A fire warning and post-disaster floor damage prediction method and system based on BIM and deep learning

Abstract

The present invention provides a method and system for fire warning and post-disaster floor damage prediction based on BIM and deep learning, which collects image information from different locations in the monitoring area, pre-processes the image information, and forms a first data set; inputs the image information of the first data set into a fire recognition neural network model; determines whether a fire occurs based on the first confidence parameter; if a fire occurs, a fire alarm is issued; identifies the flame information corresponding to the fire; inputs the flame information into a floor fire damage model based on BIM to predict the degree of damage to the floor bearing capacity. The present invention monitors fires in real time based on image information and deep learning algorithms, and warns of potential threats such as combustibles; at the same time, the fire warning information is integrated into the BIM building information model to further ensure the safety of life and property for the flame location and related floor damage threats, as well as accurate firefighting measures based on BIM.

Description

Technical Field

[0001] The present invention relates to the field of fire warning and fire fighting technology, and in particular to a fire warning method and system based on BIM and deep learning. Background Art

[0002] Fires pose a huge safety hazard to people's lives and property. Fire warning and fire protection systems are important components of building security systems. Traditional fire monitoring systems collect relevant data through temperature sensors, smoke sensors, etc. to determine the occurrence of fires. They often have a slow response speed and low efficiency in identifying flames at a distance, especially smokeless flames, and cannot guarantee the accuracy and timeliness of fire monitoring. At the same time, when a fire occurs, the traditional fire protection system will turn on the fire sprinkler device of the entire building or the entire floor, which cannot accurately extinguish the fire. In the case of a small fire, it will cause unnecessary property losses to other areas. When a fire occurs, it is impossible to plan an escape route to instruct on-site personnel to escape. A fire warning and fire protection system based on BIM and deep learning monitors fires in real time based on information such as images and deep learning algorithms, warns of potential threats such as combustibles, and integrates fire warning information into the BIM building information model, making full use of the three-dimensional visualization and collaborative characteristics of the BIM model to plan escape routes to help people escape, and accurately extinguish fires, thereby ensuring the safety of life and property. Summary of the invention

[0003] In view of this, the present invention provides a fire warning method and system based on BIM and deep learning, which monitors fires in real time according to information such as images and deep learning algorithms, warns of potential threats such as combustibles, and integrates fire warning information into the BIM building information model, thereby solving the technical problems of inaccurate fire extinguishing and inadequate identification of floor safety hazards.

[0004] The technical solution of the present invention is as follows:

[0005] A fire early warning and post-disaster floor damage prediction method based on BIM and deep learning, including:

[0006] Collecting image information at different locations within the monitoring area, and preprocessing the image information to form a first data set;

[0007] Inputting the image information of the first data set into a fire recognition neural network model to obtain a first confidence parameter;

[0008] determining whether a fire occurs based on the first confidence parameter; and if a fire occurs, issuing a fire alarm;

[0009] Identify flame information corresponding to the fire; input the flame information into a BIM-based floor fire damage model to predict the degree of damage to the floor bearing capacity.

[0010] Preferably, the step of inputting the image information into a fire identification neural network model to obtain a first confidence parameter comprises:

[0011] The original image size in the first dataset is compressed and cropped to 256×256×3, and convolution calculation is performed through two visual depth convolution layers C1 and C2, with the number of convolution kernels being 3×3, the number of channels being 100, the number of steps being 1, and the number of padding being 1;

[0012] The output result is calculated by the pooling layer P1 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C3, C4, and C5, the number of convolution kernels is 3×3, the number of convolution kernels is 128, the number of channels is 100, the step size is 1, and the padding number is 1;

[0013] The output result is calculated by the pooling layer P2 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C6, C7, and C8, the number of convolution kernels is 3×3, the number of channels is 200, the step size is 1, and the padding number is 1;

[0014] The output result is calculated by the pooling layer P3 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C9, C10, and C11, the number of convolution kernels is 3×3, the number of channels is 400, the step size is 1, and the padding number is 1;

[0015] The output result is calculated by the pooling layer P4 for the maximum value within the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by four visual depth convolution layers C12, C13, C14, and C15, the number of convolution kernels is 3×3, the number of channels is 400, the step size is 1, and the padding number is 1;

[0016] The output result is calculated through the pooling layer P5 for the maximum value in the range, the number of pooling cores is 2×2, and the step size is 2; the output result is calculated through the fully connected layer F1, and the number of channels is 4000; the output result is calculated through the fully connected layer F2, and the number of channels is 1000;

[0017] The output result is calculated by the fully connected layer F3 with a channel number of 100; the output result is calculated by the fully connected layer F4 with a channel number of 5; the output result is calculated by the softmax function normalized exponential layer and the output confidence Rn (n = 1, 2, 3..., 5).

