Tunnel fire spatiotemporal evolution trend perception method based on sparse temperature data

Abstract

The application discloses a tunnel flowing fire space-time evolution situation awareness method based on sparse temperature data and belongs to the technical field of fire safety and artificial intelligence. The application adopts an asymmetrically arranged sparse temperature sensor array, collects sparse temperature data along a tunnel central axis and a single side wall, extracts a global feature vector based on a time series neural network model, maps the global feature vector into a fire source probability distribution matrix of a tunnel ground two-dimensional grid space based on a spatial form reconstruction model, and embeds spatial continuity, boundary constraints and other physical priors in a training process. The application realizes cross-physical quantity real-time reconstruction from sparse temperature observation to two-dimensional planar space form of the flowing fire, overcomes the limitation that traditional methods can only invert point source positions or temperature fields, can output dynamic spreading distribution of the flowing fire in real time under sparse observation conditions, and provides intuitive and reliable situation information for tunnel fire emergency response.

Description

Technical Field

[0001] This invention relates to the fields of fire safety engineering and artificial intelligence technology, and in particular to a method for sensing the spatiotemporal evolution of tunnel flowing fire based on sparse temperature data. Background Technology

[0002] Tunnels, as vital transportation links connecting different regions, play a crucial role in improving traffic efficiency and promoting regional economic development. However, due to the relatively enclosed space and limited ventilation in tunnels, smoke and heat cannot dissipate quickly enough in the event of a fire, easily leading to severe casualties and structural damage. Therefore, tunnel fire safety has gradually become a key research focus in the fields of transportation engineering and fire safety.

[0003] Tunnel fires are caused by a variety of factors, including traffic accidents, vehicle fires, electrical system failures, and leaks during the transportation of hazardous chemicals. Among these, flowing fires, which are formed when hazardous liquid fuels leak and spread on the ground and ignite, are the most destructive type of tunnel fire due to their dynamic changes in the combustion area and unpredictable spread paths. The combustion process of flowing fires is influenced by the coupled effects of multiple factors, including fuel properties, ground slope, tunnel ventilation conditions, and boundary constraints. Their spatial morphology and spread process exhibit significant nonlinear characteristics and randomness, making the fire situation difficult to predict and control.

[0004] Existing research on tunnel fire behavior primarily relies on experimental studies and numerical simulations to establish predictive models by analyzing fire development patterns. For example, fire simulation methods based on computational fluid dynamics (CFD) can accurately describe smoke flow and temperature distribution during a fire; however, these methods are typically computationally expensive and difficult to meet real-time prediction requirements. Meanwhile, while experimental methods can provide highly reliable data, they are limited by experimental costs and safety constraints, making large-scale, multi-condition system studies difficult. Therefore, traditional methods suffer from insufficient real-time performance and difficulties in data acquisition in practical engineering applications.

[0005] With the development of big data and artificial intelligence technologies, data-driven fire prediction methods have gradually become a research hotspot. The fire process is essentially a typical spatiotemporal coupling problem, and its sensor data (such as temperature and smoke concentration) usually exhibits obvious time-series characteristics. Researchers typically use models such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and Transformers to model time-series data such as temperature and smoke. Among these, the Transformer model has significant advantages in handling long-term dependencies and can effectively extract the temporal features in the fire development process. In terms of spatial information reconstruction, convolutional neural networks (CNNs) and transposed convolutional neural networks (TCNNs) are widely used to map low-dimensional sensor data into high-dimensional spatial distribution information.

[0006] Existing research has demonstrated the feasibility of the "temporal feature extraction + spatial dimensionality enhancement reconstruction" approach by combining time-series models with TCNN structures to predict the temperature field distribution in tunnel fires and to invert parameters such as fire source location and heat release rate. However, most of these methods still fall under the category of reconstruction problems within the same physical quantity system, i.e., continuously reconstructing the temperature field based on a subset of temperature measurement points. Essentially, it is a "temperature → temperature" mapping process. Since the temperature field is a continuous physical field, its distribution is smooth and satisfies physical laws such as heat conduction and convection. Therefore, this type of reconstruction problem typically relies on strong continuity assumptions and certain physical constraints.

[0007] In contrast, flowing fires in tunnels exhibit significantly different physical characteristics. The spatial distribution of flowing fires is determined by the coupled effects of multiple factors, including fuel diffusion, combustion reaction, and ventilation conditions. Their fire source areas typically exhibit a planar structure that dynamically evolves over time, with distinct boundary features and discontinuities. From a data representation perspective, the morphology of flowing fires is better described as spatial structural information with a 0 / 1 distribution characteristic, rather than a continuous physical field. Therefore, inverting the morphology of flowing fires from temperature observation data at the tunnel top is essentially a cross-physical quantity inversion problem (temperature → morphology), meaning that the spatial distribution of the fire source is inferred from the indirect observation of temperature, and there is no direct physical closed-loop mapping relationship between them.

[0008] Furthermore, in practical engineering, the number of temperature sensors inside tunnels is limited, and the observation data exhibits significant sparsity and incompleteness, while the spatial distribution of flowing fire has high dimensionality (such as a two-dimensional grid distribution). Therefore, this problem also has the characteristic of inferring high-dimensional structural information from low-dimensional observation data, belonging to a typical complex inverse problem. Its solution is often not unique, and it places higher demands on the stability and physical consistency of the model.

[0009] However, most existing studies are based on the "single-point fire source" assumption, simplifying the fire source to a single point or equivalent heat source, mainly focusing on scalar or low-dimensional characteristics such as fire source location, intensity, or temperature field distribution. For tunnel flowing fires, a type of fire with significant spatial expansion characteristics, the fire source exhibits a planar structure that dynamically evolves over time, making it difficult for existing methods to effectively describe its spatial morphology and spread process. Furthermore, the spread of flowing fires is influenced by the coupled effects of multiple factors such as ventilation, slope, and boundary conditions, making its spatiotemporal evolution process more complex. Existing models lack specific modeling methods for this type of fire. Therefore, how to achieve efficient inversion of the spatial morphology of flowing fires based on existing temporal modeling and spatial reconstruction methods remains a subject for further research.

[0010] Therefore, there is an urgent need to propose a technical solution that can make real-time predictions of the spatial morphology and dynamic evolution of flowing fire in tunnels based on limited sensor data. Summary of the Invention

[0011] To overcome the shortcomings of existing technologies, this invention provides a method for situational awareness of the spatiotemporal evolution of tunneling flowing fire based on sparse temperature data. This invention constructs a mapping relationship between temperature characteristics and the spatial morphology of flowing fire, enabling real-time reconstruction (inversion) of the fire source spatial structure from temperature data under sparse temperature observation conditions. Its core is to complete high-dimensional structural inference across physical quantities under physical constraints.

