Comprehensive pipe gallery fire positioning method based on modal recognition and Bayesian algorithm

Abstract

The invention discloses a comprehensive pipe gallery fire positioning method based on modal recognition and a Bayesian algorithm, and the method comprises the steps: firstly, determining a comprehensive pipe gallery disaster condition for a space scene, and building a comprehensive pipe gallery fire numerical calculation model based on a computational fluid mechanics method; secondly, a physical field modal reduced-order model is constructed based on time slice data, and nonlinear modal recognition of physical information such as temperature, carbon dioxide and carbon monoxide in the utility tunnel during a fire disaster is achieved; thirdly, performing nonlinear modal feature recognition on features of physical information such as temperature, carbon dioxide and carbon monoxide by using a neural network driven by physical information; and finally, based on a Bayesian reasoning algorithm method, carrying out fire position rapid positioning. According to the method, the machine learning theory is fused, and the accuracy, interpretability and timeliness of the model are improved. The method has an important practical application value for rapid prediction and disaster source positioning of the comprehensive pipe gallery fire, and especially can rapidly respond in an emergency, thereby providing support for disaster management.

Description

Technical Field

[0001] The invention relates to a comprehensive pipe gallery fire locating method based on modal recognition and Bayesian algorithm, belonging to the technical field of fire locating. Background Art

[0002] The utility tunnel is a modern urban infrastructure, which is mainly used to centrally house and protect various underground pipelines in the city, such as electricity, communications, gas, water supply and drainage systems. This design not only optimizes the use of urban space and reduces maintenance costs, but also significantly improves the efficiency and safety of urban operations. However, the underground utility tunnel has a narrow and closed space structure, and the internal pipeline layout is complex, with a high risk of fire. Once a fire occurs, the damage is great, among which the fire in the cable compartment is the most serious. Therefore, it is necessary to study the fire problem of the cable compartment in the utility tunnel.

[0003] At present, traditional methods or machine learning methods are mainly used for parameter identification and state construction of fire. Traditional methods have strong limitations and are difficult to deal with complex scenarios. Machine learning requires sufficient training samples, and at the same time has strong dependence on training data and poor generalization ability. In actual disaster scenarios, the sensor deployment area is limited, and it is impossible to obtain sufficient data samples. Moreover, models that rely solely on data-driven models may obtain theoretical results that violate the laws of physics, resulting in large errors in the analysis results of disaster scenarios that are difficult to observe directly. However, disaster data usually contain certain physical laws, empirical rules or other boundary information. The current machine learning methods do not analyze and utilize the laws contained in the limited amount of disaster information data itself and the physical models implied behind it. Therefore, it is necessary to design a new method to predict the disaster scene of the integrated pipeline corridor and quickly locate the source of the disaster. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a fire locating method for an integrated pipe gallery based on modal recognition and Bayesian algorithm, effectively predict the diffusion of the physical field during an integrated pipe gallery fire, invert the location of the fire starting point based on Bayesian theory and MCMC algorithm, and use sensor parameters to quickly locate the disaster source.

[0005] The present invention adopts the following technical solutions to solve the above technical problems:

[0006] A fire location method for an integrated pipe gallery based on modal recognition and Bayesian algorithm comprises the following steps:

[0007] Step 1: construct a CFD numerical model of the fire scene of the integrated pipe gallery, and solve the model to obtain the physical field information in the integrated pipe gallery that changes with time;

[0008] Step 2: construct time slice data according to the physical field information in the integrated pipe gallery that changes with time obtained in step 1, and extract the physical field modal reduction model from the time slice data using a nonlinear sparse recognition algorithm;

[0009] Step 3, constructing a physical information neural network model with the fire source location and the corresponding temperature information as input and the predicted temperature field information inside the integrated pipe gallery as output, obtaining a first type of initial training set, a second type of training set and a third type of training set according to the physical field modal reduction model, pre-training the physical information neural network model using the first type of initial training set to obtain a pre-trained model, and further training the pre-trained model using the second and third types of training sets to obtain a trained model;

[0010] Step 4: Use the trained model to reconstruct the temperature field in the integrated pipe gallery, and quickly locate the fire source based on the reconstructed temperature field and Bayesian inference algorithm.

[0011] As a preferred embodiment of the present invention, the specific process of step 1 is as follows:

[0012] According to the different fire source locations, fire source intensities and ventilation conditions inside the integrated pipe gallery, the computational fluid dynamics technology is used to construct a CFD numerical model of the integrated pipe gallery fire scene, and the model is solved based on the semi-implicit algorithm of the pressure coupling equations to obtain the physical field information inside the integrated pipe gallery that changes with time, including temperature, carbon monoxide concentration, carbon dioxide concentration and oxygen concentration data information when a fire occurs.

