Intelligent green curtain wall reverse modeling method based on BIM-IoT fusion

Abstract

The application relates to an intelligent green curtain wall reverse modeling method based on BIM-IoT fusion in the technical field of data processing and artificial intelligence. The method comprises the following steps: acquiring BIM component data and IoT sensing data, and constructing a data set; performing node feature fusion and edge weight calculation based on the data set, and constructing a dynamic graph structure; constructing a machine learning-based geometric parameter prediction model, including a space-time feature encoder module, a space-time graph convolution network module, a multi-scale feature pyramid module and a physical constraint decoder module; inputting the dynamic graph structure into the machine learning-based geometric parameter prediction model for iterative training; constructing a real-time dynamic graph structure based on basic BIM component data and real-time IoT sensing data, inputting the real-time dynamic graph structure into the trained machine learning-based geometric parameter prediction model, outputting a predicted geometric parameter vector, and performing intelligent green curtain wall maintenance. The method realizes the collaborative expression of component static geometry and transient environmental changes.

Description

Technical Field

[0001] This application relates to the fields of data processing and artificial intelligence technology, and in particular to a reverse modeling method for intelligent green curtain walls based on BIM-IoT fusion. Background Technology

[0002] As a crucial component of modern architecture, curtain walls not only serve as enclosures and enhance aesthetics but also play a vital role in building energy conservation, structural safety, and environmental adaptability. With the increasing prevalence of green building and intelligent building concepts, curtain wall systems require a higher level of digital support in their design, construction, and operation and maintenance phases. Currently, BIM (Building Information Modeling) is widely used in curtain wall design and management, providing complete geometric and attribute information. However, it primarily consists of static data, making it difficult to reflect the dynamic performance changes of the curtain wall during actual operation. Meanwhile, the development of IoT (Internet of Things) technology enables sensors to collect environmental and structural parameters such as temperature, humidity, wind pressure, and illumination in real time, making dynamic performance monitoring and intelligent operation and maintenance possible.

[0003] However, in existing technologies, BIM and IoT data are often fragmented, lacking an effective integration mechanism. This makes it difficult to simultaneously consider both static geometric features and dynamic temporal characteristics in curtain wall reverse modeling. Furthermore, traditional modeling methods often rely on fixed topological relationships or single geometric matching algorithms, failing to accurately capture transient changes in components under environmental influences. This results in reconstructions prone to localized inaccuracies or lack of physical consistency. These shortcomings limit the intelligent management and green performance optimization of curtain walls throughout their entire lifecycle.

[0004] Therefore, there is an urgent need in related technologies for a method of reverse modeling of curtain walls that can take into account both static geometric features and dynamic temporal features. Summary of the Invention

[0005] Therefore, it is necessary to provide a smart green curtain wall reverse modeling method based on BIM-IoT integration that can take into account both static geometric features and dynamic temporal features, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a reverse modeling method for intelligent green curtain walls based on BIM-IoT integration. The method includes:

[0007] Acquire BIM component data and IoT sensor data to construct a dataset;

[0008] Based on the dataset, node feature fusion and edge weight calculation are performed to construct a dynamic graph structure;

[0009] A machine learning-based geometric parameter prediction model is constructed, which includes a spatiotemporal feature encoder module, a spatiotemporal graph convolutional network module, a multi-scale feature pyramid module, and a physical constraint decoder module.

[0010] The dynamic graph structure is input into the machine learning-based geometric parameter prediction model for iterative training. The multi-objective mixed loss function and parameter gradient are calculated, and the parameters are updated using the Adam optimizer.

[0011] A real-time dynamic graph structure is constructed based on basic BIM component data and real-time IoT sensor data. The real-time dynamic graph structure is input into a trained machine learning-based geometric parameter prediction model, which outputs a predicted geometric parameter vector and performs intelligent green curtain wall maintenance.

[0012] Optionally, in one embodiment of this application, the step of fusing node features and calculating edge weights based on the dataset to construct a dynamic graph structure includes:

[0013] The static geometric attribute feature vector is concatenated with the temporal difference feature of the dynamic sensing data to form a node feature vector that is sensed at any given time. The temporal difference feature of the dynamic sensing data is the difference between the feature vector of the sensing data at the current time and the feature vector of the sensing data at the previous time.

[0014] Optionally, in one embodiment of this application, the step of constructing a dynamic graph structure by fusing node features and calculating edge weights based on the dataset further includes:

[0015] Edge weights are calculated based on the geometric distance between nodes and the topological connection matrix. The geometric distance is calculated using the Euclidean norm, and the topological connection matrix indicates whether there is a physical connection between components.

[0016] Optionally, in one embodiment of this application, the spatiotemporal feature encoder module includes:

[0017] Based on edge weights and node feature vectors, the neighbor information of each node is aggregated through graph convolution operations, and the spatial encoding feature vector is extracted using the ReLU activation function.

[0018] One-dimensional convolution operations with different kernel sizes are used to extract short, medium, and long-range temporal features in parallel. The short, medium, and long-range temporal features are then fused through max pooling to form a multi-scale temporal feature vector.

[0019] Optionally, in one embodiment of this application, the spatiotemporal graph convolutional network module includes:

[0020] Based on spatially encoded feature vectors and prior edge weights, a dynamic adjacency matrix is ​​generated by calculating the dynamic association strength between nodes through a self-attention mechanism.

[0021] Optionally, in one embodiment of this application, the spatiotemporal graph convolutional network module further includes:

[0022] The node feature matrix is ​​updated by performing graph convolution operation based on the dynamic adjacency matrix, and the multi-scale temporal feature vector is injected into the node feature matrix through a gating mechanism.

[0023] Optionally, in one embodiment of this application, the multi-scale feature pyramid module includes:

[0024] Upsampling is performed on the node feature matrix of the third layer of the spatiotemporal graph convolutional network to expand the feature size and obtain high-level pyramid features.

[0025] The node feature matrix of the second layer of the spatiotemporal graph convolutional network is concatenated with the high-level pyramid features, and the features are fused through convolution operations to generate the mid-level pyramid features.

[0026] The spatial coding feature matrix is ​​concatenated with the middle-layer pyramid features, and the features are fused through convolution to generate the first-layer pyramid features.

[0027] Optionally, in one embodiment of this application, the physical constraint decoder module includes:

[0028] Based on the first-layer pyramid features and multi-scale temporal feature vectors, the predicted geometric parameter vectors are obtained by decoding through a multilayer perceptron.

[0029] Temperature and light intensity are obtained by decoding the predicted geometric parameter vector, and physical constraint loss is calculated based on the temperature and light intensity.

[0030] Optionally, in one embodiment of this application, the formula for calculating the physical constraint loss is:

[0031] ;

[0032] in, Represents the physical constraint loss term. This represents the thermal conductivity of the material at the v-th node. This represents the decoding temperature value of the v-th node. Indicates the external ambient temperature. Indicates the heat capacity coefficient of a material. Indicates the time step. This represents the decoded illumination intensity of the v-th node, where v is the node index. This represents the L2 norm.

[0033] Optionally, in one embodiment of this application, the multi-objective hybrid loss function includes geometric reconstruction loss, physical constraint loss, adversarial loss, and weight regularization loss. The geometric reconstruction loss is calculated based on the predicted geometric attribute matrix and the true geometric attribute matrix. The adversarial loss is calculated based on the discriminant output of the attribute discriminator on the true geometric attribute matrix, the predicted geometric attribute matrix, and the features of the first layer pyramid. The weight regularization loss is calculated based on the weight matrix of each layer of the spatiotemporal graph convolutional network.

