A method and system for predicting the propagation of a fire in a confined space based on a physical information neural network

Abstract

The present application relates to the technical field of confined space fire flashover prediction. A confined space fire flashover prediction method and system based on physical information neural network are provided. The steps are as follows: obtaining sensor observation data and initial fuel distribution data; inputting the space-time coordinate vector and the initial fuel distribution data into the Mamba neural network to obtain temperature, incident radiation and fuel mass fraction prediction values; calculating the space-time derivatives of the prediction values, constructing physical residual terms containing energy conservation, radiation transfer and fuel consumption equation residuals, and boundary residual terms and observation residual terms; combining the three types of residuals by weighting to construct a loss function, updating the network parameters and the physical parameters to be learned through an optimization algorithm; and determining the flashover occurrence time when the temperature reaches the flashover critical value based on the updated network prediction temperature field. The present application solves the problems of poor reliability and prediction deviation caused by the lack of physical law constraints in existing data-driven fire prediction models.

Description

Technical Field

[0001] This invention relates to the field of confined space fire flashover prediction technology, and more specifically, to a confined space fire flashover prediction method and system based on a physical information neural network. Background Technology

[0002] In the field of fire prediction technology, particularly in the prediction of flashover in confined space fires, existing technologies have proposed various methods based on sensor data and machine learning. For example, patent document CN120763505A utilizes wearable temperature sensors and radiative heat flux sensors to collect data sequences and performs predictions using a model combining convolutional neural networks (CNN) and long short-term memory networks (LSTM). This method trains the model based on historical data, outputs predicted temperature and radiative heat flux data for future moments, and issues an alarm when the predicted values ​​exceed a critical value. This method relies on a data-driven machine learning model, capable of extracting features from the fire development process and making predictions, but it does not incorporate physical laws as constraints for the model.

[0003] However, these existing methods have significant drawbacks. Because the models are entirely data-driven and lack the constraints of physical laws, their interpretability is poor, and predictions may not conform to physical principles, thus affecting reliability and stability. This is particularly true in confined space fires, where fire development involves complex physical processes such as energy conservation, radiation transfer, and fuel consumption. Purely data-driven models struggle to accurately capture these physical mechanisms, potentially leading to prediction bias. Furthermore, existing methods typically use general neural network architectures, such as CNN-LSTM, which have limited processing capabilities for long-term time-series data and do not incorporate physical information to enhance the model's generalization ability and robustness, thus limiting their effectiveness in real-world fire rescue scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting flashover fires in confined spaces based on physical information neural networks, aiming to solve the problems of poor reliability and prediction bias caused by the lack of physical constraints in existing data-driven fire prediction models.

[0005] This invention is achieved through the following technical solution:

[0006] A method for predicting flashover fires in confined spaces based on physical information neural networks includes the following steps:

[0007] Acquire sensor observation data and initial fuel distribution data in a confined space. The sensor observation data includes time-series temperature data, heat flux data, and vent gas velocity data.

[0008] The spatiotemporal coordinate vector and the initial fuel distribution data are input into the Mamba neural network. The Mamba neural network is used to perform function mapping to obtain the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction.

[0009] The predicted temperature, predicted incident radiation, and predicted fuel mass fraction are calculated to obtain the corresponding time derivative and spatial derivative.

[0010] Based on the time derivative and spatial derivative, physical residual terms are constructed, which include residuals of the energy conservation equation, residuals of the radiation transfer equation, and residuals of the fuel consumption equation; wherein, the residuals of the energy conservation equation combine gas density, specific heat capacity, thermal conductivity, fuel calorific value, activation energy parameter, and radiative heat flux vector.

[0011] Construct boundary residuals and observation residuals. The boundary residuals include boundary residuals of the energy equation, boundary residuals of the radiation equation, boundary residuals of the fuel consumption equation, and boundary residuals of the vent. The observation residuals include temperature observation residuals and thermal radiation flux observation residuals.

[0012] The physical residual, boundary residual, and observation residual are combined in a weighted sum form to construct a loss function, which is used to measure the degree of violation of physical laws and the fitting error of observation data.

[0013] The loss function is minimized by an optimization algorithm, and the parameters of the Mamba neural network and the physical parameters to be learned are updated. The physical parameters to be learned include the absorption coefficient, the Arrhenius reaction rate coefficient, and the activation energy parameter.

[0014] Based on the updated Mamba neural network, the temperature field change in the confined space is predicted, and the time of flashover is determined when the predicted temperature field reaches the flashover critical temperature.

[0015] Optionally, the specific process for acquiring sensor observation data and initial fuel distribution data in the confined space is as follows:

[0016] Temperature and heat flux data are collected in real time by deploying temperature and heat flux sensors at multiple locations within a confined space.

[0017] Gas velocity data at the ventilation opening is collected by a gas velocity sensor installed at the ventilation opening in the confined space.

[0018] Based on the initial configuration information of the fuel in the confined space, initial fuel distribution data is obtained, which includes fuel type, location and initial fuel mass fraction.