[0018] Preferably, the first confidence parameter is an output matrix Rn;

[0019] The determining whether a fire has occurred based on the first confidence parameter includes: if the maximum value max(Rn) in the output matrix Rn is equal to R1 or max(Rn) is equal to R2 or max(Rn) is equal to R3, it is determined that a fire has occurred; if the maximum value max(Rn) is equal to R4 or max(Rn) is equal to R5, it is determined that there is no fire.

[0020] Preferably, the fire identification neural network model includes a training process;

[0021] The training process includes a weight adjustment phase, the task of which is to adjust the parameters according to the deviation value of the loss function until the model converges and reaches the target prediction accuracy. The loss function is as follows:

[0022] Rule 1: L = L CE (y i ,y i * )+αL MSE (y i ,y i * )

[0023] Rule 2:

[0024] Rule 3:

[0025] Among them, L is the loss function and L ≥ 0, y is the true distribution, that is, the true label of the sample, y * is the predicted distribution, L CE is the cross entropy loss function, L MSE is the mean square error loss function, α is the weight value, in order to adjust the proportion of the mean square error loss function in the entire loss function, N is the number of training batches, and C is the number of categories.

[0026] Preferably, the preprocessing of the image information includes preprocessing the image information using a variable bounding box.

[0027] Preferably, the flame information includes: the real length a, real width b, real height c of the flame, and the distance x between the flame and the floor; the characteristic parameters of the floor and the flame are combined to obtain a combined parameter X, and feature scaling is used to make the parameter value between [-1, 1].

[0028] Preferably, the inputting the flame information into a BIM-based floor fire damage model to predict the damage degree of the floor bearing capacity includes predicting the damage degree of the floor bearing capacity using a linear regression model:

[0029] Rule 1: f(x) = w T X+b

[0030] Rule 2: D = f(x) / Q

[0031] Where f(x) is the predicted floor bearing capacity, w T is the weight parameter, X is the merging parameter, b is the offset, Q is the initial bearing capacity strength of the floor before the fire, and D is the characteristic value of the bearing capacity damage degree.

[0032] Preferably, if a fire occurs, a fire alarm is issued, which also includes: triggering a fire alarm; triggering a solenoid valve in a fire sprinkler system to open a fire sprinkler head at the location where the fire occurs to extinguish the fire; and opening an escape sign on the escape route.

[0033] In addition, a fire warning and fire fighting system is proposed which is implemented by the above-mentioned fire warning and post-disaster floor damage prediction method based on BIM and deep learning. The system is characterized in that the system includes: a fire monitoring module based on a camera, an infrared thermal imager, a smoke detector, a video information storage device and a smoke information storage device; a fire information processing system; a computer intelligent terminal; a BIM model storage device; a visualization terminal; a cloud server; a fire fighting system control terminal, and a fire fighting and rescue module based on a fire sprinkler system, a fire alarm, a broadcast and an escape sign.

[0034] In the solution of the present invention, the fire warning and post-disaster floor damage prediction method based on BIM and deep learning collects image information at different locations in the monitoring area, pre-processes the image information, and forms a first data set; inputs the image information of the first data set into the fire recognition neural network model to obtain a first confidence parameter; determines whether a fire occurs based on the first confidence parameter; if a fire occurs, a fire alarm is issued; identifies the flame information corresponding to the fire; inputs the flame information into the floor fire damage model based on BIM to predict the degree of damage to the floor bearing capacity, so as to further perform corresponding warnings. The present invention performs real-time fire monitoring based on image information and deep learning algorithms, and warns of potential threats such as combustibles; at the same time, the fire warning information is integrated into the BIM building information model to further predict and warn of the flame location and related floor damage threats, and accurate fire fighting measures are carried out based on BIM, further ensuring the safety of life and property. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of a method for predicting fire early warning and post-disaster floor damage based on BIM and deep learning in an embodiment of the present invention;

[0036] Figure 2 This is a structural diagram of a fire warning and fire fighting system based on BIM and deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] like Figure 1 As shown, the present invention implements a fire early warning and post-disaster floor damage prediction method based on BIM and deep learning, including:

[0039] Collecting image information at different locations within the monitoring area, and preprocessing the image information to form a first data set;

[0040] Inputting the image information of the first data set into a fire recognition neural network model to obtain a first confidence parameter;

[0041] determining whether a fire occurs based on the first confidence parameter; and if a fire occurs, issuing a fire alarm;

[0042] Identify flame information corresponding to the fire; input the flame information into a BIM-based floor fire damage model to predict the degree of damage to the floor bearing capacity.