[0034] The aforementioned reverse modeling method for intelligent green curtain walls based on BIM-IoT fusion first acquires BIM component data and IoT sensor data to construct a dataset. Then, based on the dataset, node feature fusion and edge weight calculation are performed to construct a dynamic graph structure. Next, a machine learning-based geometric parameter prediction model is constructed, comprising a spatiotemporal feature encoder module, a spatiotemporal graph convolutional network module, a multi-scale feature pyramid module, and a physical constraint decoder module. The dynamic graph structure is then input into the machine learning-based geometric parameter prediction model for iterative training, calculating a multi-objective hybrid loss function and parameter gradients, and updating parameters using the Adam optimizer. Finally, a real-time dynamic graph structure is constructed based on basic BIM component data and real-time IoT sensor data. This real-time dynamic graph structure is input into the trained machine learning-based geometric parameter prediction model, outputting a predicted geometric parameter vector, and used for intelligent green curtain wall maintenance. In other words, by heterogeneously fusing BIM geometric attributes with IoT sensing data, a dynamic feature modeling method based on temporal difference and geometric attribute splicing is proposed, realizing the collaborative expression of static geometry of components and transient environmental changes. A dynamic graph structure combining topological relationships and spatial distances is constructed, and an adaptive attention mechanism is introduced to dynamically adjust adjacency relationships, enabling the modeling process to simultaneously possess physical rationality and implicit spatiotemporal dependency expression capabilities. A multi-scale geometric perception feature pyramid structure is proposed, which fully preserves local geometric details while ensuring global semantics, solving the problem that traditional feature extraction methods easily lose key geometric information such as curvature and hole positions. Physical laws such as energy conservation are introduced into the decoding and loss function design, and adversarial regularization and multi-objective optimization are combined to achieve a unity of geometric accuracy and physical rationality, and improve the robustness and generalization ability of the model under small sample conditions. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a reverse modeling method for a smart green curtain wall based on BIM-IoT fusion in one embodiment.

[0036] Figure 2 This is a schematic diagram showing the comparison results of different methods in geometric reconstruction accuracy in one embodiment;

[0037] Figure 3This is a schematic diagram showing the comparison of the degree to which different methods satisfy physical constraints in one embodiment;

[0038] Figure 4 This is a schematic diagram showing the comparison of the running time of various methods under different numbers of components in one embodiment;

[0039] Figure 5 This is a schematic diagram showing the comparison of reconstruction accuracy results of different methods under different numbers of components in one embodiment;

[0040] Figure 6 This is a schematic diagram showing the results of a comparison of transient change capture capabilities in one embodiment;

[0041] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] In one embodiment, such as Figure 1 As shown, a reverse modeling method for intelligent green curtain walls based on BIM-IoT integration is provided, including the following steps:

[0044] S101: Acquire BIM component data and IoT sensor data to build a dataset.

[0045] In this embodiment, multi-source data is acquired from the BIM (Building Information Modeling) model and IoT (Internet of Things) sensor network of the intelligent green curtain wall system. BIM component data is exported in batches through the application programming interface (API) of BIM design software (such as Revit or ArchiCAD), extracting the static geometric attributes of each curtain wall component, including dimensions (such as length, width, and height), shape parameters (such as curvature and angle), and material type (such as glass and aluminum alloy). These attributes are stored in vector form, forming a geometric attribute feature vector. Simultaneously, topological connections are extracted from the BIM model to construct a topological connection matrix, representing the physical connection status between components, such as bolted connections and welding. IoT sensor data is collected in real time by sensor nodes deployed on the curtain wall components, monitoring environmental parameters such as temperature, humidity, wind pressure, and illumination. Sensor data is recorded in time-series form, with the sampling frequency set according to application requirements, such as once per second or once every ten seconds, forming a sensor data feature vector. The collected data is divided into a training set for model training and a validation set for model validation, with a ratio of 7:3.

[0046] S102: Based on the dataset, perform node feature fusion and edge weight calculation to construct a dynamic graph structure.

[0047] In this embodiment, during the reverse modeling task of the intelligent green curtain wall, the BIM component diagram contains static topological relationships and geometric attributes, while IoT sensor data provides dynamic temporal information such as temperature, humidity, wind pressure, and illumination. Conventional processing methods directly stitch these two heterogeneous features together, weakening the correlation between spatial structure and temporal dynamics, failing to effectively express the physical interaction relationships between components, and thus limiting the model's ability to model curtain wall behavior in complex environments. Therefore, this invention models the curtain wall system as a dynamic graph structure. In a dynamic graph structure, nodes represent components, and edges represent the physical interactions between components. Node features integrate geometric and sensory information, while edge weights integrate geometric distance and topological priors. Definition For a set of nodes, there are a total of individual curtain wall components, Given a set of edges, the method for calculating the weight of each edge is as follows: The calculation method is the same. The node feature matrix is ​​composed of eigenvectors. Composition, node feature matrix Each row corresponds to a feature vector of a node at time t. By constructing a dynamic graph structure, where node features are fused with temporal differences of geometric attributes and sensor information, and edge weights are calculated based on spatial coordinates and topological connections, the ability to model spatial-temporal relationships between components is enhanced.

[0048] Specifically, in one embodiment of this application, the step of constructing a dynamic graph structure by fusing node features and calculating edge weights based on the dataset includes:

[0049] The static geometric attribute feature vector is concatenated with the temporal difference feature of the dynamic sensing data to form a node feature vector that is sensed at any given time. The temporal difference feature of the dynamic sensing data is the difference between the feature vector of the sensing data at the current time and the feature vector of the sensing data at the previous time.

[0050] In one embodiment of this application, for each curtain wall component node, the static geometric attribute feature vector is concatenated with the temporal differential feature of dynamic sensing data to form a node feature vector that is aware of the current moment. The static geometric attribute feature vector includes attributes such as size, shape, and material type, while the dynamic sensing data includes monitored values ​​such as temperature, humidity, wind pressure, and illumination. The temporal differential feature is obtained by calculating the difference between the current sensing data feature vector and the previous sensing data feature vector. The concatenation operation enhances the perception of transient changes and decouples static and dynamic features, as shown below:

[0051] ;

[0052] in, Let v represent the feature vector of the v-th node at time t, with dimension 1. It integrates geometric properties with dynamic sensing change information; This represents the geometric attribute feature vector of the v-th node, with dimension 1. This includes static attributes such as size, shape, and material type; The dimension of the geometric attribute feature vector depends on the number of geometric attributes selected and is set according to the actual application. For example, if it includes length, width, height, area, volume, and material type, then... ; This represents the feature vector of the sensing data of the v-th node at time t, with dimension . This includes dynamic monitoring values ​​such as temperature, humidity, wind pressure, and light intensity; The dimension of the sensor data feature vector depends on the number of sensor types and is set according to the actual number of sensor channels. For example, if monitoring four parameters: temperature, humidity, wind pressure, and light intensity, then... ; The dimension of the temporal difference feature is equal to the dimension of the sensor data feature vector. Same, that is ; This represents the feature vector of the sensing data of the v-th node at time t-1, used to calculate instantaneous changes; The time-difference operator is represented as follows: This is used to extract the changes in sensor data between adjacent time points; This indicates a vector concatenation operation; v is the node index.

