Long tunnel fire situation real-time prediction method based on image recognition and physical information fusion

Abstract

The application discloses a long and large tunnel fire situation real-time prediction method based on image recognition and physical information fusion, which comprises the following steps: 1, a complex ventilation mode is built under the scale size long and large slope tunnel fire simulation device, and a long and large slope tunnel fire data set is constructed; 2, a long and large slope tunnel fire parameter prediction network based on a Vision Transformer architecture is constructed; 3, a total loss function of the long and large slope tunnel fire parameter prediction network is constructed and trained, so that an optimal fire parameter prediction model is obtained, which is used for predicting long and large slope tunnel fire images and physical information, and long and large slope tunnel fire parameters are obtained. The application can realize dynamic simulation of long and large slope tunnel fires under complex ventilation modes, and can predict the heat release rate of long and large slope tunnel fires under the coupling effect of complex ventilation modes in the construction and operation stages, and the temperature and heat flow distribution at different positions under the ceiling according to flame images and physical information.

Description

Technical Field

[0001] This invention belongs to the field of tunnel fire safety and intelligent sensing technology, specifically involving a real-time prediction method for fire situations in long tunnels based on image recognition and physical information fusion. It is applicable to the dynamic simulation and key parameter prediction of fires in long and steep tunnels with complex ventilation modes, such as hydropower engineering tunnels, auxiliary chambers, and urban highway tunnels. Background Technology

[0002] As one of the most critical underground structures in hydropower projects, the safety and reliability of hydraulic tunnels directly affect the long-term stable operation of the entire engineering system. Hydraulic tunnels are characterized by their massive scale, encompassing ultra-long tunnel groups, auxiliary chambers, single-ended tunnels, and complex ventilation and drainage systems. Compared to general traffic tunnels, hydraulic tunnels not only undertake multiple functions such as water conveyance, flood discharge, and power transmission, but are also located in high-altitude, low-pressure, steep-slope, and complex geological environments, making their operating environment exceptionally complex and placing higher demands on safety protection. Furthermore, during construction and operation and maintenance, the extensive use of fuel-powered / new energy machinery, welding and cutting equipment, and highly flammable materials such as explosives poses significant fire safety hazards. Simultaneously, the dense layout of construction cables, transformers, and control equipment within the tunnels makes them highly susceptible to fire in the event of an electrical short circuit. Key fire parameters, such as the heat release rate, temperature distribution below the tunnel ceiling, and heat flow distribution, directly affect personnel evacuation, smoke extraction coordination, and firefighting decisions. Therefore, how to quickly and accurately predict fire parameters in complex hydropower tunnel environments is a core issue that urgently needs to be addressed in the field of fire safety.

[0003] Existing technologies mainly rely on video surveillance and sensor location-based measurement methods. However, such methods have the following shortcomings: (1) The limited number of sensors leads to sparse data collection, making it difficult to fully reflect the spatiotemporal evolution characteristics of the fire scene; (2) In high-radiation and high-temperature smoke environments, sensors are easily interfered with or damaged; (3) The combustion characteristics and smoke spread patterns of tunnel fires are affected by multiple factors such as tunnel slope, atmospheric pressure, fuel type, and fire source location. It is difficult to achieve accurate prediction across working conditions and all time periods by simply relying on empirical formulas or a single physical model.

[0004] With the rapid development of computer vision and deep learning technologies, fire parameter inversion based on flame images has provided a new approach for predicting fire parameters in hydropower engineering tunnels. This type of method can overcome the limitations of traditional point sensors, utilizing image features to capture the spatiotemporal variation patterns of flames, thereby inferring key fire parameters. However, most existing methods only target a single parameter, lacking the ability to predict multiple parameters in real time, exhibiting insufficient generalization in complex environments, and have not yet formed a systematic solution for practical engineering applications. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing a real-time prediction method for fire situations in long tunnels based on image recognition and physical information fusion. The aim is to rapidly and accurately predict the heat release rate, temperature below the ceiling, and heat flow distribution in long, steep tunnels under complex ventilation modes. This enables real-time and accurate output of multiple fire parameters under complex ventilation modes, meeting the real-time response requirements during fire development and providing timely data support for tunnel fire risk assessment and emergency decision-making, thereby improving the safety and reliability of tunnel fire emergency response.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] The present invention provides a real-time prediction method for fire situations in long tunnels based on image recognition and physical information fusion, characterized by the following steps:

[0008] Step 1: Construct a scaled-down fire simulation device for long and steep tunnels under complex ventilation modes to build a fire dataset for long and steep tunnels, including: a dataset of physical information parameters of long and steep tunnels. P Long slope tunnel fire image dataset X Dataset of experimental values ​​for fires in long and steep tunnels Y ;

[0009] Step 2: Construct a fire parameter prediction network for long and steep tunnels based on the Vision Transformer architecture, including: a physical information encoding module, a flame image feature extraction module, an HRR feature fusion module, an HRR prediction head, a T / Heatflux feature fusion module, and a T / Heat flux prediction head, which respectively... P and X Encoding and feature extraction are performed to obtain the first... i Under the first working condition m Physical information fusion vector at time step and flame feature vector and will and After fusion processing, the first... i Under the first working condition m Flame parameter feature fusion vector at time step ,Will Input the HRR prediction header to get the first i Under the first working condition m Predicted heat release rate at time The HRR embedding layer will Convert to the first i Under the first working condition m The heat release rate eigenvector at time t. ;Will , With the corresponding temperature location coordinate feature vector The fusion yields the first i Under the first working condition m Temperature feature fusion vector at time point At the same time, , With the corresponding heat flow location coordinate eigenvector The fusion yields a heat flow feature fusion vector. ;Will and The data is input into the T prediction header and the Heat Flux prediction header respectively, and the corresponding results are obtained. i Under the first working condition m Predicted temperature value at time and heat flow prediction values ;

[0010] Step 3: Construct the total loss function for the fire parameter prediction network in long and steep tunnels. The Including heat release rate loss function Temperature distribution loss function and heat flux distribution loss function ;

[0011] Step 4, using the above The fire parameter prediction network is trained, and the network parameters are continuously adjusted through the backpropagation algorithm until the desired parameters are obtained. The model converges to obtain the optimal fire parameter prediction model; it is used to predict the fire images and physical information of long and steep tunnels, and outputs the corresponding predicted values ​​of heat release rate, temperature distribution and heat flow distribution in real time.

