Intelligent fire-fighting big data analysis method based on cloud service

Abstract

The invention relates to an intelligent fire-fighting big data analysis method based on cloud service, and the method comprises the following steps: obtaining multi-source gas concentration time sequence data collected by a gas sensor array at a fire scene, and obtaining a smooth gas concentration time sequence curve; analyzing the concentration change of different gases according to the gas concentration time sequence curve; establishing a coupling relation matrix of the environmental parameters and the chemical reaction chain; through the coupling relation matrix, the diffusion velocity and mixing degree of different gases under different environmental conditions are calculated; identifying typical gas concentration change modes caused by different types of dangerous sources, and generating preliminary estimation results of the types and positions of the dangerous sources; analyzing the type and position of a hazard source of the current fire scene, and generating real-time positioning information of the hazard source; and adjusting the coupling coefficient in real time, and generating a dynamically adjusted dangerous source identification result. According to the invention, through fusion of multi-source data, the chemical reaction chain model and the machine learning algorithm, accurate, real-time and dynamic identification of the hazard source of the fire scene is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a cloud service-based smart firefighting big data analysis method. Background Art

[0002] There is a key technical problem in the study of the component characteristics of gas sensor arrays in smart firefighting. At the fire scene, the concentration of multiple gases will show a complex change pattern over time, especially when the concentration of a certain type of gas fluctuates violently, how to accurately identify the source of danger becomes a major challenge.

[0003] The dramatic fluctuations in gas concentration may be due to a variety of reasons, such as changes in the combustion materials, the spread of fire, the occurrence of chemical reactions, etc. It is difficult to accurately determine the type and location of the hazard source by simply relying on the changing trend of gas concentration. At this time, how to combine the temporal change law of gas concentration with the chemical reaction chain becomes the key to the problem. Different combustion materials will produce different gas components during the combustion process, and there may be complex chemical reactions between these gas components. If the characteristics of the chemical reaction chain can be deeply analyzed and matched with the changing law of gas concentration, it is possible to accurately identify the hazard source. However, the analysis of the chemical reaction chain needs to consider many factors, such as temperature, pressure, catalysts, etc., which further increases the complexity of the problem. How to process and analyze sensor data in real time in a complex environment and generate reliable firefighting tactics reports in a timely manner is also a major technical challenge. This requires comprehensive consideration of the timeliness, accuracy and interpretability of the data, and combining it with the experience and knowledge of fire experts to provide strong support for firefighting decisions. Summary of the invention

[0004] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a cloud-based intelligent firefighting big data analysis method. The cloud-based intelligent firefighting big data analysis method realizes accurate, real-time and dynamic identification of fire scene hazards by integrating multi-source data, chemical reaction chain models and machine learning algorithms.

[0005] The cloud service-based intelligent firefighting big data analysis method of the present invention comprises the following steps:

[0006] S1. Obtain multi-source gas concentration time series data collected by the gas sensor array at the fire scene, and pre-process the concentration change trends of different gases respectively to remove noise interference and obtain a smooth gas concentration time series curve;

[0007] S2. Analyze the concentration changes of different gases according to the gas concentration time series curve, and combine with the pre-established chemical reaction chain model, and if there is a chemical reaction of the target gas, generate a mapping relationship of the chemical reaction chain related to the target gas;

[0008] S3. Obtain environmental parameters at the fire scene, combine the mapping relationship of the chemical reaction chain, establish a coupling relationship matrix between environmental parameters and chemical reaction chains, and determine the impact of environmental parameters on gas concentration fluctuations;

[0009] S4. Calculate the diffusion speed and mixing degree of different gases under different environmental conditions through the coupling relationship matrix, generate a predicted value of the target gas concentration fluctuation, and analyze the degree of deviation between the predicted value of the target gas concentration fluctuation and the actual measured value;

[0010] S5. If the target gas is related to the location and type of the hazard source, the relationship between the gas concentration fluctuation and the hazard source is extracted from the historical fire data, the typical gas concentration change patterns caused by different types of hazard sources are identified, the classification model is trained, and the preliminary estimation results of the hazard source type and location are generated;

[0011] S6. Analyze the type and location of the hazard source at the current fire scene based on the preliminary estimation result and the real-time collected gas concentration time series data, and generate real-time location information of the hazard source;

[0012] S7. Update the predicted value of gas concentration fluctuation through the real-time positioning information, adjust the coupling coefficient between environmental parameters and chemical reaction chain in real time in combination with the changing trend of environmental factors, and generate dynamically adjusted hazard source identification results.

[0013] Preferably, the step S1 specifically includes:

[0014] Acquire various gas concentration time series data from the gas sensor array, and perform normalization processing on the concentration time series data according to the sensitivity coefficients of various gas sensors to obtain first gas concentration data;

[0015] The first gas concentration data is smoothed by using a sliding mean filter, and according to a preset noise interference amplitude threshold, the data points exceeding the preset noise interference amplitude threshold are subjected to wavelet transform denoising to obtain second gas concentration data;

[0016] For the second gas concentration data, a cubic spline interpolation function is used to fit discrete data points in the time series segment to obtain third gas concentration data;

[0017] For the third gas concentration data, the first-order concentration change gradient value is calculated according to the concentration threshold value set according to the physical characteristics of each gas, and the gradient value is smoothed using a Gaussian kernel function to obtain a smoothed concentration time series curve of each gas.

[0018] Preferably, the step S2 specifically includes:

[0019] Acquire target gas concentration data according to the gas concentration time series curve, calculate similarity values ​​between target gas molecules and other gas molecules through molecular structure feature vectors stored in a chemical reaction chain database, and obtain candidate chemical reaction chain data;

[0020] For the candidate chemical reaction chain data, a pre-established gas phase reaction rate constant database is used to calculate the reaction rate constant value of each elementary reaction step in the reaction chain within a temperature range to obtain reaction chain kinetic data;

[0021] Calculate the Pearson correlation coefficient between the reactant concentration change value and the measured gas concentration time series curve within the temperature range according to the reaction chain kinetic data, and group the reaction chains whose correlation coefficient is greater than the coefficient threshold by using a hierarchical clustering algorithm to obtain reaction chain association data;

[0022] For the reaction chain associated data, the standard formation enthalpy change and standard entropy change values ​​of reactants and products are obtained from the thermodynamic database, the chemical reaction equilibrium constant of the reaction chain under temperature and pressure conditions is calculated, and a reaction chain mapping model is established through a support vector regression method.

[0023] Preferably, the step S3 specifically includes:

[0024] The effects of environmental parameters on gas concentration fluctuations include gas reaction rate, gas diffusion rate, and gas mixing degree;

[0025] A sensor network is used to obtain temperature data, humidity data and pressure data at the fire scene, and a temperature distribution field, a humidity distribution field and a pressure distribution field are obtained by radial basis function calculation according to the temperature data, the humidity data and the pressure data;

[0026] For the temperature distribution field, the humidity distribution field and the pressure distribution field, the activation energy value and the pre-exponential factor value of the reactants are extracted through a chemical reaction chain mapping relationship, and the reaction rate influence coefficient is obtained according to the activation energy value and the pre-exponential factor value combined with the Arrhenius equation;

[0027] According to the reaction rate influence coefficient, the gas diffusion flux value is calculated using Fick's law, and a deep neural network model is established through the gas diffusion flux value to obtain gas diffusion velocity distribution data;

[0028] The Shannon entropy value is calculated for the gas diffusion velocity distribution data through the gas concentration time series curve, and a multivariate autoregressive model is established according to the Shannon entropy value to obtain the environmental parameter influence coefficient matrix.