[0012] The technical solution of the present invention is as follows:

[0013] This invention provides a method for situational awareness of the spatiotemporal evolution of tunneling fires based on a sparse temperature sensor array, comprising the following steps:

[0014] 1) Construct a training dataset containing sparse temperature observation sequences and fire source labels; wherein, the sparse temperature observation sequences are obtained by multiple temperature sensor groups deployed along the longitudinal direction of the tunnel, each temperature sensor group includes two sensors, used to obtain transverse temperature data at the same longitudinal position of the tunnel, and the fire source labels are two-dimensional matrices used to indicate whether there is a flame in each grid cell in the tunnel.

[0015] 2) Construct a joint model that includes a temporal neural network model and a spatial morphology reconstruction model;

[0016] The joint model constructs physical features reflecting the spatial distribution characteristics of fire sources based on sparse temperature observation sequences. The original temperature data and the constructed physical features are concatenated to form an enhanced temporal input tensor, which is then input into a temporal neural network model to obtain a global feature vector. The global feature vector is then input into a spatial morphology reconstruction model to output a fire source probability distribution matrix corresponding to the spatial size of the tunnel grid.

[0017] 3) The joint model is trained using the training dataset. During the training process, physical constraints are applied to the fire source probability distribution matrix to suppress non-physical prediction results.

[0018] 4) When flowing fire occurs, temperature data inside the tunnel is collected in real time, and the real-time fire source probability distribution matrix is ​​obtained using the trained joint model, so as to realize the real-time reconstruction of the two-dimensional spatial morphology of flowing fire in the tunnel across physical quantities.

[0019] According to a preferred embodiment of the present invention, the temperature sensor groups are arranged at a preset interval, and each temperature sensor group consists of a first sensor located on the central axis of the tunnel and a second sensor located at a preset position between the central axis of the tunnel and a single side wall.

[0020] According to a preferred embodiment of the present invention, the data in the training dataset is obtained by a flowing fire simulation experiment. During the experiment, temperature data inside the tunnel is acquired in real time by a temperature sensor, and video of the flowing fire development process is acquired by a camera.

[0021] For each frame of video captured by the camera, the flame region is identified by methods based on brightness threshold, color space segmentation or image semantic segmentation, and the original flame image is converted into a binary image.

[0022] Then, the flame recognition results of the target frame and several adjacent video frames before and after it are averaged to form the pixel-level flame occurrence probability map corresponding to the target frame image.

[0023] Map each pixel in the flame occurrence probability map to a tunnel grid cell; for each grid cell, if the flame occurrence probability in its region is greater than a preset threshold, mark the grid cell as 1, otherwise mark it as 0;

[0024] By aligning with timestamps, the fire source label at each moment is paired with the sparse temperature observation at the same moment.

[0025] According to a preferred embodiment of the present invention, during the flowing fire simulation experiment, different working conditions are constructed by changing the working conditions of the tunnel foundation and the flowing fire, thereby obtaining a training dataset containing sparse temperature observation sequences and fire source labels under different working conditions.

[0026] Different tunnel foundation conditions are constructed by changing the slope and gradient of the tunnel floor; different flowing fire conditions are constructed by adjusting the peristaltic pump to change the fuel leakage rate and leakage duration.

[0027] According to a preferred embodiment of the present invention, the construction of physical features reflecting the spatial distribution characteristics of fire sources includes: calculating the temperature difference between the first sensor and the second sensor at the same longitudinal position in the tunnel as a lateral temperature difference feature; calculating the temperature difference between the first sensor and the second sensor at adjacent longitudinal positions as a longitudinal temperature difference feature; calculating the temperature ratio between the first sensor and the second sensor at the same longitudinal position as a relative temperature distribution feature; and calculating the temperature change of the same sensor at adjacent time steps as a temperature-time change rate feature.

[0028] According to a preferred embodiment of the present invention, in step 2), the temporal neural network model projects the features of each time step into a unified latent space, then adds position encoding, and then aggregates the time dimension to obtain a global feature vector.

[0029] According to a preferred embodiment of the present invention, in step 2), the tunnel surface is discretized into a regular grid; the spatial morphology reconstruction model inputs the global feature vector into a fully connected mapping layer and projects it into a low-resolution feature map; then, multi-layer transposed convolution is used to upsample step by step to restore the low-resolution feature map to the target spatial resolution, and finally outputs a single-channel fire source probability distribution matrix; each element in the fire source probability distribution matrix represents the probability that a flame exists in the corresponding grid cell.

[0030] According to a preferred embodiment of the present invention, in step 3), the physical constraints are embedded into the model training process by constructing a joint loss function, wherein the joint loss function includes a data consistency loss term and at least one physical constraint loss term:

[0031] The data consistency loss term is used to constrain the difference between the predicted fire source probability distribution matrix and the actual fire source labels;

[0032] The physical constraint loss term includes at least one of the following: spatial continuity loss term, boundary constraint loss term, temporal evolution consistency loss term, propagation direction prior loss term, and area change rate constraint loss term.

[0033] According to a preferred embodiment of the present invention, in step 4), the spatial continuity loss term is used to constrain the fire source regions in the fire source probability distribution matrix to have spatial connectivity, and the total variation regularization is used to penalize the prediction difference between adjacent grid cells.

[0034] The boundary loss term is used to constrain the distribution of fire sources to not exceed the boundary of the tunnel sidewall, and to penalize the predicted values ​​that exceed the effective area mask of the tunnel.

[0035] The temporal evolution consistency loss term is used to constrain the continuity of changes between the fire source probability distribution matrices output by adjacent time steps, and to avoid non-physical morphological abrupt changes.

[0036] The consistency loss term for the direction of fire spread is used to constrain the main direction of fire spread based on the tunnel ventilation direction or slope conditions.

[0037] The area change rate constraint loss term is used to constrain the rate of change of the fire source area within a preset reasonable threshold range per unit time.