[0012] The real-time prediction method for fire situations in long tunnels based on image recognition and physical information fusion, as described in this invention, is characterized in that step 1 includes:

[0013] Step 1.1: Construct a scaled-down fire simulation device for a long, steep tunnel under a complex ventilation mode, including: a long, steep tunnel, a complex ventilation and smoke extraction system, a lifting multi-type fire source module, and a fire parameter acquisition unit;

[0014] Step 1.1.1 The long slope tunnel includes a long slope tunnel experimental section (1) and a hydraulic lifting system (2); the long slope tunnel experimental section (1) is constructed with a high temperature resistant stainless steel frame, and a high temperature resistant transparent tempered glass observation window is set on the front; the tunnel slope is adjusted by the hydraulic lifting system (2) to simulate fire scenarios under different slopes;

[0015] Step 1.1.2: The complex ventilation and smoke exhaust system includes a longitudinal smoke exhaust fan system (3), a lateral smoke exhaust fan system (4) and a top smoke exhaust fan system (5), which are used to simulate the smoke exhaust conditions under independent or coupled ventilation modes in the experimental section of a long and steep tunnel.

[0016] Step 1.1.2.1 The longitudinal smoke exhaust fan system (3) includes: a first variable frequency fan and a first frequency converter; the first variable frequency fan is arranged at one end of the long slope tunnel test section, and the fan speed is adjusted by the first frequency converter to realize the continuous and controllable change of longitudinal wind speed, which is used to simulate the smoke exhaust conditions under different longitudinal ventilation modes;

[0017] Step 1.1.2.2, the lateral smoke exhaust fan system (4) includes: a second variable frequency fan, a second frequency converter, a lateral suction port, and a lateral valve; multiple lateral suction ports are arranged along the length direction on the side wall of the long slope tunnel experimental section, and the independent opening and closing of each suction port is controlled by the lateral valve; each lateral suction port is connected to the second variable frequency fan through a pipe, and the lateral suction intensity and suction position are precisely controlled by the second frequency converter to simulate the smoke exhaust conditions under different lateral smoke exhaust modes;

[0018] Step 1.1.2.3, The top smoke exhaust fan system (5) includes: a third variable frequency fan, a third frequency converter, a top suction port, and a top valve; multiple top suction ports are arranged at the top of the long slope tunnel test section, and their opening and closing states are controlled by the top valve; each top suction port is connected to the third variable frequency fan through a pipe, and the independent emission or linkage emission mode of multiple smoke exhaust ports is switched through the third frequency converter to simulate the smoke exhaust conditions under different top smoke exhaust modes;

[0019] Step 1.1.3: The lifting multi-category fire source module includes: an electric lifting platform (6) and a multi-category fire source module (7);

[0020] Step 1.1.3.1: The electric lifting platform (6) adjusts the height of the fire source to simulate different fire sources;

[0021] Step 1.1.3.2: The multi-category ignition module (7) is used to switch between three types of ignition systems: solid, liquid, and gas. The solid ignition system includes an electronic balance and a solid burner. The liquid ignition system includes an electronic balance and a liquid burner. The gas ignition system includes a flow meter, a gas burner, and a connecting hose.

[0022] Step 1.1.4: The fire parameter acquisition unit includes: a thermocouple array (8), a heat flow meter array (9), a digital camera (10), and a computer (11); the thermocouple array (8) is arranged at a spacing of 0.1 m along the tunnel centerline on the long slope tunnel test section for collecting temperature distribution; the heat flow meter array (9) is arranged at a spacing of 0.3 m along the tunnel centerline on the long slope tunnel test section for collecting heat flow distribution; the digital camera (10) is used to collect fire images; the computer (11) is used to collect solid and liquid fuel mass changes, as well as temperature and heat flow data;

[0023] Step 1.2: Construct a long and steep tunnel fire simulation dataset using a scaled-down long and steep tunnel fire simulation device, including: a dataset of physical information parameters of long and steep tunnels. Long slope tunnel fire image dataset Dataset of experimental values ​​for fires in long and steep tunnels ;

[0024] Step 1.2.1, Obtain the first i Physical information parameters of long and steep tunnels under various working conditions were collected, including: tunnel slope, atmospheric pressure, fuel type, ignition source height, longitudinal ventilation velocity, lateral suction velocity, top suction velocity, burner aspect ratio, location coordinates of temperature and heat flux measurement points, and their time series distribution; thereby constructing a set of tunnel physical information parameters. ,in, Indicates the first i A set of physical information parameters of the tunnel under various working conditions, and ; Indicates the first i Under the first working condition The first moment e One tunnel physical information parameter; Indicates the first i Total number of data collection times under each working condition; I Indicates the total number of operating conditions;

[0025] Step 1.2.2: Use a digital camera (10) to acquire a dataset of images of fires in long, steep tunnels. ,in, Indicates the first i A dataset of tunnel fire images under various working conditions, and , Indicates the first i Under the first working condition m Images of the tunnel fire at that moment; H and W It is the height and width of the tunnel fire image. C This represents the number of channels in the tunnel fire image;

[0026] Step 1.2.3: Use an electronic balance or flow meter to measure the mass change of solid / liquid fuel or the flow rate change of gas under different operating conditions and calculate the experimental value of heat release rate. Use a heat flow meter and thermocouple array to obtain the experimental values ​​of temperature distribution and heat flow distribution, respectively, thereby constructing a dataset for fire experiments in long and steep tunnels. ,in, Indicates the first i Dataset of experimental values ​​of tunnel fire under various working conditions, and ;in, Indicates the first i Under the first working condition m Experimental value of heat release rate at time. Indicates the first i Under the first working condition m Temperature experimental value at time, Indicates the first i Under the first working condition m Experimental values ​​of heat flux at any given time; Indicates the first i The total number of moments under each working condition.