[0029] Preferably, the step S4 specifically includes:

[0030] According to the coupling relationship matrix, the values ​​of the temperature field, the pressure field and the humidity field at the spatial grid points are obtained, and the concentration gradient data of the target gas in the three-dimensional space is obtained by the numerical calculation;

[0031] The concentration gradient data is used to calculate the mixed entropy value of the target gas and the ambient gas, and a Bayesian network including three nodes of temperature, pressure and humidity is established through the mixed entropy value to obtain gas concentration prediction data;

[0032] Dividing the gas concentration prediction data into a plurality of time windows, using the time windows to calculate deviation feature vectors between the prediction value and the measured value, and grouping the deviation feature vectors by a hierarchical clustering algorithm to obtain deviation feature data;

[0033] Whether the standard deviation exceeds a preset upper limit of the standard deviation threshold is determined based on the deviation characteristic data. If it exceeds the preset upper limit of the standard deviation threshold, Fourier transform is used to extract the spectral characteristics of the deviation characteristic data, and the target gas concentration change mode is determined by the spectral characteristics.

[0034] Preferably, the step S5 specifically includes:

[0035] Acquire gas concentration time series data according to the three-dimensional spatial coordinates of the hazard source, and extract low-frequency approximate coefficients and high-frequency detail coefficients of the time series data at several decomposition scales using wavelet transform to obtain first characteristic data;

[0036] For the first feature data, a random forest classifier is used to establish a mapping relationship between the first feature data and the type of hazard source, and classifier parameters are optimized through cross-validation to obtain second classification data;

[0037] Calculating the Euclidean distance matrix between the gas concentration fluctuation characteristics of different types of hazardous sources according to the second classification data, grouping the hazardous sources using a spectral clustering algorithm, extracting a typical fluctuation template for each group, and obtaining third template data;

[0038] If the Pearson correlation coefficient between the measured gas concentration curve and the third template data is greater than the preset Pearson correlation coefficient determination threshold, the naive Bayes method is used to calculate the posterior probability of the location of the hazardous source to determine the spatial location area of ​​the hazardous source and the type of hazardous substance.

[0039] Preferably, the step S6 specifically includes:

[0040] The bicubic interpolation algorithm is used to spatially interpolate the concentration values ​​between the gas concentration sampling points to obtain continuous concentration distribution field data;

[0041] The continuous concentration distribution field data is filtered using a Gaussian kernel function to obtain first filtered field data;

[0042] Divide the concentric circle search area centered on the initial position of the hazard source according to the first filter field data, and use a particle swarm algorithm to iteratively search the concentration gradient in the search area to obtain second gradient field data;

[0043] The statistical moment of gas concentration in the search area is calculated for the second gradient field data, and a hidden Markov chain is used to estimate the state transition of the position coordinates and type identification of the hazard source, and the spatial coordinates of the hazard source are determined.

[0044] Preferably, the step S7 specifically includes:

[0045] Real-time adjustment of the coupling coefficient between environmental parameters and chemical reaction chains includes adjustment of reaction rate constants;

[0046] Generate dynamically adjusted hazard source identification results including hazard source type, location and hazard level information;

[0047] Constructing an environmental change trend vector based on the actual measured values ​​of environmental parameters collected by the temperature sensor, the pressure sensor and the humidity sensor, and obtaining a predicted value of future environmental parameter changes through the environmental change trend vector;

[0048] According to the predicted value of the future environmental parameter change, a reaction rate constant is calculated using a temperature correction term and a pressure correction term, and a Kalman filter is established through the reaction rate constant and a recursive least square method to obtain a predicted value of the gas concentration;

[0049] For the predicted value of gas concentration, the gradient vector and Hessian matrix of the concentration field are calculated, and the coordinates of the location of the hazard source are obtained by support vector regression using a kernel function as a radial basis;

[0050] For the location coordinates of the hazardous source, Bayesian probability estimation is used to update the posterior distribution of the hazardous source type, and a hazard index vector is constructed according to the gas concentration exceeding the standard multiple, the spreading speed and the diffusion range to obtain the hazard level.

[0051] The cloud service-based intelligent firefighting big data analysis method described in the present invention has the following advantages:

[0052] The cloud service-based intelligent firefighting big data analysis method of the present invention obtains multi-source gas concentration time series data at the fire scene and performs preprocessing to remove noise interference and obtain a smooth gas concentration time series curve, thereby improving the accuracy of the data; combined with the chemical reaction chain model and environmental parameters, the diffusion speed and mixing degree of different gases under different environmental conditions can be calculated to generate a predicted value of the target gas concentration fluctuation, which helps firefighters to understand the gas concentration change at the fire scene in advance and make more accurate judgments; by analyzing the relationship between gas concentration fluctuations and hazardous sources, typical gas concentration change patterns caused by different types of hazardous sources can be identified, and a classification model can be trained to generate preliminary estimation results of the type and location of the hazardous source, which greatly shortens the identification time of the hazardous source; according to the preliminary estimation results and the real-time collected gas concentration time series data, the type and location of the hazardous source at the current fire scene can be further analyzed, and real-time positioning information of the hazardous source can be generated, which provides strong support for quickly locating the fire scene and eliminating the hazardous source; by integrating advanced technologies such as the Internet of Things, big data, and cloud computing, comprehensive perception and intelligent analysis of the fire scene are realized, which helps to promote the intelligent development of firefighting work and improve the efficiency and accuracy of firefighting work. The present invention realizes accurate, real-time and dynamic identification of dangerous sources at fire scenes by integrating multi-source data, chemical reaction chain models and machine learning algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a smart firefighting big data analysis method based on cloud services described in the present invention. DETAILED DESCRIPTION

[0054] like Figure 1 As shown, the cloud service-based intelligent firefighting big data analysis method described in the present invention includes the following steps:

[0055] S1. Obtain multi-source gas concentration time series data collected by the gas sensor array at the fire scene, and pre-process the concentration change trends of different gases respectively to remove noise interference and obtain a smooth gas concentration time series curve;

[0056] S2. Analyze the concentration changes of different gases according to the gas concentration time series curve, and combine with the pre-established chemical reaction chain model. If there is a chemical reaction of the target gas, generate a mapping relationship of the chemical reaction chain related to the target gas;

[0057] S3. Obtain environmental parameters at the fire scene, combine the mapping relationship of the chemical reaction chain, establish a coupling relationship matrix between environmental parameters and chemical reaction chains, and determine the impact of environmental parameters on gas concentration fluctuations;

[0058] S4. Calculate the diffusion speed and mixing degree of different gases under different environmental conditions through the coupling relationship matrix, generate the predicted value of the target gas concentration fluctuation, and analyze the deviation degree between the predicted value of the target gas concentration fluctuation and the actual measured value;

[0059] S5. If the target gas is related to the location and type of the hazard source, the relationship between the gas concentration fluctuation and the hazard source is extracted from the historical fire data, the typical gas concentration change patterns caused by different types of hazard sources are identified, the classification model is trained, and the preliminary estimation results of the hazard source type and location are generated;

[0060] S6. Based on the preliminary estimation results and in combination with the real-time collected gas concentration time series data, the types and locations of the dangerous sources at the current fire scene are analyzed to generate real-time location information of the dangerous sources;

[0061] S7. Update the predicted value of gas concentration fluctuation through real-time positioning information, adjust the coupling coefficient of environmental parameters and chemical reaction chain in real time according to the changing trend of environmental factors, and generate dynamically adjusted hazard source identification results.