[0019] Optionally, the specific process of inputting the spatiotemporal coordinate vector and the initial fuel distribution data into a Mamba neural network, and obtaining the predicted temperature, predicted incident radiation, and predicted fuel mass fraction through function mapping by the Mamba neural network is as follows:

[0020] Using the serialization architecture of the Mamba neural network, the spatiotemporal coordinate vector and the initial fuel distribution data are taken as input, forward propagation calculation is performed, and the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction are output.

[0021] The Mamba neural network adopts an architecture that combines a selective state-space model with a convolutional neural network. It learns the spatiotemporal dynamic correlation in the fire development process through sequential state-space modeling, and realizes a continuous differentiable mapping from spatiotemporal coordinates to physical fields.

[0022] The parameters of the Mamba neural network are initialized using the Xavier initializer, and convergence is accelerated by a pre-training starter.

[0023] Optionally, the specific process of calculating the predicted temperature, predicted incident radiation, and predicted fuel mass fraction to obtain the corresponding time derivative and spatial derivative is as follows:

[0024] Automatic differentiation is performed on the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction to obtain time derivatives and spatial derivatives. The time derivatives include the time derivatives of temperature, incident radiation, and fuel mass fraction. The spatial derivatives include spatial gradients and Laplace operators. The spatial gradients include the spatial gradients of temperature, incident radiation, and fuel mass fraction. The Laplace operators include the Laplace operator of temperature and the Laplace operator of incident radiation.

[0025] The automatic differential calculation is performed during each forward inference process, and partial derivatives are calculated based on the chain rule. These partial derivatives are used to subsequently construct physical residuals, boundary residuals, and observation residuals.

[0026] Optionally, the specific process of constructing the physical residual term based on the time derivative and spatial derivative is as follows:

[0027] Based on the time and spatial derivatives of the predicted temperature, the spatial derivative of the predicted incident radiation, and the time derivative of the predicted fuel mass fraction, the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction are substituted into predefined physical equations to calculate the residuals of the energy conservation equation, the radiation transfer equation, and the fuel consumption equation.

[0028] Among them, the residual of the energy conservation equation The expression for is shown in equation (1):

[0029]

[0030] in, The density of the gas; Specific heat capacity of the gas; This is a predicted temperature value; The thermal conductivity of the gas; This is the partial derivative of the temperature prediction with respect to time. The spatial gradient of the predicted temperature value; The gas velocity field is estimated based on the gas velocity data from the ventilation opening; This refers to the calorific value of the fuel. The Arrhenius reaction rate coefficient; This is a predicted value for fuel mass fraction; The activation energy parameter; This is the universal gas constant; It is the radiative heat flux vector; The divergence of the radiative heat flux vector;

[0031] The residual of the radiation transfer equation The expression for is shown in equation (2):

[0032]

[0033] in, The absorption coefficient to be learned; The Stefan-Boltzmann constant; This is the predicted value of the incident radiation;

[0034] The residual of the fuel consumption equation The expression for is shown in equation (3):

[0035]

[0036] in, This is the time derivative of the predicted fuel mass fraction.

[0037] Optionally, the specific process of constructing the boundary residual term and the observation residual term is as follows:

[0038] Based on the predicted temperature, predicted incident radiation, and predicted fuel mass fraction and their spatial derivatives, boundary residual terms are calculated according to the boundary conditions. The boundary residual terms include the boundary residuals of the energy equation, the radiation equation, the fuel consumption equation, and the vent boundary residuals.

[0039] Based on the predicted temperature and thermal radiation flux, the predicted values ​​are compared with the temperature and thermal radiation flux data in the sensor observation data to calculate the observation residuals, which include the temperature observation residuals and the thermal radiation flux observation residuals.

[0040] Optionally, the specific process of constructing the loss function is as follows:

[0041] The physical residual, boundary residual, and observation residual are combined in a weighted sum form to construct the loss function. ;in, Represents the parameters of the Mamba neural network; This represents the set of physical parameters to be learned;

[0042] The expression for the loss function is shown in equation (4) below:

[0043]

[0044] in, The physical residual term represents the sum of squares and average of the residuals of the energy conservation equation, the radiation transfer equation, and the fuel consumption equation, and is calculated as shown in equation (5) below:

[0045]

[0046] This is the number of sampling points used to calculate the physical residual; The boundary residual term represents the sum of the squares of the boundary residuals of the energy equation, radiation equation, fuel consumption equation, and vent, and is calculated as shown in equation (6) below:

[0047]

[0048] This represents the number of boundary sampling points used to calculate the boundary residuals; The boundary residuals of the energy equation; The boundary residuals of the radiation equation; The boundary residuals of the fuel consumption equation; For the residual at the ventilation opening boundary; The observation residual term represents the sum of squares and average of the temperature observation residual and the thermal radiation flux observation residual, and is calculated as shown in equation (7) below:

[0049]

[0050] This refers to the number of temperature observation points; For temperature observation residuals; The number of observation points for thermal radiation flux; For the residual of thermal radiation flux observation;

[0051] , and These are the weighting coefficients for the physical residual term, the boundary residual term, and the observation residual term, respectively, used to balance the contribution of the degree of violation of physical laws and the fitting error of the observation data to the loss function.

