High-rise building indoor fire source positioning method based on glass surface temperature

Abstract

A high-rise building indoor fire source positioning method based on glass surface temperature comprises the steps that m wireless thermocouples are arranged on the outer surface of glass, and the horizontal distance between the wireless thermocouple located at the center point position in a wireless thermocouple arrangement matrix and the walls of the two sides of a fire room is measured; the collected temperature data signals are stored every a set time period t and sent to a data processing end for average processing; and taking the averaged real-time temperature glass data and the horizontal distance between the central wireless thermocouple and the walls at the two ends as input data to obtain a data set, inputting the data set into a prediction program combined by the trained multi-layer perceptron and a KNN algorithm, and obtaining a prediction position block of a fire source through back propagation and category voting. According to the invention, the method achieves the accurate recognition of the position of a fire source, is not limited by a fire-fighting system and installation, is high in deployment flexibility, is high in real-time performance and reliability, meets the application requirements of various actual scenes, and is higher in universality and real-time performance.

Description

Technical Field

[0001] The invention relates to a method for locating indoor fire sources in high-rise buildings based on glass surface temperature, and belongs to the technical field of fire source locating. Background Art

[0002] In the process of accelerating urbanization and urbanization construction, the original building form cannot meet the living and working environment requirements of the urban population, and the building floors are gradually increasing. There are a large number of vertical shafts such as pipe wells, cable wells, and smoke exhaust ducts in high-rise buildings. Once a fire occurs, these shafts increase the vertical spread of fire in high-rise buildings and accelerate the spread of smoke. When carrying out fire fighting and rescue in high-rise buildings, it is often necessary to implement fire fighting operations from the outside. However, the water spraying height of existing fire trucks cannot meet the needs of fire fighting and rescue in high-rise buildings, and fire rescue personnel cannot complete the fire fighting of high-rise buildings in time. Positioning the indoor fire source of high-rise buildings can greatly improve the efficiency of fire fighting and rescue, and reduce casualties and property losses caused by fires.

[0003] At present, room fire source positioning technologies mainly include visual fire positioning technology, wireless sensor network positioning technology and optical fiber sensor positioning technology. These positioning technologies all require fixed position sensors, which need to be installed during building construction. Fixed deployment has poor flexibility and is likely to fail during a fire. In addition, these positioning technologies have a high time cost and cannot meet the real-time requirements in actual application scenarios. Summary of the invention

[0004] The present invention provides a method for locating indoor fire sources in high-rise buildings based on glass surface temperature. The method has high deployment flexibility, strong real-time reliability, and can meet the application of various practical scenarios and has a wide range of applications.

[0005] In order to achieve the above object, the present invention provides a method for locating indoor fire sources in high-rise buildings based on glass surface temperature, comprising the following steps:

[0006] Step 1: m wireless thermocouples are arranged on the outer surface of the fire room glass, and the square array is arranged on the largest piece of glass in the fire room; each wireless thermocouple is connected to a temperature data signal acquisition module;

[0007] Step 2: Measure the horizontal distance from the wireless thermocouple at the center point of the wireless thermocouple array to the walls on both sides of the fire room, and record them as D1 and D2;

[0008] Step 3: Divide the fire room into r blocks of equal area. The number of blocks is the same as the number of wireless thermocouples set, and they are arranged in a square matrix. They are numbered as the fire source location category output by the program.

[0009] Step 4: Adjust the temperature data signal acquisition module to receive the temperature data signal of each thermocouple at a frequency q;

[0010] Step 5: Start the fire source and temperature data signal acquisition module of the fire room, collect and read the temperature data of each thermocouple in real time, and save the collected temperature data signal every set time period t and send it to the data processing end;

[0011] Step 6: After receiving the temperature data, the data processing end averages all the temperature data obtained by each wireless thermocouple in each set time period t to obtain m average temperature data;

[0012] Step 7: The data processing end combines the obtained m average temperature data into a 1*m matrix and adds the D1 and D2 distance data to the 1*m matrix to obtain a 1*(m+2) matrix;

[0013] Step 8: Input the 1*(m+2) matrix obtained in step 7 into the prediction program combining the trained multi-layer perceptron and KNN algorithm, and obtain the predicted location block of the fire source through back propagation and category voting.

[0014] Furthermore, the specific process of obtaining the temperature data matrix in step 6 is as follows:

[0015] (1) The data processing end receives the temperature data and obtains:

[0016]

[0017] T m(tq) is the tqth data collected by wireless thermocouple m in time period t; the number of temperature data collected by a single wireless thermocouple in time period t is tq;

[0018] (2) Take the average value of each row of temperature data in the matrix of step (1) to form a row matrix:

[0019]

[0020] T m is the average temperature data value of the wireless thermocouple m in the time period t.

