Integrated decoration multi-source heterogeneous data cloud processing and prediction system
By combining spatiotemporal feature embedding and dynamic knowledge graph construction, the problem of spatiotemporal mismatch of multi-source heterogeneous data is solved, enabling efficient and accurate prediction of decoration construction progress and material consumption, and improving the adaptability and accuracy of the prediction system.
Patent Information
- Application Number
- CN202611245942.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing integrated decoration cloud prediction systems fail to effectively consider the inherent relationship between drawing spatial coordinates, material physical properties, and construction time flow when processing drawing topology data, material attribute data, and construction sensor flow data. This leads to spatiotemporal mismatch of multi-source heterogeneous data, affecting the prediction network's perception of the dynamic evolution logic of construction procedures and causing prediction bias.
By using a spatiotemporal feature joint embedding structure, drawing topology data, material attribute data, and construction sensor flow data are mapped to the same high-dimensional feature space to construct a dynamic knowledge graph. Dynamic graph convolution and temporal prediction structures are used to update node states with construction time sequence as the driving force, thereby realizing the spatiotemporal alignment of heterogeneous data and the synchronous coupling of dynamic knowledge graph.
It improves the accuracy of prediction results, reduces redundant calculations, ensures that the prediction process conforms to the physical process evolution logic of decoration construction, and enhances the independence and accuracy of construction progress and material loss prediction.
Smart Images

Figure CN123115649A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electrical data processing technology, specifically relating to an integrated decoration multi-source heterogeneous data cloud processing and prediction system. Background Technology
[0002] Existing integrated cloud-based decoration prediction systems process drawing topology data, material attribute data, and construction sensor flow data by directly stitching together independently extracted features. Drawing topology data is processed using a graph neural network to extract spatial structure features; material attribute data is processed using a fully connected layer to extract discrete attribute features; and construction sensor flow data is processed using a recurrent neural network to extract temporal features. These three independently extracted feature vectors are then directly stitched together at the feature layer and input into the prediction network. The prediction network outputs predictions of decoration construction progress and material waste based on the stitched features.
[0003] The above-mentioned method of directly splicing after independent extraction does not take into account the inherent relationship between the spatial coordinates of the drawings, the physical properties of the materials, and the construction time flow. Due to differences in sampling frequency and representation dimensions, heterogeneous data suffers from spatiotemporal mismatch, which causes the prediction network to be unable to perceive the dynamic evolution logic of the construction process. This presents the core technical problem of prediction bias caused by spatiotemporal mismatch of multi-source heterogeneous data. Summary of the Invention
[0004] The purpose of this invention is to provide an integrated cloud-based system for processing and predicting multi-source heterogeneous data in decoration, which can solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] An integrated cloud-based system for processing and predicting multi-source heterogeneous data in decoration includes: a data receiving structure configured to receive drawing topology data, material attribute data, and construction sensor stream data from a decoration scene; a spatiotemporal feature joint embedding structure configured to map the spatial coordinate information in the drawing topology data, the physical attribute dimension in the material attribute data, and the timestamp in the construction sensor stream data to the same high-dimensional feature space to achieve spatiotemporal alignment of heterogeneous data, wherein the spatiotemporal embedding dimension corresponds to the physical coordinates of the decoration entity and the progress of the construction stage; a dynamic knowledge graph construction structure configured to construct a dynamic knowledge graph of the decoration process based on the spatiotemporally aligned feature stream, wherein nodes in the dynamic knowledge graph represent decoration entities and materials, and edges represent construction procedures and physical associations; and a dynamic graph convolution and temporal prediction structure configured to update the node state of the dynamic knowledge graph driven by the construction time sequence, and input the updated graph features into a temporal prediction network to output a joint prediction result of decoration construction progress and material loss, wherein the dynamic graph update is synchronously coupled with the physical progress of the construction procedures.
[0007] Preferably, the data receiving structure is further configured to perform time-scale alignment and spatial benchmark unification on the drawing topology data, the material attribute data, and the construction sensor stream data; extract the absolute coordinate sequence from the drawing topology data; parse the relative timestamps from the construction sensor stream data; anchor the absolute coordinate sequence and the relative timestamps based on the decoration construction process list; and generate a multi-source data stream under a unified spatiotemporal benchmark. The multi-source data stream under the unified spatiotemporal benchmark serves as the input to the spatiotemporal feature joint embedding structure, eliminating sampling frequency differences and clock drift between heterogeneous data sources.
[0008] Preferably, the spatiotemporal feature joint embedding structure is configured to construct a joint embedding network including a spatial location encoder, an attribute encoder, and a time encoder. The spatial coordinate information is input into the spatial location encoder to generate a spatial feature vector, the physical attribute dimension is input into the attribute encoder to generate an attribute feature vector, and the timestamp is input into the time encoder to generate a time feature vector. A cross-attention fusion operation is performed on the spatial feature vector, the attribute feature vector, and the time feature vector. The fused high-dimensional vector is used as a unified representation in the high-dimensional feature space. The weight matrix of the cross-attention fusion operation is constrained by the physical coordinates of the decorative entity.