[0052] Based on the same inventive concept, this invention also provides a confined space fire flashover prediction system based on a physical information neural network, used to implement the aforementioned confined space fire flashover prediction method based on a physical information neural network, comprising:

[0053] The data acquisition module is used to acquire sensor observation data and initial fuel distribution data in a confined space. The sensor observation data includes time-series temperature data, heat flux data, and vent gas velocity data. The initial fuel distribution data includes fuel type, location, and initial fuel mass fraction.

[0054] The neural network processing module is used to input the spatiotemporal coordinate vector and the initial fuel distribution data into the Mamba neural network, and to obtain the temperature prediction value, incident radiation prediction value and fuel mass fraction prediction value through the function mapping of the Mamba neural network. The Mamba neural network adopts an architecture that combines a selective state space model and a convolutional neural network, and learns the spatiotemporal dynamic correlation in the fire development process through sequential state space modeling.

[0055] The derivative calculation module is used to calculate the predicted temperature, predicted incident radiation, and predicted fuel mass fraction, and obtain the corresponding time derivative and spatial derivative through automatic differentiation. The time derivative includes the temperature time derivative, the incident radiation time derivative, and the fuel mass fraction time derivative, and the spatial derivative includes the spatial gradient and the Laplace operator.

[0056] The residual construction module is used to construct physical residual terms, boundary residual terms, and observation residual terms based on the time derivative and spatial derivative. The physical residual terms include residuals of the energy conservation equation, residuals of the radiation transfer equation, and residuals of the fuel consumption equation. The boundary residual terms include boundary residuals of the energy equation, boundary residuals of the radiation equation, boundary residuals of the fuel consumption equation, and boundary residuals of the vent. The observation residual terms include temperature observation residuals and thermal radiation flux observation residuals.

[0057] The loss function construction module is used to combine the physical residual term, boundary residual term and observation residual term in a weighted sum form to construct a loss function, which is used to measure the degree of violation of physical laws and the fitting error of observation data.

[0058] The optimization and update module is used to minimize the loss function through an optimization algorithm and update the parameters of the Mamba neural network and the physical parameters to be learned, wherein the physical parameters to be learned include the absorption coefficient, the Arrhenius reaction rate coefficient, and the activation energy parameter.

[0059] The flashover prediction module is used to predict temperature field changes in a confined space based on an updated Mamba neural network. When the predicted temperature field value reaches the flashover critical temperature, the flashover time is determined.

[0060] Based on the same inventive concept, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described confined space fire flashover prediction method based on physical information neural network.

[0061] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described confined space fire flashover prediction method based on a physical information neural network.

[0062] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0063] By introducing residual constraints from physical equations such as energy conservation, radiation transfer, and fuel consumption, physical laws are deeply embedded into the neural network training process. This soft constraint improves the physical consistency and interpretability of the prediction results, effectively overcoming the shortcomings of pure data-driven models that may violate physical principles due to the lack of physical constraints. This significantly enhances the interpretability and reliability of the model.

[0064] By constructing a loss function using physical equations, the neural network is guided to learn the inherent laws of real physical processes during training, rather than relying solely on limited observation data. This enables a more accurate description of the nonlinear dynamic characteristics of fire development (such as energy transfer and fuel consumption), improving prediction accuracy and generalization performance for unseen scenarios or sparsely dataed regions.

[0065] Using the Mamba neural network as the core architecture, its state-space model characteristics can efficiently capture long-sequence dependencies, overcoming the limitations of traditional CNN-LSTM models in long-term fire data modeling, thus more stably predicting the entire fire development process, especially more accurately determining the flashover critical point.

[0066] By combining physical residuals, boundary residuals, and observation residuals, a complete constraint system covering the governing equations, boundary conditions, and measured data is constructed. This ensures that the model remains consistent with physical reality in multiple dimensions, including spatial distribution, boundary effects, and observation fitting, thereby improving the stability and applicability of predictions in complex and confined spatial environments. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the confined space fire flashover prediction method based on a physical information neural network according to an embodiment of the present invention.

[0068] Figure 2 This is a schematic diagram of the structure of a confined space fire flashover prediction system based on a physical information neural network according to an embodiment of the present invention;

[0069] Figure 3 This is a schematic diagram of the neural network architecture of a confined space fire flashover prediction system based on physical information neural networks, according to an embodiment of the present invention. Detailed Implementation

[0070] The following is a detailed description of the embodiments, in conjunction with the accompanying drawings.

[0071] Reference Figure 1 A method for predicting flashover fires in confined spaces based on physical information neural networks includes the following steps:

[0072] Step 1: Acquire sensor observation data and initial fuel distribution data in the confined space. The sensor observation data includes time-series temperature data, heat flux data, and vent gas flow velocity data.

[0073] In some embodiments, the specific process of acquiring sensor observation data and initial fuel distribution data in a confined space is as follows:

[0074] Temperature and heat flux data are collected in real time by deploying temperature and heat flux sensors at multiple locations within a confined space.

