Tunnel fire spreading identification and temperature field prediction method and system

Abstract

The invention relates to a tunnel fire spread identification and temperature field prediction method and system, and the method comprises the steps: integrating various working condition data obtained in parallel for simulation and test, carrying out the preprocessing of the integrated data to construct a complete data set, and dividing the data set into a training set for model training and a test set for model test; inputting the training set into a time domain convolutional network model, and executing model training and parameter tuning to construct a TCN-based fire spread recognition model and a temperature field prediction model; for the fire spreading recognition model, the fire spreading direction and the number of fire sources are output; for the temperature field and HRR prediction model, the temperature and the HRR are output; verifying the trained TCN model by using the test set and a verification set in the test data to obtain an optimal prediction model; the problems that a current algorithm applied to the tunnel fire does not consider the fire spreading characteristic, advanced temperature data cannot be obtained, and the development situation of the fire in the tunnel cannot be accurately and visually captured are solved.

Description

Technical Field

[0001] This invention belongs to the field of tunnel fire identification and prediction technology, and relates to a method and system for identifying tunnel fire spread and predicting temperature field, particularly a method and system for identifying tunnel fire spread and predicting temperature field based on TCN convolutional network. Background Technology

[0002] Tunnel fires possess complex dynamic characteristics and pose significant safety hazards. In scenarios where vehicles are congested, fires are prone to spread, leading to increased temperatures, reduced visibility, and higher concentrations of toxic substances within confined spaces. This poses a greater risk than a single-vehicle fire, seriously threatening the lives and property of people inside the tunnel.

[0003] Because tunnel fires are time-varying in their development, predicting fire progression and temperature fields is crucial for providing advance warnings to fire and rescue operations. However, traditional empirical models typically focus on steady-state single-vehicle fires, paying less attention to fire spread, resulting in inaccurate descriptions of fire spread and temperature field distribution. Furthermore, heat release rate (HRR), a critical parameter for fire risk assessment, cannot be directly obtained from real-world fire scenarios using current technology, rendering many theoretical models of temperature decay lacking universality and timeliness.

[0004] With the development of computer technology, deep learning algorithms have been introduced into the field of tunnel fire safety. Data-driven methods have made it possible to predict the complex and ever-changing temperature field of fire spread in real time and to obtain the HRR (Heat Rate) in real time. However, current algorithms applied to tunnel fires do not consider the characteristics of fire spread and have failed to form methods for fire spread identification and temperature field prediction. This makes it difficult to accurately and intuitively capture the development trend of fires within tunnels and to provide more precise and timely decision support for fire rescue operations. Therefore, it is urgent to establish a fire spread identification and temperature prediction model by combining deep learning algorithms, obtain advanced HRR data, and provide support for temperature field prediction. Summary of the Invention

[0005] In view of this, in order to solve the problems that current algorithms applied to tunnel fires do not consider the characteristics of fire spread, cannot obtain advanced temperature data, and cannot accurately and intuitively capture the development trend of fires in tunnels, thus failing to provide more accurate and timely decision support for fire rescue work, this invention provides a method and system for identifying the spread of fire and predicting the temperature field in tunnels.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for identifying the spread of fire in tunnels and predicting the temperature field includes the following steps:

[0008] S1. Multi-source data acquisition: Simulation time series data generated by the fire dynamics simulator and test time series data collected by the scaled model fire test are acquired in parallel. The simulation time series data includes temperature, wind speed, heat release rate (HRR) and sensor location information. A database is established based on the real-time HRR, temperature and wind speed data obtained by the simulator and the temperature field data obtained by the scaled model. This database contains different fire spread conditions, temperature and HRR.

[0009] S2. Data Fusion and Preprocessing: Integrate the various working condition data from the simulation and experiment in step S1, preprocess the integrated data to construct a complete dataset, and divide the dataset into a training set for model training and a test set for model testing.

[0010] S3, TCN Model Joint Training: The training set divided in step S2 is input into the Temporal Convolutional Network (TCN) model to perform model training and parameter tuning in order to build a fire spread identification model and a temperature field prediction model based on TCN. This process aims to simultaneously obtain the ability to predict HRR and future temperature field, as well as the ability to identify fire spread characteristics (such as the number of fire sources and the direction of spread).

[0011] S4. Model Validation and Application: The trained TCN model is validated using the test set and validation set in step S2 to obtain the optimal prediction model; real-time tunnel environment data is input into the optimal prediction model, and the current fire spread identification results, as well as future temperature field and HRR prediction data are output in parallel.

[0012] Furthermore, the temperature data during the fire development process in step S1 has obvious temporal correlation and belongs to typical time series data. Therefore, the TCN algorithm for processing time series data is used to train the temperature prediction model, the core idea of ​​which is causal convolution.

[0013] Furthermore, step S2 specifically includes:

[0014] S21. Data Cleaning: The simulation time series data is generated by FDS software and does not require data cleaning. In contrast, for scaled-down model test data with missing or abnormal temperature data, linear interpolation is used to fill in the missing values, and... Criteria for outlier detection and removal; smoothing of repeated temperature values ​​using moving averages; fire data is transformed into a format suitable for training and testing TCN convolutional network models, improving model accuracy and generalization performance.