[0062] Furthermore, in this embodiment, step S1 specifically includes:

[0063] Acquire various gas concentration time series data from the gas sensor array, and perform normalization processing on the concentration time series data according to the sensitivity coefficients of various gas sensors to obtain first gas concentration data;

[0064] The first gas concentration data is smoothed by a sliding mean filter, and according to a preset noise interference amplitude threshold, the data points exceeding the preset noise interference amplitude threshold are subjected to wavelet transform denoising to obtain the second gas concentration data;

[0065] For the second gas concentration data, a cubic spline interpolation function is used to fit the discrete data points in the time series segment to obtain the third gas concentration data;

[0066] For the third gas concentration data, the first-order concentration change gradient value is calculated according to the concentration threshold value set according to the physical characteristics of each gas, and the gradient value is smoothed by using the Gaussian kernel function to obtain the smoothed concentration time series curve of each gas;

[0067] Specifically, the concentration data of carbon monoxide, carbon dioxide, ammonia and nitrogen oxides are collected from the gas sensor array, and the gas concentration values ​​of multiple consecutive time points are obtained for each sensor position at a sampling frequency of ten times per second, and then the collected concentration data are normalized according to the sensitivity coefficients of various gas sensors to obtain the first gas concentration data;

[0068] The first gas concentration data is initially smoothed using a sliding mean filter with a window length of five data points. When the concentration change rate between adjacent data points exceeds the preset noise interference amplitude threshold of each type of gas, the fourth-order wavelet transform is used to denoise the abnormal data points to obtain the second gas concentration data;

[0069] The second gas concentration data is divided into a time series segment every ten seconds according to the data collection period, and the discrete data points in each time series segment are fitted using a cubic spline interpolation function, which uses natural boundary conditions to obtain the third gas concentration data;

[0070] For the third gas concentration data, the concentration threshold is set according to the physical characteristics of each gas to divide it into segments, and the first-order concentration change gradient value is calculated within each time series step;

[0071] For the first-order concentration change gradient value, a Gaussian kernel function with a standard deviation of 0.5 is used to perform final smoothing on the gradient curve to obtain the smoothed concentration time series curves of various gases;

[0072] Here is an example:

[0073] In the gas monitoring at the fire scene, the concentration change trend of various harmful gases such as carbon monoxide, carbon dioxide, ammonia, nitrogen oxides, etc. has an important impact on the safety protection of firefighters;

[0074] The gas sensor array is arranged at different locations at the fire scene, collecting 10 data points per second in a high temperature environment. Due to the interference of temperature and humidity, the raw gas concentration data collected by the sensor often fluctuates and has noise;

[0075] For the carbon monoxide gas sensor, its sensitivity coefficient is 2.5. When the measured carbon monoxide concentration is 500ppm, the normalized value is 0.5. Other gases are also normalized according to their respective sensitivity coefficients.

[0076] In the sliding mean filtering process, the normalized gas concentration data is slid with a window length of 5 data points, and the data fluctuation is smoothed by calculating the average value of the data in the window. For example, the original values ​​of the carbon monoxide concentration at 5 consecutive data points are 0.48, 0.52, 0.49, 0.51, and 0.50, respectively. The value obtained after sliding average is 0.50;

[0077] When the concentration change of a data point exceeds 0.1, the fourth-order wavelet transform is used to denoise the abnormal point, thereby eliminating the data anomaly caused by temperature fluctuations;

[0078] In the data segmentation processing, every 10 seconds of data is regarded as a time series segment, and the cubic spline interpolation function is used to fit the 100 discrete data points in this time period;

[0079] The interpolation function uses natural boundary conditions at the boundary of the sequence segment, that is, the second-order derivative is zero, which ensures the smooth transition of the curve at the boundary;

[0080] For example, in the time series segment from 0 to 10 seconds, the carbon monoxide concentration rises from 0.2 to 0.8, and the continuous concentration change curve during this period is obtained by cubic spline interpolation;

[0081] In the gas concentration segmentation threshold setting, the thresholds are set according to the physical properties of gases such as carbon monoxide and carbon dioxide. The concentration threshold of carbon monoxide is set to 0.6, and the concentration threshold of carbon dioxide is set to 0.8. When calculating the first-order gradient, the concentration change rate between two adjacent data points reflects the speed of gas diffusion;

[0082] A Gaussian kernel function with a standard deviation of 0.5 was used to smooth the gradient curve to obtain a smoother concentration change trend curve, eliminating the slight fluctuations introduced during data acquisition and processing. After normalization, denoising, interpolation and smoothing, the original gas concentration data was obtained to obtain a time series curve that can accurately reflect the change rules of various gas concentrations at the fire scene, providing reliable data support for safety assessment at the fire scene.

[0083] Furthermore, in this embodiment, step S2 specifically includes:

[0084] The target gas concentration data is obtained according to the gas concentration time series curve, and the similarity value between the target gas molecule and other gas molecules is calculated by the molecular structure feature vector stored in the chemical reaction chain database to obtain the candidate chemical reaction chain data;

[0085] For the candidate chemical reaction chain data, the reaction rate constant value of each elementary reaction step in the reaction chain within the temperature range is calculated using the pre-established gas phase reaction rate constant database to obtain the reaction chain kinetic data;

[0086] The Pearson correlation coefficient between the reactant concentration change value and the measured gas concentration time series curve within the temperature range is calculated based on the reaction chain kinetic data, and the reaction chains with correlation coefficients greater than the coefficient threshold are grouped using a hierarchical clustering algorithm to obtain the reaction chain association data;

[0087] For the reaction chain related data, the standard formation enthalpy change and standard entropy change of reactants and products are obtained from the thermodynamic database, the chemical reaction equilibrium constant of the reaction chain under temperature and pressure conditions is calculated, and the reaction chain mapping model is established by the support vector regression method;

[0088] Specifically, the concentration data of the target gas including carbon monoxide, carbon dioxide, ammonia and nitrogen oxides are extracted from the gas concentration time series curve, and the similarity between the target gas molecules and other gas molecules is calculated through the molecular structure feature vectors stored in the chemical reaction chain database to form multiple candidate chemical reaction paths in which the target gas participates as a reactant or a product, and the reaction chains are sorted according to the reactant conversion rate to obtain the first reaction chain structure data;

[0089] The first reaction chain structure data is screened according to the activation energy required for breaking the molecular bonds of the reactants in each reaction chain, and the reaction rate constants of each elementary reaction step in each reaction chain in the temperature range of 298 to 1500 Kelvin are calculated using a pre-established gas phase reaction rate constant database, and the reaction rate values ​​are calculated in combination with the reactant concentrations to generate the second reaction chain kinetic data;

[0090] For the second reaction chain kinetic data, based on the change value of reactant concentration in the fire temperature range of 298 to 1500 Kelvin and the measured gas concentration time series curve, the Pearson correlation coefficient was used to calculate the correlation, and the reaction chains with correlation coefficient greater than 0.8 were grouped using the hierarchical clustering algorithm to obtain the third reaction chain correlation data;

[0091] For the third reaction chain related data, the standard formation enthalpy change and standard entropy change of reactants and products are obtained from the thermodynamic database, and the chemical reaction equilibrium constant of each reaction chain under the current temperature and pressure conditions is calculated;

[0092] The concentration of the product under the theoretical equilibrium state is calculated according to the reaction equilibrium constant, and the degree of agreement between the theoretical concentration and the measured concentration data is evaluated by the root mean square error. The mapping relationship between the reaction chain and the target gas is established by the support vector regression method with the kernel function as the radial basis function.

[0093] At the fire scene, chemical reactions between multiple gases form a complex reaction network. Identifying key reaction chains by analyzing the trend of gas concentration changes is of great significance for the development of fire. The chemical reaction chain database stores the gas molecular structure feature vectors.

