A building construction progress monitoring method and system based on big data

Abstract

The application relates to the technical field of building construction, in particular to a building construction progress monitoring method and system based on big data, which comprises the following steps: S100, based on an end-edge-cloud collaborative architecture, performing lightweight feature extraction on multi-source heterogeneous original data through an edge computing node arranged in a physical execution scene, generating a structured state vector and uploading the structured state vector to a cloud control layer; S200, acquiring a wide-area environment feature vector, introducing a time decay factor to correct the wide-area environment feature vector in terms of timeliness, performing cross-modal semantic alignment on the corrected wide-area environment feature vector and the structured state vector, and generating a node real-time state perception vector fused with multi-dimensional information. Under the premise of guaranteeing the correctness of physical operation logic, the application breaks through the full-automatic industrial control closed loop from multi-modal data perception, accurate graph network prediction to adaptive scheduling of on-site hardware devices.

Description

Technical Field

[0001] This invention relates to the field of building construction technology, and specifically to a method and system for monitoring building construction progress based on big data. Background Technology

[0002] In large-scale and complex physical execution scenarios (such as large-scale construction projects), the evolution of project nodes is not only controlled by the allocation of machinery and personnel within a local space, but also constrained by dynamic environmental factors such as weather changes, supply chain logistics, and traffic control over a wide area. Existing progress monitoring and dynamic scheduling systems suffer from the following core technical deficiencies when dealing with complex node evolution problems:

[0003] 1. Existing data processing architectures lack deep coupling mechanisms for multimodal sensing data and environmental time-series data. Current systems mostly employ independent visual processing or single sensor threshold alarms, failing to perform unified structured dimensionality reduction on local multi-source heterogeneous data (such as video streams, time-series trajectories, and text logs). Furthermore, when introducing external wide-area environmental data, existing systems typically use a linear processing method of static numerical superposition, ignoring the "time-dependent decay" characteristic of environmental factors such as meteorological or logistical anomalies on the current node's state. This leads to severe semantic misalignment problems during cross-modal feature fusion, resulting in the data source input to the prediction model containing a large amount of temporal noise.

[0004] 2. Existing project schedule network models cannot achieve a joint representation of multi-dimensional physical constraints in the underlying mathematical space. Traditional topology graphs based on the Critical Path Method (CPM) or Project Review and Approval Technique (PERT) can only reflect single temporal logical dependencies between processes, completely stripping away the three-dimensional interference in physical space and resource competition arising from shared hardware within the same time window. When using general graph neural networks (GNNs) to predict schedule delay evolution, the lack of feature weighting of spatial bounding box overlap and resource conflict entropy in the underlying adjacency matrix, coupled with the lack of causal convolution operators and graph Laplace regularization constraints for temporal irreversibility, leads to the model output delay propagation paths frequently violating underlying physical common sense (e.g., incorrect predictions of preceding tasks even before subsequent logical tasks are completed), significantly reducing the engineering usability of the prediction results.

[0005] 3. Existing scheduling systems have not yet formed a closed-loop control link for issuing control commands to physical hardware. After detecting node progress deviations, current technologies typically only generate visual early warning reports or manual scheduling suggestions at the software application layer. They lack a dynamic topology optimization mechanism under hard logic transfer mask constraints, and are unable to serialize the reconstructed resource scheduling requirements into communication protocol commands recognizable by the underlying industrial controllers (such as PLCs or IIoT gateways). This forces the system's data processing flow to terminate at the application layer, making it impossible to directly drive automated access control or anti-collision navigation of tower cranes in physical spaces, and failing to substantially solve the problem of automated closed-loop control from state perception to physical execution in complex engineering projects. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a construction progress monitoring method based on big data, including the following steps: S100, based on the edge-cloud collaborative architecture, lightweight feature extraction is performed on multi-source heterogeneous raw data through edge computing nodes deployed in the physical execution scenario, generating a structured state vector and uploading it to the cloud control layer; S200. Obtain the wide-area environment feature vector, and introduce a time decay factor to correct the timeliness of the wide-area environment feature vector. Align the corrected wide-area environment feature vector with the structured state vector across modal semantics to generate a node real-time state perception vector that integrates multi-dimensional information. S300. Extract the topological attributes corresponding to each subdivided physical execution task in the physical execution scenario and construct a multidimensional composite graph adjacency matrix; wherein, the multidimensional composite graph adjacency matrix is ​​obtained by weighted fusion of the spatial constraint matrix representing physical interference relationship, the logical dependency matrix representing temporal sequence logic, and the resource constraint matrix representing shared hardware devices; S400. The real-time state perception vector of the node is used as the initial node feature and synchronously input with the adjacency matrix of the multidimensional composite graph into the pre-constructed spatiotemporal graph convolutional network for aggregation calculation, and the state offset of each subdivided physical execution task within the prediction window is output. S500. When the state offset is detected to exceed the preset safety threshold, the critical path is dynamically reconstructed based on the heuristic algorithm under the constraint of the adjacency matrix of the multidimensional composite graph, and a resource scheduling control strategy is generated and output to the corresponding physical execution terminal based on the reconstruction result.