[0015] S22, Data labeling; The direction and quantity of fire spread are labeled by encoding; No spread, horizontal spread, and vertical spread are represented by labels 0, 1, and 2 respectively; The number of fire sources is represented by labels 1 and 2. The label of the number of fire sources is changed to 2 when the spread occurs, so as to form a fire spread identification dataset.

[0016] S23. Normalization processing: In order to avoid slowing down the learning and convergence speed of the TCN neural network, the min-max normalization method shown in Equation (4) is adopted to transform the data into a unified interval, such as [0, 1], so as to uniformly process multiple types of data and improve the prediction accuracy of the model.

[0017] (4)

[0018] In the formula: x i Initialize the data of type i; x i * represents the normalized value; max x i The maximum value of data of type i; minx i The minimum value of data of type i;

[0019] S24. Dataset Partitioning: Divide the dataset into training and testing sets in a ratio of 50-80%:20-50%. For each operating condition, record temperature data using 150-200 thermocouples. Store the data in a CSV file as input to the TCN. Collect data three times per second to form the temperature dataset.

[0020] Furthermore, step S3 specifically includes:

[0021] S31, TCN Model Structure Design: Fire spread pattern recognition and temperature field prediction require training two models: a regression prediction model and a classification model. For the regression task, time, temperature, wind speed, and sensor coordinates (x and y) are used as input variables X1, and HRR and temperature at the future time step are used as output variables Y1. For the classification task, time, temperature, wind speed, sensor coordinates (x and y), and HRR are used as input features X2, and the number of fire sources is used as output feature Y. 2，1 The direction of spread is used as the output feature Y 2，2The number of fire sources is a binary classification task (0 and 1), which requires a binary classification output neuron and obtains the probability through the Sigmoid activation function. For the prediction of the direction of fire spread, since there are three cases (0, 1, 2), three output neurons are needed to form a multi-class problem, and the probability distribution of each possibility is obtained by using the Softmax activation function. Therefore, the input layer of the regression model needs 5 input neurons and the output layer needs 2 output neurons. The input layer of the classification model needs 6 neurons and the output layer needs 4 neurons. The Sigmoid and Softmax activation functions are shown in Equations (6) and (7).

[0022] (6)

[0023] (7)

[0024] In the formula: z is the input value; i is the output value of the i-th neuron in the previous layer of the neural network; S i This represents the output probability of the i-th neuron in the previous layer of the neural network.

[0025] S32, Model Training and Evaluation Metrics; The MSE loss function shown in Equation (8) is used to evaluate the training effect of the regression TCN neural model, and the cross-entropy loss loss shown in Equation (9) is used to optimize the training process of the model:

[0026] (8)

[0027] (9).

[0028] Furthermore, the prediction accuracy of the prediction model in step S4 is calculated as shown in equation (10):

[0029] (10)

[0030] In the formula: To improve prediction accuracy; Let i be the initial value of the i-th data. is the predicted value obtained by the model, and n is the number of data points.

[0031] A prediction system based on tunnel fire spread identification and temperature field prediction methods includes:

[0032] The multi-source data acquisition module includes a simulation database module and an experimental database module. The simulation database module is used to store time-series data under various fire conditions generated by the fire dynamics simulator FDS. The simulation time-series data includes temperature, wind speed, HRR, and sensor location information. The experimental database module is used to store experimental time-series data under various fire conditions obtained through scaled-down model fire tests.

[0033] The data preprocessing module has its input connected to the output of the simulation database module and the experimental database module, respectively. It is used to fuse and preprocess the simulation data and experimental data, and to divide the complete dataset into training set and test set according to a preset ratio.

[0034] The TCN identification and prediction model module has its input end connected to the output end of the aforementioned data preprocessing module. It is used to receive the aforementioned training set for model training and parameter tuning, and to use the validation set for model validation to generate the optimal prediction model. Real-time tunnel environment data is input into the optimal prediction model, and the current fire spread identification results, as well as future temperature field and HRR prediction data, are output in parallel.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. The tunnel fire spread identification and temperature field prediction method disclosed in this invention proposes a fire spread identification, temperature, and HRR prediction model based on a TCN (Temporal Convolutional Neural Network). Temperature and HRR data obtained through numerical simulation are preprocessed and divided into training and testing sets, which are then used for learning and validating the TCN model. The trained TCN model demonstrates high accuracy and good generalization ability in temperature and HRR prediction, and also exhibits high accuracy in fire spread identification, confirming its effectiveness in fire spread detection. This provides a basis for firefighting strategy formulation and improves the intelligence of emergency rescue, aiming to maximize the safety of people's lives and property.

[0037] 2. In the tunnel fire spread identification and temperature field prediction method disclosed in this invention, the temperature and HRR prediction models based on TCN showed good generalization in the test set, with the prediction errors for both temperature and HRR controlled within 19%. Given that the HRR prediction is based on 183 temperature data points as input features, its prediction error is even lower, consistently below 5%.