[0027] Furthermore, step 2 includes:

[0028] Step 2.1: The physical information encoding module uses the first fully connected layer to... After performing the dimensionality upgrade operation, we obtain the first... Under the first working condition m Physical information parameters characteristics at time Then, the ReLU activation function is applied to... Perform a nonlinear transformation to obtain the first... i Under the first working condition m Physical information at time point nonlinear parameter characteristics , Then it is input into the second fully connected layer for processing, to obtain the... i Under the first working condition m Physical information feature fusion vector at time step ;

[0029] Step 2.2, the flame image feature extraction module includes: a flame image embedding layer, a Transformer encoder, and an MLP head, for... Perform feature extraction processing to obtain the first... i Under the first working condition m Flame feature vector at time step ;

[0030] Step 2.3, and The input is fed into the HRR feature fusion module for a series of processes, including splicing and dimensionality reduction, to obtain the [missing information]. Under the first working condition Flame parameter feature fusion vector at time step :

[0031] Step 2.4, Input into the HRR prediction header to obtain the first... i Under the first working condition m One-dimensional heat release rate prediction at time t. ;

[0032] Step 2.5: The T and Heat flux feature fusion module includes: an HRR embedding layer and a position embedding layer;

[0033] Step 2.6, and The data are input into the T prediction header and the Heat Flux prediction header respectively for processing, and the results are obtained respectively. Under the first working condition One-dimensional temperature prediction values ​​corresponding to different position coordinates at different times and heat flow prediction values .

[0034] Furthermore, step 2.2 includes:

[0035] Step 2.2.1, the flame image embedding layer... Perform position embedding processing to obtain the first i Under the first working condition m Vector sequence of embedded location information at time step ;

[0036] Step 2.2.2, the Transformer encoder... Encode to obtain the first i Under the first working condition m The first moment b Embedded vectors ;

[0037] Step 2.2.3 Residual joins and layer normalization are performed in the input MLP header, and the output is the first... i Under the first working condition m Flame feature vector at time step .

[0038] Furthermore, step 2.2.1 includes:

[0039] Step 2.2.1.1, Divided into h A fixed-size image block ,in, Indicates the firsti Under the first working condition m The first moment b Image blocks, h Indicates the number of image patches, and , B Indicates the size of the image patch, Flattened into a length of a one-dimensional vector ;

[0040] Step 2.2.1.2, using equation (1) Perform a linear transformation to obtain the first... i Under the first working condition m The first moment b Embedded vectors , The dimension of the embedding vector:

[0041] (1)

[0042] In equation (1), The weight matrix is ​​the linear transformation matrix. It is the bias vector;

[0043] Step 2.2.1.3: Using equation (2) to obtain The final embedding vector representation Thus, a vector sequence with embedded location information is obtained. :

[0044] (2)

[0045] In equation (2), Represents the position embedding matrix The first in Row position embedding vector.

[0046] Furthermore, step 2.2.2 includes:

[0047] Step 2.2.2.1: Using equation (3), obtain the first... i Under the first working condition m The first moment b query vectors , No. b Key vectors and the b Value vectors :

[0048] (3)

[0049] In equation (3), These are the three weight matrices to be learned;

[0050] Step 2.2.2.2: Using equation (4) to obtain With the i Under the first working condition m The first moment c Key vectors Attention weights between :

[0051] (4)

[0052] In equation (4), Indicates the first i Under the first working condition m The first moment u One key vector; It is the scaling factor;

[0053] Step 2.2.2.3: Using equation (5), we obtain the first... i Under the first working condition m The first moment b The output of a self-attention mechanism ;

[0054] (5)

[0055] In equation (5), Indicates the first i Under the first working condition m The c-th value vector at time t;

[0056] Step 2.2.2.4: Using equation (6), we obtain the first... i Under the first working condition m The b-th embedding vector at time b ;

[0057] (6)

[0058] In equation (6), Presentation layer normalization operation; Indicates residual connection;

[0059] Step 2.2.2.5 Input feedforward neural network Thus, by using equation (7), we can obtain the first... i Under the first working condition m The first moment b Embedded vectors :

[0060] (7)

[0061] In equation (7), yes The weight matrices of the two linear layers, yes The bias vectors of the two linear layers, This indicates a residual connection.

[0062] Furthermore, step 2.5 includes:

[0063] Step 2.5.1, the HRR embedding layer will Convert to the first i Under the first working condition m The heat release rate eigenvector at time t. ;

[0064] Step 2.5.2, the position embedding layer will... i Under the first working condition m Convert the coordinates of the temperature and heat flow measurement points at time 1 to the position coordinates of the first time. i Under the first working condition m Temperature position coordinate feature vector at time and heat flow location coordinate eigenvector ;

[0065] Step 2.5.3, will , Separately and , After splicing, we get the first... i Under the first working condition m Temperature feature fusion vector at time point Fusion vector with heat flow features .

[0066] Furthermore, step 3 includes:

[0067] Step 3.1: Construct the first equation using equation (8). i Heat release rate loss function under each operating condition :

[0068] (8)

[0069] Step 3.2: Construct the first equation using equation (9). i Temperature distribution loss function under various operating conditions :

[0070] (9)

[0071] Step 3.3: Construct the first equation using equation (10). i Heat flux distribution loss function under various operating conditions :

[0072] (10)

[0073] Step 3.4: Construct the total loss function of the fire parameter prediction network for long and steep tunnels using equation (11). :

[0074] (11)

[0075] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program that supports the processor in executing the real-time prediction method for fire parameters in long and steep tunnels, and the processor is configured to execute the program stored in the memory.

[0076] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, performs the steps of the real-time prediction method for fire parameters in long and steep tunnels.

[0077] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0078] 1. This invention constructs a fire simulation experimental device for long and steep tunnels under complex ventilation modes, and builds a fire dataset for long and steep tunnels that includes physical information, fire images, and experimental data. Compared with traditional experimental datasets that are limited to a single fuel or stable combustion stage, the dataset established by this invention can comprehensively cover the dynamic characteristics of multiple fire scenarios in long and steep tunnels under complex ventilation modes, providing high-quality basic support for subsequent model training, thereby effectively improving the applicability and universality of the model.