[0043] Preferably, the step of inputting the image information into a fire identification neural network model to obtain a first confidence parameter comprises:

[0044] The original image size in the first dataset is compressed and cropped to 256×256×3, and convolution calculation is performed through two visual depth convolution layers C1 and C2, with the number of convolution kernels being 3×3, the number of channels being 100, the number of steps being 1, and the number of padding being 1;

[0045] The output result is calculated by the pooling layer P1 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C3, C4, and C5, the number of convolution kernels is 3×3, the number of convolution kernels is 128, the number of channels is 100, the step size is 1, and the padding number is 1;

[0046] The output result is calculated by the pooling layer P2 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C6, C7, and C8, the number of convolution kernels is 3×3, the number of channels is 200, the step size is 1, and the padding number is 1;

[0047] The output result is calculated by the pooling layer P3 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C9, C10, and C11, the number of convolution kernels is 3×3, the number of channels is 400, the step size is 1, and the padding number is 1;

[0048] The output result is calculated by the pooling layer P4 for the maximum value within the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by four visual depth convolution layers C12, C13, C14, and C15, the number of convolution kernels is 3×3, the number of channels is 400, the step size is 1, and the padding number is 1;

[0049] The output result is calculated through the pooling layer P5 for the maximum value in the range, the number of pooling cores is 2×2, and the step size is 2; the output result is calculated through the fully connected layer F1, and the number of channels is 4000; the output result is calculated through the fully connected layer F2, and the number of channels is 1000;

[0050] The output result is calculated by the fully connected layer F3 with a channel number of 100; the output result is calculated by the fully connected layer F4 with a channel number of 5; the output result is calculated by the softmax function normalized exponential layer and the output confidence Rn (n = 1, 2, 3..., 5).

[0051] Specifically, in this embodiment, by establishing a flame image dataset of different combustibles in advance, a fire recognition neural network model (FireR-Net) based on deep learning is established and trained. The model has the ability to determine whether a fire has occurred and to obtain combustible information when a fire occurs, and the model is stored in a fire information processing system.

[0052] Furthermore, flame videos and images of different combustibles are obtained by downloading from the Internet, shooting flames, and recording flame tests, and the original training data set S is obtained by marking according to different combustibles. The marking types are divided into five categories: smokeless flame, smoke flame, smoke, light source, and background. Image processing is performed by image coding compression, image enhancement, image filling, random erasure, and image transformation to expand the original training data set S, and the training data set S is shuffled. The cross-validation method is used to divide the training data set S into K parts, and K-1 parts are selected in turn as training sets, and the remaining part is used as a validation set SV. Flame videos of different combustibles are obtained by recording flame tests, and marked to obtain a test data set ST. The data set S contains 30,000 labeled image information, and the test data set ST contains 3,000 labeled image information.

[0053] Furthermore, flame videos of different combustibles are obtained by recording flame tests, and the videos need to be recorded under different background environments, including shooting in different functional rooms indoors. To ensure the difference of the images, image information needs to be extracted from the videos. The method is to capture images of several frames in the video, and the capture interval is d frames.

[0054] Rule 1: d>12n, n is a positive integer and n∈[5,10]

[0055] Furthermore, the fire information processing system is trained by the training data set S and the deep learning model, and the ability of the fire information processing system to identify flames is tested by the test data set ST. If the judgment accuracy is less than 95%, the deep learning model is repeatedly optimized.

[0056] Further, set:

[0057] Rule 1:

[0058] Rule 2: h'(n) = mR(h(n))

[0059] Rule 3: mR(x) = max(ax,x), where a < -0.001

[0060] Rule 4:

[0061] Rule 5: F = WP + b

[0062] Rule 6:

[0063] Where f represents the image information matrix after compression and cropping of the original image, g represents the visual depth convolution layer matrix, h(n) represents the result of the convolution operation of f and g, mR(x) represents an mRelu activation function, which is used to remove overfitting and part of the convolution operation result. It is different from the traditional Relu function. mR(x) can retain the local subtle features of the image. h'(n) represents the result of the convolution operation after the activation function operation. The matrix P represents the operation result of the pooling layer. The pooling layer does not change the number of convolution layers. The role of the pooling layer is to increase the receptive field to prevent overfitting, reduce the amount of model calculations, and improve the calculation speed. Indicates that the matrix h'(n) is operated with a step length of 2, and the number of pooling cores is 2×2. The matrix F represents the calculation result of the fully connected layer, where W represents the weight matrix, b represents the bias matrix, and R n Represents confidence, which is obtained through softmax normalized exponential operation and is a common multi-classifier model.