[0021] Furthermore, the specific process of step seven is: adding D1 and D2 to the row matrix obtained in step six to form a new matrix:

[0022] [T1,T2,T3,...T m ,D1,D2].

[0023] Furthermore, the training process of the prediction program combining the multilayer perceptron and the KNN algorithm in step eight is as follows:

[0024] (1) Use fire simulation software to simulate fire data and construct a fire data set. The specific process is as follows:

[0025] a. Build a fire combustion model and set parameters, including room geometry, fire source location, fire source heat release rate and fire source area;

[0026] b. Set m temperature measuring points in a square array of equal rows and columns on the glass of the fire room, record the temperature at the measuring point, and record the distance D1 and D2 from the center of the wireless thermocouple to the two ends of the wall;

[0027] c. Set the simulation time and obtain the temperature data set of the measured points after simulation {T mt}, T mt Represents the temperature data of the mth measuring point at the tth second;

[0028] d. Package the temperature measurement point data into a data set, where the vertical direction represents different times and the horizontal direction represents different wireless thermocouple temperature parameters and distance parameters:

[0029]

[0030] (2) Use train_test_split to divide the data set into training set and test set;

[0031] (3) Create a multi-layer perceptron model (MLP), set the structure of the hidden layer, the activation function ReLU, and the maximum number of iterations; specifically, set the input feature dimension according to the experimental data parameters, that is, the number of temperature measurement points and the number of distance parameters; set the structure of the hidden layer according to the input feature dimension, set the activation function ReLU and the maximum number of iterations;

[0032] (4) Call the fit method to train the multi-layer perceptron model and optimize the weights and biases of the network to minimize the loss function. The steps are:

[0033] a. Input data training set and initialize parameters, including weights and biases;

[0034] b. Pass the input data to each neuron in the input layer;

[0035] c. In the hidden layer, multiply the input data by the weight matrix of the current layer and add the bias term;

[0036] d. The weighted sum is transformed nonlinearly through the activation function Relu to obtain the output of the hidden layer:

[0037] H=σ(XW h ++b h )

[0038]

[0039] In the formula, H is the hidden layer, σ is the activation function, X is the input layer, and W h is the weight, b h is the bias term, The probability result obtained by the multi-layer perceptron, the softmax function is used to convert the output into a probability distribution;

[0040] e. In the output layer, multiply the output of the last hidden layer by the weight matrix of the output layer and add the bias term;

[0041] f. Use the Softmax activation function to calculate the final prediction output;

[0042] g. Calculate the error between the model prediction value and the actual label, that is, the loss value;

[0043] h. Use the chain rule to calculate the gradient of the loss function with respect to the weights and biases;

[0044] i. Update the network parameters using the calculated gradients;

[0045] j. Input the validation set and get the loss value;

[0046] k. Repeat steps b to j until the loss value of the validation dataset no longer decreases or reaches the number of data training times;

[0047] (5) Use the predict_proba method to read the output probability of the MLP model. The output probability represents the predicted probability of each fire source location category;

[0048] (6) Create a KNN classifier and set the number of neighbors k;

[0049] (7) Train the KNN classifier using features extracted from the multi-layer perceptron model;

[0050] (8) Call the fit method to train the KNN model. The steps are to store the training data and training labels. During the fit process, the KNN model will not modify or learn any data. It simply saves the data inside the model.

[0051] (9) Use the predict method to predict the test data. The steps are:

[0052] a. For each test sample input, calculate its Euclidean distance with all training samples:

[0053]

[0054] Where n is the dimension; x i is the training sample, y i is the test sample;

[0055] b. Select k nearest training samples based on the distance obtained in step a;

[0056] c. Voting or weighted voting on the labels of the selected k neighbors to determine the final predicted fire source location category of the test sample. The formula is:

[0057]

[0058] In the formula, is the prediction result; Π(y i =c) is the indicator function, if the category y of the i-th neighbor i The function is equal to 1 if c is true, and equal to 0 otherwise;

[0059] (10) Use the classification_report function to calculate the classification accuracy. The formula is:

[0060]

[0061] Where Accuracy is the accuracy; n 准确 is the number of correctly predicted test sets, n 全部 is the total number of test sets;

[0062] (11) Keep increasing the k value of the classifier until the accuracy stops increasing;

[0063] (12) After training is completed, save the parameters of the multilayer perceptron and KNN model.