[0007] Furthermore, step S100 specifically includes: Unstructured video stream data is collected by visual sensing devices deployed in physical execution scenarios, time-series data of device operation trajectory is collected by IoT sensors, and on-site text log data is collected by mobile terminals. The edge computing node is used to perform modal decomposition on the unstructured video stream data, device operation trajectory time series data and on-site text log data, and map them to preset attribute fields to generate the structured state vector; The specific process of lightweight feature extraction is as follows: A pre-trained lightweight convolutional neural network model is loaded into the edge computing node to perform target detection and key point localization on the unstructured video stream data and extract the geometric feature information of the physical task. The time-series data of the device's operating trajectory are sampled using a sliding window, and statistical features of the device's motion pattern are extracted using a feature mapping function; The geometric feature information and the statistical features are concatenated to construct the structured state vector.

[0008] Furthermore, step S100 also includes a data filtering mechanism at the edge: The edge computing node calculates the Euclidean distance between the structured state vector generated at the current time and the state vector at the previous upload time. Determine whether the Euclidean distance exceeds a preset mutation threshold; If the limit is exceeded, the upload action is performed; if the limit is not exceeded, only the local state cache is updated on the edge side without uploading to the cloud control layer. The structured state vector contains at least the following fields: The node task unique identifier ID, the normalized task completion percentage, the status bit of the current hardware resources occupied, and the spatial coordinates of the node's associated location.

[0009] Furthermore, the specific process of obtaining the wide-area environment feature vector in step S200 is as follows: By using public API interfaces or preset data crawling engines, meteorological early warning data, key raw material supply chain logistics data, and regional traffic control data associated with the physical execution scenario can be obtained in real time. The above data is vectorized using a feature encoder and aligned according to timestamps to construct the wide-area environment feature vector. ; The specific formula for the time decay factor to correct the timeliness of the wide-area environment feature vector is as follows: ; in, This is the corrected environmental feature vector. This is the initial environmental feature vector. The preset attenuation coefficient, This represents the time difference between the time the environmental event occurred and the current sampling time. The specific process of cross-modal semantic alignment is as follows: Map the structured state vector to a query matrix. The corrected wide-area environment feature vectors are mapped to key matrices respectively. Sum matrix ; Calculate the query matrix AND key matrix The dot product is used to obtain the cross-modal attention weight distribution; Using the attention weight distribution to log the value matrix Perform weighted aggregation, and concatenate the aggregation result with the structured state vector to generate the real-time state perception vector of the node; The calculation of the cross-modal attention weight distribution also introduces a spatial masking operator: Based on the spatial coordinates of each subdivided physical execution task, the weight distribution is weighted by spatial distance to enhance the impact weight of local environmental anomalies on node tasks that are physically close.

[0010] Furthermore, the specific process of constructing the adjacency matrix of the multidimensional composite graph in step S300 is as follows: Based on the BIM model, the three-dimensional spatial coordinates and bounding box information of each subdivided physical execution task are extracted to construct the spatial constraint matrix. ; The directed graph of the construction schedule is analyzed to extract the pre- and post-task dependencies and construct the logical dependency matrix. ; Identify the list of key equipment requirements for each task node, extract conflicting node pairs that share the same hardware device, and construct the resource constraint matrix. ; The above matrices are linearly weighted and fused according to preset scene weights to generate the multidimensional composite graph adjacency matrix. ; The spatial constraint matrix The element values ​​are determined according to the following rules: Calculate the Euclidean distance between any two node tasks in three-dimensional space. : like If the physical interference threshold is less than the preset threshold, or if the bounding boxes corresponding to the two node tasks overlap, then the corresponding matrix element is set to the association weight based on the negative correlation of distance; otherwise, it is set to 0. The logical dependency matrix Represented using an asymmetric directed adjacency matrix: If node For nodes The immediate preceding process, then the matrix elements Set as the first preset value, and Set to 0; The first preset value is dynamically assigned based on the urgency threshold of the preceding process for the subsequent process; The resource constraint matrix The construction process includes: Define a resource demand vector to record the occupancy status of tower cranes, construction hoists, and special operation personnel for each node task; Calculate the resource conflict entropy of any two node tasks within the same time window, and fill the correlation weights of the resource constraint matrix according to the magnitude of the conflict entropy; The weight parameters of the weighted fusion Determined in the following ways: Based on the type of work in the current construction phase, the corresponding initial weight values ​​are matched from the preset parameter library; The weight parameters are trained by backpropagation using historical construction delay samples to minimize the fitting error of the adjacency matrix of the multidimensional composite graph to the relationship between nodes.

[0011] Further, in step S400, the spatiotemporal graph convolutional network includes at least one spatial graph convolutional layer. The spatial graph convolutional layer uses the multidimensional composite graph adjacency matrix to perform spatial dimension aggregation calculation on the node features, and its single-layer feature update mechanism satisfies the following formula: ; in, For the first The hidden state matrix of the nodes in the layer, with the real-time state awareness vector of the nodes as the initial input. ; ,in It is a multidimensional composite graph adjacency matrix. It is the identity matrix; for The corresponding degree matrix; For the first The learnable weight parameter matrix of the layer; It is a non-linear activation function; The spatiotemporal graph convolutional network further includes temporal convolutional layers that are stacked alternately with the spatial graph convolutional layers: The temporal convolutional layer employs a one-dimensional causal convolution operator to extract local features from the hidden state sequence of the same physical execution task along the time step dimension. The one-dimensional causal convolution operator at any time The receptive field only covers a time interval The historical state sequence within, where This is the preset size of the convolution kernel window.