[0094] Here is an example:

[0095] Taking the carbon monoxide molecule as an example, its characteristic vector contains information such as the carbon-oxygen single bond length of 124 picometers and the molecular polar moment of 0.112 Debye. By calculating the cosine similarity between the target gas and other gas molecules, molecular pairs with a similarity greater than 0.9 may undergo chemical reactions;

[0096] When screening reaction chains, the activation energy required for molecular bond breaking is an important reference indicator. For example, the activation energy for breaking the carbon-oxygen single bond in a carbon monoxide molecule is 1076 kJ / mol, and the activation energy for breaking the carbon-oxygen double bond in a carbon dioxide molecule is 532 kJ / mol.

[0097] The gas phase reaction rate constant database records the reaction rate constants at different temperatures. For example, the reaction rate constant of carbon monoxide and oxygen at 298 Kelvin is 2.3×10 -4 cubic meter per mole per second. As the temperature rises to 1500 Kelvin, the reaction rate constant increases to 3.8×10 -2 cubic meter per mole per second;

[0098] Based on the measured gas concentration time series curve, the Pearson correlation coefficient between the change trend of the reactant concentration and the theoretical prediction value can be calculated. For the reaction of carbon monoxide oxidation to produce carbon dioxide, when the correlation coefficient is greater than 0.8, it means that the reaction does occur in the fire scene;

[0099] The hierarchical clustering algorithm classifies reaction chains with similar correlation coefficients into one category. For example, the oxidation reaction of carbon monoxide and the oxidation reaction of ammonia are both exothermic reactions and can be classified into the same category.

[0100] The thermodynamic database provides the thermodynamic parameters of reactants and products, such as the standard enthalpy of formation of carbon monoxide is -110.5 kJ / mol, the standard entropy is 197.6 J / mol Kelvin, the standard enthalpy of formation of carbon dioxide is -393.5 kJ / mol, the standard entropy is 213.7 J / mol Kelvin;

[0101] Based on these parameters, the equilibrium constant of the reaction at different temperatures can be calculated, and then the concentration of the product in the theoretical equilibrium state can be obtained;

[0102] When the root mean square error between the theoretical concentration and the measured concentration is less than 10%, the support vector regression of the radial basis kernel function is used to establish the reaction chain mapping relationship. The scaling parameter of the kernel function is set to 0.1 and the penalty coefficient is set to 100 to achieve a balance between the accuracy and generalization of the reaction chain prediction.

[0103] Furthermore, in this embodiment, step S3 specifically includes:

[0104] The effects of environmental parameters on gas concentration fluctuations include gas reaction rate, gas diffusion rate, and gas mixing degree;

[0105] A sensor network is used to obtain the temperature data, humidity data and pressure data of the fire scene. Based on the temperature data, humidity data and pressure data, the temperature distribution field, humidity distribution field and pressure distribution field are obtained through radial basis function calculation.

[0106] For the temperature distribution field, humidity distribution field and pressure distribution field, the activation energy value and pre-exponential factor value of the reactants are extracted through the chemical reaction chain mapping relationship, and the reaction rate influence coefficient is obtained based on the activation energy value and pre-exponential factor value combined with the Arrhenius equation;

[0107] According to the reaction rate influence coefficient, the gas diffusion flux value is calculated using Fick's law, and a deep neural network model is established through the gas diffusion flux value to obtain the gas diffusion velocity distribution data;

[0108] For the gas diffusion velocity distribution data, the Shannon entropy value is calculated through the gas concentration time series curve. Based on the Shannon entropy value, a multivariate autoregressive model is established to obtain the environmental parameter influence coefficient matrix.

[0109] Specifically, multi-point time series data of temperature, humidity and pressure are collected from the fire scene sensor network, and a three-dimensional interpolation algorithm based on radial basis function is used to spatially interpolate the sampling point data to construct the temperature distribution field, humidity distribution field and pressure distribution field, and the three-dimensional gradient vector of each distribution field is calculated to obtain the first environmental field data;

[0110] For the first environmental field data, the activation energy and pre-exponential factor of the reactants are extracted based on the chemical reaction chain mapping relationship, the influence coefficient of temperature on the reaction rate constant is calculated in combination with the Arrhenius equation, and the correction factor of pressure on the collision frequency of gas molecules is calculated using the state equation to generate the second reaction kinetics data;

[0111] For the second reaction kinetics data, Fick's first and second laws are used to calculate the diffusion flux of gas under the action of temperature gradient, pressure gradient and humidity gradient. A deep neural network consisting of three convolutional layers and two fully connected layers is established to predict the distribution of gas diffusion velocity in space and obtain the third diffusion field data.

[0112] For the third diffusion field data, the Shannon entropy is calculated according to the gas concentration time series curve as the gas mixing index, and a multivariate autoregressive model with temperature, pressure and humidity as independent variables is established to quantify the influence coefficient of environmental parameters on the gas mixing degree.

[0113] The second reaction kinetics data, the third diffusion field data and the gas mixing index are combined to construct the coupling relationship matrix between the environmental parameters and the chemical reaction chain. The singular value decomposition method is used to extract the main influencing factors and determine the influence of the environmental parameters on the gas concentration fluctuation.

[0114] Here is an example:

[0115] There is a complex coupling relationship between the environmental parameters and the chemical reaction chain at the fire scene. The sensor network is arranged at different locations of the fire scene to collect environmental parameters. Taking the temperature sensor as an example, 27 measuring points are evenly arranged in a space of 3 meters × 3 meters × 3 meters. Each measuring point collects one temperature value per second, and the measurement range is 298 to 1500 Kelvin.

[0116] When radial basis function is used for three-dimensional interpolation, Gaussian kernel function is selected as basis function, and the influence radius is set to 1 meter. The continuous temperature distribution field of the entire space is obtained by interpolation;

[0117] The relationship between gas reaction rate and temperature follows the Arrhenius equation. Taking the carbon monoxide oxidation reaction as an example, its activation energy is 125 kJ / mol and the pre-exponential factor is 4.6×10 11 cubic meter per mole per second;

[0118] When the temperature rises from 298 Kelvin to 1500 Kelvin, the reaction rate constant increases by about 5 orders of magnitude. The effect of pressure on the reaction rate is mainly reflected in the frequency of molecular collisions. According to the equation of state, the collision frequency increases by about 0.3 times for every 100 kPa increase in pressure;

[0119] The diffusion process of gases is described by Fick's diffusion law. For the diffusion of carbon monoxide in air, its diffusion coefficient is 2×10 at 298 Kelvin and 101 kPa. -5 Square meter per second, the diffusion coefficient increases by about 15% for every 100 Kelvin increase in temperature, and by about 12% for every 10 kPa decrease in pressure;

[0120] The deep neural network contains three convolutional layers, each with 64, 128, and 256 convolution kernels of size 3×3×3, which are used to extract spatial features;

[0121] The number of neurons in the two fully connected layers is 512 and 256, respectively, which are used to predict the diffusion speed distribution;

[0122] The degree of gas mixing is characterized by Shannon entropy. Taking the mixing of two gases as an example, when the volume ratio is 1:1, the maximum entropy value is 0.693, and the entropy value decreases when it deviates from this ratio. The multivariate autoregressive model uses the temperature, pressure, and humidity time series with a lag order of 3 as input to predict the change of mixing entropy;

[0123] The dimension of the coupling relationship matrix is ​​the product of the number of parameters and the number of influencing indicators. The main influencing patterns can be extracted through singular value decomposition;

[0124] When the cumulative contribution rate of singular values ​​reaches 90%, the corresponding singular vector reflects the main way in which environmental parameters affect gas concentration fluctuations. Temperature plays a role by affecting the reaction rate, and its maximum singular value is 0.85;

[0125] Pressure exerts its influence by affecting the collision frequency and diffusion speed, with a singular value of 0.62. Humidity has a relatively small effect, with a singular value of 0.31.