[0038] 3. In the tunnel fire spread identification and temperature field prediction method disclosed in this invention, the error in temperature prediction gradually increases with the increase of the lead time. When the lead time is set to 10 seconds, the prediction accuracy is the highest, at 99.77% and 91.37% under both natural ventilation and critical wind speed conditions, respectively. However, when the prediction time is extended to 40 seconds, the accuracy is significantly reduced compared to the short-term 10-second prediction, with the prediction accuracy under natural ventilation and critical wind speed conditions decreasing to 94.90% and 86.53%, respectively. Nevertheless, the temperature prediction model can predict the temperature field 40 seconds in advance, which is more valuable for practical applications, and this level is still within an acceptable range.

[0039] 4. In the tunnel fire spread identification and temperature field prediction method disclosed in this invention, the TCN fire spread identification model achieves an accuracy rate exceeding 94% when identifying fire spread in different directions under different ventilation conditions. The critical wind speed has a significant impact on the identification of longitudinal fire spread, as it may cause chaotic temperature distribution and shift the highest temperature area, resulting in a 3.5% decrease in identification accuracy compared to natural ventilation conditions. In contrast, the critical wind speed has a smaller impact on lateral spread identification, with identification errors under different wind speed conditions controlled within 1%, and the accuracy rate maintained above 95.5%, indicating strong robustness of the model in identifying lateral spread.

[0040] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0041] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0042] Figure 1 This is a flowchart of the tunnel fire spread identification and temperature field prediction method of the present invention;

[0043] Figure 2 This is a schematic diagram of the fire spread identification and temperature / HRR prediction model based on TCN of the present invention, wherein... Figure 2 (a) is a flowchart of the regression model. Figure 2 (b) is a flowchart of the classification model;

[0044] Figure 3 This is a schematic diagram of the TCN convolutional network structure of the present invention;

[0045] Figure 4This is a schematic diagram illustrating the principle of causal convolution in this invention.

[0046] Figure 5 This is a schematic diagram illustrating the principle of dilated convolution in this invention.

[0047] Figure 6 This is a diagram showing the iterative loss of the regression prediction model in this invention.

[0048] Figure 7 This is a comparison chart of the HRR and temperature prediction performance of the test set in this invention. Figure 7 (a) is a comparison chart of the actual HRR and the predicted value. Figure 7 (b) is a comparison chart of actual temperature and predicted value;

[0049] Figure 8 This is a comparison diagram of the actual temperature field and the predicted temperature field of the present invention, wherein... Figure 8 (a) is a contour map of the actual temperature field. Figure 8 (b) is a cloud map of the temperature field predicted 40 seconds in advance. Detailed Implementation

[0050] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0051] Tunnel fires possess complex dynamic characteristics and pose significant safety hazards, making research on fire spread identification and key parameter prediction crucial. Traditional fire monitoring methods struggle to accurately and in real-time perceive fire conditions and predict key parameters. In recent years, with the rapid development of deep learning technology, more and more researchers have begun to focus on the application of deep learning algorithms in fire scenarios. Temporal Convolutional Networks (TCNs), as a novel deep learning model, possess powerful time series prediction capabilities.

[0052] Simply reconstructing the real-time temperature field has relatively limited value in guiding rescue efforts. Predicting the future temperature field based on the current development of a fire, and constructing a predicted temperature cloud map accordingly, would provide invaluable reference information for rescue decision-making. Temperature cloud maps are typically generated based on a certain number of known temperature data points, calculated and plotted using an interpolation model. Similarly, achieving advanced temperature field prediction requires relying on temperature measurement data at a future point in time, combined with a real-time temperature cloud map generation model. Given the significant temporal correlation of temperature data during fire development, classifying it as typical time-series data, the TCN algorithm for processing time-series data is used to train the temperature prediction model, obtaining advanced temperature data and thus providing solid data support for advanced temperature field prediction.

[0053] like Figure 1 The method for constructing the tunnel fire spread identification and temperature field prediction model, as shown, includes the following steps:

[0054] S1. Multi-source data acquisition: Simulation time series data generated by the fire dynamics simulator and test time series data collected from the scaled-down model fire test are acquired in parallel. The simulation time series data includes temperature, wind speed, heat release rate (HRR), and sensor location information. A database is established based on the real-time HRR, temperature, wind speed, and other data obtained from the simulator and the temperature field data obtained from the scaled-down model. This database contains different fire spread conditions, temperatures, and HRRs. The temperature field data is obtained by measuring the temperature sensor array.

[0055] Specifically, temperature data during the fire's development exhibits significant temporal correlation, classifying it as typical time-series data. Therefore, the TCN algorithm for processing time-series data is employed to train the temperature prediction model. The core idea of ​​the TCN algorithm is to utilize causal convolution to process time-series data. The principle of causal convolution strictly adheres to causal relationships, meaning that the result is determined solely by its cause and is unaffected by any future events. During the TCN operation, for any given time t at any layer, its output depends only on the values ​​at time t and prior to it in the next layer. This process can be precisely expressed by formula (1). This design implies that when processing time-series data, the model does not "predict the future" but makes decisions solely based on current and past information, thus avoiding the possibility of information leakage.