[0053] It should be noted that nodes represent independent physical components in the curtain wall system, such as glass panels, aluminum alloy frame units, photovoltaic modules, or ventilation control valves. Each node corresponds to an identifiable entity object in the BIM model. (Geometric attribute feature vector) These are static geometric attributes extracted from the BIM model, including dimensions, shape, material type, etc. For example, for a glass panel component, It may include length, width, height, area, volume, and material encoding. In one embodiment, if a length of 2.5m, a width of 1.2m, a height of 0.01m, an area of ​​3.0m², a volume of 0.03m³, and a material type of 1 represent tempered glass, then the geometric attribute feature vector... Dimension Sensor data feature vector These are dynamic monitoring values ​​acquired in real time from IoT sensors. These sensors are directly installed on the curtain wall components. For example, for a glass panel node, temperature, humidity, wind pressure, and light intensity sensors may be installed. In one embodiment, at time t, the temperature is 25.5°C, humidity is 60%, wind pressure is 1013.25 Pa, and light intensity is 500 lux. The sensor data feature vector... Dimension .

[0054] Conventional methods directly concatenate geometric attribute feature vectors and sensor data feature vector ,Right now This approach ignores dynamic trends, which can easily lead to a lag in the model's response to transient events such as sudden wind loads. In this embodiment, however, a time-series difference operator is used. Extract instantaneous changes and construct feature vectors Differential quantification can enhance transient perception by quantifying the abrupt intensity of environmental disturbances. At the same time, the separation of geometric attributes and dynamic changes can decouple static / dynamic states and avoid feature confusion.

[0055] In one embodiment of this application, the step of constructing a dynamic graph structure by fusing node features and calculating edge weights based on the dataset further includes:

[0056] Edge weights are calculated based on the geometric distance between nodes and the topological connection matrix. The geometric distance is calculated using the Euclidean norm, and the topological connection matrix indicates whether there is a physical connection between components.

[0057] In one embodiment of this application, edge weights are calculated based on the three-dimensional spatial coordinate vectors between nodes and the topological connection matrix to characterize the physical interaction strength between components. The edge weights are obtained by multiplying the inverse weighted geometric distance with the topological connection state, where the geometric distance is calculated using the Euclidean norm. The topological connection matrix indicates whether there is a physical connection between components. Normalization ensures a reasonable weight distribution, with physically adjacent and connected components having higher interaction weights, expressed as follows:

[0058] ;

[0059] in, The table shows the edge weights between nodes v and v, which are scalars reflecting the strength of the physical interaction between the two nodes; u is the node index distinct from v. This represents the three-dimensional spatial coordinate vector of the u-th node, such as the three-dimensional spatial coordinate vector of the u-th node in a glass panel. The unit of measurement is uniformly the meter; Represents the three-dimensional spatial coordinate vector of the v-th node; This represents the three-dimensional spatial coordinate vector of the k-th node, which is the three-dimensional spatial coordinate vector of the neighboring nodes of the u-th node, and is used to calculate the normalized distance weight. To distinguish the node index from v and u; This represents the L2 norm, which is equivalent to the Euclidean norm. Let represent the set of neighboring nodes of the u-th node; This represents the connection status between the u-th node and the v-th node in the topology connection matrix. The value is 0 or 1, where 1 indicates that there is a connection and 0 indicates that there is no connection.

[0060] It should be noted that, The term is a normalization term, summed by the reciprocals of the distances between the u-th node and all its neighboring nodes, ensuring a uniform weight distribution and avoiding weight imbalance due to differences in the magnitude of component spacing. The topology connection matrix is ​​extracted from the connection relationship library of the BIM model. If the u-th component and the v-th component have a physical connection, such as a bolted connection or welded connection, then... If it is an independent component, such as adjacent glass panels not being directly connected, then .

[0061] S103: Construct a machine learning-based geometric parameter prediction model, which includes a spatiotemporal feature encoder module, a spatiotemporal graph convolutional network module, a multi-scale feature pyramid module, and a physical constraint decoder module.

[0062] In this embodiment, a machine learning-based geometric parameter prediction model is constructed, including a spatiotemporal feature encoder module, a spatiotemporal graph convolutional network module, a multi-scale feature pyramid module, and a physical constraint decoder module. Specifically, conventional graph convolutional methods, when processing curtain wall component data, neglect the propagation characteristics of sensor data under geometric constraints, and the use of a fixed time window makes it difficult to capture multi-scale dynamic changes, resulting in insufficient extraction of spatial and temporal features and an inability to effectively model the complex interaction behavior between components, thus affecting the accuracy of inverse modeling. Therefore, by designing a geometry-aware spatiotemporal encoder, spatial and temporal features are processed independently, the encoding paths are decoupled to avoid feature confusion, and multi-scale convolution is used to capture dynamic changes, improving the parallel efficiency and expressive power of feature extraction.

[0063] Conventional spatiotemporal graph convolutional networks employ a fixed neighborhood sampling strategy, which cannot adapt to the dynamically changing physical coupling relationships between curtain wall components. Graph convolution with a fixed structure struggles to capture the implicit connections between components in complex environments, resulting in insufficient ability to model the dynamic behavior of the curtain wall system and affecting the accuracy of inverse modeling. Therefore, an adaptive neighborhood aggregation mechanism is designed, incorporating topological prior knowledge into self-attention computation to dynamically adjust the connection weights between nodes. Furthermore, a gating mechanism is used to fuse temporal features, enhancing the model's ability to model the dynamic physical coupling of curtain wall components.

[0064] In reverse modeling, while deep network features possess rich semantic information, they lose geometric details, leading to a decrease in the accuracy of reconstructing the geometric parameters of curtain wall components. Conventional methods directly use deep features for decoding, failing to effectively preserve component geometric details and affecting the accuracy of reverse modeling. Therefore, by constructing a multi-scale feature pyramid structure and employing a top-down path to fuse deep semantic features with shallow geometric features, and by using skip connections to retain feature information at different levels, the problem of detail loss in reverse modeling is solved.

[0065] Conventional decoders focus solely on the reconstruction accuracy of geometric parameters during curtain wall reverse modeling, neglecting the physical laws that curtain wall systems must follow, particularly the law of conservation of energy. This can lead to prediction results that violate fundamental thermodynamic principles, reducing the engineering reliability and practicality of the reverse modeling results. Therefore, by designing a physically constrained decoder, thermodynamic constraints are applied to the geometric parameter predictions to ensure that the prediction results conform to the law of conservation of energy, thereby improving the physical rationality and engineering applicability of the reverse modeling.

[0066] In one embodiment of this application, the spatiotemporal feature encoder module includes:

[0067] S201: Based on edge weights and node feature vectors, the neighbor information of each node is aggregated through graph convolution operations, and the spatial encoding feature vector is extracted using the ReLU activation function.

[0068] S202: Short, medium, and long-range temporal features are extracted in parallel using one-dimensional convolution operations with different kernel sizes. The short, medium, and long-range temporal features are then fused through max pooling to form a multi-scale temporal feature vector.

[0069] In one embodiment of this application, firstly, based on edge weights and node feature vectors, the neighbor information of each node is aggregated through graph convolution operations. The aggregated features are then processed using the ReLU activation function to extract spatially encoded feature vectors. The graph convolution operation uses edge weights to weight and sum the features of neighboring nodes, and performs linear projection through the weight matrix to capture the spatial dependencies between components and enhance nonlinear expressive power, as shown below:

[0070] ;

[0071] in, This represents the spatial encoding feature vector of the v-th node, with dimension . , representing its spatial context information; Represents the ReLU activation function; Let v represent the set of neighboring nodes of the v-th node; The spatiotemporal feature encoding weight matrix is ​​a trainable parameter with dimension . , used for linear projection.