[0064] The detailed structure of the fire recognition neural network model (FireR-Net) based on deep learning is shown in the following table:

[0065]

[0066]

[0067] Preferably, the fire identification neural network model includes a training process;

[0068] The training process includes a weight adjustment phase, the task of which is to adjust the parameters according to the deviation value of the loss function until the model converges and reaches the target prediction accuracy. The loss function is as follows:

[0069] Rule 1: L = L CE (y i ,y i * )+αL MSE (y i ,y i * )

[0070] Rule 2:

[0071] Rule 3:

[0072] Among them, L is the loss function and L ≥ 0, y is the true distribution, that is, the true label of the sample, y * is the predicted distribution, L CE is the cross entropy loss function, L MSE is the mean square error loss function, α is the weight value, in order to adjust the proportion of the mean square error loss function in the entire loss function, N is the number of training batches, and C is the number of categories.

[0073] Furthermore, the degree of gradient descent is judged based on whether the value of L is constantly approaching 0, and the deep learning model is forward-passed, and the RMSprop optimizer is used to adjust the weight matrix W.

[0074] Rule 1:

[0075] Rule 2:

[0076] where w i is the element in the weight matrix, θ i It is the defined value used to change w iThe gradient of , β is a hyperparameter, and γ is the learning rate. The difference between the RMSprop optimizer and the traditional SGD (stochastic gradient descent) optimizer is that the learning rate can be continuously adjusted, making it easier to find the minimum value of the loss function.

[0077] Preferably, the first confidence parameter is an output matrix Rn;

[0078] The determining whether a fire has occurred based on the first confidence parameter includes: if the maximum value max(Rn) in the output matrix Rn is equal to R1 or max(Rn) is equal to R2 or max(Rn) is equal to R3, it is determined that a fire has occurred; if the maximum value max(Rn) is equal to R4 or max(Rn) is equal to R5, it is determined that there is no fire.

[0079] Preferably, the preprocessing of the image information includes preprocessing the image information using a variable bounding box.

[0080] Furthermore, the features of the variable bounding box include three features: base point, scale, and width and height. The base point of the variable bounding box slides on the pixels of the image and performs dense sampling. Assume that the width and height of the input image are W and H, and each pixel of the image is used as the base point to generate bounding boxes of different sizes. The number of bounding boxes is Num.

[0081] Rule 1:

[0082] Rule 2:

[0083] Rule 3: Num = WH(n+m-1)

[0084] Where X is the width of the variable bounding box, Y is the height of the variable bounding box, ψ is the size ratio with n types, and s is the aspect ratio with m types. By controlling the position of the base point, the value of the size ratio and the value of the aspect ratio, Num bounding boxes of different sizes are obtained.

[0085] Furthermore, the images framed in the Num bounding boxes are input into the fire recognition neural network model (FireR-Net) based on deep learning for prediction. For each image, an output matrix Rn is obtained, which contains the confidence of each type of prediction result. If the maximum value of the output matrix max(Rn)=R4 or max(Rn)=R5, the bounding box is discarded. At the same time, a threshold λ is set. If max(Rn)<λ, the bounding box is discarded.

[0086] Furthermore, the Jaccard parameter is used to measure the similarity between variable bounding boxes, that is, the size of the intersection of the two divided by the size of the union of the two.

[0087] Rule 1:

[0088] Rule 2: J(μ,ν)∈(0,1)

[0089] Where J is the intersection-union ratio.

[0090] Furthermore, for the images framed by the remaining bounding boxes, the NMS non-maximum suppression method is used to search for local maxima to identify and locate the flame. The images framed by the remaining bounding boxes are sorted in descending order according to the confidence value to obtain a descending matrix K, and an empty preferred bounding box matrix L is set to store the preferred bounding boxes.

[0091] Rule 1: Compare the intersection-and-union ratio J of K1 and K2 in the descending matrix K. If J(μ,ν)>κ, K2 is discarded. If J(μ,ν)≤κ, K2 is retained.

[0092] Rule 2: Compare the intersection and union ratio J values ​​of K1 and K3 in the descending matrix K, repeat the operation of Rule 1 until the entire matrix K is traversed, and move K1 to the preferred bounding box matrix L.

[0093] Repeat rules 1 and 2 until all bounding box values ​​in the descending matrix K are transferred to the preferred bounding box matrix L. The preferred bounding box in L is the identified fire information and its location information, including the size of the variable frame and the coordinates of the base point.

