Machine learning-based building scheme generation method for automatically splitting single building of refrigeration house

Abstract

The invention discloses a method for generating a single building scheme of a refrigeration house automatically split building based on machine learning. The method comprises the following steps: S1, acquiring a data set of a refrigeration house building and carrying out standardization processing; s2, extracting feature vectors of the building by using a self-attention mechanism; s3, constructing a generative adversarial network to generate a preliminary building splitting scheme; s4, further optimizing the generated preliminary building splitting scheme by using a self-attention mechanism to obtain an optimized building splitting scheme; s5, the optimized building splitting scheme is matched with the function partition requirements of the refrigeration house, all the function areas are reasonably distributed in the layout after splitting, and a final building splitting scheme is obtained; and S6, converting the final building splitting scheme into a planar graph. According to the method, an efficient and scientific optimization scheme can be provided in the design of the automatic split building monomers of the refrigeration house, and remarkable technical values and economic benefits are brought to practical application.

Description

Technical Field

[0001] The present invention relates to the fields of architectural design and artificial intelligence, and in particular to a method for generating a building plan for automatically splitting a cold storage building based on machine learning. Background Art

[0002] As an important facility of cold chain logistics, cold storage is widely used in food, medicine, chemical industry and other fields. The design of cold storage directly affects its storage efficiency, operating cost and energy consumption. In the cold storage design process, space division and functional zoning are key factors that determine the design effect and utilization rate. However, traditional cold storage design methods generally rely on manual experience for space layout and division, which leads to problems such as long design cycle, poor flexibility and low space utilization, especially in cold storage with large-scale and high-density storage requirements.

[0003] With the rapid development of artificial intelligence (AI) technology, especially the maturity of deep learning technology, more and more intelligent design methods have begun to be applied in the field of architectural design. Generative adversarial networks and self-attention mechanisms are AI technologies that have been widely used in the field of architecture in recent years. Generative adversarial networks can generate realistic and design-compliant architectural plans through adversarial training of generators and discriminators. The self-attention mechanism can help the model capture important and interrelated features in spatial layout and design, and optimize the accuracy of building segmentation.

[0004] Nevertheless, the research and application of generative adversarial networks and self-attention mechanisms for cold storage building design are still in their infancy. No existing technology can provide a complete solution for automatically splitting building units for cold storage based on generative adversarial networks and self-attention mechanisms. In addition, the feature vectors extracted using traditional attention mechanisms often contain noise, and the cold storage building splitting solutions generated by them are limited in effectiveness.

[0005] In summary, the traditional cold storage design method has certain limitations in terms of space division and functional requirements matching. Especially in the scenario of large cold storage and high-density storage requirements, manual design cannot quickly generate a partitioning scheme that meets the specifications. At the same time, although the existing attention mechanism can extract feature vectors, the effect is poor. Therefore, a new method is urgently needed to overcome the above defects, realize the automatic splitting of cold storage building units, and further optimize the space utilization of the scheme. Summary of the invention

[0006] One purpose of the present invention is to propose a method for generating building plans for automatically splitting building units of cold storage based on machine learning. The present invention can provide an efficient and scientific optimization solution in the generation of building plans for automatically splitting building units of cold storage, bringing significant technical value and economic benefits to practical applications.

[0007] According to an embodiment of the present invention, a method for generating a cold storage building plan based on machine learning for automatic splitting of buildings comprises the following steps:

[0008] S1, obtain the dataset of cold storage buildings and perform standardization;

[0009] S2. Use the self-attention mechanism to analyze the building's plane dimensions, structural form, and fire zoning requirements, identify key spatial splitting features, and generate feature vectors for building splitting;

[0010] S3. Construct a generative adversarial network. The generator generates a spatial splitting solution by inputting the feature vector of building splitting. The discriminator is used to evaluate the rationality and conformity of the generated solution. The generator is trained through the generative adversarial network to generate multiple preliminary building splitting solutions.

[0011] S4. Use the self-attention mechanism to further optimize the generated preliminary building splitting scheme, adjust the allocation ratio of each area to meet the functional requirements of each area, and obtain the optimized building splitting scheme;

[0012] S5. Match the optimized building splitting scheme with the functional zoning requirements of the cold storage, so that all functional areas are reasonably allocated in the split layout, and obtain the final building splitting scheme;

[0013] S6. Convert the final building split plan into a floor plan, draw the distribution of each functional area, provide corresponding structural descriptions, and mark the fire zones.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Obtain the length L, width W, total area A, number of floors N, total height H, column spacing C, and fire partition area requirement A = {A1, A2, ..., A p}, the area requirement of the refrigerated area is F1, the area requirement of the frozen area is F2, the area requirement of the channel area is F3, and the characteristic vector V is obtained by normalization.