[0009] Preferably, the dynamic knowledge graph construction structure is configured to divide decorative entities and materials into region-level nodes and component-level nodes. The region-level nodes and the component-level nodes are connected by inclusion relationship edges. Construction procedures are transformed into temporal execution edges, and physical associations are transformed into spatial constraint edges. The feature vectors of the region-level nodes and the component-level nodes are initialized based on the unified representation. An adjacency matrix is constructed based on the temporal execution edges and the spatial constraint edges. The element values of the adjacency matrix change dynamically with the update of the construction sensing flow data to reflect the topological evolution of the decoration construction process.
[0010] Preferably, the dynamic graph convolution and temporal prediction structure is configured to introduce a temporal gating mechanism during the node state update process of the dynamic knowledge graph. The temporal gating mechanism controls the fusion ratio of historical node states and current input features. The gating value of the temporal gating mechanism is jointly determined by the interval of construction time sequence and the attribute features of the node. The node features updated by the temporal gating mechanism are then subjected to graph pooling to obtain a global construction state vector, which is then input into the temporal prediction network.
[0011] Preferably, the time-series prediction network includes a progress prediction branch and a loss prediction branch. The global construction state vector is decoupled in terms of features. The decoupled time-dependent features are input into the progress prediction branch, and the decoupled material-related features are input into the loss prediction branch. The progress prediction branch and the loss prediction branch share the underlying time-series convolutional kernels. Orthogonal constraints are introduced in the output layer to minimize the feature interference between the progress prediction branch and the loss prediction branch. The progress prediction branch outputs the construction progress time-series curve, and the loss prediction branch outputs the material loss distribution matrix.
[0012] Preferably, the process of anchoring the absolute coordinate sequence and the relative timestamp based on the decoration construction process list includes: identifying event trigger tags in the construction sensor stream data; matching the event trigger tags with key milestone nodes in the decoration construction process list; performing reverse time-series calibration on the absolute coordinate sequence in the drawing topology data based on the time reference of the successfully matched key milestone nodes; correcting the timestamp offset in the construction sensor stream data caused by transmission delay; and ensuring the consistency of multi-source data streams under the unified spatiotemporal reference at the event triggering level.
[0013] Preferably, the spatial position encoder uses a multi-scale radial basis function to encode the spatial coordinate information, capturing local spatial topological features and global spatial distribution features. The time encoder uses a periodic trigonometric function transformation to process the timestamp to extract construction cycle features. The cross-attention fusion operation introduces a spatial distance attenuation mask, which suppresses the attention weights between distant nodes based on the physical distance between decorative entities, so that the unified representation conforms to the spatial physical attenuation law of the decorative scene during the fusion stage.
[0014] Preferably, the process of dynamically changing the element values of the adjacency matrix as the construction sensor flow data is updated includes: when the construction sensor flow data indicates that a certain component-level node has completed installation, cutting off the temporal execution edge associated with the component-level node, activating the spatial constraint edge between the component-level node and adjacent completed nodes, and simultaneously injecting a state decay factor into the component-level node. The state decay factor reduces the message passing intensity of the component-level node in the dynamic knowledge graph as the construction timeline progresses, thus concentrating computational resources on uncompleted nodes.
[0015] Preferably, the process of determining the gate value of the time-series gating mechanism by the interval of construction time sequence and the attribute features of nodes includes: calculating the reciprocal of the interval of construction time sequence as a time sensitivity coefficient; extracting the material activity vector from the attribute features of nodes; performing a Hadamard product operation on the time sensitivity coefficient and the material activity vector; mapping the operation result to a gate value through a S-shaped activation function; and increasing the gate value when the interval of construction time sequence increases and the material activity vector represents the construction activity state, so as to improve the fusion ratio of the current input features.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0017] 1. By using a spatiotemporal feature joint embedding structure, the spatial coordinate information in the drawing topology data, the physical attribute dimension in the material attribute data, and the timestamp in the construction sensor flow data are mapped to the same high-dimensional feature space, completing the spatiotemporal alignment of heterogeneous data and overcoming the spatiotemporal mismatch defects caused by independent extraction. Based on the spatiotemporally aligned feature flow, a dynamic knowledge graph is constructed. The dynamic graph convolution and temporal prediction structure update the node state of the dynamic knowledge graph with the construction time sequence as the driving force, so that the prediction process fits the physical process evolution logic of decoration construction, avoids the problem that directly splicing features cannot perceive the construction dynamics, and improves the accuracy of the prediction results.
[0018] 2. A temporal gating mechanism is introduced to control the fusion ratio of historical node states and current input features. Combined with a state decay factor, the message transmission intensity of completed component-level nodes is reduced during construction sequence advancement, concentrating computational resources on unfinished nodes and reducing redundant computation. A spatial distance decay mask is introduced in the cross-attention fusion operation to suppress the attention weights between distant nodes, making feature fusion conform to the spatial physical decay law of the decoration scene. Feature decoupling and orthogonal constraints minimize feature interference between progress prediction and loss prediction, ensuring the independence of multi-task prediction outputs. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the overall operation of the integrated decoration multi-source heterogeneous data cloud processing and prediction system of the present invention.
[0020] Figure 2 This is a flowchart of the unified spatiotemporal reference processing for the data receiving structure of the present invention;
[0021] Figure 3 This is a flowchart of the spatiotemporal feature joint embedding network encoding and fusion process of the present invention;
[0022] Figure 4 This is a flowchart illustrating the construction and dynamic updating of the dynamic knowledge graph of the decoration process according to the present invention.