[0079] 2. This invention proposes a fire parameter prediction network based on the Vision Transformer architecture and physical information fusion. It uses a self-attention mechanism to extract features from fire images of long and steep tunnels under complex ventilation modes. This not only overcomes the shortcomings of traditional convolutional neural networks in long-distance dependency modeling, but also fully captures the spatiotemporal evolution characteristics of flames under complex ventilation conditions. Furthermore, it introduces physical information about fires in long and steep tunnels and deeply fuses it with flame image features, thereby introducing physical constraints in the prediction process and avoiding overfitting and physical inconsistencies that may occur when relying solely on data-driven models.

[0080] 3. This invention constructs a comprehensive loss function. This method enables joint prediction of heat release rate, temperature distribution, and heat flow distribution, avoiding the redundancy problem of having to build multiple models separately in traditional methods. This method not only improves prediction efficiency, but also enhances overall prediction performance through feature sharing and information complementarity between different tasks, thus achieving efficient, real-time, and accurate prediction of key parameters for fires in long and steep tunnels. Attached Figure Description

[0081] Figure 1 This is a flowchart of the fire parameter prediction network for long and steep tunnels according to the present invention.

[0082] Figure 2 This is a schematic diagram of the fire simulation test platform for a long, steep tunnel, as described in this invention.

[0083] Figure 3 This is a diagram illustrating the partitioning of the dataset for this example.

[0084] Figure 4 This is a schematic diagram of the network architecture for predicting fire parameters in a long, steep tunnel in this example.

[0085] Figure 5 This is a schematic diagram of the Vision Transformer-Small network architecture for this example;

[0086] Figure 6 This is a diagram showing the ablation experiment results of the model in this example;

[0087] Numbered in the diagram: 1. Experimental section of long and steep tunnel; 2. Hydraulic lifting system; 3. Longitudinal smoke exhaust fan system; 4. Lateral smoke exhaust fan system; 5. Top smoke exhaust fan system; 6. Electric lifting platform; 7. Fire source module; 8. Thermocouple; 9. Heat flow meter; 10. Digital camera; 11. Computer. Detailed Implementation

[0088] In this embodiment, to quickly and accurately predict the heat release rate, temperature distribution, and heat flow distribution of fires in long and steep tunnels under complex ventilation modes during the construction and maintenance phase, a real-time prediction method for fire situations in long tunnels based on image recognition and physical information fusion is proposed. This method uses a self-attention mechanism to extract flame image features under complex ventilation modes, further incorporating fire physical information and deeply fusing it with image features. This introduces physical constraints into the prediction process, avoiding overfitting and physical inconsistencies that may occur with purely data-driven models. Specifically, as... Figure 1 As shown, the method includes the following steps:

[0089] Step 1: Construct a scaled-down fire simulation device for long, steep tunnels under complex ventilation conditions, such as... Figure 2 As shown, this is used to construct a tunnel fire dataset;

[0090] Step 1.1: Construct a scaled-down fire simulation device for a long and steep tunnel under a complex ventilation mode, including: a tunnel test section, a complex ventilation and smoke extraction system, a lifting multi-type fire source module, and a fire parameter acquisition unit;

[0091] Step 1.1.1: Construct a long, steeply sloping tunnel experimental section using a high-temperature resistant stainless steel frame, with a high-temperature resistant transparent tempered glass observation window on the front; adjust the tunnel slope using a hydraulic lifting system to simulate fire scenarios at different slopes;

[0092] Step 1.1.2: The complex ventilation and smoke exhaust system includes: a longitudinal smoke exhaust fan system, a lateral smoke exhaust fan system and a top smoke exhaust fan system, which are used to simulate smoke exhaust conditions under independent or coupled ventilation modes in the experimental section of a long and steep tunnel;

[0093] Step 1.1.2.1: The longitudinal smoke exhaust fan system includes: a first variable frequency fan and a first frequency converter; the first variable frequency fan is arranged at one end of the long and steep tunnel test section, and the fan speed is adjusted by the first frequency converter to realize the continuous and controllable change of longitudinal wind speed, which is used to simulate the smoke exhaust conditions under different longitudinal ventilation modes.

[0094] Step 1.1.2.2: The lateral smoke exhaust fan system includes: a second variable frequency fan, a second frequency converter, lateral suction ports, and lateral valves; multiple lateral suction ports are arranged along the length of the sidewall of the long and steep tunnel experimental section, and the independent opening and closing of each suction port is controlled by the lateral valves; each lateral suction port is connected to the second variable frequency fan through a pipe, and the lateral suction intensity and suction position are precisely controlled by the second frequency converter to simulate the smoke exhaust conditions under different lateral smoke exhaust modes;

[0095] Step 1.1.2.3: The top smoke exhaust fan system includes: a third variable frequency fan, a third variable frequency drive, a top suction port, and a top valve; multiple top suction ports are arranged at the top of the long and steep tunnel test section, and their opening and closing status is controlled by the top valve; each top suction port is connected to the third variable frequency fan through a pipeline, and the third variable frequency drive realizes the switching between independent emission or linkage emission modes of multiple smoke exhaust ports, which is used to simulate the smoke exhaust conditions under different top smoke exhaust modes;

[0096] Step 1.1.3: The lifting multi-category fire source module includes: an electric lifting platform and a multi-category fire source module;

[0097] Step 1.1.3.1: Adjust the height of the fire source using the electric lifting platform to simulate different fire sources;

[0098] Step 1.1.3.2: The multi-category ignition module is used to switch between three types of ignition systems: solid, liquid, and gas. The solid ignition system includes an electronic balance and a solid burner; the liquid ignition system includes an electronic balance and a liquid burner; the gas ignition system includes a flow meter, a gas burner, and a connecting hose. In this embodiment, a liquid ignition system is used.

