external smoke.
[0010] Furthermore, the camera array captures images of external smoke overflowing from each end of the tunnel from multiple perspectives, including:
[0011] Images of external smoke escaping from each end of the tunnel were captured by cameras in a safe area outside the tunnel, away from the fire source.
[0012] Images of external smoke taken by a camera outside the tunnel at any angle, at any distance from the tunnel, and at any height;
[0013] External smoke images can be captured directly by a camera, or they can be extracted frame by frame from captured external smoke videos.
[0014] Furthermore, the step of segmenting, denoising, and grayscale processing the external smoke image via the server includes:
[0015] The main body of the external smoke in the image is segmented using object detection algorithms or semantic segmentation algorithms;
[0016] Reduce the impact of background noise in images using noise reduction algorithms;
[0017] The image is converted to grayscale format via a server.
[0018] Furthermore, the target detection algorithm identifies and locates the target smoke region from the image, and segments it from the background along the boundary of the target smoke.
[0019] Furthermore, the semantic segmentation algorithm identifies the target smoke region in the image by using bounding boxes or by finding all pixels belonging to the target smoke.
[0020] Furthermore, the noise reduction algorithm is used to reduce image noise caused by random interference factors such as camera installation, placement, and parameter settings.
[0021] Furthermore, converting an image to grayscale format reduces the number of color channels in the image to reduce computational load; the final image may not necessarily be in grayscale format.
[0022] Furthermore, the smoke images input into the deep learning algorithm can be smoke images taken from a single viewpoint, or any combination of smoke images taken from different angles of the tunnel, at different distances from the tunnel entrance, at different heights, or at different tunnel entrances.
[0023] Furthermore, the deep learning algorithm is a recurrent neural network or a visual attention model, which can identify the temporal sequence relationship between the input signals and extract features that a single signal does not have from this temporal sequence relationship; the image input to the deep learning algorithm is a smoke image taken from a single viewpoint, or any combination of smoke images taken at different angles to the tunnel, at different distances from the tunnel port, at different heights, or at different ports of the tunnel.
[0024] The system for implementing the intelligent prediction method for tunnel fires based on external smoke images is characterized by comprising:
[0025] A camera array is used to capture images of smoke outside the tunnel from multiple perspectives;
[0026] The data acquisition module is used to receive external smoke images and transmit them to the server, including on-site computer terminals and cloud servers;
[0027] The server is used to segment, denoise, and perform grayscale processing on external smoke images. It uses convolutional neural networks to predict the fire source power and location of tunnel fires in real time, and uses deep learning algorithms such as recurrent neural networks or visual attention models to process external smoke images.
[0028] Compared with existing technologies, the beneficial effects of the present invention are as follows:
[0029] This invention uses a camera to capture images of external smoke overflowing from the tunnel's openings in a safe area outside the tunnel. These images are received by a data acquisition module and transmitted to a server. The server then processes the images, performing segmentation, noise reduction, and grayscale adjustments. Finally, the processed images are used to create a training dataset, which is then input into a deep learning algorithm capable of processing time-series logic signals. This allows for real-time prediction of the power and location of fire sources within the tunnel based on the external smoke images, providing a safety assessment basis for personnel evacuation and firefighting rescue operations in tunnel fires. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the overall structure of the intelligent tunnel fire prediction method and system based on external smoke images, as shown in the embodiment.
[0031] Figure 2 These are side views of external smoke from both sides of the tunnel under different conditions in the embodiments;
[0032] Figure 3 A flowchart illustrating the processing of external smoke images in an embodiment;
[0033] Figure 4 This is a schematic diagram of the deep learning algorithm structure for an example.
[0034] Figure 5A comparison chart of the heat release rate curve preset for the embodiment and the heat release rate curve predicted by the deep learning algorithm;
[0035] Figure 6 This is a comparison chart of the actual fire source location and the fire source location predicted by the deep learning algorithm in an example. Detailed Implementation
[0036] This invention provides a method and system for intelligent prediction of tunnel fires based on external smoke images. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.
[0037] Please see Figure 1 The intelligent prediction method for tunnel fires based on external smoke images in this embodiment includes the following steps:
[0038] S10. Using camera array 20, images of external smoke overflowing from each port of the tunnel are captured from multiple perspectives.
[0039] S20. Receive external smoke images through the built-in or external data acquisition module 30 of the camera, and then transmit the data acquisition module to the server 40, including the on-site computer terminal 41 and the cloud server 42.
[0040] S30. The external smoke image is segmented, denoised, and grayscale processed by the server to extract the effective calculation area of the external smoke.