[0064] Furthermore, the time period t is set to have a value range of 10 to 60 seconds, and the frequency q is set to have a value range of 1 Hz or more. The specific frequency is set according to the time period t to ensure that there is more than one temperature data collected in the time period t.

[0065] The present invention obtains the glass temperature data of the fire room through wireless thermocouples, uses the temperature difference calculation of multiple wireless thermocouples, and the classification model of the multi-layer sensor combined with the KNN algorithm to better handle nonlinear relationships, thereby realizing accurate identification of the fire source location. At the same time, the present invention can be applied to any form of building fire, is not restricted by the fire protection system and installation, and is used for fire protection to locate the fire source anytime and anywhere. It has high deployment flexibility and strong real-time reliability, meets the application of various practical scenarios, and has strong universality and real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a work flow chart of the present invention;

[0067] Figure 2 It is a schematic diagram of the horizontal distance between the central wireless thermocouple of the present invention and the walls at both ends;

[0068] Figure 3 It is a workflow diagram of the prediction program combining the multi-layer perceptron and the KNN algorithm of the present invention;

[0069] Figure 4 It is a comparison chart of the predicted results and the actual results of the embodiment of the present invention. DETAILED DESCRIPTION

[0070] The present invention will be further described below in conjunction with the accompanying drawings.

[0071] like Figure 1 As shown, a method for locating indoor fire sources in high-rise buildings based on glass surface temperature comprises the following steps:

[0072] Step 1: m wireless thermocouples are arranged on the outer surface of the fire room glass, and the square array is arranged on the largest piece of glass in the fire room; each wireless thermocouple is connected to a temperature data signal acquisition module;

[0073] Step 2: Figure 2 As shown, measure the horizontal distance from the wireless thermocouple at the center point of the wireless thermocouple array to the walls on both sides of the fire room and record them as D1 and D2;

[0074] Step 3: Divide the fire room into r blocks of equal area. The number of blocks is the same as the number of wireless thermocouples set, and they are arranged in a square matrix. They are numbered as the fire source location category output by the program.

[0075] Step 4: Adjust the temperature data signal acquisition module to receive the temperature data signal of each thermocouple at a frequency q;

[0076] Step 5: Start the fire source and temperature data signal acquisition module of the fire room, collect and read the temperature data of each thermocouple in real time, and save the collected temperature data signal every set time period t and send it to the data processing end;

[0077] Step 6: After receiving the temperature data, the data processing end averages all the temperature data obtained by each wireless thermocouple in each set time period t to obtain m average temperature data;

[0078] Step 7: The data processing end combines the obtained m average temperature data into a 1*m matrix and adds the D1 and D2 distance data to the 1*m matrix to obtain a 1*(m+2) matrix;

[0079] Step 8: Input the 1*(m+2) matrix obtained in step 7 into the prediction program combining the trained multi-layer perceptron and KNN algorithm, and obtain the predicted location block of the fire source through back propagation and category voting.

[0080] Embodiment: (1) A fire room model was established according to the size of a real room. The room model is 2.8m high, 5m long and 5m wide. The starting height of the room window is 1m, 1.8m wide and 1.5m high. It is kept closed during a fire. The door is 0.8m wide and 2m high. It is set on the opposite side of the room window and is kept open during a fire. The room floor is set to tile material, and the wall is set to concrete material. The upper surface of the combustible is foam with a thickness of 0.05m, and the remaining surfaces and interior are made of yellow pine material. The combustibles are set at nine locations in the room. For each group of experiments, the fire source is set in the form of a vent on the upper surface of a different single combustible. The maximum heat release rate of each fire source is set to be between 300 and 400kW, the simulation time is set to 600s, and the position of the window is changed to conduct multiple groups of experiments to obtain sufficient data.

[0081] (2) Randomly select 3,000 sets of data from the obtained data as the training data set, and use 80% of them as the training set and 20% as the test set:

[0082] (3) Figure 3 As shown, create an MLPClassifier instance, set the structure of the hidden layer to 64 neurons in the first layer and 32 neurons in the second layer, use Relu as the activation function, and set the maximum number of iterations to 1000; use the training set to train the multi-layer perceptron model (MLP model): adjust the weights and biases through forward propagation and back propagation algorithms to minimize the prediction error, usually the cross entropy loss;

[0083] (4) After training is completed, the MLP model is used to perform forward propagation on the training set and the test set to obtain the output of the hidden layer, i.e., the feature representation;

[0084] (5) The hidden layers of the MLP transform the input data into intermediate representations, which are used for feature extraction. The outputs of these hidden layers are considered as new feature sets for subsequent KNN classifiers. The purpose of feature extraction is to transform the data into the abstract representation learned by the MLP.