[0126] Furthermore, in this embodiment, step S4 specifically includes:

[0127] According to the coupling relationship matrix, the values ​​of the temperature field, pressure field and humidity field at the spatial grid points are obtained, and the concentration gradient data of the target gas in three-dimensional space is obtained through numerical calculation;

[0128] The concentration gradient data is used to calculate the mixed entropy value of the target gas and the ambient gas. A Bayesian network including three nodes of temperature, pressure and humidity is established through the mixed entropy value to obtain the gas concentration prediction data.

[0129] The gas concentration prediction data is divided into multiple time windows, and the deviation feature vector between the predicted value and the measured value is calculated using the time window. The deviation feature vector is grouped by the hierarchical clustering algorithm to obtain the deviation feature data.

[0130] Determine whether the standard deviation exceeds a preset upper limit of the standard deviation threshold according to the deviation characteristic data. If it exceeds the preset upper limit of the standard deviation threshold, extract the spectrum characteristics of the deviation characteristic data by Fourier transform, and determine the change mode of the target gas concentration by the spectrum characteristics;

[0131] Specifically, the values ​​of the temperature field, pressure field and humidity field at the spatial grid points are read from the environmental parameter and chemical reaction chain coupling matrix, the concentration gradient of the target gas in the three-dimensional space is calculated based on Fick's diffusion law, and the mean free path and collision frequency of the gas molecules are calculated through the gas kinetic equation to obtain the first spatiotemporal diffusion data;

[0132] For the first spatiotemporal diffusion data, Shannon entropy and Gibbs mixing entropy are used to calculate the mixing uniformity of the target gas and the ambient gas, and a Bayesian conditional probability network containing three nodes of temperature, pressure and humidity is established to predict the trend of gas concentration changes, and the second time series prediction data is obtained;

[0133] The second time series prediction data is divided into multiple time windows to calculate the mean square error, standard deviation and maximum deviation between the predicted concentration value and the measured concentration value in each window, and the deviation feature vectors of all time windows are grouped using a hierarchical clustering algorithm to obtain third deviation feature data;

[0134] The third deviation characteristic data is judged according to a preset standard deviation threshold and a maximum deviation threshold. When the standard deviation exceeds the preset upper threshold or the maximum deviation exceeds the preset lower threshold, Fourier transform is used to extract the spectrum characteristics of the deviation sequence.

[0135] The main frequency components and energy distribution of the deviation sequence are calculated based on the spectrum characteristics to determine whether the target gas concentration fluctuation presents a periodic change mode, a monotonically increasing mode or an oscillating attenuation mode, and obtain the deviation mode determination result;

[0136] Here is an example:

[0137] In the prediction of environmental parameters and gas concentrations at the fire scene, the coupling relationship matrix between environmental parameters and chemical reaction chains records the temperature, pressure and humidity values ​​at the spatial grid points. Taking a space of 3 meters × 3 meters × 3 meters as an example, it is divided into 27 grid points, and the temperature range is 298 to 1500 Kelvin, the pressure range is 90 to 110 kPa, and the relative humidity range is 20% to 90% at each grid point;

[0138] When calculating the concentration gradient of the target gas based on Fick's diffusion law, a grid spacing of 0.5 m was selected, and the diffusion coefficient of carbon monoxide under standard conditions was calculated to be 2×10 -5 square meters per second, the gas dynamics equation calculated the molecular mean free path to be 70 nanometers and the collision frequency to be 7×10 9 times per second;

[0139] The calculation of gas mixing uniformity uses Shannon entropy and Gibbs mixing entropy. For binary gas mixtures, when the volume ratio of the two gases is 1:1, Shannon entropy reaches a maximum value of 0.693, and gradually decreases as the ratio deviates from 1:1;

[0140] The conditional probabilities between the three nodes of temperature, pressure and humidity in the Bayesian network are obtained through historical data training. When the temperature increases by 100 Kelvin, the probability of carbon monoxide concentration increasing is 0.8, and when the pressure decreases by 10 kPa, the probability of concentration increasing is 0.6;

[0141] In the time window division, 10 seconds is selected as the window length, and adjacent windows overlap for 5 seconds. The statistical characteristics of the predicted value and the measured value are calculated for each window, and the mean square error threshold is set to 0.01, the standard deviation threshold is set to 0.1, and the maximum deviation threshold is set to 0.2;

[0142] Hierarchical clustering uses Euclidean distance as the similarity measure and clusters the deviation feature vectors into three categories, corresponding to the three states of stability, fluctuation and drastic change;

[0143] When extracting the spectrum features of the deviation sequence through Fourier transform, the sampling frequency is 1 Hz and the transform length is 128 points;

[0144] The criterion for determining the main frequency component is that the energy proportion exceeds 50%. When a main frequency component of 0.1 Hz appears, it indicates that the deviation fluctuates with a period of 10 seconds.

[0145] When the spectrum shows a continuous upward trend, it indicates that the deviation presents a monotonically increasing pattern; when the energy of the high-frequency component gradually decays, it indicates that the deviation presents an oscillating decay pattern;

[0146] In practical applications, taking the prediction of carbon monoxide concentration as an example, when the ambient temperature rises from 500 Kelvin to 800 Kelvin, the deviation sequence between the predicted value and the measured value shows a main frequency component of 0.05 Hz after Fourier transformation, and the energy accounts for 65%, indicating that the concentration prediction deviation presents a fluctuation pattern with a period of 20 seconds;

[0147] At the same time, the standard deviation of the deviation sequence is 0.15, exceeding the preset threshold of 0.1, and the maximum deviation is 0.25, exceeding the preset threshold of 0.2. Comprehensive judgment shows that the forecast result has significant periodic deviation.

[0148] Furthermore, in this embodiment, step S5 specifically includes:

[0149] The gas concentration time series data is obtained according to the three-dimensional spatial coordinates of the hazardous source, and the low-frequency approximate coefficients and high-frequency detail coefficients of the time series data are extracted at several decomposition scales by using wavelet transform to obtain the first characteristic data;

[0150] For the first characteristic data, a random forest classifier is used to establish a mapping relationship between the first characteristic data and the hazard source type, and the classifier parameters are optimized through cross-validation to obtain the second classification data;

[0151] The Euclidean distance matrix between the gas concentration fluctuation characteristics of different types of hazardous sources is calculated based on the second classification data, the hazardous sources are grouped using a spectral clustering algorithm, and typical fluctuation templates of each group are extracted to obtain third template data;

[0152] If the Pearson correlation coefficient between the measured gas concentration curve and the third template data is greater than the preset Pearson correlation coefficient determination threshold, the naive Bayes method is used to calculate the posterior probability of the location of the hazardous source to determine the spatial location area of ​​the hazardous source and the type of hazardous substance;

[0153] Specifically, the three-dimensional spatial coordinates of the hazardous source, the type of hazardous source material and the corresponding gas concentration time series data are read from the historical fire database, the statistical characteristics of the gas concentration fluctuation curve are extracted according to the fixed sampling period, and the low-frequency approximate coefficients and high-frequency detail coefficients of the fluctuation curve are extracted at three decomposition scales using wavelet transform to obtain the first-level feature data;

[0154] For the first-level feature data, the relative distance vector and azimuth vector are calculated according to the spatial coordinates of the hazard source and the coordinates of the gas sampling point, and a sample matrix containing fluctuation characteristics, distance characteristics and azimuth characteristics is constructed. The correspondence between the feature samples and the hazard source types is established using a random forest classifier, and the classifier parameters are optimized through five-fold cross validation to obtain the second classification mapping data.

