High-rise residential area fire-fighting site automatic generation design method based on deep learning

Abstract

The invention discloses an automatic generation design method for a high-rise residential area fire-fighting site based on deep learning. The method comprises the following steps: acquiring a data sample; cleaning and standardizing the data sample, and dividing the data sample into a real sample serving as a training set and a real sample serving as a test set; inputting the training set into a generator of a pix2pix model to generate an image, inputting the image and a real sample of the training set corresponding to the image into a discriminator at the same time, and outputting a discrimination result according to the authenticity of the image; the generator determines how to adjust own parameters by calculating a loss function according to feedback information of the discriminator so as to improve the quality and reliability of the generated image; inputting the test set into a trained pix2pix model, and generating a high-rise residential area fire-fighting site layout scheme image; and evaluating the high-rise residential area fire-fighting site layout scheme image and optimizing a pix2pix model. According to the invention, the fire-fighting site layout scheme can be efficiently and reasonably generated.

Description

Technical Field

[0001] The present invention relates to the technical field of building auxiliary design, and in particular to a method for realizing automatic generation and design of fire fighting sites in high-rise residential areas based on deep learning technology. Background Art

[0002] The overall layout design of high-rise residential areas is a complex system engineering, involving many factors, such as building spacing, functional zoning, traffic organization, and fire evacuation. Among them, the rational planning of fire evacuation sites is a crucial link. In the traditional architectural design process, the overall layout of high-rise residential areas mainly relies on the professional knowledge and experience of designers, and is designed through hand-drawn drawings and the use of relevant design software (such as AutoCAD, etc.). This traditional method will face the problem of low efficiency and susceptibility to the subjective factors of the designer, which leads to uneven quality of design solutions. In addition, with the continuous updating of building codes and standards, and the improvement of people's requirements for the safety of living environment, traditional design methods are unable to cope with the complex fire evacuation site planning requirements.

[0003] With the development of computer technology, some methods based on artificial intelligence technology have been applied to architectural design to improve work efficiency, such as the study of automatic generation of kindergarten campus layout based on deep learning, and the hybrid method of deep learning and ABM modeling to generate automated 2D building layout (this method first generates a bubble map that meets the topological conditions through agent-based modeling, and then converts the bubble map into a heat map through a rule algorithm, and finally uses the pix2pix algorithm to convert the heat map into a building space layout). There are also studies on the automatic generation of building layouts using deep learning and graphics algorithms (accurately segmenting building areas through the input-controlled spatial attention U-Net model, and refining the layout using image operations), etc. This shows that there are many studies on the use of artificial intelligence in the field of architecture, but there is a lack of methods for using deep learning models to automatically generate fire sites in high-rise residential areas. The reasonable planning of fire evacuation sites is an important part of architectural design layout. Therefore, the automatic generation design method of high-rise residential fire sites based on deep learning is of great significance. Summary of the invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide a method for automatically generating and designing fire sites in high-rise residential areas based on deep learning, which can more efficiently, accurately and flexibly generate a general plan layout that meets the fire evacuation requirements, thereby improving the design quality and efficiency.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for automatically generating and designing firefighting sites in high-rise residential areas based on deep learning, comprising the following steps: S1: Obtain data samples for high-rise residential site design; S2: Cleaning and standardizing the data samples, and dividing the data samples into real samples for training and real samples for testing. S3: Using the pix2pix model in the generative adversarial network as the core architecture, inputting the training set into the generator of the pix2pix model to generate an image, and simultaneously inputting the image and the real sample of the training set corresponding to it into the discriminator, and outputting the discrimination result according to the authenticity of the image; the generator determines how to adjust its own parameters by calculating the loss function according to the feedback information of the discriminator to improve the quality and reliability of the generated image; S4: inputting the test set into the trained pix2pix model to generate a fire site layout plan image of a high-rise residential area; S5: Evaluate and optimize the pix2pix model for the high-rise residential area fire site layout plan image.

[0006] Furthermore, S1 also includes a checklist for checking the data integrity of the data sample, the checklist includes scale markings, entrance and exit markings, location, area, and entrance and exit locations of the fire scene, and samples that do not meet the conditions of the checklist are judged as not meeting the requirements.

[0007] Furthermore, S1 also includes formulating labeling specifications and standards, and determining samples that do not meet the labeling standards as not meeting the sample collection standards.

[0008] Furthermore, the steps of cleaning and standardization treatment include: Remove noise and irrelevant information from sample images so that only elements related to building layout and firefighting sites are retained in the images; Resize sample images to a uniform size and resolution; The collected samples that meet the screening criteria are structured and coded.