[0094] Preferably, the flame information includes: the real length a, real width b, real height c of the flame, and the distance x between the flame and the floor; the characteristic parameters of the floor and the flame are combined to obtain a combined parameter X, and feature scaling is used to make the parameter value between [-1, 1].

[0095] Furthermore, the size information of the room where the fire occurred is obtained based on the BIM model information, and the length of the room is A, the width is B, and the height is C.

[0096] Furthermore, the true length a, true width b and true height c of the flame are obtained according to the size of the preferred bounding box in L, and a reduction factor is set.

[0097] Rule 1:

[0098] Rule 2:

[0099] Rule 3:

[0100] Furthermore, the real length, real width and real height of the flame are input into the BIM model, and the BIM model calculates the distance x from the flame to the floor.

[0101] Furthermore, a BIM-based floor fire damage model is established and trained. In the event of a fire, the flame transfers the temperature to the floor through heat conduction, heat radiation and heat convection. The floor is an important structural component in a building, used to directly support various objects in the building. In a fire, a large amount of heat accumulates on the top, and the floor is often severely damaged. We need to understand the damage of the floor in a timely manner after the fire and take reinforcement measures in a timely manner.

[0102] Furthermore, the floor slab contains a temperature sensor and a strain sensor, and the temperature value of the floor slab bottom is obtained through the temperature sensor, and the concrete strain, the surface steel bar strain and the steel bar strain at the bottom of the slab are obtained through the strain sensor.

[0103] Furthermore, in the BIM-based floor fire damage model, the BIM floor model contains the characteristic parameters of the floor, including the thickness of the protective layer of the floor, the concrete strength grade, the steel bar strength, the span ratio, the concrete strain, the steel bar strain on the plate surface, the steel bar strain on the bottom of the plate, the reinforcement ratio on the plate surface, the plate deflection, and the temperature on the bottom of the plate, which are used as parameters in the floor fire damage model, and then the flame characteristic parameters and the true length a, true width b, true height c of the flame, and the distance x between the flame and the floor are introduced. The characteristic parameters of the floor and the flame characteristic parameters are merged to obtain the merged parameter X, and the characteristic scaling is used to make the parameter value between [-1,1].

[0104] Furthermore, the dataset Z was obtained through literature review and experiments. The cross-validation method was used to divide the training dataset Z into K parts, and K-1 parts were selected as training sets in turn, and the remaining part was used as the validation set ZV. Flame videos of different combustibles were obtained by recording flame tests and marked to obtain the test dataset ZT.

[0105] Preferably, the inputting the flame information into a BIM-based floor fire damage model to predict the damage degree of the floor bearing capacity includes predicting the damage degree of the floor bearing capacity using a linear regression model:

[0106] Rule 1: f(x) = w T X+b

[0107] Rule 2: D = f(x) / Q

[0108] Where f(x) is the predicted floor bearing capacity, w T is the weight parameter, X is the merging parameter, b is the offset, Q is the initial bearing capacity strength of the floor before the fire, and D is the characteristic value of the bearing capacity damage degree.

[0109] Furthermore, the training process of the BIM-based floor fire damage model includes a weight adjustment stage, the task of which is to adjust the parameters according to the deviation value of the loss function, using the Huber Loss function, and the loss is as follows:

[0110] Rule 1:

[0111] Rule 2:

[0112] Where L is the loss function, y represents the true value, f(x) represents the predicted value, and δ is a hyperparameter.

[0113] Furthermore, the gradient descent method is used to adjust the weights of the model until convergence.

[0114] Rule 1:

[0115] Where w is the weight parameter, N is the number of training batches, and α is the learning rate.

[0116] Furthermore, the training of the BIM-based floor fire damage model is completed, and the degree of damage to the floor bearing capacity can be predicted.

[0117] Furthermore, the identified flame information parameters including the flame's real length a, real width b, real height c, the distance x between the flame and the floor, and the floor's BIM model parameters including the floor's protective layer thickness, concrete strength grade Q, steel bar strength q, slab span ratio λ, concrete strain ε1, slab surface steel bar strain ε2, slab bottom steel bar strain ε3, slab surface reinforcement ratio ρ, and slab deflection γ are input into the BIM-based floor fire damage model to obtain the floor bearing capacity damage degree characteristic value D.

[0118] Preferably, if a fire occurs, a fire alarm is issued, which also includes: triggering a fire alarm; triggering a solenoid valve in a fire sprinkler system to open a fire sprinkler head at the location where the fire occurs to extinguish the fire; and opening an escape sign on the escape route.