[0075] Gas velocity data at the ventilation opening is collected by a gas velocity sensor installed at the ventilation opening in the confined space.

[0076] Based on the initial configuration information of the fuel in the confined space, initial fuel distribution data is obtained, which includes fuel type, location and initial fuel mass fraction.

[0077] Step 2: Input the spatiotemporal coordinate vector and the initial fuel distribution data into the Mamba neural network, and perform function mapping through the Mamba neural network to obtain the predicted temperature, predicted incident radiation, and predicted fuel mass fraction.

[0078] In some embodiments, the specific process of inputting the spatiotemporal coordinate vector and the initial fuel distribution data into a Mamba neural network, and performing function mapping through the Mamba neural network to obtain the predicted temperature, predicted incident radiation, and predicted fuel mass fraction is as follows:

[0079] The Mamba neural network, using a sequential architecture, takes the spatiotemporal coordinate vector and the initial fuel distribution data as input, performs forward propagation calculations, and outputs the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction (when observations are sparse, a low-dimensional output strategy can be used, outputting only the upper-level room temperature and a low-dimensional HRR representation, and constructing the corresponding residuals accordingly). The Mamba neural network employs an architecture combining a selectively structured state-space model and a convolutional neural network. Through sequential state-space modeling, it learns the spatiotemporal dynamic correlations during fire development, achieving a continuous differentiable mapping from spatiotemporal coordinates to the physical field. The parameters of the Mamba neural network are initialized using a Xavier initializer, and convergence is accelerated through a pre-training initiator. (Spatiotemporal coordinate vector) Representing spatial location in a confined space (such as three-dimensional coordinates) ) and time point These coordinates are generated through discretization sampling, covering the entire confined space and the time period of fire development. Initial fuel distribution data includes fuel type, spatial distribution of fuel location, and initial fuel mass fraction. This initial fuel distribution data is obtained based on initial fuel configuration information within the confined space, such as information determined through site surveys or design drawings. Input data is preprocessed before being fed into the network, including normalization, to ensure numerical stability. For example, spatiotemporal coordinates may be scaled according to spatial scale and time range, while fuel distribution data may be converted to a scalar field. The Mamba neural network employs a sequential architecture combining Selective State-Space Models (SSMs) and Convolutional Neural Networks (CNNs). SSMs are used to capture long-range dependencies in the time series, while CNNs are used to extract spatial features, thus jointly learning the spatiotemporal dynamic correlations during fire development. The network input is a spatiotemporal coordinate vector. The input includes initial fuel distribution data (as an additional conditional parameter). Internally, the input is first transformed into high-dimensional features through an embedding layer, then processed for the time dimension through an SSM layer, and finally processed for the spatial dimension through a CNN layer. Ultimately, the network outputs three scalar values: a predicted temperature value... Predicted value of incident radiation and fuel mass fraction prediction value Network parameters (such as weights and biases) are initialized using a Xavier initializer (also known as a Glorot initializer) to maintain gradient stability. Furthermore, convergence is accelerated via a pre-trained starter (warm-star), for example, by initially training the network using historical fire data or simulated data to initialize the state-space parameters.

[0080] During the forward propagation process, the input data is transformed through each layer of the network. Specifically:

[0081] The SSM layer uses state-space equations (such as linear time-invariant systems) to process time series and outputs hidden states; the CNN layer applies convolutional kernels to extract spatial features and introduces nonlinearity through activation functions (such as ReLU); the final output layer maps the features to the predicted values ​​of three physical fields, ensuring that the output is continuous and differentiable, facilitating subsequent automatic differentiation. For example, given a spatiotemporal point... Network output , and These values ​​represent the predicted temperature, incident radiation, and fuel mass fraction at that point.

[0082] The output physics predictions are used in subsequent steps, including automatic differentiation to calculate the physical residuals. Since the network output is continuously differentiable, the derivative calculation module can accurately calculate the time and spatial derivatives. Through this process, the Mamba neural network achieves a mapping from spatiotemporal coordinates and fuel distribution to key physics fields, providing a foundation for flashover prediction.

[0083] Step 3: Calculate the predicted temperature, predicted incident radiation, and predicted fuel mass fraction to obtain the corresponding time derivative and spatial derivative.

[0084] In some embodiments, the specific process of calculating the predicted temperature, predicted incident radiation, and predicted fuel mass fraction to obtain the corresponding time derivative and spatial derivative is as follows:

[0085] Automatic differentiation is performed on the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction to obtain time and spatial derivatives. The time derivatives include the time derivatives of temperature, incident radiation, and fuel mass fraction. The spatial derivatives include the spatial gradient and the Laplace operator. The spatial gradient includes the spatial gradients of temperature, incident radiation, and fuel mass fraction. The Laplace operator includes the Laplace operator of temperature and the Laplace operator of incident radiation. For cases with sparse or low-dimensional output, dimensionality reduction / approximate derivative expressions are used (e.g., for time difference or only constructing upper-level energy conservation residuals) to avoid unidentifiable situations.

[0086] The automatic differential calculation is performed during each forward inference process, and partial derivatives are calculated based on the chain rule. These partial derivatives are used to subsequently construct physical residuals, boundary residuals, and observation residuals.