[0009] Furthermore, the discriminator in S3 outputs a discrimination result according to the authenticity of the image: the discriminator outputs a probability value representing the authenticity of the image, and the value range is usually between 0 and 1. A value close to 0 indicates that the discriminator believes that the image is a generated fake image, and a value close to 1 indicates that the discriminator believes that the image is a real image.

[0010] Furthermore, the generator in S3 determines to adjust its own parameters by calculating the loss function based on the feedback information of the discriminator. When the loss function values ​​of the generator and the discriminator no longer change significantly in several consecutive rounds of training, or the loss function values ​​drop below a preset threshold, the model is considered to have reached a convergence state.

[0011] Furthermore, S2 also includes the steps of: The real samples corresponding to each training set are made to have only basic information: pictures of surrounding roads and sites are used as training set data together with the real samples; For each test set, the real samples are produced with only basic information: pictures of surrounding roads and venues as test set data.

[0012] Further, it also includes: using training set data with only basic information to generate a required high-rise residential fire site layout design; Using test set data with only basic information, the built-in encoder part is used to extract features from the image, compressing the image information into a feature vector containing key information of the site and surrounding roads to provide a basis for subsequent layout generation.

[0013] Furthermore, the S4 also includes, based on the extracted feature vector, referring to the layout patterns of many high-rise residential fire sites learned in the training phase through the decoder part of the generator, and gradually restoring the feature vector to image form through deconvolution or transposed convolution operations, while incorporating the planning principles of fire sites: according to the fire protection regulations, a safe distance must be maintained between fire sites and buildings. During the generation process, once the red line position of the site is detected, the fire site position is automatically planned at an appropriate distance; and according to the surrounding road information, it is ensured that the fire passage can be smoothly connected to the road, and the corresponding connection channel layout is generated.

[0014] Implementing the embodiments of the present invention has the following beneficial effects: (1) Efficiency: The present invention uses the deep learning pix2pix model to quickly and automatically generate fire site layout plans for high-rise residential areas, greatly shortening the design cycle and improving design efficiency. It can provide designers with multiple feasible design plans in a short period of time, meeting the requirements of the rapidly developing construction market for design efficiency.

[0015] (2) Reasonableness of fire evacuation: By studying a large number of actual cases and introducing the constraint loss of the fire evacuation site, the layout plan generated by the present invention can comprehensively consider various factors of the fire evacuation site, such as area, shape, location and relationship with buildings, etc., to ensure that residents can evacuate quickly and safely in emergency situations such as fire, effectively improve the fire safety performance of residential areas, and make up for the shortcomings of existing technologies in fire evacuation planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of elements included in data collection in an embodiment of the present invention; Figure 2 is a schematic diagram of a real sample of data preprocessing in an embodiment of the present invention; Figure 3 is a schematic diagram of a data preprocessing data input sample in an embodiment of the present invention; Figure 4 is a schematic diagram of a data input sample in an embodiment of the present invention; Figure 5 Schematic diagram of the pix2pix model of the present invention. DETAILED DESCRIPTION

[0017] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

[0018] Reference Figures 1 to 5 Schematic diagram shown.

[0019] A method for automatically generating and designing firefighting sites in high-rise residential areas based on deep learning in an embodiment of the present invention is implemented through the following steps.

[0020] S1: Data Collection Obtain high-rise residential site design data samples from urban open source data platforms (such as government websites).

[0021] During the acquisition process, a data sample integrity checklist is developed. The checklist lists the required data, including scale marking, entrance and exit marking, location of the fire site, area, entrance and exit location, etc. For each sample collected, check whether these data exist one by one. If any key data is missing, the sample is judged to be non-compliant. At the same time, marking specifications and standards are developed. The entrance and exit markings should clearly point to the actual entrance and exit locations. The marking text should be easy to identify and should not be vague, ambiguous or inconsistent with the actual situation. If the above requirements are not met, it will be judged as not meeting the sample collection standards. Samples that do not meet the sample collection conditions are as follows: Figure 1 shown.

[0022] S2: Data preprocessing The collected data samples are cleaned and standardized to remove noise and irrelevant information, text annotations, redundant lines, etc. in the sample images, so that the images only retain elements related to the building layout and fire-fighting sites: residential buildings, entrances and exits, internal and external roads, fire-fighting sites, and public activities.