[0013] As a preferred solution of the present invention, in step 2, according to the physical field information in the integrated pipe gallery that changes with time obtained in step 1, the slicing time is set to 0.1s, and the time slice data of the fire source location, temperature, carbon monoxide concentration, carbon dioxide concentration and oxygen concentration in the physical field information are extracted every 0.1s;

[0014] The formula of the physical field modal reduction model is as follows:

[0015]

[0016] Where X is the state variable, Θ(X) is the feature library function matrix, Θ1(X), Θ2(X), Θ n (X) are all feature library functions, Ξ is the sparse regression coefficient matrix used to describe the physical field modal reduction model, ∈1(X), ∈2(X), ∈ n (X) are sparse regression coefficients.

[0017] As a preferred solution of the present invention, the specific process of step 3 is as follows:

[0018] Construct a physical information neural network model with the fire source location and corresponding temperature information as input and the predicted temperature field information inside the integrated pipe gallery as output;

[0019] A point is randomly selected at other locations of the utility corridor except the boundary area as the first sampling point, the position of the first sampling point and the corresponding physical field information are obtained as a piece of training data in the first type of initial training set, and the process of randomly selecting the first sampling point and obtaining the corresponding training data is repeated multiple times to generate the first type of initial training set;

[0020] A point is randomly selected in the boundary area of ​​the utility corridor as the second sampling point, the position of the second sampling point and the corresponding physical field information are obtained as a piece of training data in the second type of training set, and the process of randomly selecting the second sampling point and obtaining the corresponding training data is repeated multiple times to generate the second type of training set;

[0021] A first-class loss function is constructed according to the heat conduction equation and the boundary of the integrated pipe gallery, and the physical information neural network model is pre-trained using the first-class initial training set and the first-class loss function to obtain model parameters that minimize the first-class loss function as parameters of the pre-training model; wherein the heat conduction equation formula is as follows:

[0022]

[0023] In the formula, k is the thermal conductivity, T is the temperature, φ(x,y) is the heat source term, which represents the internal heat source or heat generated per unit volume, and (x,y) is the position coordinate;

[0024] The formula of the first type of loss function is as follows:

[0025]

[0026] In the formula, represents the first type of loss function, w pde 、w bc is a hyperparameter, L pde is the internal partial differential equation loss, L bc is the boundary condition loss, P pde is the first type of initial training set, T θ is the temperature field output by the model, θ is the model parameter, φ rated is the heat source intensity distribution function, P bc is the second type of training set, T0 is the vent location temperature or ambient temperature, and n is the boundary normal direction;

[0027] The second type of loss function is constructed according to the heat conduction equation and the boundary of the integrated pipe gallery. The pre-trained model is trained again using the second type of training set and the second type of loss function to obtain the model parameters that minimize the second type of loss function as the parameters of the trained model. The formula of the second type of loss function is as follows:

[0028]

[0029] In the formula, represents the second type of loss function, w data is a hyperparameter, L data is the loss term based on the temperature sensor monitoring value, T obd is the temperature sensor monitoring value, P data This is the third type of training set; the third type of training set consists of training data based on the temperature monitored by the temperature sensors deployed in the integrated pipe gallery and the location coordinates of the temperature sensors.

[0030] As a preferred solution of the present invention, the specific process of step 4 is as follows:

[0031] Step 4.1, the temperature monitored by the temperature sensor arranged in the integrated pipe gallery, the corresponding time and the location coordinates of the temperature sensor are used as the input of the Bayesian inference algorithm;

[0032] Step 4.2: Assuming that the fire source is located in the center of the utility corridor, the temperature field in the utility corridor is reconstructed based on the model trained in step 3.

[0033] Step 4.3, calculating the likelihood probability of the temperature monitored by each temperature sensor under the temperature field reconstructed in step 4.2;

[0034] Step 4.4, based on the likelihood probability and the temperature monitored by the temperature sensor, the fire source location is calculated using the Markov Chain Monte Carlo method;

[0035] Step 4.5, determine whether the fire source position calculated in step 4.4 is consistent with the fire source position set in step 4.2. If so, the fire source position calculated in step 4.4 is output as the final fire source location; otherwise, proceed to step 4.6;

[0036] Step 4.6, replace the fire source location set in step 4.2 with the fire source location calculated in step 4.4, and repeat steps 4.2 to 4.5.