[0038] According to a preferred embodiment of the present invention, step 4) further includes a step of visualizing the output planar fire source distribution information: the output fire source probability distribution matrix is ​​displayed in real time in the form of a two-dimensional heat map, wherein the fire source area is marked with a first color and the non-fire source area is marked with a second color, and the tunnel geometric boundary information is superimposed to generate a visualized image or video.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1) Existing technologies are mostly based on the assumption of a single ignition source, which can only predict the location of the ignition source or the heat release rate, making it difficult to describe the flowing fire, a fire morphology with spatial expansion characteristics and dynamic evolution. This invention discretizes the tunnel surface area into a two-dimensional grid space and uses an ignition source probability distribution matrix or a binary ignition source distribution matrix to characterize the actual spatial coverage of the flowing fire. Through this technique, this invention can directly output the two-dimensional planar distribution of the flowing fire on the tunnel surface, expanding the ignition source representation from the traditional point location to a planar structure with spatial boundaries, coverage, and spread patterns. Therefore, this invention can more accurately reflect the irregular spatial morphology formed during the fuel leakage, diffusion, combustion, and spread of flowing fires, facilitating rescue personnel's intuitive judgment of the main fire source area, the fire front position, and the potential spread direction, thus improving the completeness and accuracy of tunnel fire situational awareness.

[0041] 2) Traditional numerical simulation methods suffer from high computational costs and poor real-time performance, making them difficult to meet the needs of engineering applications. This invention pre-constructs training samples using numerical simulation data, experimental data, or a combination of both, to train a joint model comprising a temporal neural network model and a spatial morphology reconstruction model. In practical applications, only the real-time collected sparse temperature observation sequence needs to be input into the trained joint model, and the current fire source probability distribution matrix can be quickly obtained through forward computation. Compared to traditional computational fluid dynamics methods that require repeated solutions to fluid control equations, combustion reaction equations, and heat transfer equations, this invention eliminates the need for costly numerical solutions for every real-time condition after deployment, significantly reducing the online computational burden and improving prediction response speed. Therefore, it is more suitable for engineering scenarios with high real-time requirements, such as early identification, dynamic monitoring, and emergency decision-making in tunnel fires.

[0042] 3) Considering that in actual tunnel environments, temperature observation data typically originates from a limited number of sensors, exhibiting significant sparsity, while the spatial morphology of flowing fire is a high-dimensional structural information, this problem belongs to the complex inverse problem of inferring high-dimensional structures from low-dimensional observations. Existing methods struggle to achieve stable and accurate inversion while ensuring physical plausibility. This invention employs a sparse temperature sensor array deployed longitudinally along the tunnel, with a central axis sensor and a single-sided offset sensor at each longitudinal measuring point to acquire longitudinal temperature change information and lateral temperature difference information. Based on this, this invention further constructs physical characteristics such as lateral temperature difference, longitudinal temperature difference, temperature ratio, and temperature-time change rate, and concatenates these physical characteristics with the original temperature data to form an enhanced time-series input tensor. Through this technique, this invention not only utilizes the absolute magnitude of temperature but also fully leverages the relative changes between different measuring points and different time steps, enabling sparse temperature data to more fully characterize information such as fire source location, fire source offset, spread direction, and development speed. Therefore, this invention can improve the stability and accuracy of real-time reconstruction of flowing fire morphology with fewer sensors, reduce reliance on high-density sensor arrays, and lower the construction, maintenance and replacement costs of the system in the tunnel.

[0043] 4) Existing methods generally lack physical constraint modeling for tunnel boundary conditions and fire evolution processes, resulting in limited applicability of prediction results in real-world scenarios. This invention introduces physical constraint mechanisms during model training, including spatial continuity constraints, boundary constraints, temporal evolution consistency constraints, prior constraints on spread direction, and area change rate constraints, which, together with data consistency loss, constitute a joint loss function. Spatial continuity constraints suppress isolated grids and discrete noise in the prediction results, making the predicted fire source area more consistent with the physical characteristics of continuous flowing fire. Boundary constraints limit the distribution of predicted fire sources from exceeding the effective geometric area of ​​the tunnel, improving the consistency between the prediction results and tunnel boundary conditions. Temporal evolution consistency constraints prevent unreasonable jumps in fire source morphology between adjacent time steps, making the prediction results more consistent with the continuous expansion and gradual evolution of flowing fire. Prior constraints on spread direction and area change rate constraints further utilize ventilation direction, slope conditions, and fuel diffusion laws to limit the main development trend of the fire source. Therefore, this invention improves the physical rationality, temporal continuity, and engineering credibility of the model output results.

[0044] 5) Addressing the problem that traditional temperature field reconstruction methods primarily address mapping similar physical quantities and struggle to directly obtain the spatial structure of fire sources, this invention proposes a real-time cross-physical quantity reconstruction method for reconstructing the two-dimensional spatial morphology of flowing fire from sparse temperature observation data (i.e., the tunnel flowing fire spatiotemporal evolution situation perception method of this invention). Specifically, this invention does not perform continuous interpolation or reconstruction of the tunnel temperature field. Instead, it directly maps the indirect thermal response information collected by temperature sensors, after temporal feature extraction and spatial morphology reconstruction, into a fire source probability distribution matrix of flowing fire on the tunnel surface. This approach overcomes the limitations of traditional methods, which can only obtain low-dimensional information such as temperature distribution, fire source location, or heat release rate, making the monitoring results closer to the actual needs of fire source coverage, fire source boundaries, and spread patterns in fire emergency response. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the application of the method in an embodiment of the present invention.

[0046] Figure 2 This is a comparison diagram of the flow fire space morphology inversion effect for Condition 1 and Condition 2.

[0047] Figure 3 This is a comparison diagram of the flow fire space morphology inversion effect for conditions 3 and 4.

[0048] Figure 4 This is a comparison chart of the HRR inversion effect of dynamically diffused flowing fire. Detailed Implementation

[0049] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0050] The present invention will be further described in detail below with reference to its technical solutions. The embodiments described are merely examples of the present disclosure and do not limit the scope of the invention.

[0051] like Figure 1 The diagram shown is a complete process diagram of an embodiment of the present invention. This embodiment takes a highway tunnel with a length of 100m and a width of 10m as an example for illustration, but the present invention is not limited to this specific size.