[0087] Step 4: Based on the time derivative and spatial derivative, construct the physical residual terms, which include the residuals of the energy conservation equation, the radiation transfer equation, and the fuel consumption equation; wherein, the residuals of the energy conservation equation combine gas density, specific heat capacity, thermal conductivity, fuel calorific value, activation energy parameter, and radiative heat flux vector.

[0088] In some embodiments, the specific process of constructing the physical residual term based on the time derivative and spatial derivative is as follows:

[0089] Based on the time and spatial derivatives of the predicted temperature, the spatial derivative of the predicted incident radiation, and the time derivative of the predicted fuel mass fraction, the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction are substituted into predefined physical equations to calculate the residuals of the energy conservation equation, the radiation transfer equation, and the fuel consumption equation.

[0090] Among them, the residual of the energy conservation equation The expression for is shown in equation (1):

[0091]

[0092] in, The density of the gas; Specific heat capacity of the gas; This is a predicted temperature value; The thermal conductivity of the gas; This is the partial derivative of the temperature prediction with respect to time. The spatial gradient of the predicted temperature value; The gas velocity field is estimated based on the gas velocity data from the ventilation opening; This refers to the calorific value of the fuel. The Arrhenius reaction rate coefficient; This is a predicted value for fuel mass fraction; The activation energy parameter; This is the universal gas constant; It is the radiative heat flux vector; The divergence of the radiative heat flux vector;

[0093] The residual of the radiation transfer equation The expression for is shown in equation (2):

[0094]

[0095] in, The absorption coefficient to be learned; The Stefan-Boltzmann constant; This is the predicted value of the incident radiation;

[0096] The residual of the fuel consumption equation The expression for is shown in equation (3):

[0097]

[0098] in, This is the time derivative of the predicted fuel mass fraction.

[0099] Step 5: Construct boundary residuals and observation residuals. The boundary residuals include boundary residuals of the energy equation, radiation equation, fuel consumption equation, and vent. The observation residuals include temperature observation residuals and thermal radiation flux observation residuals.

[0100] In some embodiments, the specific process of constructing the boundary residual term and the observation residual term is as follows:

[0101] Based on the predicted temperature, predicted incident radiation, and predicted fuel mass fraction and their spatial derivatives, boundary residual terms are calculated according to the boundary conditions. The boundary residual terms include the boundary residuals of the energy equation, the radiation equation, the fuel consumption equation, and the vent boundary residuals.

[0102] Based on the predicted temperature and thermal radiation flux, the predicted values ​​are compared with the temperature and thermal radiation flux data in the sensor observation data to calculate the observation residuals, which include the temperature observation residuals and the thermal radiation flux observation residuals.

[0103] Step 6: Combine the physical residual, boundary residual, and observation residual in a weighted sum form to construct a loss function, which is used to measure the degree of violation of physical laws and the fitting error of observation data.

[0104] In some embodiments, the specific process of constructing the loss function is as follows:

[0105] The physical residual, boundary residual, and observation residual are combined in a weighted sum form to construct the loss function. ;in, Represents the parameters of the Mamba neural network; This represents the set of physical parameters to be learned, including the absorption coefficient. Arrhenius reaction rate coefficient and activation energy parameters ;

[0106] The expression for the loss function is shown in equation (4) below:

[0107]

[0108] in, The physical residual term represents the sum of squares and average of the residuals of the energy conservation equation, the radiation transfer equation, and the fuel consumption equation, and is calculated as shown in equation (5) below:

[0109]

[0110] This is the number of sampling points used to calculate the physical residual; The boundary residual term represents the sum of the squares of the boundary residuals of the energy equation, radiation equation, fuel consumption equation, and vent, and is calculated as shown in equation (6) below:

[0111]

[0112] This represents the number of boundary sampling points used to calculate the boundary residuals; The boundary residuals of the energy equation; The boundary residuals of the radiation equation; The boundary residuals of the fuel consumption equation; For the residual at the ventilation opening boundary; The observation residual term represents the sum of squares and average of the temperature observation residual and the thermal radiation flux observation residual, and is calculated as shown in equation (7) below:

[0113]

[0114] This refers to the number of temperature observation points; For temperature observation residuals; The number of observation points for thermal radiation flux; For the residual of thermal radiation flux observation;

[0115] , and These are the weighting coefficients for the physical residual term, the boundary residual term, and the observation residual term, respectively, used to balance the contribution of the degree of violation of physical laws and the fitting error of the observation data to the loss function.

[0116] Step 7: Minimize the loss function using an optimization algorithm, and update the parameters of the Mamba neural network and the physical parameters to be learned, including the absorption coefficient, the Arrhenius reaction rate coefficient, and the activation energy parameter.