[0012] Furthermore, the output of the state offset of each of the subdivided physical execution tasks within the prediction window is specifically manifested as the output of a two-dimensional joint prediction vector. : in, This represents the absolute time offset of a node task from the baseline timeline within a given prediction window. This represents the confidence probability of the absolute time offset value occurring.

[0013] The training process of the pre-constructed spatiotemporal graph convolutional network includes: Obtain a sample set containing historical node states and actual offset results for end-to-end training; The prediction error is calculated using a joint loss function, which includes a mean squared error loss term to constrain the accuracy of time offset prediction, and a graph Laplacian regularization term to maintain the smoothness of the composite topology. Based on the joint loss function, the weight parameters in the spatiotemporal graph convolutional network are updated using the backpropagation algorithm.

[0014] Further, in step S500, the preset safety threshold is a dynamically calculated floating threshold, and its determination process includes: Obtain the global available time buffer of the current critical path in the adjacency matrix of the multidimensional composite graph; Based on the global available time buffer and the resource scarcity of the current node, the dynamic security threshold corresponding to each subdivided physical execution task is calculated in real time using the exponential smoothing method. When the absolute time offset value output by the spatiotemporal graph convolutional network is greater than the dynamic safety threshold, the graph topology dynamic reconstruction mechanism is triggered. The heuristic algorithm-based topology dynamic reconstruction of the critical path under the constraints of the adjacency matrix of the multidimensional composite graph specifically includes: An improved multi-objective ant colony optimization algorithm is adopted, with the minimum global total delay time after reconstruction and the minimum conflict entropy in the resource constraint matrix as the joint optimization objectives; During the optimization iteration process of the ant colony optimization algorithm, the logical dependency matrix is... As a hard state transition constraint, the state transition probability of an individual ant is forced to zero on paths that violate the logic of temporal sequence. Based on the Pareto optimal solution set obtained through optimization, the edge connection weights and execution timing in the adjacency matrix of the multidimensional composite graph are updated to complete the dynamic topology reconstruction.

[0015] Furthermore, based on the reconstruction results, a resource scheduling control policy is generated and output to the corresponding physical execution terminal, specifically including: The adjacency matrix of the reconstructed multidimensional composite graph is analyzed, and the node tasks that have undergone temporal changes or resource allocation are extracted as target nodes. The resource demand change vector corresponding to the target node is compiled into a machine-readable industrial control instruction set; The control command set is sent to the automated access control, intelligent tower crane control system or material sorting terminal in the physical execution scenario through the industrial IoT gateway, so as to automatically trigger the corresponding physical space locking, resource release or path navigation.

[0016] On the other hand, a construction progress monitoring system based on big data includes: an edge cloud data acquisition module and a multimodal semantic alignment module, a composite graph matrix construction module, a spatiotemporal graph prediction module, and an adaptive reconstruction control module.

[0017] Beneficial effects

[0018] This invention overcomes the technical defects of existing neural network predictions that violate temporal and physical common sense by constructing a multidimensional composite graph adjacency matrix that integrates spatial interference, logical dependence and resource constraints at the bottom layer and introducing it into a spatiotemporal graph convolutional network containing a one-dimensional causal convolution operator, and realizes high-confidence quantitative prediction of the time offset of engineering nodes.

[0019] This invention introduces a time decay factor and a cross-modal attention mechanism to align features between wide-area environmental data and local state data. It also utilizes a multi-objective improved ant colony algorithm with hard constraints imposed by logical masks to generate a topology reconstruction strategy. Finally, the strategy is compiled into a set of machine instructions that can be directly called by the underlying industrial gateway. Thus, while ensuring the correctness of the physical operation logic, it opens up a fully automated industrial control closed loop from multimodal data perception and accurate graph network prediction to adaptive scheduling of on-site hardware equipment. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a diagram of the edge-cloud collaborative system architecture of the present invention; Figure 3 This is a schematic diagram illustrating the principle of constructing the adjacency matrix of the multidimensional composite graph of the present invention. Figure 4 This is a derivation diagram of the spatiotemporal graph convolutional network of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings: Example

[0024] like Figures 1-4 As shown, a construction progress monitoring method based on big data includes the following steps: S100, based on the edge-cloud collaborative architecture, performs lightweight feature extraction on multi-source heterogeneous raw data through edge computing nodes deployed in physical execution scenarios, generates structured state vectors, and uploads them to the cloud control layer; S200. Obtain the wide-area environment feature vector and introduce a time decay factor to correct the timeliness of the wide-area environment feature vector. Align the corrected wide-area environment feature vector with the structured state vector across modal semantics to generate a node real-time state perception vector that integrates multi-dimensional information. S300. Extract the topological attributes corresponding to each subdivided physical execution task in the physical execution scenario and construct a multidimensional composite graph adjacency matrix. The multidimensional composite graph adjacency matrix is ​​obtained by weighted fusion of the spatial constraint matrix representing physical interference relationship, the logical dependency matrix representing temporal order logic, and the resource constraint matrix representing shared hardware devices. S400: The real-time state perception vector of the node is used as the initial node feature and synchronously input with the adjacency matrix of the multidimensional composite graph into the pre-constructed spatiotemporal graph convolutional network for aggregation calculation, and the state offset of each subdivided physical execution task within the prediction window is output. S500 When the detected state offset exceeds the preset safety threshold, the critical path is dynamically reconstructed based on the heuristic algorithm under the constraint of the adjacency matrix of the multidimensional composite graph, and the resource scheduling control strategy is generated and output to the corresponding physical execution terminal based on the reconstruction result.