[0119] In addition, if Figure 2 As shown, the present invention also proposes a fire warning and fire fighting system that adopts the above-mentioned fire warning based on BIM and deep learning and post-disaster floor damage prediction method, characterized in that the system includes: a fire monitoring module based on a camera, an infrared thermal imager, a smoke detector, a video information storage device and a smoke information storage device; a fire information processing system; a computer intelligent terminal; a BIM model storage device; a visualization terminal; a cloud server; a fire fighting system control terminal, and a fire fighting rescue module based on a fire sprinkler system, a fire alarm, a broadcast and an escape sign.

[0120] Specifically, the computer intelligent terminal reads the building BIM model from the BIM model storage, and writes the fire information, combustion material type information, and camera number information obtained through the fire information processing system into the BIM model for integration.

[0121] Furthermore, the computer intelligent terminal reads the smoke information from the smoke information storage device and writes it into the BIM model for integration.

[0122] Furthermore, the computer intelligent terminal dispatches the fire protection system and plans the escape route according to the integrated BIM information, and forms fire protection instructions which are transmitted to the fire protection system control terminal.

[0123] Furthermore, the fire protection system control terminal receives instructions from the computer intelligent terminal and opens the fire sprinkler head where the fire occurs to extinguish the fire through the solenoid valve in the fire sprinkler system, and at the same time opens the fire sprinkler head where the smoke spreads, and controls the size of the sprinkler flow through the solenoid valve.

[0124] Furthermore, the fire protection system control terminal controls the opening of escape signs on the escape route to provide guidance for the evacuation of personnel.

[0125] Furthermore, the fire protection system control terminal starts broadcasting to provide guidance on the evacuation of personnel.

[0126] Furthermore, the computer intelligent terminal transmits the integrated BIM model to the visualization terminal for display. At this time, the integrated BIM model contains information such as building BIM model, fire information, type of combustible material information, camera number information, smoke information, smoke detector number information, fire system dispatch information and escape route planning information.

[0127] Furthermore, the visualization terminal transmits the visualization information to the cloud, and fire rescue personnel and escape personnel can check the fire situation and escape routes through mobile devices such as mobile phones.

[0128] Furthermore, escapees can connect to the cloud through mobile devices such as mobile phones, mark their locations in the BIM model, and send out distress messages. The BIM model will display the location information of the escapees, making it easier for firefighters to carry out rescue operations.

[0129] like Figure 2As shown, the present invention relates to a fire warning and fire protection system based on BIM and deep learning, including a fire information processing system, a BIM model information storage device, a computer intelligent terminal, a fire protection system control terminal, a fire sprinkler system, a video information storage device, a smoke information storage device, a visualization terminal, a cloud server, a mobile device, a camera, a smoke detector, an exhaust fan, a broadcast, an escape sign, and a fire smoke exhaust window. The camera and the smoke detector obtain the image and smoke information data in the building in real time; the fire information processing system obtains the image data through the video information storage device, obtains the smoke data through the smoke data storage device, and determines whether a fire occurs, the information of the combustible material, and the location of the fire according to the built-in deep learning algorithm; the computer intelligent terminal obtains the building BIM three-dimensional model from the BIM model storage device, obtains the fire information and location information from the fire information processing system, obtains the smoke information from the smoke data storage device, performs fire protection system scheduling and escape route planning, and at the same time integrates the fire information, smoke information, fire protection system scheduling and escape route planning information into the building BIM three-dimensional model; the fire protection system control terminal reads the computer intelligent terminal's scheduling information The fire alarm and the fire sprinkler system at the fire location are controlled by the speed command and the escape route planning command to accurately extinguish the fire. At the same time, the fire sprinkler head at the smoke spreading location is opened, the size of the sprinkler flow is controlled by the solenoid valve, and the escape sign and broadcast are controlled to help the on-site personnel escape. At the same time, the fire smoke exhaust window and exhaust fan are opened; the computer intelligent terminal inputs the integrated BIM three-dimensional model into the visualization terminal for visualization display; fire rescue personnel and escape personnel can check the fire situation and escape route through mobile devices such as mobile phones; escape personnel can connect to the cloud through mobile devices such as mobile phones, mark their location in the BIM model, and send out distress messages. The BIM model will display the location information of the escape personnel to facilitate the rescue of firefighters. The present invention realizes fire warning and fire rescue through image information and smoke information. Through images and deep learning neural network models, it can quickly and accurately determine whether a fire has occurred, obtain fire information and location information, and integrate the building BIM three-dimensional model with fire information, location information and smoke information. It can accurately dispatch the fire protection system in the building, accurately extinguish the fire in the place where the fire occurs, control the exhaust fan and fire smoke exhaust window to timely remove harmful smoke to reduce casualties, and can plan the escape route, control the escape sign and broadcast to indicate the escape route. At the same time, the fire dispatch information and escape route planning information are integrated into the BIM model for visual display, which helps firefighters to quickly and efficiently grasp the fire situation and deploy rescue plans, and helps escapers find escape routes to escape. The present invention can effectively improve the efficiency of building fire extinguishing and rescue, and reduce the loss of property and life caused by fire.