[0117] In some embodiments, a gradient descent-type optimization algorithm, such as the Adam (Adaptive Moment Estimation) optimizer, can be selected to minimize the loss function. Optimizer hyperparameter settings include the learning rate. First-order moment attenuation factor Second-order moment attenuation factor and numerical stability constant The loss function is calculated using Automatic Differentiation (AD). Regarding network parameters and physical parameters gradient and Automatic differentiation calculates the gradient precisely based on the chain rule after each forward inference and passes the gradient value to the optimizer. The gradient calculation covers all physical residuals, boundary residuals, and observation residuals, ensuring that the loss function is minimized while taking into account both data fitting and physical constraints.

[0118] Update parameters using Adam optimizer and The specific update formulas are shown in equations (8) and (9) below:

[0119] For parameters :

[0120]

[0121] For parameters :

[0122]

[0123] in, and For parameters First and second moment estimates; and For parameters Estimate the first and second moments. Initially, , , and All vectors are initialized to zero.

[0124] Repeat the following steps until the loss function converges or the maximum number of iterations is reached:

[0125] A batch of spatiotemporal points is sampled from the training data, including interior points (for physical residuals), boundary points (for boundary residuals), and observation points (for observation residuals); the loss function is calculated via forward propagation. Backpropagation to calculate gradient and Update parameters and Monitor the loss value; if the change in the loss value is less than a preset threshold or the number of iterations reaches the upper limit, stop the optimization.

[0126] Through the above optimization process, the parameters of the Mamba neural network are... and physical parameters The joint optimization ensures that the model's predicted temperature field, incident radiation, and fuel mass fraction not only fit the observed data but also strictly adhere to physical laws such as energy conservation, radiation transfer, and fuel consumption, thereby improving the accuracy and reliability of flashover prediction.

[0127] Step 8: Based on the updated Mamba neural network, predict the temperature field changes in the confined space. When the predicted temperature field value reaches the flashover critical temperature, determine the flashover time.

[0128] In some embodiments, the time-varying sequence of temperature prediction values ​​can be extracted from the network output. Specifically, for each spatial location, the evolution curve of the temperature prediction value over time is analyzed to monitor whether the temperature shows an upward trend and approaches or reaches the flashover critical temperature. The flashover critical temperature is a preset value determined based on fire dynamics theory or experimental data (e.g., the typical critical temperature for flashover in confined spaces is approximately 600°C). The critical temperature can be adjusted according to the specific scenario and serves as a threshold for determining the occurrence of flashover. When the temperature prediction value at any spatial location reaches or exceeds the flashover critical temperature, flashover is determined to have occurred. The flashover occurrence time is defined as the time point when the temperature prediction value first reaches the critical temperature. If multiple locations reach the critical temperature simultaneously, the earliest time point is taken as the flashover occurrence time. This determination is based on global monitoring of the temperature field prediction values ​​to ensure timely detection of flashover signs. The final output is the time point of flashover occurrence, which may be accompanied by visualization results of the temperature field prediction (such as temperature distribution maps or time series curves) to support fire rescue decisions.

[0129] Based on the same inventive concept, and corresponding to any of the above embodiments, refer to... Figure 2 This invention provides a confined space fire flashover prediction system based on a physical information neural network, used to implement the aforementioned confined space fire flashover prediction method based on a physical information neural network, comprising:

[0130] The data acquisition module is used to acquire sensor observation data and initial fuel distribution data in a confined space. The sensor observation data includes time-series temperature data, heat flux data, and vent gas velocity data. The initial fuel distribution data includes fuel type, location, and initial fuel mass fraction.

[0131] The neural network processing module is used to input the spatiotemporal coordinate vector and the initial fuel distribution data into the Mamba neural network, and to obtain the temperature prediction value, incident radiation prediction value and fuel mass fraction prediction value through the function mapping of the Mamba neural network. The Mamba neural network adopts an architecture that combines a selective structured state space model with a convolutional neural network, and learns the spatiotemporal dynamic correlation in the fire development process through sequential state space modeling.

[0132] The derivative calculation module is used to calculate the predicted temperature, predicted incident radiation, and predicted fuel mass fraction, and obtain the corresponding time derivative and spatial derivative through automatic differentiation. The time derivative includes the temperature time derivative, the incident radiation time derivative, and the fuel mass fraction time derivative, and the spatial derivative includes the spatial gradient and the Laplace operator.

[0133] The residual construction module is used to construct physical residual terms, boundary residual terms, and observation residual terms based on the time derivative and spatial derivative. The physical residual terms include residuals of the energy conservation equation, residuals of the radiation transfer equation, and residuals of the fuel consumption equation. The boundary residual terms include boundary residuals of the energy equation, boundary residuals of the radiation equation, boundary residuals of the fuel consumption equation, and boundary residuals of the vent. The observation residual terms include temperature observation residuals and thermal radiation flux observation residuals.

[0134] The loss function construction module is used to combine the physical residual term, boundary residual term and observation residual term in a weighted sum form to construct a loss function, which is used to measure the degree of violation of physical laws and the fitting error of observation data.

[0135] The optimization and update module is used to minimize the loss function through an optimization algorithm and update the parameters of the Mamba neural network and the physical parameters to be learned. The physical parameters to be learned include the absorption coefficient, the Arrhenius reaction rate coefficient, and the activation energy parameter. When the observation is sparse, threshold / power law or low-dimensional parameterization can be used in combination with narrow priors to ensure accuracy.