[0056] (1)

[0057] In the formula: For input variables, This represents the filter size, * indicates the convolution operation for extracting feature information, and p represents the convolution kernel size. Let be the convolution kernel for the i-th element, and t be the time series. This indicates that convolution operations are performed only on past data. The processing procedure for causal convolution is as follows: Figure 4 As shown.

[0058] Dilated convolution is a special type of convolution operation that introduces dilation between ordinary convolution kernels. By setting the dilation rate, the receptive field of the convolution kernel—that is, the temporal range that a single convolution can cover—can be increased without adding additional parameters. Specifically, in a dilated convolution layer, several zeros are inserted between the kernel's normal positions, thus effectively performing sparse sampling on the input sequence during convolution. Figure 5 As shown, as the number of network layers increases, the expansion factor of each layer can increase, thereby forming a larger effective receptive field, which helps to capture long-distance temporal dependencies.

[0059] The dilation convolution operation can be expressed as equation (2):

[0060] (2)

[0061] in, For the first The coefficient of thermal expansion of the layer, The size of the one-dimensional convolution kernel. This refers to the previous timing information. When When =1, dilated convolution degenerates into ordinary convolution; the sampling interval of the convolution kernel on the input sequence will also increase accordingly, thereby expanding the receptive field and the corresponding output mapping to the input range is wider.

[0062] Residual connections in TCN are primarily used to address the vanishing or exploding gradient problems that may occur during the training of deep networks. In traditional deep neural networks dealing with very deep hierarchical structures, as network depth increases, gradients may gradually decrease during backpropagation, making it difficult to update parameters in earlier layers, or causing training instability due to excessively large gradients. Residual connections introduce direct paths, allowing information to be directly passed from one layer to the next without undergoing multiple nonlinear transformations, thereby improving information flow and gradient propagation. The formula for the residual connection module is as follows:

[0063] (3)

[0064] In the formula: x is the input data of the network module; F(x) is the output result obtained by a series of transformation operations (such as convolution, pooling, batch normalization, etc.) performed on the input x; Activation is the activation function, which is used to transform the result after linear transformation into a nonlinear representation, thereby increasing the complexity of the model; O is the final output of the residual module, which is obtained by adding the original input x to the transformed result F(x) and then passing it through the activation function.

[0065] like Figure 3 As shown, the difference module consists of two consecutive sets of dilated causal convolutional layers, interspersed with weight normalization layers, ReLU activation functions, and Dropout layers. The input features first pass through the first set of dilated causal convolutional layers, then undergo normalization, ReLU activation, and Dropout operations, before entering the second set of the same structure, finally yielding the output features.

[0066] In terms of model testing, the HRR of multi-fire-source tunnel fires in the experiment was calibrated based on the weightlessness method and the calculated ignition range and ignition time to obtain the HRR of multi-fire-source combustion. For numerical simulation, to ensure the consistency between the fire development process and results in the numerical simulation and the experiment, some characteristic points were selected from the HRR change curves during the combustion process, and these points and their corresponding time data were input into the FDS software for numerical simulation comparison. Through standardization, all data were kept within the [0,1] interval. Then, using the combustion curves obtained from the scaled-down model test, the FDS acquisition frequency was set to 3 times / second, obtaining 1050 rows of data including temperature, wind speed, and HRR for each working condition to expand the dataset. This dataset was used as input to establish multiple working conditions for calculating and analyzing the characteristics of HRR curves under different spacing and longitudinal wind speeds.

[0067] S2. Data Fusion and Preprocessing: Integrate the various working condition data from the simulation and experiment in step S1, preprocess the integrated data to construct a complete dataset, and divide the dataset into a training set for model training and a test set for model testing.

[0068] Specifically, S21, data cleaning: The simulation time series data is generated by FDS software, ensuring its data integrity and eliminating the need for data cleaning. In contrast, the scaled-down model test data is affected by factors such as thermocouple damage and unstable data transmission, resulting in missing and abnormal temperature data. To address this, linear interpolation was used to fill in the missing values, and [the following was applied]... Outliers are detected and removed using criteria; in addition, repeated temperature values ​​are smoothed using a moving average method. Through this series of preprocessing steps, the tunnel fire data is transformed into a format suitable for training and testing the TCN convolutional network model, thereby improving the model's accuracy and generalization performance.

[0069] S22. Data labeling; Since TCN is a supervised learning model, it needs to be provided with labeled training samples so that the model can learn how to predict the target variable based on the input features.

[0070] Fire spread identification involves the location of the fire source, the direction of fire spread, the number of fire sources, and temperature data from various sensors. The number of fire sources is a dynamic process, changing with the initial development of the fire. Therefore, the number of fire sources and the direction of fire spread are difficult to represent in the data labels. They are therefore assigned as labels for the direction and number of fire spreads through independent encoding. The fire spread status includes no spread, lateral spread, and longitudinal spread, represented by labels 0, 1, and 2, respectively. The number of fire sources is represented by labels 1 and 2, with the label for the number of fire sources changed to 2 at the moment of spread. In summary, the fire spread identification dataset should be processed into the format shown in Table 1.