[0130] At least one camera must be installed in every room of the building, and the shooting range must cover every corner of the room. At the same time, several smoke detectors must be installed. Cameras and smoke detectors must be arranged at intervals in the public areas of the building to ensure that the shooting range can cover every corner of the room. The camera collects image information of different areas and stores it in the video information storage, and the smoke detector collects smoke information and stores it in the smoke information storage.

[0131] The building BIM information model is established using Autodesk's Revit software. The established model includes building walls, columns, beams, floor slabs, doors, windows, cameras, fire sprinkler systems and other components. The established BIM information model is stored in the BIM model storage.

[0132] A flame image dataset of different combustibles was established, and a fire recognition neural network model (FireR-Net) based on deep learning was established and trained. The model has the ability to determine whether a fire has occurred and to obtain combustible information when a fire occurs. The model is then stored in the fire information processing system.

[0133] The flame videos and images of different combustibles are obtained by downloading from the Internet and recording flame tests, and the original training data set S is obtained by annotating according to different combustibles; the original training data set S is expanded by image coding compression, image enhancement and restoration, and image transformation to obtain four sub-training data sets S1, S2, S3, and S4. The flame videos of different combustibles are obtained by recording flame tests, and the test data set T is obtained by annotating according to different combustibles. The fire information processing system is trained by the training data sets S, S1, S2, S3, and S4 and the deep learning model, and the ability of the information processing system to identify flames is tested by the test data set T. If the judgment accuracy is less than 95%, the deep learning model is repeatedly optimized.

[0134] The fire information processing system trained by the deep model can determine whether a fire has occurred, the type of burning material, and the camera number information, and transmit it to the computer intelligent terminal. The computer intelligent terminal reads the smoke information from the smoke information storage and writes it into the BIM model for fusion. The computer intelligent terminal reads the building BIM model from the BIM model storage, and writes the fire information, burning material type information, and camera number information obtained by the fire information processing system into the BIM model for fusion.

[0135] The computer intelligent terminal dispatches the fire system and plans the escape route according to the integrated BIM information, and forms a fire command to be transmitted to the fire system control terminal. The fire system control terminal receives the command from the computer intelligent terminal, opens the fire sprinkler head at the fire occurrence point through the solenoid valve in the fire sprinkler system to extinguish the fire, and opens the fire sprinkler head at the smoke spreading point at the same time, and controls the size of the sprinkler flow through the solenoid valve. The fire system control terminal controls the opening of the escape sign on the escape route to guide the evacuation of personnel. The fire system control terminal controls the opening of the broadcast to guide the evacuation of personnel.

[0136] The computer intelligent terminal transmits the integrated BIM model to the visualization terminal for display. At this time, the integrated BIM model contains the following information: building BIM model, fire information, type of burning material information, camera number information, smoke information, smoke detector number information, fire system dispatch information and escape route planning information. The visualization terminal transmits the visualization information to the cloud, and fire rescue personnel and escape personnel can view the fire situation and escape route through mobile devices such as mobile phones. The escape personnel can connect to the cloud through mobile devices such as mobile phones, mark their location in the BIM model, and send out a distress message. The BIM model will display the location information of the escape personnel, which is convenient for firefighters to rescue.

[0137] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the scope of protection determined by the claims of the present invention.