[0016] Optionally, S2 includes the following steps:

[0017] S21. Perform linear transformation on the eigenvector V:

[0018] Q=W q V;

[0019] K=W k V;

[0020] V=W v V;

[0021] Among them, Q is the query vector, K is the key vector, V is the value vector, and Wq , W v , W v are the weight matrices for query, key, and value respectively;

[0022] Perform weighted summation on the query vector, key vector, and value vector obtained to obtain the attention result after attention:

[0023]

[0024] Among them, softmax is the normalization function, d k is the query and key dimension, V ’ for the attentional outcome after attention;

[0025] S22. One limitation of the attention mechanism is that the attention result is always obtained through weighted summation. When the model lacks the context to be paid attention to, the attention result generated will be a noise vector.

[0026] Therefore, the superposition attention mechanism is further used to optimize the attention results generated by S21. The superposition attention mechanism generates information vectors and attention gates to filter the attention results after attention:

[0027]

[0028]

[0029] in, is the weight matrix, b I 、b D is the bias term, I is the information vector, and D is the attention gate;

[0030]

[0031] Among them, ⊙ represents the bitwise multiplication, Represents the final feature vector obtained by the stacked attention mechanism.

[0032] Optionally, S3 includes the following steps:

[0033] S31. Build a generative adversarial network, including two networks: the generator and the discriminator;

[0034] S32, the final feature vector obtained by superimposing the attention mechanism Input to generator G:

[0035]

[0036] Among them, S genrepresents a space splitting scheme generated by a generator, wherein the space splitting scheme includes the size and position of each functional area in the cold storage;

[0037] S33. Input the space splitting scheme generated by the generator into the discriminator to obtain a rationality score:

[0038] D(S gen )=σ(W d S gen +b d );

[0039] Among them, σ is the Sigmoid activation function, W d is the weight matrix of the discriminator, b d is the bias term, D(S gen ) is the rationality score of the discriminator output;

[0040] S34, the generator and the discriminator are continuously optimized through the adversarial training process. The discriminator evaluates the rationality of the splitting scheme generated by the generator and reversely transfers the error to the generator, thereby promoting the generator to generate a splitting scheme that is more in line with the specification;

[0041] The loss functions of the generator and discriminator during training are:

[0042]

[0043]

[0044] S35. Update the parameters of the generator G and the discriminator D through the back propagation algorithm, use gradient descent to calculate the gradient of the loss function to the network parameters, and update the model parameters:

[0045]

[0046]

[0047] Among them, η is the learning rate, θ G and θ D are the parameters of the generator and the discriminator respectively;

[0048] S36. After T rounds of training, use the generator to output a preliminary building splitting solution S initial .

[0049] Optionally, S4 includes the following steps:

[0050] S41. Preliminary building split plan S initial Perform linear transformation to obtain the query matrix Q, key matrix K and value matrix V, and calculate the self-attention weight matrix W attention :

[0051]

[0052] S42. Obtain the optimized building splitting solution S through weighted summation optimized :

[0053] S optimized =W attention V;

[0054] S43, by iterating S41 and S42, the generated building splitting scheme is optimized so that the proportion and layout of each functional area are maximized and the optimized building splitting scheme S' is obtained optimized .

[0055] Optionally, S5 includes the following steps:

[0056] S51, optimize the building splitting plan S' optimized The area and spatial layout information of each region in is extracted and recorded as S regions = {R1, R2, R3}, where R i (i∈{1, 2, 3}) represents the information of each functional area in the building splitting scheme;

[0057] S52. Using R i Further obtain the area information A of each functional area i and location information L i ;

[0058] S53, calculate the error value E of each area i :

[0059] E i =|A i -A i_target ||L i -L i_target |;

[0060] Among them, A i_target Indicates the target area in the corresponding functional zoning requirements, L i_target Indicates the spatial layout requirements of the corresponding functional zones;

[0061] S54. For each region R i , by minimizing the error value E i To adjust the area and layout information:

[0062]

[0063] S55. After the final matching of the functional zoning and the building splitting scheme is completed, the final building splitting scheme S is output. final , including the final layout, area and location arrangement of each functional area.