[0052] 1. Sensor deployment and data acquisition

[0053] Temperature sensors are deployed along the length of the tunnel (i.e., longitudinally) at the top of the tunnel, with a set of measuring points every 10 m, for a total of 11 sets of measuring points. Each set of measuring points includes one sensor located at the tunnel's central axis and another sensor 1 m from the sidewall, thus forming a dual-measuring-point structure deployed transversely along the tunnel. The sensors can be K-type thermocouples or fiber Bragg grating temperature sensors; in this embodiment, the sampling frequency is 1 Hz. During the fire, temperature data from each measuring point is collected at 1-second sampling intervals, and 20 consecutive seconds of temperature data are selected as the subsequent model input to construct a three-dimensional input data tensor of size (20, 11, 2). This deployment scheme is also applicable to subsequent experimental data acquisition: in the tunnel fire experiment, based on the highway tunnel studied in this embodiment, the experimental platform is scaled down and constructed, and temperature sensors are installed in the same positional relationship to form a sparse temperature sensor array (11 sets of measuring points, 2 temperature sensors per set). The collected temperature data and video image data are aligned using timestamps to form paired training samples.

[0054] 2. Tunnel Spatial Discretization and Fire Source Representation

[0055] The spatial morphology data of flowing fire was primarily obtained through tunnel fire experiments. Specifically, high-speed cameras were deployed on the tunnel ceiling and sidewalls to continuously record the entire process of fuel leakage, diffusion, and ignition of the flowing fire. The cameras were positioned to cover the flame area on the tunnel floor as completely as possible with minimal obstruction, ensuring accurate extraction of the spatial extent and morphological changes of the flame. Simultaneously, temperature sensors were installed on the tunnel ceiling according to the aforementioned deployment plan, synchronously recording temperature data at each measuring point at a sampling frequency of 1 Hz.

[0056] After acquiring video data, the first step is to calibrate the scale between the images and the actual physical space. Specifically, before the experiment begins, calibration objects of known size (such as rulers or equally spaced markers) are placed on the tunnel floor, and images are taken using a camera. Using these calibrated images, the actual physical length corresponding to each pixel in the image can be calculated (e.g., how many meters each pixel corresponds to), thus establishing a mapping relationship between "pixel coordinates and actual spatial coordinates." In actual processing, multiple calibration points with known distances can be selected for fitting to improve calibration accuracy. Through this calibration step, the pixel lengths, areas, and other information measured in subsequent images can be converted into actual physical dimensions.

[0057] Based on this, the experimental video is processed frame by frame. Specifically, the original video is divided into segments at fixed time intervals; in this embodiment, a time step of 1 second is used to convert the video into a continuous image sequence, thereby obtaining the instantaneous spatial distribution of the flowing fire at different times. For each frame, image processing methods (such as segmentation methods based on brightness thresholds or color spaces) are used to identify the flame region, and the original flame image is converted into a binary image, where the pixel value corresponding to the flame region is 1, and the non-flame region is 0. Through this process, the pixel-level spatial distribution of the flame in the image can be obtained.

[0058] To obtain fire source tags corresponding to the dynamic evolution of flowing fire, this embodiment processes the experimental video frame-by-frame or time-by-time according to preset time intervals, rather than only selecting a fixed time period after combustion enters the quasi-steady stage for statistical analysis. Specifically, the experimental video is divided into time steps matching the temperature sampling frequency, for example, obtaining a continuous image sequence with a time step of 1 second; for each target time, the corresponding video frame is extracted, and the flame region is identified using methods based on brightness thresholds, color space segmentation, or image semantic segmentation to obtain the pixel-level flame identification result for that target time. The pixel-level flame identification result mainly serves as an intermediate result in the fire source tag generation process, used for subsequent spatial calibration and mesh transformation.

[0059] Furthermore, to reduce the impact of flame boundary flickering, image noise, and transient disturbances on tag stability, several adjacent video frames within a short-time sliding window can be selected for local statistical analysis, centered on or ending at the target time. The flame recognition results for the same pixel location within this short-time sliding window across multiple frames are averaged to obtain a pixel-level flame occurrence probability map corresponding to the target time. This short-time sliding window is only used to suppress flame boundary jitter and image recognition errors; its window length is less than the time required for a significant change in the overall spread pattern of the flowing fire, thus not altering the dynamic correspondence between the fire source tag and the target time. The resulting pixel-level flame occurrence probability map retains the dynamic characteristics of the flowing fire's spatial morphology at the current time while improving the stability and robustness of the flame boundary region tags.

[0060] After constructing the pixel-level flame occurrence probability map, it is converted into the discrete grid representation of the tunnel surface used in this invention. Specifically, the tunnel surface is first divided into a regular two-dimensional grid (e.g., a 500×50 grid) according to a preset resolution. Then, based on the correspondence between pixels and actual space obtained from the aforementioned calibration, each pixel in the image is mapped to the corresponding grid cell. For each grid cell, if a flame pixel exists in its corresponding region (or the flame probability exceeds a preset threshold, e.g., 0.5), the grid cell is marked as 1; otherwise, it is marked as 0. Through this process, the continuous spatial distribution of flames can be transformed into a discrete two-dimensional matrix representation, serving as the fire source label for model training. Through timestamp alignment, the fire source label at each moment is paired with the sparse temperature observation value at the same moment to form a complete training sample (input: 20 s temperature window; output: fire source distribution matrix at the current moment).

[0061] 3. Data Construction and Preprocessing

[0062] This embodiment prioritizes experimental data acquisition. Regarding experimental setup, the leak point location can be fixed, for example, at a predetermined position on the tunnel floor's central axis. This reduces the interference of leak location changes on the evolution of the fire source morphology, facilitating focused research on the correspondence between the spatial morphology of flowing fire and its temporal response under different foundation conditions, leak conditions, and ventilation conditions. It should be noted that the core objective of this invention is not to establish a dedicated prediction model for a specific slope or gradient, but rather to construct a high-quality training dataset containing various typical operating conditions, enabling the model to learn the mapping relationship between sparse temperature temporal data and the spatial morphology of flowing fire.

[0063] To generate diverse flowing fire spatial morphologies, this embodiment can construct different foundation conditions by adjusting the slope and gradient of the tunnel floor. Specifically, the following typical foundation conditions can be set:

[0064] (1) Horizontal base (0% slope): Simulates a straight tunnel section. The flowing fire mainly spreads by the gravity of the fuel itself and the feedback of combustion heat. The flame shape is approximately circular or elliptical and expands outward.

[0065] (2) Single slope base: The tunnel is set to slope downhill along its length, and the slope can be selected in the range of 1% to 5%. When going downhill, the fuel accelerates forward under the action of gravity, and the flame is "teardrop" shaped and spreads faster.

[0066] (3) Lateral slope base: Set a single-sided inclination (e.g., lower on the left and higher on the right) in the cross section of the tunnel to simulate the lateral inclination of the tunnel caused by construction errors or settlement. At this time, the flowing fire will spread to the lower side and the flame shape will show obvious lateral deviation.