[0023] Figure 5This is a flowchart of the dynamic graph convolution and temporal gating node state update process of the present invention;
[0024] Figure 6 This is a flowchart of the multi-task prediction output of the temporal prediction network of the present invention. Detailed Implementation
[0025] refer to Figure 1 In one embodiment, an integrated decoration multi-source heterogeneous data cloud processing and prediction system is deployed on a cloud server cluster, establishing communication connections with various data acquisition devices at the decoration construction site through an edge computing gateway. The data receiving structure receives three types of heterogeneous data from the decoration scene through the edge gateway: drawing topology data, material attribute data, and construction sensor flow data. The drawing topology data originates from vector format files exported from Building Information Modeling (BIM) software, containing the three-dimensional spatial coordinates, geometric contour parameters, and topological connections between all entities in the decoration project. The spatial coordinate information is represented by absolute coordinate values in the world coordinate system, and the topological connections include spatial relationships such as inclusion, adjacency, and connection between entities. The material attribute data originates from a decoration material database, containing the physical attribute dimensions of various decoration materials, specifically including continuous attributes such as density, thermal conductivity, compressive strength, abrasion resistance, water absorption rate, theoretical consumption per unit area, and drying time, as well as discrete attributes such as material type, fire resistance rating, and environmental protection rating. Construction sensor stream data originates from various data acquisition devices deployed at the construction site, including accelerometers installed on construction machinery, RFID readers deployed in material storage areas, smart bracelets worn by workers, and process execution data uploaded by handheld terminals. The construction sensor stream data exists in time series form, and each data sample includes a relative timestamp, equipment identifier, data type, and numerical information.
[0026] The spatiotemporal feature joint embedding structure receives three types of heterogeneous data from the data receiving structure. It maps the spatial coordinate information from the drawing topology data, the physical attribute dimension from the material attribute data, and the timestamps from the construction sensor flow data to the same high-dimensional feature space, completing the spatiotemporal alignment of the heterogeneous data. The spatiotemporal feature joint embedding structure constructs a joint embedding network containing a spatial position encoder, an attribute encoder, and a time encoder. These encoders encode the core features of the three types of data, generating spatial feature vectors, attribute feature vectors, and time feature vectors of the same dimension. The joint embedding network performs a cross-attention fusion operation on the three types of feature vectors, generating a fused high-dimensional vector as a unified representation in the high-dimensional feature space. The weight matrix of the cross-attention fusion operation is constrained by the physical coordinates of the decorative entity, ensuring that the spatiotemporal embedding dimension corresponds to the physical coordinates of the decorative entity and the construction stage progress.
[0027] The dynamic knowledge graph construction structure receives a unified representation stream output from a spatiotemporal feature joint embedding structure, and constructs a dynamic knowledge graph of the decoration process based on this feature stream. Nodes in the dynamic knowledge graph represent decoration entities and materials, while edges represent construction procedures and physical relationships. The dynamic knowledge graph construction structure divides decoration entities and materials into region-level nodes and component-level nodes. Region-level nodes correspond to functional zones in the decoration project, and component-level nodes correspond to specific decoration components and materials within those functional zones. Region-level nodes and component-level nodes are connected by inclusion edges, indicating that a component belongs to a specific functional zone. The dynamic knowledge graph construction structure transforms decoration construction procedures into temporal execution edges, representing the sequential execution order of procedures, and transforms the physical relationships between decoration entities into spatial constraint edges, representing the physical connections or positional constraints between entities. The dynamic knowledge graph construction structure initializes the feature vectors of region-level nodes and component-level nodes based on the unified representation, and constructs an adjacency matrix based on temporal execution edges and spatial constraint edges. The element values of the adjacency matrix represent the connection strength between nodes, and the element values change dynamically with the updates of construction sensor flow data to reflect the topological evolution of the decoration construction process.
[0028] The dynamic graph convolutional and temporal prediction structure receives the dynamic adjacency matrix and node feature vectors output by the dynamic knowledge graph construction structure, and updates the node states of the dynamic knowledge graph driven by the construction timeline. During the node state update process, the dynamic graph convolutional and temporal prediction structure introduces a temporal gating mechanism to control the fusion ratio of historical node states and current input features. The gating value of the temporal gating mechanism is jointly determined by the interval of the construction timeline and the attribute features of the node. The dynamic graph convolutional and temporal prediction structure performs graph pooling on the node features updated by the temporal gating mechanism, aggregating the local features of all nodes to obtain a global construction state vector. The global construction state vector is input into the temporal prediction network, which includes a progress prediction branch and a loss prediction branch, outputting a construction progress timeline curve and a material loss distribution matrix, respectively. The dynamic update process of the knowledge graph is synchronously coupled with the physical advancement of the construction procedures, ensuring that changes in node states reflect the actual progress of the construction process in real time.
[0029] In this embodiment, the system achieves spatiotemporal alignment by mapping three types of heterogeneous data to the same high-dimensional feature space, avoiding the spatiotemporal mismatch problem caused by directly splicing independently extracted features. Based on the spatiotemporally aligned features, a dynamic knowledge graph is constructed, which can accurately represent the dynamic relationships between entities during the decoration construction process. A node state update mechanism driven by construction time sequence enables the prediction process to conform to the physical process evolution logic of decoration construction, providing an accurate feature basis for the joint prediction of construction progress and material loss.