[0071] Table 1 Fire Spread Identification Dataset

[0072]

[0073] Temperature and HRR predictions are regression predictions, and their label formats differ. For temperature data, the input feature X is historical temperature data, the size of which depends on the sliding window size, and the output feature Y is the predicted temperature value at a future point in time or over a period of time based on this historical data. For HRR, the input feature X is the temperature values ​​from all temperature sensors within the fire area, and the output feature Y is the HRR. Its dataset format is similar to the fire spread identification dataset label, except that the spread direction and the number of fire sources are removed, so it will not be listed separately.

[0074] S23. Normalization: When processing input variables, differences in dimensions and orders of magnitude often adversely affect model predictions, which may slow down the learning and convergence speed of the TCN neural network. To avoid this problem, normalization is usually used to transform the data into a uniform interval, such as [0, 1]. This transformation eliminates the differences in dimensions and units between different features, enhancing the comparability between features. Given the volatility and complexity of fire monitoring data, in order to maintain data consistency and ensure the comparability of the data input to the TCN neural network, the min-max normalization method shown in Equation (4) is adopted. This method can uniformly process multiple types of data and improve the prediction accuracy of the model.

[0075] (4)

[0076] In the formula: x i Initialize the data of type i; x i * represents the normalized value; max x iThe maximum value of data of type i; minx i It represents the minimum value of data of type i.

[0077] S24. Dataset partitioning: The dataset is divided into training and test sets according to a certain ratio. This method is also called the retention method. Usually, the data is randomly divided into training and test sets according to the ratio of 8:2, 7:3, 6:4, or 5:5. The training set is used to generate the model, and the test set is used to test the accuracy and error of the model to verify the effectiveness of the model. In this embodiment, the dataset is divided into 80% training set and 20% test set. Temperature data for each working condition is recorded by 183 thermocouples and stored in CSV format as input information for TCN. Data is collected 3 times per second, that is, 3*183 temperature data are collected per second, with an actual duration of 350s, which constitutes the temperature dataset as shown in Equation (5). In the formula, T represents the temperature data, the first subscript represents the sampling time, and the second subscript represents the thermocouple code.

[0078] (5)

[0079] Using the time-series temperature and wind speed data from Table 1 as input X to the TCN convolutional network, and the number of lanes and fire sources as output Y. Figure 1 As shown, temperature sensor data collected from each fire scenario is stored in a database. The database contains 183 columns and 1050 rows, representing the temperature and corresponding time collected from all sensors. Each 40 rows of continuous data are grouped into samples at 10-second intervals. For example, rows 1-40, 11-50, and 21-60 constitute the first three samples. Since the test set comprises 20%, the last sample in each fire scenario for the temperature and HRR prediction model is the data in rows 841-1050. Each fire scenario generates 102 samples, resulting in 49*102=4998 samples for all 49 fire scenarios. The labels for each sample are the desired outputs: for the fire spread identification model, the fire spread direction and number of fire sources are outputs; for the temperature and HRR prediction model, temperature and HRR are outputs. Since the fire size and number of fire sources vary over time, the text file storing the temperature field is converted to a standard binary file format in NumPy to speed up data loading.

[0080] S3. Joint Training of TCN Models: The training set divided in step S2 is input into the Temporal Convolutional Network (TCN) model to perform model training and parameter tuning, in order to construct a TCN-based fire spread identification model and temperature field prediction model. This process aims to simultaneously obtain the ability to identify fire spread characteristics and predict HRR and temperature field. Among them, the identification of fire spread characteristics includes the judgment of the number of fire sources and the direction of fire spread. The fire spread identification model is a multi-label classification model used to identify the number of fire sources and the direction of fire spread. The temperature field prediction model is a regression prediction model used to predict the temperature and HRR at future times.

[0081] Specifically, the S31 and TCN model structures are designed; the number of neurons in the input and output layers of the TCN convolutional neural network needs to be determined based on the actual problem. Based on the observational data obtainable in actual tunnel fire scenarios, time (t), temperature (T), wind speed (u), and temperature sensor coordinates (x and y coordinates) are used as input variables, i.e., input features X; HRR, number of fire sources (1 and 2), fire spread direction (non-spreading, lateral, and longitudinal), and temperature some time later are used as output variables, i.e., output features Y. Among these, HRR and temperature as output variables belong to a regression prediction model, while the number of fire sources and fire spread direction as output variables belong to a classification model. Figure 3 The algorithm structure of the TCN convolutional network is shown. Here, K is the kernel size and b is the dilation base.