[0072] Subsequently, for the feature sequence of each node, short-, medium-, and long-range temporal features are extracted in parallel using one-dimensional convolution operations with different kernel sizes. Then, these multi-scale features are fused through max pooling to form a multi-scale temporal feature vector, thereby capturing the dependencies between different time scales and improving computational efficiency. This is represented as:

[0073] ;

[0074] in, This represents the multi-scale temporal feature vector of the v-th node, with dimension . It integrates dynamic features at different time scales; This indicates the max pooling operation, used to select the most significant response from multi-scale features; The three convolution operations with kernel sizes of 3, 5, and 7 are used to extract short, medium, and long-range temporal features, respectively. Indicates the kernel size; This indicates the length of the time window and controls the number of historical time steps. In one embodiment, the time window length... Based on sensor data sampling frequency Settings, and typically take into account the length of the time window. The settings need to cover the typical cycle of the curtain wall response. Hz indicates sampling on the order of seconds. Covering 10 seconds of history, if Hz indicates sampling frequency at 10-second intervals. Coverage lasts for 150 seconds.

[0075] It should be noted that, The term represents the node v starting from the node t. The feature sequence from time t to time t is constructed with a length of . The time-series slices are used to capture hysteresis effects, such as the delayed response to thermal stress caused by temperature changes, while also identifying periodic patterns, such as the diurnal periodic fluctuations in light intensity.

[0076] In one embodiment of this application, the spatiotemporal graph convolutional network module includes:

[0077] Based on spatially encoded feature vectors and prior edge weights, a dynamic adjacency matrix is ​​generated by calculating the dynamic association strength between nodes through a self-attention mechanism.

[0078] In one embodiment of this application, firstly, based on the spatially encoded feature vector and prior edge weights, a self-attention mechanism is used to calculate the dynamic association strength between nodes. The self-attention mechanism transforms the spatially encoded features using a query transformation weight matrix and a key transformation weight matrix, and scales the attention score. Then, it is combined with the prior edge weights and normalized using the Softmax function to generate dynamic adjacency matrix weights, thereby adaptively capturing implicit relationships based on topological priors, as shown below:

[0079] ;

[0080] in, This represents the weight of the u-th node and the v-th node in the dynamic adjacency matrix, which is the element in the u-th row and v-th column of the dynamic adjacency matrix. It is a scalar that reflects the strength of the dynamic relationship. This represents the Softmax function, used to normalize attention scores; This represents the query transformation weight matrix, with dimensions [missing information]. , are trainable parameters; This represents the key transformation weight matrix, with dimension 1. , are trainable parameters; This represents the spatial encoding feature vector of the v-th node; This represents the spatial encoding feature vector of the u-th node; This represents the scaling factor, used to control the numerical range of the attention score; a value of 64 is preferred. This indicates the transpose operation.

[0081] It should be noted that conventional self-attention calculation methods do not consider edge weights and ignore physical topology, which can easily lead to non-physical connections, such as spurious associations between distant glass panels. However, by using prior edge weights... As an attention bias, it can force In cases where nodes without topological connections have attention scores approaching 0, the model adaptively learns implicit associations, such as wind pressure transmission paths, based on the topological skeleton. This allows the model to capture dynamic couplings while maintaining physical rationality, thus resolving the contradiction between engineering interpretability and model flexibility.

[0082] In one embodiment of this application, the spatiotemporal graph convolutional network module further includes:

[0083] The node feature matrix is ​​updated by performing graph convolution operation based on the dynamic adjacency matrix, and the multi-scale temporal feature vector is injected into the node feature matrix through a gating mechanism.

[0084] In one embodiment of this application, a graph convolution operation using a dynamic adjacency matrix is ​​performed to update the node feature matrix. Simultaneously, a gating mechanism is used to inject a multi-scale temporal feature matrix into the node features. This gating mechanism transforms the multi-scale temporal features using a temporal feature transformation weight matrix and combines it with the current node features through a hyperbolic tangent activation function and element-wise multiplication, thereby achieving a fusion of spatial dependence and temporal evolution and enhancing spatiotemporal expressiveness. This is expressed as follows:

[0085] ;

[0086] in, Represents the spatiotemporal graph convolutional network. The node feature matrix of the layer has a dimension of ; Represents the spatiotemporal graph convolutional network. The node feature matrix of the layer has a dimension of ; This is the layer index of the spatiotemporal graph convolutional network, which has a total of 3 layers. Represents a dynamic adjacency matrix with dimension . The element in the u-th row and v-th column of the dynamic adjacency matrix is ; Represents the spatiotemporal graph convolutional network. The weight matrix of the layer has dimensions of , are trainable parameters; This represents the time-series feature injection gate function, calculated as follows: ; Represents a multi-scale temporal feature matrix with dimension . The v-th row of the multi-scale temporal feature matrix represents the multi-scale temporal feature vector of the v-th node. ; This represents the weight matrix for time-series feature transformation, with dimension 1. , are trainable parameters; Represents the hyperbolic tangent activation function; This represents element-wise multiplication; The number of nodes, i.e. the total number of curtain wall components, depends on the number of components in the specific curtain wall system.

[0087] It should be noted that the total number of layers in the spatiotemporal graph convolution is set to 3. Characterized by shallow features, dominated by geometric details. Characterizing mid-level features, dominated by local interactions. The deep features are represented by global semantics. At the same time, setting the total number of spatiotemporal graph convolutions to 3 is only a preferred implementation of the present invention. That is, while maintaining the feature extraction performance, the number of model parameters is reduced. In specific implementations, the total number of layers can be expanded to a greater number.

[0088] It should also be noted that, The project aggregates neighbor information based on a dynamic adjacency matrix and models the physical coupling between components. The method selectively enhances temporal features through a gating mechanism, while using a hyperbolic tangent activation function to filter out irrelevant temporal noise, such as instantaneous illumination interference. This enables the independent encoding of spatial dependence and temporal evolution, avoiding feature confusion.

[0089] In one embodiment of this application, the multi-scale feature pyramid module includes:

[0090] S301: The node feature matrix of the third layer of the convolutional network is upsampled to expand the feature size and obtain high-level pyramid features.

[0091] S302: The node feature matrix of the second layer of the spatiotemporal graph convolutional network is concatenated with the high-level pyramid features, and the features are fused through convolution operations to generate the mid-level pyramid features.

[0092] S303: The spatial coding feature matrix is ​​concatenated with the middle-layer pyramid features, and the features are fused through convolution to generate the first-layer pyramid features.

[0093] In one embodiment of this application, firstly, an upsampling operation is performed on the node feature matrix of the third layer of the spatiotemporal graph convolutional network to expand the feature size and obtain high-level pyramid features. The upsampling uses bilinear interpolation to match the size of the high-level semantic features with that of the middle-level features, as shown below:

[0094] ;

[0095] in, This represents the pyramid feature of the third layer of the spatiotemporal graph convolutional network, with dimension [missing information]. This is the upsampling result of high-level semantic features; Set the dimensions for the pyramid features. To maintain feature consistency; This indicates an upsampling operation, which uses bilinear interpolation to enlarge the feature size; This represents the node feature matrix of the third layer of the spatiotemporal graph convolutional network, with dimension 1. It contains high-level semantic information.