[0037] A computer device comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, the steps of the integrated pipe gallery fire locating method based on modal recognition and Bayesian algorithm are implemented.

[0038] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the integrated pipe gallery fire locating method based on modal recognition and Bayesian algorithm are implemented.

[0039] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0040] 1. The present invention analyzes and utilizes the rules contained in a limited number of disaster information data and the physical models behind them, breaking through the limitations of traditional data-driven machine learning data modeling, breaking the situation of black box modeling of neural networks, and greatly improving the accuracy and interpretability of the model. By analyzing the models of physical information such as temperature, carbon dioxide and carbon monoxide and the corresponding characteristic parameters, the location information of the fire starting point in the tunnel is reversely predicted through Bayesian theory and MCMC algorithm, so as to achieve the purpose of rapid response in case of fire.

[0041] 2. The present invention fully considers the physical characteristics of the closed scene of the utility corridor, takes the known boundary conditions as the modal constraints, and solves the reduced-order modal model of the physical field. It is integrated with the machine learning method, which has important practical application value for the rapid prediction and disaster source location of the utility corridor fire, especially in emergency situations, and can respond quickly to provide support for disaster management. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of a fire location method for an integrated pipe gallery based on modal recognition and Bayesian algorithm of the present invention;

[0043] Figure 2 It is the CFD numerical model solution flow chart;

[0044] Figure 3 It is a flow chart of the physical field modal reduction model constructed based on time slice data;

[0045] Figure 4 It is a structural diagram of the physical information neural network model;

[0046] Figure 5 It is a flow chart of nonlinear modal feature recognition based on physical information neural network model;

[0047] Figure 6 It is a Bayesian inference flowchart;

[0048] Figure 7 It is a numerical simulation diagram of the fire in the integrated pipe gallery, where (a) is the temperature of the pipe gallery and (b) is the smoke in the pipe gallery;

[0049] Figure 8 is the accuracy of fire source tracing of each model;

[0050] Fig. 9 It is the deduction of the temperature field at different fire source locations in the comprehensive pipeline corridor. DETAILED DESCRIPTION

[0051] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0052] The fire location method of the integrated pipe gallery based on modal recognition and Bayesian algorithm proposed in this invention mainly consists of four parts. The specific ideas are as follows: Figure 1 shown.

[0053] 1. CFD numerical model of comprehensive pipeline corridor disaster scenario

[0054] Computational fluid dynamics (CFD) is the calculation and research of fluids. It has been gradually developed and improved on the basis of multiple disciplines such as fluid mechanics theory and computer science computing. The main research goal is to study the laws of fluid mechanics in a more scientific and efficient way, focusing on the discretization and solution analysis process of integral and differential terms in fluid control equations.

[0055] The fire model of the integrated pipe gallery in the present invention is mainly constructed using general CFD software for the construction of various scenes of the integrated pipe gallery. The calculation method adopts numerical methods to solve a set of low-speed flow Navier-Stokes equations. In the calculation process, the heat transfer process, smoke flow and fire spread in the fire are analyzed according to thermal drive. The software divides the space of the analysis object into multiple extremely small three-dimensional grids or calculation units, and solves the parameters of each unit respectively, including temperature, gas density, flow rate, pressure and component concentration. During the calculation, the partial differential equations of fluid mechanics such as conservation of mass, conservation of momentum and conservation of energy are used to approximate finite differences. The finite volume method is used to realize thermal radiation calculation and analyze the turbulence phenomenon in the fluid by dividing the grid, so as to track and predict the movement of the gas in the fire, and the growth and spread of the fire are deduced in combination with the parameter characteristics of various materials in the construction model. In this software, the calculation is performed by a semi-implicit algorithm based on the pressure coupling equation group. The algorithm is a pressure correction method. It mainly follows the idea of ​​"pressure prediction-pressure correction" to solve the pressure field in the discretized fluid unit grid, and further calculate and solve the momentum equation. The research problem involves the diffusion and transmission between different gas components, and the process of gas injection from the pipeline into the pipeline gallery space has certain turbulent characteristics. The corresponding solution process of this process is shown in the figure below: Figure 2 shown.

[0056] Through the above simulation, the changes in temperature, carbon monoxide, carbon dioxide and oxygen concentrations in various situations when a fire occurs in the utility corridor can be obtained, and the accuracy and effectiveness of the model used in the present invention are verified through a full-scale utility corridor fire test. The test process simulates many fire scenes caused by cable joint fires, wire short circuits, etc. in the cable compartment of the utility corridor, restoring the real scene of the utility corridor fire to the greatest extent.