[0023] The sample images are uniformly adjusted to a size and resolution of 512 pixels * 512 pixels, 300 pixels / inch, to meet the input requirements of the deep learning model. The collected samples that meet the screening conditions are then structured and encoded, converted into a form suitable for computer processing, and different color blocks are used to mark areas with different functions to distinguish residential buildings, entrances and exits, internal and external roads, fire sites, and public activities. Different color blocks have different RGB codes, so that the pix2pix model can learn the regional functional characteristics represented by different RGB codes (i.e., different color blocks), so as to better understand and generate fire site layouts that meet the requirements: the fire site layout images generated by the generator will be compared with these annotated real sample images; the discriminator will judge the authenticity of the generated image based on this comparison, and the generator will continuously adjust its own parameters based on the feedback information of the discriminator. In this way, sample images with different color blocks annotated provide clear guidance for model training, which helps the model learn the correct fire site layout mode and improve the quality and accuracy of the images generated by the model.

[0024] The processed images include all samples of residential buildings, entrances and exits, internal and external roads, fire sites, and public activities, which are real samples.

[0025] 60% of them are taken as the real samples of the training set, and 40% are taken as the real samples of the test set.

[0026] The real samples corresponding to each training set are made with only basic information: surrounding roads and venues (such as Figure 3 ) together with the real samples as the training set data input samples (such as Figure 4 ).

[0027] For each test set, the real samples are created with only basic information: pictures of surrounding roads and venues are used as test set data input samples.

[0028] S3: pix2pix model construction and training The present invention adopts the pix2pix model in the generative adversarial network (GAN) as the core architecture (e.g. Figure 5 ). The pix2pix model mainly consists of two parts: the generator and the discriminator.

[0029] The network structure of the generator adopts an encoder-decoder architecture. The encoder part consists of multiple convolutional layers, which are used to extract the feature information of the input image. Each convolutional layer contains a certain number of convolution kernels (such as 64, 128, 256, etc.), the size of the convolution kernel is 3x3 or 4x4, and the step size is 2. The image resolution is gradually reduced through the convolution operation, while the number of feature channels is increased, thereby extracting the high-level semantic features of the image.

[0030] The decoder part consists of multiple deconvolutional layers. The deconvolution formula is:

[0031] in: Represents the index position of the element in the deconvolution kernel matrix, i, j represents the coordinates of the output feature map Y, k h and k w is the height and width of the convolution kernel, s is the stride, Indicates rounding down, X p,q It is the pixel value of the input feature map, which is usually filled with zero (i.e. zero-padding) when it is out of range.

[0032] The function of the deconvolution layer is to gradually restore the feature information extracted by the encoder to a general layout image with the same size as the target image. The convolution kernel size and step size of the deconvolution layer are set opposite to those of the convolution layer. The deconvolution operation gradually restores the resolution of the image, while reducing the number of feature channels, and finally generates a general layout image of a high-rise residential area with a reasonable layout of fire evacuation sites. In order to further improve the performance of the generator, the encoder and decoder directly transfer the input low-level features through skip connections. This design can effectively improve the detail quality of the generated image. The goal of the generator is to generate a result that is as close to the target image as possible. Its loss function consists of adversarial loss and L1 loss: (1) Adversarial Loss:

[0033] Where: x is the input image, y is the target image, G(x,z) is the generator output, and z is random noise.

[0034] (2) L1 Reconstruction Loss:

[0035] Where: G(x) represents the image generated by the generator, based on the input image x and random noise z.

[0036] (3) Joint loss function:

[0037] in: is a weight hyperparameter used to balance the two losses.

[0038] The network structure of the discriminator is also based on a convolutional neural network. Its purpose is to determine whether the input image is a real high-rise residential area master plan layout case image or a fake image generated by the generator. This process optimizes the following objectives through an adversarial loss function:

[0039] in: represents the loss function of the discriminator, D(x,y) represents the output value of the discriminator, that is, the probability that the input (x,y) is a real image, G(x,z) represents the output of the generator, that is, the image generated based on the input x and random noise z, E x,y represents the expectation of the real samples x and y, E x,z represents the expectation of the input x and random noise z.

[0040] The discriminator uses the PatchGAN discriminator architecture. The discriminator of PatchGAN calculates the discrimination result of the local image area. The formula can be expressed as . Where: W is the convolution kernel weight, represents the convolution operation, ⊕ represents the concatenation of the input and target images, b is the bias term, and σ is the activation function (such as Leaky ReLU). After PatchGAN extracts the feature information of the image through the convolution layer, it performs classification judgment through the fully connected layer and outputs a probability value between 0 and 1, indicating the possibility that the input image is a real image.