[0025] Furthermore, the specific implementation process of step S100 is as follows: In the specific implementation steps, edge computing nodes and multi-source data acquisition devices are pre-deployed within the physical execution scenario. The data acquisition devices acquire multi-source heterogeneous raw data of the physical entities in real time. This multi-source heterogeneous raw data consists of unstructured video stream data. Equipment operation trajectory time series data and on-site text log data Composition. Edge computing nodes establish physical connections with data acquisition devices through local area communication protocols, receiving and buffering the aforementioned multi-source heterogeneous raw data.

[0026] Combined with appendix Figure 2 As shown, the edge-cloud collaborative architecture of this invention encompasses a cloud control layer (Cloud), an edge computing layer (Edge), and a physical execution layer (Device / Terminal). Specifically, the data acquisition devices include video surveillance (Cam), IoT sensors (Sen), and system log input terminals; the physical execution terminals include automated access control systems, intelligent tower cranes, or industrial PLC control terminals. Furthermore, the edge computing layer includes a local feature cache M1 for temporarily storing the current structured state vector at the edge and comparing it with historical state vectors to measure mutations.

[0027] Edge computing nodes have built-in ARM-based microprocessors and hardware acceleration modules, and load pre-trained lightweight convolutional neural network models and temporal feature mapping functions into local main memory.

[0028] Edge computing nodes receive unstructured video stream data Then, a keyframe image matrix is ​​extracted according to a preset sampling frame rate. This keyframe image matrix is ​​input into a lightweight convolutional neural network model, where depthwise separable convolutional layers extract local texture and boundary features, outputting the geometric feature information of the physical task object. Geometric feature information The calculation process includes using a regression branch to output the three-dimensional coordinates of the center point of the target bounding box. and length, width and height dimensions In mathematical terms, if the network mapping function is... Then the visual geometric features are represented as ,in Preset visual feature dimensions.

[0029] For equipment operation trajectory time series data The edge computing node initiates the sliding window sampling mechanism. Let the current time be... Set a sliding window with a fixed time span. Extract the timing signal sequence within the window. ,in Representative moment The device's sensor physical parameters. Edge computing nodes call feature mapping functions to calculate the statistical features of the window sequence and extract the mean parameter. With variance parameter This characterizes the equipment's motion mode. The specific calculation formula is as follows: ; ; The above statistical features are vectorized and concatenated to generate a time-series feature vector. Meanwhile, regarding the on-site text log data... The word embedding algorithm is used to extract entity word vectors associated with the physical entity state. .

[0030] After completing the feature extraction of the multimodal data, the edge computing node allocates contiguous storage space in the memory control unit to store the visual geometric feature information. Temporal feature vectors and entity word vectors By performing dimensionality reduction and residual concatenation with preset fixed attribute fields, a structured state vector is constructed. Structured state vector The data structure is forcibly defined as a one-dimensional tensor, and the internal fields are included in a fixed byte order: a node task unique identifier ID field, and a mapping to a range. The fields include the normalized task completion percentage field, the hardware resource occupancy status field represented by one-hot encoding, and the spatial location coordinate field associated with the node.

[0031] To constrain the data throughput of the edge communication link, the edge computing node generates the current time... Structured state vector Then, the data filtering mechanism on the edge side is triggered. The processor reads the previous upload time from the cache. Historical structured state vector And calculate the Euclidean distance between the two in the high-dimensional feature space. As a measure of mutation, the Euclidean distance is calculated using the following formula: ; in, The total feature dimension of the structured state vector. For the current vector, the first... dimensional components, The first of the history vectors dimensional components, The preset penalty weight coefficients for the corresponding feature dimensions.

[0032] The edge computing nodes will calculate the Euclidean distance. Compared with the preset mutation threshold stored in the local register The comparison is performed. The conditional branch decision logic is as follows: like The edge computing node determines that the state of the current physical execution scenario has undergone a substantial evolution, and transmits the current structured state vector through the wide area network interface. Encapsulated into TCP / IP packets, uploaded to the cloud control layer, and simultaneously... Write the overwrite to the local cache as the new benchmark value; if If the physical state is determined to be in a steady state range, the processor only updates the local state cache record at the edge and discards the upload process instruction.

[0033] Furthermore, the specific implementation process of step S200 is as follows: In the actual execution process, the cloud control layer asynchronously calls the APIs of the meteorological monitoring system, supply chain logistics management platform, and regional traffic control center through a preset external data gateway interface. The acquired raw wide-area environmental data includes real-time precipitation, average wind speed, GPS latitude and longitude coordinates and arrival times of key material vehicles, and traffic congestion coefficients around the construction area. The cloud control layer uses a multi-dimensional feature encoder to vectorize the above non-numerical and discrete data, constructing an initial wide-area environmental feature vector. ,in denoted as the dimension of the feature space.

[0034] Considering the persistence and time-dependent effects of environmental factors on the state of physical nodes, the system introduces a time decay factor. Dynamic corrections are performed. The cloud control layer extracts the timestamps of environmental events from the data packet headers. And calculate its relationship with the current system sampling time. Time difference The system calls the exponential decay function to calculate the corrected environmental feature vector. The calculation formula is as follows: ; in, This is a preset attenuation coefficient, the value of which is obtained by retrieving it from the parameter table according to the category of environmental factor. For example, the attenuation coefficient for meteorological factors is significantly smaller than that for traffic control factors, to reflect the lag between water accumulation elimination and traffic diversion in the physical scenario.