[0155] For the second classification mapping data, the Euclidean distance matrix between the gas concentration fluctuation characteristics caused by different types of hazardous sources is calculated, the hazardous sources are grouped using a spectral clustering algorithm, and each group of typical fluctuation templates is extracted to obtain the third template feature data;

[0156] For the third template feature data, the Pearson correlation coefficient is used to calculate the similarity between the measured concentration curve of the target gas and the typical fluctuation template of the hazardous source. When the correlation coefficient is greater than the preset judgment threshold, the fourth matching association data is generated;

[0157] According to the fourth matching association data, the naive Bayes method is used to calculate the posterior probability that the location of the hazard source falls in different spatial grid areas, and the spatial location area and hazardous substance type of the hazard source are determined in combination with the conditional probability distribution of the hazard source type;

[0158] Here is an example:

[0159] In the identification of hazardous sources at the fire scene, the historical database records the characteristics of gas concentration changes caused by different types of hazardous sources;

[0160] Taking the methanol storage tank as an example, its spatial coordinates are expressed in a three-dimensional rectangular coordinate system. The origin of the coordinates is set at the entrance of the fire scene, the position coordinates of the storage tank are (10 meters, 5 meters, 2 meters), and the gas concentration sampling points are arranged around the storage tank at intervals of 1 meter.

[0161] The sampling period of the gas concentration time series data is 1 second, and a continuous concentration change curve is obtained within the 120-second observation time;

[0162] When performing wavelet transform on the concentration fluctuation curve, the db4 wavelet basis function is selected to extract low-frequency approximate coefficients and high-frequency detail coefficients at three decomposition scales respectively;

[0163] The first level of decomposition reflects the changing trend of the 120-second scale, the second level of decomposition reflects the fluctuation characteristics of the 60-second scale, and the third level of decomposition reflects the local details of the 30-second scale;

[0164] For the fire caused by methanol tank leakage, the low-frequency coefficient shows an overall increasing trend in concentration, and the high-frequency coefficient shows obvious periodic fluctuations;

[0165] In the construction of feature samples, the distance feature includes the Euclidean distance from the sampling point to the hazard source. The distance between the methanol storage tank and the surrounding 8 sampling points is distributed in the range of 1 to 3 meters;

[0166] The orientation feature includes the horizontal azimuth angle and the elevation angle, forming a 16-dimensional position feature vector;

[0167] The random forest classifier contains 100 decision trees with a maximum depth of 8 layers, and the optimal parameter combination is determined by five-fold cross validation;

[0168] The clustering results of hazard source types show that flammable liquid hazard sources, such as methanol and ethanol storage tanks, exhibit similar concentration fluctuation characteristics, and their typical templates show a form of rapid rise and then stabilization;

[0169] The concentration fluctuation of gaseous flammable substances, such as natural gas pipelines, shows periodic oscillation characteristics, while the concentration of solid flammable substances, such as plastic products, changes relatively slowly. The correlation coefficient threshold with the typical template is set to 0.8, and a match is considered successful if it exceeds this threshold;

[0170] In Bayesian probability inference, the fire scene space is divided into 27 grid areas, each with a size of 2 m × 2 m × 2 m;

[0171] When it is detected that the target gas concentration shows a rapid rising characteristic and the correlation coefficient with the typical template of flammable liquids is 0.85, the location of the hazardous source is most likely to fall in the 9 grid areas within 2 meters from the sampling point, among which the posterior probability of the central grid is the highest, reaching 0.45, and the probability of the hazardous source type being judged as flammable liquid is 0.82, realizing the probabilistic estimation of the location and type of the hazardous source.

[0172] Furthermore, in this embodiment, step S6 specifically includes:

[0173] The bicubic interpolation algorithm is used to spatially interpolate the concentration values ​​between the gas concentration sampling points to obtain continuous concentration distribution field data;

[0174] The continuous concentration distribution field data are filtered using a Gaussian kernel function to obtain first filtered field data;

[0175] According to the first filter field data, a concentric circle search area centered on the initial position of the hazard source is divided, and a particle swarm algorithm is used to iteratively search the concentration gradient in the search area to obtain the second gradient field data;

[0176] The statistical moment of gas concentration in the search area is calculated for the second gradient field data, and the hidden Markov chain is used to estimate the state transition of the location coordinates and type identification of the hazard source, and the spatial coordinates of the hazard source are determined;

[0177] Specifically, gas concentration data are collected from the sensor network, and a concentration distribution matrix is ​​constructed according to the coordinates of the spatially arranged sampling points. The concentration values ​​between the sampling points are spatially interpolated using a bicubic interpolation algorithm with a cubic polynomial weight as a parameter to generate a continuous concentration distribution field with a spatial resolution of one-fourth of the spacing between the sampling points.

[0178] The concentration field data is filtered using a Gaussian kernel function with a standard deviation of half the sampling point spacing to obtain the first filtered field data;

[0179] For the first filtered field data, combined with the initial position of the hazard source generated by the classifier, three layers of concentric circle search areas with radii of integer multiples of the sampling point spacing are divided with the initial position as the center, and the particle swarm algorithm with a population size of fifty is used to iteratively search the concentration gradient in each search area to obtain the second gradient field data;

[0180] For the second gradient field data, the statistical moment of the gas concentration in each search area is calculated based on the sliding time window, and the state transition of the hazardous source position coordinates and type identification is estimated by using a hidden Markov chain whose observation probability matrix size is the number of hazardous source types to obtain the third state estimation data;

[0181] For the third state estimation data, the normalized mutual correlation coefficient between the gas concentration field at the current moment and the preset hazard source characteristic template is calculated, and the spatial coordinates of the hazard source are determined according to the peak position and value of the mutual correlation coefficient matrix;

[0182] Based on the spatial coordinates and type identification of the hazardous source, combined with the real-time change trend of the gas concentration field, the hazardous source location information is updated according to a fixed data sampling cycle to generate real-time tracking data on the location and type of the hazardous source;

[0183] Here is an example:

[0184] In the location of dangerous sources at the fire scene, the sensor network is evenly arranged in space at intervals of 2 meters to form a three-dimensional sampling grid of 8×8×4;

[0185] For the discrete concentration data obtained at the sampling points, a bicubic interpolation algorithm is used to generate a continuous concentration distribution field with a resolution of 0.5 m;

[0186] The interpolation algorithm uses a cubic polynomial weight function to calculate the concentration value of the interpolation point based on the data of 16 adjacent sampling points, ensuring the continuity and smoothness of the spatial concentration distribution;

[0187] Gaussian filtering uses a kernel function with a standard deviation of 1 meter, which effectively eliminates sensor noise interference;

[0188] After the hazard source is initially located, three concentric circles are divided into search areas with the initial location as the center, with radii of 2 meters, 4 meters, and 6 meters respectively;

[0189] The population size of the particle swarm algorithm is 50, and each particle carries position and velocity information, tracking the direction with the largest concentration gradient in the search area;

[0190] Taking the methanol tank leakage as an example, when the search area radius is 4 meters, after 20 iterations, the particle swarm finds the local maximum of the concentration gradient at 3.2 meters from the initial position, indicating the possible location of the hazard source;

[0191] During the state estimation process, a 30-second sliding time window was used to calculate statistical features;

[0192] For flammable liquid hazardous sources, the skewness of their concentration distribution is usually greater than 0.8, showing an obvious right-skewed characteristic;

[0193] The kurtosis is greater than 3.5, indicating that the distribution curve is steeper than the normal distribution;

[0194] The dimension of the observation probability matrix of the hidden Markov chain is 5 × 5, corresponding to the five main types of hazard sources: flammable liquids, flammable gases, flammable solids, explosives, and toxic substances;

[0195] The hazard source feature template is a 5×5×5 three-dimensional array that describes the typical concentration distribution patterns of different types of hazard sources in the surrounding space;

[0196] When calculating the normalized cross-correlation coefficient between the current concentration field and the characteristic template, the sliding step size is set to 0.5 m, and the cross-correlation peak is searched in the entire search space;

[0197] When the correlation coefficient exceeds 0.85, it is considered that the position matching the feature template has been found. For the methanol tank leakage scenario, within 120 seconds after the fire, the tracking error of the hazard source position was kept within 1 meter, and the accuracy of type recognition reached 90%;

[0198] The data update cycle is set to 1 second, matching the sampling frequency of the sensor. In each update cycle, the location coordinates and type identification of the hazard source are dynamically adjusted based on the latest concentration field data and state estimation results;

[0199] When the position change is less than 0.5 meters within five consecutive cycles and the type judgment remains consistent, it means that the positioning result tends to be stable. Through this real-time tracking mechanism, the hazard source positioning information can timely reflect the dynamic changes of the fire situation.