[0096] Next, the node feature matrix of the second layer of the spatiotemporal graph convolutional network is concatenated with the upsampled high-level pyramid features, and then these features are fused through convolution operations to generate the mid-level pyramid features, represented as:

[0097] ;

[0098] in, This represents the second-layer pyramid feature of the spatiotemporal graph convolutional network, with dimension [missing information]. It integrates the characteristics of middle and high-level management. This represents a convolution operation used for feature fusion and dimensionality reduction, and is preferably employed. Convolutional kernels perform feature fusion and dimensionality reduction to avoid spatial redundancy while maintaining computational efficiency; This represents the node feature matrix of the second layer of the spatiotemporal graph convolutional network, with dimension 1. It contains mid-level semantic information.

[0099] Finally, the spatial encoding feature matrix is ​​concatenated with the mid-level pyramid features, and then these features are fused through convolution to generate the first-level pyramid features, preserving shallow geometric details and mid-level semantic information, thus achieving multi-scale feature fusion, represented as:

[0100] ;

[0101] in, This represents the first-layer pyramid feature of the spatiotemporal graph convolutional network, with dimension [missing information]. As the final output, multi-scale details are preserved; Indicates the convolution operation; Indicates feature concatenation operation; Represents the spatial encoding feature matrix, with dimension . It contains shallow geometric information, and the v-th row of the spatial coding feature matrix represents the spatial coding feature vector of the v-th node. .

[0102] Spatiotemporal Graph Convolutional Network First Layer Pyramid Features Employing a pyramid fusion approach, the reverse feature recovery path is used to achieve top-down semantic transfer. This utilizes the third layer of pyramid features in the spatiotemporal graph convolutional network. Global semantics, such as the overall deformation pattern of the curtain wall, are passed to the middle layer. Simultaneously, a cross-layer skip connection method is used to utilize the second layer pyramid features of the spatiotemporal graph convolutional network. By integrating mid-to-high-level features, component-level interactions are captured, such as frame-panel collaborative deformation. Additionally, the spatiotemporal graph convolutional network utilizes first-layer pyramid features. Directly based on spatial coding feature matrix The obtained high-resolution geometric information can effectively recover component-level details, such as bolt hole offsets, which is something that conventional codec structures such as U-Net cannot achieve. It can effectively avoid the situation where conventional codec structures lose key geometric parameters such as radius of curvature due to multiple downsampling in curtain wall reverse modeling.

[0103] In one embodiment of this application, the physical constraint decoder module includes:

[0104] S401: Based on the first layer pyramid features and multi-scale temporal feature vectors, the predicted geometric parameter vector is obtained by decoding through a multilayer perceptron.

[0105] S402: Decode the temperature value and light intensity based on the predicted geometric parameter vector, and calculate the physical constraint loss based on the temperature value and light intensity.

[0106] In one embodiment of this application, firstly, based on the first-layer pyramid features and multi-scale temporal feature vectors, a predicted geometric parameter vector is obtained through multilayer perceptron decoding. The multilayer perceptron then maps the concatenated features to the geometric attribute space, realizing the prediction from features to geometric parameters, as shown below:

[0107] ;

[0108] in, This represents the predicted geometric parameter vector for the v-th node, with dimension . This includes static attributes such as size, shape, and material type; This represents a multilayer perceptron used for mapping features to geometric parameters; This represents the first-layer pyramid feature of the spatiotemporal graph convolutional network at node v, in vector format, and is also the first-layer pyramid feature of the spatiotemporal graph convolutional network. The Okay, that is, the first layer pyramid features of the spatiotemporal graph convolutional network. The feature vector of the v-th node.

[0109] It should be noted that the predicted geometric parameter vector Characterization of the first The geometric attribute vectors reconstructed from the reverse modeling process of each curtain wall component have dimensions similar to the original geometric attribute feature vectors. They are completely identical, and there is a one-to-one correspondence between them in terms of semantics and structure; that is, the predicted geometric parameter vectors. It represents the geometric parameters of the components inferred by the model based on spatiotemporal characteristics, and... It represents the relationship between predicted and actual values, and is the geometric attribute of the current state inferred by the model based on dynamic sensing data, such as the actual state after being affected by the environment. In one embodiment, it is assumed that the geometric attribute feature vector of a certain glass panel node... After reverse modeling, the geometric attribute feature vector predicted by the model may be: This indicates that the glass panel's length has slightly increased from 2.5 to 2.52 due to thermal expansion, and its width has slightly decreased from 1.2 to 1.18. The area and volume have changed accordingly, but the material type remains the same, and the code is still 1.

[0110] It should also be noted that during the training process, the predicted geometric parameter vector... Geometric Attribute Eigenvectors Losses are calculated to optimize model parameters. The predicted results must be directly compared with the original BIM component parameters to support engineering decisions, such as deformation detection and maintenance prompts. If certain geometric properties are immutable, such as material type, the predicted value should be consistent with the original value. If the component undergoes deformation, such as bending or displacement, the predicted value reflects the actual state and differs from the original static value.

[0111] It should also be noted that the multilayer sensor receives... Using these features as input yields better technical results than simply receiving the first-layer pyramid features of a spatiotemporal graph convolutional network. Due to the time-delay nature of curtain wall geometric deformation, such as glass thermal bending after 2 hours of sunlight, this can be addressed by combining multi-scale temporal feature vectors with temporal features. It can effectively perceive the impact of dynamic environments.

[0112] Next, the temperature value and light intensity are obtained by decoding the predicted geometric parameter vector. Then, the physical constraint loss term is calculated. The physical constraint loss term is calculated through the energy conservation equation to measure the deviation between the prediction result and the energy conservation, and to force the prediction result to conform to the basic thermodynamic principles.

[0113] In one embodiment of this application, the formula for calculating the physical constraint loss is:

[0114] ;

[0115] in, The physical constraint loss term is a scalar that measures the deviation of the prediction result from the law of conservation of energy. This represents the thermal conductivity coefficient of the material at the v-th node. It is a priori physical parameter that can be obtained from the material library. For example, if glass = 1.05 W / m²K, then set it to... ; This represents the decoding temperature value of the v-th node; This represents the external ambient temperature, which is a known constant and can be obtained from real-time data from weather stations. The unit is uniformly measured in degrees Celsius, for example, it is set to 32.5. The specific heat capacity coefficient of a material is a priori physical parameter, calculated by multiplying the specific heat capacity of the material by its density. This indicates the time step, which is consistent with the interval of sensor data acquisition, and is uniformly measured in seconds, such as 15. This represents the decoded illumination intensity of the v-th node, which is the decoded value from the illumination sensor.

[0116] In the specific implementation, the temperature value is decoded. With decoding light intensity The decoding method adopts accomplish, The weight matrix for decoding geometric parameters is a trainable parameter. The weight vector is a geometric parameter decoder and consists of trainable parameters.

[0117] It should be noted that the law of conservation of energy requires that the heat inflow into the curtain wall system equals the change in heat storage. This characterization represents the heat conduction through the curtain wall components. The term characterizes the increase in thermal energy due to light radiation, where the material's thermal capacity coefficient is... Integrating the heat capacity and absorption coefficient of the material, and the physical constraint loss term The temperature and decoded light intensity corresponding to the forced prediction geometric parameters satisfy the energy balance, thereby ensuring that the reverse modeling results conform to the basic principles of thermodynamics. Physical constraints are directly integrated into the machine learning loss function, making the model output not only geometrically accurate but also physically reasonable, improving engineering credibility and practicality, and avoiding problems that may violate physical laws in conventional methods.