[0136] The flashover prediction module is used to predict temperature field changes in a confined space based on an updated Mamba neural network. When the predicted temperature field value reaches the flashover critical temperature, the flashover time is determined.

[0137] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the confined space fire flashover prediction method based on physical information neural network of the embodiment.

[0138] Alternatively, the aforementioned electronic device may be a server.

[0139] In addition, this embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the confined space fire flashover prediction method based on a physical information neural network according to the embodiment.

[0140] It is understood that the processor in the embodiments of the present invention may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0141] The method steps in the embodiments of the present invention can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0142] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a storage medium or transmitted through a storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive (SSD)).

Claims

1. A method for predicting flashover in confined spaces based on physical information neural networks, characterized in that, Includes the following steps: Acquire sensor observation data and initial fuel distribution data in a confined space. The sensor observation data includes time-series temperature data, heat flux data, and vent gas velocity data. The spatiotemporal coordinate vector and the initial fuel distribution data are input into the Mamba neural network. The Mamba neural network is used to perform function mapping to obtain the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction. The predicted temperature, predicted incident radiation, and predicted fuel mass fraction are calculated to obtain the corresponding time derivative and spatial derivative. Based on the time derivative and spatial derivative, physical residual terms are constructed, including the residuals of the energy conservation equation, the radiative transfer equation, and the fuel consumption equation. The energy conservation equation residuals incorporate gas density, specific heat capacity, thermal conductivity, fuel calorific value, activation energy parameter, and radiative heat flux vector. The specific process for constructing the physical residual terms based on the time derivative and spatial derivative is as follows: Based on the time and spatial derivatives of the predicted temperature, the spatial derivative of the predicted incident radiation, and the time derivative of the predicted fuel mass fraction, the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction are substituted into predefined physical equations to calculate the residuals of the energy conservation equation, the radiation transfer equation, and the fuel consumption equation. Among them, the residual of the energy conservation equation The expression for is shown in equation (1): in, The density of the gas; Specific heat capacity of the gas; This is a predicted temperature value; The thermal conductivity of the gas; This is the partial derivative of the temperature prediction with respect to time. The spatial gradient of the predicted temperature value; The gas velocity field is estimated based on the gas velocity data from the ventilation opening; This refers to the calorific value of the fuel. The Arrhenius reaction rate coefficient; This is a predicted value for fuel mass fraction; The activation energy parameter; This is the universal gas constant; It is the radiative heat flux vector; The divergence of the radiative heat flux vector; The residual of the radiation transfer equation The expression for is shown in equation (2): in, The absorption coefficient to be learned; The Stefan-Boltzmann constant; This is the predicted value of the incident radiation; The residual of the fuel consumption equation The expression for is shown in equation (3): in, The time derivative of the predicted fuel mass fraction; Boundary residuals and observation residuals are constructed. Based on the predicted temperature, incident radiation, and fuel mass fraction and their spatial derivatives, boundary residuals are calculated according to boundary conditions. Based on the predicted temperature and thermal radiation flux, observation residuals are calculated by comparing them with temperature and thermal flux data from sensor observations. The boundary residuals include boundary residuals of the energy equation, radiation equation, fuel consumption equation, and vent boundary. The observation residuals include temperature observation residuals and thermal radiation flux observation residuals. The physical residual, boundary residual, and observation residual are combined in a weighted sum form to construct a loss function, which is used to measure the degree of violation of physical laws and the fitting error of observation data. The loss function is minimized by an optimization algorithm, and the parameters of the Mamba neural network and the physical parameters to be learned are updated. The physical parameters to be learned include the absorption coefficient, the Arrhenius reaction rate coefficient, and the activation energy parameter. Based on the updated Mamba neural network, the temperature field change in the confined space is predicted, and the time of flashover is determined when the predicted temperature field reaches the flashover critical temperature.

2. The confined space fire flashover prediction method based on physical information neural network as described in claim 1, characterized in that... The specific process for acquiring sensor observation data and initial fuel distribution data in the confined space is as follows: Temperature and heat flux data are collected in real time by deploying temperature and heat flux sensors at multiple locations within a confined space. Gas velocity data at the ventilation opening is collected by a gas velocity sensor installed at the ventilation opening in the confined space. Based on the initial configuration information of the fuel in the confined space, initial fuel distribution data is obtained, which includes fuel type, location and initial fuel mass fraction.

3. The confined space fire flashover prediction method based on physical information neural network as described in claim 1, characterized in that, The specific process of inputting the spatiotemporal coordinate vector and the initial fuel distribution data into the Mamba neural network, and obtaining the predicted temperature, predicted incident radiation, and predicted fuel mass fraction through function mapping by the Mamba neural network is as follows: Using the serialization architecture of the Mamba neural network, the spatiotemporal coordinate vector and the initial fuel distribution data are taken as input, forward propagation calculation is performed, and the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction are output. The Mamba neural network adopts an architecture that combines a selective state-space model with a convolutional neural network. It learns the spatiotemporal dynamic correlation in the fire development process through sequential state-space modeling, and realizes a continuous differentiable mapping from spatiotemporal coordinates to physical fields. The parameters of the Mamba neural network are initialized using the Xavier initializer, and convergence is accelerated by a pre-training starter.