[0064] Optionally, the S6 includes the following steps:

[0065] S61. Final building split plan S final Each region R i Perform spatial layout transformation to obtain the plane coordinate information C of each functional area i =(x i ,y i ), where x i ,y i Respectively represent functional areas R i The horizontal and vertical coordinate positions in a two-dimensional plane;

[0066] S62, according to the plane coordinates C i and region R i Area requirement A i , draw the outer boundary of each area on the design drawing to form the outline of each functional area, and calculate the geometric center G of each area i and the boundary position B i To draw the specific position relationship;

[0067] S63. According to the structural requirements of the cold storage building, the structural characteristic data is transferred to the plan view, and important structural elements such as the bearing structure, column spacing and load-bearing walls of each area are further drawn;

[0068] S64. Mark each functional area in the plan and mark the fire partition area on the drawing according to the fire partition specification. Fire partition area D fire = {d1, d2, d3}, representing the boundaries and areas of each fire zone in the building;

[0069] S65. After marking the functional categories and fire zones of each area, conduct fire protection compliance verification to ensure that the area requirements of each fire zone match the fire protection requirements of the cold storage;

[0070] S66, Generate a complete floor plan file G final The file contains the layout of all functional areas, structural descriptions and fire partition markings. The file uses standard CAD format to facilitate further construction or verification.

[0071] The beneficial effects of the present invention are:

[0072] (1) The method for generating a building plan for automatically splitting a cold storage building based on machine learning can effectively overcome the problems of low space utilization, low design efficiency, and poor flexibility in the prior art. First, by standardizing the data of the cold storage building and converting it into a unified standardized coordinate system, accurate and unified input is provided for the subsequent generation of building splitting plans. This standardization method ensures that the data of different cold storage buildings can be compared and analyzed under the same framework, thereby avoiding the deviation caused by improper data processing in traditional design methods.

[0073] (2) The self-attention mechanism is used to comprehensively analyze multiple design constraints such as building plan size, structural form, and fire partitions, and can extract key spatial splitting features from multiple dimensions. By capturing the correlation between the various parts of the building, the mechanism can effectively optimize the spatial layout so that the functional requirements of each area are met while maximizing the utilization of the space. This global optimization capability is difficult to achieve in traditional design. Through the self-attention mechanism, the allocation ratio of each area can be accurately adjusted to ensure that the functional areas in the cold storage are reasonably laid out. Based on the feature vector generated by the attention mechanism, the superposition attention mechanism can filter out irrelevant noise vectors, thereby providing a higher quality building feature vector for the generative adversarial network.

[0074] (3) By constructing a generative adversarial network, the generator can generate multiple possible spatial splitting schemes based on the input architectural data, while the discriminator evaluates the rationality of these schemes and continuously optimizes the output of the generator. Compared with traditional design methods, constructing a generative adversarial network can not only automatically generate splitting schemes that meet the requirements, but also quickly iterate and optimize the design, improve design efficiency, and reduce the time and cost of manual intervention. At the same time, the adversarial training mechanism of the generative adversarial network can effectively avoid the generation of unreasonable or non-standard schemes, thereby ensuring the reliability of the design results. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0076] Figure 1 A flowchart of a method for generating a building plan for automatically splitting a cold storage building based on machine learning proposed by the present invention;

[0077] Figure 2 This is a multi-round generative adversarial network training diagram in the method for generating a building plan for automatically splitting a cold storage building based on machine learning proposed by the present invention. DETAILED DESCRIPTION

[0078] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0079] refer to Figure 1-Figure 2 , a method for generating a cold storage building plan based on machine learning, comprising the following steps:

[0080] S1, obtain the dataset of cold storage buildings and perform standardization;

[0081] S2. Use the self-attention mechanism to analyze the building's plane dimensions, structural form, and fire zoning requirements, identify key spatial splitting features, and generate feature vectors for building splitting;

[0082] S3. Construct a generative adversarial network. The generator generates a spatial splitting solution by inputting the feature vector of building splitting. The discriminator is used to evaluate the rationality and conformity of the generated solution. The generator is trained through the generative adversarial network to generate multiple preliminary building splitting solutions.