Claims

1. A fire warning and post-disaster floor damage prediction method based on BIM and deep learning, characterized in that: include: Collecting image information at different locations within the monitoring area, and preprocessing the image information to form a first data set; Inputting the image information of the first data set into the fire recognition neural network model to obtain a first confidence parameter; comprising: The original image size in the first dataset is compressed and cropped to 256×256×3, and convolution calculation is performed through two visual depth convolution layers C1 and C2, with the number of convolution kernels being 3×3, the number of channels being 100, the number of steps being 1, and the number of padding being 1; The output result is calculated by the pooling layer P1 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C3, C4, and C5, the number of convolution kernels is 3×3, the number of convolution kernels is 128, the number of channels is 100, the step size is 1, and the padding number is 1; The output result is calculated by the pooling layer P2 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C6, C7, and C8, the number of convolution kernels is 3×3, the number of channels is 200, the step size is 1, and the padding number is 1; The output result is calculated by the pooling layer P3 for the maximum value in the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by three visual depth convolution layers C9, C10, and C11, the number of convolution kernels is 3×3, the number of channels is 400, the step size is 1, and the padding number is 1; The output result is calculated by the pooling layer P4 for the maximum value within the range, the number of pooling kernels is 2×2, and the step size is 2; the output result is convolved by four visual depth convolution layers C12, C13, C14, and C15, the number of convolution kernels is 3×3, the number of channels is 400, the step size is 1, and the padding number is 1; The output result is calculated through the pooling layer P5 for the maximum value in the range, the number of pooling cores is 2×2, and the step size is 2; the output result is calculated through the fully connected layer F1, and the number of channels is 4000; the output result is calculated through the fully connected layer F2, and the number of channels is 1000; The output result is calculated by the fully connected layer F3, with a channel number of 100; the output result is calculated by the fully connected layer F4, with a channel number of 5; the output result is calculated by the softmax function normalized exponential layer, and the output confidence Rn (n = 1, 2, 3..., 5); The fire identification neural network model includes a training process; The training process includes a weight adjustment phase, the task of which is to adjust the parameters according to the deviation value of the loss function until the model converges and reaches the target prediction accuracy. The loss function is as follows: Rule 1: L = L CE (y i ,y i * ) + αL MSE (y i ,y i * ) Rule 2: Rule 3: Among them, L is the loss function and L ≥ 0, y is the true distribution, that is, the true label of the sample, y * is the predicted distribution, L CE is the cross entropy loss function, L MSE is the mean square error loss function, α is the weight value, in order to adjust the proportion of the mean square error loss function in the entire loss function, N is the number of training batches, and C is the number of categories; determining whether a fire occurs based on the first confidence parameter; and if a fire occurs, issuing a fire alarm; Identify flame information corresponding to the fire; input the flame information into a BIM-based floor fire damage model to predict the degree of damage to the floor bearing capacity.

2. The method for fire early warning and post-disaster floor damage prediction based on BIM and deep learning according to claim 1 is characterized in that: The first confidence parameter is an output matrix Rn; The determining whether a fire has occurred based on the first confidence parameter includes: if the maximum value max(Rn) in the output matrix Rn is equal to R1 or max(Rn) is equal to R2 or max(Rn) is equal to R3, it is determined that a fire has occurred; if the maximum value max(Rn) is equal to R4 or max(Rn) is equal to R5, it is determined that there is no fire.

3. The method for fire early warning and post-disaster floor damage prediction based on BIM and deep learning according to claim 1 is characterized in that: The preprocessing of the image information includes preprocessing the image information using a variable bounding box.

4. The method for fire early warning and post-disaster floor damage prediction based on BIM and deep learning according to claim 1, characterized in that: The flame information includes: the real length a, the real width b, the real height c of the flame, and the distance x between the flame and the floor; the characteristic parameters of the floor and the flame are combined to obtain a combined parameter X, and feature scaling is used to make the parameter value between [-1, 1].

5. The method for fire early warning and post-disaster floor damage prediction based on BIM and deep learning according to claim 1, characterized in that: The inputting of the flame information into the BIM-based floor fire damage model to predict the damage degree of the floor bearing capacity includes predicting the damage degree of the floor bearing capacity using a linear regression model: Rule 1: f(x) = w T X+b Rule 2: D = f(x) / Q Where f(x) is the predicted floor bearing capacity, w T is the weight parameter, X is the merging parameter, b is the offset, Q is the initial bearing capacity strength of the floor before the fire, and D is the characteristic value of the bearing capacity damage degree.

6. The method for fire early warning and post-disaster floor damage prediction based on BIM and deep learning according to claim 1, characterized in that: If a fire occurs, a fire alarm is issued, which also includes: triggering a fire alarm to alarm; triggering a solenoid valve in a fire sprinkler system to open a fire sprinkler head at the location where the fire occurs to extinguish the fire; and opening an escape sign on the escape path.

7. A fire warning and fire fighting system, used to implement the fire warning and post-disaster floor damage prediction method based on BIM and deep learning as described in any one of claims 1 to 6, characterized in that: The system includes: a fire monitoring module based on a camera, an infrared thermal imager, a smoke detector, a video information storage device and a smoke information storage device; a fire information processing system; a computer intelligent terminal; a BIM model storage device; a visualization terminal; a cloud server; a fire protection system control terminal, and a fire protection and rescue module based on a fire sprinkler system, a fire alarm, a broadcast and an escape sign.

Images