[0041] S40. Process external smoke images through convolutional neural networks to predict the fire source power and fire source location of tunnel fires in real time based on the shape, color, and texture of external smoke; process external smoke images through deep learning algorithms suitable for processing signals containing temporal logic, such as recurrent neural networks or visual attention models (Vision Transformer), to predict the fire source power and fire source location of tunnel fires in real time based on the changing trends of external smoke morphology characteristics.
[0042] Specifically, because the shape, color, and texture of external smoke are all controlled by the fire inside the tunnel, much information about the fire can be obtained by observing the characteristics of the external smoke. For example... Figure 2 As shown, Figure 2 (a) is a side view of the external smoke on both sides of the tunnel after changing its heat release rate while keeping the location of the fire source and the smoke production rate unchanged. Figure 2 (b) is a side view of the external smoke on both sides of the tunnel after changing the location of the fire source while keeping the heat release rate and smoke production rate of the fire source in the tunnel constant. Figure 2(c) Side views of external smoke on both sides of the tunnel after changing the smoke production rate of the fire source while keeping the heat release rate and location of the fire source constant. When the location and smoke production rate of the fire source remain constant, the higher the heat release rate, the darker the color of the external smoke and the wider it becomes. When the heat release rate and smoke production rate of the fire source remain constant, changing the location of the fire source will change the inclination angle of the external smoke at both ends of the tunnel. However, if the change in the location of the fire source is small, the corresponding change in the external smoke is difficult to capture with the naked eye, but it can be captured by deep learning algorithms. When the heat release rate and location of the fire source remain constant, an increase in the smoke production rate of the fire source will also make the color of the external smoke darker and the width wider. Therefore, changes in the smoke production rate will affect our judgment of changes in the heat release rate. However, an increase in the heat release rate will cause the temperature inside the tunnel to rise faster and the smoke flow to accelerate, while an increase in the smoke production rate will not produce such an effect. Therefore, there is a subtle difference between the changes in external smoke caused by an increase in the smoke production rate and the changes in external smoke caused by an increase in the heat release rate. This also allows deep learning algorithms to shield themselves from the influence of smoke production rate, enabling accurate prediction of the heat release rate at a fire site. Based on this, such as... Figure 1 As shown, this invention deploys cameras 20 at both ends of the tunnel to capture images of external smoke from multiple angles, ensuring comprehensive acquisition of the real-time morphological characteristics of the external smoke. These images are then received by the information acquisition module 30 and sent to the server 40, which includes a field server terminal 41 and a cloud server 42. The server 40 segments the images to extract effective computational regions; it denoises the images to reduce the impact of image noise caused by background and other random factors; and it performs grayscale processing to reduce the number of color channels and computational load. The external smoke images are then input into a convolutional neural network via the server 40. The convolutional neural network uses convolutional layers to extract features from the images, reducing the length and width of the feature tensor and increasing its depth. Finally, the feature tensor is transformed into a vector and input into a fully connected layer to calculate the fire source power and location. The server 40 inputs external smoke images into deep learning algorithms, such as recurrent neural networks (RNNs) or visual attention models. The former (RNN) processes the sequence relationships in the signal by updating memory and iteratively adding computations, then inputs the processed feature tensor into a fully connected layer to obtain the fire source power and location. The latter (visual attention model) uses an attention mechanism to encode the sequence signal, which is then input into a fully connected layer to obtain the fire source power and location. Furthermore, because later-stage fires cause greater damage to sensors inside the tunnel, pose a greater risk to rescue personnel, and are more likely to trigger secondary disasters, this invention is crucial for providing accurate real-time prediction of fire source power and location within the tunnel from a safe area outside the tunnel.
[0043] In one implementation, step S30 includes the following steps:
[0044] S31. Server 40 uses object detection or semantic segmentation algorithms to segment the main part of the external smoke in the image, leaving only the effective calculation area.
[0045] S32. The server 40 performs noise reduction on the external smoke image to reduce the impact of image noise caused by background and other random factors.
[0046] S33. The external smoke image is converted to grayscale by server 40 to reduce the number of color channels in the image.
[0047] Specifically, server 40 segments the external smoke portion of the external smoke image. Noise reduction algorithms are used to reduce background noise and mitigate image noise caused by factors such as camera installation, placement, and parameter settings. These algorithms include, but are not limited to, filtering algorithms, image noise modeling-based algorithms, and deep learning-based algorithms. The image is converted to a format including, but not limited to, grayscale values to reduce the number of color channels.
[0048] In a preferred implementation, the location of the external smoke portion in the external smoke image can be determined by a target detection algorithm, or the external smoke portion can be cut out along the boundary using a semantic segmentation algorithm. The segmented external smoke portion is preferably the portion near the tunnel opening, as the morphological characteristics of this portion are primarily controlled by the fire situation inside the tunnel.