[0099] Step 1.1.4: Fire parameter acquisition unit, including: thermocouple array, heat flow meter array, digital camera and computer; in this embodiment, the heat flow meter array is arranged at 0.3 m intervals along the tunnel centerline on the long slope tunnel test section to collect heat flow distribution; the digital camera is used to acquire fire images; the computer is used to collect solid and liquid fuel mass changes and temperature and heat flow data;

[0100] Step 1.2: Construct a long and steep tunnel fire simulation dataset using a scaled-down long and steep tunnel fire simulation device, including: a dataset of physical information parameters of long and steep tunnels. Long slope tunnel fire image dataset Dataset of experimental values ​​for fires in long and steep tunnels ;

[0101] Step 1.2.1: Obtain the first i This embodiment includes physical information parameters for tunnels with different lengths and gradients under various working conditions, including: longitudinal ventilation velocity, lateral smoke exhaust velocity, burner aspect ratio, and the location coordinates of heat flow measurement points; thereby constructing a set of tunnel physical information parameters. ,in, Indicates the first i A set of physical information parameters of the tunnel under various working conditions, and ; Indicates the first i Under the first working condition m The first moment One tunnel physical information parameter; Indicates the first i Total number of data collection times under each working condition; This indicates the total number of operating conditions.

[0102] Step 1.2.2: Acquire image dataset of long and steep tunnel fires using digital camera 10. Take, among which, Indicates the first A dataset of tunnel fire images under various working conditions, and , Indicates the first i Under the first working condition m Images of the tunnel fire at that moment; H and W It is the height and width of the tunnel fire image.C This represents the number of channels in the tunnel fire image;

[0103] Step 1.2.3: Use an electronic balance to measure the mass change of liquid fuel under different operating conditions and calculate the experimental value of heat release rate. Use a heat flow meter to obtain the experimental value of heat flow distribution, thereby constructing a tunnel fire experimental dataset. ,in, Indicates the first i Dataset of experimental values ​​of tunnel fire under various working conditions, and ;in, Indicates the first i Under the first working condition m Experimental value of heat release rate at time. Indicates the first i Under the first working condition m Experimental values ​​of heat flux at any given time; Indicates the first i The total number of moments under each working condition.

[0104] Step 1.2.4: For Data preprocessing includes three aspects: data cleaning, data standardization and normalization, and data partitioning, aiming to improve data quality and enhance the efficiency and accuracy of model training.

[0105] Step 1.2.4.1: In the data cleaning stage, considering the occasional "second skipping" phenomenon in the electronic balance acquisition software used to measure mass loss during long-term recording, resulting in missing values ​​in some time series data, it is necessary to process the missing values ​​first. To ensure the continuity and accuracy of the data, this paper uses the "linear interpolation method" commonly used in machine learning to impute the missing data. This method is a widely used technique in numerical computation and data modeling. Its basic idea is to perform linear fitting between adjacent data points to construct a continuous function to estimate the missing values.

[0106] Secondly, during the data standardization phase, considering the consistency requirements of the model input for image dimensions, the original image data was first uniformly adjusted to a standard size of 224 × 224, such as... Figure 3 As shown. This size is widely applicable to various classic convolutional neural network models, effectively improving computational efficiency and memory utilization while ensuring that key image features are not destroyed. Next, the image pixel values ​​were standardized to make their distribution close to a standard normal distribution (mean of 0, standard deviation of 1) to accelerate the gradient descent process of the model and improve training stability and generalization ability.

[0107] Finally, the dataset was divided into training, validation, and test sets in a 6:2:2 ratio, as follows: Figure 3As shown in the figure, this partitioning strategy, while ensuring sufficient model learning, provides a scientific basis for performance evaluation and parameter tuning, thereby helping to improve the model's generalization ability and robustness in practical applications.

[0108] Step 2: Construct a long-slope fire parameter prediction network based on the Vision Transformer architecture. The model architecture diagram is shown below. Figure 4 As shown, it includes: a physical information encoding module, a flame image feature extraction module, an HRR feature fusion module, an HRR prediction head, a heat flux feature fusion module, and a heat flux prediction head, which respectively address the following: and Encoding and feature extraction are performed to obtain the first... Under various working conditions Physical information fusion vector at time step and flame feature vector and will and The flame parameter feature fusion vector is obtained by performing fusion processing. ,Will Input the corresponding prediction head to obtain the predicted value of heat release rate. and heat flow prediction values .

[0109] Step 2.1: The physical information encoding module uses the first fully connected layer to... After performing the dimensionality upgrade operation, we obtain the first... Under the first working condition m The physical information parameters at time step are then processed using the ReLU activation function. Perform a nonlinear transformation to obtain the first... Under the first working condition Physical information at time nonlinear parameter characteristics , Then it is input into the second fully connected layer for processing, to obtain the... i Under the first working condition m Physical information feature fusion vector at time step ;

[0110] Step 2.2: In this example, the flame image feature extraction module is ViT-Small, and its structural diagram is shown below. Figure 5 As shown, it includes a flame image embedding layer, a Transformer encoder, and an MLP head, for... Perform feature extraction processing to obtain the first... Under the first working condition Flame feature vector at time step .

[0111] Step 2.2.1: Flame Image Embedding Layer Pair Perform position embedding processing to obtain the first i Under the first working condition m Vector sequence of embedded location information at time step ;

[0112] Step 2.2.1.1: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Divided into A fixed-size image block ,in, Indicates the first Under the first working condition The first moment Image blocks, h Indicates the number of image patches, and , Indicates the size of the image patch, Flattened into a length of a one-dimensional vector ;

[0113] Step 2.2.1.2: Use equation (1) to... Perform a linear transformation to obtain the first... Under the first working condition The b-th embedding vector at time b , The dimension of the embedding vector:

[0114] (1)

[0115] In equation (1), The weight matrix is ​​the linear transformation matrix. This is the bias vector.