[0118] S104: Input the dynamic graph structure into the machine learning-based geometric parameter prediction model, perform iterative training, calculate the multi-objective mixed loss function and parameter gradient, and update the parameters using the Adam optimizer.

[0119] In this embodiment, under small-sample training scenarios, the inverse modeling model is prone to overfitting the training data, resulting in insufficient generalization ability under missing data or noise interference. Conventional regularization methods such as Dropout or weight decay are difficult to effectively constrain the quality of the generated distribution and cannot ensure the consistency between the predicted geometric parameters and the true distribution. In this invention, a machine learning-based geometric parameter prediction model is constructed by inputting a dynamic graph structure, and an adversarial training framework is built using an attribute discriminator to distinguish the correlation between the true geometric attributes and the prediction results. By constraining the generated distribution through adversarial loss, the robustness and generalization ability of the model under missing data are improved.

[0120] The discriminator function is defined as follows: it receives the concatenated vector of the true geometric attribute feature vector and the first-level pyramid features, performs a linear transformation through the discriminator weight matrix, and outputs true / false probability values ​​using the sigmoid activation function. This distinguishes the correlation between true and predicted geometric attributes and features, expressed as:

[0121] ;

[0122] in, For the discriminator function, The discriminant's judgment result for the v-th node is represented by an output value between 0 and 1. The closer the value is to 1, the more likely the discriminant considers the input to be real data. This refers to the set of parameters for the discriminator, specifically the set of trainable parameters. This represents the discriminator weight matrix, with dimension 1. , are trainable parameters; This represents the Sigmoid activation function.

[0123] It should be noted that, The representation identifies the true geometric feature pairs, and the target output is 1. The representation discriminates and predicts geometric feature pairs, with the target output being 0. Both share the same discriminator, and the prediction geometric parameter vector is improved through adversarial training. Distribution reliability.

[0124] The model training uses the backpropagation algorithm and gradient descent optimizer to minimize the mixed loss function. At the beginning of the training process, all trainable parameters are initialized. Each iteration inputs batch data, including the node feature matrix, the edge weight matrix composed of the weights of the edges between each node, the time series corresponding to the sensor data, and the true geometric attribute feature matrix. The forward propagation process of the model calculates the node features, dynamic graph structure, spatiotemporal coding features, and pyramid features, and outputs the predicted geometric parameter vector.

[0125] Then, the mixed loss function is calculated, the backpropagation process of the model calculates the gradient of the loss with respect to the parameters, and the parameters are updated using the Adam optimizer.

[0126] Training iterations continue until a stopping condition is met, such as reaching the maximum number of iterations or the loss function converging, meaning the loss value no longer decreases significantly on the validation set. During training, the validation set is used to monitor model performance to prevent overfitting, and training is stopped when the validation loss does not improve for several consecutive epochs.

[0127] In one embodiment of this application, the multi-objective hybrid loss function includes geometric reconstruction loss, physical constraint loss, adversarial loss, and weight regularization loss. The geometric reconstruction loss is calculated based on the predicted geometric attribute matrix and the real geometric attribute matrix. The adversarial loss is calculated based on the discriminant output of the attribute discriminator on the real geometric attribute matrix, the predicted geometric attribute matrix, and the features of the first layer pyramid. The weight regularization loss is calculated based on the weight matrix of each layer of the spatiotemporal graph convolutional network.

[0128] In one embodiment of this application, the reverse modeling process needs to simultaneously satisfy geometric reconstruction accuracy, consistency of physical laws, and model generalization ability. Conventional single mean squared error loss cannot couple with prior physical knowledge and is prone to overfitting in small sample scenarios, resulting in prediction results with small numerical errors but impractical engineering. Therefore, this invention integrates geometric reconstruction loss, physical constraint loss, adversarial loss, and weight regularization loss through a multi-objective loss function to form a hybrid loss function as the overall optimization objective, thereby achieving multi-task constraint integration.

[0129] The geometric reconstruction loss is obtained by calculating the L1 norm error between the predicted and true geometric attribute matrices. The L1 norm is used to enhance robustness to outliers and measure the accuracy of the predicted geometric parameters, expressed as:

[0130] ;

[0131] in, This represents the geometric reconstruction loss, which measures the deviation between the predicted geometric parameters and the true values. This represents the L1 norm.

[0132] Based on the discriminator's discriminative outputs on the true and predicted geometric attribute matrices and the features of the first-layer pyramid, adversarial loss is calculated. This adversarial loss is calculated using a logarithmic function and an expectation operation, measuring the discriminator's discriminative ability. Adversarial training further improves the reliability of the predicted geometric parameter distribution, and is expressed as:

[0133] ;

[0134] in, The adversarial loss is a scalar that measures the discriminator's ability to distinguish between true and predicted geometric attributes. This represents the expected operation, which corresponds to the mean of the batch data in actual calculations. That is, the mean of all nodes within the batch is taken during training. It is a logarithmic function, with the default base being the natural constant; The geometric attribute feature matrix representing all nodes is a stack of the true geometric parameters of all nodes, and the geometric attribute feature vectors are also represented. yes The vth element; The predicted geometric attribute matrix for all nodes is a stack of predicted geometric parameters for all nodes, and the predicted geometric parameter vector is a vector representing the predicted geometric attributes of all nodes. yes The vth element.

[0135] It should be noted that, The discriminant's judgment results on the true geometric attributes and corresponding features. The discriminant's judgment results on predicted geometric attributes and corresponding features. Item and The terms represent the discrimination results of the true geometric attribute matrix and predicted geometric attribute matrix of all nodes, respectively, and the feature pyramid matrix. This is a batch-level aggregation, meaning it's an average or sum of the discrimination outputs of all nodes within a batch. Item and The item is the judgment result of a single node. In actual calculations, Similarly, The global discrimination result is the arithmetic mean of the local discrimination results.

[0136] It should also be noted that the adversarial loss implements the joint feature-attribute discrimination constraint. The input to the discriminator is a concatenated vector of geometric attributes and pyramid features, and the adversarial loss forces the prediction of the geometric parameter vector. Features of the first layer pyramid of the spatiotemporal graph convolutional network The correlation is matched with the true distribution, thereby decoupling irrelevant variations and suppressing the first-layer pyramid features of the spatiotemporal graph convolutional network. The noise in the image interferes with the geometric reconstruction.

[0137] The weight regularization loss is obtained by calculating the sum of the Frobenius norms of the weight matrices of each layer in the spatiotemporal graph convolutional network. The Frobenius norm is used to penalize and constrain the weight magnitude to prevent overfitting, as expressed in:

[0138] ;

[0139] in, This represents the weighted regularization loss; Represents the spatiotemporal graph convolutional network. The weight matrix of the layer has dimensions of , are trainable parameters; This represents the Frobenius norm.

[0140] It should be noted that the weight matrix It is the core transformation matrix of graph convolution. Term constraint weight matrix The Frobenius norm is equivalent to restricting the weight magnitude, thus improving generalization, while the temporal convolution kernel parameters are not included in the weight regularization loss. This limits the computational scope and avoids weakening dynamic modeling capabilities.