[0030] refer to Figure 2In a preferred embodiment, after receiving three types of heterogeneous data, the data receiving structure performs time-scale alignment and spatial benchmark unification processing on the drawing topology data, material attribute data, and construction sensor flow data to generate a multi-source data stream under a unified spatiotemporal benchmark, which serves as the input to the spatiotemporal feature joint embedding structure. The data receiving structure first extracts the absolute coordinate sequence from the drawing topology data, traverses all decorative entities in the building information model file, extracts the three-dimensional absolute coordinate values of the vertices of each entity, and organizes them into an absolute coordinate sequence according to entity type and topological relationship. Each coordinate sequence corresponds to the spatial location information of a decorative entity. The data receiving structure parses the relative timestamps in the construction sensor flow data, converting the relative timestamp of each data sample into an absolute time value based on the commencement time of the decoration project, while simultaneously recording the sampling frequency and equipment clock offset of each data sample.
[0031] The data receiving structure anchors absolute coordinate sequences and relative timestamps based on a list of decoration construction procedures. This list, pre-stored in a cloud database, includes all construction procedures for the decoration project, their execution order, key milestone nodes for each procedure, and the spatial area corresponding to each milestone node. The data receiving structure identifies event trigger tags in the construction sensor stream data. These tags are automatically generated by on-site acquisition equipment when specific construction events occur, or manually uploaded by workers via handheld terminals. The structure performs string and semantic matching between the identified event trigger tags and the key milestone nodes in the decoration construction procedure list to determine the key milestone node corresponding to each event trigger tag. Based on the time reference of the successfully matched key milestone nodes, the data receiving structure performs reverse temporal calibration on the absolute coordinate sequences in the drawing topology data, assigning a corresponding construction time period to the absolute coordinate sequence of each decoration entity. This time period represents the expected execution time range of the construction procedure corresponding to that entity.
[0032] The data receiving structure corrects timestamp offsets caused by transmission delays in construction sensor stream data. For each construction sensor stream data sample, the difference between the standard time of the key milestone node corresponding to its event trigger tag and the relative timestamp of the sample is calculated. This difference is used as the transmission delay correction amount to correct the timestamps of that sample and all subsequent samples uploaded by the same device. The data receiving structure uses a linear interpolation method to resample heterogeneous data with different sampling frequencies, unifying the sampling frequency of all data to a preset reference sampling frequency to eliminate sampling frequency differences between heterogeneous data sources. The data receiving structure uses a clock synchronization algorithm to correct clock drift of different acquisition devices. Based on the time reference of multiple key milestone nodes, the clock drift rate of each device is calculated, and the timestamps of all data samples uploaded by that device are dynamically corrected according to the clock drift rate.
[0033] In this embodiment, the correspondence between key milestone nodes in decoration construction and event trigger tags is shown in Table 1.
[0034] Table 1. Correspondence between Key Milestones and Event Trigger Tags in Decoration Construction
[0035] Construction phase Key milestones Event triggering tag Time base type Spatial reference area Preparation stage Construction drawings handover completed Drawing handover confirmation Absolute time base The entire construction area Grassroots processing stage Wall base leveling completed Wall leveling inspection passed relative time base Corresponding wall area Grassroots processing stage Ground base hardening completed Ground hardening test passed relative time base Corresponding ground area Finishing stage Tile laying begins Tile delivery inspection completed relative time base Corresponding paving area Finishing stage Wall latex paint application completed Latex paint drying test passed relative time base Corresponding wall area Installation phase Doors and windows installed Door and window sealing test passed relative time base Corresponding door and window areas Completion Phase Overall completion and acceptance completed Submission of completion acceptance report Absolute time base The entire construction area
[0036] Table 1 lists typical key milestones and their corresponding event trigger tags during the decoration construction process, along with the time reference type and corresponding spatial reference region for each milestone. The data receiving structure uses this correspondence to match event trigger tags with key milestones, ensuring the accuracy of the spatiotemporal anchoring process. For special construction procedures not listed in Table 1, the data receiving structure uses semantic similarity calculation to match their event trigger tags with the most similar key milestones and records the matching results for subsequent model optimization.
[0037] In this embodiment, the data receiving structure eliminates sampling frequency differences and clock drift between heterogeneous data sources by using time-scale alignment and unified spatial reference processing, ensuring the consistency of multi-source data streams in both time and space dimensions. Based on the anchoring mechanism of event-triggered tags and key milestone nodes, precise correspondence between drawing spatial information and construction time information is achieved, providing a unified spatiotemporal reference for subsequent joint embedding of spatiotemporal features.
[0038] refer to Figure 3 In a preferred embodiment, the joint embedding network constructed using the spatiotemporal feature joint embedding structure comprises three parallel encoding branches: a spatial location encoder, an attribute encoder, and a temporal encoder, as well as a cross-attention fusion layer. The spatial location encoder uses multi-scale radial basis functions to encode spatial coordinate information, capturing the local spatial topological features and global spatial distribution features of the decorative entity. For any three-dimensional spatial coordinate point of a decorative entity... Spatial position encoder adopts Encoding is performed using radial basis functions at different scales. The encoding process is represented by the following formula:
[0039]
[0040] in, Spatial coordinate points The corresponding spatial feature vector, For spatial feature dimensions; For the first The center coordinates of the radial basis functions are evenly distributed throughout the entire space of the decoration project; For the first The scaling parameter of each radial basis function controls the size of the spatial receptive field of that function; For the first The learnable weight matrix corresponds to each radial basis function. By adjusting the scaling parameter... The spatial position encoder can capture spatial features at different scales, including smaller ones, by taking values that allow it to capture spatial features at different scales. Corresponding to local fine spatial features, larger Corresponding to global distribution spatial characteristics.