[0035] Obtaining the corrected wide-area environment feature vector Then, the system initiates a cross-modal semantic alignment process based on a scaled dot product attention mechanism. First, the system uses three independent learnable linear transformation matrices... Project the input data. Specifically, project the structured state vector generated in step S100. Mapped to query matrix The corrected wide-area environment feature vector Mapped to key matrices respectively Sum matrix : ; Subsequently, the system calculates and The transpose of the dot product is used to measure the strength of the association between physical node states and environmental features in the semantic space, and a scaling factor is introduced. To prevent gradient vanishing, the attention weight distribution is obtained by normalization using the Softmax function. : ;

[0036] Based on this, the system will distribute the weights. AND-value matrix By performing a weighted summation, we obtain the aligned feature vector of the environment mode under the current node state constraint. .

[0037] To maintain the stability of feature evolution, the system will align feature vectors. With the original structured state vector Residual connections are performed, and the final real-time state awareness vector of the nodes is generated through layer normalization. Its mapping logic is defined as follows: ; Among them, among them, Represents a linear mapping layer. The representation layer normalization function is used to adjust the aligned high-dimensional feature dimensions to match the normalization process. Consistent. The generated node real-time state awareness vector. It is stored in a dynamic memory buffer and used as the input feature matrix for subsequent spatiotemporal graph neural network computation units.

[0038] Furthermore, the specific implementation process of step S300 is as follows: The system first defines a complex graph structure to represent the relationships between physical execution tasks. , where the set of nodes This corresponds to the various sub-tasks performed in the physical scene. To accurately characterize the interactions between nodes, the system constructs a multi-dimensional composite graph adjacency matrix. The construction of this matrix is ​​not based on a single topological connection, but rather on a spatial constraint matrix. Logical dependency matrix and resource constraint matrix Perform heterogeneous fusion.

[0039] First, the system extracts the three-dimensional geometric center coordinates of each subdivided physical execution task by parsing the pre-stored BIM (Building Information Modeling) data. And the 3D bounding box parameters. Spatial constraint matrix. elements in Used to measure nodes With nodes Euclidean distance between If the distance is less than the preset physical interference radius... If the bounding boxes of two nodes overlap in any dimension, then the value of the element is calculated using the following Gaussian kernel function: ; If the interference condition is not met, then Through this calculation, the matrix It can capture the cascading compression effect of neighboring nodes in physical space when displacement or delay occurs.

[0040] Secondly, the system parses the PERT (Plan-Do-Check-Act) diagram in the construction schedule, extracts the directed dependencies between task nodes, and constructs a logical dependency matrix. Because the process has irreversible temporal characteristics, the matrix... It is an asymmetric matrix. If the nodes... For nodes The immediate preceding process will then use the matrix elements. Assign weight values This weight value is positively correlated with the urgency of the process. The calculation logic is as follows: ; in, For nodes Free time difference, The total project duration is represented by [value]. This design ensures that logical connections on the critical path have higher numerical significance in the matrix.

[0041] Next, the system identifies the resource constraint matrix for each task node's allocation plan for large hardware equipment (such as tower cranes and construction elevators) and special work teams. If node With nodes The same physical resources need to be shared within the same time window. If so, then there is an implicit relationship between the two. The system defines resource conflict entropy. Determine matrix elements : ; in, To compete for resources in the same time period The total number of tasks. This matrix reveals the state coupling phenomenon caused by resource contention among spatially non-adjacent nodes.

[0042] Finally, the system employs a weighted fusion strategy to generate a multidimensional composite graph adjacency matrix. The cloud-based control layer adjusts the weighting coefficients based on the current construction stage (such as the civil engineering stage or the decoration stage). , , Perform the following fusion operation: ; To ensure the convergence of the subsequent graph convolutional neural network, the system performs Laplacian normalization on the fused matrix. First, the degree matrix is ​​calculated. diagonal elements This leads to the standardized symmetric adjacency matrix: ; in It is the identity matrix. This normalized matrix... As a bottom-level topological constraint, it will be synchronously input into the prediction model along with the perception vector output from step S200.

[0043] Furthermore, the specific implementation process of step S400 is as follows: In the specific implementation process, the system instantiates a multi-layered stacked spatiotemporal graph convolutional network model. Let the current time step be... The system will output the node real-time state perception vector in step S200. As the first layer input feature matrix of the network, i.e. ,in To further break down the total number of nodes physically executing tasks, This represents the feature dimension of the state-aware vector. Simultaneously, the normalized multidimensional composite graph adjacency matrix generated in step S300 is... It is solidified into network topology constraint parameters.

[0044] Combined with appendix Figure 4 As shown in the inference diagram of the spatiotemporal graph convolutional network, the initial input of the perceptual vector in the first layer of the network is labeled Input1. Inside the network, the spatiotemporal graph convolutional units alternately stack spatial graph convolutional layers (Spatial GCN) and one-dimensional causal temporal convolutional layers (1D Causal TCN), and the node hidden state matrix output between layers is labeled Hl (where l represents the number of feature update layers). After multi-layer feature extraction, the hidden states at multiple time steps are compressed through a global average pooling layer (GAP) and finally connected to a fully connected output layer (FC) for mapping, so as to output the absolute time offset value ΔT and confidence probability P of each subdivided physical execution task within the prediction window.