[0049] Furthermore, the denoising process in step S32 can be used depending on the specific circumstances. Step S32 can be skipped when image noise is not significant or the background has been removed. The external smoke portion should not be too large or too small. Too large a portion will waste computational resources and may even introduce image noise; while too small a portion will result in the loss of external smoke feature information.
[0050] In one implementation, step S40 specifically includes the following steps:
[0051] S41. Process external smoke images through convolutional neural networks, and predict the fire source power and fire source location of tunnel fires in real time based on the shape, color and texture of external smoke.
[0052] S42. External smoke images are processed using deep learning algorithms suitable for processing signals containing temporal logic, such as recurrent neural networks or visual attention models. Based on the changing trends of external smoke morphology, the fire source power and location of tunnel fires are predicted in real time.
[0053] Specifically, after step S30, the effective computational region of the external smoke image is extracted. The processed external smoke image is input into a convolutional neural network to obtain the tunnel fire source power and location corresponding to each frame. The processed external smoke image is then input into a recurrent neural network or other deep learning algorithms, such as a visual attention model, to obtain the tunnel fire development trend over a future period.
[0054] In a preferred implementation, the convolutional neural network can be VGG16 or ResNet. The former (VGG16) has a neat network structure, which is conducive to hardware acceleration and reduces the number of parameters to be trained by using multiple small-sized convolutional kernels; the latter (ResNet) can use residual connections, which greatly increases the upper limit of the number of layers of the convolutional neural network.
[0055] Furthermore, the order in which steps S41 and S42 are implemented is not important, nor do they necessarily need to be completed. Depending on the specific circumstances, only step S41 or only step S42 may be implemented, or steps S41 and S42 may be performed simultaneously.
[0056] The advantages of this invention are:
[0057] (1) Without entering the fire scene, the morphological characteristics of external smoke are obtained only in a safe area outside the tunnel, so as to indirectly predict the power and location of the fire source in the tunnel.
[0058] (2) The camera used only needs to be placed outside the tunnel, and its shooting accuracy is not affected by the development of the fire. It can realize real-time prediction of the power and location of the fire source.
[0059] (3) Using deep learning algorithms, the power and location information of fire sources in the tunnel can be predicted synchronously based on the morphological characteristics of external smoke.
[0060] (4) The required equipment and systems are easy to install, simple to operate and easy to maintain, which can save a lot of maintenance costs.
[0061] The intelligent prediction method for tunnel fires based on external smoke images of the present invention will be illustrated below by way of example.
[0062] Please see Figure 1 In a numerical model of a tunnel measuring 100m × 8.6m × 6.6m, a car fire source measuring 9m × 3m × 4.5m, located 45m from the left side of the tunnel, was ignited. The fire source... Figure 5The solid line in the diagram represents the trend of heat release rate. Camera array 20 is positioned at both ends of the tunnel, with two cameras at each end capturing images of the external smoke from both frontal and side views. The cameras are connected to the information acquisition module 30 via wired connections. After receiving the images of the external smoke captured by the cameras, the information acquisition module 30 transmits them wirelessly to the cloud server 42.
[0063] The image was processed within the cloud server 42 as follows Figure 3 The processing is as shown. First, the portion of the external smoke near the tunnel opening is segmented, measuring 8.8m × 8.8m, and then this portion is converted to grayscale format. Four images of the external smoke from both ends of the tunnel, taken from both frontal and side views, are simultaneously input into a convolutional neural network to obtain the location of the fire source and the heat release rate of the fire for each frame. Note: The location of the fire source refers to the distance between the center of the fire source and the left end of the tunnel.
[0064] The convolutional neural network used is VGG16, and its specific structure is as follows: Figure 4 As shown in the figure, VGG16 has a neat network structure, which is beneficial for hardware acceleration. It uses multiple small-sized convolutional kernels instead of large-sized convolutional kernels, reducing the number of parameters to be trained. VGG16 has 13 convolutional layers, 5 max-pooling layers, and 3 fully connected layers. The convolutional layers use 3×3 small-sized convolutional kernels, and batch normalization is performed after each convolution calculation. The activation function is ReLU. Each max-pooling layer is followed by a Dropout layer with a Dropout rate of 0.3. The fully connected layers are also followed by a ReLU activation function, followed by a Dropout layer with the same Dropout rate of 0.3.