[0067] By adjusting the aforementioned slope and gradient parameters, and combining this with precise control of the fuel leakage rate (typically ranging from 0.5 to 5 L / s) and leakage duration (10 to 60 s) using a peristaltic pump, various flowing fire conditions with different diffusion ranges, spread rates, frontal morphologies, and spatial distributions can be generated. Based on the synchronized video and temperature data obtained under these different conditions, and after the aforementioned image processing and discretization steps, a flowing fire spatial morphology dataset covering rich baseline conditions can be constructed for subsequent model training and validation.

[0068] Different slope shapes and gradients affect the fuel flow path, ignition source coverage, flame front position, and ignition source area change rate, thus further influencing the temporal response of various temperature measurement points during the development of a flowing fire. In other words, although environmental factors such as slope shape and gradient are not explicitly required as separate input parameters, their impact on the morphology of a flowing fire is reflected in information such as the temperature rise rate, longitudinal temperature difference, lateral temperature difference, temperature ratio, and temporal change pattern collected by temperature sensors. Therefore, during model training, the time-series temperature data obtained under different slope shapes, gradients, leakage rates, leakage durations, ventilation conditions, and other operating conditions, along with their corresponding ignition source labels, can be uniformly constructed into a training dataset for training the same joint inversion model.

[0069] Specifically, each training sample includes a sparse temperature observation sequence within a preset time window, and a fire source probability distribution matrix corresponding to the target time. For samples generated under different slope shapes and gradients, although their fire source labels differ, their temperature time-series responses also change. The joint model learns the statistical mapping relationship between different temperature time-series responses and the corresponding flowing fire spatial morphology through training on a large number of samples. Therefore, this invention does not require building independent models for each slope shape or gradient, nor does it require pre-determining the slope shape or gradient of the tunnel during the inference stage; within the slope shape, gradient, and related working conditions covered by the training data, the trained joint model can reconstruct (invert) the flowing fire spatial morphology at the current moment based on the real-time collected sparse temperature time-series data.

[0070] By constructing the dataset as described above, this invention can incorporate the influence of different slope shapes, gradients, and leakage conditions on the evolution of flowing fire into the training sample distribution, enabling the model to learn the correspondence between temperature response and fire source morphology under multiple working conditions within a unified framework, thereby improving the applicability and generalization ability of the model in the complex environment of actual tunnels.

[0071] During data preprocessing, outlier detection and missing value handling are first performed on the raw temperature data collected by the temperature sensors. For outlier data that significantly exceeds the sensor's range or does not conform to the continuity of temperature changes, interpolation using nearby time steps, interpolation using adjacent measurement points, or removal of corresponding samples can be used for processing. For short-term missing data, linear interpolation, forward imputation, or estimation based on adjacent sensor data can be used for completion. After outlier handling and missing value completion, the temperature measurement data of each point are aligned with the fire source label according to a unified timestamp to ensure that the temperature observation sequence in the same training sample has a clear temporal correspondence with the fire source distribution matrix at the target time.

[0072] Furthermore, the time-aligned temperature data is normalized to eliminate the impact of differences in initial ambient temperature, combustion intensity, and sensor response range under different operating conditions on the model training stability. In a preferred embodiment, the original temperature of each temperature measurement point can first be converted into a temperature rise value relative to the initial ambient temperature. That is, the average temperature over a period of time before the fire occurs is used as the reference temperature, and the real-time temperature is subtracted from the reference temperature to obtain the temperature rise sequence. Subsequently, based on the mean and standard deviation obtained from the training dataset, the temperature rise sequence is standardized to convert it into a normalized input with a mean close to 0 and a relatively stable variance. Its form can be expressed as:

[0073]

[0074] in, This represents the temperature or temperature rise value at the t-th time step, the ith longitudinal measuring point, and the c-th transverse measuring point. and ε represents the mean and standard deviation of the temperature or temperature rise at the corresponding measurement point in the training dataset, respectively; ε is a very small positive number to prevent the denominator from being zero. This represents the normalized temperature input. Besides the standardization method described above, maximum-minimum normalization can also be used to scale the temperature or temperature rise data to a preset range.

[0075] After the above preprocessing, each training sample corresponds to a normalized temperature input tensor T. This normalized temperature input tensor serves as the foundational data for subsequent "temperature-driven temporal feature extraction and physical feature construction." In other words, when constructing physical features such as lateral temperature difference, longitudinal temperature difference, and temperature-time rate of change, this invention preferably uses normalized temperature data for calculation. For relative features such as temperature ratios, which are sensitive to the sign and scale of values, calculations can be performed based on temperature data that has been temperature-rated and maintains a non-negative scale. The resulting ratio feature is then normalized according to the statistical parameters of the training set and concatenated with other features. This approach allows for a continuous and consistent data processing flow between the original temperature data preprocessing, physical feature construction, and temporal model input.

[0076] 4. Temperature-driven temporal feature extraction and physical feature construction

[0077] In this embodiment, after the aforementioned data cleaning, time alignment, and normalization processes, a normalized temperature input tensor is obtained:

[0078]

[0079] in, Represents the normalized temperature input tensor; R represents the real number domain; the first dimension of the exponent represents the time step, the second dimension represents the longitudinal measurement point number, and the third dimension represents the transverse measurement point type; it can be agreed that the first channel in the third dimension is the temperature of the central axis measurement point, and the second channel is the temperature of the single-sided measurement point.

[0080] 4.1 Original Temperature Tensor Expansion

[0081] To facilitate subsequent feature construction, for any given time... ,remember:

[0082] The temperature sequence along the central axis is as follows:

[0083]

[0084] The temperature sequence of a single measuring point is

[0085]

[0086] Therefore, the original input of a single sample can also be represented as:

[0087]

[0088] Where X0 represents the basic input matrix obtained by expanding the normalized temperature data; the 22-dimensional vector at each time step is obtained by concatenating 11 central axis temperatures and 11 unilateral temperatures.

[0089] 4.2 Physical Features Construction

[0090] Since this invention relies only on sparse 11×2 measurement point data, in order to enhance the input's ability to characterize the spatial morphology of flowing fire, the following physical characteristics are further constructed based on the original temperature.