[0041] The time encoder uses periodic trigonometric function transformations to process timestamp information, extracting periodic and long-term trend characteristics during the construction process. For any given timestamp... The encoding process of the time encoder is represented by the following formula:
[0042]
[0043] in, For timestamps The corresponding time feature vector, The time feature dimension; For the first The periodic parameters of the trigonometric functions correspond to different time periods in the decoration construction process, including daily, weekly, monthly, and the total construction period. Through periodic trigonometric function transformations, the time encoder can map continuous timestamps into periodic high-dimensional feature vectors, effectively capturing the time patterns in the construction process.
[0044] The attribute encoder uses a three-layer multilayer perceptron to encode material attribute data, mapping discrete and continuous attributes into attribute feature vectors of a unified dimension. For any material's physical attribute vector... ,in For the physical attribute dimension, the encoding process of the attribute encoder is represented by the following formula:
[0045]
[0046] in, Material property vector The corresponding attribute feature vector, For attribute feature dimensions; , The weight matrix is a learnable weight matrix; , A learnable bias vector; For the hidden layer dimension; This is a linear rectified activation function. The attribute encoder maps physical attributes of different dimensions and types into a unified high-dimensional feature representation through multiple nonlinear transformations.
[0047] The cross-attention fusion layer of the joint embedding network on spatial feature vectors Attribute feature vector With time feature vector Perform a cross-attention fusion operation. During the fusion process, a spatial distance attenuation mask is introduced to suppress the attention weights between distant nodes based on the physical distance between decorative entities, ensuring that the feature fusion conforms to the spatial and physical attenuation rules of the decorative scene. Spatial distance attenuation mask. The calculation process is represented by the following formula:
[0048]
[0049] in, For the first The decorative entity and the first Mask values between decorative entities; , The first The and the first The center coordinates of each decorative entity; This is the distance attenuation coefficient, controlling the degree to which spatial distance affects attention weights. As the physical distance between two decorative entities increases, the mask value... Reduce, thus suppressing the attention weights between them.
[0050] The computational process of the cross-attention fusion operation is represented by the following formula:
[0051]
[0052]
[0053]
[0054] in, For the query matrix, given the spatial feature vectors Obtained through linear transformation; , These are the key matrix and the value matrix, respectively, derived from the attribute feature vectors. With time feature vector The result is obtained by linear transformation after splicing; , , The weight matrix is a learnable weight matrix; For the dimensions of query and key; The dimension of the value; This refers to the Hadamard product operation; It is a normalized exponential function; This is the unified representation vector after fusion; The weight matrix is a learnable weight matrix; The number of decorative entities. The weight matrix of the cross-attention fusion operation is constrained by the physical coordinates of the decorative entities, and is achieved through a spatial distance attenuation mask. Implementing this constraint results in higher attention weighting between decorative entities that are closer together.
[0055] In this embodiment, the correspondence between the scale parameters of the multi-scale radial basis function encoding and the spatial feature capture range is shown in Table 2.
[0056] Table 2. Correspondence between scale parameters and spatial feature capture range of multi-scale radial basis function encoding.
[0057] Scale parameters (Unit: meters) Spatial feature capture range (unit: meters) Feature type Typical application scenarios 0.1 0-0.3 Local fine features Component edges and joint locations 0.5 0-1.5 Component-level features The overall shape of a single decorative component 2.0 0-6.0 Regional features Component layout within functional zones 5.0 0-15.0 Global distribution characteristics Spatial structure of the entire construction area
[0058] Table 2 lists the spatial feature capture range, feature types, and typical application scenarios corresponding to different scale parameters in multi-scale radial basis function encoding. The spatial position encoder uses the four scale parameters in Table 2 for multi-scale encoding, which can comprehensively capture various spatial features in the decoration scene, from local fine-grained to globally distributed. For different types of decoration projects, the values of the scale parameters can be adjusted according to the project scale to adapt to different spatial ranges.
[0059] In this embodiment, the spatiotemporal feature joint embedding structure maps spatial coordinates, timestamps, and material attributes into feature vectors of a unified dimension through multi-scale radial basis function encoding, periodic trigonometric function encoding, and multilayer perceptron encoding, respectively. A cross-attention fusion operation using a spatial distance attenuation mask is introduced to achieve effective fusion of the three types of features, and the fusion process conforms to the spatial physical attenuation law of the decorative scene. The generated unified representation vector simultaneously contains spatial, attribute, and temporal information, providing an accurate feature foundation for subsequent dynamic knowledge graph construction.
[0060] refer to Figure 4 In a preferred embodiment, the dynamic knowledge graph construction structure divides decorative entities and materials into two levels: region-level nodes and component-level nodes. Region-level nodes correspond to functional zones in the decoration project, including living rooms, bedrooms, kitchens, bathrooms, balconies, etc. Each region-level node contains spatial extent information and overall construction progress information for that functional zone. Component-level nodes correspond to specific decorative components and materials within the functional zones, including walls, floors, ceilings, doors, windows, cabinets, tiles, latex paint, wood flooring, etc. Each component-level node contains the component's geometric parameters, material properties, and construction status information. Region-level nodes and component-level nodes are connected by inclusion edges. The weight of the inclusion edge represents the importance of the component within the region, determined by the component's area proportion and construction complexity.