[0116] Step 2.2.1.3: Obtain using equation (2) The final embedding vector representation Thus, a vector sequence with embedded location information is obtained. :

[0117] (2)

[0118] In equation (2), Represents the position embedding matrix The first in Row position embedding vector;

[0119] Step 2.2.2: Transformer encoder pair Encode to obtain the first i Under the first working condition m The first moment b Embedded vectors ;

[0120] Step 2.2.2.1: Use equation (3) to obtain the first... i Under the first working condition m The first moment b query vectors , No. b Key vectors and the b Value vectors :

[0121] (3)

[0122] In equation (3), These are the three weight matrices to be learned.

[0123] Step 2.2.2.2: Obtain using equation (4) With the Under the first working condition The c-th key vector at time c Attention weights between :

[0124] (4)

[0125] In equation (4), Indicates the first i Under the first working condition m The first moment u One key vector; It is the scaling factor;

[0126] Step 2.2.2.3: Use equation (5) to obtain the first... i Under the first working condition m The first moment b The output of a self-attention mechanism ;

[0127] (5)

[0128] In equation (5), Indicates the first Under the first working condition The c-th value vector at time c.

[0129] Step 2.2.2.4: Use equation (6) to obtain the first... Under the first working condition The b-th embedding vector at time b ;

[0130] (6)

[0131] In equation (6), Presentation layer normalization operation; Indicates residual connection;

[0132] Step 2.2.2.5: Input feedforward neural network Thus, by using equation (7), we can obtain the first... i Under the first working condition m The first moment b Embedded vectors :

[0133] (7)

[0134] In equation (7), yes The weight matrices of the two linear layers, yes The bias vectors of the two linear layers, Indicates residual connection;

[0135] Step 2.2.3: Residual joins and layer normalization are performed in the input MLP header, and the output is the first... i Under the first working condition m Flame feature vector at time step ;

[0136] Step 2.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and The input is fed into the HRR feature fusion module for a series of processes, including splicing and dimensionality reduction, to obtain the [missing information]. Under the first working condition 256-dimensional flame parameter feature fusion vector at time step :

[0137] Step 2.4: Input into the HRR prediction header to obtain the first... Under the first working condition One-dimensional heat release rate prediction at time t. .

[0138] Step 2.5: The Heat flux feature fusion module includes: HRR embedding layer and location embedding layer;

[0139] Step 2.5.1: The HRR embedding layer will Convert to the first Under the first working condition 32-dimensional heat release rate eigenvector at time step ;

[0140] Step 2.5.2: [The remaining text appears to be incomplete and requires further context.] Under the first working condition The location coordinates of the heat flow measurement point at any given time are converted into a heat flow feature fusion vector. ;

[0141] Step 2.5.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation , and After splicing, we get the first... Under the first working condition 304-dimensional heat flux feature fusion vector at time step .

[0142] Step 2.6: The data is input into the Heat flux prediction header for processing to obtain the first... Under the first working condition One-dimensional heat flux prediction values ​​corresponding to different position coordinates at different times .

[0143] Step 3: Construct the total loss function for the fire parameter prediction network , Including heat release rate loss function and heat flux distribution loss function ;

[0144] Step 3.1: Construct the first equation using equation (8). i Heat release rate loss function under each operating condition :

[0145] (8)

[0146] Step 3.2: Construct the first equation using equation (9). i Heat flux distribution loss function under various operating conditions :

[0147] (9)

[0148] Step 3.3: Construct the total loss function of the fire parameter prediction network for long and steep tunnels using equation (10). :

[0149] (10)

[0150] Step 4: Utilize A parameter prediction network for fires on long slopes is trained, and the network parameters are continuously adjusted using the backpropagation algorithm until... The model converged to obtain the optimal prediction model for long-slope fire parameters. The key hyperparameters of the final model were set as follows: learning rate of 0.0005, batch size of 128, and epochs of 120. Experiments were conducted on a system with an Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz (2 processors), 256 GB of RAM, and an NVIDIA GeForce RTX 4090 D (with 24 GB of dedicated memory). This prediction model can output predicted values ​​of heat release rate and heat flow distribution under corresponding working conditions based on the input tunnel flame image, thus quickly and accurately outputting key fire parameters, significantly shortening the prediction time, and effectively overcoming the limitation of traditional physical modeling, which is only applicable to the stable combustion stage, achieving reliable prediction under complex unsteady-state conditions.

[0151] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the methods described above, and the processor is configured to execute the program stored in the memory.

[0152] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0153] To verify the advancement of the prediction framework proposed in this study, we conducted the following ablation experiments. Specifically, without fusing any physical information, we used only flame images as input to predict the heat release rate (HRR) and heat flux distribution, respectively. The experimental results are as follows: Figure 6 As shown, in the heat release rate (HRR) prediction task, the baseline model without fused physical information performed slightly worse than the final improved model with fused physical information. This phenomenon indicates that flame images already contain a large number of physical features highly correlated with HRR, thus making HRR prediction based solely on visual information feasible to some extent. In contrast, in the heat flux distribution prediction task, the improved model with fused physical information significantly outperformed the baseline model. This is mainly attributed to the significant spatial differences in heat flux distribution; the introduction of physical information such as location coordinates significantly enhanced the model's ability to model spatial distribution characteristics.