[0091] (1) Lateral temperature difference characteristics

[0092] For each longitudinal position i, calculate the temperature difference between the central axis and the measuring point on one side:

[0093]

[0094] in, This represents the lateral temperature difference characteristic at the t-th time step and the i-th longitudinal position; symbol The value represents the difference; the superscript lat indicates the lateral direction. This feature reflects the lateral temperature gradient and is used to characterize the degree of lateral offset of the fire source and the asymmetry of heat distribution.

[0095] This yields the transverse temperature difference vector:

[0096]

[0097] (2) Longitudinal temperature difference characteristics

[0098] Calculate the first-order difference along the tunnel length for the central axis and single-sided measuring points respectively.

[0099]

[0100]

[0101] in, This represents the normalized temperature difference between two adjacent longitudinal measuring points along the central axis at the t-th time step. This represents the normalized temperature difference between two adjacent longitudinal measuring points in the unilateral offset direction at time step t; the superscript lon indicates the longitudinal direction. The longitudinal temperature difference reflects the distribution change of heat along the tunnel length and is used to characterize the longitudinal spread trend of the fire source. This part yields a total of 20 features.

[0102] (3) Temperature ratio characteristics

[0103] To reduce the influence of different absolute temperature levels and enhance the expression of relative relationships, a ratio of the central axis temperature to the temperature on one side is constructed for each longitudinal position:

[0104]

[0105] in, This represents the temperature ratio characteristic at the t-th time step and the ith longitudinal position. To prevent extremely small positive numbers with a denominator of zero, such as 10 -6 This characteristic reflects the relative non-uniformity of temperature distribution within the same cross-section.

[0106] (4) Characteristics of the rate of change over time

[0107] Calculate the rate of temperature change for continuous time steps:

[0108]

[0109]

[0110] in, This represents the normalized temperature change of the i-th central axis measuring point at time step t relative to the previous time step. This represents the normalized temperature change at time step t for the i-th unilaterally biased measuring point relative to the previous time step; V represents the temperature-time rate of change characteristic. This characteristic is used to characterize the local heating rate, thereby reflecting the fire source development speed and spread stage.

[0111] 4.3 Feature splicing and input shape

[0112] In a preferred embodiment, the following feature vector is constructed for each time step t:

[0113] (1) Original temperature: 22 dimensions

[0114] (2) Lateral temperature difference: 11 dimensions

[0115] (3) Longitudinal temperature difference: 20 dimensions

[0116] (4) Temperature ratio: 11 dimensions

[0117] (5) Rate of change over time: 22 dimensions (zero padding is allowed when t=1)

[0118] The feature dimension of a single time step is:

[0119]

[0120] Therefore, the constructed temporal input tensor is:

[0121]

[0122] In another embodiment, the spatial structure can be preserved, and the original temperature and features such as lateral temperature difference and ratio can be stacked according to channels to form:

[0123]

[0124] in This represents the enhanced input tensor that preserves the longitudinal measurement point structure. C is the number of feature channels. For example, when using 6 types of channels, namely the original central axis temperature, the original single-sided temperature, the lateral temperature difference, the lateral absolute temperature difference, the temperature ratio, and the time change rate, C=6.

[0125] 5. Temporal Feature Modeling

[0126] In this embodiment, a self-attention-based temporal modeling network is used to encode the above features.

[0127] When the input is First, the 86-dimensional features at each time step are projected onto a unified latent space through a linear mapping layer, for example:

[0128]

[0129] in, This represents the initial latent feature matrix obtained after linear mapping; d is the embedding dimension, which can be 128 or 256, for example. Then, positional encoding is added and input into a temporal coding network, outputting a temporal feature representation:

[0130]

[0131] Here, Z represents the temporal feature matrix after processing by the temporal coding network, containing dependency information between different time steps. Further aggregation of the time dimension, such as using features from the last time step, average pooling, or weighted pooling, yields the global feature vector:

[0132]

[0133] Here, h represents the global feature vector characterizing the development state of the flowing fire within the current time window. The temporal dimension aggregation method can employ techniques such as latent features from the last time step, temporal average pooling, weighted pooling, or attention pooling. This global feature vector h serves as the input to the subsequent spatial morphology reconstruction model.

[0134] 6. Construction of Spatial Morphology Reconstruction Model

[0135] The prediction target of this invention is not a continuous temperature field, but rather the two-dimensional spatial form of flowing fire on the ground. The tunnel floor is discretized into a regular 500×50 grid, therefore the output target is:

[0136]

[0137] Where Y represents the probability distribution matrix of the actual fire source at the target time; when Y is close to 1, it indicates that there is a high probability that there is a flame in the corresponding grid cell; when Y is close to 0, it indicates that there is a high probability that there is no flame in the corresponding grid cell.

[0138] 6.1 Feature Upgrading

[0139] The global feature vector obtained in step 5 The input is a fully connected mapping layer, which is then projected onto a low-resolution feature map. For example, it can be mapped first as follows:

[0140]

[0141] Where F0 represents the low-resolution feature map obtained by upscaling the global feature vector; 25×5 is the initial spatial size, and c0 is the number of channels, such as 64 or 128.

[0142] 6.2 Transposed Convolution Reconstruction

[0143] Then, multi-layer transposed convolutions are used for step-by-step upsampling to restore the low-resolution feature map to the target spatial resolution. An example process is as follows:

[0144]

[0145] After upsampling at each level, a convolutional layer and a nonlinear activation layer are applied to recover the local spatial structure. The final output is a single-channel prediction matrix.

[0146]

[0147] in, This represents the fire source probability distribution matrix predicted by the model, where each element represents the probability that the corresponding grid belongs to the fire source region.

[0148] 6.3 Binarization Output

[0149] During the training phase, It can be directly used as a probability graph in loss calculation; during the inference phase, a threshold can be used. Binarization:

[0150]

[0151] in, This represents the predicted binary fire source distribution matrix. A value of 0.5 can be used, or the optimal threshold can be determined through a validation set.

[0152] This step does not continuously reconstruct the temperature field, but directly maps low-dimensional temperature features to a fire source spatial structure with clear boundaries. The output object is a two-dimensional mask map of flowing fire rather than a temperature value field, thus it belongs to real-time reconstruction across physical quantities.

[0153] 7. Construction of Physical Constraint Embedding and Inversion Mechanism

[0154] To ensure that the predicted flowing fire spatial morphology not only closely matches the sample labels but also conforms to the actual fire development patterns, this embodiment constructs a joint loss function during the training process, incorporating both data consistency terms and physical constraint terms into the optimization objective.