[0085] (6) Create a KNeighborsClassifier instance, set cross validation, select the k value of [1, 3, 5, 7, 9, 11], and use the features extracted from the MLP model to train the KNN classifier; KNN does not require a learning process during the training phase, but stores training data;

[0086] (7) Use the KNN classifier to predict the test data: KNN finds the nearest K neighbors based on the distance metric, and then decides the prediction category by majority voting; for the test sample, calculate the Euclidean distance between it and all samples in the training set:

[0087]

[0088] According to the calculated distance, the training samples are sorted from small to large, and the k training samples closest to the test sample are selected;

[0089] (8) Vote for the category labels of the selected k neighbors. The category label of each neighbor will increase by one vote. That is, if three neighbors have category 1 and two neighbors have category 3, then the category of the sample is 1. Calculate the number of votes for each category label, and then select the category with the most votes as the predicted category of the test sample:

[0090]

[0091] The experimental data table is as follows:

[0092]

[0093]

[0094] Through the above embodiments, the following conclusions are drawn: when the fire source is located on one side of the glass, the temperature of the wireless thermocouple on the horizontal line along the glass will always show a trend of decreasing from the fire source side to the non-fire source side; when the fire source is located at the vertical position of the center of the glass surface, the temperature of the horizontal line along the glass will not show any increasing or decreasing form; this change trend can be used as a basis for judging the horizontal position of the fire source;

[0095] Under the combined effects of smoke and heat radiation, when the fire source is far from the glass, the smoke has a greater impact on the glass temperature; when the fire source is close to the glass, the heat radiation and fire plume have a greater impact on the glass temperature; when the fire source is far from the glass, the temperature difference between the highest horizontal thermocouple and the center horizontal thermocouple is smaller than the temperature difference between the center horizontal thermocouple and the lowest horizontal thermocouple, and the opposite is true when the fire source is close to the glass. When the fire source is located on the center plane of the room perpendicular to the glass, the difference between the two is basically the same; this feature can be used as a basis for judging the longitudinal position of the fire source.

[0096] Such as experimental data table and Figure 4 As shown, the present invention can show a high accuracy rate with a small number of training sets, and has strong universality, and can have a high accuracy rate for different building structures. The present invention is suitable for high-rise building fire fighting and rescue, and can be installed immediately. Compared with other fire source positioning methods, the present invention has higher universality, timeliness and efficiency, and has a higher level of accuracy.

Claims

1. A method for locating indoor fire sources in high-rise buildings based on glass surface temperature, characterized in that: The steps include: Step 1: m wireless thermocouples are arranged on the outer surface of the fire room glass, and the square array is arranged on the largest piece of glass in the fire room; each wireless thermocouple is connected to a temperature data signal acquisition module; Step 2: Measure the horizontal distance from the wireless thermocouple at the center point of the wireless thermocouple array to the walls on both sides of the fire room, and record them as D1 and D2; Step 3: Divide the fire room into r blocks of equal area. The number of blocks is the same as the number of wireless thermocouples set, and they are arranged in a square matrix. They are numbered as the fire source location category output by the program. Step 4: Adjust the temperature data signal acquisition module to receive the temperature data signal of each thermocouple at a frequency q; Step 5: Start the fire source and temperature data signal acquisition module of the fire room, collect and read the temperature data of each thermocouple in real time, and save the collected temperature data signal every set time period t and send it to the data processing end; Step 6: After receiving the temperature data, the data processing end averages all the temperature data obtained by each wireless thermocouple in each set time period t to obtain m average temperature data; Step 7: The data processing end combines the obtained m average temperature data into a 1*m matrix and adds the D1 and D2 distance data to the 1*m matrix to obtain a 1*(m+2) matrix; Step 8: Input the 1*(m+2) matrix obtained in step 7 into the prediction program combining the trained multi-layer perceptron and KNN algorithm, and obtain the predicted location block of the fire source through back propagation and category voting.

2. The method for locating indoor fire sources in high-rise buildings based on glass surface temperature according to claim 1, characterized in that: The specific process of obtaining the temperature data matrix in step 6 is as follows: (1) The data processing end receives the temperature data and obtains: T m(tq) is the tqth data collected by wireless thermocouple m in time period t; the number of temperature data collected by a single wireless thermocouple in time period t is tq; (2) Take the average value of each row of temperature data in the matrix of step (1) to form a row matrix: T m is the average temperature data value of the wireless thermocouple m in the time period t.