[0083] S4. Use the self-attention mechanism to further optimize the generated preliminary building splitting scheme, adjust the allocation ratio of each area to meet the functional requirements of each area, and obtain the optimized building splitting scheme;

[0084] S5. Match the optimized building splitting scheme with the functional zoning requirements of the cold storage, so that all functional areas are reasonably allocated in the split layout, and obtain the final building splitting scheme;

[0085] S6. Convert the final building split plan into a floor plan, draw the distribution of each functional area, provide corresponding structural descriptions, and mark the fire zones.

[0086] In this implementation, S1 includes the following steps:

[0087] S11. Obtain the length L, width W, total area A, number of floors N, total height H, column spacing C, and fire partition area requirement A = {A1, A2, ..., A p}, the area requirement of the refrigerated area is F1, the area requirement of the frozen area is F2, the area requirement of the channel area is F3, and the characteristic vector V is obtained by normalization.

[0088] In this implementation, S2 includes the following steps:

[0089] S21. Perform linear transformation on the eigenvector V:

[0090] Q=W q V;

[0091] K=W kV;

[0092] V=W v V;

[0093] Among them, Q is the query vector, K is the key vector, V is the value vector, and W q , W v , W v are the weight matrices for query, key, and value respectively;

[0094] Perform weighted summation on the query vector, key vector, and value vector obtained to obtain the attention result after attention:

[0095]

[0096] Among them, softmax is the normalization function, d k is the query and key dimension, V ’ for the attentional outcome after attention;

[0097] S22. One limitation of the attention mechanism is that the attention result is always obtained through weighted summation. When the model lacks the context to be paid attention to, the attention result generated will be a noise vector.

[0098] Therefore, the superposition attention mechanism is further used to optimize the attention results generated by S21. The superposition attention mechanism generates information vectors and attention gates to filter the attention results after attention:

[0099]

[0100] in, is the weight matrix, b I 、b D is the bias term, I is the information vector, and D is the attention gate;

[0101]

[0102] Among them, ⊙ represents the bitwise multiplication, Represents the final feature vector obtained by the stacked attention mechanism.

[0103] In this implementation, S3 includes the following steps:

[0104] S31. Build a generative adversarial network, including two networks: the generator and the discriminator;

[0105] S32, the final feature vector obtained by superimposing the attention mechanism Input to generator G:

[0106]

[0107] Among them, S gen represents a space splitting scheme generated by a generator, wherein the space splitting scheme includes the size and position of each functional area in the cold storage;

[0108] S33. Input the space splitting scheme generated by the generator into the discriminator to obtain a rationality score:

[0109] D(S gen )=σ(W d S gen +b d );

[0110] Among them, σ is the Sigmoid activation function, W d is the weight matrix of the discriminator, b d is the bias term, D(S gen ) is the rationality score of the discriminator output;

[0111] S34, the generator and the discriminator are continuously optimized through the adversarial training process. The discriminator evaluates the rationality of the splitting scheme generated by the generator and reversely transfers the error to the generator, thereby promoting the generator to generate a splitting scheme that is more in line with the specification;

[0112] The loss functions of the generator and discriminator during training are:

[0113]

[0114] S35. Update the parameters of the generator G and the discriminator D through the back propagation algorithm, use gradient descent to calculate the gradient of the loss function to the network parameters, and update the model parameters:

[0115]

[0116]

[0117] Among them, η is the learning rate, θ G and θ D are the parameters of the generator and the discriminator respectively;

[0118] S36. After T rounds of training, use the generator to output a preliminary building splitting solution S initial .

[0119] In this implementation, S4 includes the following steps:

[0120] S41. Preliminary building split plan S initial Perform linear transformation to obtain the query matrix Q, key matrix K and value matrix V, and calculate the self-attention weight matrix W attention :

[0121]

[0122] S42. Obtain the optimized building splitting solution S through weighted summation optimized :

[0123] S optimized =W attention V;

[0124] S43, by iterating S41 and S42, the generated building splitting scheme is optimized so that the proportion and layout of each functional area are maximized and the optimized building splitting scheme S' is obtained optimized .

[0125] In this implementation, S5 includes the following steps:

[0126] S51, optimize the building splitting plan S' optimized The area and spatial layout information of each region in is extracted and recorded as S regions = {R1, R2, R3}, where R i (i∈{1, 2, 3}) represents the information of each functional area in the building splitting scheme;

[0127] S52. Using R i Further obtain the area information A of each functional area i and location information L i ;

[0128] S53, calculate the error value E of each area i :

[0129] E i =|A i -A i_target ||L i -L i_target |;

[0130] Among them, A i_target Indicates the target area in the corresponding functional zoning requirements, L i_target Indicates the spatial layout requirements of the corresponding functional zones;

[0131] S54. For each region R i , by minimizing the error value E i To adjust the area and layout information:

[0132]

[0133] S55. After the final matching of the functional zoning and the building splitting scheme is completed, the final building splitting scheme S is output. final , including the final layout, area and location arrangement of each functional area.