[0155] Let the total loss function be:

[0156]

[0157] Where L represents the joint loss function that needs to be minimized during model training; This represents the data consistency loss term, used to constrain the difference between the predicted fire source probability distribution matrix and the actual fire source probability distribution matrix; This represents the spatial continuity loss term, used to suppress isolated grids and non-physical discrete noise in the prediction results; This represents the boundary constraint loss term, used to ensure that the predicted fire source does not exceed the effective area of ​​the tunnel. , , These are the weighting coefficients for each loss term.

[0158] 6.1 Data Inconsistency Loss

[0159] The data consistency term is used to constrain the predicted fire source distribution to be consistent with the actual labels. It can be a combination of mean squared error loss, intersection-over-union loss, binary cross-entropy loss, and Dice loss.

[0160]

[0161] in, Indicates the mean square error loss; Indicates the crossover and union ratio loss; This represents the binary cross-entropy loss; This indicates Dice's loss.

[0162] The root mean square error loss can be expressed as:

[0163]

[0164] in, This represents the probability value predicted by the model that a flame exists in the i-th length direction grid and the j-th width direction grid; This indicates the actual fire source label for the corresponding grid cell.

[0165] The crossover ratio loss can be expressed as:

[0166]

[0167] in, This represents a very small positive number set to prevent the denominator from being zero.

[0168] The binary cross-entropy loss can be expressed as:

[0169]

[0170] Where log represents the natural logarithm function.

[0171] Dice loss can be expressed as:

[0172]

[0173] By using the aforementioned data consistency loss term, the predicted fire source probability distribution matrix can be made to closely approximate the actual fire source labels in terms of overall location, area range, and boundary morphology.

[0174] 6.2 Spatial continuity loss term

[0175] In practice, flowing fire regions are usually connected, so a neighborhood smoothing term is constructed to suppress isolated pixels:

[0176]

[0177] This term is essentially in the form of Total Variation, which can reduce non-physical discrete noise in the prediction results.

[0178] 6.3 Boundary Constraint Loss Term

[0179] Let the effective area mask of the tunnel be... Within the effective area Invalid region Then, any source of fire that crosses the boundary will be punished:

[0180]

[0181] This ensures that the predicted fire source does not exceed the tunnel's geometric boundaries.

[0182] 7. Model Training and Prediction

[0183] The model is trained using the dataset constructed earlier, combined with the constructed joint loss function. By iteratively optimizing the model parameters, the predicted results gradually approximate the actual fire source distribution.

[0184] In practical applications, real-time temperature data is input into the trained model to output the spatial distribution of flowing fire at the current moment. To verify the feasibility of the proposed method and the accuracy of the prediction results, the model was trained and tested based on numerical simulation data, and the model output results were compared and analyzed. The specific implementation results are shown below:

[0185] (1) Verification of the spatial morphology inversion effect of flowing fire:

[0186] Four typical verification conditions were selected: 1) Lateral slope = 2%, longitudinal slope = 0%, leakage rate = 1 L / s; 2) Lateral slope = -2%, longitudinal slope = 2%, leakage rate = 1 L / s; 3) Lateral slope = 2%, longitudinal slope = 0%, leakage rate = 1.4 L / s; 4) Lateral slope = 2%, longitudinal slope = -2%, leakage rate = 0.8 L / s. The model prediction results were compared and analyzed with the actual ignition source distribution, as shown in the attached figure. Figure 2 and Figure 3 As shown in the figure, the actual fire source distribution (GT), the model prediction probability distribution (PredProb), the binarized prediction result (PredMask), and the difference distribution between the actual and predicted fire sources (|GT−Prob|) are presented respectively. The comparison results show that:

[0187] 1) The method of the present invention can successfully realize the inversion from temperature data to the fire source distribution matrix.

[0188] The model outputs a two-dimensional matrix of size (500, 50), which is consistent with the preset grid space, verifying that the method of the present invention can stably output the spatial distribution results of fire sources and has good feasibility.

[0189] 2) The predicted fire source area is highly consistent with the actual fire source area in spatial location.

[0190] A comparison between GT and PredMask shows that the predicted fire source areas are basically consistent with the actual fire source distribution in terms of location, shape and range, especially in the fire source front and main area, where there is a high degree of overlap.

[0191] 3) The predicted probability distribution can accurately characterize the boundary features of the fire source.

[0192] The PredProb results show that the predicted probability is close to 1 inside the fire source area, close to 0 in the non-fire area, and exhibits a continuous transitional distribution at the fire source boundary. This indicates that the model can not only identify the fire source area, but also reasonably express the uncertainty of the boundary.

[0193] 4) The error is mainly concentrated in the boundary area of ​​the fire source, and the overall error is relatively small.

[0194] The |GT−Prob| distribution shows that the error is mainly concentrated at the fire source boundary, while the prediction error is smaller inside the fire source and in areas far from the fire source. This indicates that the method of the present invention has high accuracy in overall structure identification.

[0195] In summary, the results show that the method of the present invention can effectively invert the spatial morphology of flowing fire under sparse temperature observation conditions, verifying the feasibility of the present invention.

[0196] (2) Prediction and verification of dynamic flowing fire evolution process:

[0197] Furthermore, to verify the predictive capability of this invention for the dynamic evolution of flowing fire, the predicted fire source distribution matrix is ​​transformed into fire source area, and based on the relationship between area and heat release rate (HRR), it is mapped to an equivalent heat release rate versus time curve, as shown in the attached figure. Figure 4 As shown in the figure, the comparison results of the actual HRR curves and the model-predicted HRR curves under different operating conditions are presented. The results show that:

[0198] 1) The method of the present invention can accurately capture the development trend of flowing fire.

[0199] The predicted curve and the actual curve are highly consistent in overall trend, both reflecting the phased changes in the fire development process, including the initial slow development phase, the middle rapid growth phase, and the later stable or fluctuating phase.

[0200] 2) It can effectively reflect the process of fire source area growth.

[0201] Since HRR is directly correlated with the fire source area, the predicted HRR curve shows a trend consistent with the growth trend of the fire source area. The results indicate that the method of this invention can accurately capture the process of flowing fire expanding from localized combustion to a large-scale spread, demonstrating the model's excellent ability to characterize the dynamic evolution of spatial morphology.

[0202] 3) The predicted results have a high degree of consistency with the actual values ​​in terms of numerical values.

[0203] As can be seen from the curve comparison, the predicted value basically coincides with the actual value for most of the time period, and the error remains within a small range. Only in the local fluctuation stage is there a slight deviation, indicating that the method of the present invention has high prediction accuracy.