[0057] 2. Physical field modal reduction model based on time slice data

[0058] Based on the above-mentioned model, this part first extracts the physical field information in the integrated pipeline corridor that changes with time, including information on temperature, carbon dioxide, carbon monoxide and other data, constructs time slice data, and then uses the nonlinear sparse recognition algorithm to extract the nonlinear physical field model from the time slice data, thereby achieving the result of modal reduction of the physical field.

[0059] Nonlinear sparse identification algorithm is an efficient method to extract dynamic reduced-order models from time-sliced ​​data (including spatial data, temporal data, and physical field data). Specifically, the behavior of a dynamic system can usually be formalized as a differential equation, where the rate of change of the system state over time is described by an unknown function f. A feature library containing a variety of functions (such as polynomials, trigonometric functions, etc.) that may affect the system dynamics is constructed from the system state data. This library is designed to cover all potential dynamic behaviors that the system may follow. It is assumed that the actual system dynamics can be accurately described by a few functions in the feature library. Therefore, the goal is to find a sparse vector that can best represent these key features while making the dynamic model constructed from the feature library as close as possible to the dynamics of the real system.

[0060] The model process is as follows Figure 3 As shown. It starts with data collection and preprocessing to ensure data quality and consistency. Next, a rich feature library is built based on this data, including polynomials, cross terms, and trigonometric functions of the original variables. Through sparse regression analysis, such as LASSO or sequential threshold least squares, the algorithm selects key features that describe the dynamics of the system from the feature library. These features are then used to build a dynamic model, usually a set of differential equations that describe how the physical field data changes over time and space. The model is then verified in the third step, and analysis is performed to explore the physical mechanisms that dominate the changes in physical field information. The core advantage of this algorithm is that it can reveal simplified dynamic laws from complex time-sliced ​​data, and it exhibits powerful analytical capabilities even in scenarios where traditional modeling methods are difficult to apply.

[0061] 3. Nonlinear modal feature recognition based on neural networks driven by physical information

[0062] At present, various fire solution models or calculation software require a large amount of experimental data to verify the solved model in actual use, but it is not possible to integrate various fidelity data well during the verification process. Especially when it comes to solving inverse problems, the boundary conditions and fluid parameters are unknown, and it is necessary to calculate the model parameters and reconstruct the flow field through the measurement data of limited positions. The current mainstream CFD calculation method has poor accuracy. In addition, the finite element method has high requirements on the quality of mesh division, and adjusting the mesh structure and even the mesh division itself is also very time-consuming.

[0063] In the context of multimodal problems, this method shows significant advantages and can handle complex phenomena in fields such as physics, engineering, and biology, which are usually described by a set of nonlinear partial differential equations and involve multiple interacting physical processes or multiple time and space scales. Physical information-driven neural networks integrate deep neural networks with physical laws, and can provide a general solution strategy for complex nonlinear partial differential equation problems without relying on prior assumptions or linear simplifications. This method breaks through the limitations of traditional computational methods and can not only solve problems that only rely on initial and boundary conditions, but also handle complex problems with internal structures by including potential physical laws. It can not only obtain data distribution laws through training samples like traditional neural networks, but also learn the physical laws behind the data. Compared with relying solely on data-driven, physical information-driven neural networks rely on physical information to constrain the solution process during training, and can learn the overall model laws more accurately.

[0064] The algorithm implementation of a neural network driven by physical information usually includes the following stages: Figure 4 As shown in the figure: First, one or more neural networks are defined to approximate the solution of one or more variables of the partial differential equation; then, in the optimization process, in addition to minimizing the difference between the network output and the real observation, the residual of the equation is further minimized by introducing a residual term derived from the physical equation. This process essentially ensures that the network output satisfies the physical law under consideration. Specifically, this residual term is generated based on the physical laws (e.g., momentum equation, mass conservation equation, etc.) and the derivative information obtained by automatic differentiation technology.

[0065] In actual disaster scenarios, the development process of different factors can mostly be described by multiple partial differential equations, and the analytical or numerical solutions of their functions contain a large amount of physical field information. Therefore, by using the efficient fitting ability of neural networks for nonlinear functions, a universal function solver for partial differential equations can be constructed, and complex multimodal problems can be transformed into problems for processing nonlinear partial differential equations without making prior assumptions or linear simplifications on the model. By using the automatic differentiation function, the neural network is constrained in combination with information such as the coordinates and parameters of the physical model, so as to obtain a neural network with physical information.