[0134] In this implementation, S6 includes the following steps:

[0135] S61. Final building split plan S final Each region R i Perform spatial layout transformation to obtain the plane coordinate information C of each functional area i =(x i ,y i ), where x i ,y i Respectively represent functional areas R i The horizontal and vertical coordinate positions in a two-dimensional plane;

[0136] S62, according to the plane coordinates C i and region R i Area requirement A i , draw the outer boundary of each area on the design drawing to form the outline of each functional area, and calculate the geometric center G of each area i and the boundary position B i To draw the specific position relationship;

[0137] S63. According to the structural requirements of the cold storage building, the structural characteristic data is transferred to the plan view, and important structural elements such as the bearing structure, column spacing and load-bearing walls of each area are further drawn;

[0138] S64. Mark each functional area in the plan and mark the fire partition area on the drawing according to the fire partition specification. Fire partition area D fire = {d1, d2, d3}, representing the boundaries and areas of each fire zone in the building;

[0139] S65. After marking the functional categories and fire zones of each area, conduct fire protection compliance verification to ensure that the area requirements of each fire zone match the fire protection requirements of the cold storage;

[0140] S66, Generate a complete floor plan file G final The file contains the layout of all functional areas, structural descriptions and fire partition markings. The file uses standard CAD format to facilitate further construction or verification.

[0141] Example:

[0142] In a new cold storage project of a large domestic cold chain logistics enterprise, the traditional cold storage design process often relies on manual experience, which is not only inefficient, but also prone to problems such as unreasonable spatial layout and long design cycle. The implementers decided to adopt the method of the present invention to realize the automatic splitting of cold storage building units.

[0143] In August 2024, a new cold storage project of a large domestic cold chain logistics company was launched. The project is located in the suburbs of Beijing, with a total construction area of ​​50,000 square meters and a building height of 30 meters. It is designed as a single structure, including three functional areas: refrigerated area, frozen area, and channel area, with the corresponding functional areas accounting for 40%, 35%, and 15% respectively. At first, the implementers adopted manual design, which has been proven to have low space utilization and is time-consuming.

[0144] To solve this problem, the implementers adopted an intelligent division method and collected the basic data of the cold storage project through architectural design software, including the building length of 200 meters, the building width of 250 meters, the building area of ​​50,000 square meters, the total height of 30 meters, the structural form of prestressed concrete structure, and the column spacing of 12 meters. All data were standardized into a unified coordinate system through preprocessing to ensure that subsequent processing can be analyzed based on the same data format.

[0145] Entering the feature extraction stage, the self-attention mechanism is used to analyze the input building data and automatically identify the key features of the building, including the building length, width, number of floors, total height, fire partition requirements, etc. By analyzing the building's plane size, structural form and fire partition requirements, the model can extract the core features that affect the spatial splitting and convert these features into feature vectors for the generator to process in the next step.

[0146] By constructing a generative adversarial network, the generator receives standardized building data, including building area, structural features, functional area requirements, etc., and generates multiple space splitting schemes. The discriminator is responsible for evaluating the rationality of each generated scheme, including whether the spatial layout meets the functional requirements, whether it meets the fire zoning standards, whether it meets the column spacing restrictions, etc. After 50 rounds of adversarial training, the generator generated a building splitting scheme with a space utilization rate of 89%; after 500 rounds of adversarial training, the generator finally generated multiple reasonable and compliant building splitting schemes, with a space utilization rate of 98%. The generated splitting scheme is further optimized through the self-attention mechanism, so that the area allocation and spatial layout of each functional area are reasonably adjusted.

[0147] After the generated and optimized split plan is determined, the final building split plan is converted into a building floor plan. The floor plan clearly marks the location, area and corresponding structural description of each functional area. At the same time, the fire partition mark is accurately added between the refrigeration area and the freezing area, and the design plan meets the requirements of the national fire protection code. The implementer can provide guidance for subsequent construction based on the final floor plan and structural description.