[0204] 4) It has a good response capability to sudden changes and acceleration phases of fire.

[0205] During the rapid development phase of a fire, the predicted curve can follow the changes in the actual curve in a timely manner, indicating that the method of the present invention can effectively capture the fire development information reflected in the temperature time series changes.

[0206] In summary, the above results show that the method of the present invention can stably output the fire source distribution matrix based on finite temperature sensor data, verifying the feasibility of the method; the fire source distribution predicted by the present invention has a high degree of consistency with the actual distribution in space, indicating that the present invention has good prediction accuracy; in the dynamic evolution process, the present invention can accurately reflect the changes and development trends of flowing fire area, and has strong engineering application value.

[0207] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A tunnel running fire spatiotemporal evolution situation awareness method based on a sparse temperature sensor array, characterized in that, Includes the following steps: 1) Construct a training dataset containing sparse temperature observation sequences and fire source labels; wherein, the sparse temperature observation sequences are obtained by multiple temperature sensor groups deployed along the longitudinal direction of the tunnel, each temperature sensor group includes two sensors, used to acquire transverse temperature data at the same longitudinal position of the tunnel, and the fire source labels are two-dimensional matrices used to indicate whether there is a flame in each grid cell in the tunnel; the temperature sensor groups are deployed at a preset interval, and each temperature sensor group consists of a first sensor located on the central axis of the tunnel and a second sensor located at a preset position between the central axis of the tunnel and a single side wall; 2) Construct a joint model that includes a temporal neural network model and a spatial morphology reconstruction model; The joint model constructs physical features reflecting the spatial distribution characteristics of fire sources based on sparse temperature observation sequences. The original temperature data and the constructed physical features are concatenated to form an enhanced temporal input tensor, which is then input into a temporal neural network model to obtain a global feature vector. The global feature vector is then input into a spatial morphology reconstruction model to output a fire source probability distribution matrix corresponding to the spatial size of the tunnel grid. The physical features that reflect the spatial distribution characteristics of fire sources include: Calculate the temperature difference between the first and second sensors at the same longitudinal position in the tunnel, and use it as the lateral temperature difference feature; Calculate the temperature difference between the first sensor and the second sensor at adjacent longitudinal positions, and use it as the longitudinal temperature difference feature; Calculate the temperature ratio of the first sensor and the second sensor at the same longitudinal position as a relative temperature distribution characteristic; Calculate the temperature change of the same sensor at adjacent time steps, and use it as a feature of the temperature-time change rate. The temporal neural network model projects the features of each time step into a unified latent space, then adds position encoding, and then aggregates the time dimension to obtain a global feature vector. In step 2), the tunnel surface is discretized into a regular grid; the spatial morphology reconstruction model inputs the global feature vector into a fully connected mapping layer and projects it into a low-resolution feature map; then, it uses multi-layer transposed convolution to upsample step by step to restore the low-resolution feature map to the target spatial resolution, and finally outputs a single-channel fire source probability distribution matrix; each element in the fire source probability distribution matrix represents the probability of flames existing in the corresponding grid cell; 3) The joint model is trained using the training dataset. During the training process, physical constraints are applied to the fire source probability distribution matrix to suppress non-physical prediction results. 4) When flowing fire occurs, temperature data inside the tunnel is collected in real time, and the real-time fire source probability distribution matrix is ​​obtained using the trained joint model, so as to realize the real-time reconstruction of the two-dimensional spatial morphology of flowing fire in the tunnel across physical quantities.

2. The method according to claim 1, characterized in that, In step 1), the data in the training dataset was obtained through a flowing fire simulation experiment. During the experiment, temperature data inside the tunnel was acquired in real time through a temperature sensor, and video of the flowing fire development process was acquired through a camera. For each frame of video captured by the camera, the flame region is identified by methods based on brightness threshold, color space segmentation or image semantic segmentation, and the original flame image is converted into a binary image. Then, the flame recognition results of the target frame and several adjacent video frames before and after it are averaged to form the pixel-level flame occurrence probability map corresponding to the target frame image. Map each pixel in the flame occurrence probability map to a tunnel grid cell; For each grid cell, if the probability of flames appearing in its area is greater than a preset threshold, the grid cell is marked as 1; otherwise, it is marked as 0. By aligning with timestamps, the fire source label at each moment is paired with the sparse temperature observation at the same moment.

3. The method according to claim 2, characterized in that, During the flowing fire simulation experiment, different working conditions were constructed by changing the working conditions of the tunnel foundation and the flowing fire, thereby obtaining training datasets containing sparse temperature observation sequences and fire source labels under different working conditions. Different tunnel foundation conditions are constructed by changing the slope and gradient of the tunnel floor; different flowing fire conditions are constructed by adjusting the peristaltic pump to change the fuel leakage rate and leakage duration.

4. The method according to claim 1, characterized in that, In step 3), the physical constraints are embedded into the model training process by constructing a joint loss function, which includes a data consistency loss term and at least one physical constraint loss term. The data consistency loss term is used to constrain the difference between the predicted fire source probability distribution matrix and the actual fire source labels; The physical constraint loss term includes at least one of the following: spatial continuity loss term, boundary constraint loss term, temporal evolution consistency loss term, propagation direction prior loss term, and area change rate constraint loss term.

5. The method according to claim 4, characterized in that, In step 4), the spatial continuity loss term is used to constrain the fire source regions in the fire source probability distribution matrix to have spatial connectivity, and the total variation regularization is used to penalize the prediction difference between adjacent grid cells. The boundary loss term is used to constrain the distribution of fire sources to not exceed the boundary of the tunnel sidewall, and to penalize the predicted values ​​that exceed the effective area mask of the tunnel. The temporal evolution consistency loss term is used to constrain the continuity of changes between the fire source probability distribution matrices output by adjacent time steps, and to avoid non-physical morphological abrupt changes. The consistency loss term for the direction of fire spread is used to constrain the main direction of fire spread based on the tunnel ventilation direction or slope conditions. The area change rate constraint loss term is used to constrain the rate of change of the fire source area within a preset reasonable threshold range per unit time.

6. The method according to claim 1, characterized in that, Step 4) also includes the step of visualizing the output planar fire source distribution information: The output fire source probability distribution matrix is ​​displayed in real time as a two-dimensional heat map, where fire source areas are marked with the first color and non-fire source areas are marked with the second color, and tunnel geometric boundary information is superimposed to generate a visual image or video.

Images