[0061] The dynamic knowledge graph structure transforms decoration construction procedures into temporal execution edges. These edges connect two component-level nodes with a sequential execution order, indicating that the completion of one component's construction is a prerequisite for the start of the next. For example, a temporal execution edge exists between the wall surface preparation node and the wall latex paint application node, meaning that wall latex paint application can only begin after the wall surface preparation is completed. The dynamic knowledge graph structure also transforms the physical relationships between decoration entities into spatial constraint edges. These edges connect two component-level nodes with a physical connection or positional constraint relationship, indicating that the construction quality of one component affects the construction effect of another. For example, a spatial constraint edge exists between the floor leveling node and the wood flooring installation node, indicating that the flatness of the floor directly affects the quality of the wood flooring installation.
[0062] The dynamic knowledge graph construction structure initializes the feature vectors of region-level and component-level nodes based on the unified representation vector output by the spatiotemporal feature joint embedding structure. For each node, its initial feature vector is equal to the unified representation vector of the corresponding decorative entity or material. The dynamic knowledge graph construction structure constructs the initial adjacency matrix based on temporal execution edges and spatial constraint edges. ,in The total number of nodes. The element values of the initial adjacency matrix. Defined as: if node With nodes If there are temporally executed edges or spatially constrained edges, then ;otherwise Information about relational edges is integrated into the initial feature vector of the node, achieved through feature transfer between region-level nodes and component-level nodes.
[0063] The adjacency matrix elements dynamically change as the construction sensor flow data updates to reflect the topological evolution of the decoration construction process. When the construction sensor flow data indicates that a component-level node has been installed, the dynamic knowledge graph construction structure cuts off all temporal execution edges associated with that component-level node, setting the adjacency matrix element values corresponding to these edges to 0. Simultaneously, it activates the spatial constraint edges between that component-level node and adjacent completed nodes, setting the adjacency matrix element values corresponding to these edges from 0 to 1. The dynamic knowledge graph construction structure injects a state decay factor into this component-level node. This state decay factor reduces the message passing intensity of the node in the dynamic knowledge graph as the construction sequence progresses, concentrating computational resources on unfinished nodes. (State decay factor) The calculation process is represented by the following formula:
[0064]
[0065] in, For nodes In time State decay factor; For nodes The corresponding construction procedure completion time; This is the decay time constant, controlling the rate at which the state decays. When hour, , representing a node When under construction, the message transmission strength is at its maximum; when hour, Over time The increase in decreases exponentially, indicating that the node Construction has been completed, and the intensity of its message transmission is gradually decreasing.
[0066] In this embodiment, the correspondence between node types and edge types in the dynamic knowledge graph is shown in Table 3.
[0067] Table 3. Correspondence between node types and edge types in dynamic knowledge graphs
[0068] Node type edge type Connection Node Type The meaning of edge Dynamic update rules Regional nodes Including relation edges Component-level nodes The area contains components Static, does not change during construction. Component-level nodes Timing execution edge Component-level nodes sequence of processes Cut the corresponding edge after construction is completed. Component-level nodes Spatial constraint edges Component-level nodes Physical association constraints Activate the corresponding edge after adjacent nodes are completed.
[0069] Table 3 lists the edge types, connected node types, edge meanings, and dynamic update rules corresponding to different node types in the dynamic knowledge graph. Edges containing relationships are static edges, remaining unchanged throughout the construction process; temporal execution edges and spatial constraint edges are dynamic edges, their connection states changing dynamically as the construction process progresses. The dynamic knowledge graph construction structure adjusts the adjacency matrix element values in real time according to the update rules in Table 3, ensuring that the dynamic knowledge graph accurately reflects the topological evolution of the construction process.
[0070] In this embodiment, the dynamic knowledge graph construction structure, through hierarchical node design and multi-type edge design, can comprehensively represent various relationships between entities during the decoration construction process. The dynamic update mechanism of the adjacency matrix enables the dynamic knowledge graph to reflect the topological changes during the construction process in real time. The introduction of the state decay factor reduces the message passing intensity of completed nodes, reduces redundant calculations, and improves the system's operating efficiency.
[0071] refer to Figure 5 In a preferred embodiment, the dynamic graph convolutional and temporal prediction structure employs a multi-layer dynamic graph convolutional network to update the node states of the dynamic knowledge graph. Each layer of the dynamic graph convolutional network introduces a temporal gating mechanism during the node state update process, controlling the fusion ratio of historical node states with current input features. The gating value of the temporal gating mechanism is jointly determined by the construction time interval and the node's attribute features. For any given node... Its time step gating value The calculation process is represented by the following formula:
[0072]
[0073]
[0074]
[0075] in, For the current time step Compared to the previous time step Construction sequence intervals between them; The time sensitivity coefficient is calculated from the reciprocal of the construction sequence interval; It is a very small positive number, used to avoid the denominator being zero; For nodes The material activity vector in the attribute features represents the activity level of the material corresponding to the node during the construction process; This refers to the Hadamard product operation; The sigmoid activation function maps the result to... Interval. When the interval of construction time sequence increases and the material activity vector represents the active state of construction, the gate value... Increase the fusion ratio of the current input features in node state updates; when the construction time interval decreases or the material activity vector represents the inactive construction state, the gate value is increased. Reduce the size and retain more historical node states.