[0041] During the training process, the training set data input samples prepared in the data preprocessing step are first input into the generator. Based on the training set data input samples with only basic information, the generator will add other color blocks to generate images, which is the required high-rise residential fire site layout design. Then, the generated image and the corresponding training set real samples are simultaneously input into the discriminator. The discriminator outputs the discrimination result based on the authenticity of the image, and the mean square error is used to calculate the discriminant result. and structural similarity To measure the quality of the generated image. In the MSE formula, N represents the number of samples or the total number of pixels, y i represents the pixel value of the i-th real image, represents the pixel value of the i-th generated image, Represents the square error between the true value and the predicted value. In the SSIM formula: C1 and C2 are parameters to be adjusted, and their settings depend on the dynamic range of the image (such as the maximum value of the image pixel value), X represents the generated image (the image to be evaluated), and y represents the true image (the target image) μ x , μ y represents the mean of images X and y, represents the variance of images X and y, σ x,y Represents the covariance of images X and y.

[0042] After that, the discriminator outputs a probability value indicating the authenticity of the image, which usually ranges from 0 to 1. A value close to 0 means that the discriminator believes that the image is a fake image, and a value close to 1 means that the discriminator believes that the image is a real image. The generator determines how to adjust its own parameters based on the feedback information from the discriminator by calculating the loss function to improve the quality and reliability of the generated image. The whole process can be regarded as the training of the model through the alternating optimization of the generator and the discriminator. At each iteration: Update the parameters of the discriminator :

[0043] Update the parameters of the generator :

[0044] in: represents the parameters (weights) of the discriminator, represents the parameters (weights) of the generator, represents the loss function of the discriminator, The total loss function of the generator, Represents the learning rate.

[0045] When the loss function values ​​of the generator and the discriminator no longer change significantly in several consecutive rounds of training, or when the loss function value drops below a preset threshold, the model is considered to have reached a convergence state, that is, the model training is completed. The generated fire site layout image can also be manually evaluated to check whether the fire site layout in the image meets the specifications and whether it is practical. When the manual evaluation results recognize the quality and rationality of the generated image, the model training can also be considered to have met the requirements.

[0046] S4: Automatic generation of fire site layout After completing the training of the model, the data input samples of the test set previously prepared in the data preprocessing step are fed into the trained pix2pix model.

[0047] The generator will first receive the basic information in the test set, use its built-in encoder part to extract features from the image, and compress the image information into a feature vector containing key information such as the site and surrounding roads to provide a basis for subsequent layout generation.

[0048] Based on the extracted feature vectors, the decoder part of the generator will refer to the layout patterns of many high-rise residential fire sites learned in the training phase, and gradually restore the feature vectors to image form through a series of deconvolution or transposed convolution operations. At the same time, the planning principles of fire sites are incorporated: According to fire regulations, a certain safety distance must be maintained between fire sites and buildings. During the generation process, once the red line position of the site is detected, the fire site location is automatically "planned" at an appropriate distance; for example, after understanding the surrounding road information, ensure that the fire passage can be smoothly connected to the road and generate the corresponding connection channel layout; in addition, it is required that the net width of the fire lane should not be less than 4 meters, the size of the fire climbing site should not be less than 15 meters × 10 meters, and the distance from the outer wall of the building should not be greater than 10 meters.

[0049] According to the trained pix2pix model, after inputting the data sample of the test set, the model will automatically generate an image of the layout plan of the fire protection site in a high-rise residential area.

[0050] If the generated image is not clear enough, you can draw the generated image to enhance the clarity of the image. At the same time, according to the actual design needs, for example, the boundaries of the fire evacuation site should be accurately defined to ensure that it meets the actual requirements; the generated fire passage layout needs to consider the turning radius of the fire truck, and the evacuation site area should be calculated according to the estimated number of residents and the per capita area standard. According to these actual needs, the generated building and site elements are refined and optimized.

[0051] S5: Solution evaluation and optimization In order to evaluate the feasibility of the generated fire site layout plan for high-rise residential areas, the following evaluation indicators should be set for scheme evaluation: fire evacuation distance index (the distance from any point in the building to the nearest fire evacuation site boundary should meet the maximum evacuation distance requirements specified in the fire code), fire evacuation site area index (ensure that the area of ​​the evacuation site is not less than the minimum area requirement calculated based on the building scale and population density), building layout compactness index (measure the space utilization efficiency between buildings to avoid too loose or crowded layout), traffic smoothness index (evaluate the connectivity and traffic capacity of the internal road network of the residential area to ensure that fire trucks and personnel can be evacuated quickly in an emergency) and coordination index with the surrounding environment (consider the integration of the building layout and fire evacuation site with the surrounding natural environment and urban landscape), etc.