[0066] At the same time, as can be seen from the previous step, through the simulation of the integrated pipe gallery, we can obtain the changes in physical information such as temperature, carbon monoxide, carbon dioxide and oxygen concentration in various situations when a fire occurs, and then construct a data set (x, t, θ) with this physical information, where x is the position information of a point in space, t is the time information at that moment, and θ is the physical information of the point. For a given set of (x, t), the neural network will give the corresponding θ solution. The nonlinear modal feature recognition process of the neural network driven by physical information is as follows: Figure 5 shown.

[0067] The trained neural network model driven by physical information is used to reconstruct the temperature field and perform modal deduction in the integrated pipe gallery area. On the one hand, the model can reconstruct the spatial temperature field at any given time and deduce the temperature value at the coordinates of each sensor position at the relevant time; at the same time, the model can identify the fire source development mode and temperature diffusion model parameters based on the temperature field and fire source intensity in the previous period, thereby obtaining the fire development state in the pipe gallery space in the next time period and realizing the deduction of the fire.

[0068] 4. Rapidly locate the fire location based on Bayesian inference algorithm

[0069] Solving inverse problems is becoming increasingly important in various scenarios. Essentially, inverse problem solving is estimating the unknown parameters in a system of partial differential equations (PDEs) using a set of observations. In fire scenarios, the conditions for inverse problems are generally lacking or insufficient, as the solution to the problem may not exist or be unique. In addition, due to the relationship between the parameters and the observations, small perturbations in the data may lead to large changes in the solution. However, in practice, the observations are always noisy and may be limited in number or resolution. Therefore, quantifying the uncertainty introduced in the solution is often a challenge in inverse problems.

[0070] The Bayesian method is used here to solve these problems. The Bayesian algorithm is essentially a method that transforms prior information into posterior information by continuously updating information, and realizes probability deduction based on Bayes' theorem.

[0071]

[0072] Among them, x is the model parameter, y is the measured value, p(x|y) is the posterior probability density function of x, p(x) is the prior probability density function of x, which is generally derived from expert experience, subjective judgment, etc., and p(y|x) is the likelihood function, which indicates the degree of fit between the model parameters and the observed data. The specific process is as follows Figure 6 shown.

[0073] (1) Prior Probability

[0074] Prior information usually comes from experience or subjective judgment. For example, the location of the fire source in the fire identification of the integrated pipe corridor is usually directly selected from the length range of the pipe corridor. Prior information is often rough and must be expressed in the form of a probability density function before it can be used.

[0075] (2) Likelihood function

[0076] The likelihood function represents the degree of model parameters and is usually constructed as a normal distribution density function, an exponential distribution density function, and a Dirac function. Their expressions are as follows:

[0077] normal distribution:

[0078]

[0079] Laplace distribution:

[0080]

[0081] Dirac function:

[0082] p(y|x)=δ(yf(x))

[0083] Among them, σ 2 is the variance of the noise, and n is the number of measured data.

[0084] Since the observation error in fire is generally white noise, we assume that the observation noise ε i Subordinate to the normal distribution N(0,σ 2 ). The likelihood function is written as:

[0085]

[0086] According to the above formula, we can get:

[0087]

[0088] Where λ is the proportionality constant.

[0089] (3) Sampling of the Posterior Probability Density Function

[0090] Since the mechanism of fire occurrence is relatively complex, it is difficult to deduce the system model. The results of the posterior distribution generally have a high dimension and poor analysis effect. To solve the high-dimensional inversion problem, the Metropolis-Hastings (M-H) sampling method is selected here to sample the posterior probability density function. The M-H algorithm is the basis of the MCMC method, which relies on the Markov process to achieve the purpose of sampling from the probability distribution function and is the most widely used MCMC sampling method at present.

[0091] To establish a Markov chain with π(x) as the stationary distribution, an arbitrary non-transfer probability q(x) and a function a(x), 0 < a(x) < 1, are selected. For any combination (x, x * ), a transition kernel is formed: p(x, x * ) = q(x, x * )α(x, x * ), x ≠ x * . If the Markov chain is at position x at time t, that is, X(t) = x, then q(x) will first generate a potential transition state x → x * , and then judge whether to transfer according to the acceptance probability α(x, x * ). That is, after finding the potential transfer point x * , if u ≤ α(x, x * ), then accept α(x, x * ) as the state value of the Markov chain at the next moment; if u > α(x, x * ), reject the candidate sample and maintain the current state. Thus, after having the transfer point x * , a number can be randomly and uniformly generated from the interval [0, 1], then:

[0092]

[0093] q(x) is generally called the proposal distribution. To make the posterior distribution π(x) stationary, α needs to be selected so that the corresponding p(x|x * ) has π(x) as the equilibrium distribution:

[0094]

[0095] At this time,

[0096]

[0097] When the expected distribution satisfies q(x, x * ) = q(x * , x), α(x, x * ) can be simplified to:

[0098]

[0099] The specific sampling steps are generally as follows:

[0100] 1. Initialization: t = 0, generate a random parameter x0 for the current x t Assign a value and set the iteration termination step t = T.