Claims

1. A method for real-time prediction of fire situation in long tunnels based on image recognition and physical information fusion, characterized in that, Includes the following steps: Step 1: Build a scaled-down fire simulation device for long and steep tunnels under complex ventilation modes to construct a fire dataset for long and steep tunnels, including: a physical information parameter dataset P for long and steep tunnels, a fire image dataset X for long and steep tunnels, and a fire experimental value dataset Y for long and steep tunnels. Step 2: Construct a fire parameter prediction network for long and steep tunnels based on the Vision Transformer architecture, including: a physical information encoding module, a flame image feature extraction module, an HRR feature fusion module, an HRR prediction head, a T and heat flux feature fusion module, and a T and heat flux prediction head. Encode and extract features from P and X respectively to obtain the physical information fusion vector at time m under the i-th working condition. and flame feature vector and will and After fusion processing, the flame parameter feature fusion vector at time m under the i-th working condition is obtained. ,Will Input the HRR prediction header to obtain the predicted heat release rate at time m under the i-th operating condition. The HRR embedding layer will Convert to the heat release rate feature vector at time m under the i-th operating condition. ;Will , With the corresponding temperature location coordinate feature vector The fusion yields the temperature feature fusion vector at time m under the i-th operating condition. At the same time, , With the corresponding heat flow location coordinate eigenvector The fusion yields a heat flow feature fusion vector. ;Will and The values ​​are input into the T prediction head and the Heat Flux prediction head respectively, and the predicted temperature value at time m under the i-th operating condition is obtained accordingly. and heat flow prediction ; Step 3: Construct the total loss function for the fire parameter prediction network in long and steep tunnels. The Including heat release rate loss function Temperature distribution loss function and heat flux distribution loss function ; Step 4, using the above The fire parameter prediction network is trained, and the network parameters are continuously adjusted through the backpropagation algorithm until the desired parameters are obtained. The model converges to obtain the optimal fire parameter prediction model; it is used to predict the fire images and physical information of long and steep tunnels, and outputs the corresponding predicted values ​​of heat release rate, temperature distribution and heat flow distribution in real time.

2. The method for real-time prediction of fire situation in long tunnels based on image recognition and physical information fusion as described in claim 1, characterized in that, Step 1 includes: Step 1.1: Construct a scaled-down fire simulation device for a long, steep tunnel under a complex ventilation mode, including: a long, steep tunnel, a complex ventilation and smoke extraction system, a lifting multi-type fire source module, and a fire parameter acquisition unit; Step 1.1.1: The long gradient tunnel includes a long gradient tunnel test section and a hydraulic lifting system; the long gradient tunnel test section is constructed with a high-temperature resistant stainless steel frame, and a high-temperature resistant transparent tempered glass observation window is set on the front; the tunnel gradient is adjusted by the hydraulic lifting system to simulate fire scenarios under different gradients; Step 1.1.2: The complex ventilation and smoke exhaust system includes a longitudinal smoke exhaust fan system, a lateral smoke exhaust fan system and a top smoke exhaust fan system, which are used to simulate smoke exhaust conditions under independent or coupled ventilation modes in the experimental section of a long and steep tunnel. Step 1.1.2.1: The longitudinal smoke exhaust fan system includes: a first variable frequency fan and a first frequency converter; the first variable frequency fan is arranged at one end of the long and steep tunnel test section, and the fan speed is adjusted by the first frequency converter to realize the continuous and controllable change of longitudinal wind speed, which is used to simulate the smoke exhaust conditions under different longitudinal ventilation modes. Step 1.1.2.2: The lateral smoke exhaust fan system includes: a second variable frequency fan, a second frequency converter, lateral suction ports, and lateral valves; multiple lateral suction ports are arranged along the length of the sidewall of the long slope tunnel experimental section, and the independent opening and closing of each suction port is controlled by the lateral valves; each lateral suction port is connected to the second variable frequency fan through a pipe, and the lateral suction intensity and suction position are precisely controlled by the second frequency converter to simulate the smoke exhaust conditions under different lateral smoke exhaust modes; Step 1.1.2.3: The top smoke exhaust fan system includes: a third variable frequency fan, a third frequency converter, a top suction port, and a top valve; multiple top suction ports are arranged at the top of the long and steep tunnel test section, and their opening and closing status is controlled by the top valve; each top suction port is connected to the third variable frequency fan through a pipe, and the third frequency converter realizes the switching of independent or linked emission modes of multiple smoke exhaust ports, which is used to simulate the smoke exhaust conditions under different top smoke exhaust modes; Step 1.1.3: The lifting multi-category fire source module includes: an electric lifting platform and a multi-category fire source module; Step 1.1.3.1: The electric lifting platform adjusts the height of the fire source to simulate different fire sources; Step 1.1.3.2: The multi-type ignition source module is used to switch between three types of ignition source systems: solid, liquid, and gas. The solid ignition source system includes an electronic balance and a solid burner; the liquid ignition source system includes an electronic balance and a liquid burner; and the gas ignition source system includes a flow meter, a gas burner, and a connecting hose. Step 1.1.4: The fire parameter acquisition unit includes: a thermocouple array, a heat flow meter array, a digital camera, and a computer; the thermocouple array is arranged at 0.1 m intervals along the tunnel centerline on the long, steep tunnel test section for acquiring temperature distribution; the heat flow meter array is arranged at 0.3 m intervals along the tunnel centerline on the long, steep tunnel test section for acquiring heat flow distribution; the digital camera is used to acquire fire images; the computer is used to acquire data on changes in the mass of solid and liquid fuels, as well as temperature and heat flow data; Step 1.2: Construct a long and steep tunnel fire simulation dataset using a scaled-down long and steep tunnel fire simulation device, including: a dataset of physical information parameters of long and steep tunnels. Long slope tunnel fire image dataset Dataset of experimental values ​​for fires in long and steep tunnels ; Step 1.2.1: Obtain the physical information parameters of tunnels with different lengths and gradients under the i-th working condition, including: tunnel gradient, atmospheric pressure, fuel type, ignition source height, longitudinal ventilation velocity, lateral suction velocity, top suction velocity, burner aspect ratio, location coordinates of temperature and heat flow measurement points, and time series distribution; thereby constructing a set of tunnel physical information parameters. ,in, Let represent the set of tunnel physical information parameters under the i-th working condition, and ; Indicates the i-th working condition. The e-th tunnel physical information parameter at time e; I represents the total number of data collection times under the i-th operating condition; I represents the total number of operating conditions. Step 1.2.2: Use a digital camera (10) to acquire a dataset of images of fires in long, steep tunnels. ,in, Let represent the image dataset of a tunnel fire under the i-th working condition, and , This represents the tunnel fire image at time m under the i-th working condition; H and W are the height and width of the tunnel fire image, and C is the number of channels in the tunnel fire image; Step 1.2.3: Use an electronic balance or flow meter to measure the mass change of solid / liquid fuel or the flow rate change of gas under different operating conditions and calculate the experimental value of heat release rate. Use a heat flow meter and thermocouple array to obtain the experimental values ​​of temperature distribution and heat flow distribution, respectively, thereby constructing a dataset for fire experiments in long and steep tunnels. ,in, Let represent the dataset of experimental values ​​for tunnel fires under the i-th working condition, and ;in, This represents the experimental value of the heat release rate at time m under the i-th operating condition. This represents the experimental temperature value at time m under the i-th operating condition. This represents the experimental heat flux value at time m under the i-th operating condition; This represents the total number of moments under the i-th operating condition.