[0082] Specifically, the fire spread pattern recognition and temperature field prediction in this embodiment require training such as Figure 2 (a) regression prediction model and as Figure 2 (b) Two classification models. For the regression task, time, temperature, wind speed, and sensor coordinates (x and y) are used as input variables X1, and HRR and temperature at the future time step are used as output variables Y1; for the classification task, time, temperature, wind speed, sensor coordinates (x and y), and HRR are used as input features X2, and the number of fire sources is used as output feature Y. 2，1 The direction of spread is used as the output feature Y 2，2 The fire source count is a binary classification task (0 and 1), requiring one binary classification output neuron, and the probability is obtained through the Sigmoid activation function. For fire spread direction prediction, since there are three cases (0, 1, 2), three output neurons are needed to form a multi-class classification problem, and the probability distribution of each possibility is obtained using the Softmax activation function. Therefore, the input layer of the regression model requires 5 input neurons and the output layer requires 2 output neurons; the input layer of the classification model requires 6 neurons and the output layer requires 4 neurons. The Sigmoid and Softmax activation functions are shown in equations (6) and (7), respectively.

[0083] (6)

[0084] (7)

[0085] In the formula: z is the input value; i is the output value of the i-th neuron in the previous layer of the neural network; S i is the output probability of the i-th neuron in the previous layer of the neural network.

[0086] Regarding the number of neural network layers, research shows that if the functional relationship simulated by the neural network is continuous within its domain, one or two hidden layers are usually sufficient to solve most regression approximation problems. Considering the requirements of the inverse problem, the size of the sample set, and the training difficulty, the regression prediction model will use a neural network with one TCN hidden layer; the multi-label classification model will use a neural network with two TCN hidden layers, incorporating activation functions as shown in equations (6) and (7), and outputting the probability results of identifying the number of fire sources and the direction of fire spread. Its value range is [0,1], and the larger the output value, the greater the likelihood that the neural network model considers the input data to belong to this label. Finally, the TCN network model for HRR inversion, i.e., temperature prediction and fire spread identification, is as follows: Figure 2 As shown.

[0087] S32. Model Training and Evaluation Metrics; The most widely used evaluation metrics in regression prediction models are Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). MSE refers to the expected value of the squared difference between the predicted and actual parameter values, as shown in equation (8). The smaller the MSE value, the higher the accuracy of the neural network prediction, making it the most commonly used loss function for regression prediction models. RMSE is the arithmetic square root of the mean square error, which can describe the deviation between the predicted and actual values ​​from different gradients. MAE is the average of the absolute errors between the predicted and actual values. Compared with the other two metrics, MAE can better reflect the actual situation of the prediction error, but it is not sensitive to large deviations in the prediction. MSE is selected as the loss function to evaluate the training effect of the regression TCN neural model.

[0088] (8)

[0089] Furthermore, to optimize the model training process, the cross-entropy loss function is adopted, and its mathematical expression is as follows:

[0090] (9)

[0091] In the formula: For the true label of the i-th data point, The model identifies the labels, and N represents the number of labels. All algorithm experiments were run on PyTorch, and the neural network training environment is shown in Table 2.

[0092] Table 2 Training Environment for TCN Neural Network

[0093]

[0094] The loss function quantifies the deviation between actual observations and model predictions. The model training process aims to improve prediction accuracy by reducing the value of the loss function. Generally, a lower loss function value usually means higher prediction accuracy. During model training, the network parameters are iteratively updated based on feedback from the loss function to continuously improve model performance.

[0095] Both the regression and classification models employed the Adam optimizer (adaptive moment estimator) with an initial learning rate of 0.0001. For the regression task, the number of neurons in the hidden layers was initially set to 64; in the classification task, the number of neurons in the two hidden layers was initialized to 64 and 128, respectively. The final TCN neural network parameter configuration is shown in Table 3.

[0096] Table 3 Hyperparameter Selection

[0097]

[0098] S4. Model Validation and Application: The trained TCN model is validated using the test set and validation set in step S2 to obtain the optimal prediction model; real-time tunnel environment data is input into the optimal prediction model, and the current fire spread identification results, as well as future temperature field and HRR prediction data are output in parallel.

[0099] The calculation method for the prediction accuracy of the tunnel fire spread identification and temperature field prediction model is shown in Equation (10):

[0100] (10)

[0101] In the formula: To improve prediction accuracy; Let i be the initial value of the i-th data. is the predicted value obtained by the model, and n is the number of data points.

[0102] A prediction system based on tunnel fire spread identification and temperature field prediction methods includes:

[0103] The multi-source data acquisition module includes a simulation database module and an experimental database module. The simulation database module is used to store time-series data under various fire conditions generated by the fire dynamics simulator FDS. The simulation time-series data includes temperature, wind speed, HRR, and sensor location information. The experimental database module is used to store experimental time-series data under various fire conditions obtained through scaled-down model fire tests.

[0104] The data preprocessing module has its input connected to the output of the simulation database module and the experimental database module, respectively. It is used to fuse and preprocess the simulation data and experimental data, and to divide the complete dataset into training set and test set according to a preset ratio.

[0105] The TCN identification and prediction model module has its input end connected to the output end of the aforementioned data preprocessing module. It is used to receive the aforementioned training set for model training and parameter tuning, and to use the validation set for model validation to generate the optimal prediction model. Real-time tunnel environment data is input into the optimal prediction model, and the current fire spread identification results, as well as future temperature field and HRR prediction data, are output in parallel.