[0101] 2. Let t = t + 1, from the conditional probability distribution θ(x|x t-1 ) Generate candidate samples x ′ .

[0102] 3. Calculate the acceptance probability α.

[0103] 4. Generate a random number α from a uniform distribution t .

[0104] 5. If α t ≤α, then accept the candidate sample, then there is x t =x ′ , otherwise reject the sample and let x t =x t-1 .

[0105] 6. Stop iteration when t = T, otherwise jump to step 2 to continue iteration.

[0106] (4) Positioning algorithm

[0107] On the basis of the above, using the fire development and physical information diffusion model and corresponding characteristic parameters extracted in the second step, the location of the fire starting point can be identified by identifying the maximum values ​​of parameters such as temperature, carbon dioxide and carbon monoxide.

[0108] Example

[0109] 1) CFD numerical model of comprehensive utility corridor disaster scenario

[0110] According to the fire protection zoning of the integrated pipe gallery and the needs of the present invention, a 1:1 three-dimensional model is established according to the size of the integrated pipe gallery. The model is 200m long, 3m wide and 3m high. The ambient temperature is set to a fixed value of 20°C, and the air pressure is set to standard atmospheric pressure. Taking the fire source position as 0, the fire source level as 2kw, and the ventilation condition as 1m / s as an example, the numerical simulation results are as follows: Figure 7 as shown in (a) and (b).

[0111] 2) Physical field modal reduction of time-sliced ​​data and physical information-driven nonlinear eigenmode identification

[0112] Slices are set in the above model, and time slice data of physical field information such as position, temperature, carbon dioxide concentration, carbon monoxide concentration and oxygen concentration in the model are extracted every 0.1s. A point is randomly selected as a sampling point at other locations of the integrated pipe corridor except the boundary area and its position coordinates are determined. The sampling point location and related fire parameter information are used as the first type of initial training data. Subsequently, the random sampling process is repeated many times to obtain multiple sets of training data. Finally, a large amount of training data constitutes the first type of initial training data set including position information; the second type of training set data selection is similar to the first type, only the sampling location is converted to the boundary area, and the second type of initial training data set can be obtained after repeated sampling.

[0113] According to the relevant heat conduction equations and boundary conditions satisfied by the internal area of ​​the utility tunnel, a first-class loss function based on physical drive is constructed.

[0114] The heat conduction differential equation is:

[0115]

[0116] The first type of loss function is:

[0117]

[0118] In the formula, w pde 、w bc is a hyperparameter, L pde is the internal partial differential equation loss, L bc is the boundary condition loss, P pde is the first type of initial training set, T θ is the temperature field output by the model, θ is the model parameter, φ rated is the heat source intensity distribution function, P bc is the second type of training set, T0 is the vent location temperature or ambient temperature, and n is the boundary normal direction.

[0119] According to the heat conduction equation and boundary conditions in the integrated pipe gallery space, a second type of loss function driven by physics and constrained by sensor data is constructed. During the training process, the fire source intensity and spatial temperature field are solved as unknown parameters, and the fire source intensity and other information preliminarily judged after data fusion are used as initial values. The selected training data is used to minimize the loss function, thereby optimizing the model to fit the corresponding relationship between the measured temperature value at the location of the corresponding sensor and the temperature field of the pipe gallery space area, and finally obtain a model that conforms to the physical laws of pipe gallery fire development.

[0120] Based on the heat conduction differential equation and boundary conditions in the tunnel, the second loss function constrained by physical drive and sensor measurement values ​​is set as:

[0121]

[0122] In the formula, w data is a hyperparameter, L data is the loss term based on the temperature sensor monitoring value, T obs is the temperature sensor monitoring value, P data It is the third type of training set.