[0141] The geometric reconstruction loss, physical constraint loss, adversarial loss, and weighted regularization loss are weighted and summed to form a hybrid loss function as the overall optimization objective. The weighting coefficients balance the importance of different loss terms, achieving multi-task constraint integration, expressed as:

[0142] ;

[0143] in, This represents the mixed loss function, which is the total loss function for model training; The physical constraint loss weights are used to balance the importance of the law of conservation of energy, and are usually optimized based on experimental tuning. 3; To counteract the loss of regularization weights and improve the model's generalization ability, the optimal settings are: ; The weights are used for weight regularization to prevent overfitting; the optimal settings are as follows. .

[0144] S105: Construct a real-time dynamic graph structure based on basic BIM component data and real-time IoT sensor data, input the real-time dynamic graph structure into a trained machine learning-based geometric parameter prediction model, output the predicted geometric parameter vector, and perform intelligent green curtain wall maintenance.

[0145] In this embodiment, the initial geometric attributes and topological connections of the curtain wall components are loaded from the BIM model as a basic reference. Real-time IoT sensor data, including temperature, humidity, wind pressure, and illumination, are collected and preprocessed into temporal differential features. Then, a dynamic graph structure is constructed, where node features are fused with geometric attributes and sensor temporal differentials, and edge weights are calculated based on spatial coordinates and topological connections. The processed data is then input into a trained machine learning-based geometric parameter prediction model. Through a spatiotemporal feature encoder, a spatiotemporal graph convolutional network, a multi-scale feature pyramid, and a physical constraint decoder module, the model outputs a predicted geometric parameter vector. The prediction result represents the current geometric state of each component, such as dimensional changes and deformation. The reverse modeling results can be used for structural health monitoring, maintenance decision support, or BIM model updates. In one embodiment, for a glass panel, the model may predict that its length and width will slightly increase due to thermal expansion, thereby detecting potential structural changes.

[0146] In one embodiment of this application, different methods are compared in terms of geometric reconstruction accuracy to evaluate the geometric reconstruction accuracy of different methods in the reverse modeling task of intelligent green curtain walls. The results are as follows: Figure 2 As shown in the figure, five methods were compared: direct stitching, graph convolutional network, temporal convolutional network, no physical constraints, and the method of this invention. Geometric reconstruction accuracy was measured by mean absolute error (MAE), measured in meters. A lower error value indicates a more accurate reconstruction result, closer to the true geometric parameters. Box plots are used to show the error distribution of each method; the boxes represent the interquartile range, the median represents the median, the dashed lines represent the data range, and individual points may represent outliers. The figure shows that the MAE of the method of this invention is significantly lower than the other comparison methods. The direct stitching method has a wider error distribution and a higher median, indicating insufficient reconstruction accuracy due to the simple stitching of heterogeneous features. Graph convolutional network and temporal convolutional network methods show improved errors, but still exhibit significant volatility, indicating insufficient stability when processing complex spatiotemporal data. While the no-physical-constraint method has a lower error, it has more outliers, potentially leading to inaccurate predictions in certain situations due to the lack of physical constraints. In contrast, the method of this invention features a compact and lowest-positioned box, with both the median and overall distribution concentrated in the low-error region. This demonstrates that the accuracy and stability of geometric reconstruction are significantly improved through dynamic graph construction, heterogeneous feature fusion, and a physically constrained decoder. This reflects the advantages of this invention in capturing spatial-temporal correlations between components and ensuring physical plausibility.

[0147] In one embodiment of this application, a comparison is made of the degree to which different methods satisfy physical constraints, comparing the performance of different methods in satisfying physical constraints, especially the law of conservation of energy. The results are as follows: Figure 3As shown in the figure, the experiment compared five methods: direct splicing, graph convolutional network, temporal convolutional network, no physical constraint, and the method of this invention. The degree of satisfaction of physical constraints was measured by energy conservation deviation, in kilojoules. The lower the deviation value, the more the prediction result conforms to the basic principles of thermodynamics. Box plots are used in the figure to show the deviation distribution of each method. The boxes represent the main data range, the dashed lines show the overall distribution, and outliers may indicate large deviations in some cases. The experimental results show that the energy conservation deviation of the method of this invention is much lower than that of other methods. The direct splicing method has the highest deviation and a wide distribution, indicating that this method completely ignores physical laws, resulting in prediction results that often violate energy conservation. The graph convolutional network and temporal convolutional network methods have reduced deviations, but still have high medians and many outliers, indicating that these methods partially capture physical relationships but not sufficiently. The no-physical-constraint method has a higher deviation, because although this method may have high accuracy, the lack of physical constraints leads to large deviations in the energy calculation of the predicted values. The method of this invention produces the most compact and lowest-positioned box with almost no outliers, indicating that by using a physical constraint decoder and energy conservation loss term, the prediction results are effectively ensured to conform to the laws of thermodynamics, thus improving engineering credibility and practicality.

[0148] In one embodiment of this application, the running time and accuracy of various methods are compared under different numbers of components. The scalability of different methods in handling curtain wall systems of different scales is analyzed, including running efficiency and reconstruction accuracy. The results are as follows: Figure 4 , Figure 5 As shown in the figure, five methods were compared in the experiment: direct stitching method, graph convolutional network method, temporal convolutional network method, method without physical constraints, and the method of this invention. The number of components increased from few to many, and the running time was measured in seconds; a lower time indicates higher efficiency. Reconstruction accuracy is a dimensionless indicator; a higher value indicates better accuracy. In terms of running time, the direct stitching method had the shortest time, but this came at the cost of accuracy. The graph convolutional network method and the temporal convolutional network method had faster time increases, indicating higher computational complexity. The method without physical constraints had a moderate time, but its accuracy was unstable. The time increase of the method of this invention was reasonable; although slightly higher than the direct stitching method, it was much lower than other complex methods, indicating that a balance between efficiency and performance was achieved by optimizing the spatiotemporal encoder and pyramid structure. Regarding reconstruction accuracy, the accuracy of all methods decreased slightly with the increase of the number of components, but the accuracy of the method of this invention remained at the highest level with the smallest decrease, indicating that its multi-scale feature fusion and physical constraints effectively maintained the modeling accuracy of large-scale systems, demonstrating the high efficiency and reliability of this invention in practical engineering applications.

[0149] In one embodiment of this application, a comparative analysis of transient change capture capabilities is performed. The dynamic response of a curtain wall under a sudden temperature change environment is simulated, and the capture capabilities of different methods for transient changes in component length (unit: millimeters) are compared. The results are as follows: Figure 6 As shown. The horizontal axis represents time (seconds) covering a 200-second observation period, including three typical abrupt changes (three-color background regions). The experiment uses 500Hz high-frequency sampling to obtain the true change curves. Experimental results show that conventional methods have a response lag of about 15 seconds in the temperature rise region (50-100 seconds), while this technique achieves near-synchronous tracking with a lag of less than about 3 seconds. Conventional methods still have a significant lag in the temperature drop region, while this technique accurately restores the extreme values ​​of change. At the same time, this technique retains the true fluctuation characteristics while filtering out invalid noise, indicating that the temporal difference operator enhances transient perception, and multi-scale convolution extracts short, medium, and long-range features in parallel, enabling the model to have both fast response capability and noise robustness.