[0045] Each basic computational unit of a spatiotemporal graph convolutional network consists of a spatial graph convolutional layer and a temporal convolutional layer. In the spatial feature aggregation stage, the... Layer space graph convolution utilization Perform feature smoothing and mapping within the local receptive field. The forward propagation calculation formula is as follows: ; in, No. The learnable parameter matrix of layer-space graph convolution. It is a nonlinear activation function (the ReLU function is used in the specific implementation). The physical meaning of this spatial aggregation operation is: to force each task node to absorb the current state of its neighboring nodes that interfere with it in three-dimensional space, have hard dependencies in temporal logic, and compete for equipment resources, thereby realizing the spatial cascading transmission simulation of delay risk at the underlying data structure.

[0046] After completing single-frame spatial aggregation, the network performs temporal convolution along the time dimension. Considering the strict irreversibility of time passage in the physical execution scenario, the system employs a one-dimensional causal convolution operator to prevent data from future moments from leaking into the current receptive field. For nodes... At time step The characteristic sequence, the first The computational logic of multi-layer temporal convolution is expressed as follows: ; in, The window size of the temporal convolution kernel. These are the learnable weights for one-dimensional causal convolution. This operator ensures the learnable weights of the nodes. At any moment The hidden state is determined only by the interval It evolved from the historical state within.

[0047] After alternating extraction using multiple layers of spatial graph convolution followed by one-dimensional causal temporal convolution, the system compresses the hidden states across multiple time steps using a global average pooling layer and then connects them to the prediction output, which consists of two fully connected layers. For any node in the network... The predicted output is mapped to generate a two-dimensional joint prediction vector. .in, As a continuous variable, it represents the absolute time offset of the task from the baseline schedule within a preset forecast window (such as the next 7 days); Discrete probability mapping values ​​(normalized to the Sigmoid function) (Interval), representing the confidence level of the occurrence of an offset value.

[0048] During the training phase of the pre-built spatiotemporal graph convolutional network, the system adopts an end-to-end supervised learning paradigm. The cloud-based control layer acquires a dataset containing historical environmental fluctuations and real-world engineering delay records as training samples. To balance the accuracy of time prediction with the physical stability of the graph topology when optimizing parameters, the system defines a joint loss function. It consists of the mean square error loss term. Cross-entropy loss term and the Thulaplace regularization term composition: ; ; in, The true offset label of the sample. Let be the hidden representation matrix of the last layer of the network. It is the Laplace matrix of the composite graph. The introduction of this as a hard penalty term forces highly interconnected groups of nodes in a composite topology graph to output smooth predicted states, avoiding isolated mutations that violate physical laws. It is the Laplacian matrix of the composite graph.

[0049] The processor uses the backpropagation algorithm to calculate the gradient of the joint loss function with respect to the network weight parameters, and iteratively updates it using the Adam optimizer. and This continues until the loss function converges to its minimum value.

[0050] Furthermore, the specific implementation process of step S500 is as follows: The cloud control layer receives the two-dimensional joint prediction vectors of each subdivided physical execution task output in step S400. Then, a dynamic security threshold determination mechanism is initiated. The system does not rely on fixed empirical parameters, but rather determines the threshold by parsing the adjacency matrix of the current multidimensional composite graph. Extracting nodes from the corresponding directed acyclic graph. The system calculates the global available time buffer size using exponential smoothing, combined with the historical variance of the node. Dynamic security threshold : ; in This is the smoothing coefficient. When the system detects the predicted absolute time offset value... And its confidence probability When the probability exceeds the preset baseline, the graph topology dynamic reconstruction procedure is triggered.

[0051] The system employs an improved multi-objective ant colony optimization algorithm to reconstruct the topology of local execution sequences. The system initializes several artificial ant individuals to search for alternative paths in parallel within the graph network. During the state transitions of the ant individuals, the system converts the logical dependency matrix generated in step S300... Convert to hard state transition mask matrix For any node To the node transition probability Its calculation formula is strictly redefined as: ; in, For pheromone concentration, (The heuristic factor is negatively correlated with the resource conflict entropy between the two nodes). For ants The set of feasible nodes. The value selection logic is limited by :like The node is indicated in the middle. Must be on the node If executed previously, then forced ,otherwise This hard constraint mechanism eliminates all optimization paths that violate the logical sequence at the underlying mathematical model level. After multiple iterations, the algorithm outputs a set of Pareto optimal topological sequence solutions, with the joint optimization objective being to minimize the total global delay time. Minimize global resource conflict entropy The system selects the optimal solution to update. By determining the edge connection weights and execution timing, topology reconstruction is completed.

[0052] After the topology reconfiguration is complete, the cloud control layer initiates a physical-layer resource scheduling and control policy transition. The system then analyzes the reconfigured... The system extracts the set of target nodes that cause timing changes and resource reallocation. For this set of target nodes, the system inputs the changed equipment occupancy demand vector (such as the intervention time of a specific tower crane or the access control duration for special operations teams) into the underlying compiler. Based on the communication specifications of programmable logic controllers (PLCs), the compiler serializes the demand vector into a machine-readable industrial control instruction set conforming to the Modbus TCP or OPCUA protocol. Finally, the cloud control layer directly distributes the industrial control instruction set to the corresponding physical execution terminals through industrial IoT gateways deployed in the physical field. For example, the instruction set triggers the electromagnetic relay state toggles in an automated access control system to lock or release specific personnel, or writes the path's three-dimensional coordinates into the servo motor register of the intelligent tower crane control system to execute anti-collision path navigation, thereby achieving an industrial-grade closed loop from data prediction to adaptive control of the physical space.