[0123] According to the monitoring data of the temperature sensor, the trained model is used to reconstruct the temperature field and perform modal deduction in the integrated pipe gallery area. On the one hand, the model can reconstruct the spatial temperature field according to any given time, and deduce the temperature value at the coordinates of each sensor position at the relevant time; at the same time, the model can identify the fire source development mode and temperature diffusion model parameters based on the temperature field and fire source intensity in the previous period, so as to obtain the fire development state in the pipe gallery space in the next time period and realize the deduction of the fire.

[0124] 3) Rapid location of fire based on Bayesian inference algorithm

[0125] Based on temperature, the fire temperature characteristic parameters identified by the Bayesian fire estimation model were analyzed, including the mean value and standard deviation during the sampling process, as shown in Table 1.

[0126] Table 1 Results of rapid fire location identification based on Bayesian inference algorithm

[0127]

[0128] According to the algorithm results, the improved Bayesian inference algorithm can accurately estimate the fire state parameters based on the limited temperature monitoring values. Since the temperature value above the fire source location rises the most compared with other areas in the fire scene, the fire source location can be easily identified. When the fire source is located between two sensors, the corresponding fire source location can also be identified based on the current air flow state and the association between different sensors. The fire source location identified by the algorithm is almost the same as the actual location of the simulated fire source. In addition, the temperature of the pipe gallery is basically stable at 180s, and the temperature rise is not obvious, which is basically consistent with the temperature distribution in the fitting, proving the accuracy of the improved Bayesian inference algorithm for fire identification.

[0129] Taking five typical working conditions sampling data as an example, the accuracy of tracing the fire source location is analyzed. Figure 8 As shown in the figure, the improved Bayesian inference algorithm has higher accuracy and deduction efficiency. After the improvement, all sampling processes can converge within 100 generations. The fire location fast positioning algorithm based on the Bayesian inference algorithm can accurately trace the fire source for the integrated pipe gallery fire.

[0130] The spatial temperature distribution in the integrated pipe gallery can also be calculated using the fire characteristic parameters estimated by the improved Bayesian inference algorithm. Fig. 9 As shown in the figure, when a fire occurs, the fire source is concentrated near the origin. As the fire continues to develop, the fire spreads, the spatial temperature continues to rise, and the fire impact range is gradually expanded. When a fire occurs at different locations in the tunnel, the improved Bayesian algorithm can quickly locate the fire location and realize the fire source tracing in the tunnel scene based on relevant factors such as the temperature field and ambient wind. The algorithm deduction results are similar to the numerical simulation results, and the temperature field is basically the same.

[0131] Based on the same inventive concept, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the aforementioned integrated pipe gallery fire location method based on modal recognition and Bayesian algorithm are implemented.

[0132] Based on the same inventive concept, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the aforementioned integrated pipe gallery fire location method based on modal recognition and Bayesian algorithm.

[0133] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0134] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0135] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0136] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0137] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A fire location method for a comprehensive pipe gallery based on modal recognition and Bayesian algorithm, characterized in that: The steps include: Step 1: construct a CFD numerical model of the fire scene of the integrated pipe gallery, and solve the model to obtain the physical field information in the integrated pipe gallery that changes with time; Step 2: construct time slice data according to the physical field information in the integrated pipe gallery that changes with time obtained in step 1, and extract the physical field modal reduction model from the time slice data using a nonlinear sparse recognition algorithm; Step 3, constructing a physical information neural network model with the fire source location and the corresponding temperature information as input and the predicted temperature field information inside the integrated pipe gallery as output, obtaining a first type of initial training set, a second type of training set and a third type of training set according to the physical field modal reduction model, pre-training the physical information neural network model using the first type of initial training set to obtain a pre-trained model, and further training the pre-trained model using the second and third types of training sets to obtain a trained model; Step 4: Use the trained model to reconstruct the temperature field in the integrated pipe gallery, and quickly locate the fire source based on the reconstructed temperature field and Bayesian inference algorithm.

2. The method for locating fire in a utility tunnel based on modal recognition and Bayesian algorithm according to claim 1 is characterized in that: The specific process of step 1 is as follows: According to the different fire source locations, fire source intensities and ventilation conditions inside the integrated pipe gallery, the computational fluid dynamics technology is used to construct a CFD numerical model of the integrated pipe gallery fire scene, and the model is solved based on the semi-implicit algorithm of the pressure coupling equations to obtain the physical field information inside the integrated pipe gallery that changes with time, including temperature, carbon monoxide concentration, carbon dioxide concentration and oxygen concentration data information when a fire occurs.