3. The method for locating indoor fire sources in high-rise buildings based on glass surface temperature according to claim 2, characterized in that: The specific process of step seven is: add D1 and D2 to the row matrix obtained in step six to form a new matrix: <h2 style=";text-align:left;direction:ltr">[T1,T2,T3,...T<h2 style=";text-align:left;direction:ltr"> m <h2 style=";text-align:left;direction:ltr"> ,D1,D2]。 4. The method for locating indoor fire sources in high-rise buildings based on glass surface temperature according to claim 3 is characterized in that: The training process of the prediction program combining the multilayer perceptron and the KNN algorithm in step eight is as follows: (1) Use fire simulation software to simulate fire data and construct a fire data set. The specific process is as follows: a. Build a fire combustion model and set parameters, including room geometry, fire source location, fire source heat release rate and fire source area; b. Set m temperature measuring points in a square array of equal rows and columns on the glass of the fire room, record the temperature at the measuring point, and record the distance D1 and D2 from the center of the wireless thermocouple to the two ends of the wall; c. Set the simulation time and obtain the temperature data set of the measured points after simulation {T mt }, T mt Represents the temperature data of the mth measuring point at the tth second; d. Package the temperature measurement point data into a data set, where the vertical direction represents different times and the horizontal direction represents different wireless thermocouple temperature parameters and distance parameters: (2) Use train_test_split to divide the data set into training set and test set; (3) Create a multi-layer perceptron model, set the structure of the hidden layer, the activation function ReLU, and the maximum number of iterations; specifically, set the input feature dimension according to the experimental data parameters obtained, that is, the number of temperature measurement points and the number of distance parameters; set the structure of the hidden layer according to the input feature dimension, set the activation function ReLU and the maximum number of iterations; (4) Call the fit method to train the multi-layer perceptron model and optimize the network weights and biases to minimize the loss function. The steps are: a. Input data training set and initialize parameters, including weights and biases; b. Pass the input data to each neuron in the input layer; c. In the hidden layer, multiply the input data by the weight matrix of the current layer and add the bias term; d. The weighted sum is transformed nonlinearly through the activation function Relu to obtain the output of the hidden layer: H=σ(XW h ++b h ) In the formula, H is the hidden layer, σ is the activation function, X is the input layer, and W h is the weight, b h is the bias term, The probability result obtained by the multi-layer perceptron, the softmax function is used to convert the output into a probability distribution; e. In the output layer, multiply the output of the last hidden layer by the weight matrix of the output layer and add the bias term; f. Use the Softmax activation function to calculate the final prediction output; g. Calculate the error between the model prediction value and the actual label, that is, the loss value; h. Use the chain rule to calculate the gradient of the loss function with respect to the weights and biases; i. Update the network parameters using the calculated gradients; j. Input the validation set and get the loss value; k. Repeat steps b to j until the loss value of the validation dataset no longer decreases or reaches the number of data training times; (5) Use the predict_proba method to read the output probability of the multilayer perceptron model. The output probability represents the predicted probability of each fire source location category; (6) Create a KNN classifier and set the number of neighbors k; (7) Train the KNN classifier using features extracted from the multi-layer perceptron model; (8) Call the fit method to train the KNN model, and the steps are to store the training data and training labels; (9) Use the predict method to predict the test data. The steps are: a. For each test sample input, calculate its Euclidean distance with all training samples: Where n is the dimension; x i is the training sample, y i is the test sample; b. Select k nearest training samples based on the distance obtained in step a; c. Voting or weighted voting on the labels of the selected k neighbors to determine the final predicted fire source location category of the test sample. The formula is: In the formula, is the prediction result; Π(y i =c) is the indicator function, if the category y of the i-th neighbor i The function is equal to 1 if c is true, and equal to 0 otherwise; (10) Use the classification_report function to calculate the classification accuracy. The formula is: Where Accuracy is the accuracy; n 准确 is the number of correctly predicted test sets, n 全部 is the total number of test sets; (11) Keep increasing the k value of the classifier until the accuracy stops increasing; (12) After training is completed, save the parameters of the multilayer perceptron and KNN model.

5. The method for locating indoor fire sources in high-rise buildings based on glass surface temperature according to claim 1, characterized in that: The setting time period t ranges from 10 to 60 seconds, and the setting frequency q ranges from 1 Hz to above.

Images