[0053] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring construction progress based on big data, characterized in that, Includes the following steps: S100, based on the edge-cloud collaborative architecture, performs lightweight feature extraction on multi-source heterogeneous raw data through edge computing nodes deployed in physical execution scenarios, generates structured state vectors, and uploads them to the cloud control layer; S200. Obtain the wide-area environment feature vector, and introduce a time decay factor to correct the timeliness of the wide-area environment feature vector. Align the corrected wide-area environment feature vector with the structured state vector across modal semantics to generate a node real-time state perception vector that integrates multi-dimensional information. S300. Extract the topological attributes corresponding to each subdivided physical execution task in the physical execution scenario and construct a multidimensional composite graph adjacency matrix; wherein, the multidimensional composite graph adjacency matrix is ​​obtained by weighted fusion of the spatial constraint matrix representing physical interference relationship, the logical dependency matrix representing temporal sequence logic, and the resource constraint matrix representing shared hardware devices; S400. The real-time state perception vector of the node is used as the initial node feature and synchronously input with the adjacency matrix of the multidimensional composite graph into the pre-constructed spatiotemporal graph convolutional network for aggregation calculation, and the state offset of each subdivided physical execution task within the prediction window is output. S500. When the state offset is detected to exceed the preset safety threshold, the critical path is dynamically reconstructed based on the heuristic algorithm under the constraint of the adjacency matrix of the multidimensional composite graph, and a resource scheduling control strategy is generated and output to the corresponding physical execution terminal based on the reconstruction result.

2. The construction progress monitoring method based on big data according to claim 1, characterized in that, Step S100 specifically includes: Unstructured video stream data is collected by visual sensing devices deployed in physical execution scenarios, time-series data of device operation trajectory is collected by IoT sensors, and on-site text log data is collected by mobile terminals. The edge computing node is used to perform modal decomposition on the unstructured video stream data, device operation trajectory time series data and on-site text log data, and map them to preset attribute fields to generate the structured state vector; The specific process of lightweight feature extraction is as follows: A pre-trained lightweight convolutional neural network model is loaded into the edge computing node to perform target detection and key point localization on the unstructured video stream data and extract the geometric feature information of the physical task. The time-series data of the device's operating trajectory are sampled using a sliding window, and statistical features of the device's motion pattern are extracted using a feature mapping function; The geometric feature information and the statistical features are concatenated to construct the structured state vector.

3. The construction progress monitoring method based on big data according to claim 2, characterized in that, Step S100 also includes a data filtering mechanism on the edge side: The edge computing node calculates the Euclidean distance between the structured state vector generated at the current time and the state vector at the previous upload time. Determine whether the Euclidean distance exceeds a preset mutation threshold; If the limit is exceeded, the upload action is performed; if the limit is not exceeded, only the local state cache is updated on the edge side without uploading to the cloud control layer. The structured state vector contains at least the following fields: The node task unique identifier ID, the normalized task completion percentage, the status bit of the current hardware resources occupied, and the spatial coordinates of the node's associated location.

4. The construction progress monitoring method based on big data according to claim 3, characterized in that, The specific process of obtaining the wide-area environment feature vector in step S200 is as follows: By using public API interfaces or preset data crawling engines, meteorological early warning data, key raw material supply chain logistics data, and regional traffic control data associated with the physical execution scenario can be obtained in real time. The above data is vectorized using a feature encoder and aligned according to timestamps to construct the wide-area environment feature vector. ; The specific formula for the time decay factor to correct the timeliness of the wide-area environment feature vector is as follows: ； in, This is the corrected environmental feature vector. This is the initial environmental feature vector. The preset attenuation coefficient, This represents the time difference between the time the environmental event occurred and the current sampling time. The specific process of cross-modal semantic alignment is as follows: Map the structured state vector to a query matrix. The corrected wide-area environment feature vectors are mapped to key matrices respectively. Sum matrix ; Calculate the query matrix AND key matrix The dot product is used to obtain the cross-modal attention weight distribution; Using the attention weight distribution to log the value matrix Perform weighted aggregation, and concatenate the aggregation result with the structured state vector to generate the real-time state perception vector of the node; The calculation of the cross-modal attention weight distribution also introduces a spatial masking operator: Based on the spatial coordinates of each subdivided physical execution task, the weight distribution is weighted by spatial distance to enhance the impact weight of local environmental anomalies on node tasks that are physically close.