[0076] node At time step Post-update status The calculation process is represented by the following formula:
[0077]
[0078]
[0079] in, For nodes In the Output features of layer dynamic graph convolutional networks; For nodes The set of neighboring nodes; For time step Nodes in the adjacency matrix With nodes The element values between; For nodes At time step State decay factor; , For the first Learnable weight matrix and bias vector of layer dynamic graph convolutional network; This represents the total number of layers in the dynamic graph convolutional network. For nodes In the previous time step The node state is determined by a time-gating mechanism, which adaptively balances the contributions of historical and current information during the node state update process, enabling the node state to accurately reflect the dynamic changes in the construction process.
[0080] The dynamic graph convolutional and temporal prediction structure performs graph pooling on all node features updated through a multi-layer dynamic graph convolutional network to obtain a global construction state vector. The graph pooling operation employs an attention pooling mechanism, assigning different pooling weights to each node based on its construction state and importance. Global construction state vector. The calculation process is represented by the following formula:
[0081]
[0082]
[0083]
[0084] in, For nodes At time step Pooling weights; For nodes Importance scores are generated by a small multilayer perceptron. Calculated; This represents the total number of nodes. The attention pooling mechanism can automatically focus on nodes that are of high importance during the construction process, improving the representational ability of the global construction state vector.
[0085] refer to Figure 6 Global construction state vector The data is input into a temporal prediction network, which consists of three parts: a shared bottom layer, a schedule prediction branch, and a loss prediction branch. The shared bottom layer comprises two temporal convolutional layers, used to extract general temporal features from the global construction state vector. The schedule prediction branch and the loss prediction branch each consist of one temporal convolutional layer and one fully connected layer, used to predict construction schedule and material loss, respectively. The temporal prediction network first decouples the features of the global construction state vector, inputting the decoupled temporal dependency features into the schedule prediction branch and the decoupled material association features into the loss prediction branch. The feature decoupling process is represented by the following formula:
[0086]
[0087] in, These are the decoupled temporal dependency features; These are the decoupled material-related features; , These are learnable decoupling weight vectors, used to extract temporal dependency features and material association features, respectively.
[0088] The schedule prediction branch and the loss prediction branch share the underlying temporal convolutional kernel to reduce the number of model parameters and improve the model's generalization ability. Orthogonal constraints are introduced at the output layer to minimize feature interference between the schedule prediction and loss prediction branches. Orthogonal constraint loss function. The calculation process is represented by the following formula:
[0089]
[0090] in, , These are the weight matrices of the output layers for the schedule prediction branch and the loss prediction branch, respectively. It is the identity matrix; The Frobenius norm is used. By minimizing the orthogonal constraint loss function, the output features of the progress prediction branch and the loss prediction branch are made as orthogonal as possible, reducing feature interference between the two branches.
[0091] The progress prediction branch outputs the construction progress time series curve. , indicating the future The predicted construction progress at each time step. The calculation process of the construction progress time series curve is expressed by the following formula:
[0092]
[0093] in, To share the underlying temporal convolutional layers; Temporal convolutional layers for the progress prediction branch; A fully connected layer for the schedule prediction branch; To predict the time step, each element of the construction progress time series curve represents the percentage of construction completed at the corresponding time step, with a value range of [value range missing]. .
[0094] The loss prediction branch outputs a material loss distribution matrix. , indicating the future The predicted loss values of various materials in different regions at each time step. The calculation process of the material loss distribution matrix is expressed by the following formula:
[0095]
[0096] in, # is the temporal convolutional layer for the loss prediction branch; #imgpt168# is the fully connected layer for the loss prediction branch. The rows of the material loss distribution matrix correspond to different material types, and the columns correspond to different region-level nodes. Each element value represents the predicted loss of the corresponding material in the corresponding region.
[0097] In this embodiment, the dynamic graph convolution and temporal prediction structure achieves adaptive node state updates by introducing a temporal gating mechanism, enabling the node states to accurately reflect the dynamic changes in the construction process. The attention pooling mechanism effectively aggregates node features, generating a global construction state vector with strong representational capabilities. The introduction of feature decoupling and orthogonal constraints minimizes feature interference between progress prediction and loss prediction, ensuring the independence of multi-task prediction outputs. The shared underlying temporal convolution kernel design reduces the number of model parameters, improving the model's training efficiency and generalization ability.
Claims
1. An integrated cloud-based system for processing and predicting multi-source heterogeneous data related to decoration, characterized in that: include: The data receiving structure is configured to receive drawing topology data, material attribute data, and construction sensor stream data in the decoration scene; The spatiotemporal feature joint embedding structure is configured to map the spatial coordinate information in the drawing topology data, the physical attribute dimension in the material attribute data, and the timestamp in the construction sensor flow data to the same high-dimensional feature space to complete the spatiotemporal alignment of heterogeneous data, wherein the spatiotemporal embedding dimension corresponds to the physical coordinates of the decorative entity and the progress of the construction stage. The dynamic knowledge graph construction structure is configured to construct a dynamic knowledge graph of the decoration process based on the feature flow after spatiotemporal alignment. The nodes in the dynamic knowledge graph represent decoration entities and materials, and the edges represent construction procedures and physical associations. The dynamic graph convolution and temporal prediction structure is configured to update the node state of the dynamic knowledge graph driven by the construction time sequence, and input the updated graph features into the temporal prediction network to output the joint prediction result of decoration construction progress and material loss. The dynamic update of the graph is synchronously coupled with the physical progress of the construction process.
2. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 1, characterized in that, The data receiving structure is further configured to perform time-scale alignment and spatial benchmark unification on the drawing topology data, the material attribute data, and the construction sensor stream data; extract the absolute coordinate sequence from the drawing topology data; parse the relative timestamps from the construction sensor stream data; anchor the absolute coordinate sequence and the relative timestamps based on the decoration construction process list; and generate a multi-source data stream under a unified spatiotemporal benchmark. The multi-source data stream under the unified spatiotemporal benchmark serves as the input to the spatiotemporal feature joint embedding structure, eliminating sampling frequency differences and clock drift between heterogeneous data sources.
3. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 1, characterized in that, The spatiotemporal feature joint embedding structure is configured to construct a joint embedding network including a spatial location encoder, an attribute encoder, and a temporal encoder. The spatial coordinate information is input into the spatial location encoder to generate a spatial feature vector, the physical attribute dimension is input into the attribute encoder to generate an attribute feature vector, and the timestamp is input into the temporal encoder to generate a temporal feature vector. A cross-attention fusion operation is performed on the spatial feature vector, the attribute feature vector, and the temporal feature vector. The fused high-dimensional vector is used as a unified representation in the high-dimensional feature space. The weight matrix of the cross-attention fusion operation is constrained by the physical coordinates of the decorative entity.
4. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 2, characterized in that, The dynamic knowledge graph construction structure is configured to divide decorative entities and materials into region-level nodes and component-level nodes. The region-level nodes and component-level nodes are connected by inclusion edges. Construction procedures are transformed into temporal execution edges, and physical associations are transformed into spatial constraint edges. The feature vectors of the region-level nodes and component-level nodes are initialized based on the unified representation. An adjacency matrix is constructed based on the temporal execution edges and the spatial constraint edges. The element values of the adjacency matrix change dynamically with the update of the construction sensor flow data to reflect the topological evolution of the decoration construction process.
5. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 1, characterized in that, The dynamic graph convolution and temporal prediction structure is configured to introduce a temporal gating mechanism during the node state update process of the dynamic knowledge graph. The temporal gating mechanism controls the fusion ratio of historical node states and current input features. The gating value of the temporal gating mechanism is jointly determined by the interval of construction time sequence and the attribute features of the node. The node features updated by the temporal gating mechanism are then subjected to graph pooling to obtain a global construction state vector, which is then input into the temporal prediction network.
6. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 5, characterized in that, The time-series prediction network includes a progress prediction branch and a loss prediction branch. The global construction state vector is decoupled in terms of features. The decoupled time-dependent features are input into the progress prediction branch, and the decoupled material-related features are input into the loss prediction branch. The progress prediction branch and the loss prediction branch share the underlying time-series convolution kernels. Orthogonal constraints are introduced in the output layer to minimize the feature interference between the progress prediction branch and the loss prediction branch. The progress prediction branch outputs the construction progress time-series curve, and the loss prediction branch outputs the material loss distribution matrix.
7. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 2, characterized in that, The process of anchoring the absolute coordinate sequence and the relative timestamp based on the decoration construction process list includes: identifying event trigger tags in the construction sensor stream data; matching the event trigger tags with key milestone nodes in the decoration construction process list; performing reverse time-series calibration on the absolute coordinate sequence in the drawing topology data based on the time reference of the successfully matched key milestone nodes; correcting the timestamp offset in the construction sensor stream data caused by transmission delay; and ensuring the consistency of multi-source data streams under the unified spatiotemporal reference at the event trigger level.
8. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 3, characterized in that, The spatial location encoder uses multi-scale radial basis functions to encode the spatial coordinate information, capturing local spatial topological features and global spatial distribution features. The time encoder uses periodic trigonometric function transformation to process the timestamp to extract construction cycle features. The cross-attention fusion operation introduces a spatial distance attenuation mask, which suppresses the attention weight between distant nodes based on the physical distance between decorative entities, so that the unified representation conforms to the spatial physical attenuation law of the decorative scene during the fusion stage.
9. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 4, characterized in that, The process of dynamically changing the element values of the adjacency matrix as the construction sensor flow data is updated includes: when the construction sensor flow data indicates that a certain component-level node has been installed, cutting off the timing execution edge associated with the component-level node, activating the spatial constraint edge between the component-level node and the adjacent completed nodes, and injecting a state decay factor into the component-level node. The state decay factor reduces the message passing strength of the component-level node in the dynamic knowledge graph as the construction timeline progresses, thus concentrating computing resources on unfinished nodes.
10. The integrated decoration multi-source heterogeneous data cloud processing and prediction system according to claim 5, characterized in that, The process of determining the gate value of the time-series gating mechanism by the interval of construction time sequence and the attribute features of nodes includes: calculating the reciprocal of the interval of construction time sequence as a time sensitivity coefficient; extracting the material activity vector from the attribute features of nodes; performing a Hadamard product operation on the time sensitivity coefficient and the material activity vector; mapping the operation result to a gate value through a S-shaped activation function; and increasing the gate value when the interval of construction time sequence increases and the material activity vector represents the construction activity state, so as to improve the fusion ratio of the current input features.