[0065] like Figure 5 As shown, Figure 5 This compares the preset heat release rate curve based on the fire source power with the heat release rate curve predicted by VGG16 during the duration of a tunnel fire. The figure shows that VGG16 can predict the heat release rate of a fire inside the tunnel in real time using external smoke images, ensuring high accuracy. Similarly, such as... Figure 6 As shown, Figure 6 This is a comparison between the actual location of the fire source and the location predicted by VGG16. As can be seen from the image, VGG16 can predict the location of the fire source in real time based on external smoke images, with an error of no more than 1 meter.
[0066] Based on the above methods, the present invention also provides a tunnel fire intelligent prediction system based on external smoke images, including: a camera array for capturing external smoke images of the tunnel from multiple perspectives;
[0067] The data acquisition module is used to receive external smoke images and transmit them to the server, including on-site computer terminals and cloud servers;
[0068] The server is used to segment, denoise, and perform grayscale processing on external smoke images. It processes external smoke images through convolutional neural networks and predicts the fire source power and location of tunnel fires in real time based on the shape, color, and texture of external smoke. It also processes external smoke images through deep learning algorithms suitable for processing signals containing temporal logic, such as recurrent neural networks or visual attention models.
[0069] In summary, this invention uses a camera to capture images of external smoke billowing from inside the tunnel. The data acquisition module receives these images and transmits them to a server. The server processes the images through segmentation, noise reduction, and grayscale adjustments before inputting them into deep learning algorithms suitable for processing signals containing temporal logic, such as convolutional neural networks, recurrent neural networks, or visual attention models. Since the shape, color, and texture of the external smoke are controlled by the fire situation inside the tunnel, the deep learning algorithm can determine the fire source power and location based on the images of the external smoke. This allows for accurate prediction of the fire situation at any time during the fire, especially in the later stages, without the need for deploying and maintaining sensors inside the tunnel. This provides a safety assessment basis for personnel evacuation and rescue operations.
[0070] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.
Claims
1. A method for intelligent prediction of tunnel fires based on external smoke images, characterized in that, Includes the following steps: Images of external smoke overflowing from each end of the tunnel were captured from multiple perspectives using a camera array. The camera's data acquisition module receives external smoke images and transmits them to a server, which includes a field computer terminal and a cloud server. The server segments, denoises, and performs grayscale processing on the external smoke image, extracting the effective computational region of the external smoke image; By processing external smoke images using VGG16, the power and location of tunnel fires can be predicted in real time based on the shape, color, and texture characteristics of the external smoke. The smoke images input to VGG16 are any combination of smoke images taken at different angles, distances from tunnel entrances, heights, or at different tunnel entrances.
2. The intelligent prediction method for tunnel fires based on external smoke images according to claim 1, characterized in that, The camera array captures images of external smoke overflowing from each end of the tunnel from multiple perspectives, including: Images of external smoke escaping from each end of the tunnel were captured by cameras in a safe area outside the tunnel, away from the fire source. Images of external smoke taken by a camera outside the tunnel at any angle, at any distance from the tunnel, and at any height; External smoke images can be captured directly by a camera, or they can be extracted frame by frame from captured external smoke videos.
3. The intelligent prediction method for tunnel fires based on external smoke images according to claim 1, characterized in that, The process of segmenting, denoising, and grayscale processing the external smoke image via a server includes: The main body of the external smoke in the image is segmented using object detection algorithms or semantic segmentation algorithms; Reduce the impact of background noise in images using noise reduction algorithms; The image is converted to grayscale format via a server.
4. The intelligent prediction method for tunnel fires based on external smoke images according to claim 3, characterized in that, The target detection algorithm identifies and locates the target smoke region in the image, and segments it from the background along the boundary of the target smoke.
5. The intelligent prediction method for tunnel fires based on external smoke images according to claim 3, characterized in that, The semantic segmentation algorithm identifies the target smoke region in the image by using bounding boxes or by finding all pixels belonging to the target smoke.
6. The intelligent prediction method for tunnel fires based on external smoke images according to claim 3, characterized in that, The noise reduction algorithm is used to reduce image noise caused by random interference factors such as camera installation, placement, and parameter settings.
7. The intelligent prediction method for tunnel fires based on external smoke images according to claim 3, characterized in that, Converting an image to grayscale reduces the number of color channels to reduce computation; however, the final image may not necessarily be in grayscale format.
8. A system for implementing the intelligent prediction method for tunnel fires based on external smoke images as described in claim 1, characterized in that, include: A camera array is used to capture images of smoke outside the tunnel from multiple perspectives; The data acquisition module is used to receive external smoke images and transmit them to the server, including on-site computer terminals and cloud servers; The server is used to segment, denoise, and perform grayscale processing on external smoke images. It uses a convolutional neural network to predict the fire source power and location of tunnel fires in real time and processes external smoke images using VGG16.