[0052] The generated plan is optimized based on the evaluation results: if the fire evacuation distance does not meet the requirements, the evacuation distance can be shortened by adjusting the location of the building or the layout of the evacuation site; if the evacuation site area is insufficient, the site scope can be appropriately expanded or the layout of the building can be adjusted to increase the available evacuation space; if there is a problem with traffic flow, the road network can be optimized, such as increasing road width, setting up roundabouts, etc.

[0053] The optimization process can be done by manually adjusting the generated solution, or by feeding the evaluation results back to the pix2pix model to fine-tune the model so that it can generate a layout solution that better meets the requirements.

[0054] The above disclosure is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for automatically generating and designing firefighting sites in high-rise residential areas based on deep learning, characterized in that: The following steps are involved: S1: Obtain data samples for high-rise residential site design; S2: Cleaning and standardizing the data samples, and dividing the data samples into real samples for training and real samples for testing. S3: using the pix2pix model in the generative adversarial network as the core architecture, inputting the training set into the generator of the pix2pix model to generate an image, and simultaneously inputting the image and the real sample of the training set corresponding to it into the discriminator, and outputting the discrimination result according to the authenticity of the image; The generator determines how to adjust its own parameters based on the feedback from the discriminator by calculating the loss function to improve the quality and reliability of the generated image. S4: inputting the test set into the trained pix2pix model to generate a fire site layout plan image of a high-rise residential area; S5: Evaluate and optimize the pix2pix model for the high-rise residential area fire site layout plan image.

2. The method for automatically generating and designing high-rise residential firefighting sites based on deep learning according to claim 1 is characterized in that: The S1 also includes a checklist for checking the data integrity of the data sample, the checklist includes scale markings, entrance and exit markings, the location, area, and entrance and exit locations of the fire scene, and samples that do not meet the conditions of the checklist are judged as not meeting the requirements.

3. The method for automatically generating and designing high-rise residential firefighting sites based on deep learning according to claim 2 is characterized in that: The S1 also includes formulating labeling specifications and standards, and determining that samples that do not meet the labeling standards do not meet the sample collection standards.

4. The method for automatically generating and designing high-rise residential firefighting sites based on deep learning according to claim 1 is characterized in that: The steps of cleaning and standardization include: Remove noise and irrelevant information from sample images so that only elements related to building layout and firefighting sites are retained in the images; Resize sample images to a uniform size and resolution; The collected samples that meet the screening criteria are structured and coded.

5. The method for automatically generating and designing firefighting sites in high-rise residential areas based on deep learning according to claim 1 is characterized in that: The discriminator in S3 outputs a discrimination result according to the authenticity of the image: the discriminator outputs a probability value indicating the authenticity of the image, and the value range is between 0 and 1. A value close to 0 indicates that the discriminator believes that the image is a generated fake image, and a value close to 1 indicates that the discriminator believes that the image is a real image.

6. The method for automatically generating and designing high-rise residential firefighting sites based on deep learning according to claim 5 is characterized in that: The generator in S3 determines to adjust its own parameters by calculating the loss function based on the feedback information of the discriminator. When the loss function values ​​of the generator and the discriminator no longer change significantly in several consecutive rounds of training, or when the loss function values ​​drop below a preset threshold, the model is considered to have reached a convergence state.

7. The method for automatically generating and designing high-rise residential firefighting sites based on deep learning according to claim 1 is characterized in that: The S2 also includes the steps of: The real samples corresponding to each training set are made to have only basic information: pictures of surrounding roads and sites are used as training set data together with the real samples; For each test set, the real samples are produced with only basic information: pictures of surrounding roads and venues as test set data.

8. The deep learning-based automatic generation and design method for high-rise residential fire protection sites according to claim 7 is characterized in that: Also includes: Use training set data with only basic information to generate the required high-rise residential fire site layout design; Using test set data with only basic information, the built-in encoder part is used to extract features from the image, compressing the image information into a feature vector containing key information of the site and surrounding roads to provide a basis for subsequent layout generation.

9. The deep learning-based automatic generation and design method for high-rise residential fire protection sites according to claim 8 is characterized in that: The S4 also includes, based on the extracted feature vector, referring to the layout patterns of many high-rise residential fire sites learned in the training phase through the decoder part of the generator, and gradually restoring the feature vector to image form through deconvolution or transposed convolution operations, while incorporating the planning principles of fire sites: according to the fire regulations, a safe distance must be maintained between fire sites and buildings. During the generation process, once the red line position of the site is detected, the fire site position is automatically planned at an appropriate distance; and according to the surrounding road information, it is ensured that the fire passage can be smoothly connected to the road, and the corresponding connection channel layout is generated.

Images