3. The method for locating fire in a utility tunnel based on modal recognition and Bayesian algorithm according to claim 1 is characterized in that: In step 2, according to the physical field information in the integrated pipe gallery that changes with time obtained in step 1, the slicing time is set to 0.1s, and the time slice data of the fire source location, temperature, carbon monoxide concentration, carbon dioxide concentration and oxygen concentration in the physical field information are extracted every 0.1s; The formula of the physical field modal reduction model is as follows: Where X is the state variable, Θ(X) is the feature library function matrix, Θ1(X), Θ2(X), Θ n (X) are all feature library functions, Ξ is the sparse regression coefficient matrix used to describe the physical field modal reduction model, ∈1(X), ∈2(X), ∈ n (X) are sparse regression coefficients.

4. The method for locating fire in a utility tunnel based on modal recognition and Bayesian algorithm according to claim 3 is characterized in that: The specific process of step 3 is as follows: Construct a physical information neural network model with the fire source location and corresponding temperature information as input and the predicted temperature field information inside the integrated pipe gallery as output; A point is randomly selected at other locations of the utility corridor except the boundary area as the first sampling point, the position of the first sampling point and the corresponding physical field information are obtained as a piece of training data in the first type of initial training set, and the process of randomly selecting the first sampling point and obtaining the corresponding training data is repeated multiple times to generate the first type of initial training set; A point is randomly selected in the boundary area of ​​the utility corridor as the second sampling point, the position of the second sampling point and the corresponding physical field information are obtained as a piece of training data in the second type of training set, and the process of randomly selecting the second sampling point and obtaining the corresponding training data is repeated multiple times to generate the second type of training set; Constructing a first-class loss function according to the heat conduction equation and the boundary of the integrated pipe gallery, pre-training the physical information neural network model using the first-class initial training set and the first-class loss function, and obtaining model parameters that minimize the first-class loss function as parameters of the pre-training model; The heat conduction equation is as follows: In the formula, k is the thermal conductivity, T is the temperature, φ(x,y) is the heat source term, which represents the internal heat source or heat generated per unit volume, and (x,y) is the position coordinate; The formula of the first type of loss function is as follows: In the formula, represents the first type of loss function, w pde 、w bc is a hyperparameter, L pde is the internal partial differential equation loss, L bc is the boundary condition loss, P pde is the first type of initial training set, T θ is the temperature field output by the model, θ is the model parameter, φ rated is the heat source intensity distribution function, P bc is the second type of training set, T0 is the vent location temperature or ambient temperature, and n is the boundary normal direction; The second type of loss function is constructed according to the heat conduction equation and the boundary of the integrated pipe gallery. The pre-trained model is trained again using the second type of training set and the second type of loss function to obtain the model parameters that minimize the second type of loss function as the parameters of the trained model. The formula of the second type of loss function is as follows: In the formula, represents the second type of loss function, w data is a hyperparameter, L data is the loss term based on the temperature sensor monitoring value, T obs is the temperature sensor monitoring value, P data This is the third type of training set; the third type of training set consists of training data based on the temperature monitored by the temperature sensors deployed in the integrated pipe gallery and the location coordinates of the temperature sensors.

5. The method for locating fire in a utility tunnel based on modal recognition and Bayesian algorithm according to claim 1 is characterized in that: The specific process of step 4 is as follows: Step 4.1, the temperature monitored by the temperature sensor arranged in the integrated pipe gallery, the corresponding time and the location coordinates of the temperature sensor are used as the input of the Bayesian inference algorithm; Step 4.2: Assuming that the fire source is located in the center of the utility corridor, the temperature field in the utility corridor is reconstructed based on the model trained in step 3. Step 4.3, calculating the likelihood probability of the temperature monitored by each temperature sensor under the temperature field reconstructed in step 4.2; Step 4.4, based on the likelihood probability and the temperature monitored by the temperature sensor, the fire source location is calculated using the Markov Chain Monte Carlo method; Step 4.5, determine whether the fire source position calculated in step 4.4 is consistent with the fire source position set in step 4.

2. If so, the fire source position calculated in step 4.4 is output as the final fire source location; otherwise, proceed to step 4.6; Step 4.6, replace the fire source location set in step 4.2 with the fire source location calculated in step 4.4, and repeat steps 4.2 to 4.

5.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: When the processor executes the computer program, the steps of the integrated pipe gallery fire location method based on modal recognition and Bayesian algorithm as described in any one of claims 1 to 5 are implemented.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the integrated pipe gallery fire location method based on modal recognition and Bayesian algorithm as described in any one of claims 1 to 5 are implemented.

Images