3. The method for real-time prediction of fire situation in long tunnels based on image recognition and physical information fusion as described in claim 2, characterized in that, Step 2 includes: Step 2.1: The physical information encoding module uses the first fully connected layer to... After performing the dimensionality upgrade operation, we obtain the first... Physical information parameter characteristics at time m under various working conditions Then, the ReLU activation function is applied to... By performing a nonlinear transformation, the physical information nonlinear parameter characteristics at time m under the i-th working condition are obtained. , The data is then input into the second fully connected layer for processing to obtain the physical information feature fusion vector at time m under the i-th working condition. ; Step 2.2, the flame image feature extraction module includes: a flame image embedding layer, a Transformer encoder, and an MLP head, for... Feature extraction is performed to obtain the flame feature vector at time m under the i-th working condition. ; Step 2.3, and The input is fed into the HRR feature fusion module for concatenation and dimensionality reduction processing to obtain the first... Under the first working condition Flame parameter feature fusion vector at time step : Step 2.4, The data is input into the HRR prediction header to obtain the one-dimensional heat release rate prediction value at time m under the i-th operating condition. ; Step 2.5: The T and Heat flux feature fusion module includes: an HRR embedding layer and a location embedding layer; Step 2.6, and The data are input into the T prediction header and the Heat Flux prediction header respectively for processing, and the results are obtained respectively. Under the first working condition One-dimensional temperature prediction values ​​corresponding to different position coordinates at different times and heat flow prediction .

4. The method for real-time prediction of fire situation in long tunnels based on image recognition and physical information fusion as described in claim 3, characterized in that, Step 2.2 includes: Step 2.2.1, the flame image embedding layer... Perform position embedding processing to obtain a vector sequence of embedded position information at time m under the i-th working condition. ; Step 2.2.2, Transformer encoder pair Encode the vector to obtain the b-th embedding vector at time m under the i-th working condition. h represents the number of image patches; Step 2.2.3 The input MLP header is processed through residual connections and layer normalization, and the output is the flame feature vector at time m under the i-th working condition. .

5. The method for real-time prediction of fire situation in long tunnels based on image recognition and physical information fusion as described in claim 4, characterized in that, Step 2.2.1 includes: Step 2.2.1.1, Divide the image into h fixed-size blocks ,in, Let represent the b-th image patch at time m under the i-th working condition, and h represent the number of image patches. B represents the size of the image patch. Flattened into a length of a one-dimensional vector ; Step 2.2.1.2, using equation (1) Perform a linear transformation to obtain the b-th embedding vector at time m under the i-th working condition. , The dimension of the embedding vector: (1) In equation (1), The weight matrix is ​​the linear transformation matrix. It is the bias vector; Step 2.2.1.3: Using equation (2) to obtain The final embedding vector representation Thus, a vector sequence with embedded location information is obtained. : (2) In equation (2), Represents the position embedding matrix The first in Row position embedding vector.

6. The method for real-time prediction of fire situation in long tunnels based on image recognition and physical information fusion as described in claim 5, characterized in that, Step 2.2.2 includes: Step 2.2.2.1: Use equation (3) to obtain the b-th query vector at time m under the i-th working condition. The b-th key vector and the b-th value vector : (3) In equation (3), These are the three weight matrices to be learned; Step 2.2.2.2: Using equation (4) to obtain With the c-th key vector at time m under the i-th working condition Attention weights between : (4) In equation (4), This represents the u-th key vector at time m under the i-th working condition; It is the scaling factor; Step 2.2.2.3: Use equation (5) to obtain the output of the b-th self-attention mechanism at time m under the i-th working condition. ; (5) In equation (5), This represents the c-th value vector at time m under the i-th working condition; Step 2.2.2.4: Use equation (6) to obtain the b-th embedding vector at time m under the i-th working condition. ; (6) In equation (6), Presentation layer normalization operation; Indicates residual connection; Step 2.2.2.5 Input feedforward neural network Thus, the b-th embedding vector at time m under the i-th working condition is obtained using equation (7). : (7) In equation (7), yes The weight matrices of the two linear layers, yes The bias vectors of the two linear layers, This indicates a residual connection.

7. A method for real-time prediction of fire situation in long tunnels based on image recognition and physical information fusion as described in claim 6, characterized in that, Step 2.5 includes: Step 2.5.1, the HRR embedding layer will Convert to the heat release rate feature vector at time m under the i-th operating condition. ; Step 2.5.2: The location embedding layer converts the location coordinates of the temperature and heat flow measurement points at time m under the i-th operating condition into a temperature location coordinate feature vector at time m under the i-th operating condition. and heat flow location coordinate eigenvector ; Step 2.5.3, will , Separately and , After concatenation, the temperature feature fusion vector at time m under the i-th operating condition is obtained. Fusion vector with heat flow features .

8. A method for real-time prediction of fire situation in long tunnels based on image recognition and physical information fusion as described in claim 7, characterized in that, Step 3 includes: Step 3.1: Construct the heat release rate loss function under the i-th operating condition using equation (8). : (8) Step 3.2: Construct the temperature distribution loss function for the i-th operating condition using equation (9). : (9) Step 3.3: Construct the heat flow distribution loss function for the i-th operating condition using equation (10). : (10) Step 3.4: Construct the total loss function of the fire parameter prediction network for long and steep tunnels using equation (11). : (11)。 9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the real-time prediction method for fire parameters in long and steep tunnels as described in any one of claims 1-8, and the processor is configured to execute the program stored in the memory.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the real-time prediction method for fire parameters in long and steep tunnels as described in any one of claims 1-8.

Images