[0106] Results Analysis

[0107] 1) Performance Analysis of Fire Spread Recognition Model

[0108] The TCN model is used to determine the most likely label as the final recognition result for a specific parameter. For example, in the task of identifying the number of fire sources, if the output result is [0.2, 0.8], it means that the probability of one fire source is 0.2, and the probability of two fire sources is 0.8. The model will then select "two fire sources" as the final prediction result. The same processing logic is used to identify the direction of fire spread. By comparing the model's final output with the true value, the accuracy of the recognition can be determined: if they match, the recognition is considered successful; if they do not match, the recognition fails. The test set contains 1000 data samples, and the parameters in these samples are evenly distributed. The detailed recognition results of the TCN model for these parameters are shown in Table 4.

[0109] Table 4. Results of Tunnel Fire Spread Identification

[0110]

[0111] As shown in Table 4, the TCN model achieved 100% accuracy when there was only one fire source and the tunnel was under natural ventilation. Under longitudinal critical ventilation conditions, the model's accuracy dropped to 96%. This phenomenon is consistent with existing temperature field analysis results, indicating that the critical wind speed affects the downstream movement of the highest temperature area, a change somewhat similar to the fire spread pattern under longitudinal critical ventilation conditions. When there were two fire sources in the tunnel and longitudinal fire spread occurred, the use of the critical wind speed also led to a decrease in accuracy, lower than under natural ventilation conditions. In contrast, the recognition rate for lateral fire spread was relatively stable, less affected by longitudinal wind speed, with only a 1% difference in accuracy regardless of ventilation conditions. Overall, the TCN fire spread recognition model demonstrates high reliability in fire spread identification.

[0112] 2) Performance analysis of temperature field and HRR prediction model

[0113] Predicting temperature field and HRR is a regression prediction problem. The loss behavior in the test set during model training is as follows: Figure 6 As shown, the loss value fluctuates somewhat on both the training and test sets, with the test set showing slightly greater fluctuations than the training set – a common phenomenon in neural network model training. The loss value decreases continuously with increasing epochs, eventually stabilizing between 400 and 600 epochs. Training efficiency is high before 200 epochs, resulting in a rapid decrease in loss value, followed by a slow decline. The loss value stabilizes after 450 epochs, indicating that the model has converged.

[0114] Figure 7 This demonstrates the predictive performance of the regression prediction model for heat release rate (HRR) and temperature under a specific operating condition. (By...) Figure 7 (a) shows that although the actual HRR value contains significant noise, the predicted value exhibits a relatively smooth curve and maintains a consistent trend with the actual HRR. This indicates that the regression model possesses accurate inversion and strong predictive ability. Figure 7 (b) shows that the regression model successfully captured the temperature change trend at this location and showed smaller fluctuations compared to the actual value, confirming that the model has good generalization ability.

[0115] Table 5 shows the overall prediction accuracy of all 183 temperature sensors in the tunnel under all working conditions. Error analysis revealed that although the overall error tended to increase gradually as the prediction time increased, even under such circumstances, the maximum error rate was limited to within 20%, which is within the acceptable error range of the model.

[0116] Table 5 Temperature Prediction Accuracy

[0117]

[0118] The temperature prediction results output by the prediction model are arranged into a dataset as shown in Equation (11). This dataset is arranged by time, and each time point contains the temperature values ​​at all locations.

[0119] (11)

[0120] After obtaining the matrix, the data is read through a temperature field reconstruction model to perform dynamic visualization prediction of the temperature field. The numerical simulation and experimental conditions for lateral spread under critical longitudinal wind speeds are used for prediction, and the actual and predicted temperature fields are as follows: Figure 8 As shown. Figure 8 (a) is a contour map of the actual temperature field. Figure 8 (b) is a temperature field contour map predicted 40 seconds in advance, from Figure 8 It can be seen that when TCN predicts and reconstructs the temperature field 40 seconds in advance, the actual temperature field can be accurately visualized and predicted by the TCN prediction model and the temperature field reconstruction model, indicating that the TCN temperature prediction model has good generalization ability and the time-varying temperature field prediction method has high reliability in the prediction of ceiling temperature field.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying the spread of fire in tunnels and predicting the temperature field, characterized in that, Includes the following steps: S1. Multi-source data acquisition: Simulation time series data generated by the fire dynamics simulator and test time series data collected by the scaled model fire test are acquired in parallel. The simulation time series data includes temperature, wind speed, heat release rate (HRR) and sensor location information. A database is established based on the real-time HRR, temperature and wind speed data obtained by the simulator and the temperature field data obtained by the scaled model. This database contains different fire spread conditions, temperature and HRR. S2. Data Fusion and Preprocessing: Integrate the various working condition data from the simulation and experiment in step S1, preprocess the integrated data to construct a complete dataset, and divide the dataset into a training set for model training and a test set for model testing. S3, TCN Model Joint Training: The training set divided in step S2 is input into the Temporal Convolutional Network (TCN) model to perform model training and parameter tuning in order to build a fire spread identification model and a temperature field prediction model based on TCN; this process aims to simultaneously obtain the ability to predict HRR and future temperature field, as well as the ability to identify fire spread characteristics. S4. Model Validation and Application: The trained TCN model is validated using the test set and validation set in step S2 to obtain the optimal prediction model; real-time tunnel environment data is input into the optimal prediction model, and the current fire spread identification results, as well as future temperature field and HRR prediction data are output in parallel.