[0200] Furthermore, in this embodiment, step S7 specifically includes:

[0201] Real-time adjustment of the coupling coefficient between environmental parameters and chemical reaction chains includes adjustment of reaction rate constants;

[0202] Generate dynamically adjusted hazard source identification results including hazard source type, location and hazard level information;

[0203] An environmental change trend vector is constructed based on the measured values ​​of environmental parameters collected by the temperature sensor, the pressure sensor and the humidity sensor, and a predicted value of future environmental parameter changes is obtained through the environmental change trend vector;

[0204] According to the predicted value of future environmental parameter changes, the reaction rate constant is calculated using the temperature correction term and the pressure correction term. The Kalman filter is established through the reaction rate constant and the recursive least squares method to obtain the predicted value of gas concentration.

[0205] For the predicted value of gas concentration, the gradient vector and Hessian matrix of the concentration field are calculated, and the coordinates of the hazard source location are obtained by support vector regression using the kernel function as the radial basis;

[0206] For the location coordinates of the hazardous source, the Bayesian probability estimation is used to update the posterior distribution of the hazardous source type, and the hazard index vector is constructed according to the gas concentration exceeding the standard multiple, the spreading speed and the diffusion range to obtain the hazard level;

[0207] Specifically, according to the real-time location information of the hazard source and the measured values ​​of the environmental parameters, the changes in temperature, pressure and humidity per unit time are used to construct the environmental change trend vector, the change values ​​of the environmental parameters at future moments are predicted by the third-order autoregressive equation of the sliding time window, and the temperature correction term and the pressure correction term of the reaction rate constant are calculated based on the activation energy and the pre-exponential factor to obtain the first rate parameter data;

[0208] For the first rate parameter data, the gas reaction rate is calculated using the modified reaction rate constant, the linear coupling coefficient between the gas concentration and the environmental parameter is updated by the recursive least square method, and the Kalman filter of the measurement equation and the state equation is established to predict the gas concentration fluctuation trend to obtain the second concentration prediction data;

[0209] For the second concentration prediction data, the first-order spatial derivative and the second-order spatial derivative of the gas concentration field around the hazardous source are calculated, the gradient vector and Hessian matrix of the concentration field are constructed, and the support vector regression with the kernel function as the radial basis is used to update the coordinates of the hazardous source position to obtain the third coordinate data;

[0210] For the third coordinate data, the posterior distribution of the hazard source type is updated by using Bayesian probability estimation, and the type identification is determined in combination with the hazard source type characteristics to obtain the fourth type of data;

[0211] Based on the three dimensions of gas concentration exceeding the standard multiple, spreading speed and diffusion range, a hazard index vector is constructed. The hazard level is quantitatively graded according to the Euclidean norm of the index vector, and a dynamic identification result including the hazard source type identification, spatial location coordinates and hazard level is generated.

[0212] Here is an example:

[0213] During the dynamic identification of hazardous sources at a fire scene, the rapid changes in environmental parameters have a significant impact on the gas concentration distribution;

[0214] The environmental change trend vector uses a 10-second sliding time window to record the change rate of temperature, pressure, and humidity. When the temperature rise rate exceeds 20 Kelvin per minute, a third-order autoregressive equation is used to predict the temperature change trend within the next 30 seconds.

[0215] Taking the carbon monoxide oxidation reaction as an example, its activation energy is 125 kJ / mol, the pre-exponential factor is 4.6×10^11, and the temperature rises from 500 Kelvin to 800 Kelvin, resulting in a reaction rate constant increase of about 1000 times;

[0216] The coupling coefficients of gas concentration and environmental parameters were updated by recursive least squares method, with the initial covariance matrix set to the identity matrix and the forgetting factor set to 0.95;

[0217] The state vector of the Kalman filter contains the gas concentration and the concentration change rate. The measurement equation is constructed based on the actual measured value of the sensor. The state equation describes the evolution of the concentration over time.

[0218] For carbon monoxide, the state noise covariance is 0.01, the measurement noise covariance is 0.02, and the filter updates the state estimate once per second;

[0219] In the update of the hazard source location, the spatial derivative of the gas concentration field reflects the direction and severity of the concentration change;

[0220] Taking a 3m × 3m × 3m spatial area as an example, the concentration gradient vector is calculated on a 0.5m grid, where the components in the x, y, and z directions reflect the spatial distribution characteristics of the concentration change;

[0221] The eigenvalues ​​of the Hessian matrix characterize the curvature characteristics of the concentration field. Positive eigenvalues ​​indicate a convex distribution of concentration, while negative eigenvalues ​​indicate a concave distribution.

[0222] Support vector regression uses radial basis kernel function with kernel parameter of 0.1 and penalty coefficient of 100. Bayesian probability estimation of hazard source type is based on historical data to establish prior distribution, and real-time observation data is used to update posterior probability;

[0223] Taking flammable liquid hazardous sources as an example, when a rapidly rising concentration curve is observed and the spatial distribution shows obvious point source characteristics, the posterior probability of this type is significantly increased;

[0224] In the hazard index vector, the concentration excess multiple is calculated based on the occupational exposure limit, the spread speed is determined by the expansion rate of the high concentration area per unit time, and the diffusion range is calculated based on the spatial volume where the concentration exceeds the threshold;

[0225] When the Euclidean norm of the indicator vector is less than 1.0, it is judged as a mild hazard, when the norm is between 1.0 and 2.0, it is a moderate hazard, and when it is greater than 2.0, it is a severe hazard;

[0226] The dynamic identification results are updated once a second and include the spatial coordinates of the hazard source, type determination results and hazard level, enabling real-time tracking of hazard source characteristics and dynamic assessment of the degree of hazard.

[0227] In the description of the present invention, it is necessary to understand that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the scope of protection of the present invention.

[0228] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present invention.

Claims

1. A cloud service-based intelligent firefighting big data analysis method, characterized in that: The following steps are involved: S1. Obtain multi-source gas concentration time series data collected by the gas sensor array at the fire scene, and pre-process the concentration change trends of different gases respectively to remove noise interference and obtain a smooth gas concentration time series curve; S2. Analyze the concentration changes of different gases according to the gas concentration time series curve, and combine with the pre-established chemical reaction chain model, and if there is a chemical reaction of the target gas, generate a mapping relationship of the chemical reaction chain related to the target gas; S3. Obtain environmental parameters at the fire scene, combine the mapping relationship of the chemical reaction chain, establish a coupling relationship matrix between environmental parameters and chemical reaction chains, and determine the impact of environmental parameters on gas concentration fluctuations; S4. Calculate the diffusion speed and mixing degree of different gases under different environmental conditions through the coupling relationship matrix, generate a predicted value of the target gas concentration fluctuation, and analyze the degree of deviation between the predicted value of the target gas concentration fluctuation and the actual measured value; S5. If the target gas is related to the location and type of the hazard source, the relationship between the gas concentration fluctuation and the hazard source is extracted from the historical fire data, the typical gas concentration change patterns caused by different types of hazard sources are identified, the classification model is trained, and the preliminary estimation results of the hazard source type and location are generated; S6. Analyze the type and location of the hazard source at the current fire scene based on the preliminary estimation result and the real-time collected gas concentration time series data, and generate real-time location information of the hazard source; S7. Update the predicted value of gas concentration fluctuation through the real-time positioning information, adjust the coupling coefficient between environmental parameters and chemical reaction chain in real time in combination with the changing trend of environmental factors, and generate dynamically adjusted hazard source identification results.