4. The confined space fire flashover prediction method based on physical information neural network as described in claim 1, characterized in that, The specific process of calculating the predicted temperature, predicted incident radiation, and predicted fuel mass fraction to obtain the corresponding time derivative and spatial derivative is as follows: Automatic differentiation is performed on the predicted temperature, the predicted incident radiation, and the predicted fuel mass fraction to obtain time derivatives and spatial derivatives. The time derivatives include the time derivatives of temperature, incident radiation, and fuel mass fraction. The spatial derivatives include spatial gradients and Laplace operators. The spatial gradients include the spatial gradients of temperature, incident radiation, and fuel mass fraction. The Laplace operators include the Laplace operator of temperature and the Laplace operator of incident radiation. The automatic differential calculation is performed during each forward inference process, and partial derivatives are calculated based on the chain rule. These partial derivatives are used to subsequently construct physical residuals, boundary residuals, and observation residuals.

5. The confined space fire flashover prediction method based on physical information neural network as described in claim 1, characterized in that, The specific process for constructing the loss function is as follows: The physical residual, boundary residual, and observation residual are combined in a weighted sum form to construct the loss function. ;in, Represents the parameters of the Mamba neural network; This represents the set of physical parameters to be learned; The expression for the loss function is shown in equation (4) below: in, The physical residual term represents the sum of squares and average of the residuals of the energy conservation equation, the radiation transfer equation, and the fuel consumption equation, and is calculated as shown in equation (5) below: This is the number of sampling points used to calculate the physical residual; The boundary residual term represents the sum of the squares of the boundary residuals of the energy equation, radiation equation, fuel consumption equation, and vent, and is calculated as shown in equation (6) below: This represents the number of boundary sampling points used to calculate the boundary residuals; The boundary residuals of the energy equation; The boundary residuals of the radiation equation; The boundary residuals of the fuel consumption equation; For the residual at the ventilation opening boundary; The observation residual term represents the sum of squares and average of the temperature observation residual and the thermal radiation flux observation residual, and is calculated as shown in equation (7) below: This refers to the number of temperature observation points; For temperature observation residuals; The number of observation points for thermal radiation flux; For the residual of thermal radiation flux observation; , and These are the weighting coefficients for the physical residual term, the boundary residual term, and the observation residual term, respectively, used to balance the contribution of the degree of violation of physical laws and the fitting error of the observation data to the loss function.

6. A confined space fire flashover prediction system based on a physical information neural network, used to implement the confined space fire flashover prediction method based on a physical information neural network as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire sensor observation data and initial fuel distribution data in a confined space. The sensor observation data includes time-series temperature data, heat flux data, and vent gas velocity data. The initial fuel distribution data includes fuel type, location, and initial fuel mass fraction. The neural network processing module is used to input the spatiotemporal coordinate vector and the initial fuel distribution data into the Mamba neural network, and to obtain the temperature prediction value, incident radiation prediction value and fuel mass fraction prediction value through the function mapping of the Mamba neural network. The Mamba neural network adopts an architecture that combines a selective state space model and a convolutional neural network, and learns the spatiotemporal dynamic correlation in the fire development process through sequential state space modeling. The derivative calculation module is used to calculate the predicted temperature, predicted incident radiation, and predicted fuel mass fraction, and obtain the corresponding time derivative and spatial derivative through automatic differentiation. The time derivative includes the temperature time derivative, the incident radiation time derivative, and the fuel mass fraction time derivative, and the spatial derivative includes the spatial gradient and the Laplace operator. The residual construction module is used to construct physical residual terms, boundary residual terms, and observation residual terms based on the time derivative and spatial derivative. The physical residual terms include residuals of the energy conservation equation, residuals of the radiation transfer equation, and residuals of the fuel consumption equation. The boundary residual terms include boundary residuals of the energy equation, boundary residuals of the radiation equation, boundary residuals of the fuel consumption equation, and boundary residuals of the vent. The observation residual terms include temperature observation residuals and thermal radiation flux observation residuals. The loss function construction module is used to combine the physical residual term, boundary residual term and observation residual term in a weighted sum form to construct a loss function, which is used to measure the degree of violation of physical laws and the fitting error of observation data. The optimization and update module is used to minimize the loss function through an optimization algorithm and update the parameters of the Mamba neural network and the physical parameters to be learned, wherein the physical parameters to be learned include the absorption coefficient, the Arrhenius reaction rate coefficient, and the activation energy parameter. The flashover prediction module is used to predict temperature field changes in a confined space based on an updated Mamba neural network. When the predicted temperature field value reaches the flashover critical temperature, the flashover time is determined.

7. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program and the processor runs the computer program to enable the electronic device to perform the confined space fire flashover prediction method based on a physical information neural network as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the confined space fire flashover prediction method based on a physical information neural network as described in any one of claims 1-5.

Images