2. The method for identifying tunnel fire spread and predicting temperature field as described in claim 1, characterized in that, In step S1, the temperature data during the fire development process has obvious time correlation and belongs to typical time series data. Therefore, the TCN algorithm for processing time series data is used to train the temperature prediction model. Its core idea is causal convolution.

3. The method for identifying tunnel fire spread and predicting temperature field as described in claim 2, characterized in that, Step S2 specifically includes: S21. Data Cleaning: The simulation time series data is generated by FDS software and does not require data cleaning. In contrast, for scaled-down model test data with missing or abnormal temperature data, linear interpolation is used to fill in the missing values, and... Criteria for outlier detection and removal; smoothing of repeated temperature values ​​using moving averages; fire data is transformed into a format suitable for training and testing TCN convolutional network models, improving model accuracy and generalization performance. S22, Data labeling; The direction and quantity of fire spread are labeled by encoding; No spread, horizontal spread, and vertical spread are represented by labels 0, 1, and 2 respectively; The number of fire sources is represented by labels 1 and 2. The label of the number of fire sources is changed to 2 when the spread occurs, so as to form a fire spread identification dataset. S23. Normalization processing: In order to avoid slowing down the learning and convergence speed of the TCN neural network, the min-max normalization method shown in Equation (4) is adopted to transform the data into a unified interval, such as [0, 1], so as to uniformly process multiple types of data and improve the prediction accuracy of the model. (4) In the formula: x i Initialize the data of type i; x i * represents the normalized value; max x i For data of type i, find the maximum value; min x i The minimum value of data of type i; S24. Dataset partitioning: Divide the dataset into training and testing sets according to a ratio of 50-80%: 20-50%; record temperature data for each working condition using 150-200 thermocouples, and store the data in a CSV file as input information for TCN; collect data 3 times per second to form the temperature dataset.

4. The method for identifying tunnel fire spread and predicting temperature field as described in claim 3, characterized in that, Step S3 specifically includes: S31, TCN model structure design; Fire spread situation recognition and temperature field prediction require training two models: a regression prediction model and a classification model; The number of fire sources is a binary classification task (0 and 1), requiring one binary classification output neuron, and the probability is obtained through the Sigmoid activation function; For fire spread direction prediction, since there are three cases (0, 1, 2), three output neurons are needed to form a multi-class problem, and the probability distribution of each possibility is obtained by using the Softmax activation function; Therefore, the input layer of the regression model requires 5 input neurons, and the output layer requires 2 output neurons; The input layer of the classification model requires 6 neurons, and the output layer requires 4 neurons; Among them, the Sigmoid and Softmax activation functions are shown in equations (6) and (7): (6) (7) In the formula: z is the input value; i is the output value of the i-th neuron in the previous layer of the neural network; S i This represents the output probability of the i-th neuron in the previous layer of the neural network. S32, Model Training and Evaluation Metrics; The MSE loss function shown in Equation (8) is used to evaluate the training effect of the regression TCN neural model, and the cross-entropy loss loss shown in Equation (9) is used to optimize the training process of the model: (8) (9)。 5. The method for identifying tunnel fire spread and predicting temperature field as described in claim 4, characterized in that, In step S31, for the regression task, time, temperature, wind speed, and sensor coordinates (x and y) are used as input variables X1, and HRR and temperature at future time steps are used as output variables Y1; for the classification task, time, temperature, wind speed, sensor coordinates (x and y), and HRR are used as input features X2, and the number of fire sources is used as output feature Y. 2，1 The direction of spread is used as the output feature Y 2，2 .

6. The method for identifying tunnel fire spread and predicting temperature field as described in claim 5, characterized in that, The calculation method for the prediction accuracy of the prediction model in step S4 is shown in equation (10): (10) In the formula: To improve prediction accuracy; Let i be the initial value of the i-th data. is the predicted value obtained by the model, and n is the number of data points.

7. A prediction system based on the tunnel fire spread identification and temperature field prediction method according to any one of claims 1 to 6, characterized in that, include: The multi-source data acquisition module includes a simulation database module and an experimental database module. The simulation database module is used to store time-series data under various fire conditions generated by the fire dynamics simulator FDS. The simulation time-series data includes temperature, wind speed, HRR, and sensor location information. The experimental database module is used to store experimental time-series data under various fire conditions obtained through scaled-down model fire tests. The data preprocessing module has its input connected to the output of the simulation database module and the experimental database module, respectively. It is used to fuse and preprocess the simulation data and experimental data, and to divide the complete dataset into training set and test set according to a preset ratio. The TCN identification and prediction model module has its input end connected to the output end of the aforementioned data preprocessing module. It is used to receive the aforementioned training set for model training and parameter tuning, and to use the validation set for model validation to generate the optimal prediction model. Real-time tunnel environment data is input into the optimal prediction model, and the current fire spread identification results, as well as future temperature field and HRR prediction data, are output in parallel.

Images