5. The construction progress monitoring method based on big data according to claim 4, characterized in that, The specific process of constructing the adjacency matrix of the multidimensional composite graph in step S300 is as follows: Based on the BIM model, the three-dimensional spatial coordinates and bounding box information of each subdivided physical execution task are extracted to construct the spatial constraint matrix. ; Analyze the directed graph of the construction schedule, extract the pre- and post-task dependencies, and construct the logical dependency matrix. ; Identify the list of key equipment requirements for each task node, extract conflicting node pairs that share the same hardware device, and construct the resource constraint matrix. ; The above matrices are linearly weighted and fused according to preset scene weights to generate the multidimensional composite graph adjacency matrix. ; The spatial constraint matrix The element values ​​are determined according to the following rules: Calculate the Euclidean distance between any two node tasks in three-dimensional space. : like If the physical interference threshold is less than the preset threshold, or if the bounding boxes corresponding to the two node tasks overlap, then the corresponding matrix element is set to the association weight based on the negative correlation of distance; otherwise, it is set to 0. The logical dependency matrix Represented using an asymmetric directed adjacency matrix: If node For nodes The immediate preceding process, then the matrix elements Set as the first preset value, and Set to 0; The first preset value is dynamically assigned based on the urgency threshold of the preceding process for the subsequent process; The resource constraint matrix The construction process includes: Define a resource demand vector to record the occupancy status of tower cranes, construction hoists, and special operation personnel for each node task; Calculate the resource conflict entropy between any two node tasks within the same time window, and fill the correlation weights of the resource constraint matrix according to the magnitude of the conflict entropy; The weight parameters of the weighted fusion , , Determined in the following ways: Based on the type of work in the current construction phase, the corresponding initial weight values ​​are matched from the preset parameter library; The weight parameters are trained by backpropagation using historical construction delay samples to minimize the fitting error of the adjacency matrix of the multidimensional composite graph to the relationship between nodes.

6. The construction progress monitoring method based on big data according to claim 5, characterized in that, In step S400, the spatiotemporal graph convolutional network includes at least one spatial graph convolutional layer. The spatial graph convolutional layer uses the adjacency matrix of the multidimensional composite graph to perform spatial dimension aggregation calculation on the node features. Its single-layer feature update mechanism satisfies the following formula: ； in, For the first The hidden state matrix of the nodes in the layer, with the real-time state awareness vector of the nodes as the initial input. ; ,in It is a multidimensional composite graph adjacency matrix. It is the identity matrix; for The corresponding degree matrix; For the first The learnable weight parameter matrix of the layer; It is a non-linear activation function; The spatiotemporal graph convolutional network further includes temporal convolutional layers that are stacked alternately with the spatial graph convolutional layers: The temporal convolutional layer employs a one-dimensional causal convolution operator to extract local features from the hidden state sequence of the same physical execution task along the time step dimension. The one-dimensional causal convolution operator at any time The receptive field only covers a time interval The historical state sequence within, where This is the preset size of the convolution kernel window.

7. The construction progress monitoring method based on big data according to claim 6, characterized in that, The output represents the state offset of each of the subdivided physical execution tasks within the prediction window, specifically manifested as an output two-dimensional joint prediction vector. : in, This represents the absolute time offset of a node task from the baseline timeline within a given prediction window. This represents the confidence probability of the absolute time offset value occurring; The training process of the pre-constructed spatiotemporal graph convolutional network includes: Obtain a sample set containing historical node states and actual offset results for end-to-end training; The prediction error is calculated using a joint loss function, which includes a mean squared error loss term to constrain the accuracy of time offset prediction, and a graph Laplacian regularization term to maintain the smoothness of the composite topology. Based on the joint loss function, the weight parameters in the spatiotemporal graph convolutional network are updated using the backpropagation algorithm.

8. A construction progress monitoring method based on big data according to claim 7, characterized in that, In step S500, the preset safety threshold is a dynamically calculated floating threshold, and its determination process includes: Obtain the global available time buffer of the current critical path in the adjacency matrix of the multidimensional composite graph; Based on the global available time buffer and the resource scarcity of the current node, the dynamic security threshold corresponding to each subdivided physical execution task is calculated in real time using the exponential smoothing method. When the absolute time offset value output by the spatiotemporal graph convolutional network is greater than the dynamic safety threshold, the graph topology dynamic reconstruction mechanism is triggered. The heuristic algorithm-based topology dynamic reconstruction of the critical path under the constraints of the adjacency matrix of the multidimensional composite graph specifically includes: An improved multi-objective ant colony optimization algorithm is adopted, with the minimum global total delay time after reconstruction and the minimum conflict entropy in the resource constraint matrix as the joint optimization objectives; During the optimization iteration process of the ant colony optimization algorithm, the logical dependency matrix is... As a hard state transition constraint, the state transition probability of an individual ant is forced to zero on paths that violate the logic of temporal sequence. Based on the Pareto optimal solution set obtained through optimization, the edge connection weights and execution timing in the adjacency matrix of the multidimensional composite graph are updated to complete the dynamic topology reconstruction.

9. A construction progress monitoring method based on big data according to claim 8, characterized in that, Based on the reconstruction results, a resource scheduling and control policy is generated and output to the corresponding physical execution terminal, specifically including: The reconstructed multidimensional composite graph adjacency matrix is ​​analyzed, and the node tasks that have undergone temporal changes or resource allocation are extracted as target nodes. The resource demand change vector corresponding to the target node is compiled into a machine-readable industrial control instruction set; The control command set is sent to the automated access control, intelligent tower crane control system or material sorting terminal in the physical execution scenario through the industrial IoT gateway, so as to automatically trigger the corresponding physical space locking, resource release or path navigation.

10. A construction progress monitoring system based on big data, used to implement the construction progress monitoring method based on big data as described in any one of claims 1-9, characterized in that, include: The module includes edge cloud data acquisition, multimodal semantic alignment, composite graph matrix construction, spatiotemporal graph prediction, and adaptive reconstruction control.

Images