[0148] By comparing with the traditional method, the specific data are as follows Table 1:

[0149] Table 1 Comparison of key performances of the present invention and the traditional method in the process of optimizing the layout of wear-resistant materials for labor protection shoes

[0150] project Manual design method Automatic splitting method Design cycle 40 days 2 days Space Utilization 93％ 98％ Refrigerated area ratio 40％ 43％ Frozen area ratio 35％ 37％ Channel area ratio 15％ 15％ Fire zone compliance 100％ 100％

[0151] In the entire embodiment, the method of the present invention not only solves the problems of low space utilization and long design cycle in the traditional manual splitting process in the cold storage design, but also effectively improves the design efficiency and the scientific nature of the space layout. After the application of the present invention, the design cycle of the cold storage project is shortened from 40 days of traditional design to 2 days, and the space utilization rate in the design process is increased from 93% of the traditional manual design to 98%, especially the space allocation of the cold storage area and the freezing area. After optimization through the superimposed attention mechanism, the proportion of the cold storage area is increased to 43%, and the proportion of the freezing area is increased to 37%. The fire zoning fully meets the requirements of national standards, and all functional areas (including cold storage, freezing, and channel functional areas) are reasonably laid out, and there are no unreasonable designs.

[0152] The present invention introduces a self-attention mechanism to conduct in-depth analysis of building plan dimensions, building height, structural features, functional requirements, and column spacing information, and converts these features into feature vectors required for space segmentation. Compared with traditional methods, the present invention can automatically identify key features and generate more reasonable space partitions, thereby reducing the deviation and inaccuracy in manual design.

[0153] In order to solve the problem that the attention mechanism generates noise vectors during the feature extraction process, the present invention further adopts an optimized superposition attention mechanism to extract the feature vectors of buildings. The superposition attention mechanism generates attention gates and information vectors to filter the feature vectors previously generated by the attention mechanism, thereby generating more accurate feature vectors.

[0154] The present invention adopts a generative adversarial network to process the extracted building feature vectors. The building splitting scheme generated by the generative adversarial network is optimized and evaluated by the discriminator, which further improves the rationality and conformity of the design scheme, ensuring that each split area can effectively meet the specific functional requirements of the cold storage.

[0155] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for generating a cold storage building plan based on machine learning, characterized in that: The steps include: S1, obtain the dataset of cold storage buildings and perform standardization; S2. Use the self-attention mechanism to analyze the building's plane dimensions, structural form, and fire zoning requirements, identify key spatial splitting features, and generate feature vectors for building splitting; S3. Construct a generative adversarial network. The generator generates a spatial splitting solution by inputting the feature vector of building splitting. The discriminator is used to evaluate the rationality and conformity of the generated solution. The generator is trained through the generative adversarial network to generate multiple preliminary building splitting solutions. S4. Use the self-attention mechanism to further optimize the generated preliminary building splitting scheme, adjust the allocation ratio of each area to meet the functional requirements of each area, and obtain the optimized building splitting scheme; S5. Match the optimized building splitting scheme with the functional zoning requirements of the cold storage, so that all functional areas are reasonably allocated in the split layout, and obtain the final building splitting scheme; S6. Convert the final building split plan into a floor plan, draw the distribution of each functional area, provide corresponding structural descriptions, and mark the fire zones.

2. The method for generating a cold storage building plan based on machine learning automatically splitting a building monomer according to claim 1 is characterized in that: The S1 comprises the following steps: S11. Obtain the length L, width W, total area A, number of floors N, total height H, column spacing C, and fire partition area requirement A = {A1, A2, ..., A p }, the area requirement of the refrigerated area is F1, the area requirement of the frozen area is F2, the area requirement of the channel area is F3, and the characteristic vector V is obtained by normalization.

3. The method for generating a cold storage building plan based on machine learning and automatic splitting of buildings according to claim 1 is characterized in that: The S2 comprises the following steps: S21. Perform linear transformation on the eigenvector V: Q=W q V; K=W k V; V=W v V; Among them, Q is the query vector, K is the key vector, V is the value vector, and W q , W v , W v are the weight matrices for query, key, and value respectively; Perform weighted summation on the query vector, key vector, and value vector obtained to obtain the attention result after attention: Among them, softmax is the normalization function, d k is the query and key dimension, V ’ for the attentional outcome after attention; S22: The attention result after attention generated by S21 is further optimized by using the superposition attention mechanism. The superposition attention mechanism screens the attention result after attention by generating information vectors and attention gates: in, is the weight matrix, b I 、b D is the bias term, I is the information vector, and D is the attention gate; Among them, ⊙ represents the bitwise multiplication, Represents the final feature vector obtained by the stacked attention mechanism.