[0150] In the aforementioned reverse modeling method for intelligent green curtain walls based on BIM-IoT fusion, firstly, BIM component data and IoT sensor data are acquired to construct a dataset; then, node feature fusion and edge weight calculation are performed based on the dataset to construct a dynamic graph structure; next, a machine learning-based geometric parameter prediction model is constructed, which includes a spatiotemporal feature encoder module, a spatiotemporal graph convolutional network module, a multi-scale feature pyramid module, and a physical constraint decoder module; then, the dynamic graph structure is input into the machine learning-based geometric parameter prediction model for iterative training, calculating a multi-objective hybrid loss function and parameter gradients, and updating parameters using the Adam optimizer; finally, a real-time dynamic graph structure is constructed based on basic BIM component data and real-time IoT sensor data, and the real-time dynamic graph structure is input into the trained machine learning-based geometric parameter prediction model to output predicted geometric parameter vectors for intelligent green curtain wall maintenance. In other words, by heterogeneously fusing BIM geometric attributes with IoT sensing data, a dynamic feature modeling method based on temporal difference and geometric attribute splicing is proposed, realizing the collaborative expression of static geometry of components and transient environmental changes. A dynamic graph structure combining topological relationships and spatial distances is constructed, and an adaptive attention mechanism is introduced to dynamically adjust adjacency relationships, enabling the modeling process to simultaneously possess physical rationality and implicit spatiotemporal dependency expression capabilities. A multi-scale geometric perception feature pyramid structure is proposed, which fully preserves local geometric details while ensuring global semantics, solving the problem that traditional feature extraction methods easily lose key geometric information such as curvature and hole positions. Physical laws such as energy conservation are introduced into the decoding and loss function design, and adversarial regularization and multi-objective optimization are combined to achieve a unity of geometric accuracy and physical rationality, and improve the robustness and generalization ability of the model under small sample conditions.

[0151] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0152] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program can implement a reverse modeling method for intelligent green curtain walls based on BIM-IoT integration. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0153] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0154] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0155] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0156] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0157] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0158] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0159] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0160] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A BIM-IoT fusion-based intelligent green curtain wall reverse modeling method, characterized in that, The method comprises: acquiring BIM component data and IoT sensor data, and constructing a dataset; performing node feature fusion and edge weight calculation based on the dataset, and constructing a dynamic graph structure; constructing a machine learning-based geometric parameter prediction model, which comprises a spatio-temporal feature encoder module, a spatio-temporal graph convolution network module, a multi-scale feature pyramid module, and a physical constraint decoder module, wherein the spatio-temporal feature encoder module is used to independently process spatial and temporal features respectively, and adopts multi-scale convolution to capture dynamic changes; the spatio-temporal graph convolution network module is used to integrate topological prior knowledge into self-attention calculation, dynamically adjust the connection weight between nodes, and fuse temporal features through a gating mechanism; the multi-scale feature pyramid module adopts a top-down path to fuse deep semantic features and shallow geometric features, and retains feature information at different levels through a skip connection; and the physical constraint decoder module is used to predict geometric parameters in combination with thermodynamic constraints; inputting the dynamic graph structure into the machine learning-based geometric parameter prediction model, performing iterative training, calculating a multi-objective hybrid loss function and a parameter gradient, and updating parameters using an Adam optimizer; based on the basic BIM component data and real-time IoT sensor data, constructing a real-time dynamic graph structure, inputting the real-time dynamic graph structure into the trained machine learning-based geometric parameter prediction model, outputting a predicted geometric parameter vector, and performing intelligent green curtain wall maintenance; the node feature fusion and edge weight calculation based on the dataset to construct a dynamic graph structure comprises: concatenating a static geometric attribute feature vector and a time series difference feature of dynamic sensor data to form a time-aware node feature vector, wherein the time series difference feature of dynamic sensor data is a difference between a current time sensor data feature vector and a previous time sensor data feature vector; calculating edge weights based on geometric distances between nodes and a topological connection matrix, wherein the geometric distances are calculated using Euclidean norm, and the topological connection matrix represents whether there is a physical connection between components.

2. The BIM-IoT fusion-based intelligent green curtain wall reverse modeling method according to claim 1, characterized in that, the spatio-temporal feature encoder module comprises: aggregating neighbor information of each node based on edge weights and node feature vectors through graph convolution operation, and extracting a spatial encoding feature vector using a ReLU activation function; extracting short, medium and long range temporal features in parallel using one-dimensional convolution operation with different kernel sizes, and fusing the short, medium and long range temporal features through a max-pooling operation to form a multi-scale temporal feature vector.

3. The BIM-IoT fusion-based intelligent green curtain wall reverse modeling method according to claim 2, characterized in that, the spatio-temporal graph convolution network module comprises: based on the spatial encoding feature vector and the prior edge weight, calculating the dynamic correlation strength between nodes through the self-attention mechanism to generate a dynamic adjacency matrix.

4. The BIM-IoT fusion-based intelligent green curtain wall reverse modeling method according to claim 3, characterized in that, the spatio-temporal graph convolution network module further comprises: updating the node feature matrix based on the dynamic adjacency matrix through graph convolution operation, and injecting the multi-scale temporal feature vector into the node feature matrix through a gating mechanism.

5. The BIM-IoT fusion-based intelligent green curtain wall reverse modeling method according to claim 4, characterized in that, the multi-scale feature pyramid module comprises: performing an upsampling operation based on the node feature matrix of the third layer of the spatio-temporal graph convolution network to expand the feature size and obtain high-level pyramid features; The node feature matrix of the second layer of the spatio-temporal graph convolution network is spliced with the high-level pyramid feature, and feature fusion is performed through convolution operation to generate a middle-level pyramid feature; The spatial encoding feature matrix is spliced with the middle-level pyramid feature, and feature fusion is performed through convolution operation to generate a first-level pyramid feature.

6. The BIM-IoT fusion-based intelligent green curtain wall reverse modeling method according to claim 5, characterized in that, The physical constraint decoder module comprises: A predicted geometry parameter vector is decoded based on the first-level pyramid feature and a multi-scale time sequence feature vector through a multi-layer perception machine; A temperature value and an illumination intensity are decoded based on the predicted geometry parameter vector, and a physical constraint loss is calculated based on the temperature value and the illumination intensity.

7. The BIM-IoT fusion-based intelligent green curtain wall reverse modeling method according to claim 6, characterized in that, The calculation formula of the physical constraint loss is: ； wherein, represents a physical constraint loss term, represents a heat transfer coefficient of the v-th node, represents a decoded temperature value of the v-th node, represents an external environment temperature, represents a heat capacity coefficient of the material, represents a time step, represents a decoded light intensity of the v-th node, v being a node index, represents an L2 norm.

8. The BIM-IoT fusion-based intelligent green curtain wall reverse modeling method according to claim 1, characterized in that, The multi-target hybrid loss function comprises a geometry reconstruction loss, a physical constraint loss, an adversarial loss and a weight regularization loss, wherein the geometry reconstruction loss is calculated based on a predicted geometry attribute matrix and a real geometry attribute matrix, the adversarial loss is calculated based on the discrimination output of an attribute discriminator on the real geometry attribute matrix, the predicted geometry attribute matrix and the discrimination output of the first-level pyramid feature, the weight regularization loss is calculated based on the weight matrix of each layer of the spatio-temporal graph convolution network, and the discrimination output of the first-level pyramid feature is the first-level pyramid feature of the spatio-temporal graph convolution network, which is generated by the multi-scale feature pyramid module.

Images