2. The cloud service-based intelligent firefighting big data analysis method according to claim 1 is characterized in that: The step S1 specifically includes: Acquire various gas concentration time series data from the gas sensor array, and perform normalization processing on the concentration time series data according to the sensitivity coefficients of various gas sensors to obtain first gas concentration data; The first gas concentration data is smoothed by using a sliding mean filter, and according to a preset noise interference amplitude threshold, the data points exceeding the preset noise interference amplitude threshold are subjected to wavelet transform denoising to obtain second gas concentration data; For the second gas concentration data, a cubic spline interpolation function is used to fit discrete data points in the time series segment to obtain third gas concentration data; For the third gas concentration data, the first-order concentration change gradient value is calculated according to the concentration threshold value set according to the physical characteristics of each gas, and the gradient value is smoothed using a Gaussian kernel function to obtain a smoothed concentration time series curve of each gas.

3. The cloud service-based intelligent firefighting big data analysis method according to claim 1 is characterized in that: The step S2 specifically includes: Acquire target gas concentration data according to the gas concentration time series curve, calculate similarity values ​​between target gas molecules and other gas molecules through molecular structure feature vectors stored in a chemical reaction chain database, and obtain candidate chemical reaction chain data; For the candidate chemical reaction chain data, a pre-established gas phase reaction rate constant database is used to calculate the reaction rate constant value of each elementary reaction step in the reaction chain within a temperature range to obtain reaction chain kinetic data; Calculate the Pearson correlation coefficient between the reactant concentration change value and the measured gas concentration time series curve within the temperature range according to the reaction chain kinetic data, and group the reaction chains whose correlation coefficient is greater than the coefficient threshold by using a hierarchical clustering algorithm to obtain reaction chain association data; For the reaction chain associated data, the standard formation enthalpy change and standard entropy change values ​​of reactants and products are obtained from the thermodynamic database, the chemical reaction equilibrium constant of the reaction chain under temperature and pressure conditions is calculated, and a reaction chain mapping model is established through a support vector regression method.

4. The cloud service-based intelligent firefighting big data analysis method according to claim 1 is characterized in that: The step S3 specifically includes: The effects of environmental parameters on gas concentration fluctuations include gas reaction rate, gas diffusion rate, and gas mixing degree; A sensor network is used to obtain temperature data, humidity data and pressure data at the fire scene, and a temperature distribution field, a humidity distribution field and a pressure distribution field are obtained by radial basis function calculation according to the temperature data, the humidity data and the pressure data; For the temperature distribution field, the humidity distribution field and the pressure distribution field, the activation energy value and the pre-exponential factor value of the reactants are extracted through a chemical reaction chain mapping relationship, and the reaction rate influence coefficient is obtained according to the activation energy value and the pre-exponential factor value combined with the Arrhenius equation; According to the reaction rate influence coefficient, the gas diffusion flux value is calculated using Fick's law, and a deep neural network model is established through the gas diffusion flux value to obtain gas diffusion velocity distribution data; The Shannon entropy value is calculated for the gas diffusion velocity distribution data through the gas concentration time series curve, and a multivariate autoregressive model is established according to the Shannon entropy value to obtain the environmental parameter influence coefficient matrix.

5. The cloud service-based intelligent firefighting big data analysis method according to claim 1 is characterized in that: The step S4 specifically includes: According to the coupling relationship matrix, the values ​​of the temperature field, the pressure field and the humidity field at the spatial grid points are obtained, and the concentration gradient data of the target gas in the three-dimensional space is obtained by the numerical calculation; The concentration gradient data is used to calculate the mixed entropy value of the target gas and the ambient gas, and a Bayesian network including three nodes of temperature, pressure and humidity is established through the mixed entropy value to obtain gas concentration prediction data; Dividing the gas concentration prediction data into a plurality of time windows, using the time windows to calculate deviation feature vectors between the prediction value and the measured value, and grouping the deviation feature vectors by a hierarchical clustering algorithm to obtain deviation feature data; Whether the standard deviation exceeds a preset upper limit of the standard deviation threshold is determined based on the deviation characteristic data. If it exceeds the preset upper limit of the standard deviation threshold, Fourier transform is used to extract the spectral characteristics of the deviation characteristic data, and the target gas concentration change mode is determined by the spectral characteristics.

6. The cloud service-based intelligent firefighting big data analysis method according to claim 1 is characterized in that: The step S5 specifically includes: Acquire gas concentration time series data according to the three-dimensional spatial coordinates of the hazard source, and extract low-frequency approximate coefficients and high-frequency detail coefficients of the time series data at several decomposition scales using wavelet transform to obtain first characteristic data; For the first feature data, a random forest classifier is used to establish a mapping relationship between the first feature data and the type of hazard source, and classifier parameters are optimized through cross-validation to obtain second classification data; Calculating the Euclidean distance matrix between the gas concentration fluctuation characteristics of different types of hazardous sources according to the second classification data, grouping the hazardous sources using a spectral clustering algorithm, extracting a typical fluctuation template for each group, and obtaining third template data; If the Pearson correlation coefficient between the measured gas concentration curve and the third template data is greater than the preset Pearson correlation coefficient determination threshold, the naive Bayes method is used to calculate the posterior probability of the location of the hazardous source to determine the spatial location area of ​​the hazardous source and the type of hazardous substance.

7. The cloud service-based intelligent firefighting big data analysis method according to claim 1 is characterized in that: The step S6 specifically includes: The bicubic interpolation algorithm is used to spatially interpolate the concentration values ​​between the gas concentration sampling points to obtain continuous concentration distribution field data; The continuous concentration distribution field data is filtered using a Gaussian kernel function to obtain first filtered field data; Divide the concentric circle search area centered on the initial position of the hazard source according to the first filter field data, and use a particle swarm algorithm to iteratively search the concentration gradient in the search area to obtain second gradient field data; The statistical moment of gas concentration in the search area is calculated for the second gradient field data, and a hidden Markov chain is used to estimate the state transition of the position coordinates and type identification of the hazard source, and the spatial coordinates of the hazard source are determined.

8. The cloud service-based intelligent firefighting big data analysis method according to claim 1 is characterized in that: The step S7 specifically includes: Real-time adjustment of the coupling coefficient between environmental parameters and chemical reaction chains includes adjustment of reaction rate constants; Generate dynamically adjusted hazard source identification results including hazard source type, location and hazard level information; Constructing an environmental change trend vector based on the actual measured values ​​of environmental parameters collected by the temperature sensor, the pressure sensor and the humidity sensor, and obtaining a predicted value of future environmental parameter changes through the environmental change trend vector; According to the predicted value of the future environmental parameter change, a reaction rate constant is calculated using a temperature correction term and a pressure correction term, and a Kalman filter is established through the reaction rate constant and a recursive least square method to obtain a predicted value of the gas concentration; For the predicted value of gas concentration, the gradient vector and Hessian matrix of the concentration field are calculated, and the coordinates of the location of the hazard source are obtained by support vector regression using a kernel function as a radial basis; For the location coordinates of the hazardous source, Bayesian probability estimation is used to update the posterior distribution of the hazardous source type, and a hazard index vector is constructed according to the gas concentration exceeding the standard multiple, the spreading speed and the diffusion range to obtain the hazard level.

Images