4. The method for generating a cold storage building plan based on machine learning and automatic splitting of buildings according to claim 1 is characterized in that: The S3 comprises the following steps: S31. Build a generative adversarial network, including two networks: the generator and the discriminator; S32, the final feature vector obtained by superimposing the attention mechanism Input to generator G: Among them, S gen represents a space splitting scheme generated by a generator, wherein the space splitting scheme includes the size and position of each functional area in the cold storage; S33. Input the space splitting scheme generated by the generator into the discriminator to obtain a rationality score: D(S gen )=σ(W d S gen +b d ); Among them, σ is the Sigmoid activation function, W d is the weight matrix of the discriminator, b d is the bias term, D(S gen ) is the rationality score of the discriminator output; S34, the generator and the discriminator are continuously optimized through the adversarial training process. The discriminator evaluates the rationality of the splitting scheme generated by the generator and reversely transfers the error to the generator, thereby promoting the generator to generate a splitting scheme that is more in line with the specification; The loss functions of the generator and discriminator during training are: S35. Update the parameters of the generator G and the discriminator D through the back propagation algorithm, use gradient descent to calculate the gradient of the loss function to the network parameters, and update the model parameters: Among them, η is the learning rate, θ G and θ D are the parameters of the generator and the discriminator respectively; S36. After T rounds of training, use the generator to output a preliminary building splitting solution S initial .

5. The method for generating a cold storage building plan based on machine learning and automatic splitting of buildings according to claim 1 is characterized in that: The S4 comprises the following steps: S41. Preliminary building split plan S initial Perform linear transformation to obtain the query matrix Q, key matrix K and value matrix V, and calculate the self-attention weight matrix W attention : S42. Obtain the optimized building splitting solution S through weighted summation optimized : S optimized =W attention V; S43, by iterating S41 and S42, the generated building splitting scheme is optimized so that the proportion and layout of each functional area are maximized and the optimized building splitting scheme S' is obtained optimized .

6. The method for generating a cold storage building plan based on machine learning automatically splitting a building monomer according to claim 1 is characterized in that: The S5 comprises the following steps: S51, optimize the building splitting plan S' optimized The area and spatial layout information of each region in is extracted and recorded as S regions = {R1, R2, R3}, where R i (i∈{1, 2, 3}) represents the information of each functional area in the building splitting scheme; S52. Using R i Further obtain the area information A of each functional area i and location information L i ; S53, calculate the error value E of each area i : AND i |A i -TO i_target ||L i -L i_target |? Among them, A i_target Indicates the target area in the corresponding functional zoning requirements, L i_target Indicates the spatial layout requirements of the corresponding functional zones; S54. For each region R i , by minimizing the error value E i To adjust the area and layout information: S55. After the final matching of the functional zoning and the building splitting scheme is completed, the final building splitting scheme S is output. final , including the final layout, area and location arrangement of each functional area.

7. The method for generating a cold storage building plan based on machine learning automatically splitting a building unit according to claim 1 is characterized in that: The S6 comprises the following steps: S61. Final building split plan S final Each region R i Perform spatial layout transformation to obtain the plane coordinate information C of each functional area i =(x i ,y i ), where x i ,y i Respectively represent functional areas R i The horizontal and vertical coordinate positions in a two-dimensional plane; S62, according to the plane coordinates C i and region R i Area requirement A i , draw the outer boundary of each area on the design drawing to form the outline of each functional area, and calculate the geometric center G of each area i and the boundary position B i To draw the specific position relationship; S63. According to the structural requirements of the cold storage building, the structural characteristic data is transferred to the plan view, and important structural elements such as the bearing structure, column spacing and load-bearing walls of each area are further drawn; S64. Mark each functional area in the plan and mark the fire partition area on the drawing according to the fire partition specification. Fire partition area D fire = {d1, d2, d3}, representing the boundaries and areas of each fire zone in the building; S65. After marking the functional categories and fire zones of each area, conduct fire protection compliance verification to ensure that the area requirements of each fire zone match the fire protection requirements of the cold storage; S66, Generate a complete floor plan file G final The file contains the layout of all functional areas, structural descriptions and fire partition markings. The file uses standard CAD format to facilitate further construction or verification.

Images